[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
+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",