Sync from v0.13

This commit is contained in:
2026-01-19 10:38:50 +08:00
parent b2ef04d792
commit 5aef6c175a
3714 changed files with 854317 additions and 89342 deletions

View File

@@ -0,0 +1,164 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import argparse
import torch
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
from vllm.model_executor.layers.fused_moe.config import FusedMoEQuantConfig
from .common import Config
from .mk_objects import (
MK_ALL_PREPARE_FINALIZE_TYPES,
MK_FUSED_EXPERT_TYPES,
MK_SINGLE_GPU_PREPARE_FINALIZE_TYPES,
)
def make_config_arg_parser(description: str):
def to_pf_class_type(s: str) -> mk.FusedMoEPrepareAndFinalize:
for pf in MK_ALL_PREPARE_FINALIZE_TYPES:
if pf.__name__ == s:
return pf
raise ValueError(f"Cannot find a PrepareFinalize type that matches {s}")
def to_experts_class_type(s: str) -> mk.FusedMoEPermuteExpertsUnpermute:
for fe in MK_FUSED_EXPERT_TYPES:
if fe.__name__ == s:
return fe
raise ValueError(f"Cannot find a FusedExperts type that matches {s}")
def to_quant_torch_dtype(s: str) -> torch.dtype:
if s == "torch.float8_e4m3fn":
return torch.float8_e4m3fn
raise ValueError(f"Unsupported quant type {s}")
parser = argparse.ArgumentParser(description=description)
parser.add_argument(
"--world-size",
type=int,
default=2,
help="Number of ranks that participate in all2all",
)
parser.add_argument(
"--pf-type",
type=to_pf_class_type,
required=True,
help=(
"Choose a PrepareFinalize Type : "
f"{[x.__name__ for x in MK_ALL_PREPARE_FINALIZE_TYPES]}"
),
)
parser.add_argument(
"--experts-type",
type=to_experts_class_type,
required=True,
help=(
f"Choose a FusedExpert type : {[x.__name__ for x in MK_FUSED_EXPERT_TYPES]}"
),
)
parser.add_argument(
"-m",
nargs="+",
type=int,
default=[64],
help="num tokens per rank",
)
parser.add_argument(
"-k",
type=int,
default=7168,
help="hidden-size",
)
parser.add_argument(
"-n",
type=int,
default=1024,
help="N dimension of the first fused-moe matmul",
)
parser.add_argument(
"--num-experts", type=int, default=32, help="Global num experts"
)
parser.add_argument("--topk", nargs="+", type=int, default=[4, 1], help="num topk")
parser.add_argument(
"--fused-moe-chunk-size",
type=int,
help="Fused moe chunk size used for the non-batched fused experts impl.",
)
# Quant args
parser.add_argument(
"--quant-dtype", type=to_quant_torch_dtype, help="Quant datatype"
)
parser.add_argument(
"--per-token-quantized-activations",
action="store_true",
help=("The input activations must be per-token quantized"),
)
parser.add_argument(
"--per-channel-quantized-weights",
action="store_true",
help="The weights must be per-channel quantized.",
)
parser.add_argument(
"--block-shape", nargs="+", type=int, help="Quantization block shape"
)
# Torch trace profile generation args
parser.add_argument(
"--torch-trace-dir-path",
type=str,
default=None,
help="Get torch trace for single execution",
)
return parser
def _validate_args(args: argparse.Namespace):
if args.quant_dtype is not None:
assert args.quant_dtype == torch.float8_e4m3fn
if args.block_shape is not None:
assert len(args.block_shape) == 2, (
f"block shape must have 2 elements. got {args.block_shape}"
)
if args.experts_type in MK_SINGLE_GPU_PREPARE_FINALIZE_TYPES:
assert args.world_size == 1, "Single GPU objects need world size set to 1"
if args.torch_trace_dir_path is not None:
from pathlib import Path
assert Path(args.torch_trace_dir_path).is_dir(), (
f"Please create {args.torch_trace_dir_path}"
)
def make_config(args: argparse.Namespace) -> Config:
_validate_args(args)
quant_config = None
if args.quant_dtype is not None:
quant_config = FusedMoEQuantConfig(
quant_dtype=args.quant_dtype,
per_act_token_quant=args.per_token_quantized_activations,
per_out_ch_quant=args.per_channel_quantized_weights,
block_shape=args.block_shape,
)
return Config(
Ms=args.m,
K=args.k,
N=args.n,
E=args.num_experts,
topks=args.topk,
dtype=torch.bfloat16, # hard-code
quant_config=quant_config,
prepare_finalize_type=args.pf_type,
fused_experts_type=args.experts_type,
fused_moe_chunk_size=args.fused_moe_chunk_size,
world_size=args.world_size,
torch_trace_dir_path=args.torch_trace_dir_path,
)

View File

@@ -0,0 +1,668 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from dataclasses import dataclass
from typing import Any
import torch
import vllm._custom_ops as ops
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
from tests.kernels.moe.utils import make_test_weights, per_token_cast_to_fp8
from tests.kernels.quantization.nvfp4_utils import (
FLOAT4_E2M1_MAX,
FLOAT8_E4M3_MAX,
dequantize_nvfp4_to_dtype,
)
from tests.kernels.utils import torch_experts
from vllm.config import VllmConfig
from vllm.distributed import (
get_dp_group,
get_pcp_group,
get_tensor_model_parallel_world_size,
)
from vllm.forward_context import set_forward_context
from vllm.model_executor.layers.fused_moe.config import (
FusedMoEConfig,
FusedMoEParallelConfig,
FusedMoEQuantConfig,
)
from vllm.model_executor.layers.fused_moe.fused_moe import fused_topk
from vllm.utils.import_utils import has_deep_ep, has_deep_gemm, has_pplx
from .mk_objects import (
TestMoEQuantConfig,
expert_info,
make_fused_experts,
make_prepare_finalize,
prepare_finalize_info,
)
from .parallel_utils import ProcessGroupInfo
def _describe_tensor(t: torch.Tensor | None, name: str) -> str:
if t is None:
return f"{name} : None"
else:
return f"{name} : {t.shape} {t.dtype} {t.device}"
@dataclass
class Config:
Ms: list[int] | int
K: int
N: int
E: int
topks: list[int] | int
dtype: torch.dtype
quant_config: TestMoEQuantConfig | None
prepare_finalize_type: mk.FusedMoEPrepareAndFinalize
fused_experts_type: mk.FusedMoEPermuteExpertsUnpermute
fused_moe_chunk_size: int | None
world_size: int
torch_trace_dir_path: str | None = None
def __post_init__(self):
if self.quant_config is None:
self.quant_config = TestMoEQuantConfig(None, False, False, None)
def describe(self) -> str:
s = ""
s += "== Config:\n"
s += f" world_size={self.world_size}\n"
s += f" PF={self.prepare_finalize_type.__name__}\n"
s += f" FE={self.fused_experts_type.__name__}\n"
s += f" E={self.E}\n"
s += f" Ms={self.Ms}\n"
s += f" N={self.N}\n"
s += f" K={self.K}\n"
s += f" topk={self.topks}\n"
s += f" dtype={self.dtype}\n"
s += f" fused_moe_chunk_size={self.fused_moe_chunk_size}\n"
s += " Quant:\n"
if self.quant_config is not None:
s += f" q_dtype={self.quant_dtype}\n"
s += f" q_block_shape={self.quant_block_shape}\n"
s += f" q_per_out_ch_quant={self.is_per_out_ch_quant}\n"
s += f" q_per_act_token={self.is_per_act_token_quant}\n"
else:
s += " quant=None\n"
return s
@property
def M(self) -> int:
assert isinstance(self.Ms, int)
return self.Ms
@property
def quant_dtype(self) -> torch.dtype | str | None:
assert self.quant_config is not None
return self.quant_config.quant_dtype
@property
def is_per_act_token_quant(self) -> bool:
assert self.quant_config is not None
return self.quant_config.per_act_token_quant
@property
def is_per_tensor_act_quant(self) -> bool:
return not self.is_per_act_token_quant and self.quant_block_shape is None
@property
def is_per_out_ch_quant(self) -> bool:
assert self.quant_config is not None
return self.quant_config.per_out_ch_quant
@property
def quant_block_shape(self) -> list[int] | None:
assert self.quant_config is not None
return self.quant_config.block_shape
@property
def topk(self) -> int:
assert isinstance(self.topks, int)
return self.topks
@property
def num_local_experts(self) -> int:
return self.E // self.world_size
def make_env_data(self) -> tuple[VllmConfig, dict[Any, Any]]:
"""
make env data for vllm launch.
