[Perf][Kernel] Integrate Mega-mHC from DeepGEMM for DeepSeek V4.1 (reopen of #56255) (#56962)

Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
Signed-off-by: Juntian Liu <Juntianl777@gmail.com>
Co-authored-by: Yongye Zhu <zyy1102000@gmail.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Thien Tran <gau.nernst@yahoo.com.sg>
This commit is contained in:
Juntian Liu
2026-09-15 04:01:06 -07:00
committed by GitHub
co-authored by Yongye Zhu Claude Opus 5 Thien Tran
parent 5be911ee87
commit bb55077411
6 changed files with 316 additions and 41 deletions
+6 -1
View File
@@ -27,6 +27,7 @@ from vllm.platforms import current_platform
from vllm.utils.deep_gemm import (
fp8_gemm_nt,
get_tma_aligned_size,
is_deep_gemm_e8m0_used,
per_block_cast_to_fp8,
should_use_deepgemm_for_fp8_linear,
)
@@ -291,7 +292,11 @@ def test_w8a8_block_fp8_deep_gemm_matmul(M, N, K, block_size, out_dtype, seed):
A_fp8, As_fp8 = per_token_group_quant_fp8(
A_fp32, block_size[1], column_major_scales=True, tma_aligned_scales=True
)
B_fp8, Bs_fp8 = per_block_cast_to_fp8(B_fp32, block_size=block_size)
B_fp8, Bs_fp8 = per_block_cast_to_fp8(
B_fp32,
block_size=block_size,
use_ue8m0=is_deep_gemm_e8m0_used(),
)
As = As_fp8.to(torch.float32)
Bs = Bs_fp8.to(torch.float32)
+104 -1
View File
@@ -11,6 +11,7 @@ import vllm.model_executor.kernels.mhc # noqa: F401
import vllm.model_executor.layers.mhc as mhc_layers
from vllm.model_executor.kernels.mhc.tilelang import (
_torch_hc_prenorm_gemm,
mhc_post_tilelang,
mhc_pre_delayed_tilelang,
)
from vllm.model_executor.kernels.mhc.tilelang_kernels import (
@@ -39,6 +40,10 @@ from vllm.models.deepseek_v4.nvidia.model import (
from vllm.models.deepseek_v41.nvidia.model import (
DeepseekV4DecoderLayer as DeepseekV41DecoderLayer,
)
from vllm.models.deepseek_v41.nvidia.ops.mega_mhc import (
is_mega_mhc_supported,
mhc_shifted_post_pre_deep_gemm,
)
from vllm.platforms import current_platform
from vllm.utils.torch_utils import set_random_seed
@@ -495,7 +500,7 @@ def test_deepseek_v41_decoder_mixes_match_torch(
reference,
)
monkeypatch.setattr(
"vllm.models.deepseek_v41.nvidia.model.mhc_fused_post_pre_delayed_tilelang",
"vllm.models.deepseek_v41.nvidia.model.mhc_shifted_post_pre",
fused_reference,
)
expected = decoder(x, positions, None, **kwargs)
@@ -580,6 +585,104 @@ def test_mhc_pre_delayed_custom_op_supports_compile(carried):
)
@pytest.mark.skipif(not HAS_TILELANG_MHC, reason="TileLang MHC support required")
@pytest.mark.skipif(
not current_platform.is_device_capability_family(100),
reason="DeepGEMM Mega mHC requires SM100-family CUDA",
)
@pytest.mark.parametrize("num_tokens", [0, 1, 17, 128, 1024])
@pytest.mark.parametrize("hidden_size", [5120, 7168])
def test_deep_gemm_mega_mhc_correctness(num_tokens, hidden_size):
set_random_seed(0)
hc_mult = 4
if not is_mega_mhc_supported(hidden_size, 4):
pytest.skip("DeepGEMM Mega mHC is not available")
x = torch.randn(num_tokens, hidden_size, dtype=torch.bfloat16, device=DEVICE)
residual = torch.randn(
num_tokens, hc_mult, hidden_size, dtype=torch.bfloat16, device=DEVICE
)
previous_mix = torch.rand(num_tokens, hc_mult, device=DEVICE)
post_mix = torch.rand(num_tokens, hc_mult, 1, device=DEVICE)
res_mix = torch.rand(num_tokens, hc_mult, hc_mult, device=DEVICE)
fn = torch.randn(24, hc_mult * hidden_size, device=DEVICE) * 0.01
scale = torch.randn(3, device=DEVICE) * 0.1
base = torch.randn(24, device=DEVICE) * 0.1
norm_weight = torch.empty(
hidden_size, dtype=torch.bfloat16, device=DEVICE
).uniform_(0.9, 1.1)
expected_residual = (
mhc_post_tilelang(x, residual, post_mix, res_mix) if num_tokens else residual
)
(
expected_post_mix,
expected_res_mix,
expected_y_bf16,
expected_previous_mix,
) = mhc_pre_delayed_tilelang(
expected_residual,
fn,
scale,
base,
2e-5,
3e-4,
2e-6,
1.25,
10,
pre_mix=previous_mix,
norm_weight=norm_weight,
norm_eps=7e-6,
)
args = (
x,
residual,
previous_mix,
post_mix,
res_mix,
fn,
scale,
base,
2e-5,
3e-4,
1.25,
2e-6,
10,
norm_weight,
7e-6,
)
actual = mhc_shifted_post_pre_deep_gemm(*args)
expected = (
expected_residual,
expected_post_mix,
expected_res_mix,
expected_y_bf16,
expected_previous_mix,
)
for i, (result, ref) in enumerate(zip(actual, expected, strict=True)):
