add qwen3
This commit is contained in:
Chranos
0
vllm-v0.6.2/vllm/model_executor/layers/__init__.py
Normal file
0
vllm-v0.6.2/vllm/model_executor/layers/__init__.py
Normal file
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
302
vllm-v0.6.2/vllm/model_executor/layers/activation.py
Normal file
302
vllm-v0.6.2/vllm/model_executor/layers/activation.py
Normal file
@@ -0,0 +1,302 @@
|
||||
"""Custom activation functions."""
|
||||
import math
|
||||
from typing import Optional
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
|
||||
from vllm.distributed import (divide, get_tensor_model_parallel_rank,
|
||||
get_tensor_model_parallel_world_size)
|
||||
from vllm.model_executor.custom_op import CustomOp
|
||||
from vllm.model_executor.utils import set_weight_attrs
|
||||
from vllm.utils import LazyDict
|
||||
|
||||
|
||||
@CustomOp.register("fatrelu_and_mul")
|
||||
class FatreluAndMul(CustomOp):
|
||||
"""An activation function for FATReLU.
|
||||
|
||||
The function computes x -> FATReLU(x[:d]) * x[d:] where
|
||||
d = x.shape[-1] // 2.
|
||||
This is used in openbmb/MiniCPM-S-1B-sft.
|
||||
|
||||
Shapes:
|
||||
x: (num_tokens, 2 * d) or (batch_size, seq_len, 2 * d)
|
||||
return: (num_tokens, d) or (batch_size, seq_len, d)
|
||||
"""
|
||||
|
||||
def __init__(self, threshold: float = 0.):
|
||||
super().__init__()
|
||||
self.threshold = threshold
|
||||
|
||||
def forward_native(self, x: torch.Tensor) -> torch.Tensor:
|
||||
d = x.shape[-1] // 2
|
||||
x1 = x[..., :d]
|
||||
x2 = x[..., d:]
|
||||
x1 = F.threshold(x1, self.threshold, 0.0)
|
||||
return x1 * x2
|
||||
|
||||
def forward_cuda(self, x: torch.Tensor) -> torch.Tensor:
|
||||
from vllm import _custom_ops as ops
|
||||
|
||||
d = x.shape[-1] // 2
|
||||
output_shape = (x.shape[:-1] + (d, ))
|
||||
out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
|
||||
ops.fatrelu_and_mul(out, x, self.threshold)
|
||||
return out
|
||||
|
||||
|
||||
@CustomOp.register("silu_and_mul")
|
||||
class SiluAndMul(CustomOp):
|
||||
"""An activation function for SwiGLU.
|
||||
|
||||
The function computes x -> silu(x[:d]) * x[d:] where d = x.shape[-1] // 2.
|
||||
|
||||
Shapes:
|
||||
x: (num_tokens, 2 * d) or (batch_size, seq_len, 2 * d)
|
||||
return: (num_tokens, d) or (batch_size, seq_len, d)
|
||||
"""
|
||||
|
||||
def forward_native(self, x: torch.Tensor) -> torch.Tensor:
|
||||
"""PyTorch-native implementation equivalent to forward()."""
|
||||
d = x.shape[-1] // 2
|
||||
return F.silu(x[..., :d]) * x[..., d:]
|
||||
|
||||
def forward_cuda(self, x: torch.Tensor) -> torch.Tensor:
|
||||
from vllm import _custom_ops as ops
|
||||
|
||||
d = x.shape[-1] // 2
|
||||
output_shape = (x.shape[:-1] + (d, ))
|
||||
out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
|
||||
ops.silu_and_mul(out, x)
|
||||
return out
|
||||
|
||||
def forward_xpu(self, x: torch.Tensor) -> torch.Tensor:
|
||||
from vllm._ipex_ops import ipex_ops as ops
|
||||
|
||||
d = x.shape[-1] // 2
|
||||
output_shape = (x.shape[:-1] + (d, ))
|
||||
out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
|
||||
ops.silu_and_mul(out, x)
|
||||
return out
|
||||
|
||||
|
||||
@CustomOp.register("gelu_and_mul")
|
||||
class GeluAndMul(CustomOp):
|
||||
"""An activation function for GeGLU.
|
||||
|
||||
The function computes x -> GELU(x[:d]) * x[d:] where d = x.shape[-1] // 2.
|
||||
|
||||
Shapes:
|
||||
x: (batch_size, seq_len, 2 * d) or (num_tokens, 2 * d)
|
||||
return: (batch_size, seq_len, d) or (num_tokens, d)
|
||||
"""
|
||||
|
||||
def __init__(self, approximate: str = "none"):
|
||||
super().__init__()
|
||||
self.approximate = approximate
|
||||
if approximate not in ("none", "tanh"):
|
||||
raise ValueError(f"Unknown approximate mode: {approximate}")
|
||||
|
||||
def forward_native(self, x: torch.Tensor) -> torch.Tensor:
|
||||
"""PyTorch-native implementation equivalent to forward()."""
|
||||
d = x.shape[-1] // 2
|
||||
return F.gelu(x[..., :d], approximate=self.approximate) * x[..., d:]
|
||||
|
||||
def forward_cuda(self, x: torch.Tensor) -> torch.Tensor:
|
||||
from vllm import _custom_ops as ops
|
||||
|
||||
d = x.shape[-1] // 2
|
||||
output_shape = (x.shape[:-1] + (d, ))
|
||||
out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
|
||||
if self.approximate == "none":
|
||||
ops.gelu_and_mul(out, x)
|
||||
elif self.approximate == "tanh":
|
||||
ops.gelu_tanh_and_mul(out, x)
|
||||
return out
|
||||
|
||||
def forward_xpu(self, x: torch.Tensor) -> torch.Tensor:
|
||||
from vllm._ipex_ops import ipex_ops as ops
|
||||
|
||||
d = x.shape[-1] // 2
|
||||
output_shape = (x.shape[:-1] + (d, ))
|
||||
out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
|
||||
if self.approximate == "none":
|
||||
ops.gelu_and_mul(out, x)
|
||||
elif self.approximate == "tanh":
|
||||
ops.gelu_tanh_and_mul(out, x)
|
||||
return out
|
||||
|
||||
def extra_repr(self) -> str:
|
||||
return f'approximate={repr(self.approximate)}'
|
||||
|
||||
|
||||
@CustomOp.register("gelu_new")
|
||||
class NewGELU(CustomOp):
|
||||
|
||||
def forward_native(self, x: torch.Tensor) -> torch.Tensor:
|
||||
"""PyTorch-native implementation equivalent to forward()."""
|
||||
c = math.sqrt(2.0 / math.pi)
|
||||
return 0.5 * x * (1.0 + torch.tanh(c *
|
||||
(x + 0.044715 * torch.pow(x, 3.0))))
|
||||
|
||||
def forward_cuda(self, x: torch.Tensor) -> torch.Tensor:
|
||||
from vllm import _custom_ops as ops
|
||||
|
||||
out = torch.empty_like(x)
|
||||
ops.gelu_new(out, x)
|
||||
return out
|
||||
|
||||
def forward_xpu(self, x: torch.Tensor) -> torch.Tensor:
|
||||
from vllm._ipex_ops import ipex_ops as ops
|
||||
|
||||
return ops.gelu_new(x)
|
||||
|
||||
|
||||
@CustomOp.register("gelu_fast")
|
||||
class FastGELU(CustomOp):
|
||||
|
||||
def forward_native(self, x: torch.Tensor) -> torch.Tensor:
|
||||
"""PyTorch-native implementation equivalent to forward()."""
|
||||
return 0.5 * x * (1.0 + torch.tanh(x * 0.7978845608 *
|
||||
(1.0 + 0.044715 * x * x)))
|
||||
|
||||
def forward_cuda(self, x: torch.Tensor) -> torch.Tensor:
|
||||
from vllm import _custom_ops as ops
|
||||
|
||||
out = torch.empty_like(x)
|
||||
ops.gelu_fast(out, x)
|
||||
return out
|
||||
|
||||
def forward_xpu(self, x: torch.Tensor) -> torch.Tensor:
|
||||
from vllm._ipex_ops import ipex_ops as ops
|
||||
|
||||
return ops.gelu_fast(x)
|
||||
|
||||
|
||||
@CustomOp.register("quick_gelu")
|
||||
class QuickGELU(CustomOp):
|
||||
# https://github.com/huggingface/transformers/blob/main/src/transformers/activations.py#L90
|
||||
def forward_native(self, x: torch.Tensor) -> torch.Tensor:
|
||||
"""PyTorch-native implementation equivalent to forward()."""
|
||||
return x * torch.sigmoid(1.702 * x)
|
||||
|
||||
def forward_cuda(self, x: torch.Tensor) -> torch.Tensor:
|
||||
from vllm import _custom_ops as ops
|
||||
|
||||
out = torch.empty_like(x)
|
||||
ops.gelu_quick(out, x)
|
||||
return out
|
||||
|
||||
def forward_xpu(self, x: torch.Tensor) -> torch.Tensor:
|
||||
from vllm._ipex_ops import ipex_ops as ops
|
||||
|
||||
out = torch.empty_like(x)
|
||||
ops.gelu_quick(out, x)
|
||||
return out
|
||||
|
||||
# TODO implement forward_xpu for QuickGELU
|
||||
# def forward_xpu(self, x: torch.Tensor) -> torch.Tensor:
|
||||
|
||||
|
||||
@CustomOp.register("relu2")
|
||||
class ReLUSquaredActivation(CustomOp):
|
||||
"""
|
||||
Applies the relu^2 activation introduced in https://arxiv.org/abs/2109.08668v2
|
||||
"""
|
||||
|
||||
def forward_native(self, x: torch.Tensor) -> torch.Tensor:
|
||||
"""PyTorch-native implementation equivalent to forward()."""
|
||||
return torch.square(F.relu(x))
|
||||
|
||||
def forward_cuda(self, x: torch.Tensor) -> torch.Tensor:
|
||||
return self.forward_native(x)
|
||||
|
||||
|
||||
class ScaledActivation(nn.Module):
|
||||
"""An activation function with post-scale parameters.
|
||||
|
||||
This is used for some quantization methods like AWQ.
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
act_module: nn.Module,
|
||||
intermediate_size: int,
|
||||
input_is_parallel: bool = True,
|
||||
params_dtype: Optional[torch.dtype] = None,
|
||||
):
|
||||
super().__init__()
|
||||
self.act = act_module
|
||||
self.input_is_parallel = input_is_parallel
|
||||
if input_is_parallel:
|
||||
tp_size = get_tensor_model_parallel_world_size()
|
||||
intermediate_size_per_partition = divide(intermediate_size,
|
||||
tp_size)
|
||||
else:
|
||||
intermediate_size_per_partition = intermediate_size
|
||||
if params_dtype is None:
|
||||
params_dtype = torch.get_default_dtype()
|
||||
self.scales = nn.Parameter(
|
||||
torch.empty(intermediate_size_per_partition, dtype=params_dtype))
|
||||
set_weight_attrs(self.scales, {"weight_loader": self.weight_loader})
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
return self.act(x) / self.scales
|
||||
|
||||
def weight_loader(self, param: nn.Parameter, loaded_weight: torch.Tensor):
|
||||
param_data = param.data
|
||||
if self.input_is_parallel:
|
||||
tp_rank = get_tensor_model_parallel_rank()
|
||||
shard_size = param_data.shape[0]
|
||||
start_idx = tp_rank * shard_size
|
||||
loaded_weight = loaded_weight.narrow(0, start_idx, shard_size)
|
||||
assert param_data.shape == loaded_weight.shape
|
||||
param_data.copy_(loaded_weight)
|
||||
|
||||
|
||||
_ACTIVATION_REGISTRY = LazyDict({
|
||||
"gelu":
|
||||
lambda: nn.GELU(),
|
||||
"gelu_fast":
|
||||
lambda: FastGELU(),
|
||||
"gelu_new":
|
||||
lambda: NewGELU(),
|
||||
"gelu_pytorch_tanh":
|
||||
lambda: nn.GELU(approximate="tanh"),
|
||||
"relu":
|
||||
lambda: nn.ReLU(),
|
||||
"relu2":
|
||||
lambda: ReLUSquaredActivation(),
|
||||
"silu":
|
||||
lambda: nn.SiLU(),
|
||||
"quick_gelu":
|
||||
lambda: QuickGELU(),
|
||||
})
|
||||
|
||||
|
||||
def get_act_fn(act_fn_name: str) -> nn.Module:
|
||||
"""Get an activation function by name."""
|
||||
act_fn_name = act_fn_name.lower()
|
||||
if act_fn_name not in _ACTIVATION_REGISTRY:
|
||||
raise ValueError(
|
||||
f"Activation function {act_fn_name!r} is not supported.")
|
||||
|
||||
return _ACTIVATION_REGISTRY[act_fn_name]
|
||||
|
||||
|
||||
_ACTIVATION_AND_MUL_REGISTRY = LazyDict({
|
||||
"gelu": lambda: GeluAndMul(),
|
||||
"silu": lambda: SiluAndMul(),
|
||||
})
|
||||
|
||||
|
||||
def get_act_and_mul_fn(act_fn_name: str) -> nn.Module:
|
||||
"""Get an activation-and-mul (i.e. SiluAndMul) function by name."""
|
||||
act_fn_name = act_fn_name.lower()
|
||||
if act_fn_name not in _ACTIVATION_AND_MUL_REGISTRY:
|
||||
raise ValueError(
|
||||
f"Activation function {act_fn_name!r} is not supported.")
|
||||
|
||||
return _ACTIVATION_AND_MUL_REGISTRY[act_fn_name]
|
||||
46
vllm-v0.6.2/vllm/model_executor/layers/fused_moe/__init__.py
Normal file
46
vllm-v0.6.2/vllm/model_executor/layers/fused_moe/__init__.py
Normal file
@@ -0,0 +1,46 @@
|
||||
from contextlib import contextmanager
|
||||
from typing import Any, Dict, Optional
|
||||
|
||||
from vllm.model_executor.layers.fused_moe.layer import (
|
||||
FusedMoE, FusedMoEMethodBase, FusedMoeWeightScaleSupported)
|
||||
from vllm.triton_utils import HAS_TRITON
|
||||
|
||||
_config: Optional[Dict[str, Any]] = None
|
||||
|
||||
|
||||
@contextmanager
|
||||
def override_config(config):
|
||||
global _config
|
||||
old_config = _config
|
||||
_config = config
|
||||
yield
|
||||
_config = old_config
|
||||
|
||||
|
||||
def get_config() -> Optional[Dict[str, Any]]:
|
||||
return _config
|
||||
|
||||
|
||||
__all__ = [
|
||||
"FusedMoE",
|
||||
"FusedMoEMethodBase",
|
||||
"FusedMoeWeightScaleSupported",
|
||||
"override_config",
|
||||
"get_config",
|
||||
]
|
||||
|
||||
if HAS_TRITON:
|
||||
# import to register the custom ops
|
||||
import vllm.model_executor.layers.fused_moe.fused_marlin_moe # noqa
|
||||
import vllm.model_executor.layers.fused_moe.fused_moe # noqa
|
||||
from vllm.model_executor.layers.fused_moe.fused_moe import (
|
||||
fused_experts, fused_moe, fused_topk, get_config_file_name,
|
||||
grouped_topk)
|
||||
|
||||
__all__ += [
|
||||
"fused_moe",
|
||||
"fused_topk",
|
||||
"fused_experts",
|
||||
"get_config_file_name",
|
||||
"grouped_topk",
|
||||
]
|
||||
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,218 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"5120": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"9216": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"13312": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"17408": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"25600": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"33792": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"41984": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"50176": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"58368": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,218 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"5120": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"9216": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"13312": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"17408": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"25600": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"33792": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"41984": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"50176": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"58368": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,218 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"5120": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"9216": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"13312": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"17408": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"25600": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"33792": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"41984": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"50176": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"58368": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,218 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 5
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 5
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 5
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 256,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 256,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 256,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 256,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 256,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 256,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 256,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"5120": {
|
||||
"BLOCK_SIZE_M": 256,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"9216": {
|
||||
"BLOCK_SIZE_M": 256,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"13312": {
|
||||
"BLOCK_SIZE_M": 256,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"17408": {
|
||||
"BLOCK_SIZE_M": 256,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"25600": {
|
||||
"BLOCK_SIZE_M": 256,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"33792": {
|
||||
"BLOCK_SIZE_M": 256,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"41984": {
|
||||
"BLOCK_SIZE_M": 256,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"50176": {
|
||||
"BLOCK_SIZE_M": 256,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"58368": {
|
||||
"BLOCK_SIZE_M": 256,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,218 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 5
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"5120": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"9216": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"13312": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"17408": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"25600": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"33792": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"41984": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"50176": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"58368": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,218 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 256,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"5120": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"9216": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"13312": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"17408": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"25600": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"33792": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"41984": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"50176": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"58368": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,218 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"5120": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"9216": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"13312": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"17408": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"25600": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"33792": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"41984": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"50176": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"58368": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,218 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"5120": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"9216": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"13312": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"17408": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"25600": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"33792": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"41984": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"50176": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"58368": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,218 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"5120": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"9216": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"13312": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"17408": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"25600": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"33792": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"41984": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"50176": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"58368": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 5
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 5
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 5
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 5
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,218 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"5120": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"9216": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"13312": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"17408": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"25600": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"33792": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"41984": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"50176": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"58368": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,218 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"5120": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"9216": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"13312": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"17408": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"25600": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"33792": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"41984": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"50176": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"58368": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 5
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 256,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 5
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 5
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 256,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,130 @@
|
||||
{
|
||||
"3328": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 32,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 8,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"768": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 8,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1792": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2560": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"2816": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3584": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 8,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3840": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1280": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"2304": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,218 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 5
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"5120": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"9216": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"13312": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"17408": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"25600": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"33792": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"41984": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"50176": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"58368": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,130 @@
|
||||
{
|
||||
"3840": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 8,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1792": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 8,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3584": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"2816": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 32,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1280": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"768": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 8,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3328": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"2560": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 8,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 32,
|
||||
"GROUP_SIZE_M": 8,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2304": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 5
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 5
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 5
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 256,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 256,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,130 @@
|
||||
{
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 32,
|
||||
"GROUP_SIZE_M": 8,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1792": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 8,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3328": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 2
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"2560": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 32,
|
||||
"GROUP_SIZE_M": 8,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"768": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"2816": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"2304": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 8,
|
||||
"num_warps": 8,
|
||||
"num_stages": 2
|
||||
},
|
||||
"1280": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3840": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"3584": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 32,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 5
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 5
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,200 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 2,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 16,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 2,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 2,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 16,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 1,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 16,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 1,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 16,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 4,
|
||||
"num_warps": 2,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 16,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 4,
|
||||
"num_warps": 2,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 4,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 4,
|
||||
"num_warps": 4,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 4,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 4,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 4,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 32,
|
||||
"kpack": 2
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,138 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 5
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 2
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,200 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 2,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 16,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 2,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 16,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 4,
|
||||
"num_warps": 2,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 2,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 4,
|
||||
"num_warps": 4,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 4,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 4,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 5
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,200 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 16,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 2,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 16,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 2,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 2,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 2,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 2,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 16,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 4,
|
||||
"num_warps": 2,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 2,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 4,
|
||||
"num_warps": 4,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 4,
|
||||
"num_warps": 4,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 4,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 4,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 32,
|
||||
"kpack": 2
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,173 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_ctas": 1,
|
||||
"num_stages": 2
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 32,
|
||||
"GROUP_SIZE_M": 2,
|
||||
"num_warps": 4,
|
||||
"num_ctas": 1,
|
||||
"num_stages": 7
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 32,
|
||||
"GROUP_SIZE_M": 128,
|
||||
"num_warps": 2,
|
||||
"num_ctas": 1,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 2,
|
||||
"num_ctas": 1,
|
||||
"num_stages": 1
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 2,
|
||||
"num_ctas": 1,
|
||||
"num_stages": 1
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 2,
|
||||
"num_warps": 4,
|
||||
"num_ctas": 1,
|
||||
"num_stages": 2
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 2,
|
||||
"num_warps": 4,
|
||||
"num_ctas": 1,
|
||||
"num_stages": 2
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 2,
|
||||
"num_warps": 4,
|
||||
"num_ctas": 1,
|
||||
"num_stages": 2
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 8,
|
||||
"num_warps": 8,
|
||||
"num_ctas": 1,
|
||||
"num_stages": 2
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 8,
|
||||
"num_warps": 8,
|
||||
"num_ctas": 1,
|
||||
"num_stages": 2
|
||||
},
|
||||
"192": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 8,
|
||||
"num_warps": 8,
|
||||
"num_ctas": 1,
|
||||
"num_stages": 2
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 16,
|
||||
"num_ctas": 1,
|
||||
"num_stages": 2
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 128,
|
||||
"num_warps": 2,
|
||||
"num_ctas": 1,
|
||||
"num_stages": 8
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 4,
|
||||
"num_warps": 8,
|
||||
"num_ctas": 1,
|
||||
"num_stages": 3
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 16,
|
||||
"num_ctas": 1,
|
||||
"num_stages": 2
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 16,
|
||||
"num_ctas": 1,
|
||||
"num_stages": 2
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 32,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_ctas": 1,
|
||||
"num_stages": 2
|
||||
},
|
||||
"6144": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 32,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_ctas": 1,
|
||||
"num_stages": 2
|
||||
},
|
||||
"8192": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 16,
|
||||
"num_ctas": 1,
|
||||
"num_stages": 2
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,200 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 16,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 2,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 32,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 2,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 16,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 4,
|
||||
"num_warps": 2,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 4,
|
||||
"num_warps": 2,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 4,
|
||||
"num_warps": 4,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 4,
|
||||
"num_warps": 4,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 4,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 4,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 32,
|
||||
"kpack": 2
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 2
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 0,
|
||||
"waves_per_eu": 0,
|
||||
"matrix_instr_nonkdim": 16,
|
||||
"kpack": 1
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 3
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 32,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 16,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 32,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,146 @@
|
||||
{
|
||||
"1": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 64,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 5
|
||||
},
|
||||
"2": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 16,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"4": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"8": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 64,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 5
|
||||
},
|
||||
"16": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"24": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"32": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"48": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"64": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"96": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"128": {
|
||||
"BLOCK_SIZE_M": 64,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 256,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 4,
|
||||
"num_stages": 4
|
||||
},
|
||||
"256": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 128,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 1,
|
||||
"num_warps": 8,
|
||||
"num_stages": 5
|
||||
},
|
||||
"512": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1024": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 64,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"1536": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"2048": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
},
|
||||
"3072": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 3
|
||||
},
|
||||
"4096": {
|
||||
"BLOCK_SIZE_M": 128,
|
||||
"BLOCK_SIZE_N": 256,
|
||||
"BLOCK_SIZE_K": 128,
|
||||
"GROUP_SIZE_M": 32,
|
||||
"num_warps": 8,
|
||||
"num_stages": 4
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,10 @@
|
||||
This directory contains tuned configurations for different settings of the fused_moe kernel.