"""
vllm_config = VllmConfig()
vllm_config.parallel_config.data_parallel_size = self.world_size
vllm_config.parallel_config.enable_expert_parallel = True
env_dict = {
"VLLM_USE_DEEP_GEMM": str(int(self.needs_deep_gemm())),
}
backend = self.all2all_backend()
vllm_config.parallel_config.all2all_backend = backend
if backend is not None:
env_dict.update({"VLLM_ALL2ALL_BACKEND": backend})
if self.fused_moe_chunk_size is not None:
env_dict.update(
{"VLLM_FUSED_MOE_CHUNK_SIZE": str(self.fused_moe_chunk_size)}
)
return vllm_config, env_dict
def is_fp8_block_quantized(self):
return (
self.quant_dtype == torch.float8_e4m3fn
and self.quant_block_shape is not None
)
def is_batched_prepare_finalize(self):
info = prepare_finalize_info(self.prepare_finalize_type)
return mk.FusedMoEActivationFormat.BatchedExperts == info.activation_format
def is_batched_fused_experts(self):
info = expert_info(self.fused_experts_type)
return mk.FusedMoEActivationFormat.BatchedExperts == info.activation_format
def is_standard_fused_experts(self):
info = expert_info(self.fused_experts_type)
return mk.FusedMoEActivationFormat.Standard == info.activation_format
def fe_supported_types(self):
info = expert_info(self.fused_experts_type)
return info.supported_dtypes
def pf_supported_types(self):
info = prepare_finalize_info(self.prepare_finalize_type)
return info.supported_dtypes
def is_block_quant_supported(self):
info = expert_info(self.fused_experts_type)
return info.blocked_quantization_support
def is_fe_supports_chunking(self):
info = expert_info(self.fused_experts_type)
return info.supports_chunking
def supports_expert_map(self):
info = expert_info(self.fused_experts_type)
return info.supports_expert_map
def supports_apply_weight_on_input(self):
info = prepare_finalize_info(self.prepare_finalize_type)
return info.supports_apply_weight_on_input
def needs_deep_gemm(self):
info = expert_info(self.fused_experts_type)
return info.needs_deep_gemm
def needs_pplx(self):
info = prepare_finalize_info(self.prepare_finalize_type)
return info.backend == "pplx"
def needs_deep_ep(self):
info = prepare_finalize_info(self.prepare_finalize_type)
return (
info.backend == "deepep_high_throughput"
or info.backend == "deepep_low_latency"
)
def all2all_backend(self):
info = prepare_finalize_info(self.prepare_finalize_type)
return info.backend
def is_valid(self) -> tuple[bool, str | None]:
# Check prepare-finalize and fused-experts compatibility
if self.is_batched_prepare_finalize():
if not self.is_batched_fused_experts():
return False, "Mismatched format."
else:
if not self.is_standard_fused_experts():
return False, "Mismatched format."
use_chunking = self.fused_moe_chunk_size is not None
if use_chunking and not self.is_fe_supports_chunking():
return False, "Chunking not supported."
# Check quantization sanity
if (
int(self.is_per_act_token_quant)
+ int(self.is_per_tensor_act_quant)
+ int(self.quant_block_shape is not None)
) > 1:
# invalid quant config
return False, f"Bad quant_config {self.quant_config}."
# check type support
if self.quant_dtype is None:
if (
self.dtype not in self.pf_supported_types()
or self.dtype not in self.fe_supported_types()
):
return False, (
f"Unsupported type {self.dtype} not in "
f"{self.pf_supported_types()} and "
f"{self.fe_supported_types()}."
)
else:
if (
self.quant_dtype not in self.pf_supported_types()
or self.quant_dtype not in self.fe_supported_types()
):
return False, (
f"Unsupported quant type {self.quant_dtype} "
f"not in {self.pf_supported_types()} and "
f"{self.fe_supported_types()}."
)
# Check block quanization support
is_block_quatized = self.quant_block_shape is not None
if is_block_quatized and self.quant_dtype is None:
return False, "No block quantization support."
if is_block_quatized and not self.is_block_quant_supported():
return False, "Mismatched block quantization support."
# deep_gemm only works with block-quantized
if self.needs_deep_gemm() and not is_block_quatized:
return False, "Needs DeepGEMM but not block quantized."
