576 lines
18 KiB
Python
576 lines
18 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
|
|
|
|
import numpy
|
|
import torch
|
|
|
|
import vllm.envs as envs
|
|
from vllm import _custom_ops as ops
|
|
from vllm.logger import init_logger
|
|
from vllm.model_executor.layers.linear import LinearBase
|
|
from vllm.platforms import current_platform
|
|
from vllm.scalar_type import ScalarType, scalar_types
|
|
|
|
from .quant_utils import pack_cols, unpack_cols
|
|
|
|
logger = init_logger(__name__)
|
|
|
|
GPTQ_MARLIN_TILE = 16
|
|
GPTQ_MARLIN_MIN_THREAD_N = 64
|
|
GPTQ_MARLIN_MIN_THREAD_K = 128
|
|
GPTQ_MARLIN_MAX_PARALLEL = 16
|
|
|
|
MARLIN_SUPPORTED_GROUP_SIZES = [-1, 32, 64, 128]
|
|
|
|
# In case there is a performance issue with Marlin, the variable below can be
|
|
# changed to False, which allows Marlin to perform global reductions in fp16
|
|
# precision (instead of fp32), and therefore, save on some memory movements.
|
|
USE_FP32_REDUCE_DEFAULT = True
|
|
|
|
|
|
# For binary size and compile time, we don't support the same types for with and
|
|
# without runtime zero-point. We support common cases, i.e. AWQ and GPTQ.
|
|
# TODO: we may want to move this into the C++ so its closer to the actual impl
|
|
def query_marlin_supported_quant_types(
|
|
has_zp: bool | None = None,
|
|
include_fp_type: bool = True,
|
|
device_capability: int | None = None,
|
|
):
|
|
if device_capability is None:
|
|
capability_tuple = current_platform.get_device_capability()
|
|
device_capability = (
|
|
-1 if capability_tuple is None else capability_tuple.to_int()
|
|
)
|
|
|
|
if device_capability < 80:
|
|
return []
|
|
|
|
# - has_zp is True: return quant_types that has zero points
|
|
# - has_zp is False: return quant_types that has not zero points
|
|
# - has_zp is None: both
|
|
if has_zp is None:
|
|
types0 = query_marlin_supported_quant_types(
|
|
False, include_fp_type, device_capability
|
|
)
|
|
types1 = query_marlin_supported_quant_types(
|
|
True, include_fp_type, device_capability
|
|
)
|
|
return types0 + types1
|
|
|
|
if has_zp:
|
|
# AWQ style, unsigned + runtime zero-point
|
|
return [scalar_types.uint4]
|
|
else:
|
|
# GPTQ style, unsigned + symmetric bias
|
|
res = [scalar_types.uint4b8, scalar_types.uint8b128]
|
|
if include_fp_type:
|
|
res += [scalar_types.float8_e4m3fn, scalar_types.float4_e2m1f]
|
|
return res
|
|
|
|
|
|
def _check_marlin_supported(
|
|
quant_type: ScalarType,
|
|
group_size: int | None,
|
|
has_zp: bool,
|
|
device_capability: int | None = None,
|
|
) -> tuple[bool, str | None]:
|
|
if device_capability is None:
|
|
capability_tuple = current_platform.get_device_capability()
|
|
device_capability = (
|
|
-1 if capability_tuple is None else capability_tuple.to_int()
|
|
)
|
|
|
|
supported_types = query_marlin_supported_quant_types(
|
|
has_zp, True, device_capability
|
|
)
|
|
|
|
if quant_type not in supported_types:
|
|
return (
