From 79eb9369c5baa83db934407b0a448a5005c3dd72 Mon Sep 17 00:00:00 2001 From: Peter Pan Date: Sat, 21 Mar 2026 01:36:32 +0800 Subject: [PATCH] fix CUDAGraph memory being counted twice (#37426) Signed-off-by: Peter Pan Signed-off-by: Peter Pan Co-authored-by: Matthew Bonanni --- vllm/v1/worker/gpu_worker.py | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-) diff --git a/vllm/v1/worker/gpu_worker.py b/vllm/v1/worker/gpu_worker.py index d101edc18..39374db5b 100644 --- a/vllm/v1/worker/gpu_worker.py +++ b/vllm/v1/worker/gpu_worker.py @@ -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