# Check dependencies (turn into asserts?)
if self.needs_deep_ep() and not has_deep_ep():
return False, "Needs DeepEP, but DeepEP not available."
if self.needs_deep_gemm() and not has_deep_gemm():
return False, "Needs DeepGEMM, but DeepGEMM not available."
if self.needs_pplx() and not has_pplx(): # noqa: SIM103
return False, "Needs PPLX, but PPLX not available."
return True, None
@dataclass
class WeightTensors:
w1: torch.Tensor
w2: torch.Tensor
w1_scale: torch.Tensor | None
w2_scale: torch.Tensor | None
w1_gs: torch.Tensor | None = None
w2_gs: torch.Tensor | None = None
def describe(self):
s = ""
s += "== Weight Tensors: \n"
s += f" - {_describe_tensor(self.w1, 'w1')} \n"
s += f" - {_describe_tensor(self.w2, 'w2')} \n"
s += f" - {_describe_tensor(self.w1_scale, 'w1_scale')} \n"
s += f" - {_describe_tensor(self.w2_scale, 'w2_scale')} \n"
s += f" - {_describe_tensor(self.w1_gs, 'w1_gs')} \n"
s += f" - {_describe_tensor(self.w2_gs, 'w2_gs')} \n"
return s
def is_quantized(self) -> bool:
# or w1_scale is not None?
return (
self.w1.dtype == torch.float8_e4m3fn
or self.w1.dtype == torch.uint8
or self.w1.dtype == torch.int8
)
def to_current_device(self):
device = torch.cuda.current_device()
self.w1 = self.w1.to(device=device)
self.w2 = self.w2.to(device=device)
if self.w1_scale is not None:
self.w1_scale = self.w1_scale.to(device=device)
if self.w2_scale is not None:
self.w2_scale = self.w2_scale.to(device=device)
if self.w1_gs is not None:
self.w1_gs = self.w1_gs.to(device=device)
if self.w2_gs is not None:
self.w2_gs = self.w2_gs.to(device=device)
def slice_weights(self, rank: int, num_local_experts: int) -> "WeightTensors":
s = rank * num_local_experts
e = s + num_local_experts
w1 = self.w1[s:e, :, :]
w2 = self.w2[s:e, :, :]
w1_scale = self.w1_scale[s:e, :, :] if self.w1_scale is not None else None
w2_scale = self.w2_scale[s:e, :, :] if self.w2_scale is not None else None
w1_gs = self.w1_gs[s:e] if self.w1_gs is not None else None
w2_gs = self.w2_gs[s:e] if self.w2_gs is not None else None
return WeightTensors(w1, w2, w1_scale, w2_scale, w1_gs, w2_gs)
@staticmethod
def make(config: Config) -> "WeightTensors":
(_, w1, w1_scale, w1_gs), (_, w2, w2_scale, w2_gs) = make_test_weights(
e=config.E,
n=config.N,
k=config.K,
in_dtype=config.dtype,
quant_dtype=config.quant_dtype,
block_shape=config.quant_block_shape,
# or config.is_per_out_ch_quant
per_out_ch_quant=config.is_per_act_token_quant,
)
return WeightTensors(
w1=w1, w2=w2, w1_scale=w1_scale, w2_scale=w2_scale, w1_gs=w1_gs, w2_gs=w2_gs
)
@dataclass
class RankTensors:
hidden_states: torch.Tensor
hidden_states_scale: torch.Tensor | None
topk_weights: torch.Tensor
topk_ids: torch.Tensor
expert_map: torch.Tensor | None
def describe(self):
s = ""
s += "== Rank Tensors: \n"
s += f" - {_describe_tensor(self.hidden_states, 'HS')} \n"
s += f" - {_describe_tensor(self.hidden_states_scale, 'HS_scale')} \n"
s += f" - {_describe_tensor(self.topk_weights, 'topk_weights')} \n"
s += f" - {_describe_tensor(self.topk_ids, 'topk_ids')} \n"
s += f" - {_describe_tensor(self.expert_map, 'expert_map')} \n"
return s
@staticmethod
def make_hidden_states(
config: Config,
) -> tuple[torch.Tensor, torch.Tensor | None]:
"""
Return hidden_states
"""
m, k, dtype = (config.M, config.K, config.dtype)
a = torch.randn((m, k), device=torch.cuda.current_device(), dtype=dtype) / 15.0
if config.quant_dtype is None:
return a, None
# We dequant and use that as hidden_states so the tests are stable.
# quantizing and dequantizing yield slightly different results
# depending on the hardware. Here we, quantize and dequantize
# first - so further quantize and dequantize will yield the same
# values.
if config.is_per_tensor_act_quant:
a_q, a_scales = ops.scaled_fp8_quant(a, use_per_token_if_dynamic=False)
return a_q.float().mul(a_scales).to(dtype), a_scales
if config.is_per_act_token_quant:
a_q, a_scales = ops.scaled_fp8_quant(a, use_per_token_if_dynamic=True)
return a_q.float().mul(a_scales).to(dtype), None
assert config.quant_block_shape is not None
block_k = config.quant_block_shape[1]
a_q, a_scales = per_token_cast_to_fp8(a, block_size=block_k)
return a_q.float().view((-1, block_k)).mul(a_scales.view(-1, 1)).view(m, k).to(
dtype
), None
@staticmethod
def make(config: Config, pgi: ProcessGroupInfo):
dtype = config.dtype
topk, m, _ = (config.topk, config.M, config.K)
hidden_states, hidden_states_scale = RankTensors.make_hidden_states(config)
num_local_experts, global_num_experts = (config.num_local_experts, config.E)
score = torch.randn((m, global_num_experts), device="cuda", dtype=dtype)
topk_weights, topk_ids, _ = fused_topk(hidden_states, score, topk, False)
# distribute topk_ids evenly
for mi in range(m):
topk_ids[mi] = torch.randperm(config.E)[:topk]
topk_ids = topk_ids.to(device=torch.cuda.current_device())
expert_map = None
if config.world_size > 1 and config.supports_expert_map():
expert_map = torch.full(
(global_num_experts,), fill_value=-1, dtype=torch.int32
)
s = pgi.rank * num_local_experts
e = s + num_local_experts
expert_map[s:e] = torch.tensor(list(range(num_local_experts)))
expert_map = expert_map.to(
device=torch.cuda.current_device(), dtype=torch.int32
)
return RankTensors(
hidden_states=hidden_states,
hidden_states_scale=hidden_states_scale,
topk_weights=topk_weights,
topk_ids=topk_ids,
expert_map=expert_map,
)
def reference_moe_impl(
config: Config, weights: WeightTensors, rank_tensors: RankTensors
) -> torch.Tensor:
if config.quant_dtype == "nvfp4":
quant_blocksize = 16
dtype = config.dtype
w1_q = weights.w1
w1_blockscale = weights.w1_scale
w1_gs = weights.w1_gs
w2_q = weights.w2
w2_blockscale = weights.w2_scale
w2_gs = weights.w2_gs
a_global_scale = (
(FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX)
/ torch.amax(rank_tensors.hidden_states.flatten(), dim=-1)
).to(torch.float32)
assert w1_gs is not None
assert w2_gs is not None
assert w1_blockscale is not None
assert w2_blockscale is not None
assert w1_blockscale.shape[1] % 128 == 0
