fix: Fix DSR1 perf regression due to unnecessarily falling back to triton gemm (#28073)

python/sglang/srt/layers/quantization/fp8_utils.py

@@ -499,10 +499,10 @@ def flashinfer_gemm_w8a8_block_fp8_linear_with_fallback(
 
     input_2d = input.view(-1, input.shape[-1])
     backend = _get_flashinfer_groupwise_backend()
-    # TRTLLM backend requires K >= 256 and weight scales in UE8M0/R128c4
-    # packed format. Fall back to triton when scales are plain float32.
+    # Fall back to triton for non-supported formats.
+    # TODO: Check if flashinfer supports other output dtypes besides bf16.
     if backend == "trtllm" and (
-        input_2d.shape[1] < 256 or not getattr(weight_scale, "format_ue8m0", False)
+        input_2d.shape[1] < 256 or input_2d.dtype != torch.bfloat16
     ):
         return triton_w8a8_block_fp8_linear(
             input, weight, block_size, weight_scale, input_scale, bias

python/sglang/srt/model_loader/utils.py

@@ -249,18 +249,6 @@ def get_architecture_class_name(model_config: ModelConfig) -> str:
     return get_model_architecture(model_config)[1]
 
 
-def post_load_weights(model: nn.Module, model_config: ModelConfig):
-    # Model weight loading consists of two stages:
-    # 1. Initial weight loading.
-    # 2. Post-processing of weights, including assigning specific member variables.
-    # For `dummy_init`, only the second stage is required.
-    if hasattr(model, "post_load_weights"):
-        if model_config.hf_config.architectures[0] == "DeepseekV3ForCausalLMNextN":
-            model.post_load_weights(is_nextn=True)
-        else:
-            model.post_load_weights()
-
-
 def should_deepgemm_weight_requant_ue8m0(
     weight_block_size, output_dtype=None, weight_shape=None
 ):