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 8, 22:45 UTC
[Deprecation] Deprecate profiling envs (#33722)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Feb 4, 2026 at 00:58 UTC
d88a1df699f68e5284fe3a3170f8ae292a3e9c3f
2 files changed
+2
-49
docs/usage/security.md
+2
-2
@@ -178,7 +178,7 @@ These endpoints are **only available when the environment variable `VLLM_SERVER_
178
- `/is_sleeping` - Check if engine is sleeping
179
- `/collective_rpc` - Execute arbitrary RPC methods on the engine (extremely dangerous)
180
181
-**Profiler endpoints (only when `VLLM_TORCH_PROFILER_DIR` or `VLLM_TORCH_CUDA_PROFILE` are set):**
181
+**Profiler endpoints (only when profiling is enabled via `--profiler-config`):**
182
183
These endpoints are only available when profiling is enabled and should only be used for local development:
184
@@ -207,7 +207,7 @@ An attacker who can reach the vLLM HTTP server can:
207
- Cache manipulation that can disrupt service
208
- Detailed server configuration disclosure
209
210
-Similarly, never enable profiler endpoints (`VLLM_TORCH_PROFILER_DIR` or `VLLM_TORCH_CUDA_PROFILE`) in production.
210
+Similarly, never enable profiler endpoints in production.
211
212
**Be cautious with `--enable-tokenizer-info-endpoint`:** Only enable the `/tokenizer_info` endpoint if you need to expose tokenizer configuration information. This endpoint reveals chat templates and tokenizer settings that may contain sensitive implementation details or prompt engineering strategies.
213
vllm/envs.py
-47
@@ -855,53 +855,6 @@ environment_variables: dict[str, Callable[[], Any]] = {
855
"VLLM_LORA_RESOLVER_HF_REPO_LIST": lambda: os.getenv(
856
"VLLM_LORA_RESOLVER_HF_REPO_LIST", None
857
),
858
- # Enables torch CUDA profiling if set to 1.
859
- # Deprecated, see profiler_config.
860
- "VLLM_TORCH_CUDA_PROFILE": lambda: os.getenv("VLLM_TORCH_CUDA_PROFILE"),
861
- # Enables torch profiler if set.
862
- # Deprecated, see profiler_config.
863
- "VLLM_TORCH_PROFILER_DIR": lambda: os.getenv("VLLM_TORCH_PROFILER_DIR"),
864
- # Enable torch profiler to record shapes if set to 1.
865
- # Deprecated, see profiler_config.
866
- "VLLM_TORCH_PROFILER_RECORD_SHAPES": lambda: (
867
- os.getenv("VLLM_TORCH_PROFILER_RECORD_SHAPES")
868
- ),
869
- # Enable torch profiler to profile memory if set to 1.
870
- # Deprecated, see profiler_config.
871
- "VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY": lambda: (
872
- os.getenv("VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY")
873
- ),
874
- # Enable torch profiler to profile stack if set to 1.
875
- # Deprecated, see profiler_config.
876
- "VLLM_TORCH_PROFILER_WITH_STACK": lambda: (
877
- os.getenv("VLLM_TORCH_PROFILER_WITH_STACK")
878
- ),
879
- # Enable torch profiler to profile flops if set to 1.
880
- # Deprecated, see profiler_config.
881
- "VLLM_TORCH_PROFILER_WITH_FLOPS": lambda: (
882
- os.getenv("VLLM_TORCH_PROFILER_WITH_FLOPS")
883
- ),
884
- # Disable torch profiling of the AsyncLLMEngine process if set to 1.
885
- # Deprecated, see profiler_config.
886
- "VLLM_TORCH_PROFILER_DISABLE_ASYNC_LLM": lambda: (
887
- os.getenv("VLLM_TORCH_PROFILER_DISABLE_ASYNC_LLM")
888
- ),
889
- # Delay number of iterations before starting profiling when using
890
- # the torch/torch CUDA profiler. If set to 0, will start profiling immediately.
891
- # Deprecated, see profiler_config.
892
- "VLLM_PROFILER_DELAY_ITERS": lambda: (os.getenv("VLLM_PROFILER_DELAY_ITERS")),
893
- # Maximum number of iterations to profile when using the torch/torch CUDA profiler.
894
- # If set to 0, will not limit the number of iterations.
895
- "VLLM_PROFILER_MAX_ITERS": lambda: os.getenv("VLLM_PROFILER_MAX_ITERS"),
896
- # Control whether torch profiler gzip-compresses profiling files.
897
- # Deprecated, see profiler_config.
898
- "VLLM_TORCH_PROFILER_USE_GZIP": lambda: os.getenv("VLLM_TORCH_PROFILER_USE_GZIP"),
899
- # Control whether torch profiler dumps the self_cuda_time_total table.
900
- # Set to 0 to disable dumping the table.
901
- # Deprecated, see profiler_config.
902
- "VLLM_TORCH_PROFILER_DUMP_CUDA_TIME_TOTAL": lambda: (
903
- os.getenv("VLLM_TORCH_PROFILER_DUMP_CUDA_TIME_TOTAL")
904
- ),
858
# If set, vLLM will use Triton implementations of AWQ.
859
"VLLM_USE_TRITON_AWQ": lambda: bool(int(os.getenv("VLLM_USE_TRITON_AWQ", "0"))),
860
# If set, allow loading or unloading lora adapters in runtime,