assert w1_blockscale.shape[2] % 4 == 0
assert w2_blockscale.shape[1] % 128 == 0
assert w2_blockscale.shape[2] % 4 == 0
a_fp4, a_scale_interleaved = ops.scaled_fp4_quant(
rank_tensors.hidden_states, a_global_scale
)
a = dequantize_nvfp4_to_dtype(
a_fp4,
a_scale_interleaved,
a_global_scale,
dtype=dtype,
device=a_fp4.device,
block_size=quant_blocksize,
)
e = w1_q.shape[0]
n = w1_q.shape[1] // 2
k = w2_q.shape[1]
w1 = torch.zeros((e, 2 * n, k), device="cuda", dtype=dtype)
w2 = torch.zeros((e, k, n), device="cuda", dtype=dtype)
for idx in range(0, e):
w1[idx] = dequantize_nvfp4_to_dtype(
w1_q[idx],
w1_blockscale[idx],
w1_gs[idx],
dtype=dtype,
device=w1_q.device,
block_size=quant_blocksize,
)
w2[idx] = dequantize_nvfp4_to_dtype(
w2_q[idx],
w2_blockscale[idx],
w2_gs[idx],
dtype=dtype,
device=w2_q.device,
block_size=quant_blocksize,
)
a_scale = None
w1_scale = None
w2_scale = None
quant_dtype = None
per_act_token_quant = False
block_shape = None
else:
a = rank_tensors.hidden_states
a_scale = rank_tensors.hidden_states_scale
w1 = weights.w1
w1_scale = weights.w1_scale
w2 = weights.w2
w2_scale = weights.w2_scale
quant_dtype = config.quant_dtype
per_act_token_quant = config.is_per_act_token_quant
block_shape = config.quant_block_shape
return torch_experts(
a=a,
w1=w1,
w2=w2,
topk_weight=rank_tensors.topk_weights,
topk_ids=rank_tensors.topk_ids,
global_num_experts=config.E,
expert_map=None,
w1_scale=w1_scale,
w2_scale=w2_scale,
a1_scale=a_scale,
quant_dtype=quant_dtype,
per_act_token_quant=per_act_token_quant,
block_shape=block_shape,
apply_router_weights_on_input=config.topk == 1
and config.supports_apply_weight_on_input(),
)
def _make_gscale(num_experts: int) -> torch.Tensor:
return torch.ones(
(num_experts,), device=torch.cuda.current_device(), dtype=torch.float32
)
def make_modular_kernel(
config: Config,
vllm_config: VllmConfig,
quant_config: FusedMoEQuantConfig,
) -> mk.FusedMoEModularKernel:
def next_power_of_2(x):
import math
if x == 0:
return 1
return 2 ** math.ceil(math.log2(x))
# make moe config
moe_parallel_config: FusedMoEParallelConfig = FusedMoEParallelConfig.make(
tp_size_=get_tensor_model_parallel_world_size(),
pcp_size_=get_pcp_group().world_size,
dp_size_=get_dp_group().world_size,
vllm_parallel_config=vllm_config.parallel_config,
)
moe = FusedMoEConfig(
num_experts=config.E,
experts_per_token=config.topk,
hidden_dim=config.K,
num_local_experts=config.num_local_experts,
moe_parallel_config=moe_parallel_config,
in_dtype=config.dtype,
max_num_tokens=next_power_of_2(config.M),
)
# make modular kernel
prepare_finalize = make_prepare_finalize(
config.prepare_finalize_type, config.all2all_backend(), moe, quant_config
)
fused_experts = make_fused_experts(
config.fused_experts_type,
moe,
quant_config,
prepare_finalize.num_dispatchers(),
config.N,
)
modular_kernel = mk.FusedMoEModularKernel(
prepare_finalize=prepare_finalize,
fused_experts=fused_experts,
)
return modular_kernel
def run_modular_kernel(
pgi: ProcessGroupInfo,
vllm_config: VllmConfig,
config: Config,
weights: WeightTensors,
rank_tensors: RankTensors,
) -> torch.Tensor:
assert isinstance(config.Ms, int)
assert isinstance(config.topks, int)
# weights for rank
rank_weights = weights.slice_weights(pgi.rank, config.num_local_experts)
if config.quant_dtype == "nvfp4":
gscale = _make_gscale(config.num_local_experts)
else:
gscale = None
quant_config = FusedMoEQuantConfig.make(
config.quant_dtype,
w1_scale=rank_weights.w1_scale,
w2_scale=rank_weights.w2_scale,
a1_scale=rank_tensors.hidden_states_scale,
g1_alphas=(1 / rank_weights.w1_gs) if rank_weights.w1_gs is not None else None,
g2_alphas=(1 / rank_weights.w2_gs) if rank_weights.w2_gs is not None else None,
a1_gscale=gscale,
a2_gscale=gscale,
block_shape=config.quant_block_shape,
per_act_token_quant=config.is_per_act_token_quant,
per_out_ch_quant=config.is_per_out_ch_quant,
)
mk = make_modular_kernel(config, vllm_config, quant_config)
# impls might update the tensor in place
hidden_states = rank_tensors.hidden_states.clone()
topk_ids = rank_tensors.topk_ids.to(mk.prepare_finalize.topk_indices_dtype())
mk_kwargs = {
"hidden_states": hidden_states,
"w1": rank_weights.w1,
"w2": rank_weights.w2,
"topk_weights": rank_tensors.topk_weights,
"topk_ids": topk_ids,
"expert_map": rank_tensors.expert_map,
"global_num_experts": config.E,
"apply_router_weight_on_input": config.topk == 1
and config.supports_apply_weight_on_input(),
}
num_tokens = rank_tensors.hidden_states.shape[0]
num_tokens_across_dp = torch.tensor(
[num_tokens] * config.world_size, device="cuda", dtype=torch.int
)
with set_forward_context(
None,
vllm_config,
num_tokens=num_tokens,
num_tokens_across_dp=num_tokens_across_dp,
):
out = mk.forward(**mk_kwargs)
return out

View File

@@ -0,0 +1,196 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import copy
from enum import Enum
from itertools import product
import torch
from tqdm import tqdm
from vllm.config import VllmConfig, set_current_vllm_config
from vllm.model_executor.layers.fused_moe.config import FUSED_MOE_UNQUANTIZED_CONFIG
from vllm.platforms import current_platform
from .common import (
Config,
RankTensors,
WeightTensors,
reference_moe_impl,
run_modular_kernel,
)
from .mk_objects import (
MK_FUSED_EXPERT_TYPES,
MK_MULTI_GPU_PREPARE_FINALIZE_TYPES,
MK_QUANT_CONFIGS,
)
from .parallel_utils import ProcessGroupInfo, parallel_launch_with_config
class Result(Enum):
PASS = 1
FAIL = 2
SKIP = 3
def rank_worker(
pgi: ProcessGroupInfo,
vllm_config: VllmConfig,
cpu_group,
config: Config,
weights: WeightTensors,
):
current_platform.seed_everything(pgi.rank)
# sanity check
from vllm import envs
if config.fused_moe_chunk_size is not None:
assert config.fused_moe_chunk_size == envs.VLLM_FUSED_MOE_CHUNK_SIZE
# get weights to this device
weights.to_current_device()
Ms = config.Ms
assert isinstance(Ms, list)
TOPKs = config.topks
assert isinstance(TOPKs, list)
for m, topk in product(Ms, TOPKs):
print(f"Running m={m}, topk={topk} ...")