|
||||
For different settings of
|
||||
- E (number of experts)
|
||||
- N (intermediate size)
|
||||
- device_name (torch.cuda.get_device_name())
|
||||
the JSON file contains a mapping from M (batch size) to the chosen configuration.
|
||||
|
||||
The example configurations provided are for the Mixtral model for TP2 on H100
|
||||
and TP4 on A100. Mixtral has intermediate size N = 14336, i.e. for TP2 we have
|
||||
N = 7168 and for TP4 we have N = 3584.
|
||||
@@ -0,0 +1,359 @@
|
||||
"""Fused MoE utilities for GPTQ."""
|
||||
import functools
|
||||
from typing import Optional
|
||||
|
||||
import torch
|
||||
|
||||
from vllm import _custom_ops as ops
|
||||
from vllm.model_executor.layers.fused_moe.fused_moe import (
|
||||
fused_topk, moe_align_block_size, try_get_optimal_moe_config)
|
||||
from vllm.scalar_type import scalar_types
|
||||
from vllm.utils import direct_register_custom_op
|
||||
|
||||
|
||||
def get_scalar_type(num_bits: int, has_zp: bool):
|
||||
if has_zp:
|
||||
assert num_bits == 4
|
||||
return scalar_types.uint4
|
||||
else:
|
||||
return scalar_types.uint4b8 if num_bits == 4 else scalar_types.uint8b128
|
||||
|
||||
|
||||
def single_marlin_moe(
|
||||
hidden_states: torch.Tensor,
|
||||
w: torch.Tensor,
|
||||
scales: torch.Tensor,
|
||||
gating_output: torch.Tensor,
|
||||
topk: int,
|
||||
renormalize: bool,
|
||||
g_idx: Optional[torch.Tensor] = None,
|
||||
sort_indices: Optional[torch.Tensor] = None,
|
||||
w_zeros: Optional[torch.Tensor] = None,
|
||||
num_bits: int = 8,
|
||||
is_k_full: bool = True,
|
||||
) -> torch.Tensor:
|
||||
"""
|
||||
This function computes the multiplication of hidden_states with expert
|
||||
weights used in Marlin MoE, using weights w and top-k gating mechanism.
|
||||
Its purpose is testing and debugging the fused MoE kernel.
|
||||
|
||||
Parameters:
|
||||
- hidden_states (torch.Tensor): The input tensor to the Marlin Mul.
|
||||
- w (torch.Tensor): The set of expert weights.
|
||||
- scales (torch.Tensor): The quantization scales.
|
||||
- gating_output (torch.Tensor): The output of the gating operation
|
||||
(before softmax).
|
||||
- g_idx (Optional[torch.Tensor]): Optional act_order indices.
|
||||
- sort_indices (Optional[torch.Tensor]): Optional act_order input
|
||||
permutation.
|
||||
- topk (int): The number of top-k experts to select.
|
||||
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
||||
- w_zeros (Optional[torch.Tensor]): Optional zero points to be used for w.
|
||||
- num_bits (bool): The number of bits in expert weights quantization.
|
||||
|
||||
Returns:
|
||||
- torch.Tensor: The output tensor after applying the MoE layer.
|
||||
"""
|
||||
# Check constraints.
|
||||
assert hidden_states.shape[0] == gating_output.shape[0], (
|
||||
"Number of tokens mismatch")
|
||||
assert hidden_states.shape[1] == w.shape[1] * 16, "Hidden size mismatch"
|
||||
assert gating_output.shape[1] == w.shape[0], "Number of experts mismatch"
|
||||
assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
|
||||
assert w.is_contiguous(), "Expert weights must be contiguous"
|
||||
assert hidden_states.dtype == torch.float16
|
||||
assert num_bits in [4, 8]
|
||||
|
||||
M, K = hidden_states.shape
|
||||
E = w.shape[0]
|
||||
N = w.shape[2] // (num_bits // 2)
|
||||
|
||||
topk_weights, topk_ids = fused_topk(hidden_states, gating_output, topk,
|
||||
renormalize)
|
||||
|
||||
# This might not be an optimal config for a single MMM
|
||||
get_config_func = functools.partial(try_get_optimal_moe_config,
|
||||
w.shape,
|
||||
w.shape,
|
||||
topk_ids.shape[1],
|
||||
None,
|
||||
is_marlin=True)
|
||||
config = get_config_func(M)
|
||||
|
||||
block_size_m = config['BLOCK_SIZE_M']
|
||||
|
||||
sorted_token_ids, _, _ = moe_align_block_size(topk_ids, block_size_m, E)
|
||||
|
||||
max_workspace_size = (N // 64) * 16
|
||||
workspace = torch.zeros(max_workspace_size,
|
||||
dtype=torch.int,
|
||||
device=hidden_states.device,
|
||||
requires_grad=False)
|
||||
|
||||
has_zero_point = w_zeros is not None
|
||||
if w_zeros is None:
|
||||
w_zeros = torch.empty((0, 0),
|
||||
dtype=hidden_states.dtype,
|
||||
device=hidden_states.device,
|
||||
requires_grad=False)
|
||||
|
||||
if g_idx is None:
|
||||
g_idx = torch.empty((0, 0),
|
||||
dtype=torch.int32,
|
||||
device=hidden_states.device,
|
||||
requires_grad=False)
|
||||
|
||||
if sort_indices is None:
|
||||
sort_indices = torch.empty((0),
|
||||
dtype=torch.int32,
|
||||
device=hidden_states.device,
|
||||
requires_grad=False)
|
||||
|
||||
scalar_type = get_scalar_type(num_bits, has_zero_point)
|
||||
|
||||
intermediate_cache = torch.ops._moe_C.marlin_gemm_moe(
|
||||
hidden_states, w, sorted_token_ids, topk_weights, topk_ids, scales,
|
||||
w_zeros, g_idx, sort_indices, workspace, scalar_type.id, M, N, K,
|
||||
is_k_full, E, topk, block_size_m, True, False)
|
||||
|
||||
return torch.sum(intermediate_cache.view(*intermediate_cache.shape), dim=1)
|
||||
|
||||
|
||||
def single_marlin_moe_fake(
|
||||
hidden_states: torch.Tensor,
|
||||
w: torch.Tensor,
|
||||
scales: torch.Tensor,
|
||||
gating_output: torch.Tensor,
|
||||
topk: int,
|
||||
renormalize: bool,
|
||||
g_idx: Optional[torch.Tensor] = None,
|
||||
sort_indices: Optional[torch.Tensor] = None,
|
||||
w_zeros: Optional[torch.Tensor] = None,
|
||||
num_bits: int = 8,
|
||||
is_k_full: bool = True,
|
||||
) -> torch.Tensor:
|
||||
return torch.empty_like(hidden_states)
|
||||
|
||||
|
||||
direct_register_custom_op(
|
||||
op_name="single_marlin_moe",
|
||||
op_func=single_marlin_moe,
|
||||
mutates_args=[],
|
||||
fake_impl=single_marlin_moe_fake,
|
||||
)
|
||||
|
||||
|
||||
def fused_marlin_moe(
|
||||
hidden_states: torch.Tensor,
|
||||
w1: torch.Tensor,
|
||||
w2: torch.Tensor,
|
||||
w1_scale: torch.Tensor,
|
||||
w2_scale: torch.Tensor,
|
||||
gating_output: torch.Tensor,
|
||||
topk_weights: torch.Tensor,
|
||||
topk_ids: torch.Tensor,
|
||||
g_idx1: Optional[torch.Tensor] = None,
|
||||
g_idx2: Optional[torch.Tensor] = None,
|
||||
sort_indices1: Optional[torch.Tensor] = None,
|
||||
sort_indices2: Optional[torch.Tensor] = None,
|
||||
w1_zeros: Optional[torch.Tensor] = None,
|
||||
w2_zeros: Optional[torch.Tensor] = None,
|
||||
num_bits: int = 8,
|
||||
is_k_full: bool = True,
|
||||
) -> torch.Tensor:
|
||||
"""
|
||||
This function computes a Mixture of Experts (MoE) layer using two sets of
|
||||
weights, w1 and w2, and top-k gating mechanism.
|
||||
|
||||
Parameters:
|
||||
- hidden_states (torch.Tensor): The input tensor to the MoE layer.
|
||||
- w1 (torch.Tensor): The first set of expert weights.
|
||||
- w2 (torch.Tensor): The second set of expert weights.
|
||||
- w1_scale (torch.Tensor): Scale to be used for w1.
|
||||
- w2_scale (torch.Tensor): Scale to be used for w2.
|
||||
- gating_output (torch.Tensor): The output of the gating operation
|
||||
(before softmax).
|
||||
- g_idx1 (Optional[torch.Tensor]): The first set of act_order indices.
|
||||
- g_idx2 (Optional[torch.Tensor]): The second set of act_order indices.
|
||||
- sort_indices1 (Optional[torch.Tensor]): The first act_order input
|
||||
permutation.
|
||||
- sort_indices2 (Optional[torch.Tensor]): The second act_order input
|
||||
permutation.
|
||||
- topk_weights (torch.Tensor): Top-k weights.
|
||||
- topk_ids (torch.Tensor): Indices of topk-k elements.
|
||||
- w1_zeros (Optional[torch.Tensor]): Optional zero points to be used for w1.
|
||||
- w2_zeros (Optional[torch.Tensor]): Optional zero points to be used for w2.
|
||||
- num_bits (bool): The number of bits in expert weights quantization.
|
||||
|
||||
Returns:
|
||||
- torch.Tensor: The output tensor after applying the MoE layer.
|
||||
"""
|
||||
# Check constraints.
|
||||
assert hidden_states.shape[0] == gating_output.shape[
|
||||
0], "Number of tokens mismatch"
|
||||
assert hidden_states.shape[
|
||||
1] == w1.shape[1] * 16, "Hidden size mismatch w1"
|
||||
assert hidden_states.shape[1] == w2.shape[2] // (
|
||||
num_bits // 2), "Hidden size mismatch w2"
|
||||
assert gating_output.shape[1] == w1.shape[0], "Number of experts mismatch"
|
||||
assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
|
||||
assert w1.is_contiguous(), "Expert weights1 must be contiguous"
|
||||
assert w2.is_contiguous(), "Expert weights2 must be contiguous"
|
||||
assert hidden_states.dtype == torch.float16
|
||||
assert num_bits in [4, 8]
|
||||
|
||||
has_no_act_order = (g_idx1 is None and g_idx2 is None
|
||||
and sort_indices1 is None and sort_indices2 is None)
|
||||
has_all_act_order = (g_idx1 is not None and g_idx2 is not None
|
||||
and sort_indices1 is not None
|
||||
and sort_indices2 is not None)
|
||||
assert has_no_act_order or has_all_act_order, (
|
||||
"g_idx and sorted_indices "
|
||||
"must be all not None or must be all None")
|
||||
|
||||
has_no_zp = w1_zeros is None and w2_zeros is None
|
||||
has_all_zp = w1_zeros is not None and w2_zeros is not None
|
||||
assert has_no_zp or has_all_zp, ("zero points must be both not None or "
|
||||
"must be both None")
|
||||
|
||||
M, K = hidden_states.shape
|
||||
E = w1.shape[0]
|
||||
N = w2.shape[1] * 16
|
||||
topk = topk_ids.shape[1]
|
||||
|
||||
get_config_func = functools.partial(
|
||||
try_get_optimal_moe_config,
|
||||
w1.shape,
|
||||
w2.shape,
|
||||
topk_ids.shape[1],
|
||||
None,
|
||||
is_marlin=True,
|
||||
)
|
||||
config = get_config_func(M)
|
||||
|
||||
block_size_m = config["BLOCK_SIZE_M"]
|
||||
|
||||
sorted_token_ids, _, _ = moe_align_block_size(topk_ids, block_size_m, E)
|
||||
|
||||
max_workspace_size = (max(2 * N, K) // 64) * 16
|
||||
workspace = torch.zeros(max_workspace_size,
|
||||
dtype=torch.int,
|
||||
device="cuda",
|
||||
requires_grad=False)
|
||||
|
||||
if has_no_zp:
|
||||
w1_zeros = torch.empty((0, 0),
|
||||
dtype=hidden_states.dtype,
|
||||
device=hidden_states.device,
|
||||
requires_grad=False)
|
||||
w2_zeros = torch.empty((0, 0),
|
||||
dtype=hidden_states.dtype,
|
||||
device=hidden_states.device,
|
||||
requires_grad=False)
|
||||
|
||||
if has_no_act_order:
|
||||
g_idx1 = torch.empty((0, 0),
|
||||
dtype=torch.int32,
|
||||
device=hidden_states.device,
|
||||
requires_grad=False)
|
||||
g_idx2 = torch.empty((0, 0),
|
||||
dtype=torch.int32,
|
||||
device=hidden_states.device,
|
||||
requires_grad=False)
|
||||
sort_indices1 = torch.empty((0),
|
||||
dtype=torch.int32,
|
||||
device=hidden_states.device,
|
||||
requires_grad=False)
|
||||
sort_indices2 = torch.empty((0, 0),
|
||||
dtype=torch.int32,
|
||||
device=hidden_states.device,
|
||||
requires_grad=False)
|
||||
|
||||
scalar_type1 = get_scalar_type(num_bits, has_all_zp)
|
||||
scalar_type2 = get_scalar_type(num_bits, has_all_zp)
|
||||
|
||||
intermediate_cache2 = torch.empty(
|
||||
(M * topk_ids.shape[1], N),
|
||||
device=hidden_states.device,
|
||||
dtype=hidden_states.dtype,
|
||||
)
|
||||
|
||||
intermediate_cache1 = torch.ops._moe_C.marlin_gemm_moe(
|
||||
hidden_states,
|
||||
w1,
|
||||
sorted_token_ids,
|
||||
topk_weights,
|
||||
topk_ids,
|
||||
w1_scale,
|
||||
w1_zeros,
|
||||
g_idx1,
|
||||
sort_indices1,
|
||||
workspace,
|
||||
scalar_type1.id,
|
||||
M,
|
||||
2 * N,
|
||||
K,
|
||||
is_k_full,
|
||||
E,
|
||||
topk,
|
||||
block_size_m,
|
||||
True,
|
||||
False,
|
||||
)
|
||||
|
||||
ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, 2 * N))
|
||||
|
||||
intermediate_cache3 = torch.ops._moe_C.marlin_gemm_moe(
|
||||
intermediate_cache2,
|
||||
w2,
|
||||
sorted_token_ids,
|
||||
topk_weights,
|
||||
topk_ids,
|
||||
w2_scale,
|
||||
w2_zeros,
|
||||
g_idx2,
|
||||
sort_indices2,
|
||||
workspace,
|
||||
scalar_type2.id,
|
||||
M,
|
||||
K,
|
||||
N,
|
||||
is_k_full,
|
||||
E,
|
||||
topk,
|
||||
block_size_m,
|
||||
False,
|
||||
True,
|
||||
)
|
||||
|
||||
return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape),
|
||||
dim=1)
|
||||
|
||||
|
||||
def fused_marlin_moe_fake(
|
||||
hidden_states: torch.Tensor,
|
||||
w1: torch.Tensor,
|
||||
w2: torch.Tensor,
|
||||
w1_scale: torch.Tensor,
|
||||
w2_scale: torch.Tensor,
|
||||
gating_output: torch.Tensor,
|
||||
topk_weights: torch.Tensor,
|
||||
topk_ids: torch.Tensor,
|
||||
g_idx1: Optional[torch.Tensor] = None,
|
||||
g_idx2: Optional[torch.Tensor] = None,
|
||||
sort_indices1: Optional[torch.Tensor] = None,
|
||||
sort_indices2: Optional[torch.Tensor] = None,
|
||||
w1_zeros: Optional[torch.Tensor] = None,
|
||||
w2_zeros: Optional[torch.Tensor] = None,
|
||||
num_bits: int = 8,
|
||||
is_k_full: bool = True,
|
||||
) -> torch.Tensor:
|
||||
return torch.empty_like(hidden_states)
|
||||
|
||||
|
||||
direct_register_custom_op(
|
||||
op_name="fused_marlin_moe",
|
||||
op_func=fused_marlin_moe,
|
||||
mutates_args=[],
|
||||
fake_impl=fused_marlin_moe_fake,
|
||||
)
|
||||
776
vllm-v0.6.2/vllm/model_executor/layers/fused_moe/fused_moe.py
Normal file
776
vllm-v0.6.2/vllm/model_executor/layers/fused_moe/fused_moe.py
Normal file
@@ -0,0 +1,776 @@
|
||||
"""Fused MoE kernel."""
|
||||
import functools
|
||||
import json
|
||||
import os
|
||||
from typing import Any, Callable, Dict, Optional, Tuple
|
||||
|
||||
import torch
|
||||
import triton
|
||||
import triton.language as tl
|
||||
|
||||
import vllm.envs as envs
|
||||
from vllm import _custom_ops as ops
|
||||
from vllm.logger import init_logger
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.utils import direct_register_custom_op
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
|
||||
@triton.jit
|
||||
def fused_moe_kernel(
|
||||
# Pointers to matrices
|
||||
a_ptr,
|
||||
b_ptr,
|
||||
c_ptr,
|
||||
a_scale_ptr,
|
||||
b_scale_ptr,
|
||||
topk_weights_ptr,
|
||||
sorted_token_ids_ptr,
|
||||
expert_ids_ptr,
|
||||
num_tokens_post_padded_ptr,
|
||||
# Matrix dimensions
|
||||
N,
|
||||
K,
|
||||
EM,
|
||||
num_valid_tokens,
|
||||
# The stride variables represent how much to increase the ptr by when
|
||||
# moving by 1 element in a particular dimension. E.g. `stride_am` is
|
||||
# how much to increase `a_ptr` by to get the element one row down
|
||||
# (A has M rows).
|
||||
stride_am,
|
||||
stride_ak,
|
||||
stride_be,
|
||||
stride_bk,
|
||||
stride_bn,
|
||||
stride_cm,
|
||||
stride_cn,
|
||||
stride_bse,
|
||||
stride_bsn,
|
||||
# Meta-parameters
|
||||
BLOCK_SIZE_M: tl.constexpr,
|
||||
BLOCK_SIZE_N: tl.constexpr,
|
||||
BLOCK_SIZE_K: tl.constexpr,
|
||||
GROUP_SIZE_M: tl.constexpr,
|
||||
MUL_ROUTED_WEIGHT: tl.constexpr,
|
||||
top_k: tl.constexpr,
|
||||
compute_type: tl.constexpr,
|
||||
use_fp8_w8a8: tl.constexpr,
|
||||
use_int8_w8a16: tl.constexpr):
|
||||
"""
|
||||
Implements the fused computation for a Mixture of Experts (MOE) using
|
||||
token and expert matrices.
