[Performance][B200] Fix deepgemm prologue (#27897)

Signed-off-by: Varun Sundar Rabindranath <vsundarr@redhat.com>
Co-authored-by: Varun Sundar Rabindranath <vsundarr@redhat.com>
This commit is contained in:
Varun Sundar Rabindranath
2025-11-12 13:13:03 -08:00
committed by GitHub
co-authored by Varun Sundar Rabindranath
parent 478ee511de
commit 74a9a9faad
6 changed files with 163 additions and 48 deletions
+15 -4
View File
@@ -49,10 +49,6 @@ def is_deep_gemm_e8m0_used() -> bool:
logger.info_once("DeepGEMM E8M0 disabled: _fp8_gemm_nt_impl not found")
return False
if envs.VLLM_USE_FLASHINFER_MOE_FP8:
logger.info_once("DeepGEMM E8M0 disabled: FlashInfer MOE is enabled.")
return False
if envs.VLLM_USE_DEEP_GEMM_E8M0:
logger.info_once("DeepGEMM E8M0 enabled on current platform.")
return True
@@ -77,6 +73,7 @@ _fp8_paged_mqa_logits_impl: Callable[..., Any] | None = None
_get_paged_mqa_logits_metadata_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
_transform_sf_into_required_layout_impl: Callable[..., Any] | None = None
def _lazy_init() -> None:
@@ -86,6 +83,7 @@ def _lazy_init() -> None:
global _get_paged_mqa_logits_metadata_impl
global _get_mn_major_tma_aligned_tensor_impl
global _get_mk_alignment_for_contiguous_layout_impl
global _transform_sf_into_required_layout_impl
# fast path
if (
_fp8_gemm_nt_impl is not None
@@ -95,6 +93,7 @@ def _lazy_init() -> None:
or _fp8_paged_mqa_logits_impl is not None
or _get_paged_mqa_logits_metadata_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
):
return
@@ -124,6 +123,9 @@ def _lazy_init() -> None:
_get_mk_alignment_for_contiguous_layout_impl = getattr(
_dg, "get_mk_alignment_for_contiguous_layout", None
)
_transform_sf_into_required_layout_impl = getattr(
_dg, "transform_sf_into_required_layout", None
)
def get_num_sms() -> int:
@@ -179,6 +181,15 @@ def fp8_m_grouped_gemm_nt_masked(*args, **kwargs):
)
def transform_sf_into_required_layout(*args, **kwargs):
_lazy_init()
if _transform_sf_into_required_layout_impl is None:
return _missing(*args, **kwargs)
return _transform_sf_into_required_layout_impl(
*args, disable_ue8m0_cast=not is_deep_gemm_e8m0_used(), **kwargs
)
def fp8_mqa_logits(
q: torch.Tensor,
kv: tuple[torch.Tensor, torch.Tensor],