# Match the existing mHC BF16 tolerance: post rounding differences
# can be amplified by cancellation in the carried collapse.
atol, rtol = (1.6e-2, 1e-2) if i in (0, 3) else (1e-6, 1e-3)
torch.testing.assert_close(result, ref, atol=atol, rtol=rtol)
if num_tokens:
stream = torch.cuda.Stream()
stream.wait_stream(torch.cuda.current_stream())
with torch.cuda.stream(stream):
# DeepGEMM initializes barriers per stream before capture.
mhc_shifted_post_pre_deep_gemm(*args)
graph = torch.cuda.CUDAGraph()
with torch.cuda.graph(graph, stream=stream):
captured = mhc_shifted_post_pre_deep_gemm(*args)
for multiplier in (0.5, -1.0):
x.mul_(multiplier)
graph.replay()
eager = mhc_shifted_post_pre_deep_gemm(*args)
for result, ref in zip(captured, eager, strict=True):
torch.testing.assert_close(result, ref, atol=0, rtol=0)
torch.cuda.current_stream().wait_stream(stream)
def sinkhorn_normalize_ref(x: torch.Tensor, repeat: int, eps: float) -> torch.Tensor:
x = x.softmax(-1) + eps
x = x / (x.sum(-2, keepdim=True) + eps)
+34 -38
View File
@@ -20,7 +20,6 @@ from vllm.distributed import (
from vllm.forward_context import get_forward_context, is_forward_context_available
from vllm.logger import init_logger
from vllm.model_executor.kernels.mhc.tilelang import (
mhc_fused_post_pre_delayed_tilelang,
mhc_post_tilelang,
mhc_pre_delayed_tilelang,
)
@@ -79,6 +78,7 @@ from vllm.v1.worker.ubatching import dbo_current_ubatch_id
from ..common.engram import EngramLayout, NgramHashState
from ..common.mm_preprocess import IMAGE_SENTINEL_BASE_ID, image_sentinel_mask
from .engram import Engram, gather_engram_hashes
from .ops.mega_mhc import mhc_shifted_post_pre
if typing.TYPE_CHECKING:
from vllm.v1.attention.backends.mla.sparse_swa import DeepseekSparseSWAMetadata
@@ -403,25 +403,23 @@ class DeepseekV4DecoderLayer(nn.Module):
else:
# The collapse already reads the post-mapped streams, so the mean
# aux consumers want comes out of the same kernel.
residual, post_mix, res_mix, x, attn_pre, aux = (
mhc_fused_post_pre_delayed_tilelang(
x,
residual,
post_mix,
res_mix,
self.hc_attn_fn,
self.hc_attn_scale,
self.hc_attn_base,
self.rms_norm_eps,
self.hc_eps,
self.hc_eps,
self.hc_post_alpha,
self.hc_sinkhorn_iters,
pre_mix=pre_mix,
norm_weight=self.attn_norm.weight,
norm_eps=self.attn_norm.variance_epsilon,
capture_aux=capture_previous_aux,
)
residual, post_mix, res_mix, x, attn_pre, aux = mhc_shifted_post_pre(
x,
residual,
post_mix,
res_mix,
self.hc_attn_fn,
self.hc_attn_scale,
self.hc_attn_base,
self.rms_norm_eps,
self.hc_eps,
self.hc_eps,
self.hc_post_alpha,
self.hc_sinkhorn_iters,
pre_mix=pre_mix,
norm_weight=self.attn_norm.weight,
norm_eps=self.attn_norm.variance_epsilon,
capture_aux=capture_previous_aux,
)
if capture_previous_aux:
previous_aux = aux
@@ -433,24 +431,22 @@ class DeepseekV4DecoderLayer(nn.Module):
if self.use_sequence_parallel:
x = sp_reduce_scatter(x)
residual, post_mix, res_mix, x, ffn_pre, _ = (
mhc_fused_post_pre_delayed_tilelang(
x,
residual,
post_mix,
res_mix,
self.hc_ffn_fn,
self.hc_ffn_scale,
self.hc_ffn_base,
self.rms_norm_eps,
self.hc_eps,
self.hc_eps,
self.hc_post_alpha,
self.hc_sinkhorn_iters,
pre_mix=attn_pre,
norm_weight=self.ffn_norm.weight,
norm_eps=self.ffn_norm.variance_epsilon,
)
residual, post_mix, res_mix, x, ffn_pre, _ = mhc_shifted_post_pre(
x,
residual,
post_mix,
res_mix,
self.hc_ffn_fn,
self.hc_ffn_scale,
self.hc_ffn_base,
self.rms_norm_eps,
self.hc_eps,
self.hc_eps,
self.hc_post_alpha,
self.hc_sinkhorn_iters,
pre_mix=attn_pre,
norm_weight=self.ffn_norm.weight,
norm_eps=self.ffn_norm.variance_epsilon,
)
x = self.ffn(x, input_ids)
return x, residual, post_mix, res_mix, ffn_pre, previous_aux
@@ -0,0 +1,2 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
@@ -0,0 +1,156 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import functools
import torch
from vllm.model_executor.kernels.mhc.tilelang import (
mhc_fused_post_pre_delayed_tilelang,
)
from vllm.platforms import current_platform
from vllm.utils.deep_gemm import (
_import_deep_gemm,
is_deep_gemm_supported,
mega_mhc,
)
@functools.cache
def is_mega_mhc_supported(hidden_size: int, hc_mult: int) -> bool:
"""Return whether DeepGEMM can run the DSV4.1 shifted mHC kernel."""
if not (
is_deep_gemm_supported()
and current_platform.is_device_capability_family(100)
and hidden_size > 0
and hidden_size % 1024 == 0
and hc_mult == 4
):
return False
deep_gemm = _import_deep_gemm()
return deep_gemm is not None and callable(getattr(deep_gemm, "mega_mhc", None))
def mhc_shifted_post_pre_deep_gemm(
x: torch.Tensor,
residual: torch.Tensor,
shifted_prev_mix: torch.Tensor,
post_mix: torch.Tensor,
comb_res_mix: torch.Tensor,
fn: torch.Tensor,
mix_scales: torch.Tensor,
mix_bases: torch.Tensor,
hc_norm_eps: float,
hc_pre_eps: float,
hc_post_scale: float,
sinkhorn_eps: float,
num_sinkhorn_iters: int,
rmsnorm_weight: torch.Tensor,
rmsnorm_eps: float,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
"""Run DSV4.1 shifted post, next pre, and BF16 RMSNorm with Mega mHC."""
num_tokens, hidden_size = x.shape
hc_mult = residual.shape[1]
new_residual = torch.empty_like(residual)
new_prev_mix = x.new_empty(num_tokens, hc_mult, 1, dtype=torch.float32)
new_post_mix = torch.empty_like(post_mix)
new_comb_res_mix = torch.empty_like(comb_res_mix)
y_bf16 = x.new_empty(num_tokens, hidden_size, dtype=torch.bfloat16)
mega_mhc(
x=x,
residual=residual,
shifted_prev_mix=shifted_prev_mix.unsqueeze(-1),
post_mix=post_mix,
comb_res_mix=comb_res_mix,
fn=fn,
mix_scales=mix_scales,
mix_bases=mix_bases,
hc_mult=hc_mult,
hc_norm_eps=hc_norm_eps,
hc_pre_eps=hc_pre_eps,
hc_post_scale=hc_post_scale,
sinkhorn_eps=sinkhorn_eps,
num_sinkhorn_iters=num_sinkhorn_iters,
rmsnorm_weight=rmsnorm_weight,
rmsnorm_eps=rmsnorm_eps,
rmsnorm_scale=1.0,
new_residual=new_residual,
new_prev_mix=new_prev_mix,
new_post_mix=new_post_mix,
new_comb_res_mix=new_comb_res_mix,
y_bf16=y_bf16,
)
return (
new_residual,
new_post_mix,
new_comb_res_mix,
y_bf16,
new_prev_mix.squeeze(-1),
)
def mhc_shifted_post_pre(
x: torch.Tensor,
residual: torch.Tensor,
post_layer_mix: torch.Tensor,
comb_res_mix: torch.Tensor,
fn: torch.Tensor,
hc_scale: torch.Tensor,
hc_base: torch.Tensor,
rms_eps: float,
hc_pre_eps: float,
hc_sinkhorn_eps: float,
hc_post_mult_value: float,
sinkhorn_repeat: int,
pre_mix: torch.Tensor | None = None,
norm_weight: torch.Tensor | None = None,
norm_eps: float = 1e-6,
capture_aux: bool = False,
) -> tuple[
torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor
]:
"""Use Mega-mHC for carried mixing, retaining TileLang for aux capture."""