|
||||
|
||||
Key Parameters:
|
||||
- A: The input tensor representing tokens with shape (*, K), where '*' can
|
||||
be any shape representing batches and K is the feature dimension of
|
||||
each token.
|
||||
- B: The stacked MOE weight tensor with shape (E, N, K), where E is
|
||||
the number of experts, K is the input feature dimension, and N is
|
||||
the output feature dimension.
|
||||
- C: The output cache tensor with shape (M, topk, N), where M is the
|
||||
total number of tokens post padding, topk is the number of times
|
||||
each token is repeated, and N is the output feature dimension.
|
||||
- sorted_token_ids: A tensor containing the sorted indices of tokens,
|
||||
repeated topk times and arranged by the expert index they are
|
||||
assigned to.
|
||||
- expert_ids: A tensor containing the indices of the expert for each
|
||||
block. It determines which expert matrix from B should be used for
|
||||
each block in A.
|
||||
This kernel performs the multiplication of a token by its corresponding
|
||||
expert matrix as determined by `expert_ids`. The sorting of
|
||||
`sorted_token_ids` by expert index and padding ensures divisibility by
|
||||
BLOCK_SIZE_M, which is necessary to maintain consistency in block matrix
|
||||
multiplication across different blocks processed by the same expert.
|
||||
"""
|
||||
# -----------------------------------------------------------
|
||||
# Map program ids `pid` to the block of C it should compute.
|
||||
# This is done in a grouped ordering to promote L2 data reuse.
|
||||
pid = tl.program_id(axis=0)
|
||||
num_pid_m = tl.cdiv(EM, BLOCK_SIZE_M)
|
||||
num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
|
||||
num_pid_in_group = GROUP_SIZE_M * num_pid_n
|
||||
group_id = pid // num_pid_in_group
|
||||
first_pid_m = group_id * GROUP_SIZE_M
|
||||
group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M)
|
||||
pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m)
|
||||
pid_n = (pid % num_pid_in_group) // group_size_m
|
||||
|
||||
# ----------------------------------------------------------
|
||||
# Create pointers for the first blocks of A and B.
|
||||
# We will advance this pointer as we move in the K direction
|
||||
# and accumulate
|
||||
# `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers
|
||||
# `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers
|
||||
num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
|
||||
if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
|
||||
return
|
||||
offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
|
||||
offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
|
||||
token_mask = offs_token < num_valid_tokens
|
||||
|
||||
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
|
||||
offs_k = tl.arange(0, BLOCK_SIZE_K)
|
||||
a_ptrs = a_ptr + (offs_token[:, None] // top_k * stride_am +
|
||||
offs_k[None, :] * stride_ak)
|
||||
|
||||
off_experts = tl.load(expert_ids_ptr + pid_m)
|
||||
b_ptrs = b_ptr + off_experts * stride_be + (offs_k[:, None] * stride_bk +
|
||||
offs_bn[None, :] * stride_bn)
|
||||
if use_int8_w8a16:
|
||||
b_scale_ptrs = b_scale_ptr + off_experts * stride_bse + offs_bn[
|
||||
None, :] * stride_bsn
|
||||
b_scale = tl.load(b_scale_ptrs)
|
||||
|
||||
if use_fp8_w8a8:
|
||||
a_scale = tl.load(a_scale_ptr)
|
||||
b_scale = tl.load(b_scale_ptr + off_experts)
|
||||
|
||||
# -----------------------------------------------------------
|
||||
# Iterate to compute a block of the C matrix.
|
||||
# We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block
|
||||
# of fp32 values for higher accuracy.
|
||||
# `accumulator` will be converted back to fp16 after the loop.
|
||||
accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
|
||||
|
||||
for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
|
||||
# Load the next block of A and B, generate a mask by checking the
|
||||
# K dimension.
|
||||
a = tl.load(a_ptrs,
|
||||
mask=token_mask[:, None] &
|
||||
(offs_k[None, :] < K - k * BLOCK_SIZE_K),
|
||||
other=0.0)
|
||||
b = tl.load(b_ptrs,
|
||||
mask=offs_k[:, None] < K - k * BLOCK_SIZE_K,
|
||||
other=0.0)
|
||||
# We accumulate along the K dimension.
|
||||
if use_int8_w8a16:
|
||||
accumulator = tl.dot(a, b.to(compute_type), acc=accumulator)
|
||||
elif use_fp8_w8a8:
|
||||
accumulator = tl.dot(a, b, acc=accumulator)
|
||||
else:
|
||||
accumulator += tl.dot(a, b)
|
||||
# Advance the ptrs to the next K block.
|
||||
a_ptrs += BLOCK_SIZE_K * stride_ak
|
||||
b_ptrs += BLOCK_SIZE_K * stride_bk
|
||||
|
||||
if MUL_ROUTED_WEIGHT:
|
||||
moe_weight = tl.load(topk_weights_ptr + offs_token,
|
||||
mask=token_mask,
|
||||
other=0)
|
||||
accumulator = accumulator * moe_weight[:, None]
|
||||
if use_int8_w8a16:
|
||||
accumulator = (accumulator * b_scale).to(compute_type)
|
||||
elif use_fp8_w8a8:
|
||||
accumulator = (accumulator * a_scale * b_scale).to(compute_type)
|
||||
else:
|
||||
accumulator = accumulator.to(compute_type)
|
||||
# -----------------------------------------------------------
|
||||
# Write back the block of the output
|
||||
offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
|
||||
c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[
|
||||
None, :]
|
||||
c_mask = token_mask[:, None] & (offs_cn[None, :] < N)
|
||||
tl.store(c_ptrs, accumulator, mask=c_mask)
|
||||
|
||||
|
||||
def moe_align_block_size(
|
||||
topk_ids: torch.Tensor, block_size: int,
|
||||
num_experts: int) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
|
||||
"""
|
||||
Aligns the token distribution across experts to be compatible with block
|
||||
size for matrix multiplication.
|
||||
|
||||
Parameters:
|
||||
- topk_ids: A tensor of shape [total_tokens, top_k] representing the
|
||||
top-k expert indices for each token.
|
||||
- block_size: The block size used in block matrix multiplication.
|
||||
- num_experts: The total number of experts.
|
||||
|
||||
Returns:
|
||||
- sorted_token_ids: A tensor containing the sorted token indices according
|
||||
to their allocated expert.
|
||||
- expert_ids: A tensor indicating the assigned expert index for each block.
|
||||
- num_tokens_post_padded: The total number of tokens after padding,
|
||||
ensuring divisibility by block_size.
|
||||
|
||||
This function pads the number of tokens that each expert needs to process
|
||||
so that it is divisible by block_size.
|
||||
Padding ensures that during block matrix multiplication, the dimensions
|
||||
align correctly.
|
||||
|
||||
Example:
|
||||
Given topk_ids = [[2, 3, 4], [1, 2, 4], [1, 3, 4], [1, 2, 3]],
|
||||
block_size = 4, and num_experts = 4:
|
||||
- We initially have 12 tokens (after repeating 'top_k' times) and 4 experts,
|
||||
with each expert needing to process 3 tokens.
|
||||
- As block_size is 4, we pad 1 token for each expert.
|
||||
- First, flatten topk_ids to [2, 3, 4, 1, 2, 4, 1, 3, 4, 1, 2, 3].
|
||||
- Then append padding tokens [12, 12, 12, 12] for each block.
|
||||
- After sorting by expert index, we obtain token_ids
|
||||
[3, 6, 9, 12, 0, 4, 10, 12, 1, 7, 11, 12, 2, 5, 8, 12].
|
||||
Tokens 12 are non-existent (padding) and are ignored in
|
||||
the subsequent matrix multiplication.
|
||||
- The padding ensures that the total number of tokens is now divisible
|
||||
by block_size for proper block matrix operations.
|
||||
"""
|
||||
max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1)
|
||||
sorted_ids = torch.empty((max_num_tokens_padded, ),
|
||||
dtype=torch.int32,
|
||||
device=topk_ids.device)
|
||||
sorted_ids.fill_(topk_ids.numel())
|
||||
max_num_m_blocks = triton.cdiv(max_num_tokens_padded, block_size)
|
||||
expert_ids = torch.empty((max_num_m_blocks, ),
|
||||
dtype=torch.int32,
|
||||
device=topk_ids.device)
|
||||
num_tokens_post_pad = torch.empty((1),
|
||||
dtype=torch.int32,
|
||||
device=topk_ids.device)
|
||||
ops.moe_align_block_size(topk_ids, num_experts, block_size, sorted_ids,
|
||||
expert_ids, num_tokens_post_pad)
|
||||
return sorted_ids, expert_ids, num_tokens_post_pad
|
||||
|
||||
|
||||
def invoke_fused_moe_kernel(A: torch.Tensor, B: torch.Tensor, C: torch.Tensor,
|
||||
A_scale: Optional[torch.Tensor],
|
||||
B_scale: Optional[torch.Tensor],
|
||||
topk_weights: torch.Tensor, topk_ids: torch.Tensor,
|
||||
sorted_token_ids: torch.Tensor,
|
||||
expert_ids: torch.Tensor,
|
||||
num_tokens_post_padded: torch.Tensor,
|
||||
mul_routed_weight: bool, top_k: int,
|
||||
config: Dict[str, Any], compute_type: tl.dtype,
|
||||
use_fp8_w8a8: bool, use_int8_w8a16: bool) -> None:
|
||||
assert topk_weights.stride(1) == 1
|
||||
assert sorted_token_ids.stride(0) == 1
|
||||
|
||||
if use_fp8_w8a8:
|
||||
A, A_scale = ops.scaled_fp8_quant(A, A_scale)
|
||||
assert B_scale is not None
|
||||
elif use_int8_w8a16:
|
||||
assert B_scale is not None
|
||||
else:
|
||||
assert A_scale is None
|
||||
assert B_scale is None
|
||||
|
||||
grid = lambda META: (triton.cdiv(sorted_token_ids.shape[0], META[
|
||||
'BLOCK_SIZE_M']) * triton.cdiv(B.shape[1], META['BLOCK_SIZE_N']), )
|
||||
|
||||
fused_moe_kernel[grid](
|
||||
A,
|
||||
B,
|
||||
C,
|
||||
A_scale,
|
||||
B_scale,
|
||||
topk_weights,
|
||||
sorted_token_ids,
|
||||
expert_ids,
|
||||
num_tokens_post_padded,
|
||||
B.shape[1],
|
||||
B.shape[2],
|
||||
sorted_token_ids.shape[0],
|
||||
topk_ids.numel(),
|
||||
A.stride(0),
|
||||
A.stride(1),
|
||||
B.stride(0),
|
||||
B.stride(2),
|
||||
B.stride(1),
|
||||
C.stride(1),
|
||||
C.stride(2),
|
||||
B_scale.stride(0) if B_scale is not None and use_int8_w8a16 else 0,
|
||||
B_scale.stride(1) if B_scale is not None and use_int8_w8a16 else 0,
|
||||
MUL_ROUTED_WEIGHT=mul_routed_weight,
|
||||
top_k=top_k,
|
||||
compute_type=compute_type,
|
||||
use_fp8_w8a8=use_fp8_w8a8,
|
||||
use_int8_w8a16=use_int8_w8a16,
|
||||
**config,
|
||||
)
|
||||
|
||||
|
||||
def get_config_file_name(E: int, N: int, dtype: Optional[str]) -> str:
|
||||
device_name = current_platform.get_device_name().replace(" ", "_")
|
||||
dtype_selector = "" if not dtype else f",dtype={dtype}"
|
||||
return f"E={E},N={N},device_name={device_name}{dtype_selector}.json"
|
||||
|
||||
|
||||
@functools.lru_cache
|
||||
def get_moe_configs(E: int, N: int,
|
||||
dtype: Optional[str]) -> Optional[Dict[int, Any]]:
|
||||
"""
|
||||
Return optimized configurations for the fused MoE kernel.
|
||||
|
||||
The return value will be a dictionary that maps an irregular grid of
|
||||
batch sizes to configurations of the fused_moe kernel. To evaluate the
|
||||
kernel on a given batch size bs, the closest batch size in the grid should
|
||||
be picked and the associated configuration chosen to invoke the kernel.
|
||||
"""
|
||||
|
||||
# First look up if an optimized configuration is available in the configs
|
||||
# directory
|
||||
json_file_name = get_config_file_name(E, N, dtype)
|
||||
|
||||
config_file_path = os.path.join(
|
||||
os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name)
|
||||
if os.path.exists(config_file_path):
|
||||
with open(config_file_path) as f:
|
||||
logger.info("Using configuration from %s for MoE layer.",
|
||||
config_file_path)
|
||||
# If a configuration has been found, return it
|
||||
return {int(key): val for key, val in json.load(f).items()}
|
||||
|
||||
# If no optimized configuration is available, we will use the default
|
||||
# configuration
|
||||
logger.warning(
|
||||
("Using default MoE config. Performance might be sub-optimal! "
|
||||
"Config file not found at %s"), config_file_path)
|
||||
return None
|
||||
|
||||
|
||||
def get_default_config(
|
||||
M: int,
|
||||
E: int,
|
||||
N: int,
|
||||
K: int,
|
||||
topk: int,
|
||||
dtype: Optional[str],
|
||||
is_marlin: bool,
|
||||
) -> Dict[str, int]:
|
||||
config = {
|
||||
'BLOCK_SIZE_M': 64,
|
||||
'BLOCK_SIZE_N': 64,
|
||||
'BLOCK_SIZE_K': 32,
|
||||
'GROUP_SIZE_M': 8
|
||||
}
|
||||
# A heuristic: fused marlin works faster with this config for small M
|
||||
if M <= E or (is_marlin and M <= 32):
|
||||
config = {
|
||||
'BLOCK_SIZE_M': 16,
|
||||
'BLOCK_SIZE_N': 32,
|
||||
'BLOCK_SIZE_K': 64,
|
||||
'GROUP_SIZE_M': 1
|
||||
}
|
||||
return config
|
||||
|
||||
|
||||
def try_get_optimal_moe_config(
|
||||
w1_shape: Tuple[int, ...],
|
||||
w2_shape: Tuple[int, ...],
|
||||
top_k: int,
|
||||
dtype: Optional[str],
|
||||
M: int,
|
||||
is_marlin: bool = False,
|
||||
):
|
||||
from vllm.model_executor.layers.fused_moe import get_config
|
||||
override_config = get_config()
|
||||
if override_config:
|
||||
config = override_config
|
||||
else:
|
||||
# First try to load optimal config from the file
|
||||
E, _, N = w2_shape
|
||||
configs = get_moe_configs(E, N, dtype)
|
||||
|
||||
if configs:
|
||||
# If an optimal configuration map has been found, look up the
|
||||
# optimal config
|
||||
config = configs[min(configs.keys(), key=lambda x: abs(x - M))]
|
||||
else:
|
||||
# Else use the default config
|
||||
config = get_default_config(M, E, N, w1_shape[2], top_k, dtype,
|
||||
is_marlin)
|
||||
return config
|
||||
|
||||
|
||||
def fused_topk(
|
||||
hidden_states: torch.Tensor,
|
||||
gating_output: torch.Tensor,
|
||||
topk: int,
|
||||
renormalize: bool,
|
||||
):
|
||||
assert hidden_states.shape[0] == gating_output.shape[0], (
|
||||
"Number of tokens mismatch")
|
||||
|
||||
M, _ = hidden_states.shape
|
||||
|
||||
topk_weights = torch.empty(M,
|
||||
topk,
|
||||
dtype=torch.float32,
|
||||
device=hidden_states.device)
|
||||
topk_ids = torch.empty(M,
|
||||
topk,
|
||||
dtype=torch.int32,
|
||||
device=hidden_states.device)
|
||||
token_expert_indicies = torch.empty(M,
|
||||
topk,
|
||||
dtype=torch.int32,
|
||||
device=hidden_states.device)
|
||||
|
||||
ops.topk_softmax(
|
||||
topk_weights,
|
||||
topk_ids,
|
||||
token_expert_indicies,
|
||||
gating_output.float(), # TODO(woosuk): Optimize this.
|
||||
)
|
||||
del token_expert_indicies # Not used. Will be used in the future.