|
|
False,
|
|
f"Marlin does not support weight_bits = {quant_type}. "
|
|
f"Only types = {supported_types} "
|
|
f"are supported (for group_size = {group_size}, "
|
|
f"device_capability = {device_capability}, zp = {has_zp}).",
|
|
)
|
|
if group_size is None or group_size not in MARLIN_SUPPORTED_GROUP_SIZES:
|
|
return (
|
|
False,
|
|
f"Marlin does not support group_size = {group_size}. "
|
|
f"Only group_sizes = {MARLIN_SUPPORTED_GROUP_SIZES} "
|
|
"are supported.",
|
|
)
|
|
|
|
return True, None
|
|
|
|
|
|
def check_marlin_supported(
|
|
quant_type: ScalarType,
|
|
group_size: int,
|
|
has_zp: bool = False,
|
|
device_capability: int | None = None,
|
|
) -> bool:
|
|
cond, _ = _check_marlin_supported(quant_type, group_size, has_zp, device_capability)
|
|
return cond
|
|
|
|
|
|
def verify_marlin_supported(
|
|
quant_type: ScalarType, group_size: int, has_zp: bool = False
|
|
) -> None:
|
|
cond, err_msg = _check_marlin_supported(quant_type, group_size, has_zp)
|
|
if not cond:
|
|
assert err_msg is not None
|
|
raise ValueError(err_msg)
|
|
|
|
|
|
def verify_marlin_supports_shape(
|
|
output_size_per_partition: int,
|
|
input_size_per_partition: int,
|
|
input_size: int,
|
|
group_size: int,
|
|
) -> None:
|
|
# Validate output_size_per_partition
|
|
if output_size_per_partition % GPTQ_MARLIN_MIN_THREAD_N != 0:
|
|
raise ValueError(
|
|
f"Weight output_size_per_partition = "
|
|
f"{output_size_per_partition} is not divisible by "
|
|
f" min_thread_n = {GPTQ_MARLIN_MIN_THREAD_N}. "
|
|
"Consider reducing tensor_parallel_size or running "
|
|
"with --quantization gptq."
|
|
)
|
|
|
|
# Validate input_size_per_partition
|
|
if input_size_per_partition % GPTQ_MARLIN_MIN_THREAD_K != 0:
|
|
raise ValueError(
|
|
f"Weight input_size_per_partition = "
|
|
f"{input_size_per_partition} is not divisible "
|
|
f"by min_thread_k = {GPTQ_MARLIN_MIN_THREAD_K}. "
|
|
"Consider reducing tensor_parallel_size or running "
|
|
"with --quantization gptq."
|
|
)
|
|
|
|
if group_size < input_size and input_size_per_partition % group_size != 0:
|
|
raise ValueError(
|
|
f"Weight input_size_per_partition = {input_size_per_partition}"
|
|
f" is not divisible by group_size = {group_size}. "
|
|
"Consider reducing tensor_parallel_size or running "
|
|
"with --quantization gptq."
|
|
)
|
|
|
|
|
|
def check_marlin_supports_shape(
|
|
output_size_per_partition: int,
|
|
input_size_per_partition: int,
|
|
input_size: int,
|
|
group_size: int,
|
|
) -> tuple[bool, str | None]:
|
|
try:
|
|
verify_marlin_supports_shape(
|
|
output_size_per_partition, input_size_per_partition, input_size, group_size
|
|
)
|
|
except ValueError as e:
|
|
return False, e.__str__()
|
|
return True, None
|
|
|
|
|
|
def check_marlin_supports_layer(layer: LinearBase, group_size: int) -> bool:
|
|
output_size_per_partition = (
|
|
getattr(layer, "output_size_per_partition", None) or layer.output_size
|
|
)
