From 79ac28c5bd363d243ff2dea39306b568f59d4153 Mon Sep 17 00:00:00 2001 From: Misha Goin Date: Sat, 26 Sep 2026 11:22:33 -0400 Subject: [PATCH] [CI] Reduce CUDA graph mode test overhead (#58749) --- tests/v1/cudagraph/test_cudagraph_mode.py | 16 ++++++++++------ 1 file changed, 10 insertions(+), 6 deletions(-) diff --git a/tests/v1/cudagraph/test_cudagraph_mode.py b/tests/v1/cudagraph/test_cudagraph_mode.py index e413fd91e3..0147f90df5 100644 --- a/tests/v1/cudagraph/test_cudagraph_mode.py +++ b/tests/v1/cudagraph/test_cudagraph_mode.py @@ -54,13 +54,16 @@ def test_backend_and_cudagraph_mode_combo(backend_name, cudagraph_mode, supporte llm = LLM( model="Qwen/Qwen2-1.5B-Instruct", - max_num_seqs=256, + max_num_seqs=16, + max_num_batched_tokens=128, trust_remote_code=True, gpu_memory_utilization=0.45, max_model_len=1024, attention_config=attention_config, compilation_config=CompilationConfig( - mode=CompilationMode.VLLM_COMPILE, cudagraph_mode=cudagraph_mode + mode=CompilationMode.VLLM_COMPILE, + cudagraph_mode=cudagraph_mode, + cudagraph_capture_sizes=[1, 2, 4, 8, 16, 64, 128], ), ) llm.generate(["Hello, my name is"] * 10) @@ -72,11 +75,9 @@ attn_backend = "RocmAttn" if current_platform.is_rocm() else "FA2" combo_cases_2 = [ (attn_backend, "FULL", CompilationMode.NONE, True), - (attn_backend, "FULL", CompilationMode.VLLM_COMPILE, True), (attn_backend, "PIECEWISE", CompilationMode.NONE, True), (attn_backend, "PIECEWISE", CompilationMode.VLLM_COMPILE, True), (attn_backend, "FULL_AND_PIECEWISE", CompilationMode.NONE, True), - (attn_backend, "FULL_AND_PIECEWISE", CompilationMode.VLLM_COMPILE, True), (attn_backend, "FULL_DECODE_ONLY", CompilationMode.NONE, True), (attn_backend, "FULL_DECODE_ONLY", CompilationMode.VLLM_COMPILE, True), (attn_backend, "NONE", CompilationMode.NONE, True), @@ -100,13 +101,16 @@ def test_cudagraph_compilation_combo( llm = LLM( model="Qwen/Qwen2-1.5B-Instruct", - max_num_seqs=256, + max_num_seqs=16, + max_num_batched_tokens=128, trust_remote_code=True, gpu_memory_utilization=0.45, max_model_len=1024, attention_config=attention_config, compilation_config=CompilationConfig( - mode=compilation_mode, cudagraph_mode=cudagraph_mode + mode=compilation_mode, + cudagraph_mode=cudagraph_mode, + cudagraph_capture_sizes=[1, 2, 4, 8, 16, 64, 128], ), ) llm.generate(["Hello, my name is"] * 10)