diff --git a/tests/models/multimodal/generation/test_voxtral_realtime.py b/tests/models/multimodal/generation/test_voxtral_realtime.py index 1bd9f62303..2c13e07ff3 100644 --- a/tests/models/multimodal/generation/test_voxtral_realtime.py +++ b/tests/models/multimodal/generation/test_voxtral_realtime.py @@ -21,7 +21,7 @@ from vllm.utils.math_utils import cdiv from vllm.v1.engine.async_llm import AsyncLLM from vllm.v1.kv_cache_interface import SlidingWindowSpec -from ....utils import ROCM_ENGINE_KWARGS +from ....utils import ROCM_ENGINE_KWARGS, create_new_process_for_each_test MODEL_NAME = "mistralai/Voxtral-Mini-4B-Realtime-2602" AUDIO_LAYER_NAME = "whisper_encoder.whisper_encoder.layers.0.layers.self_attn.attn" @@ -192,6 +192,7 @@ def test_voxtral_realtime_forward(audio_assets, tokenizer, vllm_runner, monkeypa "cudagraph_mode", [CUDAGraphMode.FULL_DECODE_ONLY, CUDAGraphMode.FULL_AND_PIECEWISE], ) +@create_new_process_for_each_test(method="spawn") def test_voxtral_realtime_cudagraph( audio_assets, tokenizer, vllm_runner, monkeypatch, cudagraph_mode ): diff --git a/tests/utils_/test_gpu_sync_debug.py b/tests/utils_/test_gpu_sync_debug.py index dc8b122800..10b550c0a1 100644 --- a/tests/utils_/test_gpu_sync_debug.py +++ b/tests/utils_/test_gpu_sync_debug.py @@ -123,6 +123,24 @@ def test_suppressing_works_while_compiling(monkeypatch): with_gpu_sync_check(suppressed)() +@create_new_process_for_each_test() +def test_copy_checker_is_traceable_while_compiling(monkeypatch): + """The copy wrapper must bypass ContextVars while Dynamo is tracing.""" + monkeypatch.setattr(gsd, "_SYNC_CHECK_MODE", "error") + monkeypatch.setattr(gsd, "_sync_check_enabled", True) + gsd._install_copy_checkers() + + def convert_like(x, like): + return x.to(device=like.device, dtype=like.dtype) + + compiled = torch.compile(convert_like, fullgraph=True) + result = compiled( + torch.ones(4, device="cuda", dtype=torch.float64), + torch.ones(4, device="cuda", dtype=torch.float32), + ) + assert result.dtype == torch.float32 + + @create_new_process_for_each_test() def test_sync_debug_mode_restored_after_checked_call(monkeypatch): """The mode is armed only for the duration of a checked call. Leaving it diff --git a/vllm/utils/gpu_sync_debug.py b/vllm/utils/gpu_sync_debug.py index 53fcdfe710..8d0e7ae187 100644 --- a/vllm/utils/gpu_sync_debug.py +++ b/vllm/utils/gpu_sync_debug.py @@ -225,6 +225,16 @@ _torch_to = torch.Tensor.to _torch_cuda = torch.Tensor.cuda _torch_copy_ = torch.Tensor.copy_ + +# Once ``Tensor.to`` is replaced below, Dynamo sees the saved C++ descriptor +# through this module global instead of as a recognized tensor method. Put the +# descriptor behind a weak-referenceable Python callable so ``allow_in_graph`` +# can admit it and the downstream compiler can lower the original operation. +@torch.compiler.allow_in_graph +def _torch_to_in_graph(self, *args, **kwargs): + return _torch_to(self, *args, **kwargs) + + # torch's own parser for `Tensor.to`'s overloaded signature, returning # (device, dtype, non_blocking, memory_format) (also used by torch._dynamo). _parse_to = getattr(torch._C._nn, "_parse_to", None) @@ -232,7 +242,7 @@ _parse_to = getattr(torch._C._nn, "_parse_to", None) def _active_check_mode() -> str | None: """The mode the calling thread is being checked in, if any.""" - if not _sync_check_enabled or _allow_depth.get() or torch.compiler.is_compiling(): + if not _sync_check_enabled or torch.compiler.is_compiling() or _allow_depth.get(): return None return _checking.get() @@ -296,7 +306,7 @@ def _checked_to(self, *args, **kwargs): else: if non_blocking and (reason := _cpu_copy_stall_reason(self, device)): _report_implicit_sync("to", "H2D", reason, mode) - return _torch_to(self, *args, **kwargs) + return _torch_to_in_graph(self, *args, **kwargs) def _checked_cuda(self, *args, **kwargs):