|
|
input_size_per_partition = (
|
|
getattr(layer, "input_size_per_partition", None) or layer.input_size
|
|
)
|
|
|
|
return check_marlin_supports_shape(
|
|
output_size_per_partition=output_size_per_partition,
|
|
input_size_per_partition=input_size_per_partition,
|
|
input_size=layer.input_size,
|
|
group_size=group_size,
|
|
)[0]
|
|
|
|
|
|
def check_moe_marlin_supports_layer(layer: LinearBase, group_size: int) -> bool:
|
|
hidden_size = layer.hidden_size
|
|
intermediate_size_per_partition = layer.intermediate_size_per_partition
|
|
# apply_router_weight_on_input is not supported for moe marlin
|
|
supports_router_weight = not layer.apply_router_weight_on_input
|
|
# moe marlin requires the activation to be silu
|
|
supports_activation = layer.activation == "silu"
|
|
|
|
# gate-up: (n, k) = (intermediate_size_per_partition * 2, hidden_size)
|
|
# down: (n, k) = (hidden_size, intermediate_size_per_partition)
|
|
# moe marlin requires n % 128 == 0 and k % 64 == 0
|
|
supports_shape = (
|
|
hidden_size % 128 == 0
|
|
and intermediate_size_per_partition % max(64, group_size) == 0
|
|
)
|
|
supports_group_size = group_size in [-1, 32, 64, 128]
|
|
return (
|
|
supports_shape
|
|
and supports_group_size
|
|
and supports_router_weight
|
|
and supports_activation
|
|
)
|
|
|
|
|
|
def marlin_moe_intermediate_size(w1_packed: torch.Tensor, w2_packed: torch.Tensor):
|
|
"""
|
|
Given Marlin packed weight matrices w1_packed, and w2_packed,
|
|
return the MoE intermediate size N
|
|
"""
|
|
marlin_tile_size = 16
|
|
return w2_packed.size(1) * marlin_tile_size
|
|
|
|
|
|
def marlin_make_workspace(
|
|
output_size_per_partition: int, device: torch.device
|
|
) -> torch.Tensor:
|
|
max_workspace_size = (
|
|
output_size_per_partition // GPTQ_MARLIN_MIN_THREAD_N
|
|
) * GPTQ_MARLIN_MAX_PARALLEL
|
|
|
|
return torch.zeros(
|
|
max_workspace_size, dtype=torch.int, device=device, requires_grad=False
|
|
)
|
|
|
|
|
|
def marlin_make_workspace_new(
|
|
device: torch.device, max_blocks_per_sm: int = 1
|
|
) -> torch.Tensor:
|
|
# In the new marlin kernel, we use the num of threadblocks as workspace
|
|
# size. The num of threadblocks is sms_count * max_blocks_per_sm.
|
|
sms = torch.cuda.get_device_properties(device).multi_processor_count
|
|
return torch.zeros(
|
|
sms * max_blocks_per_sm, dtype=torch.int, device=device, requires_grad=False
|
|
)
|
|
|
|
|
|
def marlin_is_k_full(act_order: bool, is_row_parallel: bool) -> bool:
|
|
return (not act_order) or (act_order and not is_row_parallel)
|
|
|
|
|
|
def marlin_repeat_scales_on_all_ranks(
|
|
act_order: bool, group_size: int, is_row_parallel: bool
|
|
) -> bool:
|
|
# Need to repeat scales on every rank if act_ordering or
|
|
# channelwise and RowParallelLinear
|
|
is_channelwise = group_size == -1
|
|
return act_order or (is_channelwise and is_row_parallel)
|
|
|
|
|
|
def marlin_make_empty_g_idx(device: torch.device) -> torch.Tensor:
|
|
return torch.nn.Parameter(
|
|
torch.empty(0, dtype=torch.int, device=device), requires_grad=False
|
|
)
|
|
|
|
|
|
def marlin_make_empty_zp(device: torch.device) -> torch.Tensor:
|
|
return torch.nn.Parameter(
|
|
torch.empty(0, dtype=torch.int, device=device), requires_grad=False
|
|
)
|
|
|
|
|
|
def marlin_sort_g_idx(g_idx: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
|
|
g_idx_sort_indices = torch.argsort(g_idx).to(torch.int)
|
|
return g_idx[g_idx_sort_indices], g_idx_sort_indices
|
|
|
|
|
|
def get_scale_perms():
|
|
scale_perm: list[int] = []
|
|
for i in range(8):
|
|
scale_perm.extend([i + 8 * j for j in range(8)])
|
|
scale_perm_single: list[int] = []
|
|
for i in range(4):
|
|
scale_perm_single.extend([2 * i + j for j in [0, 1, 8, 9, 16, 17, 24, 25]])
|
|
return scale_perm, scale_perm_single
|
|
|
|
|
|
def marlin_permute_scales(
|
|
s: torch.Tensor, size_k: int, size_n: int, group_size: int
|
|
) -> torch.Tensor:
|
|
scale_perm, scale_perm_single = get_scale_perms()
|
|
if group_size < size_k and group_size != -1:
|
|
s = s.reshape((-1, len(scale_perm)))[:, scale_perm]
|
|
else:
|
|
s = s.reshape((-1, len(scale_perm_single)))[:, scale_perm_single]
|
|
s = s.reshape((-1, size_n)).contiguous()
|
|
|
|
return s
|
|
|
|
|
|
def marlin_permute_bias(s: torch.Tensor) -> torch.Tensor:
|
|
origin_shape = s.shape
|
|
_, scale_perm_single = get_scale_perms()
|
|
s = s.reshape((-1, len(scale_perm_single)))[:, scale_perm_single]
|
|
return s.reshape(*origin_shape).contiguous()
|
|
|
|
|
|
def marlin_moe_permute_scales(
|
|
s: torch.Tensor,
|
|
size_k: int,
|
|
size_n: int,
|
|
group_size: int,
|
|
):
|
|
num_experts = s.shape[0]
|
|
output = torch.empty(
|
|
(num_experts, s.shape[1], s.shape[2]),
|
|
device=s.device,
|
|
dtype=s.dtype,
|
|
)
|
|
|
|
for e in range(num_experts):
|
|
output[e] = marlin_permute_scales(s[e], size_k, size_n, group_size)
|
|
return output
|
|
|
|
|
|
def marlin_zero_points(
|
|
zp: torch.Tensor, size_k: int, size_n: int, num_bits: int
|
|
) -> torch.Tensor:
|
|
# Permute zero-points in a similar way to scales, but do not use the
|
|
# "single" permutation, since zero-points are applied on every MMA
|
|
scale_perm, _ = get_scale_perms()
|
|
zp = zp.reshape((-1, len(scale_perm)))[:, scale_perm]
|
|
|
|
# Interleave column dim (for the dequantize code) and pack it to int32
|
|
if num_bits == 4:
|
|
interleave = numpy.array([0, 2, 4, 6, 1, 3, 5, 7])
|
|
elif num_bits == 8:
|
|
interleave = numpy.array([0, 2, 1, 3])
|
|
else:
|
|
raise Exception("num_bits must be 4 or 8, got {}".format(num_bits))
|
|
|
|
zp = zp.reshape((-1, len(interleave)))[:, interleave].ravel()
|
|
zp = zp.reshape((-1, size_n)).contiguous()
|
|
zp = pack_cols(zp, num_bits, size_k, size_n)
|
|
|
|
return zp
|
|
|
|
|
|
def awq_to_marlin_zero_points(
|
|
q_zp_packed: torch.Tensor, size_k: int, size_n: int, num_bits: int
|
|
) -> torch.Tensor:
