# True eager backend **URL:** \ **Category:** General\ **Created:** [February 24, 2026, 9:29am UTC](https://discuss.vllm.ai/t/true-eager-backend/2396 "2026-02-24T09:29:54Z")\ **Posts on this page:** 7\ **Page:** 1
**Author:** ![fataswellassad](https://yyz2.discourse-cdn.com/flex036/user_avatar/discuss.vllm.ai/fataswellassad/32/1150_2.png) [@fataswellassad](https://discuss.vllm.ai/u/fataswellassad)\ **Post date:** [February 24, 2026, 9:29am UTC](https://discuss.vllm.ai/t/true-eager-backend/2396/1 "2026-02-24T09:29:54Z")
On vllm v0.11.1 I ran sth like VLLM\_DISABLE\_COMPILE\_CACHE=1 VLLM\_LOG\_LEVEL=debug VLLM\_TORCH\_PROFILER\_DIR=./prof/0111-eager-eager CUDA\_VISIBLE\_DEVICES=1 vllm serve /sse\_ard/dgh/models/Qwen3-4B --served-model-name Qwen3-4B --host 0.0.0.0 --port 1110 -O.backend=eager --enforce-eager And the profiler summary looks like ![image](https://canada1.discourse-cdn.com/flex036/uploads/vllm/original/2X/1/1ac5654a1d3ddcc88b6481e9e1ff95e823679206.jpeg) Am I right in assuming that these “execute\_new\_X\_cached\_Y” are products of torch.compile? There’s no way to turn it off, even with the insane cmd that I supply above. So my questions are: - what are these “execute\_new\_X\_cached\_Y” - how do I truly turn them off? I never encounter these in v0.11.0 ---
**Author:** ![RunLLM](https://yyz2.discourse-cdn.com/flex036/user_avatar/discuss.vllm.ai/runllm/32/233_2.png) [@RunLLM](https://discuss.vllm.ai/u/RunLLM)\ **Post date:** [February 24, 2026, 9:30am UTC](https://discuss.vllm.ai/t/true-eager-backend/2396/2 "2026-02-24T09:30:32Z")
The “execute\_new\_X\_cached\_Y” entries in your profiler summary are generated by TorchInductor, which is part of torch.compile. These represent compiled graph executions, and their presence indicates that torch.compile is still active. According to [vLLM documentation](https://docs.vllm.ai/en/latest/design/debug_vllm_compile/), to fully disable torch.compile and CUDAGraphs, you must use both --enforce-eager and set the compilation config mode to NONE (e.g., -cc.mode=0). The --backend=eager flag alone does not guarantee torch.compile is off. If you still see these entries after using --enforce-eager and -cc.mode=0, it may be a bug or a change in v0.11.1, as previous versions (like v0.11.0) did not show this behavior. For complete disabling, ensure both flags are set as described above. Would you like a step-by-step breakdown of the correct command and config to ensure torch.compile is fully disabled? Sources: - [How to debug the vLLM-torch.compile integration](https://docs.vllm.ai/en/latest/design/debug_vllm_compile/) - [Compilation Cache and disabling torch.compile](https://docs.vllm.ai/en/latest/design/torch_compile/) Hint: Mention @RunLLM in the post for followups. ---
**Author:** ![fataswellassad](https://yyz2.discourse-cdn.com/flex036/user_avatar/discuss.vllm.ai/fataswellassad/32/1150_2.png) [@fataswellassad](https://discuss.vllm.ai/u/fataswellassad)\ **Post date:** [February 24, 2026, 9:39am UTC](https://discuss.vllm.ai/t/true-eager-backend/2396/3 "2026-02-24T09:39:36Z")
VLLM\_DISABLE\_COMPILE\_CACHE=1 VLLM\_LOG\_LEVEL=debug VLLM\_TORCH\_PROFILER\_DIR=./prof/0111-eager-eager CUDA\_VISIBLE\_DEVICES=1 vllm serve /sse\_ard/dgh/models/Qwen3-4B --served-model-name Qwen3-4B --host 0.0.0.0 --port 1110 -O.backend=eager -O.mode=0 --enforce-eager So I did sth like this. No luck. Actually you don’t even see the log here (what adapter is used etc) ![企业微信截图_17719242689856](https://canada1.discourse-cdn.com/flex036/uploads/vllm/original/2X/1/190e8e3f88f3b0be018c8235bca125686a9a0de1.png) @RunLLM ---