|
||||
|
||||
if renormalize:
|
||||
topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
|
||||
|
||||
return topk_weights, topk_ids
|
||||
|
||||
|
||||
# This is used by the Deepseek-V2 model
|
||||
def grouped_topk(hidden_states: torch.Tensor,
|
||||
gating_output: torch.Tensor,
|
||||
topk: int,
|
||||
renormalize: bool,
|
||||
num_expert_group: int = 0,
|
||||
topk_group: int = 0):
|
||||
|
||||
assert hidden_states.shape[0] == gating_output.shape[0], (
|
||||
"Number of tokens mismatch")
|
||||
|
||||
scores = torch.softmax(gating_output, dim=-1)
|
||||
num_token = scores.shape[0]
|
||||
group_scores = scores.view(num_token, num_expert_group,
|
||||
-1).max(dim=-1).values # [n, n_group]
|
||||
group_idx = torch.topk(group_scores, k=topk_group, dim=-1,
|
||||
sorted=False)[1] # [n, top_k_group]
|
||||
group_mask = torch.zeros_like(group_scores) # [n, n_group]
|
||||
group_mask.scatter_(1, group_idx, 1) # [n, n_group]
|
||||
score_mask = group_mask.unsqueeze(-1).expand(
|
||||
num_token, num_expert_group,
|
||||
scores.shape[-1] // num_expert_group).reshape(num_token, -1) # [n, e]
|
||||
tmp_scores = scores.masked_fill(~score_mask.bool(), 0.0) # [n, e]
|
||||
topk_weights, topk_ids = torch.topk(tmp_scores,
|
||||
k=topk,
|
||||
dim=-1,
|
||||
sorted=False)
|
||||
|
||||
if renormalize:
|
||||
topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
|
||||
|
||||
return topk_weights.to(torch.float32), topk_ids.to(torch.int32)
|
||||
|
||||
|
||||
def get_config_dtype_str(dtype: torch.dtype,
|
||||
use_int8_w8a16: Optional[bool] = False,
|
||||
use_fp8_w8a8: Optional[bool] = False):
|
||||
if use_fp8_w8a8:
|
||||
return "fp8_w8a8"
|
||||
elif use_int8_w8a16:
|
||||
return "int8_w8a16"
|
||||
elif dtype == torch.float:
|
||||
# avoiding cases where kernel fails when float32 MoE
|
||||
# use fp16/bfloat16 configs
|
||||
return "float32"
|
||||
return None
|
||||
|
||||
|
||||
def inplace_fused_experts(hidden_states: torch.Tensor,
|
||||
w1: torch.Tensor,
|
||||
w2: torch.Tensor,
|
||||
topk_weights: torch.Tensor,
|
||||
topk_ids: torch.Tensor,
|
||||
use_fp8_w8a8: bool = False,
|
||||
use_int8_w8a16: bool = False,
|
||||
w1_scale: Optional[torch.Tensor] = None,
|
||||
w2_scale: Optional[torch.Tensor] = None,
|
||||
a1_scale: Optional[torch.Tensor] = None,
|
||||
a2_scale: Optional[torch.Tensor] = None) -> None:
|
||||
fused_experts_impl(hidden_states, w1, w2, topk_weights, topk_ids, True,
|
||||
use_fp8_w8a8, use_int8_w8a16, w1_scale, w2_scale,
|
||||
a1_scale, a2_scale)
|
||||
|
||||
|
||||
def inplace_fused_experts_fake(
|
||||
hidden_states: torch.Tensor,
|
||||
w1: torch.Tensor,
|
||||
w2: torch.Tensor,
|
||||
topk_weights: torch.Tensor,
|
||||
topk_ids: torch.Tensor,
|
||||
use_fp8_w8a8: bool = False,
|
||||
use_int8_w8a16: bool = False,
|
||||
w1_scale: Optional[torch.Tensor] = None,
|
||||
w2_scale: Optional[torch.Tensor] = None,
|
||||
a1_scale: Optional[torch.Tensor] = None,
|
||||
a2_scale: Optional[torch.Tensor] = None) -> None:
|
||||
pass
|
||||
|
||||
|
||||
direct_register_custom_op(
|
||||
op_name="inplace_fused_experts",
|
||||
op_func=inplace_fused_experts,
|
||||
mutates_args=["hidden_states"],
|
||||
fake_impl=inplace_fused_experts_fake,
|
||||
)
|
||||
|
||||
|
||||
def outplace_fused_experts(
|
||||
hidden_states: torch.Tensor,
|
||||
w1: torch.Tensor,
|
||||
w2: torch.Tensor,
|
||||
topk_weights: torch.Tensor,
|
||||
topk_ids: torch.Tensor,
|
||||
use_fp8_w8a8: bool = False,
|
||||
use_int8_w8a16: bool = False,
|
||||
w1_scale: Optional[torch.Tensor] = None,
|
||||
w2_scale: Optional[torch.Tensor] = None,
|
||||
a1_scale: Optional[torch.Tensor] = None,
|
||||
a2_scale: Optional[torch.Tensor] = None) -> torch.Tensor:
|
||||
return fused_experts_impl(hidden_states, w1, w2, topk_weights, topk_ids,
|
||||
False, use_fp8_w8a8, use_int8_w8a16, w1_scale,
|
||||
w2_scale, a1_scale, a2_scale)
|
||||
|
||||
|
||||
def outplace_fused_experts_fake(
|
||||
hidden_states: torch.Tensor,
|
||||
w1: torch.Tensor,
|
||||
w2: torch.Tensor,
|
||||
topk_weights: torch.Tensor,
|
||||
topk_ids: torch.Tensor,
|
||||
use_fp8_w8a8: bool = False,
|
||||
use_int8_w8a16: bool = False,
|
||||
w1_scale: Optional[torch.Tensor] = None,
|
||||
w2_scale: Optional[torch.Tensor] = None,
|
||||
a1_scale: Optional[torch.Tensor] = None,
|
||||
a2_scale: Optional[torch.Tensor] = None) -> torch.Tensor:
|
||||
return torch.empty_like(hidden_states)
|
||||
|
||||
|
||||
direct_register_custom_op(
|
||||
op_name="outplace_fused_experts",
|
||||
op_func=outplace_fused_experts,
|
||||
mutates_args=[],
|
||||
fake_impl=outplace_fused_experts_fake,
|
||||
)
|
||||
|
||||
|
||||
def fused_experts(hidden_states: torch.Tensor,
|
||||
w1: torch.Tensor,
|
||||
w2: torch.Tensor,
|
||||
topk_weights: torch.Tensor,
|
||||
topk_ids: torch.Tensor,
|
||||
inplace: bool = False,
|
||||
use_fp8_w8a8: bool = False,
|
||||
use_int8_w8a16: bool = False,
|
||||
w1_scale: Optional[torch.Tensor] = None,
|
||||
w2_scale: Optional[torch.Tensor] = None,
|
||||
a1_scale: Optional[torch.Tensor] = None,
|
||||
a2_scale: Optional[torch.Tensor] = None):
|
||||
if inplace:
|
||||
torch.ops.vllm.inplace_fused_experts(hidden_states, w1, w2,
|
||||
topk_weights, topk_ids,
|
||||
use_fp8_w8a8, use_int8_w8a16,
|
||||
w1_scale, w2_scale, a1_scale,
|
||||
a2_scale)
|
||||
return hidden_states
|
||||
else:
|
||||
return torch.ops.vllm.outplace_fused_experts(
|
||||
hidden_states, w1, w2, topk_weights, topk_ids, use_fp8_w8a8,
|
||||
use_int8_w8a16, w1_scale, w2_scale, a1_scale, a2_scale)
|
||||
|
||||
|
||||
def fused_experts_impl(hidden_states: torch.Tensor,
|
||||
w1: torch.Tensor,
|
||||
w2: torch.Tensor,
|
||||
topk_weights: torch.Tensor,
|
||||
topk_ids: torch.Tensor,
|
||||
inplace: bool = False,
|
||||
use_fp8_w8a8: bool = False,
|
||||
use_int8_w8a16: bool = False,
|
||||
w1_scale: Optional[torch.Tensor] = None,
|
||||
w2_scale: Optional[torch.Tensor] = None,
|
||||
a1_scale: Optional[torch.Tensor] = None,
|
||||
a2_scale: Optional[torch.Tensor] = None):
|
||||
# Check constraints.
|
||||
assert hidden_states.shape[1] == w1.shape[2], "Hidden size mismatch"
|
||||
assert topk_weights.shape == topk_ids.shape, "topk shape mismatch"
|
||||
assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
|
||||
assert w1.is_contiguous(), "Expert weights1 must be contiguous"
|
||||
assert w2.is_contiguous(), "Expert weights2 must be contiguous"
|
||||
assert hidden_states.dtype in [
|
||||
torch.float32, torch.float16, torch.bfloat16
|
||||
]
|
||||
|
||||
num_tokens, _ = hidden_states.shape
|
||||
E, N, _ = w1.shape
|
||||
# We execute the fused_moe kernel in chunks to circumvent this issue:
|
||||
# https://github.com/vllm-project/vllm/issues/5938
|
||||
CHUNK_SIZE = envs.VLLM_FUSED_MOE_CHUNK_SIZE
|
||||
M = min(num_tokens, CHUNK_SIZE)
|
||||
config_dtype = get_config_dtype_str(use_fp8_w8a8=use_fp8_w8a8,
|
||||
use_int8_w8a16=use_int8_w8a16,
|
||||
dtype=hidden_states.dtype)
|
||||
|
||||
get_config_func = functools.partial(
|
||||
try_get_optimal_moe_config,
|
||||
w1.shape,
|
||||
w2.shape,
|
||||
topk_ids.shape[1],
|
||||
config_dtype,
|
||||
)
|
||||
|
||||
config = get_config_func(M)
|
||||
|
||||
intermediate_cache1 = torch.empty((M, topk_ids.shape[1], N),
|
||||
device=hidden_states.device,
|
||||
dtype=hidden_states.dtype)
|
||||
intermediate_cache2 = torch.empty((M * topk_ids.shape[1], N // 2),
|
||||
device=hidden_states.device,
|
||||
dtype=hidden_states.dtype)
|
||||
intermediate_cache3 = torch.empty((M, topk_ids.shape[1], w2.shape[1]),
|
||||
device=hidden_states.device,
|
||||
dtype=hidden_states.dtype)
|
||||
|
||||
compute_type = (tl.bfloat16
|
||||
if hidden_states.dtype == torch.bfloat16 else tl.float16)
|
||||
|
||||
if inplace:
|
||||
out_hidden_states = hidden_states
|
||||
else:
|
||||
out_hidden_states = torch.empty_like(hidden_states)
|
||||
|
||||
for chunk in range((num_tokens // CHUNK_SIZE) + 1):
|
||||
begin_chunk_idx, end_chunk_idx = (chunk * CHUNK_SIZE,
|
||||
min((chunk + 1) * CHUNK_SIZE,
|
||||
num_tokens))
|
||||
curr_hidden_states = hidden_states[begin_chunk_idx:end_chunk_idx]
|
||||
tokens_in_chunk, _ = curr_hidden_states.shape
|
||||
|
||||
if tokens_in_chunk == 0:
|
||||
break
|
||||
|
||||
if tokens_in_chunk < CHUNK_SIZE and chunk > 0:
|
||||
# Adjust the intermediate cache size and config for the last
|
||||
# chunk. Note that in most cases we only have one chunk
|
||||
# so the cache size and config are already set correctly and
|
||||
# do not need to be adjusted.
|
||||
intermediate_cache1 = intermediate_cache1[:tokens_in_chunk]
|
||||
intermediate_cache2 = intermediate_cache2[:tokens_in_chunk]
|
||||
intermediate_cache3 = intermediate_cache3[:tokens_in_chunk]
|
||||
config = get_config_func(tokens_in_chunk)
|
||||
|
||||
curr_topk_ids = topk_ids[begin_chunk_idx:end_chunk_idx]
|
||||
curr_topk_weights = topk_weights[begin_chunk_idx:end_chunk_idx]
|
||||
|
||||
sorted_token_ids, expert_ids, num_tokens_post_padded = (
|
||||
moe_align_block_size(curr_topk_ids, config['BLOCK_SIZE_M'], E))
|
||||
|
||||
invoke_fused_moe_kernel(curr_hidden_states,
|
||||
w1,
|
||||
intermediate_cache1,
|
||||
a1_scale,
|
||||
w1_scale,
|
||||
curr_topk_weights,
|
||||
curr_topk_ids,
|
||||
sorted_token_ids,
|
||||
expert_ids,
|
||||
num_tokens_post_padded,
|
||||
False,
|
||||
topk_ids.shape[1],
|
||||
config,
|
||||
compute_type=compute_type,
|
||||
use_fp8_w8a8=use_fp8_w8a8,
|
||||
use_int8_w8a16=use_int8_w8a16)
|
||||
|
||||
ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
|
||||
|
||||
invoke_fused_moe_kernel(intermediate_cache2,
|
||||
w2,
|
||||
intermediate_cache3,
|
||||
a2_scale,
|
||||
w2_scale,
|
||||
curr_topk_weights,
|
||||
curr_topk_ids,
|
||||
sorted_token_ids,
|
||||
expert_ids,
|
||||
num_tokens_post_padded,
|
||||
True,
|
||||
1,
|
||||
config,
|
||||
compute_type=compute_type,
|
||||
use_fp8_w8a8=use_fp8_w8a8,
|
||||
use_int8_w8a16=use_int8_w8a16)
|
||||
|
||||
ops.moe_sum(intermediate_cache3.view(*intermediate_cache3.shape),
|
||||
out_hidden_states[begin_chunk_idx:end_chunk_idx])
|
||||
return out_hidden_states
|
||||
|
||||
|
||||
def fused_moe(
|
||||
hidden_states: torch.Tensor,
|
||||
w1: torch.Tensor,
|
||||
w2: torch.Tensor,
|
||||
gating_output: torch.Tensor,
|
||||
topk: int,
|
||||
renormalize: bool,
|
||||
inplace: bool = False,
|
||||
use_grouped_topk: bool = False,
|
||||
num_expert_group: Optional[int] = None,
|
||||
topk_group: Optional[int] = None,
|
||||
custom_routing_function: Optional[Callable] = None,
|
||||
use_fp8_w8a8: bool = False,
|
||||
use_int8_w8a16: bool = False,
|
||||
w1_scale: Optional[torch.Tensor] = None,
|
||||
w2_scale: Optional[torch.Tensor] = None,
|
||||
a1_scale: Optional[torch.Tensor] = None,
|
||||
a2_scale: Optional[torch.Tensor] = None,
|
||||
) -> torch.Tensor:
|
||||
"""
|
||||
This function computes a Mixture of Experts (MoE) layer using two sets of
|
||||
weights, w1 and w2, and top-k gating mechanism.
|
||||
|
||||
Parameters:
|
||||
- hidden_states (torch.Tensor): The input tensor to the MoE layer.
|
||||
- w1 (torch.Tensor): The first set of expert weights.
|
||||
- w2 (torch.Tensor): The second set of expert weights.
|
||||
- gating_output (torch.Tensor): The output of the gating operation
|
||||
(before softmax).
|
||||
- topk (int): The number of top-k experts to select.
|
||||
- renormalize (bool): If True, renormalize the top-k weights to sum to 1.
|
||||
- inplace (bool): If True, perform the operation in-place.
|
||||
Defaults to False.
|
||||
- num_expert_group: Optional[int]: additional parameter for grouped_topk
|
||||
- topk_group: Optional[int]: additional parameter for grouped_topk
|
||||
- use_grouped_topk: If True, use grouped_topk instead of fused_topk
|
||||
note: Deepseekv2 model uses grouped_topk
|
||||
- use_fp8_w8a8 (bool): If True, use fp8 arithmetic to compute the inner
|
||||
products for w1 and w2. Defaults to False.
|
||||
- use_int8_w8a16 (bool): If True, use fp8 arithmetic to compute the inner
|
||||
products for w1 and w2. Defaults to False.
|
||||
- w1_scale (Optional[torch.Tensor]): Optional scale to be used for
|
||||
w1.
|
||||
- w2_scale (Optional[torch.Tensor]): Optional scale to be used for
|
||||
w2.
|
||||
|
||||
Returns:
|
||||
- torch.Tensor: The output tensor after applying the MoE layer.
|
||||
"""
|
||||
# Check constraints.
|
||||
assert gating_output.shape[1] == w1.shape[0], "Number of experts mismatch"
|
||||
|
||||
if use_grouped_topk:
|
||||
assert num_expert_group is not None and topk_group is not None
|
||||
topk_weights, topk_ids = grouped_topk(hidden_states, gating_output,
|
||||
topk, renormalize,
|
||||
num_expert_group, topk_group)
|
||||
elif custom_routing_function is None:
|
||||
topk_weights, topk_ids = fused_topk(hidden_states, gating_output, topk,
|
||||
renormalize)
|
||||
else:
|
||||
topk_weights, topk_ids = custom_routing_function(
|
||||
hidden_states, gating_output, topk, renormalize)
|
||||
|
||||
return fused_experts(hidden_states,
|
||||
w1,
|
||||
w2,
|
||||
topk_weights,
|
||||
topk_ids,
|
||||
inplace=inplace,
|
||||
use_fp8_w8a8=use_fp8_w8a8,
|
||||
use_int8_w8a16=use_int8_w8a16,
|
||||
w1_scale=w1_scale,
|
||||
w2_scale=w2_scale,
|
||||
a1_scale=a1_scale,
|
||||
a2_scale=a2_scale)
|
||||
566
vllm-v0.6.2/vllm/model_executor/layers/fused_moe/layer.py
Normal file
566
vllm-v0.6.2/vllm/model_executor/layers/fused_moe/layer.py
Normal file
@@ -0,0 +1,566 @@
|
||||
from abc import abstractmethod
|
||||
from enum import Enum
|
||||
from typing import Callable, List, Optional, Tuple
|
||||
|
||||
import torch
|
||||
|
||||
from vllm.distributed import (get_tensor_model_parallel_rank,
|
||||
get_tensor_model_parallel_world_size,
|
||||
tensor_model_parallel_all_reduce)
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.custom_op import CustomOp
|
||||
from vllm.model_executor.layers.quantization.base_config import (
|
||||
QuantizationConfig, QuantizeMethodBase)
|
||||
from vllm.model_executor.utils import set_weight_attrs
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
if current_platform.is_cuda_alike():
|
||||
from .fused_moe import fused_experts
|
||||
else:
|
||||
fused_experts = None # type: ignore
|
||||
if current_platform.is_tpu():
|
||||
from .moe_pallas import fused_moe as fused_moe_pallas
|
||||
else:
|
||||
fused_moe_pallas = None # type: ignore
|
||||
logger = init_logger(__name__)
|
||||
|
||||
|
||||
class FusedMoeWeightScaleSupported(Enum):
|
||||
TENSOR = "tensor"
|
||||
CHANNEL = "channel"
|
||||
GROUP = "group"
|
||||
|
||||
|
||||
class FusedMoEMethodBase(QuantizeMethodBase):
|
||||
|
||||
@abstractmethod
|
||||
def create_weights(self, layer: torch.nn.Module, num_experts: int,
|
||||
hidden_size: int, intermediate_size: int,
|
||||
params_dtype: torch.dtype, **extra_weight_attrs):
|
||||
raise NotImplementedError
|
||||
|
||||
@abstractmethod
|
||||
def apply(self, layer: torch.nn.Module, x: torch.Tensor,
|
||||
router_logits: torch.Tensor, top_k: int, renormalize: bool,
|
||||
use_grouped_topk: bool) -> torch.Tensor:
|
||||
raise NotImplementedError
|
||||
|
||||
|
||||
@CustomOp.register("unquantized_fused_moe")
|
||||
class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
|
||||
"""MoE method without quantization."""
|
||||
|
||||
def create_weights(self, layer: torch.nn.Module, num_experts: int,
|
||||
hidden_size: int, intermediate_size: int,
|
||||
params_dtype: torch.dtype, **extra_weight_attrs):
|
||||
# Fused gate_up_proj (column parallel)
|
||||
w13_weight = torch.nn.Parameter(torch.empty(num_experts,
|
||||
2 * intermediate_size,
|
||||
hidden_size,
|
||||
dtype=params_dtype),
|
||||
requires_grad=False)
|
||||
layer.register_parameter("w13_weight", w13_weight)
|
||||
set_weight_attrs(w13_weight, extra_weight_attrs)
|
||||
|
||||
# down_proj (row parallel)
|
||||
w2_weight = torch.nn.Parameter(torch.empty(num_experts,
|
||||
hidden_size,
|
||||
intermediate_size,
|
||||
dtype=params_dtype),
|
||||
requires_grad=False)
|
||||
layer.register_parameter("w2_weight", w2_weight)
|
||||
set_weight_attrs(w2_weight, extra_weight_attrs)
|
||||
|
||||
def apply(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
x: torch.Tensor,
|
||||
router_logits: torch.Tensor,
|
||||
top_k: int,
|
||||
renormalize: bool,
|
||||
use_grouped_topk: bool,
|
||||
topk_group: Optional[int] = None,
|
||||
num_expert_group: Optional[int] = None,
|
||||
custom_routing_function: Optional[Callable] = None
|
||||
) -> torch.Tensor:
|
||||
return self.forward(x=x,
|
||||
layer=layer,
|
||||
router_logits=router_logits,
|
||||
top_k=top_k,
|
||||
renormalize=renormalize,
|
||||
use_grouped_topk=use_grouped_topk,
|
||||
topk_group=topk_group,
|
||||
num_expert_group=num_expert_group,
|
||||
custom_routing_function=custom_routing_function)
|
||||
|
||||
def forward_cuda(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
x: torch.Tensor,
|
||||
use_grouped_topk: bool,
|
||||
top_k: int,
|
||||
router_logits: torch.Tensor,
|
||||
renormalize: bool,
|
||||
topk_group: Optional[int] = None,
|
||||
num_expert_group: Optional[int] = None,
|
||||
custom_routing_function: Optional[Callable] = None
|
||||
) -> torch.Tensor:
|
||||
topk_weights, topk_ids = FusedMoE.select_experts(
|
||||
hidden_states=x,
|
||||
router_logits=router_logits,
|
||||
use_grouped_topk=use_grouped_topk,
|
||||
top_k=top_k,
|
||||
renormalize=renormalize,
|
||||
topk_group=topk_group,
|
||||
num_expert_group=num_expert_group,
|
||||
custom_routing_function=custom_routing_function)
|
||||
|
||||
return fused_experts(hidden_states=x,
|
||||
w1=layer.w13_weight,
|
||||
w2=layer.w2_weight,
|
||||
topk_weights=topk_weights,
|
||||
topk_ids=topk_ids,
|
||||
inplace=True)
|
||||
|
||||
def forward_cpu(self, *args, **kwargs):
|
||||
raise NotImplementedError(
|
||||
"The CPU backend currently does not support MoE.")
|
||||
|
||||
def forward_tpu(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
x: torch.Tensor,
|
||||
use_grouped_topk: bool,
|
||||
top_k: int,
|
||||
router_logits: torch.Tensor,
|
||||
renormalize: bool,
|
||||
topk_group: Optional[int] = None,
|
||||
num_expert_group: Optional[int] = None,
|
||||
custom_routing_function: Optional[Callable] = None
|
||||
) -> torch.Tensor:
|
||||
assert not use_grouped_topk
|
||||
assert num_expert_group is None
|
||||
assert topk_group is None
|
||||
assert custom_routing_function is None
|
||||
return fused_moe_pallas(hidden_states=x,
|
||||
w1=layer.w13_weight,
|
||||
w2=layer.w2_weight,
|
||||
topk=top_k,
|
||||
gating_output=router_logits,
|
||||
renormalize=renormalize)
|
||||
|
||||
forward_native = forward_cuda
|
||||
|
||||
def forward_mlu(
|
||||
self,
|
||||
x: torch.Tensor,
|
||||
w1: torch.Tensor,
|
||||
w2: torch.Tensor,
|
||||
router_logits: torch.Tensor,
|
||||
top_k: int,
|
||||
renormalize: bool,
|
||||
use_grouped_topk: bool,
|
||||
num_expert_group: Optional[int],
|
||||
topk_group: Optional[int],
|
||||
) -> torch.Tensor:
|
||||
from vllm._mlu_ops import fused_moe
|
||||
|
||||
assert use_grouped_topk is False and num_expert_group is None and topk_group is None, \
|
||||
f"Following params: use_grouped_topk, num_expert_group, topk_group are not support yet."
|
||||
return fused_moe(x,
|
||||
router_logits,
|
||||
w1, w2,
|
||||
None, None, # bias1, bias2
|
||||
None, # residual
|
||||
None, # input_smooth
|
||||
None, # act_smooth
|
||||
None, None, # w1_scale, w2_scale
|
||||
top_k,
|
||||
renormalize,
|
||||
True, # gated
|
||||
'silu')
|
||||
|
||||
|
||||
class FusedMoE(torch.nn.Module):
|
||||
"""FusedMoE layer for MoE models.
|
||||
|
||||
This layer contains both MergedColumnParallel weights (gate_up_proj /
|
||||
w13) and RowParallelLinear weights (down_proj/ w2).
|
||||
|
||||
Note: Mixtral uses w1, w2, and w3 for gate, up, and down_proj. We
|
||||
copy that naming convention here and handle any remapping in the
|
||||
load_weights function in each model implementation.
|
||||
|
||||
Args:
|
||||
num_experts: Number of experts in the model
|
||||
top_k: Number of experts selected for each token
|
||||
hidden_size: Input hidden state size of the transformer
|
||||
intermediate_size: Intermediate size of the experts
|
||||
params_dtype: Data type for the parameters.
|
||||
reduce_results: Whether to all all_reduce on the output of the layer
|
||||
renomalize: Whether to renormalize the logits in the fused_moe kernel
|
||||
quant_config: Quantization configure.