|
|
# AWQ zero-points are quantized and packed on the column dim.
|
|
# In addition, the values are permuted based on dequantizer.
|
|
# Here we undo both of these, and then apply marlin permutation
|
|
# and pack it back.
|
|
q_zp = unpack_cols(q_zp_packed, num_bits, size_k, size_n)
|
|
|
|
# Undo interleaving (use argsort(..) to get inverse perm)
|
|
if num_bits == 4:
|
|
undo_interleave = numpy.argsort(numpy.array([0, 2, 4, 6, 1, 3, 5, 7]))
|
|
elif num_bits == 8:
|
|
undo_interleave = numpy.argsort(numpy.array([0, 2, 1, 3]))
|
|
else:
|
|
raise Exception("num_bits must be 4 or 8, got {}".format(num_bits))
|
|
|
|
q_zp = q_zp.reshape((-1, len(undo_interleave)))[:, undo_interleave].ravel()
|
|
q_zp = q_zp.reshape((-1, size_n)).contiguous()
|
|
|
|
marlin_zp = marlin_zero_points(q_zp, size_k, size_n, num_bits)
|
|
return marlin_zp
|
|
|
|
|
|
def moe_awq_to_marlin_zero_points(
|
|
q_zp_packed: torch.Tensor, size_k: int, size_n: int, num_bits: int
|
|
):
|
|
num_experts = q_zp_packed.shape[0]
|
|
output = torch.empty(
|
|
(num_experts, q_zp_packed.shape[1], q_zp_packed.shape[2]),
|
|
device=q_zp_packed.device,
|
|
dtype=q_zp_packed.dtype,
|
|
)
|
|
for e in range(num_experts):
|
|
output[e] = awq_to_marlin_zero_points(q_zp_packed[e], size_k, size_n, num_bits)
|
|
return output
|
|
|
|
|
|
def maybe_warn_marlin_atomic_add(device, dtype):
|
|
if torch.compiler.is_dynamo_compiling():
|
|
return
|
|
device_capability = torch.cuda.get_device_capability(device)
|
|
if device_capability[0] < 9 and dtype == torch.bfloat16:
|
|
logger.info_once(
|
|
"You are running Marlin kernel with bf16 on GPUs before SM90. "
|
|
"You can consider change to fp16 to achieve better performance "
|
|
"if possible."
|
|
)
|
|
|
|
|
|
def maybe_warn_marlin_atomic_add_env():
|
|
if torch.compiler.is_dynamo_compiling():
|
|
return
|
|
if envs.VLLM_MARLIN_USE_ATOMIC_ADD:
|
|
return
|
|
logger.info_once(
|
|
"Marlin kernel can achieve better performance for small size_n "
|
|
"with experimental use_atomic_add feature. "
|
|
"You can consider set environment variable "
|
|
"VLLM_MARLIN_USE_ATOMIC_ADD to 1 if possible."
|
|
)
|
|
|
|
|
|
def should_use_atomic_add_reduce(
|
|
m: int, n: int, k: int, device: torch.device, dtype: torch.dtype
|
|
) -> bool:
|
|
# the performance of atomicAdd is better than global reduce
|
|
# only when m*n is small and k is large
|
|
if n >= 2048 or k < 2048 or device.type != "cuda":
|
|
return False
|
|
|
|
# disable atomicAdd reduce by default,
|
|
# one can enable it with VLLM_MARLIN_USE_ATOMIC_ADD=1
|
|
if not envs.VLLM_MARLIN_USE_ATOMIC_ADD:
|
|
maybe_warn_marlin_atomic_add_env()
|
|
return False
|
|
|
|
# sm8x doesn't support atomicAdd + bfloat16 natively
|
|
device_capability = torch.cuda.get_device_capability(device)
|
|
if device_capability[0] < 9 and dtype == torch.bfloat16:
|
|
maybe_warn_marlin_atomic_add(device, dtype)
|
|
return False
|
|
|
|
return True
|
|
|
|
|
|
def apply_gptq_marlin_linear(
|
|
input: torch.Tensor,
|
|
weight: torch.Tensor,
|
|
weight_scale: torch.Tensor,
|
|
weight_zp: torch.Tensor,
|
|
g_idx: torch.Tensor,
|
|
g_idx_sort_indices: torch.Tensor,
|
|
workspace: torch.Tensor,
|
|
wtype: ScalarType,
|
|
output_size_per_partition: int,
|
|
input_size_per_partition: int,
|
|
is_k_full: bool,
|
|
bias: torch.Tensor | None = None,
|
|
use_fp32_reduce: bool = USE_FP32_REDUCE_DEFAULT,
|
|
) -> torch.Tensor:
|
|
reshaped_x = input.reshape(-1, input.shape[-1])
|
|
out_shape = input.shape[:-1] + (output_size_per_partition,)
|
|
|
|
use_atomic_add = should_use_atomic_add_reduce(
|
|
m=reshaped_x.size(0),
|
|
n=output_size_per_partition,
|
|
k=reshaped_x.size(1),
|
|
device=input.device,
|
|
dtype=input.dtype,
|
|
)
|
|
|
|
output = ops.gptq_marlin_gemm(
|
|
reshaped_x,
|
|
None,
|
|
weight,
|
|
bias,
|
|
weight_scale,
|
|
None,
|
|
weight_zp,
|
|
g_idx,
|
|
g_idx_sort_indices,
|
|
workspace,
|
|
wtype,
|
|
size_m=reshaped_x.shape[0],
|
|
size_n=output_size_per_partition,
|
|
size_k=input_size_per_partition,
|
|
is_k_full=is_k_full,
|
|
use_atomic_add=use_atomic_add,
|
|
use_fp32_reduce=use_fp32_reduce,
|
|
is_zp_float=False,
|
|
)
|
|
|
|
return output.reshape(out_shape)
|
|
|
|
|
|
def apply_awq_marlin_linear(
|
|
input: torch.Tensor,
|
|
weight: torch.Tensor,
|
|
weight_scale: torch.Tensor,
|
|
weight_zp: torch.Tensor,
|
|
g_idx: torch.Tensor,
|
|
g_idx_sort_indices: torch.Tensor,
|
|
workspace: torch.Tensor,
|
|
quant_type: ScalarType,
|
|
output_size_per_partition: int,
|
|
input_size_per_partition: int,
|
|
bias: torch.Tensor | None = None,
|
|
use_fp32_reduce: bool = USE_FP32_REDUCE_DEFAULT,
|
|
) -> torch.Tensor:
|
|
reshaped_x = input.reshape(-1, input.shape[-1])
|
|
out_shape = input.shape[:-1] + (output_size_per_partition,)
|
|
|
|
use_atomic_add = should_use_atomic_add_reduce(
|
|
m=reshaped_x.size(0),
|
|
n=output_size_per_partition,
|
|
k=reshaped_x.size(1),
|
|
device=input.device,
|
|
dtype=input.dtype,
|
|
)
|
|
|
|
output = ops.gptq_marlin_gemm(
|
|
reshaped_x,
|
|
None,
|
|
weight,
|
|
bias,
|
|
weight_scale,
|
|
None,
|
|
weight_zp,
|
|
g_idx,
|
|
g_idx_sort_indices,
|
|
workspace,
|
|
quant_type,
|
|
size_m=reshaped_x.shape[0],
|
|
size_n=output_size_per_partition,
|
|
size_k=input_size_per_partition,
|
|
use_atomic_add=use_atomic_add,
|
|
use_fp32_reduce=use_fp32_reduce,
|
|
is_zp_float=False,
|
|
)
|
|
|
|
return output.reshape(out_shape)
|
|
|
|
|
|
def apply_rtn_marlin_linear(
|
|
input: torch.Tensor,
|
|
weight: torch.Tensor,
|
|
weight_scale: torch.Tensor,
|
|
workspace: torch.Tensor,
|
|
quant_type: ScalarType,
|
|
output_size_per_partition: int,
|
|
input_size_per_partition: int,
|
|
bias: torch.Tensor | None = None,
|
|
use_fp32_reduce: bool = USE_FP32_REDUCE_DEFAULT,
|
|
) -> torch.Tensor:
|
|
reshaped_x = input.reshape(-1, input.shape[-1])
|
|
out_shape = input.shape[:-1] + (output_size_per_partition,)
|
|
|
|
use_atomic_add = should_use_atomic_add_reduce(
|
|
m=reshaped_x.size(0),
|
|
n=output_size_per_partition,
|
|
k=reshaped_x.size(1),
|
|
device=input.device,
|
|
dtype=input.dtype,
|
|
)
|
|
|
|
output = ops.gptq_marlin_gemm(
|
|
reshaped_x,
|
|
None,
|
|
weight,
|
|
bias,
|
|
weight_scale,
|
|
None,
|
|
None,
|
|
None,
|
|
None,
|
|
workspace,
|
|
quant_type,
|
|
size_m=reshaped_x.shape[0],
|
|
size_n=output_size_per_partition,
|
|
size_k=input_size_per_partition,
|
|
use_atomic_add=use_atomic_add,
|
|
use_fp32_reduce=use_fp32_reduce,
|
|
is_zp_float=False,
|
|
)
|
|
|
|
return output.reshape(out_shape)
|