[feat]: CUTLASS block scaled group gemm for SM100 (#19757)

Signed-off-by: Duncan Moss <djm.moss@gmail.com>
Co-authored-by: Duncan Moss <dmoss@nvidia.com>
This commit is contained in:
Duncan Moss
2025-07-04 12:58:04 -06:00
committed by GitHub
co-authored by Duncan Moss
parent 2f35a022e6
commit 3d184b95b8
13 changed files with 726 additions and 30 deletions
@@ -7,12 +7,17 @@ import torch
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
from vllm import _custom_ops as ops
from vllm.logger import init_logger
from vllm.model_executor.layers.fused_moe.config import FusedMoEQuantConfig
from vllm.model_executor.layers.fused_moe.prepare_finalize import (
MoEPrepareAndFinalizeNoEP)
from vllm.model_executor.layers.fused_moe.utils import _fp8_perm, _resize_cache
from vllm.model_executor.layers.fused_moe.utils import (_fp8_perm,
_fp8_quantize,
_resize_cache)
from vllm.scalar_type import scalar_types
logger = init_logger(__name__)
def run_cutlass_moe_fp8(
output: torch.Tensor,
@@ -508,3 +513,130 @@ def cutlass_moe_fp4(a: torch.Tensor, a1_gscale: torch.Tensor,
out = (c2.view(m, num_topk, k) *
topk_weights.view(m, num_topk, 1).half()).sum(dim=1)
return out.to(dtype=out_dtype)
def _valid_cutlass_block_scaled_grouped_gemm(hidden_states: torch.Tensor,
w1: torch.Tensor,
w2: torch.Tensor) -> bool:
def _valid_cutlass_block_scaled_grouped_gemm_shape(M: int, N: int, K: int):
return M >= 128 and N % 128 == 0 and K % 128 == 0
m = hidden_states.size(0)
_, K, N = w2.size()
if not _valid_cutlass_block_scaled_grouped_gemm_shape(m, N, K):
logger.debug(
"CutlassBlockScaledGroupedGemm disabled: unalinged problem size.")
return False
if (w1.dtype != torch.float8_e4m3fn or w2.dtype != torch.float8_e4m3fn):
logger.debug(
"CutlassBlockScaledGroupedGemm disabled: invalid weight dtype(s).")
return False
return True
def run_cutlass_block_scaled_fused_experts(
a: torch.Tensor,
w1: torch.Tensor,
w2: torch.Tensor,
w1_scale: torch.Tensor,
w2_scale: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
) -> torch.Tensor:
w1_q = w1.transpose(1, 2)
w2_q = w2.transpose(1, 2)
w1_scale = w1_scale.transpose(1, 2)
w2_scale = w2_scale.transpose(1, 2)
assert topk_weights.shape == topk_ids.shape, "topk shape mismatch"
assert a.shape[0] == topk_ids.shape[
0], "a and topk_ids must have the same batch size"
assert w1_q.dtype == torch.float8_e4m3fn, "w1_q must be float8_e4m3fn"
assert w2_q.dtype == torch.float8_e4m3fn, "w2_q must be float8_e4m3fn"
assert a.shape[1] == w1_q.shape[1], "Hidden size mismatch w1"
assert w1_q.shape[2] == w2_q.shape[1] * 2, "Hidden size mismatch w2"
assert w1_q.shape[0] == w2_q.shape[0], "Expert number mismatch"
assert w1_q.shape[0] == w1_scale.shape[
0], "w1_scale expert number mismatch"
assert w1_q.shape[0] == w2_scale.shape[
0], "w2_scale expert number mismatch"
assert a.dtype in [torch.half, torch.bfloat16], "Invalid output dtype"
out_dtype = a.dtype
num_experts = w1_q.size(0)
m = a.size(0)
k = w1_q.size(1)
n = w2_q.size(1)
expert_offsets = torch.empty((num_experts + 1, ),
dtype=torch.int32,
device="cuda")
problem_sizes1 = torch.empty((num_experts, 3),
dtype=torch.int32,
device="cuda")
problem_sizes2 = torch.empty((num_experts, 3),
dtype=torch.int32,
device="cuda")
topk = topk_ids.size(1)
a_q, a1_scale = _fp8_quantize(a,
A_scale=None,
per_act_token=False,
block_shape=[128, 128])
device = a_q.device
a_map = torch.empty((topk_ids.numel()), dtype=torch.int32, device=device)
c_map = torch.empty((topk_ids.numel()), dtype=torch.int32, device=device)
ops.get_cutlass_moe_mm_data(
topk_ids,
expert_offsets,
problem_sizes1,
problem_sizes2,
a_map,
c_map,
num_experts,
n,
k,
)
rep_a_q = a_q.view(dtype=torch.uint8)[a_map].view(dtype=a_q.dtype)
rep_a1_scales = a1_scale[a_map]
c1 = torch.empty((m * topk, n * 2), dtype=out_dtype, device=device)
c2 = torch.empty((m * topk, k), dtype=out_dtype, device=device)
ops.cutlass_blockwise_scaled_grouped_mm(
c1,
rep_a_q,
w1_q,
rep_a1_scales,
w1_scale,
problem_sizes1,
expert_offsets[:-1],
)
intermediate = torch.empty((m * topk, n), dtype=out_dtype, device=device)
torch.ops._C.silu_and_mul(intermediate, c1)
intermediate_q, a2_scale = _fp8_quantize(intermediate,
A_scale=None,
per_act_token=False,
block_shape=[128, 128])
ops.cutlass_blockwise_scaled_grouped_mm(
c2,
intermediate_q,
w2_q,
a2_scale,
w2_scale,
problem_sizes2,
expert_offsets[:-1],
)
return (c2[c_map].view(m, topk, k) *
topk_weights.view(m, topk, 1).to(out_dtype)).sum(dim=1)
@@ -15,6 +15,9 @@ from vllm.logger import init_logger
# yapf: disable
from vllm.model_executor.layers.fused_moe.config import (
FusedMoEQuantConfig, get_config_quant_dtype)
from vllm.model_executor.layers.fused_moe.cutlass_moe import (
_valid_cutlass_block_scaled_grouped_gemm,
run_cutlass_block_scaled_fused_experts)
# yapf: enable
from vllm.model_executor.layers.fused_moe.deep_gemm_moe import (
_valid_deep_gemm, deep_gemm_moe_fp8)
@@ -1129,29 +1132,31 @@ def dispatch_fused_experts_func(inplace: bool) -> Callable[..., torch.Tensor]:
# TODO (bnell): replace this with modular op. Can get rid of inplace/outplace
# torch ops.
def fused_experts(hidden_states: torch.Tensor,
w1: torch.Tensor,
w2: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
inplace: bool = False,
activation: str = "silu",
apply_router_weight_on_input: bool = False,
use_fp8_w8a8: bool = False,
use_int8_w8a8: bool = False,
use_int8_w8a16: bool = False,
use_int4_w4a16: bool = False,
per_channel_quant: bool = False,
global_num_experts: int = -1,
expert_map: Optional[torch.Tensor] = None,
w1_scale: Optional[torch.Tensor] = None,
w2_scale: Optional[torch.Tensor] = None,
w1_zp: Optional[torch.Tensor] = None,
w2_zp: Optional[torch.Tensor] = None,
a1_scale: Optional[torch.Tensor] = None,
a2_scale: Optional[torch.Tensor] = None,
block_shape: Optional[list[int]] = None,
allow_deep_gemm: bool = False) -> torch.Tensor:
def fused_experts(
hidden_states: torch.Tensor,
w1: torch.Tensor,
w2: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
inplace: bool = False,
activation: str = "silu",
apply_router_weight_on_input: bool = False,
use_fp8_w8a8: bool = False,
use_int8_w8a8: bool = False,
use_int8_w8a16: bool = False,
use_int4_w4a16: bool = False,
per_channel_quant: bool = False,
global_num_experts: int = -1,
expert_map: Optional[torch.Tensor] = None,
w1_scale: Optional[torch.Tensor] = None,
w2_scale: Optional[torch.Tensor] = None,
w1_zp: Optional[torch.Tensor] = None,
w2_zp: Optional[torch.Tensor] = None,
a1_scale: Optional[torch.Tensor] = None,
a2_scale: Optional[torch.Tensor] = None,
block_shape: Optional[list[int]] = None,
allow_deep_gemm: bool = False,
allow_cutlass_block_scaled_grouped_gemm: bool = False) -> torch.Tensor:
# For now, disable DeepGemm for small N (<= 512) until better
# permute/unpermute ops are available.
N = w1.size(1)
@@ -1174,6 +1179,17 @@ def fused_experts(hidden_states: torch.Tensor,
a2_scale=a2_scale,
apply_router_weight_on_input=apply_router_weight_on_input,
)
elif (allow_cutlass_block_scaled_grouped_gemm and use_fp8_w8a8
and _valid_cutlass_block_scaled_grouped_gemm(hidden_states, w1, w2)):
assert apply_router_weight_on_input is False
return run_cutlass_block_scaled_fused_experts(
a=hidden_states,
w1=w1,
w2=w2,
w1_scale=w1_scale,
w2_scale=w2_scale,
topk_weights=topk_weights,
topk_ids=topk_ids)
else:
return dispatch_fused_experts_func(inplace)(
hidden_states=hidden_states,
+19 -1
View File
@@ -473,12 +473,30 @@ class Fp8MoEMethod(FusedMoEMethodBase):
logger.warning_once(
"DeepGemm not supported on the current platform.")
# Check for CutlassBlockScaledGroupedGemm support.
self.allow_cutlass_block_scaled_grouped_gemm = False
if not self.block_quant:
logger.warning_once("Model is not block quantized. Not using "
"CutlassBlockScaledGroupedGemm kernels")
elif (current_platform.is_cuda()
and current_platform.has_device_capability(100)):
logger.info_once(
"Using CutlassBlockScaledGroupedGemm kernels for Fp8MoEMethod."
)
self.allow_cutlass_block_scaled_grouped_gemm = True
else:
logger.warning_once(
"CutlassBlockScaledGroupedGemm not supported on the current "
"platform.")
self.topk_indices_dtype = None
self.fused_experts = functools.partial( # type: ignore
fused_experts,
use_fp8_w8a8=True,
block_shape=self.quant_config.weight_block_size,
allow_deep_gemm=self.allow_deep_gemm)
allow_deep_gemm=self.allow_deep_gemm,
allow_cutlass_block_scaled_grouped_gemm=(
self.allow_cutlass_block_scaled_grouped_gemm))
def create_weights(self, layer: Module, num_experts: int, hidden_size: int,
intermediate_size_per_partition: int,