Merge branch 'v0.6.2-dev_wm' into 'v0.6.2-dev'

[fix]修复单测test_medusa_correctness.py中的错误

See merge request dcutoolkit/deeplearing/vllm!39

tests/spec_decode/e2e/test_medusa_correctness.py

@@ -36,7 +36,7 @@ SPEC_MODEL = "abhigoyal/vllm-medusa-llama-68m-random"
 MAX_SPEC_TOKENS = 5
 
 # precision
-PRECISION = "float32"
+PRECISION = "float16"
 
 
 @pytest.mark.parametrize(

vllm/spec_decode/medusa_worker.py

@@ -40,6 +40,7 @@ class MedusaWorker(NonLLMProposerWorkerBase, Worker):
         self.medusa_buffers = None
         if hasattr(self.model_runner.model, 'medusa_choices'):
             self.medusa_choices = self.model_runner.model.medusa_choices
+            if self.medusa_choices is not None:
                 self.medusa_buffers = self.generate_medusa_buffers(
                     self.medusa_choices, device=self.device
                 )
@@ -342,7 +343,7 @@ class MedusaWorker(NonLLMProposerWorkerBase, Worker):
         
         # Move the tensors in the dictionary to the specified device
         medusa_buffers = {
-            k: v.clone().to(device)
+            k: (v.clone().to(device) if k != "tree_position_ids" else v.clone())
             if isinstance(v, torch.Tensor)
             else torch.tensor(v,  device=device)
             for k, v in medusa_buffers.items()

vllm/worker/model_runner.py

@@ -512,7 +512,7 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
         inter_data.input_positions[seq_idx] = list(range(context_len, seq_len))
 
         if seq_group_metadata.tree_position_ids is not None:
-            inter_data.input_positions[seq_idx] = seq_group_metadata.tree_position_ids.contiguous().cpu().tolist()
+            inter_data.input_positions[seq_idx] = seq_group_metadata.tree_position_ids.contiguous().tolist()
             inter_data.tree_attn_masks[seq_idx] = seq_group_metadata.tree_attn_masks
 
         inter_data.query_lens[