# override m and topk
cfgx = copy.deepcopy(config)
cfgx.Ms = m
cfgx.topks = topk
# inputs for rank
rank_tensors = RankTensors.make(cfgx, pgi)
# modular kernel out
mk_out = run_modular_kernel(pgi, vllm_config, cfgx, weights, rank_tensors)
with set_current_vllm_config(vllm_config):
ref_out = reference_moe_impl(cfgx, weights, rank_tensors)
torch.testing.assert_close(ref_out, mk_out, atol=3e-2, rtol=3e-2)
def make_feature_matrix(csv_file_path: str):
from dataclasses import asdict
import pandas as pd
def add_to_results(
config: Config, success: Result, results_df: pd.DataFrame | None = None
):
config_dict = asdict(config)
config_dict["prepare_finalize_type"] = config_dict[
"prepare_finalize_type"
].__name__
config_dict["fused_experts_type"] = config_dict["fused_experts_type"].__name__
config_dict["per_tensor_act_quant"] = config.is_per_tensor_act_quant
quant_config_dict = config_dict["quant_config"]
del config_dict["quant_config"]
if quant_config_dict is None:
quant_config = FUSED_MOE_UNQUANTIZED_CONFIG
quant_config_dict = asdict(quant_config)
config_dict |= quant_config_dict
result_dict = config_dict | {"success": success.name}
result_df = pd.DataFrame([result_dict])
if results_df is None:
results_df = result_df
else:
results_df = pd.concat([results_df, result_df], ignore_index=True)
return results_df
Ms = [64]
Ks = [7168] # hidden sizes
Ns = [2048]
TOPKs = [[4, 1]]
Es = [32]
DTYPEs = [torch.bfloat16]
PF_TYPES = MK_MULTI_GPU_PREPARE_FINALIZE_TYPES
FE_TYPES = MK_FUSED_EXPERT_TYPES
Q_TYPES = MK_QUANT_CONFIGS
combinations = list(
product(Ms, Ks, Ns, Es, TOPKs, DTYPEs, PF_TYPES, FE_TYPES, Q_TYPES)
)
results_df: pd.DataFrame | None = None
for m, k, n, e, topks, dtype, pf_type, experts_type, quant_config in tqdm(
combinations
):
config = Config(
Ms=[m],
K=k,
N=n,
E=e,
topks=topks,
dtype=dtype,
prepare_finalize_type=pf_type,
fused_experts_type=experts_type,
quant_config=quant_config,
world_size=2,
fused_moe_chunk_size=None,
)
success = None
if config.is_valid()[0]:
print(f"Running config : {config.describe()} ...")
try:
weights: WeightTensors = WeightTensors.make(config)
vllm_config, env_dict = config.make_env_data()
parallel_launch_with_config(
config.world_size,
rank_worker,
vllm_config,
env_dict,
config,
weights,
)
success = Result.PASS
except Exception as _:
success = Result.FAIL
else:
success = Result.SKIP
results_df = add_to_results(config, success, results_df)
if results_df is not None:
results_df.to_csv(f"{csv_file_path}")
if __name__ == "__main__":
import argparse
from pathlib import Path
parser = argparse.ArgumentParser(
description=(
"Make ModularKernel feature matrix \n"
"Example : python3 -m tests.kernels.moe.modular_kernel_tools.make_feature_matrix " # noqa: E501
"-f ./feature_matrices/feature_matrix.csv"
)
)
parser.add_argument(
"-f",
"--feature-matrix-csv-file-path",
type=str,
required=True,
help="File name to Generate a .csv file",
)
args = parser.parse_args()
csv_path = args.feature_matrix_csv_file_path
assert csv_path.endswith("csv"), (
f"Need a file path ending with .csv, got {csv_path}"
)
assert Path(csv_path).parent.is_dir(), (
f"Cannot find parent directory for {Path(csv_path).parent}"
)
make_feature_matrix(args.feature_matrix_csv_file_path)

View File

@@ -0,0 +1,509 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from dataclasses import dataclass
import torch
# Fused experts and PrepareFinalize imports
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
from vllm.model_executor.layers.fused_moe import TritonExperts
from vllm.model_executor.layers.fused_moe.all2all_utils import (
maybe_make_prepare_finalize,
)
from vllm.model_executor.layers.fused_moe.batched_deep_gemm_moe import (
BatchedDeepGemmExperts,
)
from vllm.model_executor.layers.fused_moe.config import (
FusedMoEConfig,
FusedMoEQuantConfig,
)
from vllm.model_executor.layers.fused_moe.deep_gemm_moe import DeepGemmExperts
from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
BatchedTritonExperts,
NaiveBatchedExperts,
)
from vllm.model_executor.layers.fused_moe.prepare_finalize import (
MoEPrepareAndFinalizeNoEP,
)
from vllm.model_executor.layers.fused_moe.triton_deep_gemm_moe import (
TritonOrDeepGemmExperts,
)
from vllm.model_executor.layers.quantization.utils.quant_utils import (
cutlass_fp4_supported,
)
from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
cutlass_fp8_supported,
)
from vllm.platforms import current_platform
from vllm.utils.deep_gemm import is_deep_gemm_supported
from vllm.utils.flashinfer import has_flashinfer_cutlass_fused_moe
from vllm.utils.import_utils import has_deep_ep, has_deep_gemm, has_pplx
@dataclass
class TestMoEQuantConfig:
quant_dtype: torch.dtype | str | None
per_out_ch_quant: bool
per_act_token_quant: bool
block_shape: list[int] | None
@dataclass
class PrepareFinalizeInfo:
activation_format: mk.FusedMoEActivationFormat
supported_dtypes: list[torch.dtype | str]
blocked_quantization_support: bool
backend: str | None
supports_apply_weight_on_input: bool = True
@dataclass
class ExpertInfo:
activation_format: mk.FusedMoEActivationFormat
supported_dtypes: list[torch.dtype | str]
blocked_quantization_support: bool
supports_chunking: bool
supports_expert_map: bool
needs_matching_quant: bool = False
needs_deep_gemm: bool = False
PREPARE_FINALIZE_INFO: dict[mk.FusedMoEPrepareAndFinalize, PrepareFinalizeInfo] = {}
EXPERT_INFO: dict[mk.FusedMoEPermuteExpertsUnpermute, ExpertInfo] = {}
MK_ALL_PREPARE_FINALIZE_TYPES: list[mk.FusedMoEPrepareAndFinalize] = []
MK_MULTI_GPU_PREPARE_FINALIZE_TYPES: list[mk.FusedMoEPrepareAndFinalize] = []
MK_SINGLE_GPU_PREPARE_FINALIZE_TYPES: list[mk.FusedMoEPrepareAndFinalize] = []
MK_FUSED_EXPERT_TYPES: list[mk.FusedMoEPermuteExpertsUnpermute] = []
standard_format = mk.FusedMoEActivationFormat.Standard
