From 83a7c89c3f13e4d5f1632b1e31802cfdd1f1bc28 Mon Sep 17 00:00:00 2001 From: b8zhong Date: Tue, 28 Oct 2025 22:17:08 -0700 Subject: [PATCH] followup fix for llama 4 trtllm flashinfer backend (#12314) --- python/sglang/srt/layers/quantization/modelopt_quant.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/python/sglang/srt/layers/quantization/modelopt_quant.py b/python/sglang/srt/layers/quantization/modelopt_quant.py index e32f4fedd..40fd2bb5b 100755 --- a/python/sglang/srt/layers/quantization/modelopt_quant.py +++ b/python/sglang/srt/layers/quantization/modelopt_quant.py @@ -673,7 +673,7 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase): routed_scaling_factor if routed_scaling_factor is not None else 1.0 ), use_routing_scales_on_input=use_routing_scales_on_input, - tile_tokens_dim=None, + tile_tokens_dim=8, # TODO(brayden): use the FI tile calculation routing_method_type=routing_method_type, )