biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Hardware] Replace torch.cuda.empty_cache with torch.accelerator.empty_cache (#30681)

Browse Source 
Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
Signed-off-by: Kunshang Ji <jikunshang95@gmail.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Kunshang Ji
2026-03-04 17:49:47 +08:00
committed by GitHub
parent 5dc3538736
commit 16d2ad1d38
35 changed files with 110 additions and 59 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/v1/worker/gpu_worker.py
View File

@@ -278,7 +278,7 @@ class Worker(WorkerBase):
# Now take memory snapshot after NCCL is initialized
gc.collect()
torch.cuda.empty_cache()
torch.accelerator.empty_cache()
# take current memory snapshot
self.init_snapshot = init_snapshot = MemorySnapshot(device=self.device)
@@ -585,7 +585,7 @@ class Worker(WorkerBase):
# 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 `torch.accelerator.empty_cache`.
max_num_reqs = min(
self.scheduler_config.max_num_seqs,
self.scheduler_config.max_num_batched_tokens,