rm debug log

vllm/executor/mp_distributed_executor.py

@@ -48,13 +48,13 @@ class MultiprocessingDistributedExecutor(DistributedExecutorBase):
                 f"is less than than max local gpu count ({cuda_device_count})")
 
         # Set CUDA_VISIBLE_DEVICES for the driver, inherited by workers
-        if "CUDA_VISIBLE_DEVICES" or "HIP_VISIBLE_DEVICES" not in os.environ:
-            update_environment_variables({
-                "CUDA_VISIBLE_DEVICES": (",".join(map(str, range(world_size))))
-            })
-            update_environment_variables({
-                "HIP_VISIBLE_DEVICES": (",".join(map(str, range(world_size))))
-            })
+        # if "CUDA_VISIBLE_DEVICES" or "HIP_VISIBLE_DEVICES" not in os.environ:
+        #     update_environment_variables({
+        #         "CUDA_VISIBLE_DEVICES": (",".join(map(str, range(world_size))))
+        #     })
+        #     update_environment_variables({
+        #         "HIP_VISIBLE_DEVICES": (",".join(map(str, range(world_size))))
+        #     })
 
     def _init_executor(self) -> None:
 

vllm/model_executor/layers/sampler.py

@@ -746,7 +746,6 @@ def _sample_with_torch(
         else:
             raise ValueError(f"Unsupported sampling type: {sampling_type}")
 
-    print('###sampled_token_ids', sampled_token_ids_)
     # Encapsulate arguments for computing Pythonized sampler
     # results, whether deferred or otherwise.
     maybe_deferred_args = SampleResultArgsType(

vllm/spec_decode/spec_decode_worker.py

@@ -910,7 +910,6 @@ class SpecDecodeWorker(LoRANotSupportedWorkerBase):
             accepted_token_ids, target_logprobs, select_indices_list, accept_lengths = self._verify_tokens(
                 execute_model_req.seq_group_metadata_list, proposal_scores,
                 proposals, execute_model_req.num_lookahead_slots)
-            print('###accepted_token_ids', accepted_token_ids)
             
             # move kv_caches of selected tokens to right positions
             if self.tree_decoding:
@@ -1341,7 +1340,6 @@ class SpecDecodeWorker(LoRANotSupportedWorkerBase):
             self._maybe_log_stage_times(*stage_times)
         # First `n_prefills` entries will contain prefills SamplerOutput when
         # chunked prefill is enabled, the rest is decodes in multi-step format.
-        print('###sampler_output_list', sampler_output_list)
         return sampler_output_list
 
     def _maybe_log_stage_times(self, average_time_per_proposal_tok_ms: float,

vllm/worker/model_runner.py

@@ -902,7 +902,6 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
         # Tokens and positions.
         if cuda_graph_pad_size:
             input_tokens.extend(itertools.repeat(0, cuda_graph_pad_size))
-        print('###input_tokens', input_tokens)
         assert self.runner.device is not None
         input_tokens_tensor = async_tensor_h2d(input_tokens, torch.long,
                                                self.runner.device,
@@ -917,14 +916,12 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
             for idx in range(3):
                 mrope_input_positions[idx].extend(
                     itertools.repeat(0, cuda_graph_pad_size))
-            print('###mrope_input_positions', mrope_input_positions)
             input_positions_tensor = async_tensor_h2d(mrope_input_positions,
                                                       torch.long,
                                                       self.runner.device,
                                                       self.runner.pin_memory)
         else:
             input_positions.extend(itertools.repeat(0, cuda_graph_pad_size))
-            print('###input_positions', input_positions)
             input_positions_tensor = async_tensor_h2d(input_positions,
                                                       torch.long,
                                                       self.runner.device,
@@ -932,7 +929,6 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
         # Sequence and query lengths.
         if cuda_graph_pad_size:
             seq_lens.extend(itertools.repeat(1, cuda_graph_pad_size))
-        print('###seq_lens', seq_lens)
 
         # Attention metadata.
         attn_metadata = self.attn_metadata_builder.build(
@@ -1006,7 +1002,6 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
             prompt_adapter_mapping=prompt_adapter_mapping,
             prompt_adapter_requests=prompt_adapter_requests)
         
-        print('###model_input', ret)
         return ret
 
 

vllm/zero_overhead/model_runner.py

@@ -92,7 +92,6 @@ class ZeroOverheadModelInputForGpuBuilder(ModelInputForGPUBuilder):
 
     def build(self) -> ModelInputForGPU:
         model_input = super().build()
-        print('###model_input', model_input)
         last_sampler = get_last_sampler()
         spec_step = get_spec_step()
         last_step = get_spec_last_step()
@@ -167,5 +166,4 @@ class ZeroOverheadModelInputForGpuBuilder(ModelInputForGPUBuilder):
                     )
             
 
-        print('###zero_model_input', model_input)
         return model_input

vllm/zero_overhead/sampler.py

@@ -359,7 +359,6 @@ def _sample_with_torch(
                 sampled_token_ids_tensor[long_sample_indices] = \
                     multinomial_samples[sampling_type].to(torch.long)
 
-    print('###sampled_token_ids', last_sampler.sampled_token_ids_tensor)
     # Encapsulate arguments for computing Pythonized sampler
     # results, whether deferred or otherwise.
     maybe_deferred_args = SampleResultArgsType(

vllm/zero_overhead/spec_decode/spec_decode_worker.py

@@ -545,7 +545,6 @@ class ZeroOverheadSpecDecodeWorker(SpecDecodeWorker):
             self._maybe_log_stage_times(*stage_times)
         # First `n_prefills` entries will contain prefills SamplerOutput when
         # chunked prefill is enabled, the rest is decodes in multi-step format.
-        print('###sampler_output_list', sampler_output_list)
         return sampler_output_list
 
     def _track_sequences_with_bonus_tokens(