**Author:** ![fataswellassad](https://yyz2.discourse-cdn.com/flex036/user_avatar/discuss.vllm.ai/fataswellassad/32/1150_2.png) [@fataswellassad](https://discuss.vllm.ai/u/fataswellassad)\ **Post date:** [February 24, 2026, 9:47am UTC](https://discuss.vllm.ai/t/true-eager-backend/2396/4 "2026-02-24T09:47:34Z")
VLLM\_DISABLE\_COMPILE\_CACHE=1 VLLM\_LOG\_LEVEL=debug VLLM\_TORCH\_PROFILER\_DIR=./prof/0111-eager-eager CUDA\_VISIBLE\_DEVICES=1 vllm serve /sse\_ard/dgh/models/Qwen3-4B --served-model-name Qwen3-4B --host 0.0.0.0 --port 1110 -O.backend=eager -O.mode=0 --enforce-eager So I did sth like this. No luck. Actually you don’t even see the log here (what adapter is used etc) ![企业微信截图_17719242689856](https://canada1.discourse-cdn.com/flex036/uploads/vllm/original/2X/1/190e8e3f88f3b0be018c8235bca125686a9a0de1.png) @RunLLM ---
**Author:** ![RunLLM](https://yyz2.discourse-cdn.com/flex036/user_avatar/discuss.vllm.ai/runllm/32/233_2.png) [@RunLLM](https://discuss.vllm.ai/u/RunLLM)\ **Post date:** [February 24, 2026, 9:48am UTC](https://discuss.vllm.ai/t/true-eager-backend/2396/5 "2026-02-24T09:48:04Z")
Even with `-O.backend=eager -O.mode=0 --enforce-eager`, if you still see “execute\_new\_X\_cached\_Y” in your profiler, torch.compile (and thus TorchInductor) is not fully disabled. According to [vLLM debug docs](https://docs.vllm.ai/en/latest/design/debug_vllm_compile/), both `--enforce-eager` and `-cc.mode=0` (or `-O.mode=0`) should turn off torch.compile and CUDAGraphs, and the backend should be set to “eager” for full eager execution. The code in your image confirms that if `backend="eager"`, `EagerAdaptor` should be used, not Inductor. If you still see Inductor artifacts, this may be a bug in v0.11.1, as this behavior is not expected per the official documentation and code. This was not present in v0.11.0, so it could be a regression. Would you like a step-by-step troubleshooting checklist to further isolate the issue? Sources: - [How to debug the vLLM-torch.compile integration](https://docs.vllm.ai/en/latest/design/debug_vllm_compile/) - [Compilation Cache and disabling torch.compile](https://docs.vllm.ai/en/latest/design/torch_compile/) Hint: Mention @RunLLM in the post for followups. ---
**Author:** ![fataswellassad](https://yyz2.discourse-cdn.com/flex036/user_avatar/discuss.vllm.ai/fataswellassad/32/1150_2.png) [@fataswellassad](https://discuss.vllm.ai/u/fataswellassad)\ **Post date:** [February 24, 2026, 10:28am UTC](https://discuss.vllm.ai/t/true-eager-backend/2396/6 "2026-02-24T10:28:45Z")
Is there an env var I can set for torch to disable it? @RunLLM ---
**Author:** ![RunLLM](https://yyz2.discourse-cdn.com/flex036/user_avatar/discuss.vllm.ai/runllm/32/233_2.png) [@RunLLM](https://discuss.vllm.ai/u/RunLLM)\ **Post date:** [February 24, 2026, 10:29am UTC](https://discuss.vllm.ai/t/true-eager-backend/2396/7 "2026-02-24T10:29:02Z")
There is no official environment variable in vLLM or PyTorch that fully disables torch.compile; disabling must be done via vLLM’s CLI flags/configs (e.g., `--enforce-eager` and `-cc.mode=0`). Setting `VLLM_DISABLE_COMPILE_CACHE=1` only disables the compile cache, not torch.compile itself, and there is no documented env var like `TORCH_COMPILE_DISABLE` or similar for this purpose in the provided sources. If torch.compile is still active despite these flags, it is likely a bug or regression in vLLM, not a missing environment variable. Would you like more detail on how vLLM parses these flags internally? Sources: - [How to debug the vLLM-torch.compile integration](https://docs.vllm.ai/en/latest/design/debug_vllm_compile/) - [Compilation Cache and disabling torch.compile](https://docs.vllm.ai/en/latest/design/torch_compile/) Hint: Mention @RunLLM in the post for followups.