replace torch.cuda.empty_cache with current_platform.empty_cache

Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>

Author: jikunshang

vllm/model_executor/layers/fused_moe/layer.py

@@ -297,7 +297,7 @@ def _maybe_pad_weight(self, weight: torch.Tensor) -> torch.Tensor:
                 and (weight.stride(-2) * weight.element_size()) % 512 == 0):
             num_pad = 256 // weight.element_size()
             weight = F.pad(weight, (0, num_pad), "constant", 0)[..., :-num_pad]
-            torch.cuda.empty_cache()
+            current_platform.empty_cache()
         return weight
 
     def process_weights_after_loading(self, layer: torch.nn.Module) -> None:

vllm/model_executor/layers/quantization/fp8.py

@@ -324,7 +324,7 @@ def _maybe_pad_weight(self, weight: torch.Tensor) -> torch.Tensor:
                 and (weight.stride(-2) * weight.element_size()) % 512 == 0):
             num_pad = 256 // weight.element_size()
             weight = F.pad(weight, (0, num_pad), "constant", 0)[..., :-num_pad]
-            torch.cuda.empty_cache()
+            current_platform.empty_cache()
         return weight
 
     def process_weights_after_loading(self, layer: Module) -> None:

vllm/model_executor/model_loader/bitsandbytes_loader.py

@@ -755,7 +755,7 @@ def load_weights(self, model: nn.Module,
             **stacked_quant_state_dict
         }
         self._bind_quant_states_to_params(model, stacked_quant_state_dict)
-        torch.cuda.empty_cache()
+        current_platform.empty_cache()
 
     def download_model(self, model_config: ModelConfig) -> None:
         self._prepare_weights(model_config.model, model_config.revision)

vllm/utils/__init__.py

@@ -2556,7 +2556,7 @@ def measure(self):
         # rather than `torch.cuda.memory_reserved()` .
         # After `torch.cuda.reset_peak_memory_stats()`,
         # `torch.cuda.memory_reserved()` will keep growing, and only shrink
-        # when we call `torch.cuda.empty_cache()` or OOM happens.
+        # when we call `current_platform.empty_cache()` or OOM happens.
         from vllm.platforms import current_platform
 
         self.torch_peak = current_platform.memory_stats().get(

vllm/v1/worker/gpu_worker.py

@@ -286,7 +286,7 @@ def compile_or_warm_up_model(self) -> None:
         # sampling related tensors of max possible shape to avoid memory
         # fragmentation issue.
         # NOTE: This is called after `capture_model` on purpose to prevent
-        # memory buffers from being cleared by `torch.cuda.empty_cache`.
+        # memory buffers from being cleared by `current_platform.empty_cache`.
         if get_pp_group().is_last_rank:
             max_num_reqs = min(self.scheduler_config.max_num_seqs,
                                self.scheduler_config.max_num_batched_tokens)