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

fix CUDAGraph memory being counted twice (#37426)

Browse Source 
Signed-off-by: Peter Pan <Peter.Pan@daocloud.io>
Signed-off-by: Peter Pan <peter.pan@daocloud.io>
Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>
This commit is contained in:
Peter Pan
2026-03-21 01:36:32 +08:00
committed by GitHub
parent e80cfe575d
commit 79eb9369c5
1 changed files with 2 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
vllm/v1/worker/gpu_worker.py
View File

@@ -417,9 +417,7 @@ class Worker(WorkerBase):
)
self.non_torch_memory = profile_result.non_torch_increase
self.peak_activation_memory = (
profile_result.torch_peak_increase + cudagraph_memory_estimate_applied
)
self.peak_activation_memory = profile_result.torch_peak_increase
self.cudagraph_memory_estimate = cudagraph_memory_estimate
free_gpu_memory = profile_result.after_profile.free_memory
@@ -638,6 +636,7 @@ class Worker(WorkerBase):
# slightly underestimate the memory consumption.
# So leave a small buffer (=150MiB) to avoid OOM.
redundancy_buffer_memory = 150 * (1 << 20)
non_kv_cache_memory = (
self.model_runner.model_memory_usage
+ self.peak_activation_memory