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 15, 02:15 UTC
[Misc] Respect `no_use_tqdm_on_load` flag while capturing CUDA graph (#20834)
Signed-off-by: Linkun <github@lkchen.net>
lkchen committed
Jul 11, 2025 at 23:04 UTC
f56d2996ca89989f9d80cde60650684c53e4caae
2 files changed
+5
-2
vllm/v1/worker/gpu_model_runner.py
+4
-2
@@ -2270,8 +2270,10 @@ class GPUModelRunner(LoRAModelRunnerMixin):
2270
# Only rank 0 should print progress bar during capture
2271
compilation_cases = reversed(self.cudagraph_batch_sizes)
2272
if is_global_first_rank():
2273
- compilation_cases = tqdm(list(compilation_cases),
2274
- desc="Capturing CUDA graph shapes")
2273
+ compilation_cases = tqdm(
2274
+ list(compilation_cases),
2275
+ disable=not self.load_config.use_tqdm_on_load,
2276
+ desc="Capturing CUDA graph shapes")
2277
for num_tokens in compilation_cases:
2278
# We skip EPLB here since we don't want to record dummy metrics
2279
for _ in range(
vllm/worker/model_runner.py
+1
@@ -1587,6 +1587,7 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
1587
if get_tensor_model_parallel_rank() == 0:
1588
compilation_cases = tqdm(
1589
list(compilation_cases),
1590
+ disable=not self.load_config.use_tqdm_on_load,
1591
desc="Capturing CUDA graph shapes")
1592
for batch_size, use_inputs_embeds in compilation_cases:
1593
attn_metadata = (