# Is the reason for my vllm 0.20.0 failing to start because of nixl? **URL:** \ **Category:** General\ **Created:** [May 3, 2026, 10:06am UTC](https://discuss.vllm.ai/t/is-the-reason-for-my-vllm-0-20-0-failing-to-start-because-of-nixl/2612 "2026-05-03T10:06:21Z")\ **Posts on this page:** 10\ **Page:** 1
**Author:** ![diarrheavomit](https://avatars.discourse-cdn.com/v4/letter/d/8dc957/32.png) [@diarrheavomit](https://discuss.vllm.ai/u/diarrheavomit)\ **Post date:** [May 3, 2026, 10:06am UTC](https://discuss.vllm.ai/t/is-the-reason-for-my-vllm-0-20-0-failing-to-start-because-of-nixl/2612/1 "2026-05-03T10:06:21Z")
Using the vllm/vllm-openai-rocm:latest docker image on debian 13 with rocm 7.2.2 (inside the container it’s 7.2.1.70201-81~22.04) and an amd v620. This is what happens when I try to start vllm: root@e9640e87ad42:/app# vllm serve \ --model /model \ –dtype float16 \ –max-model-len 32768 \ –max-num-seqs 1 WARNING 05-03 09:58:41 [argparse\_utils.py:257] With `vllm serve`, you should provide the model as a positional argument or in a config file instead of via the `--model` option. The `--model` option will be removed in a future version. (APIServer pid=8) INFO 05-03 09:58:41 [utils.py:299] (APIServer pid=8) INFO 05-03 09:58:41 [utils.py:299] █ █ █▄ ▄█ (APIServer pid=8) INFO 05-03 09:58:41 [utils.py:299] ▄▄ ▄█ █ █ █ ▀▄▀ █ version 0.20.0 (APIServer pid=8) INFO 05-03 09:58:41 [utils.py:299] █▄█▀ █ █ █ █ model /model (APIServer pid=8) INFO 05-03 09:58:41 [utils.py:299] ▀▀ ▀▀▀▀▀ ▀▀▀▀▀ ▀ ▀ (APIServer pid=8) INFO 05-03 09:58:41 [utils.py:299] (APIServer pid=8) INFO 05-03 09:58:41 [utils.py:233] non-default args: {‘model\_tag’: ‘/model’, ‘model’: ‘/model’, ‘dtype’: ‘float16’, ‘max\_model\_len’: 32768, ‘max\_num\_seqs’: 1} (APIServer pid=8) INFO 05-03 09:58:54 [nixl\_utils.py:20] Setting UCX\_RCACHE\_MAX\_UNRELEASED to ‘1024’ to avoid a rare memory leak in UCX when using NIXL. (APIServer pid=8) WARNING 05-03 09:58:54 [nixl\_utils.py:34] NIXL is not available (APIServer pid=8) WARNING 05-03 09:58:54 [nixl\_utils.py:44] NIXL agent config is not available (APIServer pid=8) INFO 05-03 09:58:54 [model.py:555] Resolved architecture: Qwen3ForCausalLM (APIServer pid=8) WARNING 05-03 09:58:54 [model.py:2018] Casting torch.bfloat16 to torch.float16. (APIServer pid=8) INFO 05-03 09:58:54 [model.py:1680] Using max model len 32768 (APIServer pid=8) INFO 05-03 09:58:54 [vllm.py:840] Asynchronous scheduling is enabled. (APIServer pid=8) INFO 05-03 09:58:54 [kernel.py:205] Final IR op priority after setting platform defaults: IrOpPriorityConfig(rms\_norm=[‘native’]) (EngineCore pid=407) INFO 05-03 09:59:02 [core.py:109] Initializing a V1 LLM engine (v0.20.0) with config: model=‘/model’, speculative\_config=None, tokenizer=‘/model’, skip\_tokenizer\_init=False, tokenizer\_mode=auto, revision=None, tokenizer\_revision=None, trust\_remote\_code=False, dtype=torch.float16, max\_seq\_len=32768, download\_dir=None, load\_format=auto, tensor\_parallel\_size=1, pipeline\_parallel\_size=1, data\_parallel\_size=1, decode\_context\_parallel\_size=1, dcp\_comm\_backend=ag\_rs, disable\_custom\_all\_reduce=True, quantization=None, quantization\_config=None, enforce\_eager=False, enable\_return\_routed\_experts=False, kv\_cache\_dtype=auto, device\_config=cuda, structured\_outputs\_config=StructuredOutputsConfig(backend=‘auto’, disable\_any\_whitespace=False, disable\_additional\_properties=False, reasoning\_parser=‘’, reasoning\_parser\_plugin=‘’, enable\_in\_reasoning=False), observability\_config=ObservabilityConfig(show\_hidden\_metrics\_for\_version=None, otlp\_traces\_endpoint=None, collect\_detailed\_traces=None, kv\_cache\_metrics=False, kv\_cache\_metrics\_sample=0.01, cudagraph\_metrics=False, enable\_layerwise\_nvtx\_tracing=False, enable\_mfu\_metrics=False, enable\_mm\_processor\_stats=False, enable\_logging\_iteration\_details=False), seed=0, served\_model\_name=/model, enable\_prefix\_caching=True, enable\_chunked\_prefill=True, pooler\_config=None, compilation\_config={‘mode’: \, ‘debug\_dump\_path’: None, ‘cache\_dir’: ‘’, ‘compile\_cache\_save\_format’: ‘binary’, ‘backend’: ‘inductor’, ‘custom\_ops’: [‘+sparse\_attn\_indexer’, ‘none’], ‘ir\_enable\_torch\_wrap’: True, ‘splitting\_ops’: [‘vllm::unified\_attention\_with\_output’, ‘vllm::unified\_mla\_attention\_with\_output’, ‘vllm::mamba\_mixer2’, ‘vllm::mamba\_mixer’, ‘vllm::short\_conv’, ‘vllm::linear\_attention’, ‘vllm::plamo2\_mamba\_mixer’, ‘vllm::gdn\_attention\_core’, ‘vllm::gdn\_attention\_core\_xpu’, ‘vllm::olmo\_hybrid\_gdn\_full\_forward’, ‘vllm::kda\_attention’, ‘vllm::sparse\_attn\_indexer’, ‘vllm::rocm\_aiter\_sparse\_attn\_indexer’, ‘vllm::deepseek\_v4\_attention’, ‘vllm::unified\_kv\_cache\_update’, ‘vllm::unified\_mla\_kv\_cache\_update’], ‘compile\_mm\_encoder’: False, ‘cudagraph\_mm\_encoder’: False, ‘encoder\_cudagraph\_token\_budgets’: , ‘encoder\_cudagraph\_max\_vision\_items\_per\_batch’: 0, ‘encoder\_cudagraph\_max\_frames\_per\_batch’: None, ‘compile\_sizes’: , ‘compile\_ranges\_endpoints’: [2048], ‘inductor\_compile\_config’: {‘enable\_auto\_functionalized\_v2’: False, ‘size\_asserts’: False, ‘alignment\_asserts’: False, ‘scalar\_asserts’: False, ‘combo\_kernels’: True, ‘benchmark\_combo\_kernel’: True}, ‘inductor\_passes’: {}, ‘cudagraph\_mode’: \, ‘cudagraph\_num\_of\_warmups’: 1, ‘cudagraph\_capture\_sizes’: [1, 2], ‘cudagraph\_copy\_inputs’: False, ‘cudagraph\_specialize\_lora’: True, ‘use\_inductor\_graph\_partition’: False, ‘pass\_config’: {‘fuse\_norm\_quant’: False, ‘fuse\_act\_quant’: False, ‘fuse\_attn\_quant’: False, ‘enable\_sp’: False, ‘fuse\_gemm\_comms’: False, ‘fuse\_allreduce\_rms’: False}, ‘max\_cudagraph\_capture\_size’: 2, ‘dynamic\_shapes\_config’: {‘type’: \, ‘evaluate\_guards’: False, ‘assume\_32\_bit\_indexing’: False}, ‘local\_cache\_dir’: None, ‘fast\_moe\_cold\_start’: True, ‘static\_all\_moe\_layers’: }, kernel\_config=KernelConfig(ir\_op\_priority=IrOpPriorityConfig(rms\_norm=[‘native’]), enable\_flashinfer\_autotune=True, moe\_backend=‘auto’) (EngineCore pid=407) WARNING 05-03 09:59:03 [nixl\_utils.py:34] NIXL is not available (EngineCore pid=407) WARNING 05-03 09:59:03 [nixl\_utils.py:44] NIXL agent config is not available (EngineCore pid=407) INFO 05-03 09:59:04 [parallel\_state.py:1402] world\_size=1 rank=0 local\_rank=0 distributed\_init\_method=tcp://10.88.0.5:46367 backend=nccl (EngineCore pid=407) INFO 05-03 09:59:04 [parallel\_state.py:1715] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A, EPLB rank N/A (APIServer pid=8) Traceback (most recent call last): (APIServer pid=8) File “/usr/local/bin/vllm”, line 10, in (APIServer pid=8) sys.exit(main()) (APIServer pid=8) ^^^^^^ (APIServer pid=8) File “/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/cli/main.py”, line 92, in main (APIServer pid=8) args.dispatch\_function(args) (APIServer pid=8) File “/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/cli/serve.py”, line 122, in cmd (APIServer pid=8) uvloop.run(run\_server(args)) (APIServer pid=8) File “/usr/local/lib/python3.12/dist-packages/uvloop/ **init**.py”, line 96, in run (APIServer pid=8) return \_\_asyncio.run( (APIServer pid=8) ^^^^^^^^^^^^^^ (APIServer pid=8) File “/usr/lib/python3.12/asyncio/runners.py”, line 195, in run (APIServer pid=8) return runner.run(main) (APIServer pid=8) ^^^^^^^^^^^^^^^^ (APIServer pid=8) File “/usr/lib/python3.12/asyncio/runners.py”, line 118, in run (APIServer pid=8) return self.