A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 4, 22:30 UTC
[Core] Label torch trace logging overhead with dynamo_timed (#39329)
Signed-off-by: Frederik Gossen <frgossen@meta.com>
Frederik Gossen committed
Apr 20, 2026 at 13:31 UTC
b9cf629bd0e924c69f3d8bfefbfdb77df5ffc7be
1 file changed
+2
vllm/compilation/piecewise_backend.py
+2
@@ -11,6 +11,7 @@ from typing import Any
11
12
import torch._functorch.config
13
import torch.fx as fx
14
+from torch._dynamo.utils import dynamo_timed
15
from torch._inductor.runtime.triton_heuristics import CachingAutotuner
16
from torch._logging._internal import trace_structured
17
@@ -275,6 +276,7 @@ class PiecewiseBackend:
276
277
range_entry.compiled = True
278
279
+ @dynamo_timed("vllm_log_compile_start_torch_trace_only")
280
def _log_compile_start(self, compile_range: Range):
281
"""Log compilation event for TORCH_TRACE/tlparse."""
282
is_cudagraph_size = (