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
num_experts: int,
|
||||
top_k: int,
|
||||
hidden_size: int,
|
||||
intermediate_size: int,
|
||||
params_dtype: Optional[torch.dtype] = None,
|
||||
reduce_results: bool = False,
|
||||
renormalize: bool = True,
|
||||
use_grouped_topk: bool = False,
|
||||
num_expert_group: Optional[int] = None,
|
||||
topk_group: Optional[int] = None,
|
||||
quant_config: Optional[QuantizationConfig] = None,
|
||||
tp_size: Optional[int] = None,
|
||||
prefix: str = "",
|
||||
custom_routing_function: Optional[Callable] = None,
|
||||
):
|
||||
super().__init__()
|
||||
|
||||
if params_dtype is None:
|
||||
params_dtype = torch.get_default_dtype()
|
||||
|
||||
self.tp_size = (tp_size if tp_size is not None else
|
||||
get_tensor_model_parallel_world_size())
|
||||
self.top_k = top_k
|
||||
self.num_experts = num_experts
|
||||
self.intermediate_size_per_partition = intermediate_size // self.tp_size
|
||||
self.reduce_results = reduce_results
|
||||
self.renormalize = renormalize
|
||||
self.use_grouped_topk = use_grouped_topk
|
||||
if self.use_grouped_topk:
|
||||
assert num_expert_group is not None and topk_group is not None
|
||||
self.num_expert_group = num_expert_group
|
||||
self.topk_group = topk_group
|
||||
self.custom_routing_function = custom_routing_function
|
||||
|
||||
if quant_config is None:
|
||||
self.quant_method: Optional[QuantizeMethodBase] = (
|
||||
UnquantizedFusedMoEMethod())
|
||||
else:
|
||||
self.quant_method = quant_config.get_quant_method(self, prefix)
|
||||
assert self.quant_method is not None
|
||||
|
||||
self.quant_method.create_weights(
|
||||
layer=self,
|
||||
num_experts=num_experts,
|
||||
hidden_size=hidden_size,
|
||||
intermediate_size=self.intermediate_size_per_partition,
|
||||
params_dtype=params_dtype,
|
||||
weight_loader=self.weight_loader)
|
||||
|
||||
def _load_per_tensor_weight_scale(self, shard_id: str,
|
||||
param: torch.nn.Parameter,
|
||||
loaded_weight: torch.Tensor,
|
||||
expert_id: int):
|
||||
param_data = param.data
|
||||
# for per tensor weight quantization
|
||||
if shard_id in ("w1", "w3"):
|
||||
# We have to keep the weight scales of w1 and w3 because
|
||||
# we need to re-quantize w1/w3 weights after weight loading.
|
||||
idx = 0 if shard_id == "w1" else 1
|
||||
param_data[expert_id][idx] = loaded_weight
|
||||
# If we are in the row parallel case (down_proj)
|
||||
elif shard_id == "w2":
|
||||
param_data[expert_id] = loaded_weight
|
||||
|
||||
def _load_model_weight_or_group_weight_scale(self, shard_dim: int,
|
||||
expert_data: torch.Tensor,
|
||||
shard_id: str,
|
||||
loaded_weight: torch.tensor,
|
||||
tp_rank: int):
|
||||
# Load grouped weight scales for group quantization
|
||||
# or model weights
|
||||
if shard_id == "w2":
|
||||
self._load_w2(shard_id=shard_id,
|
||||
shard_dim=shard_dim,
|
||||
loaded_weight=loaded_weight,
|
||||
expert_data=expert_data,
|
||||
tp_rank=tp_rank)
|
||||
elif shard_id in ("w1", "w3"):
|
||||
self._load_w13(shard_id=shard_id,
|
||||
shard_dim=shard_dim,
|
||||
loaded_weight=loaded_weight,
|
||||
expert_data=expert_data,
|
||||
tp_rank=tp_rank)
|
||||
|
||||
def _load_per_channel_weight_scale(self, expert_data: torch.Tensor,
|
||||
shard_dim: int, shard_id: str,
|
||||
loaded_weight: torch.tensor,
|
||||
tp_rank: int):
|
||||
# for per channel weight quantization
|
||||
if shard_id == "w2":
|
||||
expert_data.copy_(loaded_weight)
|
||||
elif shard_id in ("w1", "w3"):
|
||||
self._load_w13(shard_id=shard_id,
|
||||
shard_dim=shard_dim,
|
||||
loaded_weight=loaded_weight,
|
||||
expert_data=expert_data,
|
||||
tp_rank=tp_rank)
|
||||
|
||||
def _load_w13(self, expert_data: torch.Tensor, shard_dim: int,
|
||||
shard_id: str, loaded_weight: torch.tensor, tp_rank: int):
|
||||
|
||||
# Index the loaded weight for tp sharding.
|
||||
# gate_up_proj: "MergedColumnParallel", so tp sharding on output_dim
|
||||
shard_size = expert_data.shape[shard_dim] // 2
|
||||
loaded_weight = loaded_weight.narrow(shard_dim, shard_size * tp_rank,
|
||||
shard_size)
|
||||
# Narrow parameter and load.
|
||||
# w1, gate_proj: Load into first logical weight of w13.
|
||||
if shard_id == "w1":
|
||||
expert_data = expert_data.narrow(shard_dim, 0, shard_size)
|
||||
# w3, up_proj: Load into second logical weight of w13.
|
||||
else:
|
||||
assert shard_id == "w3"
|
||||
expert_data = expert_data.narrow(shard_dim, shard_size, shard_size)
|
||||
expert_data.copy_(loaded_weight)
|
||||
|
||||
def _load_w2(self, expert_data: torch.Tensor, shard_dim: int,
|
||||
shard_id: str, loaded_weight: torch.tensor, tp_rank: int):
|
||||
|
||||
# Index the loaded weight for tp sharding.
|
||||
# down_proj: "RowParallel" so tp sharding on input_dim
|
||||
# Narrow parameter and load.
|
||||
shard_size = expert_data.shape[shard_dim]
|
||||
loaded_weight = loaded_weight.narrow(shard_dim, shard_size * tp_rank,
|
||||
shard_size)
|
||||
# w2, down_proj: Load into only logical weight of w2.
|
||||
expert_data.copy_(loaded_weight)
|
||||
|
||||
def _load_single_value(self, param: torch.nn.Parameter,
|
||||
loaded_weight: torch.Tensor, expert_id: int):
|
||||
param_data = param.data
|
||||
|
||||
# Input scales can be loaded directly and should be equal.
|
||||
param_data[expert_id] = loaded_weight
|
||||
|
||||
def _load_g_idx(self, shard_id: str, expert_data: torch.Tensor,
|
||||
shard_dim: int, loaded_weight: torch.tensor, tp_rank: int):
|
||||
|
||||
if shard_id == "w2":
|
||||
self._load_w2(shard_id=shard_id,
|
||||
shard_dim=shard_dim,
|
||||
loaded_weight=loaded_weight,
|
||||
expert_data=expert_data,
|
||||
tp_rank=tp_rank)
|
||||
else:
|
||||
assert shard_id in ("w1", "w3")
|
||||
expert_data.copy_(loaded_weight)
|
||||
|
||||
def weight_loader(self, param: torch.nn.Parameter,
|
||||
loaded_weight: torch.Tensor, weight_name: str,
|
||||
shard_id: str, expert_id: int) -> None:
|
||||
|
||||
# compressed-tensors checkpoints with packed weights are stored flipped
|
||||
# TODO (mgoin): check self.quant_method.quant_config.quant_format
|
||||
# against known CompressionFormat enum values that have this quality
|
||||
loaded_weight = loaded_weight.t().contiguous() if (
|
||||
self.quant_method.__class__.__name__
|
||||
== "CompressedTensorsWNA16MoEMethod") else loaded_weight
|
||||
|
||||
if shard_id not in ("w1", "w2", "w3"):
|
||||
raise ValueError(f"shard_id must be ['w1','w2','w3'] but "
|
||||
f"got {shard_id}.")
|
||||
|
||||
WEIGHT_SCALE_SUPPORTED = [
|
||||
e.value for e in FusedMoeWeightScaleSupported
|
||||
]
|
||||
# Fetch the dim to shard the parameter/loaded weight
|
||||
# based on the shard id. This will be whatever
|
||||
# dimension intermediate_size is used.
|
||||
SHARD_ID_TO_SHARDED_DIM = {"w1": 0, "w2": 1, "w3": 0}
|
||||
|
||||
expert_data = param.data[expert_id]
|
||||
tp_rank = get_tensor_model_parallel_rank()
|
||||
|
||||
# is_transposed: if the dim to shard the weight
|
||||
# should be flipped. Required by GPTQ, compressed-tensors
|
||||
# should be whatever dimension intermediate_size is
|
||||
is_transposed = getattr(param, "is_transposed", False)
|
||||
shard_dim = SHARD_ID_TO_SHARDED_DIM[shard_id]
|
||||
if is_transposed:
|
||||
shard_dim = ~shard_dim
|
||||
|
||||
# Case input scale: input_scale loading is only supported for fp8
|
||||
if "input_scale" in weight_name:
|
||||
# this is needed for compressed-tensors only
|
||||
loaded_weight = loaded_weight.to(param.data.device)
|
||||
|
||||
if param.data[expert_id] != 1 and (param.data[expert_id] -
|
||||
loaded_weight).abs() > 1e-5:
|
||||
raise ValueError(
|
||||
"input_scales of w1 and w3 of a layer "
|
||||
f"must be equal. But got {param.data[expert_id]} "
|
||||
f"vs. {loaded_weight}")
|
||||
|
||||
self._load_single_value(param=param,
|
||||
loaded_weight=loaded_weight,
|
||||
expert_id=expert_id)
|
||||
return
|
||||
|
||||
# Case g_idx
|
||||
if "g_idx" in weight_name:
|
||||
self._load_g_idx(shard_dim=0,
|
||||
shard_id=shard_id,
|
||||
loaded_weight=loaded_weight,
|
||||
expert_data=expert_data,
|
||||
tp_rank=tp_rank)
|
||||
return
|
||||
|
||||
# Case weight scales and zero_points
|
||||
if ("scale" in weight_name or "zero" in weight_name):
|
||||
# load the weight scales and zp based on the quantization scheme
|
||||
# supported weight scales/zp can be found in
|
||||
# FusedMoeWeightScaleSupported
|
||||
# TODO @dsikka: once hardened, refactor to use vLLM Parameters
|
||||
# specific to each case
|
||||
quant_method = getattr(param, "quant_method", None)
|
||||
if quant_method == FusedMoeWeightScaleSupported.CHANNEL.value:
|
||||
self._load_per_channel_weight_scale(
|
||||
shard_id=shard_id,
|
||||
shard_dim=shard_dim,
|
||||
loaded_weight=loaded_weight,
|
||||
expert_data=expert_data,
|
||||
tp_rank=tp_rank)
|
||||
elif quant_method == FusedMoeWeightScaleSupported.GROUP.value:
|
||||
self._load_model_weight_or_group_weight_scale(
|
||||
shard_id=shard_id,
|
||||
shard_dim=shard_dim,
|
||||
loaded_weight=loaded_weight,
|
||||
expert_data=expert_data,
|
||||
tp_rank=tp_rank)
|
||||
elif quant_method == FusedMoeWeightScaleSupported.TENSOR.value:
|
||||
self._load_per_tensor_weight_scale(shard_id=shard_id,
|
||||
param=param,
|
||||
loaded_weight=loaded_weight,
|
||||
expert_id=expert_id)
|
||||
else:
|
||||
raise ValueError(
|
||||
f"quant method must be one of {WEIGHT_SCALE_SUPPORTED}")
|
||||
return
|
||||
|
||||
# Case weight_shape
|
||||
if "weight_shape" in weight_name:
|
||||
# only required by compressed-tensors
|
||||
self._load_single_value(param=param,
|
||||
loaded_weight=loaded_weight,
|
||||
expert_id=expert_id)
|
||||
return
|
||||
|
||||
# Case model weights
|
||||
if "weight" in weight_name:
|
||||
self._load_model_weight_or_group_weight_scale(
|
||||
shard_id=shard_id,
|
||||
shard_dim=shard_dim,
|
||||
loaded_weight=loaded_weight,
|
||||
expert_data=expert_data,
|
||||
tp_rank=tp_rank)
|
||||
return
|
||||
|
||||
@staticmethod
|
||||
def select_experts(hidden_states: torch.Tensor,
|
||||
router_logits: torch.Tensor,
|
||||
top_k: int,
|
||||
use_grouped_topk: bool,
|
||||
renormalize: bool,
|
||||
topk_group: Optional[int] = None,
|
||||
num_expert_group: Optional[int] = None,
|
||||
custom_routing_function: Optional[Callable] = None):
|
||||
from vllm.model_executor.layers.fused_moe.fused_moe import (
|
||||
fused_topk, grouped_topk)
|
||||
|
||||
# DeekSeekv2 uses grouped_top_k
|
||||
if use_grouped_topk:
|
||||
assert topk_group is not None
|
||||
assert num_expert_group is not None
|
||||
topk_weights, topk_ids = grouped_topk(
|
||||
hidden_states=hidden_states,
|
||||
gating_output=router_logits,
|
||||
topk=top_k,
|
||||
renormalize=renormalize,
|
||||
num_expert_group=num_expert_group,
|
||||
topk_group=topk_group)
|
||||
elif custom_routing_function is None:
|
||||
topk_weights, topk_ids = fused_topk(hidden_states=hidden_states,
|
||||
gating_output=router_logits,
|
||||
topk=top_k,
|
||||
renormalize=renormalize)
|
||||
else:
|
||||
topk_weights, topk_ids = custom_routing_function(
|
||||
hidden_states=hidden_states,
|
||||
gating_output=router_logits,
|
||||
topk=top_k,
|
||||
renormalize=renormalize)
|
||||
|
||||
return topk_weights, topk_ids
|
||||
|
||||
def forward(self, hidden_states: torch.Tensor,
|
||||
router_logits: torch.Tensor):
|
||||
assert self.quant_method is not None
|
||||
|
||||
# Matrix multiply.
|
||||
final_hidden_states = self.quant_method.apply(
|
||||
layer=self,
|
||||
x=hidden_states,
|
||||
router_logits=router_logits,
|
||||
top_k=self.top_k,
|
||||
renormalize=self.renormalize,
|
||||
use_grouped_topk=self.use_grouped_topk,
|
||||
topk_group=self.topk_group,
|
||||
num_expert_group=self.num_expert_group,
|
||||
custom_routing_function=self.custom_routing_function)
|
||||
|
||||
if self.reduce_results and self.tp_size > 1:
|
||||
final_hidden_states = tensor_model_parallel_all_reduce(
|
||||
final_hidden_states)
|
||||
|
||||
return final_hidden_states
|
||||
|
||||
@classmethod
|
||||
def make_expert_params_mapping(
|
||||
cls, ckpt_gate_proj_name: str, ckpt_down_proj_name: str,
|
||||
ckpt_up_proj_name: str,
|
||||
num_experts: int) -> List[Tuple[str, str, int, str]]:
|
||||
|
||||
return [
|
||||
# (param_name, weight_name, expert_id, shard_id)
|
||||
("experts.w13_" if weight_name
|
||||
in [ckpt_gate_proj_name, ckpt_up_proj_name] else "experts.w2_",
|
||||
f"experts.{expert_id}.{weight_name}.", expert_id, shard_id)
|
||||
for expert_id in range(num_experts) for shard_id, weight_name in [
|
||||
("w1", ckpt_gate_proj_name),
|
||||
("w2", ckpt_down_proj_name),
|
||||
("w3", ckpt_up_proj_name),
|
||||
]
|
||||
]
|
||||
|
||||
def _load_fp8_scale(self, param: torch.nn.Parameter,
|
||||
loaded_weight: torch.Tensor, weight_name: str,
|
||||
shard_id: str, expert_id: int) -> None:
|
||||
param_data = param.data
|
||||
|
||||
# Input scales can be loaded directly and should be equal.
|
||||
if "input_scale" in weight_name:
|
||||
if param_data[expert_id] != 1 and (param_data[expert_id] -
|
||||
loaded_weight).abs() > 1e-5:
|
||||
raise ValueError(
|
||||
"input_scales of w1 and w3 of a layer "
|
||||
f"must be equal. But got {param_data[expert_id]} "
|
||||
f"vs. {loaded_weight}")
|
||||
param_data[expert_id] = loaded_weight
|
||||
# Weight scales
|
||||
elif "weight_scale" in weight_name:
|
||||
# If we are in merged column case (gate_up_proj)
|
||||
if shard_id in ("w1", "w3"):
|
||||
# We have to keep the weight scales of w1 and w3 because
|
||||
# we need to re-quantize w1/w3 weights after weight loading.
|
||||
idx = 0 if shard_id == "w1" else 1
|
||||
param_data[expert_id][idx] = loaded_weight
|
||||
# If we are in the row parallel case (down_proj)
|
||||
else:
|
||||
param_data[expert_id] = loaded_weight
|
||||
@@ -0,0 +1,62 @@
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
from torch_xla.experimental.custom_kernel import _histogram
|
||||
|
||||
|
||||
def fused_moe(
|
||||
hidden_states: torch.Tensor,
|
||||
w1: torch.Tensor,
|
||||
w2: torch.Tensor,
|
||||
gating_output: torch.Tensor,
|
||||
topk: int,
|
||||
renormalize: bool,
|
||||
) -> torch.Tensor:
|
||||
"""
|
||||
Args:
|
||||
hidden_states: [*, hidden_size]
|
||||
w1: [num_experts, intermediate_size * 2, hidden_size]
|
||||
w2: [num_experts, hidden_size, intermediate_size]
|
||||
gating_output: [*, num_experts]
|
||||
"""
|
||||
orig_shape = hidden_states.shape
|
||||
hidden_size = hidden_states.shape[-1]
|
||||
num_tokens = hidden_states.shape[:-1].numel()
|
||||
num_experts = w1.shape[0]
|
||||
intermediate_size = w2.shape[-1]
|
||||
device = hidden_states.device
|
||||
dtype = hidden_states.dtype
|
||||
assert (num_tokens * topk) % 16 == 0, (
|
||||
"The Pallas GMM kernel requires num_tokens * topk to be a multiple of "
|
||||
f"16 but got {num_tokens * topk}")
|
||||
|
||||
hidden_states = hidden_states.view(num_tokens, hidden_size)
|
||||
gating_output = gating_output.view(num_tokens, num_experts)
|
||||
topk_weights = gating_output.softmax(dim=-1, dtype=torch.float)
|
||||
topk_weights, topk_indices = topk_weights.topk(topk, dim=-1)
|
||||
if renormalize:
|
||||
topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
|
||||
topk_weights = topk_weights.to(dtype)
|
||||
|
||||
topk_indices = topk_indices.flatten()
|
||||
topk_argsort_indices = topk_indices.argsort()
|
||||
topk_argsort_revert_indices = topk_argsort_indices.argsort()
|
||||
token_indices = torch.arange(num_tokens,
|
||||
device=device).repeat_interleave(topk)
|
||||
token_indices = token_indices[topk_argsort_indices]
|
||||
group_sizes = _histogram(topk_indices.to(torch.int32), 0, num_experts - 1)
|
||||
|
||||
# NOTE(woosuk): The GMM Pallas kernel requires a different weight layout
|
||||
# from HF Transformers.
|
||||
w1 = w1.transpose(1, 2)
|
||||
w2 = w2.transpose(1, 2)
|
||||
|
||||
x = hidden_states[token_indices]
|
||||
x = torch.ops.xla.gmm(x, w1, group_sizes)
|
||||
x = F.silu(x[..., :intermediate_size]) * x[..., intermediate_size:]
|
||||
x = torch.ops.xla.gmm(x, w2, group_sizes)
|
||||
x = x[topk_argsort_revert_indices].reshape(-1, topk, hidden_size)
|
||||
|
||||
x = x * topk_weights.unsqueeze_(dim=-1)
|
||||
x = x.sum(dim=-2)
|
||||
x = x.reshape(orig_shape)
|
||||
return x
|
||||
219
vllm-v0.6.2/vllm/model_executor/layers/layernorm.py
Normal file
219
vllm-v0.6.2/vllm/model_executor/layers/layernorm.py
Normal file
@@ -0,0 +1,219 @@
|
||||
"""Custom normalization layers."""
