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 14, 05:30 UTC
[Config] Remove Unused Environment Variable `VLLM_DISABLE_PAD_FOR_CUDAGRAPH` (#26743)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Oct 14, 2025 at 11:47 UTC
6d87a2838cfbddbfd0a13df6998e0e35bd816703
2 files changed
-8
vllm/envs.py
-7
@@ -198,7 +198,6 @@ if TYPE_CHECKING:
198
VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS: bool = False
199
VLLM_ALLREDUCE_USE_SYMM_MEM: bool = True
200
VLLM_TUNED_CONFIG_FOLDER: str | None = None
201
- VLLM_DISABLE_PAD_FOR_CUDAGRAPH: bool = False
201
VLLM_GPT_OSS_HARMONY_SYSTEM_INSTRUCTIONS: bool = False
202
VLLM_CUSTOM_SCOPES_FOR_PROFILING: bool = False
203
VLLM_NVTX_SCOPES_FOR_PROFILING: bool = False
@@ -1304,12 +1303,6 @@ environment_variables: dict[str, Callable[[], Any]] = {
1303
"VLLM_ENABLE_CUDAGRAPH_GC": lambda: bool(
1304
int(os.getenv("VLLM_ENABLE_CUDAGRAPH_GC", "0"))
1305
),
1307
- # Disable padding to CUDA graph capture batch sizes.
1308
- # TODO(wentao): https://github.com/vllm-project/vllm/issues/23378
1309
- # After the issue is fixed, we can remove this flag.
1310
- "VLLM_DISABLE_PAD_FOR_CUDAGRAPH": lambda: bool(
1311
- int(os.getenv("VLLM_DISABLE_PAD_FOR_CUDAGRAPH", "0"))
1312
- ),
1306
# Used to force set up loopback IP
1307
"VLLM_LOOPBACK_IP": lambda: os.getenv("VLLM_LOOPBACK_IP", ""),
1308
# Used to set the process name prefix for vLLM processes.
vllm/v1/worker/gpu_model_runner.py
-1
@@ -2067,7 +2067,6 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
2067
def _get_num_input_tokens(self, num_scheduled_tokens: int) -> int:
2068
if (
2069
self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
2070
- and not envs.VLLM_DISABLE_PAD_FOR_CUDAGRAPH
2070
and hasattr(self, "cudagraph_batch_sizes")
2071
and self.cudagraph_batch_sizes
2072
and num_scheduled_tokens <= self.cudagraph_batch_sizes[-1]