batched_format = mk.FusedMoEActivationFormat.BatchedExperts
common_float_types: list[torch.dtype | str] = [
torch.float8_e4m3fn,
torch.bfloat16,
torch.float16,
torch.float32,
]
common_float_and_int_types = common_float_types + [torch.int8]
nvfp4_types = ["nvfp4"]
fp8_types = [torch.float8_e4m3fn]
def register_prepare_and_finalize(
kind,
activation_format: mk.FusedMoEActivationFormat,
supported_dtypes: list[torch.dtype | str],
blocked_quantization_support: bool,
backend: str | None,
force_multigpu: bool = False,
supports_apply_weight_on_input: bool = True,
):
global PREPARE_FINALIZE_INFO
global MK_ALL_PREPARE_FINALIZE_TYPES
global MK_MULTI_GPU_PREPARE_FINALIZE_TYPES
global MK_SINGLE_GPU_PREPARE_FINALIZE_TYPES
assert kind not in PREPARE_FINALIZE_INFO
PREPARE_FINALIZE_INFO[kind] = PrepareFinalizeInfo(
activation_format,
supported_dtypes,
blocked_quantization_support,
backend,
supports_apply_weight_on_input,
)
MK_ALL_PREPARE_FINALIZE_TYPES.append(kind)
if backend is not None or force_multigpu:
MK_MULTI_GPU_PREPARE_FINALIZE_TYPES.append(kind)
else:
MK_SINGLE_GPU_PREPARE_FINALIZE_TYPES.append(kind)
def register_experts(
kind,
activation_format: mk.FusedMoEActivationFormat,
supported_dtypes: list[torch.dtype | str],
blocked_quantization_support: bool,
supports_chunking: bool,
supports_expert_map: bool,
needs_matching_quant: bool = False,
needs_deep_gemm: bool = False,
):
global EXPERT_INFO
global MK_FUSED_EXPERT_TYPES
assert kind not in EXPERT_INFO
EXPERT_INFO[kind] = ExpertInfo(
activation_format,
supported_dtypes,
blocked_quantization_support,
supports_chunking,
supports_expert_map,
needs_matching_quant,
needs_deep_gemm,
)
MK_FUSED_EXPERT_TYPES.append(kind)
def prepare_finalize_info(kind) -> PrepareFinalizeInfo:
info = PREPARE_FINALIZE_INFO.get(kind)
assert info is not None
return info
def expert_info(kind) -> ExpertInfo:
info = EXPERT_INFO.get(kind)
assert info is not None
return info
register_prepare_and_finalize(
MoEPrepareAndFinalizeNoEP,
standard_format,
common_float_types,
blocked_quantization_support=True,
backend=None,
)
register_experts(
BatchedTritonExperts,
batched_format,
common_float_types,
blocked_quantization_support=True,
supports_chunking=False,
supports_expert_map=False,
needs_matching_quant=True,
)
register_experts(
TritonExperts,
standard_format,
common_float_and_int_types,
blocked_quantization_support=True,
supports_chunking=True,
supports_expert_map=True,
needs_matching_quant=True,
)
register_experts(
NaiveBatchedExperts,
batched_format,
common_float_and_int_types,
blocked_quantization_support=True,
supports_chunking=False,
supports_expert_map=True,
)
# Disable on blackwell for now
if has_deep_ep() and not current_platform.has_device_capability(100):
from vllm.model_executor.layers.fused_moe.deepep_ht_prepare_finalize import (
DeepEPHTPrepareAndFinalize,
)
from vllm.model_executor.layers.fused_moe.deepep_ll_prepare_finalize import (
DeepEPLLPrepareAndFinalize,
)
register_prepare_and_finalize(
DeepEPHTPrepareAndFinalize,
standard_format,
common_float_types,
blocked_quantization_support=True,
backend="deepep_high_throughput",
)
register_prepare_and_finalize(
DeepEPLLPrepareAndFinalize,
batched_format,
common_float_types,
blocked_quantization_support=True,
backend="deepep_low_latency",
)
if has_pplx():
from vllm.model_executor.layers.fused_moe.pplx_prepare_finalize import (
PplxPrepareAndFinalize,
)
register_prepare_and_finalize(
PplxPrepareAndFinalize,
batched_format,
common_float_and_int_types,
blocked_quantization_support=True,
backend="pplx",
)
if has_flashinfer_cutlass_fused_moe() and current_platform.has_device_capability(100):
from vllm.model_executor.layers.fused_moe.flashinfer_cutlass_moe import (
FlashInferExperts,
)
from vllm.model_executor.layers.fused_moe.flashinfer_cutlass_prepare_finalize import ( # noqa: E501
FlashInferCutlassMoEPrepareAndFinalize,
create_flashinfer_prepare_finalize,
)
register_prepare_and_finalize(
FlashInferCutlassMoEPrepareAndFinalize,
standard_format,
nvfp4_types + fp8_types,
blocked_quantization_support=True,
backend=None,
force_multigpu=True,
supports_apply_weight_on_input=False,
)
register_experts(
FlashInferExperts,
standard_format,
nvfp4_types + fp8_types,
blocked_quantization_support=True,
supports_chunking=True,
# Note: this is a hack to get it to run for now
supports_expert_map=True,
)
else:
FlashInferCutlassMoEPrepareAndFinalize = None
if has_deep_gemm() and is_deep_gemm_supported():
register_experts(
BatchedDeepGemmExperts,
batched_format,
fp8_types,
blocked_quantization_support=True,
supports_chunking=False,
supports_expert_map=False,
needs_matching_quant=False,
needs_deep_gemm=True,
)
register_experts(
DeepGemmExperts,
standard_format,
fp8_types,
blocked_quantization_support=True,
supports_chunking=True,
supports_expert_map=True,
needs_matching_quant=False,
needs_deep_gemm=True,
)
register_experts(
TritonOrDeepGemmExperts,
standard_format,
common_float_and_int_types,
blocked_quantization_support=True,
supports_chunking=True,
supports_expert_map=True,
needs_matching_quant=True,
needs_deep_gemm=True,
)
if cutlass_fp8_supported():
from vllm.model_executor.layers.fused_moe import (
CutlassBatchedExpertsFp8,
CutlassExpertsFp8,
)
register_experts(
CutlassExpertsFp8,
standard_format,
fp8_types,
blocked_quantization_support=False,
supports_chunking=True,
supports_expert_map=False,
)
register_experts(
CutlassBatchedExpertsFp8,
batched_format,
fp8_types,
blocked_quantization_support=False,
supports_chunking=False,
supports_expert_map=False,
)
if cutlass_fp4_supported():
from vllm.model_executor.layers.fused_moe.cutlass_moe import CutlassExpertsFp4
register_experts(
CutlassExpertsFp4,
standard_format,
nvfp4_types,
blocked_quantization_support=True,
supports_chunking=True,
supports_expert_map=False,
)
