From 9e289c553c8c71a35d46547da9fca762f4382998 Mon Sep 17 00:00:00 2001 From: Dan Blanaru <48605845+DanBlanaru@users.noreply.github.com> Date: Tue, 14 Jul 2026 04:58:16 +0200 Subject: [PATCH] up FI fp8 moe topk to 32 (#44462) --- vllm/model_executor/layers/fused_moe/experts/trtllm_fp8_moe.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vllm/model_executor/layers/fused_moe/experts/trtllm_fp8_moe.py b/vllm/model_executor/layers/fused_moe/experts/trtllm_fp8_moe.py index 2fb7205918..edcea5361e 100644 --- a/vllm/model_executor/layers/fused_moe/experts/trtllm_fp8_moe.py +++ b/vllm/model_executor/layers/fused_moe/experts/trtllm_fp8_moe.py @@ -393,7 +393,7 @@ class TrtLlmFp8ExpertsMonolithic(TrtLlmFp8ExpertsBase, mk.FusedMoEExpertsMonolit n_group = num_expert_group or None selected_topk_group = topk_group or None else: - assert self.topk <= 10 + assert self.topk <= 32 fp8_quant_type = Fp8QuantizationType.DeepSeekFp8 use_shuffled_weight = True weight_layout = WeightLayout.BlockMajorK