From 1be6e937b2b49bae652370d80294f6171bd7b981 Mon Sep 17 00:00:00 2001 From: Omer Ullman Argov <118735753+omera-nv@users.noreply.github.com> Date: Mon, 13 Jul 2026 17:14:25 +0300 Subject: [PATCH] lower memory required for capturing cudagraphs for large cudagraph sizes (#48483) Signed-off-by: Omer Ullman Argov <118735753+omera-nv@users.noreply.github.com> --- vllm/v1/worker/gpu_model_runner.py | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/vllm/v1/worker/gpu_model_runner.py b/vllm/v1/worker/gpu_model_runner.py index a675fcfffe..ecf8d8ec0d 100644 --- a/vllm/v1/worker/gpu_model_runner.py +++ b/vllm/v1/worker/gpu_model_runner.py @@ -6426,7 +6426,11 @@ class GPUModelRunner( kv_cache_spec = self.get_kv_cache_spec() KVCacheSpecRegistry.check_kv_cache_spec_registry(kv_cache_spec) kv_cache_groups = get_kv_cache_groups(self.vllm_config, kv_cache_spec) - min_blocks = self.compilation_config.max_cudagraph_capture_size or 1 + # the minimum number of blocks required is 1 block *per sequence* + min_blocks = ( + min(self.max_num_reqs, self.compilation_config.max_cudagraph_capture_size) + or 1 + ) # Temporarily change num_gpu_blocks_override to allocate a minimal KV cache saved_override = self.cache_config.num_gpu_blocks_override