[Bugfix] Fix DeepGemm E8M0 accuracy degradation for Qwen3.5 FP8 on Blackwell (#38083)

Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
This commit is contained in:
Vadim Gimpelson
2026-03-26 01:21:47 -07:00
committed by GitHub
parent 71161e8b63
commit 52069012fe
10 changed files with 69 additions and 11 deletions
@@ -359,10 +359,14 @@ class W8A8BlockFp8LinearOp:
act_quant_group_shape: GroupShape,
cutlass_block_fp8_supported: bool = CUTLASS_BLOCK_FP8_SUPPORTED,
use_aiter_and_is_supported: bool = False,
use_deep_gemm: bool | None = None,
):
self.weight_group_shape = weight_group_shape
self.act_quant_group_shape = act_quant_group_shape
self.is_deep_gemm_supported = is_deep_gemm_supported()
if use_deep_gemm is not None:
self.is_deep_gemm_supported = use_deep_gemm
else:
self.is_deep_gemm_supported = is_deep_gemm_supported()
self.is_hopper = current_platform.is_device_capability(90)
self.use_deep_gemm_e8m0 = is_deep_gemm_e8m0_used()
self.is_flashinfer_supported = is_flashinfer_fp8_blockscale_gemm_supported()