|
||||
from typing import Optional, Tuple, Union
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
|
||||
from vllm.model_executor.custom_op import CustomOp
|
||||
|
||||
|
||||
@CustomOp.register("rms_norm")
|
||||
class RMSNorm(CustomOp):
|
||||
"""Root mean square normalization.
|
||||
|
||||
Computes x -> w * x / sqrt(E[x^2] + eps) where w is the learned weight.
|
||||
Refer to https://arxiv.org/abs/1910.07467
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
hidden_size: int,
|
||||
eps: float = 1e-6,
|
||||
var_hidden_size: Optional[int] = None,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
|
||||
self.hidden_size = hidden_size
|
||||
self.variance_epsilon = eps
|
||||
self.variance_size_override = (None if var_hidden_size == hidden_size
|
||||
else var_hidden_size)
|
||||
self.weight = nn.Parameter(torch.ones(hidden_size))
|
||||
|
||||
def forward_native(
|
||||
self,
|
||||
x: torch.Tensor,
|
||||
residual: Optional[torch.Tensor] = None,
|
||||
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
||||
"""PyTorch-native implementation equivalent to forward()."""
|
||||
orig_dtype = x.dtype
|
||||
x = x.to(torch.float32)
|
||||
if residual is not None:
|
||||
x = x + residual.to(torch.float32)
|
||||
residual = x.to(orig_dtype)
|
||||
|
||||
hidden_size = x.shape[-1]
|
||||
if hidden_size != self.hidden_size:
|
||||
raise ValueError("Expected hidden_size to be "
|
||||
f"{self.hidden_size}, but found: {hidden_size}")
|
||||
|
||||
if self.variance_size_override is None:
|
||||
x_var = x
|
||||
else:
|
||||
if hidden_size < self.variance_size_override:
|
||||
raise ValueError(
|
||||
"Expected hidden_size to be at least "
|
||||
f"{self.variance_size_override}, but found: {hidden_size}")
|
||||
|
||||
x_var = x[:, :, :self.variance_size_override]
|
||||
|
||||
variance = x_var.pow(2).mean(dim=-1, keepdim=True)
|
||||
|
||||
x = x * torch.rsqrt(variance + self.variance_epsilon)
|
||||
x = x.to(orig_dtype) * self.weight
|
||||
if residual is None:
|
||||
return x
|
||||
else:
|
||||
return x, residual
|
||||
|
||||
def forward_cuda(
|
||||
self,
|
||||
x: torch.Tensor,
|
||||
residual: Optional[torch.Tensor] = None,
|
||||
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
||||
if self.variance_size_override is not None:
|
||||
return self.forward_native(x, residual)
|
||||
|
||||
from vllm import _custom_ops as ops
|
||||
|
||||
if residual is not None:
|
||||
ops.fused_add_rms_norm(
|
||||
x,
|
||||
residual,
|
||||
self.weight.data,
|
||||
self.variance_epsilon,
|
||||
)
|
||||
return x, residual
|
||||
out = torch.empty_like(x)
|
||||
ops.rms_norm(
|
||||
out,
|
||||
x,
|
||||
self.weight.data,
|
||||
self.variance_epsilon,
|
||||
)
|
||||
return out
|
||||
|
||||
def forward_hpu(
|
||||
self,
|
||||
x: torch.Tensor,
|
||||
residual: Optional[torch.Tensor] = None,
|
||||
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
||||
from vllm_hpu_extension.ops import HPUFusedRMSNorm
|
||||
if HPUFusedRMSNorm is None:
|
||||
return self.forward_native(x, residual)
|
||||
if residual is not None:
|
||||
orig_shape = x.shape
|
||||
residual += x.view(residual.shape)
|
||||
# Note: HPUFusedRMSNorm requires 3D tensors as inputs
|
||||
x = HPUFusedRMSNorm.apply(residual, self.weight,
|
||||
self.variance_epsilon)
|
||||
return x.view(orig_shape), residual
|
||||
|
||||
x = HPUFusedRMSNorm.apply(x, self.weight, self.variance_epsilon)
|
||||
return x
|
||||
|
||||
def forward_xpu(
|
||||
self,
|
||||
x: torch.Tensor,
|
||||
residual: Optional[torch.Tensor] = None,
|
||||
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
||||
if self.variance_size_override is not None:
|
||||
return self.forward_native(x, residual)
|
||||
|
||||
from vllm._ipex_ops import ipex_ops as ops
|
||||
|
||||
if residual is not None:
|
||||
ops.fused_add_rms_norm(
|
||||
x,
|
||||
residual,
|
||||
self.weight.data,
|
||||
self.variance_epsilon,
|
||||
)
|
||||
return x, residual
|
||||
return ops.rms_norm(
|
||||
x,
|
||||
self.weight.data,
|
||||
self.variance_epsilon,
|
||||
)
|
||||
|
||||
def forward_mlu(
|
||||
self,
|
||||
x: torch.Tensor,
|
||||
residual: Optional[torch.Tensor] = None,
|
||||
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
||||
from vllm import _mlu_ops as mlu_ops
|
||||
|
||||
x = x.view(-1, self.weight.data.shape[0])
|
||||
if residual is not None:
|
||||
residual = residual.view(-1, self.weight.data.shape[0])
|
||||
return mlu_ops.fused_rms_norm(x, residual, self.weight.data, None, None, self.variance_epsilon, True)
|
||||
else:
|
||||
return mlu_ops.fused_rms_norm(x, residual, self.weight.data, None, None, self.variance_epsilon, False)
|
||||
|
||||
def extra_repr(self) -> str:
|
||||
s = f"hidden_size={self.weight.data.size(0)}"
|
||||
s += f", eps={self.variance_epsilon}"
|
||||
return s
|
||||
|
||||
|
||||
@CustomOp.register("gemma_rms_norm")
|
||||
class GemmaRMSNorm(CustomOp):
|
||||
"""RMS normalization for Gemma.
|
||||
|
||||
Two differences from the above RMSNorm:
|
||||
1. x * (1 + w) instead of x * w.
|
||||
2. (x * w).to(orig_dtype) instead of x.to(orig_dtype) * w.
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
hidden_size: int,
|
||||
eps: float = 1e-6,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.weight = nn.Parameter(torch.zeros(hidden_size))
|
||||
self.variance_epsilon = eps
|
||||
|
||||
@staticmethod
|
||||
def forward_static(
|
||||
weight: torch.Tensor,
|
||||
variance_epsilon: float,
|
||||
x: torch.Tensor,
|
||||
residual: Optional[torch.Tensor],
|
||||
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
||||
"""PyTorch-native implementation equivalent to forward()."""
|
||||
orig_dtype = x.dtype
|
||||
if residual is not None:
|
||||
x = x + residual
|
||||
residual = x
|
||||
|
||||
x = x.float()
|
||||
variance = x.pow(2).mean(dim=-1, keepdim=True)
|
||||
x = x * torch.rsqrt(variance + variance_epsilon)
|
||||
# Llama does x.to(float16) * w whilst Gemma is (x * w).to(float16)
|
||||
# See https://github.com/huggingface/transformers/pull/29402
|
||||
x = x * (1.0 + weight.float())
|
||||
x = x.to(orig_dtype)
|
||||
return x if residual is None else (x, residual)
|
||||
|
||||
def forward_native(
|
||||
self,
|
||||
x: torch.Tensor,
|
||||
residual: Optional[torch.Tensor] = None,
|
||||
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
||||
"""PyTorch-native implementation equivalent to forward()."""
|
||||
return self.forward_static(self.weight.data, self.variance_epsilon, x,
|
||||
residual)
|
||||
|
||||
def forward_cuda(
|
||||
self,
|
||||
x: torch.Tensor,
|
||||
residual: Optional[torch.Tensor] = None,
|
||||
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
||||
if torch.compiler.is_compiling():
|
||||
return self.forward_native(x, residual)
|
||||
|
||||
if not getattr(self, "_is_compiled", False):
|
||||
self.forward_static = torch.compile( # type: ignore
|
||||
self.forward_static)
|
||||
self._is_compiled = True
|
||||
return self.forward_native(x, residual)
|
||||
1099
vllm-v0.6.2/vllm/model_executor/layers/linear.py
Normal file
1099
vllm-v0.6.2/vllm/model_executor/layers/linear.py
Normal file
File diff suppressed because it is too large
Load Diff
161
vllm-v0.6.2/vllm/model_executor/layers/logits_processor.py
Normal file
161
vllm-v0.6.2/vllm/model_executor/layers/logits_processor.py
Normal file
@@ -0,0 +1,161 @@
|
||||
"""A layer that compute logits from hidden_stats."""
|
||||
import inspect
|
||||
from typing import Optional
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
|
||||
from vllm.distributed import (tensor_model_parallel_all_gather,
|
||||
tensor_model_parallel_gather)
|
||||
from vllm.model_executor.layers.vocab_parallel_embedding import (
|
||||
VocabParallelEmbedding)
|
||||
from vllm.model_executor.sampling_metadata import SamplingMetadata
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
|
||||
class LogitsProcessor(nn.Module):
|
||||
"""Process logits and apply logits processors from sampling metadata.
|
||||
|
||||
This layer does the following:
|
||||
1. Gather logits from model hidden_states.
|
||||
2. Scale logits if needed.
|
||||
3. Apply logits processors (if any).
|
||||
"""
|
||||
|
||||
def __init__(self,
|
||||
vocab_size: int,
|
||||
org_vocab_size: Optional[int] = None,
|
||||
scale: float = 1.0,
|
||||
logits_as_input: bool = False,
|
||||
soft_cap: Optional[float] = None) -> None:
|
||||
"""
|
||||
Args:
|
||||
scale: A scaling factor to apply to the logits.
|
||||
"""
|
||||
super().__init__()
|
||||
self.scale = scale
|
||||
self.vocab_size = vocab_size
|
||||
# Whether the input is logits (default is hidden states).
|
||||
self.logits_as_input = logits_as_input
|
||||
# original vocabulary size (without LoRA).
|
||||
self.org_vocab_size = org_vocab_size or vocab_size
|
||||
# Soft cap the logits. Used in Gemma 2.
|
||||
self.soft_cap = soft_cap
|
||||
# Whether to use gather or all-gather to gather the logits.
|
||||
self.use_gather = not current_platform.is_tpu()
|
||||
|
||||
def forward(
|
||||
self,
|
||||
lm_head: VocabParallelEmbedding,
|
||||
hidden_states: torch.Tensor,
|
||||
sampling_metadata: Optional[SamplingMetadata] = None,
|
||||
embedding_bias: Optional[torch.Tensor] = None,
|
||||
) -> Optional[torch.Tensor]:
|
||||
if self.logits_as_input:
|
||||
logits = hidden_states
|
||||
else:
|
||||
if sampling_metadata is not None:
|
||||
hidden_states = _prune_hidden_states(hidden_states,
|
||||
sampling_metadata)
|
||||
|
||||
# Get the logits for the next tokens.
|
||||
logits = self._get_logits(hidden_states, lm_head, embedding_bias)
|
||||
if logits is not None:
|
||||
if self.soft_cap is not None:
|
||||
logits = logits / self.soft_cap
|
||||
logits = torch.tanh(logits)
|
||||
logits = logits * self.soft_cap
|
||||
|
||||
if self.scale != 1.0:
|
||||
logits *= self.scale
|
||||
|
||||
# Apply logits processors (if any).
|
||||
if sampling_metadata is not None:
|
||||
logits = _apply_logits_processors(logits, sampling_metadata)
|
||||
|
||||
return logits
|
||||
|
||||
def _get_logits(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
lm_head: VocabParallelEmbedding,
|
||||
embedding_bias: Optional[torch.Tensor],
|
||||
) -> Optional[torch.Tensor]:
|
||||
# Get the logits for the next tokens.
|
||||
logits = lm_head.linear_method.apply(lm_head,
|
||||
hidden_states,
|
||||
bias=embedding_bias)
|
||||
if self.use_gather:
|
||||
# None may be returned for rank > 0
|
||||
logits = tensor_model_parallel_gather(logits)
|
||||
else:
|
||||
# Gather is not supported for some devices such as TPUs.
|
||||
# Use all-gather instead.
|
||||
# NOTE(woosuk): Here, the outputs of every device should not be None
|
||||
# because XLA requires strict SPMD among all devices. Every device
|
||||
# should execute the same operations after gathering the logits.
|
||||
logits = tensor_model_parallel_all_gather(logits)
|
||||
# Remove paddings in vocab (if any).
|
||||
if logits is not None:
|
||||
logits = logits[..., :self.org_vocab_size]
|
||||
return logits
|
||||
|
||||
def extra_repr(self) -> str:
|
||||
s = f"vocab_size={self.vocab_size}"
|
||||
s += f", forg_vocab_size={self.org_vocab_size}"
|
||||
s += f", scale={self.scale}, logits_as_input={self.logits_as_input}"
|
||||
return s
|
||||
|
||||
|
||||
def _prune_hidden_states(
|
||||
hidden_states: torch.Tensor,
|
||||
sampling_metadata: SamplingMetadata,
|
||||
) -> torch.Tensor:
|
||||
# NOTE(kzawora): The if guard is needed for Gaudi - in some scenarios
|
||||
# (warmup, profile_run) we might not have selected_token_indices,
|
||||
# so we skip pruning.
|
||||
if sampling_metadata.selected_token_indices is not None:
|
||||
return hidden_states.index_select(
|
||||
0, sampling_metadata.selected_token_indices)
|
||||
else:
|
||||
return hidden_states
|
||||
|
||||
|
||||
def _apply_logits_processors(
|
||||
logits: torch.Tensor,
|
||||
sampling_metadata: SamplingMetadata,
|
||||
) -> torch.Tensor:
|
||||
found_logits_processors = False
|
||||
logits_processed = 0
|
||||
for seq_group in sampling_metadata.seq_groups:
|
||||
seq_ids = seq_group.seq_ids
|
||||
sampling_params = seq_group.sampling_params
|
||||
logits_processors = sampling_params.logits_processors
|
||||
if logits_processors:
|
||||
found_logits_processors = True
|
||||
|
||||
for seq_id, logits_row_idx in zip(seq_ids,
|
||||
seq_group.sample_indices):
|
||||
logits_row = logits[logits_row_idx]
|
||||
past_tokens_ids = seq_group.seq_data[seq_id].output_token_ids
|
||||
prompt_tokens_ids = seq_group.seq_data[seq_id].prompt_token_ids
|
||||
|
||||
for logits_processor in logits_processors:
|
||||
parameters = inspect.signature(logits_processor).parameters
|
||||
if len(parameters) == 3:
|
||||
logits_row = logits_processor(prompt_tokens_ids,
|
||||
past_tokens_ids,
|
||||
logits_row)
|
||||
else:
|
||||
logits_row = logits_processor(past_tokens_ids,
|
||||
logits_row)
|
||||
|
||||
logits[logits_row_idx] = logits_row
|
||||
|
||||
logits_processed += len(seq_group.sample_indices) + len(
|
||||
seq_group.prompt_logprob_indices)
|
||||
|
||||
if found_logits_processors:
|
||||
# verifies that no rows in logits were missed unexpectedly
|
||||
assert logits_processed == logits.shape[0]
|
||||
return logits
|
||||
217
vllm-v0.6.2/vllm/model_executor/layers/mamba/mamba_mixer.py
Normal file
217
vllm-v0.6.2/vllm/model_executor/layers/mamba/mamba_mixer.py
Normal file
@@ -0,0 +1,217 @@
|
||||
import torch
|
||||
from torch import nn
|
||||
from torch.nn.parameter import Parameter
|
||||
|
||||
from vllm.attention.backends.abstract import AttentionMetadata
|
||||
from vllm.distributed.parallel_state import (
|
||||
get_tensor_model_parallel_rank, get_tensor_model_parallel_world_size)
|
||||
from vllm.model_executor.custom_op import CustomOp
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
|
||||
MergedColumnParallelLinear,
|
||||
RowParallelLinear)
|
||||
from vllm.model_executor.layers.mamba.ops.causal_conv1d import (
|
||||
causal_conv1d_fn, causal_conv1d_update)
|
||||
from vllm.model_executor.layers.mamba.ops.mamba_ssm import (
|
||||
selective_scan_fn, selective_state_update)
|
||||
from vllm.model_executor.models.mamba_cache import MambaCacheParams
|
||||
from vllm.model_executor.utils import set_weight_attrs
|
||||
|
||||
|
||||
# Adapted from transformers.models.mamba.modeling_mamba.MambaMixer
|
||||
@CustomOp.register("mamba_mixer")
|
||||
class MambaMixer(CustomOp):
|
||||
"""
|
||||
Compute ∆, A, B, C, and D the state space parameters and compute
|
||||
the `contextualized_states`. A, D are input independent
|
||||
(see Mamba paper [1] Section 3.5.2 "Interpretation of A"
|
||||
for why A isn't selective) ∆, B, C are input-dependent
|
||||
(this is a key difference between Mamba and the linear time
|
||||
invariant S4, and is why Mamba is called
|
||||
**selective** state spaces)
|
||||
"""
|
||||
|
||||
def __init__(self,
|
||||
hidden_size: int,
|
||||
ssm_state_size: int,
|
||||
conv_kernel_size: int,
|
||||
intermediate_size: int,
|
||||
time_step_rank: int,
|
||||
use_conv_bias: bool,
|
||||
use_bias: bool,
|
||||
use_rms_norm: bool,
|
||||
rms_norm_eps: float = 1e-5,
|
||||
activation="silu"):
|
||||
super().__init__()
|
||||
self.time_step_rank = time_step_rank
|
||||
self.ssm_state_size = ssm_state_size
|
||||
self.use_rms_norm = use_rms_norm
|
||||
self.activation = activation
|
||||
|
||||
self.conv1d = ColumnParallelLinear(
|
||||
input_size=conv_kernel_size,
|
||||
output_size=intermediate_size,
|
||||
bias=use_conv_bias,
|
||||
)
|
||||
# unsqueeze to fit conv1d weights shape into the linear weights shape.
|
||||
# Can't do this in `weight_loader` since it already exists in
|
||||
# `ColumnParallelLinear` and `set_weight_attrs`
|
||||
# doesn't allow to override it
|
||||
self.conv1d.weight.data = self.conv1d.weight.data.unsqueeze(1)
|
||||
|
||||
self.in_proj = MergedColumnParallelLinear(hidden_size,
|
||||
[intermediate_size] * 2,
|
||||
bias=use_bias)
|
||||
# selective projection used to make dt, B and C input dependent
|
||||
self.x_proj = RowParallelLinear(
|
||||
intermediate_size,
|
||||
time_step_rank + ssm_state_size * 2,
|
||||
bias=False,
|
||||
)
