[quantization] AWQ Marlin doesn't work when dtype is bfloat16 (#11494)

Signed-off-by: Kai-Hsun Chen <khchen@x.ai>
Co-authored-by: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com>

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

@@ -840,12 +840,9 @@ class AWQMoEMethod(FusedMoEMethodBase):
             self.moe_runner_config.activation == "silu"
         ), "Only SiLU activation is supported."
 
-        # The input must currently be float16
         x = dispatch_output.hidden_states
         topk_output = dispatch_output.topk_output
-
         orig_dtype = x.dtype
-        x = x.half()
 
         topk_weights, topk_ids, router_logits = topk_output
 

python/sglang/srt/model_executor/model_runner.py

@@ -527,7 +527,7 @@ class ModelRunner:
                 quantization_config := getattr(
                     self.model_config.hf_config, "quantization_config", None
                 )
-            ) is not None:
+            ) is not None and "weight_block_size" in quantization_config:
                 weight_block_size_n = quantization_config["weight_block_size"][0]
 
                 if self.tp_size % self.moe_ep_size != 0: