From 074d57fce1a1f39426fdebe8b4affac02c6e749c Mon Sep 17 00:00:00 2001 From: Nick Hill Date: Thu, 24 Sep 2026 12:07:32 -0700 Subject: [PATCH] [Bugfix] Keep JIT warmup under enforce-eager when fault tolerance is on (#58593) Signed-off-by: Nick Hill Co-authored-by: Kimi --- tests/compile/test_config.py | 15 +++++++++++++++ vllm/config/vllm.py | 23 ++++++++++++++++------- 2 files changed, 31 insertions(+), 7 deletions(-) diff --git a/tests/compile/test_config.py b/tests/compile/test_config.py index 2d4ef13e33..74ed0d34db 100644 --- a/tests/compile/test_config.py +++ b/tests/compile/test_config.py @@ -264,6 +264,21 @@ def test_enforce_eager(vllm_runner, monkeypatch): pass +@pytest.mark.parametrize("enable_fault_tolerance", [False, True]) +def test_enforce_eager_jit_warmup(enable_fault_tolerance): + """Enforce-eager disables JIT warmup unless fault tolerance is on. + + FT fault detection runs against deadlines that in-inference Triton + compilation latency spikes can blow past, so warmup stays enabled. + """ + config = VllmConfig( + model_config=ModelConfig(model="facebook/opt-125m", enforce_eager=True), + parallel_config=ParallelConfig(enable_fault_tolerance=enable_fault_tolerance), + ) + assert config.compilation_config.mode == CompilationMode.NONE + assert config.kernel_config.enable_jit_warmup == enable_fault_tolerance + + @pytest.mark.forked def test_torch_compile_disable(vllm_runner, monkeypatch): monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0") diff --git a/vllm/config/vllm.py b/vllm/config/vllm.py index 0aafd2506c..f5851e0da2 100644 --- a/vllm/config/vllm.py +++ b/vllm/config/vllm.py @@ -1666,15 +1666,24 @@ class VllmConfig: ) if self.model_config is not None and self.model_config.enforce_eager: - logger.warning_once( - "Enforce eager set, disabling torch.compile, CUDAGraphs, and JIT " - "kernel warmup. This is equivalent to setting -cc.mode=none " - "-cc.cudagraph_mode=none and " - "--kernel_config.enable_jit_warmup=False" - ) self.compilation_config.mode = CompilationMode.NONE self.compilation_config.cudagraph_mode = CUDAGraphMode.NONE - self.kernel_config.enable_jit_warmup = False + if self.parallel_config.enable_fault_tolerance: + # Keep JIT warmup: in-inference Triton compilation latency + # spikes can delay peer-fault detection past its deadline. + logger.warning_once( + "Enforce eager set, disabling torch.compile and CUDAGraphs. " + "This is equivalent to setting -cc.mode=none " + "-cc.cudagraph_mode=none" + ) + else: + logger.warning_once( + "Enforce eager set, disabling torch.compile, CUDAGraphs, and " + "JIT kernel warmup. This is equivalent to setting " + "-cc.mode=none -cc.cudagraph_mode=none and " + "--kernel_config.enable_jit_warmup=False" + ) + self.kernel_config.enable_jit_warmup = False if os.environ.get("TORCH_COMPILE_DISABLE") == "1": logger.warning_once(