Merge branch 'v0.9.2-dev-fix-zero' into 'v0.9.2-dev'

fix: zero overhead KERNEL VMFault

See merge request dcutoolkit/deeplearing/vllm!446

vllm/zero_overhead/v1/eagle.py

@@ -154,10 +154,11 @@ class V1ZeroEagleProposer(EagleProposer):
                     self.attn_metadata_cudagraph.decode.seq_lens[:attn_metadata.num_decode_tokens] = (
                         attn_metadata.decode.seq_lens)
 
+        use_ep = self.vllm_config.parallel_config.enable_expert_parallel
         with set_forward_context(per_layer_attn_metadata,
                                  self.vllm_config,
-                                 num_tokens=num_input_tokens,):
-                                 #skip_cuda_graphs=not decoding):
+                                 num_tokens=num_input_tokens,
+                                 skip_cuda_graphs= not (decoding or use_ep)):
             ret_hidden_states = self.model(
                 self.input_ids[:num_input_tokens],
                 self.positions[:num_input_tokens],