Misc fix for min_p_sampling, --cuda-graph-bs (#2761)

python/sglang/srt/layers/logits_processor.py

@@ -117,6 +117,11 @@ class LogitsProcessor(nn.Module):
         self.final_logit_softcapping = getattr(
             self.config, "final_logit_softcapping", None
         )
+        if (
+            self.final_logit_softcapping is not None
+            and self.final_logit_softcapping < 0
+        ):
+            self.final_logit_softcapping = None
 
     def forward(
         self,

python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py

@@ -1011,11 +1011,22 @@ def fused_experts_impl(
                 out_hidden_states[begin_chunk_idx:end_chunk_idx],
             )
         else:
-            torch.sum(
-                intermediate_cache3.view(*intermediate_cache3.shape),
-                dim=1,
-                out=out_hidden_states[begin_chunk_idx:end_chunk_idx],
-            )
+            if topk_ids.shape[1] == 1:
+                out_hidden_states[begin_chunk_idx:end_chunk_idx].copy_(
+                    intermediate_cache3[:, 0]
+                )
+            elif topk_ids.shape[1] == 2:
+                torch.add(
+                    intermediate_cache3[:, 0],
+                    intermediate_cache3[:, 1],
+                    out=out_hidden_states[begin_chunk_idx:end_chunk_idx],
+                ).squeeze(dim=1)
+            elif topk_ids.shape[1] > 2:
+                torch.sum(
+                    intermediate_cache3.view(*intermediate_cache3.shape),
+                    dim=1,
+                    out=out_hidden_states[begin_chunk_idx:end_chunk_idx],
+                )
 
     return out_hidden_states
 

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

@@ -1,8 +1,7 @@
 # Adapted from https://raw.githubusercontent.com/vllm-project/vllm/v0.5.5/vllm/model_executor/layers/quantization/__init__.py
 
-from typing import Callable, Dict, Optional, Type
+from typing import Dict, Type
 
-import torch
 from vllm.model_executor.layers.quantization.aqlm import AQLMConfig
 from vllm.model_executor.layers.quantization.awq import AWQConfig
 from vllm.model_executor.layers.quantization.awq_marlin import AWQMarlinConfig