\_loop.run\_until\_complete(task) (APIServer pid=8) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ (APIServer pid=8) File “uvloop/loop.pyx”, line 1518, in uvloop.loop.Loop.run\_until\_complete (APIServer pid=8) File “/usr/local/lib/python3.12/dist-packages/uvloop/ **init**.py”, line 48, in wrapper (APIServer pid=8) return await main (APIServer pid=8) ^^^^^^^^^^ (APIServer pid=8) File “/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/api\_server.py”, line 678, in run\_server (APIServer pid=8) await run\_server\_worker(listen\_address, sock, args, \*\*uvicorn\_kwargs) (APIServer pid=8) File “/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/api\_server.py”, line 692, in run\_server\_worker (APIServer pid=8) async with build\_async\_engine\_client( (APIServer pid=8) ^^^^^^^^^^^^^^^^^^^^^^^^^^ (APIServer pid=8) File “/usr/lib/python3.12/contextlib.py”, line 210, in **aenter** (APIServer pid=8) return await anext(self.gen) (APIServer pid=8) ^^^^^^^^^^^^^^^^^^^^^ (APIServer pid=8) File “/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/api\_server.py”, line 100, in build\_async\_engine\_client (APIServer pid=8) async with build\_async\_engine\_client\_from\_engine\_args( (APIServer pid=8) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ (APIServer pid=8) File “/usr/lib/python3.12/contextlib.py”, line 210, in **aenter** (APIServer pid=8) return await anext(self.gen) (APIServer pid=8) ^^^^^^^^^^^^^^^^^^^^^ (APIServer pid=8) File “/usr/local/lib/python3.12/dist-packages/vllm/entrypoints/openai/api\_server.py”, line 136, in build\_async\_engine\_client\_from\_engine\_args (APIServer pid=8) async\_llm = AsyncLLM.from\_vllm\_config( (APIServer pid=8) ^^^^^^^^^^^^^^^^^^^^^^^^^^ (APIServer pid=8) File “/usr/local/lib/python3.12/dist-packages/vllm/v1/engine/async\_llm.py”, line 217, in from\_vllm\_config (APIServer pid=8) return cls( (APIServer pid=8) ^^^^ (APIServer pid=8) File “/usr/local/lib/python3.12/dist-packages/vllm/v1/engine/async\_llm.py”, line 146, in **init** (APIServer pid=8) self.engine\_core = EngineCoreClient.make\_async\_mp\_client( (APIServer pid=8) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ (APIServer pid=8) File “/usr/local/lib/python3.12/dist-packages/vllm/tracing/otel.py”, line 178, in sync\_wrapper (APIServer pid=8) return func(\*args, \*\*kwargs) (APIServer pid=8) ^^^^^^^^^^^^^^^^^^^^^ (APIServer pid=8) File “/usr/local/lib/python3.12/dist-packages/vllm/v1/engine/core\_client.py”, line 130, in make\_async\_mp\_client (APIServer pid=8) return AsyncMPClient(\*client\_args) (APIServer pid=8) ^^^^^^^^^^^^^^^^^^^^^^^^^^^ (APIServer pid=8) File “/usr/local/lib/python3.12/dist-packages/vllm/tracing/otel.py”, line 178, in sync\_wrapper (APIServer pid=8) return func(\*args, \*\*kwargs) (APIServer pid=8) ^^^^^^^^^^^^^^^^^^^^^ (APIServer pid=8) File “/usr/local/lib/python3.12/dist-packages/vllm/v1/engine/core\_client.py”, line 900, in **init** (APIServer