Hello,
When building the Docker image for vLLM 0.15.0, I’m using PyTorch 2.10.0. Is it possible that vLLM does not support PyTorch 2.10.0 yet?
Balaxxe
February 10, 2026, 4:47am
2
I made a bug report here:
opened 12:35AM - 10 Feb 26 UTC
**Description:**
Distributed TP inference with gpt-oss-120b hangs after 1-2 req… uests on the new 2026-02-09 build (PyTorch 2.10 + Triton 3.6.0). The hang is specific to FULL CUDA graph capture mode.
**Steps to reproduce:**
```bash
./launch-cluster.sh -t vllm-node \
-n <head_ip>,<worker_ip> \
--ib-if <head_ib>,<worker_ib> \
exec vllm serve openai/gpt-oss-120b \
--port 8000 --host 0.0.0.0 \
--gpu-memory-utilization 0.70 \
-tp 2 --distributed-executor-backend ray \
--max-model-len 131072 \
--no-enable-prefix-caching \
--max-num-batched-tokens 8192
```
Send 2-3 chat completion requests. The first 1-2 succeed, then the model freezes mid-generation on a subsequent request. Throughput drops to 0 with 1 req still showing as running. After 300s it crashes with `RayChannelTimeoutError`.
**What I tested:**
- `--enforce-eager` → stable
- `--compilation-config '{"cudagraph_mode":"none"}'` → stable
- `--compilation-config '{"cudagraph_mode":"piecewise"}'` → stable
- `--compilation-config '{"cudagraph_mode":"full"}'` → hangs after 1-2 requests
- Default (`FULL_AND_PIECEWISE`) → hangs after 1-2 requests
The FULL portion of CUDA graph capture is what breaks it. Piecewise alone is fine.
**Environment:**
- 2x DGX Spark (GB10, sm_12x, aarch64)
- TP=2 across nodes via InfiniBand
- DGX OS 7.4.0, kernel 6.17.0-1008-nvidia
- vllm-node image built 2026-02-09 (PyTorch 2.10, Triton 3.6.0)
- vLLM 0.15.2rc1.dev135+g285bab475
**Previous build behavior:**
This did not happen on the previous build. The only changes in the 2026-02-09 build were PyTorch 2.10, Triton 3.6.0, and removal of the fastsafetensors patch.
**Workaround:**
`--compilation-config '{"cudagraph_mode":"piecewise"}'`
---
It’s something with vLLM or pytorch.
Have not had a chance to fully debug but I do provide a temp workaround.
Will get to it soon.
eugr
February 10, 2026, 7:06am
3
I replied in the ticket, but have you used a wheels build by the chance?--use-wheels flag).
I’m going to disable wheels builds for now.