MK_QUANT_CONFIGS: list[TestMoEQuantConfig | None] = [
None,
# per-channel / per-column weights and per-tensor activations
TestMoEQuantConfig(
quant_dtype=torch.float8_e4m3fn,
per_out_ch_quant=True,
per_act_token_quant=False,
block_shape=None,
),
# per-channel / per-column weights and per-token activations
TestMoEQuantConfig(
quant_dtype=torch.float8_e4m3fn,
per_out_ch_quant=True,
per_act_token_quant=True,
block_shape=None,
),
# per-tensor weights and per-tensor activations
TestMoEQuantConfig(
quant_dtype=torch.float8_e4m3fn,
per_out_ch_quant=False,
per_act_token_quant=False,
block_shape=None,
),
# per-tensor weights and per-token activations
TestMoEQuantConfig(
quant_dtype=torch.float8_e4m3fn,
per_out_ch_quant=False,
per_act_token_quant=True,
block_shape=None,
),
# block-quantized weights and 128 block per-token activations
TestMoEQuantConfig(
quant_dtype=torch.float8_e4m3fn,
per_out_ch_quant=False,
per_act_token_quant=False,
block_shape=[128, 128],
),
# TODO (varun) : Should we test the following combinations ?
# block-quantized weights and per-token activations
# block-quantized weights and per-tensor activations
]
if cutlass_fp4_supported() or has_flashinfer_cutlass_fused_moe():
MK_QUANT_CONFIGS += [
TestMoEQuantConfig(
quant_dtype="nvfp4",
per_out_ch_quant=False,
per_act_token_quant=False,
block_shape=None,
),
]
def make_prepare_finalize(
prepare_finalize_type: mk.FusedMoEPrepareAndFinalize,
backend: str | None,
moe: FusedMoEConfig,
quant_config: FusedMoEQuantConfig,
) -> mk.FusedMoEPrepareAndFinalize:
if backend != "naive" and backend is not None:
prepare_finalize = maybe_make_prepare_finalize(moe, quant_config)
assert prepare_finalize is not None
return prepare_finalize
elif prepare_finalize_type == FlashInferCutlassMoEPrepareAndFinalize:
return create_flashinfer_prepare_finalize(
use_dp=moe.moe_parallel_config.dp_size > 1
)
else:
return MoEPrepareAndFinalizeNoEP()
def _slice(rank: int, num_local_experts: int, t: torch.Tensor) -> torch.Tensor:
s = rank * num_local_experts
e = s + num_local_experts
return t[s:e]
def make_cutlass_strides(
e: int,
n: int,
k: int,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
ab_strides1 = torch.full((e,), k, device="cuda", dtype=torch.int64)
ab_strides2 = torch.full((e,), n, device="cuda", dtype=torch.int64)
c_strides1 = torch.full((e,), 2 * n, device="cuda", dtype=torch.int64)
c_strides2 = torch.full((e,), k, device="cuda", dtype=torch.int64)
return ab_strides1, ab_strides2, c_strides1, c_strides2
def make_fused_experts(
fused_experts_type: mk.FusedMoEPermuteExpertsUnpermute,
moe: FusedMoEConfig,
quant_config: FusedMoEQuantConfig,
num_dispatchers: int,
N: int,
) -> mk.FusedMoEPermuteExpertsUnpermute:
batch_kwargs = {
"max_num_tokens": moe.max_num_tokens,
"num_dispatchers": num_dispatchers,
}
quant_kwargs = {
"quant_config": quant_config,
}
deepgemm_kwargs = {"allow_deep_gemm": has_deep_gemm()}
torch.set_printoptions(threshold=0, edgeitems=0, linewidth=10000)
if fused_experts_type == BatchedDeepGemmExperts:
kwargs = batch_kwargs | quant_kwargs
print(f"Making BatchedDeepGemmExperts {kwargs} ...")
experts = BatchedDeepGemmExperts(**kwargs)
elif fused_experts_type == BatchedTritonExperts:
kwargs = batch_kwargs | quant_kwargs
print(f"Making BatchedTritonExperts {kwargs} ...")
experts = BatchedTritonExperts(**kwargs)
elif fused_experts_type == DeepGemmExperts:
print(f"Making DeepGemmExperts {quant_config} ...")
experts = DeepGemmExperts(quant_config)
elif fused_experts_type == TritonExperts:
kwargs = quant_kwargs
print(f"Making TritonExperts {kwargs} ...")
experts = TritonExperts(**kwargs)
elif fused_experts_type == TritonOrDeepGemmExperts:
kwargs = quant_kwargs | deepgemm_kwargs
print(f"Making TritonOrDeepGemmExperts {kwargs} ...")
experts = TritonOrDeepGemmExperts(**kwargs)
elif fused_experts_type == NaiveBatchedExperts:
kwargs = batch_kwargs | quant_kwargs
print(f"Making NaiveBatchedExperts {kwargs} ...")
experts = NaiveBatchedExperts(**kwargs)
elif fused_experts_type == CutlassExpertsFp8:
strides = make_cutlass_strides(moe.num_experts, N, moe.hidden_dim)
kwargs = {
"out_dtype": moe.in_dtype,
"ab_strides1": strides[0],
"ab_strides2": strides[1],
"c_strides1": strides[2],
"c_strides2": strides[3],
} | quant_kwargs
print(f"Making CutlassExpertsFp8 {kwargs} ...")
experts = CutlassExpertsFp8(**kwargs)
elif fused_experts_type == CutlassBatchedExpertsFp8:
strides = make_cutlass_strides(moe.num_experts, N, moe.hidden_dim)
kwargs = {
"max_experts_per_worker": moe.num_local_experts,
"num_dispatchers": num_dispatchers,
"out_dtype": moe.in_dtype,
"ab_strides1": strides[0],
"ab_strides2": strides[1],
"c_strides1": strides[2],
"c_strides2": strides[3],
} | quant_kwargs
print(f"Making CutlassBatchedExpertsFp8 {kwargs} ...")
experts = CutlassBatchedExpertsFp8(**kwargs)
elif fused_experts_type == CutlassExpertsFp4:
kwargs = {
"max_experts_per_worker": moe.num_local_experts,
"num_dispatchers": num_dispatchers,
"out_dtype": moe.in_dtype,
} | quant_kwargs
print(f"Making CutlassExpertsFp4 {kwargs} ...")
experts = CutlassExpertsFp4(**kwargs)
elif fused_experts_type == FlashInferExperts:
kwargs = {
"out_dtype": moe.in_dtype,
"ep_rank": moe.ep_rank,
"ep_size": moe.ep_size,
"tp_rank": moe.tp_rank,
"tp_size": moe.tp_size,
} | quant_kwargs
print(f"Making FlashInferExperts {kwargs} ...")