pid=8) super(). **init** ( (APIServer pid=8) File “/usr/local/lib/python3.12/dist-packages/vllm/v1/engine/core\_client.py”, line 535, in **init** (APIServer pid=8) with launch\_core\_engines( (APIServer pid=8) ^^^^^^^^^^^^^^^^^^^^ (APIServer pid=8) File “/usr/lib/python3.12/contextlib.py”, line 144, in **exit** (APIServer pid=8) next(self.gen) (APIServer pid=8) File “/usr/local/lib/python3.12/dist-packages/vllm/v1/engine/utils.py”, line 1119, in launch\_core\_engines (APIServer pid=8) wait\_for\_engine\_startup( (APIServer pid=8) File “/usr/local/lib/python3.12/dist-packages/vllm/v1/engine/utils.py”, line 1178, in wait\_for\_engine\_startup (APIServer pid=8) raise RuntimeError( (APIServer pid=8) RuntimeError: Engine core initialization failed. See root cause above. Failed core proc(s): {} root@e9640e87ad42:/app# I see there are some warnings about nixl not being available - do I need to install something like rixl to make vllm work? ---
**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:** [May 3, 2026, 10:06am UTC](https://discuss.vllm.ai/t/is-the-reason-for-my-vllm-0-20-0-failing-to-start-because-of-nixl/2612/2 "2026-05-03T10:06:39Z")
The error `RuntimeError: Engine core initialization failed. See root cause above. Failed core proc(s): {}` is a generic vLLM startup failure. The warnings about “NIXL is not available” are not the cause of the crash—NIXL is only required for advanced multi-node or disaggregated setups, not for basic single-node inference. You do **not** need to install NIXL or RIXL for standard vLLM usage on a single node or GPU ([source](https://github.com/vllm-project/vllm/issues/36615), [source](https://github.com/vllm-project/vllm/issues/32375)). The actual root cause is not shown in your logs. This type of error on ROCm platforms (especially with AMD GPUs like V620) is often due to ROCm version or driver incompatibility, or unsupported GPU architectures. There are known issues with ROCm 7.x and certain AMD GPUs, and vLLM may not support all ROCm/AMD combinations out of the box ([source](https://github.com/vllm-project/vllm/issues/36615), [source](https://github.com/vllm-project/vllm/issues/28460)). Please check the full logs for any segfaults or errors from ROCm, and verify that your GPU and ROCm version are supported by vLLM. Would you like more detail on how to debug ROCm/vLLM compatibility or how to get more informative error logs? Sources: - [https://github.com/vllm-project/vllm/issues/36615](https://github.com/vllm-project/vllm/issues/36615) - [https://github.com/vllm-project/vllm/issues/32375](https://github.com/vllm-project/vllm/issues/32375) - [https://github.com/vllm-project/vllm/issues/28460](https://github.com/vllm-project/vllm/issues/28460) Hint: Mention @RunLLM in the post for followups. ---
**Author:** ![diarrheavomit](https://avatars.discourse-cdn.com/v4/letter/d/8dc957/32.png) [@diarrheavomit](https://discuss.vllm.ai/u/diarrheavomit)\ **Post date:** [May 3, 2026, 10:18am UTC](https://discuss.vllm.ai/t/is-the-reason-for-my-vllm-0-20-0-failing-to-start-because-of-nixl/2612/3 "2026-05-03T10:18:44Z")
@RunLLM I’ve seen [[Bug]: Abnormally bad performance on AMD ROCM gfx1030 (W6800, V620, 6900XT 6800XT) · Issue #38107 · vllm-project/vllm · GitHub](https://github.com/vllm-project/vllm/issues/38107) run vllm recently using a v620, so it should be possible right? Does vllm require specifically 7.2.0 (like the one in the post)? I just checked inside the container, and it’s running 7.2.1.70201-81~22.04 amd64 (7.2.2 is the version outside of the container). Rocm-smi’s wording: Device Node IDs Temp Power Partitions SCLK MCLK Fan Perf PwrCap VRAM% GPU% 0 4 0x73a1, 55563 33.0°C 6.0W N/A, N/A, 0 0Mhz 96Mhz 0% auto 250.0W 0% 0% 1 3 0x73a1, 10554 32.0°C 7.0W N/A, N/A, 0 0Mhz 96Mhz 0% auto 250.0W 0% 0% 2 2 0x73a1, 31082 32.0°C 6.0W N/A, N/A, 0 0Mhz 96Mhz 0% auto 250.0W 0% 0% 3 1 0x73a1, 55960 34.0°C 6.0W N/A, N/A, 0 0Mhz 96Mhz 0% auto 250.0W 0% 0% Uses node to describe a gpu - but they’re all on the same single socket computer. How do I get more information on what’s causing the crash? ---
**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:** [May 3, 2026, 10:19am UTC](https://discuss.vllm.ai/t/is-the-reason-for-my-vllm-0-20-0-failing-to-start-because-of-nixl/2612/4 "2026-05-03T10:19:08Z")
Yes, vLLM can run on AMD V620 with ROCm 7.2.x, as shown in [issue #38107](https://github.com/vllm-project/vllm/issues/38107). vLLM does not require exactly ROCm 7.2.0; 7.2.1 and 7.2.2 are also supported, but compatibility can depend on the specific GPU and PyTorch build ([source](https://docs.vllm.ai/en/latest/getting_started/installation/gpu/)). The “node” in `rocm-smi` output just refers to the GPU index, not NUMA nodes. To get more information about the crash, check the full logs above the “Engine core initialization failed” message for any errors or segfaults. Also, run `rocminfo | grep gfx` inside the container to confirm your GPU architecture, and try running a simple PyTorch script to verify ROCm is working. If you see a segfault or “No HIP GPUs are available,” it may be a permissions or driver mismatch issue ([source](https://github.com/vllm-project/vllm/issues/36615)). Would you like step-by-step instructions for debugging ROCm/vLLM startup issues? Sources: - [https://github.com/vllm-project/vllm/issues/38107](https://github.com/vllm-project/vllm/issues/38107) - [https://docs.vllm.ai/en/latest/getting\_started/installation/gpu/](https://docs.vllm.ai/en/latest/getting_started/installation/gpu/) - [https://github.com/vllm-project/vllm/issues/36615](https://github.com/vllm-project/vllm/issues/36615) Hint: Mention @RunLLM in the post for followups. ---
**Author:** ![diarrheavomit](https://avatars.discourse-cdn.com/v4/letter/d/8dc957/32.png) [@diarrheavomit](https://discuss.vllm.ai/u/diarrheavomit)\ **Post date:** [May 3, 2026, 2:13pm UTC](https://discuss.vllm.ai/t/is-the-reason-for-my-vllm-0-20-0-failing-to-start-because-of-nixl/2612/5 "2026-05-03T14:13:18Z")
@RunLLM python -c “import torch; print(‘ROCm:’, torch.cuda.is\_available()); print(‘GPU:’, torch.cuda.get\_device\_name(0)); print(‘Test:’, torch.ones(1).cuda())” ROCm: True GPU: AMD Radeon Pro V620 Segmentation fault Pytorch itself just doesn’t work. But rocm llama-cpp works fine. ---
**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:** [May 3, 2026, 2:13pm UTC](https://discuss.vllm.ai/t/is-the-reason-for-my-vllm-0-20-0-failing-to-start-because-of-nixl/2612/6 "2026-05-03T14:13:41Z")
Your output shows that PyTorch with ROCm detects the GPU but segfaults on `.cuda()`, indicating a PyTorch/ROCm compatibility or installation issue. This is a known problem if the PyTorch ROCm build, ROCm runtime, or driver versions are mismatched, or if the GPU architecture is not fully supported by the installed ROCm/PyTorch combination ([source](https://github.com/vllm-project/vllm/issues/36615), [source](https://github.com/vllm-project/vllm/issues/37151)). llama.cpp uses