|
||||
# time step projection (discretization) -
|
||||
# In the forward we need to apply dt_proj without the bias,
|
||||
# as the bias is added in the selective scan kernel.
|
||||
self.dt_proj = ColumnParallelLinear(time_step_rank,
|
||||
intermediate_size,
|
||||
bias=True,
|
||||
skip_bias_add=True)
|
||||
|
||||
def weight_loader(param: Parameter, loaded_weight: torch.Tensor):
|
||||
tp_rank = get_tensor_model_parallel_rank()
|
||||
tp_size = get_tensor_model_parallel_world_size()
|
||||
param.data.copy_(
|
||||
loaded_weight.data.split(loaded_weight.shape[0] // tp_size,
|
||||
dim=0)[tp_rank])
|
||||
|
||||
def A_weight_loader(param: Parameter, loaded_weight: torch.Tensor):
|
||||
weight_loader(param, -torch.exp(loaded_weight.float()))
|
||||
|
||||
tp_size = get_tensor_model_parallel_world_size()
|
||||
self.A = nn.Parameter(
|
||||
torch.empty(
|
||||
intermediate_size // tp_size,
|
||||
ssm_state_size,
|
||||
dtype=torch.float32,
|
||||
))
|
||||
self.D = nn.Parameter(torch.ones(intermediate_size // tp_size))
|
||||
|
||||
set_weight_attrs(self.D, {"weight_loader": weight_loader})
|
||||
set_weight_attrs(self.A, {"weight_loader": A_weight_loader})
|
||||
|
||||
self.out_proj = RowParallelLinear(
|
||||
intermediate_size,
|
||||
hidden_size,
|
||||
bias=use_bias,
|
||||
input_is_parallel=True,
|
||||
)
|
||||
|
||||
self.dt_layernorm = RMSNorm(time_step_rank,
|
||||
eps=rms_norm_eps) if use_rms_norm else None
|
||||
|
||||
self.b_layernorm = RMSNorm(ssm_state_size,
|
||||
eps=rms_norm_eps) if use_rms_norm else None
|
||||
|
||||
self.c_layernorm = RMSNorm(ssm_state_size,
|
||||
eps=rms_norm_eps) if use_rms_norm else None
|
||||
|
||||
def forward_native(self, hidden_states: torch.Tensor,
|
||||
attn_metadata: AttentionMetadata,
|
||||
conv_state: torch.Tensor, ssm_state: torch.Tensor):
|
||||
pass
|
||||
|
||||
def forward_cuda(self, hidden_states: torch.Tensor,
|
||||
attn_metadata: AttentionMetadata,
|
||||
mamba_cache_params: MambaCacheParams):
|
||||
|
||||
# 1. Gated MLP's linear projection
|
||||
projected_states = self.in_proj(hidden_states)[0].transpose(-2, -1)
|
||||
hidden_states, gate = projected_states.chunk(2, dim=-2)
|
||||
|
||||
# 2. Convolution sequence transformation
|
||||
conv_weights = self.conv1d.weight.view(self.conv1d.weight.size(0),
|
||||
self.conv1d.weight.size(2))
|
||||
|
||||
if attn_metadata.query_start_loc is not None \
|
||||
and attn_metadata.context_lens_tensor is not None:
|
||||
# |---------- N-1 iteration --------|
|
||||
# |---------------- N iteration ---------------------|
|
||||
# |- tokenA -|......................|-- newTokens ---|
|
||||
# |---------- context_len ----------|
|
||||
# |-------------------- seq_len ---------------------|
|
||||
# |-- query_len ---|
|
||||
hidden_states = causal_conv1d_fn(
|
||||
hidden_states,
|
||||
conv_weights,
|
||||
self.conv1d.bias,
|
||||
activation=self.activation,
|
||||
conv_states=mamba_cache_params.conv_state,
|
||||
has_initial_state=attn_metadata.context_lens_tensor > 0,
|
||||
cache_indices=mamba_cache_params.state_indices_tensor,
|
||||
query_start_loc=attn_metadata.query_start_loc)
|
||||
else:
|
||||
hidden_states = causal_conv1d_update(
|
||||
hidden_states.transpose(0, 1),
|
||||
mamba_cache_params.conv_state,
|
||||
conv_weights,
|
||||
self.conv1d.bias,
|
||||
self.activation,
|
||||
conv_state_indices=mamba_cache_params.state_indices_tensor)
|
||||
hidden_states = hidden_states.transpose(0, 1)
|
||||
|
||||
# 3. State Space Model sequence transformation
|
||||
# 3.a. input varying initialization of time_step, B and C
|
||||
ssm_parameters = self.x_proj(hidden_states.transpose(-2, -1))[0]
|
||||
|
||||
time_step, B, C = torch.split(
|
||||
ssm_parameters,
|
||||
[self.time_step_rank, self.ssm_state_size, self.ssm_state_size],
|
||||
dim=-1,
|
||||
)
|
||||
if self.use_rms_norm:
|
||||
assert self.dt_layernorm is not None
|
||||
assert self.b_layernorm is not None
|
||||
assert self.c_layernorm is not None
|
||||
time_step = self.dt_layernorm(time_step.contiguous())
|
||||
B = self.b_layernorm(B.contiguous())
|
||||
C = self.c_layernorm(C.contiguous())
|
||||
|
||||
discrete_time_step = self.dt_proj(time_step)[0].transpose(-2, -1)
|
||||
# 3.c perform the recurrence y ← SSM(A, B, C)(x)
|
||||
time_proj_bias = (self.dt_proj.bias.float() if hasattr(
|
||||
self.dt_proj, "bias") else None)
|
||||
|
||||
if attn_metadata.query_start_loc is not None \
|
||||
and attn_metadata.context_lens_tensor is not None:
|
||||
scan_outputs = selective_scan_fn(
|
||||
hidden_states,
|
||||
mamba_cache_params.ssm_state,
|
||||
discrete_time_step,
|
||||
self.A,
|
||||
B.transpose(-2, -1),
|
||||
C.transpose(-2, -1),
|
||||
self.D.float(),
|
||||
gate,
|
||||
time_proj_bias,
|
||||
delta_softplus=True,
|
||||
cache_indices=mamba_cache_params.state_indices_tensor,
|
||||
has_initial_state=attn_metadata.context_lens_tensor > 0,
|
||||
query_start_loc=attn_metadata.query_start_loc)
|
||||
else:
|
||||
scan_outputs = selective_state_update(
|
||||
mamba_cache_params.ssm_state,
|
||||
hidden_states.transpose(0, 1),
|
||||
discrete_time_step.transpose(0, 1),
|
||||
self.A,
|
||||
B,
|
||||
C,
|
||||
self.D,
|
||||
gate.transpose(0, 1),
|
||||
time_proj_bias,
|
||||
dt_softplus=True,
|
||||
state_batch_indices=mamba_cache_params.state_indices_tensor)
|
||||
scan_outputs = scan_outputs.transpose(0, 1)
|
||||
|
||||
# 4. Final linear projection
|
||||
contextualized_states = self.out_proj(scan_outputs.transpose(-2,
|
||||
-1))[0]
|
||||
return contextualized_states
|
||||
@@ -0,0 +1,102 @@
|
||||
# Copyright (c) 2024, Tri Dao.
|
||||
# Adapted from https://github.com/Dao-AILab/causal-conv1d/blob/main/causal_conv1d/causal_conv1d_interface.py
|
||||
|
||||
from typing import Optional
|
||||
|
||||
import torch
|
||||
|
||||
from vllm import _custom_ops as ops
|
||||
from vllm.attention.backends.utils import PAD_SLOT_ID
|
||||
|
||||
|
||||
def causal_conv1d_fn(x: torch.Tensor,
|
||||
weight: torch.Tensor,
|
||||
bias: Optional[torch.Tensor] = None,
|
||||
query_start_loc: Optional[torch.Tensor] = None,
|
||||
cache_indices: Optional[torch.Tensor] = None,
|
||||
has_initial_state: Optional[torch.Tensor] = None,
|
||||
conv_states: Optional[torch.Tensor] = None,
|
||||
activation: Optional[str] = "silu",
|
||||
pad_slot_id: int = PAD_SLOT_ID):
|
||||
"""
|
||||
x: (batch, dim, seqlen) or (dim,cu_seq_len) for varlen
|
||||
sequences are concatenated from left to right for varlen
|
||||
weight: (dim, width)
|
||||
bias: (dim,)
|
||||
query_start_loc: (batch + 1) int32
|
||||
The cumulative sequence lengths of the sequences in
|
||||
the batch, used to index into sequence. prepended by 0.
|
||||
for example: query_start_loc = torch.Tensor([0,10,16,17]),
|
||||
x.shape=(dim,17)
|
||||
cache_indices: (batch) int32
|
||||
indicates the corresponding state index,
|
||||
like so: conv_state = conv_states[cache_indices[batch_id]]
|
||||
has_initial_state: (batch) bool
|
||||
indicates whether should the kernel take the current state as initial
|
||||
state for the calculations
|
||||
conv_states: (...,dim,width - 1) itype
|
||||
updated inplace if provided
|
||||
activation: either None or "silu" or "swish"
|
||||
pad_slot_id: int
|
||||
if cache_indices is passed, lets the kernel identify padded
|
||||
entries that will not be processed,
|
||||
for example: cache_indices = [pad_slot_id, 1, 20, pad_slot_id]
|
||||
in this case, the kernel will not process entries at
|
||||
indices 0 and 3
|
||||
|
||||
|
||||
out: (batch, dim, seqlen)
|
||||
"""
|
||||
if activation not in [None, "silu", "swish"]:
|
||||
raise NotImplementedError("activation must be None, silu, or swish")
|
||||
if x.stride(-1) != 1:
|
||||
x = x.contiguous()
|
||||
bias = bias.contiguous() if bias is not None else None
|
||||
|
||||
ops.causal_conv1d_fwd(x, weight, bias, conv_states, query_start_loc,
|
||||
cache_indices, has_initial_state, activation
|
||||
in ["silu", "swish"], pad_slot_id)
|
||||
return x
|
||||
|
||||
|
||||
def causal_conv1d_update(x: torch.Tensor,
|
||||
conv_state: torch.Tensor,
|
||||
weight: torch.Tensor,
|
||||
bias: Optional[torch.Tensor] = None,
|
||||
activation: Optional[str] = None,
|
||||
cache_seqlens: Optional[torch.Tensor] = None,
|
||||
conv_state_indices: Optional[torch.Tensor] = None,
|
||||
pad_slot_id: int = PAD_SLOT_ID):
|
||||
"""
|
||||
x: (batch, dim) or (batch, dim, seqlen)
|
||||
conv_state: (batch, dim, state_len), where state_len >= width - 1
|
||||
weight: (dim, width)
|
||||
bias: (dim,)
|
||||
cache_seqlens: (batch,), dtype int32.
|
||||
If not None, the conv_state is treated as a circular buffer.
|
||||
The conv_state will be updated by copying x to the conv_state
|
||||
starting at the index
|
||||
@cache_seqlens % state_len.
|
||||
conv_state_indices: (batch,), dtype int32
|
||||
If not None, the conv_state is a larger tensor along the batch dim,
|
||||
and we are selecting the batch coords specified by conv_state_indices.
|
||||
Useful for a continuous batching scenario.
|
||||
pad_slot_id: int
|
||||
if cache_indices is passed, lets the kernel identify padded
|
||||
entries that will not be processed,
|
||||
for example: cache_indices = [pad_slot_id, 1 ,20 ,pad_slot_id]
|
||||
in this case, the kernel will not process entries at
|
||||
indices 0 and 3
|
||||
out: (batch, dim) or (batch, dim, seqlen)
|
||||
"""
|
||||
if activation not in [None, "silu", "swish"]:
|
||||
raise NotImplementedError("activation must be None, silu, or swish")
|
||||
activation_val = activation in ["silu", "swish"]
|
||||
unsqueeze = x.dim() == 2
|
||||
if unsqueeze:
|
||||
x = x.unsqueeze(-1)
|
||||
ops.causal_conv1d_update(x, conv_state, weight, bias, activation_val,
|
||||
cache_seqlens, conv_state_indices, pad_slot_id)
|
||||
if unsqueeze:
|
||||
x = x.squeeze(-1)
|
||||
return x
|
||||
411
vllm-v0.6.2/vllm/model_executor/layers/mamba/ops/mamba_ssm.py
Normal file
411
vllm-v0.6.2/vllm/model_executor/layers/mamba/ops/mamba_ssm.py
Normal file
@@ -0,0 +1,411 @@
|
||||
# Copyright (c) 2024, Tri Dao, Albert Gu.
|
||||
# Adapted from https://github.com/state-spaces/mamba/blob/main/mamba_ssm/ops/triton/selective_state_update.py
|
||||
|
||||
import torch
|
||||
import triton
|
||||
import triton.language as tl
|
||||
from packaging import version
|
||||
|
||||
from vllm import _custom_ops as ops
|
||||
from vllm.attention.backends.utils import PAD_SLOT_ID
|
||||
|
||||
TRITON3 = version.parse(triton.__version__) >= version.parse("3.0.0")
|
||||
|
||||
if TRITON3:
|
||||
|
||||
@triton.jit
|
||||
def softplus(dt):
|
||||
dt = tl.where(dt <= 20.0, tl.math.log(tl.math.exp(dt) + 1), dt)
|
||||
return dt
|
||||
else:
|
||||
|
||||
@triton.jit
|
||||
def softplus(dt):
|
||||
dt = tl.where(dt <= 20.0, tl.math.log1p(tl.exp(dt)), dt)
|
||||
return dt
|
||||
|
||||
|
||||
@triton.heuristics(
|
||||
{"HAS_DT_BIAS": lambda args: args["dt_bias_ptr"] is not None})
|
||||
@triton.heuristics({"HAS_D": lambda args: args["D_ptr"] is not None})
|
||||
@triton.heuristics({"HAS_Z": lambda args: args["z_ptr"] is not None})
|
||||
@triton.heuristics({
|
||||
"HAS_STATE_BATCH_INDICES":
|
||||
lambda args: args["state_batch_indices_ptr"] is not None
|
||||
})
|
||||
@triton.heuristics(
|
||||
{"BLOCK_SIZE_DSTATE": lambda args: triton.next_power_of_2(args["dstate"])})
|
||||
@triton.jit
|
||||
def _selective_scan_update_kernel(
|
||||
# Pointers to matrices
|
||||
state_ptr,
|
||||
x_ptr,
|
||||
dt_ptr,
|
||||
dt_bias_ptr,
|
||||
A_ptr,
|
||||
B_ptr,
|
||||
C_ptr,
|
||||
D_ptr,
|
||||
z_ptr,
|
||||
out_ptr,
|
||||
state_batch_indices_ptr,
|
||||
pad_slot_id,
|
||||
# Matrix dimensions
|
||||
batch,
|
||||
nheads,
|
||||
dim,
|
||||
dstate,
|
||||
nheads_ngroups_ratio,
|
||||
# Strides
|
||||
stride_state_batch,
|
||||
stride_state_head,
|
||||
stride_state_dim,
|
||||
stride_state_dstate,
|
||||
stride_x_batch,
|
||||
stride_x_head,
|
||||
stride_x_dim,
|
||||
stride_dt_batch,
|
||||
stride_dt_head,
|
||||
stride_dt_dim,
|
||||
stride_dt_bias_head,
|
||||
stride_dt_bias_dim,
|
||||
stride_A_head,
|
||||
stride_A_dim,
|
||||
stride_A_dstate,
|
||||
stride_B_batch,
|
||||
stride_B_group,
|
||||
stride_B_dstate,
|
||||
stride_C_batch,
|
||||
stride_C_group,
|
||||
stride_C_dstate,
|
||||
stride_D_head,
|
||||
stride_D_dim,
|
||||
stride_z_batch,
|
||||
stride_z_head,
|
||||
stride_z_dim,
|
||||
stride_out_batch,
|
||||
stride_out_head,
|
||||
stride_out_dim,
|
||||
# Meta-parameters
|
||||
DT_SOFTPLUS: tl.constexpr,
|
||||
TIE_HDIM: tl.constexpr,
|
||||
BLOCK_SIZE_M: tl.constexpr,
|
||||
HAS_DT_BIAS: tl.constexpr,
|
||||
HAS_D: tl.constexpr,
|
||||
HAS_Z: tl.constexpr,
|
||||
HAS_STATE_BATCH_INDICES: tl.constexpr,
|
||||
BLOCK_SIZE_DSTATE: tl.constexpr,
|
||||
):
|
||||
pid_m = tl.program_id(axis=0)
|
||||
pid_b = tl.program_id(axis=1)
|
||||
pid_h = tl.program_id(axis=2)
|
||||
|
||||
# If HAS_STATE_BATCH_INDICES is true, then the ssm state's batch coordinate
|
||||
# is taken from the state_batch_indices_ptr Otherwise, the state coordinate
|
||||
# is the same as the batch id.
|
||||
if HAS_STATE_BATCH_INDICES:
|
||||
state_batch_indices_ptr += pid_b
|
||||
state_batch_idx = tl.load(state_batch_indices_ptr)
|
||||
state_ptr += (state_batch_idx * stride_state_batch +
|
||||
pid_h * stride_state_head)
|
||||
else:
|
||||
state_ptr += pid_b * stride_state_batch + pid_h * stride_state_head
|
||||
|
||||
x_ptr += pid_b * stride_x_batch + pid_h * stride_x_head
|
||||
dt_ptr += pid_b * stride_dt_batch + pid_h * stride_dt_head
|
||||
if HAS_DT_BIAS:
|
||||
dt_bias_ptr += pid_h * stride_dt_bias_head
|
||||
A_ptr += pid_h * stride_A_head
|
||||
B_ptr += pid_b * stride_B_batch + (pid_h //
|
||||
nheads_ngroups_ratio) * stride_B_group
|
||||
C_ptr += pid_b * stride_C_batch + (pid_h //
|
||||
nheads_ngroups_ratio) * stride_C_group
|
||||
if HAS_Z:
|
||||
z_ptr += pid_b * stride_z_batch + pid_h * stride_z_head
|
||||
out_ptr += pid_b * stride_out_batch + pid_h * stride_out_head
|
||||
|
||||
offs_m = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
|
||||
offs_n = tl.arange(0, BLOCK_SIZE_DSTATE)
|
||||
state_ptrs = state_ptr + (offs_m[:, None] * stride_state_dim +
|
||||
offs_n[None, :] * stride_state_dstate)
|
||||
x_ptrs = x_ptr + offs_m * stride_x_dim
|
||||
dt_ptrs = dt_ptr + offs_m * stride_dt_dim
|
||||
if HAS_DT_BIAS:
|
||||
dt_bias_ptrs = dt_bias_ptr + offs_m * stride_dt_bias_dim
|
||||
if HAS_D:
|
||||
D_ptr += pid_h * stride_D_head
|
||||
A_ptrs = A_ptr + (offs_m[:, None] * stride_A_dim +
|
||||
offs_n[None, :] * stride_A_dstate)
|
||||
B_ptrs = B_ptr + offs_n * stride_B_dstate
|
||||
C_ptrs = C_ptr + offs_n * stride_C_dstate
|
||||
if HAS_D:
|
||||
D_ptrs = D_ptr + offs_m * stride_D_dim
|
||||
if HAS_Z:
|
||||
z_ptrs = z_ptr + offs_m * stride_z_dim
|
||||
out_ptrs = out_ptr + offs_m * stride_out_dim
|
||||
mask = (offs_m[:, None] < dim) & (offs_n[None, :] < dstate)
|
||||
if HAS_STATE_BATCH_INDICES:
|
||||
mask &= (state_batch_idx != pad_slot_id)
|
||||
state = tl.load(state_ptrs, mask=mask, other=0.0)
|
||||
|
||||
x = tl.load(x_ptrs, mask=offs_m < dim, other=0.0).to(tl.float32)
|
||||
if not TIE_HDIM:
|
||||
dt = tl.load(dt_ptrs, mask=offs_m < dim, other=0.0).to(tl.float32)
|
||||
if HAS_DT_BIAS:
|
||||
dt += tl.load(dt_bias_ptrs, mask=offs_m < dim,
|
||||
other=0.0).to(tl.float32)
|
||||
if DT_SOFTPLUS:
|
||||
dt = softplus(dt)
|
||||
A = tl.load(A_ptrs,
|
||||
mask=(offs_m[:, None] < dim) & (offs_n[None, :] < dstate),
|
||||
other=0.0).to(tl.float32)
|
||||
dA = tl.exp(A * dt[:, None])
|
||||
else:
|
||||
dt = tl.load(dt_ptr).to(tl.float32)
|
||||
if HAS_DT_BIAS:
|
||||
dt += tl.load(dt_bias_ptr).to(tl.float32)
|
||||
if DT_SOFTPLUS:
|
||||
dt = softplus(dt)
|
||||
A = tl.load(A_ptr).to(tl.float32)
|
||||
dA = tl.exp(A * dt) # scalar, not a matrix
|
||||
|
||||
B = tl.load(B_ptrs, mask=offs_n < dstate, other=0.0).to(tl.float32)
|
||||
C = tl.load(C_ptrs, mask=offs_n < dstate, other=0.0).to(tl.float32)
|
||||
if HAS_D:
|
||||
D = tl.load(D_ptrs, mask=offs_m < dim, other=0.0).to(tl.float32)
|
||||
if HAS_Z:
|
||||
z = tl.load(z_ptrs, mask=offs_m < dim, other=0.0).to(tl.float32)
|
||||
|
||||
dB = B[None, :] * dt[:, None] if not TIE_HDIM else B * dt
|
||||
state = state * dA + dB * x[:, None]
|
||||
|
||||
mask = (offs_m[:, None] < dim) & (offs_n[None, :] < dstate)
|
||||
if HAS_STATE_BATCH_INDICES:
|
||||
mask &= (state_batch_idx != pad_slot_id)
|
||||
tl.store(state_ptrs, state, mask=mask)
|
||||
out = tl.sum(state * C[None, :], axis=1)
|
||||
if HAS_D:
|
||||
out += x * D
|
||||
if HAS_Z:
|
||||
out *= z * tl.sigmoid(z)
|
||||
tl.store(out_ptrs, out, mask=offs_m < dim)
|
||||
|
||||
|
||||
def selective_state_update(state,
|
||||
x,
|
||||
dt,
|
||||
A,
|
||||
B,
|
||||
C,
|
||||
D=None,
|
||||
z=None,
|
||||
dt_bias=None,
|
||||
dt_softplus=False,
|
||||
state_batch_indices=None,
|
||||
pad_slot_id=PAD_SLOT_ID):
|
||||
"""
|
||||
Argument:
|
||||
state: (batch, dim, dstate) or (batch, nheads, dim, dstate)
|
||||
x: (batch, dim) or (batch, nheads, dim)
|
||||
dt: (batch, dim) or (batch, nheads, dim)
|
||||
A: (dim, dstate) or (nheads, dim, dstate)
|
||||
B: (batch, dstate) or (batch, ngroups, dstate)
|
||||
C: (batch, dstate) or (batch, ngroups, dstate)
|
||||
D: (dim,) or (nheads, dim)
|
||||
z: (batch, dim) or (batch, nheads, dim)
|
||||
dt_bias: (dim,) or (nheads, dim)
|
||||
pad_slot_id: int
|
||||
if cache_indices is passed, lets the kernel identify padded
|
||||
entries that will not be processed,
|
||||
for example: cache_indices = [pad_slot_id, 1, 20, pad_slot_id]
|
||||
in this case, the kernel will not process entries at
|
||||
indices 0 and 3
|
||||
Return:
|
||||
out: (batch, dim) or (batch, nheads, dim)
|
||||
"""
|
||||
has_heads = state.dim() > 3
|
||||
if state.dim() == 3:
|
||||
state = state.unsqueeze(1)
|
||||
if x.dim() == 2:
|
||||
x = x.unsqueeze(1)
|
||||
if dt.dim() == 2:
|
||||
dt = dt.unsqueeze(1)
|
||||
if A.dim() == 2:
|
||||
A = A.unsqueeze(0)
|
||||
if B.dim() == 2:
|
||||
B = B.unsqueeze(1)
|
||||
if C.dim() == 2:
|
||||
C = C.unsqueeze(1)
|
||||
if D is not None and D.dim() == 1:
|
||||
D = D.unsqueeze(0)
|
||||
if z is not None and z.dim() == 2:
|
||||
z = z.unsqueeze(1)
|
||||
if dt_bias is not None and dt_bias.dim() == 1:
|
||||
dt_bias = dt_bias.unsqueeze(0)
|
||||
|
||||
_, nheads, dim, dstate = state.shape
|
||||
batch = x.shape[0]
|
||||
|
||||
assert x.shape == (batch, nheads, dim)
|
||||
assert dt.shape == x.shape
|
||||
assert A.shape == (nheads, dim, dstate)
|
||||
ngroups = B.shape[1]
|
||||
assert nheads % ngroups == 0, "nheads must be divisible by ngroups"
|
||||
assert B.shape == (batch, ngroups, dstate)
|
||||
assert C.shape == B.shape
|
||||
if D is not None:
|
||||
assert D.shape == (nheads, dim)
|
||||
if z is not None:
|
||||
assert z.shape == x.shape
|
||||
if dt_bias is not None:
|
||||
assert dt_bias.shape == (nheads, dim)
|
||||
if state_batch_indices is not None:
|
||||
assert state_batch_indices.shape == (batch, )
|
||||
out = torch.empty_like(x)
|
||||
grid = lambda META: (triton.cdiv(dim, META['BLOCK_SIZE_M']), batch, nheads)
|
||||
z_strides = ((z.stride(0), z.stride(1), z.stride(2)) if z is not None else
|
||||
(0, 0, 0))
