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 DeepSeek-V2.5-1210-FP8 issue (#27267)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Oct 22, 2025 at 11:00 UTC
1c160841ea5f70bae61e62ab8eea17141d654e29
1 file changed
+11
-5
vllm/model_executor/layers/fused_moe/cutlass_moe.py
+11
-5
@@ -511,13 +511,19 @@ def cutlass_moe_fp8(
511
assert quant_config is not None
512
513
if quant_config.a1_scale is not None:
514
- assert quant_config.per_act_token_quant == quant_config.a1_scale.numel() != 1
514
+ assert quant_config.per_act_token_quant == (quant_config.a1_scale.numel() != 1)
515
if quant_config.a2_scale is not None:
516
- assert quant_config.per_act_token_quant == quant_config.a2_scale.numel() != 1
516
+ assert quant_config.per_act_token_quant == (quant_config.a2_scale.numel() != 1)
517
518
- assert quant_config.w1_scale is None or (
519
- quant_config.per_out_ch_quant == (quant_config.w1_scale.size(1) == w1_q.size(1))
520
- )
518
+ if quant_config.w1_scale is not None:
519
+ if quant_config.per_out_ch_quant:
520
+ assert quant_config.w1_scale.dim() > 1 and quant_config.w1_scale.size(
521
+ 1
522
+ ) == w1_q.size(1)
523
+ else:
524
+ assert (
525
+ quant_config.w1_scale.dim() == 1 or quant_config.w1_scale.size(1) == 1
526
+ )
527
528
num_experts = global_num_experts if global_num_experts != -1 else w1_q.size(0)
529