its own ROCm backend and does not depend on PyTorch, so it can work even if PyTorch+ROCm is broken. To resolve this, ensure your PyTorch ROCm wheel matches your ROCm runtime version and GPU architecture. Try reinstalling PyTorch using the official ROCm wheels for your ROCm version ([instructions](https://docs.vllm.ai/en/latest/getting_started/installation/gpu/)). Would you like detailed steps for reinstalling PyTorch ROCm or checking version compatibility? Sources: - [https://github.com/vllm-project/vllm/issues/37151](https://github.com/vllm-project/vllm/issues/37151) - [https://docs.vllm.ai/en/latest/getting\_started/installation/gpu/](https://docs.vllm.ai/en/latest/getting_started/installation/gpu/) Hint: Mention @RunLLM in the post for followups. ---
**Author:** ![diarrheavomit](https://avatars.discourse-cdn.com/v4/letter/d/8dc957/32.png) [@diarrheavomit](https://discuss.vllm.ai/u/diarrheavomit)\ **Post date:** [May 3, 2026, 2:14pm UTC](https://discuss.vllm.ai/t/is-the-reason-for-my-vllm-0-20-0-failing-to-start-because-of-nixl/2612/7 "2026-05-03T14:14:55Z")
@RunLLM Shouldn’t that not be a problem with docker images? ---
**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:** [May 3, 2026, 2:15pm UTC](https://discuss.vllm.ai/t/is-the-reason-for-my-vllm-0-20-0-failing-to-start-because-of-nixl/2612/8 "2026-05-03T14:15:14Z")
It can still be a problem with Docker images if the PyTorch ROCm wheel inside the container does not exactly match the ROCm runtime and driver version on the host, or if the container is missing required device mounts or permissions ([source](https://github.com/vllm-project/vllm/issues/36615), [source](https://github.com/vllm-project/vllm/issues/37151)). Even with official images, version mismatches or missing `/dev/kfd` and `/dev/dri` mounts can cause segfaults. Would you like a checklist for running ROCm containers with vLLM? Sources: - [https://github.com/vllm-project/vllm/issues/37151](https://github.com/vllm-project/vllm/issues/37151) Hint: Mention @RunLLM in the post for followups. ---
**Author:** ![diarrheavomit](https://avatars.discourse-cdn.com/v4/letter/d/8dc957/32.png) [@diarrheavomit](https://discuss.vllm.ai/u/diarrheavomit)\ **Post date:** [June 3, 2026, 12:34pm UTC](https://discuss.vllm.ai/t/is-the-reason-for-my-vllm-0-20-0-failing-to-start-because-of-nixl/2612/9 "2026-06-03T12:34:51Z")
@RunLLM So after a month of troubleshooting, reinstalling, buying a new server… It was SR-IOV. Disabling SR-IOV in the bios fixed all the issues. ---
**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:** [June 3, 2026, 12:35pm UTC](https://discuss.vllm.ai/t/is-the-reason-for-my-vllm-0-20-0-failing-to-start-because-of-nixl/2612/10 "2026-06-03T12:35:13Z")
It seems likely that enabling SR-IOV (Single Root I/O Virtualization) in the BIOS can cause issues with ROCm and PyTorch, leading to GPU detection problems or segfaults. Disabling SR-IOV has been reported to resolve such issues for others as well ([source](https://github.com/vllm-project/vllm/issues/36615), [source](https://github.com/vllm-project/vllm/issues/12292)). Would you like more detail on why SR-IOV interferes with ROCm or how to verify your BIOS and kernel settings? Sources: - [https://github.com/vllm-project/vllm/issues/36615](https://github.com/vllm-project/vllm/issues/36615) - [https://github.com/vllm-project/vllm/issues/12292](https://github.com/vllm-project/vllm/issues/12292) Hint: Mention @RunLLM in the post for followups.