A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Bug] Fix B200 DeepGEMM E8M0 Accuracy Issue (#22399)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Aug 6, 2025 at 20:03 UTC
eec890c1c1cdf6d4bbf4c0563fac54abe80ab8b6
1 file changed
+2
-1
vllm/model_executor/layers/quantization/utils/fp8_utils.py
+2
-1
@@ -799,7 +799,8 @@ def requant_weight_ue8m0_inplace(
799
s_exp = s_exp[:m_cur, :k_cur]
800
w_dq = w_q.to(torch.float32) * s_exp
801
# Re-quantise using power-of-two scaling (UE8M0).
802
- w_requant, s_requant = per_block_cast_to_fp8(w_dq, [block_m, block_k])
802
+ w_requant, s_requant = per_block_cast_to_fp8(w_dq, [block_m, block_k],
803
+ use_ue8m0=True)
804
805
# Write back the results in-place.
806
w_q.copy_(w_requant)