|
||||
# We don't want autotune since it will overwrite the state
|
||||
# We instead tune by hand.
|
||||
BLOCK_SIZE_M, num_warps = ((32, 4) if dstate <= 16 else
|
||||
((16, 4) if dstate <= 32 else
|
||||
((8, 4) if dstate <= 64 else
|
||||
((4, 4) if dstate <= 128 else ((4, 8))))))
|
||||
tie_hdim = A.stride(-1) == 0 and A.stride(-2) == 0 and dt.stride(
|
||||
-1) == 0 and dt_bias.stride(-1) == 0
|
||||
with torch.cuda.device(x.device.index):
|
||||
_selective_scan_update_kernel[grid](
|
||||
state,
|
||||
x,
|
||||
dt,
|
||||
dt_bias,
|
||||
A,
|
||||
B,
|
||||
C,
|
||||
D,
|
||||
z,
|
||||
out,
|
||||
state_batch_indices,
|
||||
pad_slot_id,
|
||||
batch,
|
||||
nheads,
|
||||
dim,
|
||||
dstate,
|
||||
nheads // ngroups,
|
||||
state.stride(0),
|
||||
state.stride(1),
|
||||
state.stride(2),
|
||||
state.stride(3),
|
||||
x.stride(0),
|
||||
x.stride(1),
|
||||
x.stride(2),
|
||||
dt.stride(0),
|
||||
dt.stride(1),
|
||||
dt.stride(2),
|
||||
*(dt_bias.stride(0),
|
||||
dt_bias.stride(1)) if dt_bias is not None else 0,
|
||||
A.stride(0),
|
||||
A.stride(1),
|
||||
A.stride(2),
|
||||
B.stride(0),
|
||||
B.stride(1),
|
||||
B.stride(2),
|
||||
C.stride(0),
|
||||
C.stride(1),
|
||||
C.stride(2),
|
||||
*(D.stride(0), D.stride(1)) if D is not None else 0,
|
||||
z_strides[0],
|
||||
z_strides[1],
|
||||
z_strides[2],
|
||||
out.stride(0),
|
||||
out.stride(1),
|
||||
out.stride(2),
|
||||
dt_softplus,
|
||||
tie_hdim,
|
||||
BLOCK_SIZE_M,
|
||||
num_warps=num_warps,
|
||||
)
|
||||
if not has_heads:
|
||||
out = out.squeeze(1)
|
||||
return out
|
||||
|
||||
|
||||
def selective_scan_fn(u,
|
||||
ssm_states,
|
||||
delta,
|
||||
A,
|
||||
B,
|
||||
C,
|
||||
D=None,
|
||||
z=None,
|
||||
delta_bias=None,
|
||||
delta_softplus=False,
|
||||
query_start_loc=None,
|
||||
cache_indices=None,
|
||||
has_initial_state=None,
|
||||
pad_slot_id=PAD_SLOT_ID) -> torch.Tensor:
|
||||
"""
|
||||
u: (dim, total_length) for varlen or (batch, dim, seqlen)
|
||||
applies changes in place.
|
||||
ssm_states: (batch, dim, dstate) or (batch, nheads, dim, dstate)
|
||||
applies changes in place.
|
||||
delta: (dim, total_length) for varlen or (batch, dim, seqlen)
|
||||
A: (dim, dstate)
|
||||
B: (ngroups, dstate, total_length) for varlen or
|
||||
(batch,ngroups,dstate,seqlen)
|
||||
C: (ngroups, dstate, total_length) for varlen or
|
||||
(batch,ngroups,dstate,seqlen)
|
||||
D: (dim,)
|
||||
z: (dim, total_length) for varlen or (batch, dim, seqlen)
|
||||
dt_bias: (dim,) or (dim)
|
||||
query_start_loc: (batch + 1) int32
|
||||
The cumulative sequence lengths of the sequences in
|
||||
the batch, used to index into sequence. prepended with 0.
|
||||
for example: query_start_loc = torch.Tensor([0,10,16,17]),
|
||||
x.shape=(dim,17)
|
||||
cache_indices: (batch) int32
|
||||
A tensor with each cell is a correspondent
|
||||
input and output ssm_state index
|
||||
has_initial_state: (batch) bool
|
||||
A tensor populated with ones and zeros,
|
||||
indicate if the ssm_state at the corresponding index should be
|
||||
used as initial state. Not providing argument assumes
|
||||
there's no initial state
|
||||
pad_slot_id: int
|
||||
if cache_indices is passed, lets the kernel identify padding entries
|
||||
that will not be processed,
|
||||
for example: cache_indices = [pad_slot_id, 1 ,20 ,pad_slot_id]
|
||||
in this case, the kernel will not process entries at indices 0 and 3
|
||||
returns
|
||||
output: (dim, total_length) for varlen or (batch, dim, seqlen)
|
||||
supports inplace replacement
|
||||
"""
|
||||
if u.stride(-1) != 1:
|
||||
u = u.contiguous()
|
||||
if delta.stride(-1) != 1:
|
||||
delta = delta.contiguous()
|
||||
if D is not None:
|
||||
D = D.contiguous()
|
||||
if B.stride(-1) != 1:
|
||||
B = B.contiguous()
|
||||
if C.stride(-1) != 1:
|
||||
C = C.contiguous()
|
||||
if z is not None and z.stride(-1) != 1:
|
||||
z = z.contiguous()
|
||||
if B.dim() == 3 and query_start_loc is None:
|
||||
B = B.unsqueeze(1)
|
||||
if B.dim() == 2 and query_start_loc is not None:
|
||||
B = B.unsqueeze(0)
|
||||
if C.dim() == 3 and query_start_loc is None:
|
||||
C = C.unsqueeze(1)
|
||||
if C.dim() == 2 and query_start_loc is not None:
|
||||
C = C.unsqueeze(0)
|
||||
|
||||
ops.selective_scan_fwd(u, delta, A, B, C, D, z, delta_bias, delta_softplus,
|
||||
query_start_loc, cache_indices, has_initial_state,
|
||||
ssm_states, pad_slot_id)
|
||||
|
||||
if z is None:
|
||||
return delta # output written inplace to delta
|
||||
else:
|
||||
return z # output written inplace to z
|
||||
150
vllm-v0.6.2/vllm/model_executor/layers/pooler.py
Normal file
150
vllm-v0.6.2/vllm/model_executor/layers/pooler.py
Normal file
@@ -0,0 +1,150 @@
|
||||
from enum import IntEnum
|
||||
from typing import List, Optional
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
|
||||
from vllm.config import PoolerConfig
|
||||
from vllm.model_executor.pooling_metadata import (PoolingMetadata,
|
||||
PoolingTensors)
|
||||
from vllm.sequence import EmbeddingSequenceGroupOutput, PoolerOutput
|
||||
|
||||
|
||||
class PoolingType(IntEnum):
|
||||
"""Enumeration for different types of pooling methods."""
|
||||
LAST = 0
|
||||
ALL = 1
|
||||
CLS = 2
|
||||
STEP = 3
|
||||
MEAN = 4
|
||||
|
||||
|
||||
class Pooler(nn.Module):
|
||||
"""A layer that pools specific information from hidden states.
|
||||
|
||||
This layer does the following:
|
||||
1. Extracts specific tokens or aggregates data based on pooling method.
|
||||
2. Normalizes output if specified.
|
||||
3. Returns structured results as `PoolerOutput`.
|
||||
|
||||
Attributes:
|
||||
pooling_type: The type of pooling to use.
|
||||
normalize: Whether to normalize the pooled data.
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
pooling_type: PoolingType,
|
||||
normalize: bool,
|
||||
softmax: bool,
|
||||
step_tag_id: Optional[int] = None,
|
||||
returned_token_ids: Optional[List[int]] = None,
|
||||
):
|
||||
super().__init__()
|
||||
|
||||
self.pooling_type = pooling_type
|
||||
self.normalize = normalize
|
||||
self.softmax = softmax
|
||||
self.step_tag_id = step_tag_id
|
||||
self.returned_token_ids = returned_token_ids
|
||||
|
||||
@classmethod
|
||||
def from_config_with_defaults(
|
||||
cls,
|
||||
pooler_config: PoolerConfig,
|
||||
pooling_type: PoolingType,
|
||||
normalize: bool,
|
||||
softmax: bool,
|
||||
step_tag_id: Optional[int] = None,
|
||||
returned_token_ids: Optional[List[int]] = None,
|
||||
) -> Optional["Pooler"]:
|
||||
if pooler_config is None:
|
||||
return None
|
||||
return cls(
|
||||
pooling_type=PoolingType[pooler_config.pooling_type]
|
||||
if pooler_config.pooling_type is not None else pooling_type,
|
||||
normalize=pooler_config.normalize
|
||||
if pooler_config.normalize is not None else normalize,
|
||||
softmax=pooler_config.softmax
|
||||
if pooler_config.softmax is not None else softmax,
|
||||
step_tag_id=pooler_config.step_tag_id
|
||||
if pooler_config.step_tag_id is not None else step_tag_id,
|
||||
returned_token_ids=pooler_config.returned_token_ids
|
||||
if pooler_config.returned_token_ids is not None else
|
||||
returned_token_ids,
|
||||
)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
pooling_metadata: PoolingMetadata,
|
||||
) -> PoolerOutput:
|
||||
"""Pools specific information from hidden states based on metadata."""
|
||||
|
||||
prompt_lens = PoolingTensors.from_pooling_metadata(
|
||||
pooling_metadata, hidden_states.device).prompt_lens
|
||||
|
||||
if self.pooling_type is PoolingType.CLS:
|
||||
first_token_flat_indices = torch.zeros_like(prompt_lens)
|
||||
first_token_flat_indices[1:] += torch.cumsum(prompt_lens,
|
||||
dim=0)[:-1]
|
||||
pooled_data = hidden_states[first_token_flat_indices]
|
||||
elif self.pooling_type == PoolingType.LAST:
|
||||
last_token_flat_indices = torch.cumsum(prompt_lens, dim=0) - 1
|
||||
pooled_data = hidden_states[last_token_flat_indices]
|
||||
elif self.pooling_type == PoolingType.ALL:
|
||||
offset = 0
|
||||
pooled_data_lst = []
|
||||
for prompt_len in prompt_lens:
|
||||
pooled_data_i = hidden_states[offset:offset + prompt_len]
|
||||
|
||||
pooled_data_lst.append(pooled_data_i)
|
||||
offset += prompt_len
|
||||
|
||||
pooled_data = torch.stack(pooled_data_lst)
|
||||
elif self.pooling_type == PoolingType.MEAN:
|
||||
# Calculate mean pooling
|
||||
cumsum = torch.cumsum(hidden_states, dim=0)
|
||||
start_indices = torch.cat([
|
||||
torch.tensor([0], device=hidden_states.device),
|
||||
torch.cumsum(prompt_lens[:-1], dim=0)
|
||||
])
|
||||
end_indices = torch.cumsum(prompt_lens, dim=0)
|
||||
pooled_data = (
|
||||
cumsum[end_indices - 1] - cumsum[start_indices] +
|
||||
hidden_states[start_indices]) / prompt_lens.unsqueeze(1)
|
||||
elif self.pooling_type == PoolingType.STEP:
|
||||
returned_token_ids = self.returned_token_ids
|
||||
if returned_token_ids is not None and len(returned_token_ids) > 0:
|
||||
hidden_states = hidden_states[:, returned_token_ids]
|
||||
|
||||
logits = hidden_states.softmax(dim=-1)
|
||||
step_tag_id = self.step_tag_id
|
||||
|
||||
offset = 0
|
||||
pooled_data_lst = []
|
||||
for prompt_len, seq_data_i in zip(
|
||||
prompt_lens, pooling_metadata.seq_data.values()):
|
||||
pooled_data_i = logits[offset:offset + prompt_len]
|
||||
if step_tag_id is not None:
|
||||
token_ids = torch.tensor(seq_data_i.prompt_token_ids)
|
||||
pooled_data_i = pooled_data_i[token_ids == step_tag_id]
|
||||
|
||||
offset += prompt_len
|
||||
pooled_data_lst.append(pooled_data_i)
|
||||
|
||||
pooled_data = torch.stack(pooled_data_lst)
|
||||
else:
|
||||
raise ValueError(f"Invalid pooling type: {self.pooling_type}")
|
||||
|
||||
if self.normalize:
|
||||
pooled_data = nn.functional.normalize(pooled_data, p=2, dim=1)
|
||||
|
||||
if self.softmax:
|
||||
pooled_data = nn.functional.softmax(pooled_data, dim=-1)
|
||||
|
||||
pooled_outputs = [
|
||||
EmbeddingSequenceGroupOutput(data.tolist()) for data in pooled_data
|
||||
]
|
||||
|
||||
return PoolerOutput(outputs=pooled_outputs)
|
||||
@@ -0,0 +1,66 @@
|
||||
from typing import Dict, Type
|
||||
|
||||
from vllm.model_executor.layers.quantization.aqlm import AQLMConfig
|
||||
from vllm.model_executor.layers.quantization.awq import AWQConfig
|
||||
from vllm.model_executor.layers.quantization.awq_marlin import AWQMarlinConfig
|
||||
from vllm.model_executor.layers.quantization.base_config import (
|
||||
QuantizationConfig)
|
||||
from vllm.model_executor.layers.quantization.bitsandbytes import (
|
||||
BitsAndBytesConfig)
|
||||
from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors import ( # noqa: E501
|
||||
CompressedTensorsConfig)
|
||||
from vllm.model_executor.layers.quantization.deepspeedfp import (
|
||||
DeepSpeedFPConfig)
|
||||
from vllm.model_executor.layers.quantization.experts_int8 import (
|
||||
ExpertsInt8Config)
|
||||
from vllm.model_executor.layers.quantization.fbgemm_fp8 import FBGEMMFp8Config
|
||||
from vllm.model_executor.layers.quantization.fp8 import Fp8Config
|
||||
from vllm.model_executor.layers.quantization.gguf import GGUFConfig
|
||||
from vllm.model_executor.layers.quantization.gptq import GPTQConfig
|
||||
from vllm.model_executor.layers.quantization.gptq_marlin import (
|
||||
GPTQMarlinConfig)
|
||||
from vllm.model_executor.layers.quantization.gptq_marlin_24 import (
|
||||
GPTQMarlin24Config)
|
||||
from vllm.model_executor.layers.quantization.ipex_quant import IPEXConfig
|
||||
from vllm.model_executor.layers.quantization.marlin import MarlinConfig
|
||||
from vllm.model_executor.layers.quantization.modelopt import ModelOptFp8Config
|
||||
from vllm.model_executor.layers.quantization.neuron_quant import (
|
||||
NeuronQuantConfig)
|
||||
from vllm.model_executor.layers.quantization.qqq import QQQConfig
|
||||
from vllm.model_executor.layers.quantization.tpu_int8 import Int8TpuConfig
|
||||
|
||||
|
||||
QUANTIZATION_METHODS: Dict[str, Type[QuantizationConfig]] = {
|
||||
"aqlm": AQLMConfig,
|
||||
"awq": AWQConfig,
|
||||
"deepspeedfp": DeepSpeedFPConfig,
|
||||
"tpu_int8": Int8TpuConfig,
|
||||
"fp8": Fp8Config,
|
||||
"fbgemm_fp8": FBGEMMFp8Config,
|
||||
"modelopt": ModelOptFp8Config,
|
||||
"marlin": MarlinConfig,
|
||||
"gguf": GGUFConfig,
|
||||
"gptq_marlin_24": GPTQMarlin24Config,
|
||||
"gptq_marlin": GPTQMarlinConfig,
|
||||
"awq_marlin": AWQMarlinConfig,
|
||||
"gptq": GPTQConfig,
|
||||
"compressed-tensors": CompressedTensorsConfig,
|
||||
"bitsandbytes": BitsAndBytesConfig,
|
||||
"qqq": QQQConfig,
|
||||
"experts_int8": ExpertsInt8Config,
|
||||
"neuron_quant": NeuronQuantConfig,
|
||||
"ipex": IPEXConfig,
|
||||
}
|
||||
|
||||
|
||||
def get_quantization_config(quantization: str) -> Type[QuantizationConfig]:
|
||||
if quantization not in QUANTIZATION_METHODS:
|
||||
raise ValueError(f"Invalid quantization method: {quantization}")
|
||||
return QUANTIZATION_METHODS[quantization]
|
||||
|
||||
|
||||
__all__ = [
|
||||
"QuantizationConfig",
|
||||
"get_quantization_config",
|
||||
"QUANTIZATION_METHODS",
|
||||
]
|
||||
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Binary file not shown.
Some files were not shown because too many files have changed in this diff Show More
Reference in New Issue
Block a user