Merge branch 'v0.9.2-dev' of http://10.16.6.30/dcutoolkit/deeplearing/vllm into v0.9.2-dev

vllm/v1/worker/gpu_model_runner.py

@@ -2088,8 +2088,9 @@ class GPUModelRunner(LoRAModelRunnerMixin):
                 hidden_states = outputs
 
             if self.speculative_config and self.speculative_config.use_eagle() and not is_profile:
-                assert isinstance(self.drafter, EagleProposer)
-                self.drafter.dummy_run(num_tokens, attn_metadata)
+                #assert isinstance(self.drafter, EagleProposer)
+                if hasattr(self, 'drafter') and isinstance(self.drafter, EagleProposer):
+                    self.drafter.dummy_run(num_tokens, attn_metadata)
 
         # This is necessary to avoid blocking DP.
         # For dummy runs, we typically skip EPLB since we don't have any real
@@ -2677,10 +2678,11 @@ class GPUModelRunner(LoRAModelRunnerMixin):
         kv_caches = self.initialize_kv_cache_tensors(kv_cache_config)
 
         if self.speculative_config and self.speculative_config.use_eagle():
-            assert isinstance(self.drafter, EagleProposer)
+            #assert isinstance(self.drafter, EagleProposer)
             # validate all draft model layers belong to the same kv cache
             # group
-            self.drafter.validate_same_kv_cache_group(kv_cache_config)
+            if hasattr(self, 'drafter') and isinstance(self.drafter, EagleProposer):
+                self.drafter.validate_same_kv_cache_group(kv_cache_config)
 
         if has_kv_transfer_group():
             get_kv_transfer_group().register_kv_caches(kv_caches)