if (
pre_mix is not None
and norm_weight is not None
and not capture_aux
and x.shape[0] <= 1 << 20
and is_mega_mhc_supported(x.shape[1], residual.shape[1])
):
outputs = mhc_shifted_post_pre_deep_gemm(
x,
residual,
pre_mix,
post_layer_mix,
comb_res_mix,
fn,
hc_scale,
hc_base,
rms_eps,
hc_pre_eps,
hc_post_mult_value,
hc_sinkhorn_eps,
sinkhorn_repeat,
norm_weight,
norm_eps,
)
return *outputs, x.new_empty(0, x.shape[1])
return mhc_fused_post_pre_delayed_tilelang(
x,
residual,
post_layer_mix,
comb_res_mix,
fn,
hc_scale,
hc_base,
rms_eps,
hc_pre_eps,
hc_sinkhorn_eps,
hc_post_mult_value,
sinkhorn_repeat,
pre_mix=pre_mix,
norm_weight=norm_weight,
norm_eps=norm_eps,
capture_aux=capture_aux,
)
+14 -1
View File
@@ -159,6 +159,7 @@ _fp8_fp4_mqa_logits_impl: Callable[..., Any] | None = None
_fp8_fp4_paged_mqa_logits_impl: Callable[..., Any] | None = None
_get_paged_mqa_logits_metadata_impl: Callable[..., Any] | None = None
_tf32_hc_prenorm_gemm_impl: Callable[..., Any] | None = None
_mega_mhc_impl: Callable[..., Any] | None = None
_get_mn_major_tma_aligned_tensor_impl: Callable[..., Any] | None = None
_get_mk_alignment_for_contiguous_layout_impl: Callable[..., Any] | None = None
_get_theoretical_mk_alignment_for_contiguous_layout_impl: Callable[..., Any] | None = (
@@ -231,7 +232,7 @@ def _lazy_init() -> None:
global _grouped_impl, _grouped_masked_impl, _grouped_fp4_impl
global _fp8_fp4_mqa_logits_impl, _fp8_fp4_paged_mqa_logits_impl
global _get_paged_mqa_logits_metadata_impl
global _tf32_hc_prenorm_gemm_impl
global _tf32_hc_prenorm_gemm_impl, _mega_mhc_impl
global _get_mn_major_tma_aligned_tensor_impl
global _get_mk_alignment_for_contiguous_layout_impl
global _get_theoretical_mk_alignment_for_contiguous_layout_impl
@@ -251,6 +252,7 @@ def _lazy_init() -> None:
or _fp8_fp4_paged_mqa_logits_impl is not None
or _get_paged_mqa_logits_metadata_impl is not None
or _tf32_hc_prenorm_gemm_impl is not None
or _mega_mhc_impl is not None
or _get_mk_alignment_for_contiguous_layout_impl is not None
or _transform_sf_into_required_layout_impl is not None
or _pack_ue8m0_to_int_impl is not None
@@ -294,6 +296,7 @@ def _lazy_init() -> None:
_dg, "get_paged_mqa_logits_metadata", None
)
_tf32_hc_prenorm_gemm_impl = getattr(_dg, "tf32_hc_prenorm_gemm", None)
_mega_mhc_impl = getattr(_dg, "mega_mhc", None)
_get_mn_major_tma_aligned_tensor_impl = getattr(
_dg, "get_mn_major_tma_aligned_tensor", None
)
@@ -704,6 +707,15 @@ def tf32_hc_prenorm_gemm(
)
def mega_mhc(*args: Any, **kwargs: Any) -> None:
"""Run DeepGEMM Mega mHC with caller-owned output tensors."""
_lazy_init()
if _mega_mhc_impl is None:
_missing()
return
_mega_mhc_impl(*args, **kwargs)
def _ceil_to_ue8m0(x: torch.Tensor):
return torch.pow(2.0, torch.ceil(torch.log2(x.abs())))
@@ -797,6 +809,7 @@ __all__ = [
"per_block_cast_to_fp8",
"is_deep_gemm_e8m0_used",
"is_deep_gemm_supported",
"mega_mhc",
"get_num_sms",
"set_num_sms",
"should_use_deepgemm_for_fp8_linear",