experts = FlashInferExperts(**kwargs)
else:
raise RuntimeError(f"Unknown fused experts type: {fused_experts_type}")
torch.set_printoptions(threshold=1000, edgeitems=5, linewidth=80)
return experts

View File

@@ -0,0 +1,134 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import dataclasses
import os
import traceback
from collections.abc import Callable
from typing import Any, Concatenate
import torch
from torch.multiprocessing import spawn # pyright: ignore[reportPrivateImportUsage]
from typing_extensions import ParamSpec
from vllm.config import VllmConfig, set_current_vllm_config
from vllm.distributed import init_distributed_environment, initialize_model_parallel
from vllm.utils.network_utils import get_open_port
## Parallel Processes Utils
P = ParamSpec("P")
@dataclasses.dataclass
class ProcessGroupInfo:
world_size: int
world_local_size: int
rank: int
node_rank: int
local_rank: int
device: torch.device
def _set_vllm_config(
vllm_config: VllmConfig, world_size: int, rank: int, local_rank: int
):
import tempfile
temp_file = tempfile.mkstemp()[1]
with set_current_vllm_config(vllm_config):
init_distributed_environment(
world_size=world_size,
rank=rank,
distributed_init_method=f"file://{temp_file}",
local_rank=local_rank,
backend="nccl",
)
initialize_model_parallel(
tensor_model_parallel_size=vllm_config.parallel_config.tensor_parallel_size,
pipeline_model_parallel_size=vllm_config.parallel_config.pipeline_parallel_size,
)
cpu_group = torch.distributed.new_group(list(range(world_size)), backend="gloo")
return cpu_group
def _worker_parallel_launch(
local_rank: int,
world_size: int,
world_local_size: int,
node_rank: int,
init_method: str,
worker: Callable[Concatenate[ProcessGroupInfo, VllmConfig | None, Any, P], None],
vllm_config: VllmConfig | None,
env_dict: dict | None,
*args: P.args,
**kwargs: P.kwargs,
) -> None:
rank = node_rank * world_local_size + local_rank
torch.cuda.set_device(local_rank)
device = torch.device("cuda", local_rank)
torch.distributed.init_process_group(
backend="cpu:gloo,cuda:nccl",
init_method=init_method,
rank=rank,
world_size=world_size,
device_id=device,
)
barrier = torch.tensor([rank], device=device)
torch.distributed.all_reduce(barrier)
if env_dict is not None:
os.environ.update(env_dict)
cpu_group = None
if vllm_config is not None:
cpu_group = _set_vllm_config(vllm_config, world_size, rank, local_rank)
try:
worker(
ProcessGroupInfo(
world_size=world_size,
world_local_size=world_local_size,
rank=rank,
node_rank=node_rank,
local_rank=local_rank,
device=device,
),
vllm_config,
cpu_group,
*args,
**kwargs,
)
except Exception as ex:
print(ex)
traceback.print_exc()
raise
finally:
torch.distributed.destroy_process_group()
def parallel_launch_with_config(
world_size: int,
worker: Callable[Concatenate[ProcessGroupInfo, VllmConfig, Any, P], None],
vllm_config: VllmConfig,
env_dict: dict[Any, Any],
*args: P.args,
**kwargs: P.kwargs,
) -> None:
assert not kwargs
spawn(
_worker_parallel_launch,
args=(
world_size,
world_size,
0,
f"tcp://{os.getenv('LOCALHOST', 'localhost')}:{get_open_port()}",
worker,
vllm_config,
env_dict,
)
+ args,
nprocs=world_size,
join=True,
)

View File

@@ -0,0 +1,137 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import copy
from collections.abc import Callable
from itertools import product
from typing import Any
import torch
from vllm.config import VllmConfig
from vllm.platforms import current_platform
from .common import Config, RankTensors, WeightTensors, make_modular_kernel
from .parallel_utils import ProcessGroupInfo, parallel_launch_with_config
def do_profile(
fn: Callable,
fn_kwargs: dict[Any, Any],
pgi: ProcessGroupInfo,
config: Config,
num_warmups: int = 5,
):
for _ in range(num_warmups):
fn(**fn_kwargs)
with torch.profiler.profile(
activities=[
torch.profiler.ProfilerActivity.CPU,
torch.profiler.ProfilerActivity.CUDA,
],
with_stack=True,
record_shapes=True,
) as tprof:
fn(**fn_kwargs)
torch.cuda.synchronize(torch.cuda.current_device())
# TODO (varun): Add a descriptive trace file name
tprof.export_chrome_trace(
f"{config.torch_trace_dir_path}/m{config.M}_{pgi.rank}_trace.json"
)
def profile_modular_kernel(
pgi: ProcessGroupInfo,
vllm_config: VllmConfig,
config: Config,
weights: WeightTensors,
rank_tensors: RankTensors,
) -> None:
assert isinstance(config.Ms, int)
assert isinstance(config.topks, int)
# weights for rank
rank_weights = weights.slice_weights(pgi.rank, config.num_local_experts)
# make modular kernel
mk = make_modular_kernel(config, vllm_config, weights)
mk_kwargs = {
"hidden_states": rank_tensors.hidden_states,
"w1": rank_weights.w1,
"w2": rank_weights.w2,
"topk_weights": rank_tensors.topk_weights,
"topk_ids": rank_tensors.topk_ids,
"expert_map": rank_tensors.expert_map,
"w1_scale": rank_weights.w1_scale,
"w2_scale": rank_weights.w2_scale,
"a1_scale": rank_tensors.hidden_states_scale,
"global_num_experts": config.E,
"apply_router_weight_on_input": config.topk == 1,
}
do_profile(mk.forward, mk_kwargs, pgi, config)
def rank_worker(
pgi: ProcessGroupInfo,
vllm_config: VllmConfig,
cpu_group,
config: Config,
weights: WeightTensors,
):
current_platform.seed_everything(pgi.rank)
# sanity check
from vllm import envs
if config.fused_moe_chunk_size is not None:
assert config.fused_moe_chunk_size == envs.VLLM_FUSED_MOE_CHUNK_SIZE
# get weights to this device
weights.to_current_device()
Ms = config.Ms
assert isinstance(Ms, list)
TOPKs = config.topks
assert isinstance(TOPKs, list)
for m, topk in product(Ms, TOPKs):
print(f"Running m={m}, topk={topk} ...")
# override m and topk
cfgx = copy.deepcopy(config)
cfgx.Ms = m
cfgx.topks = topk
# inputs for rank
rank_tensors = RankTensors.make(cfgx, pgi)
profile_modular_kernel(pgi, vllm_config, cfgx, weights, rank_tensors)
def run(config: Config):
weights: WeightTensors = WeightTensors.make(config)
vllm_config, env_dict = config.make_env_data()
parallel_launch_with_config(
config.world_size, rank_worker, vllm_config, env_dict, config, weights
)
if __name__ == "__main__":
from .cli_args import make_config, make_config_arg_parser
parser = make_config_arg_parser(
description=(
"Run single prepare-finalize & fused-experts combination test"
"Example : python3 -m tests.kernels.moe.modular_kernel_tools.profile_modular_kernel " # noqa: E501
"--pf-type PplxPrepareAndFinalize --experts-type BatchedTritonExperts"
)
)
args = parser.parse_args()
assert args.torch_trace_dir_path is not None, (
"Please pass in a directory to store torch traces"
)
config = make_config(args)
run(config)