up FI fp8 moe topk to 32 (#44462)
Dan Blanaru committed
Jul 14, 2026 at 04:58 UTC
9e289c553c8c71a35d46547da9fca762f4382998
1 file changed
+1
-1
vllm/model_executor/layers/fused_moe/experts/trtllm_fp8_moe.py
+1
-1
@@ -393,7 +393,7 @@ class TrtLlmFp8ExpertsMonolithic(TrtLlmFp8ExpertsBase, mk.FusedMoEExpertsMonolit
393
n_group = num_expert_group or None
394
selected_topk_group = topk_group or None
395
else:
396
- assert self.topk <= 10
396
+ assert self.topk <= 32
397
fp8_quant_type = Fp8QuantizationType.DeepSeekFp8
398
use_shuffled_weight = True
399
weight_layout = WeightLayout.BlockMajorK