10. Troubleshooting

Symptom Cause / fix
No matching distribution / is not a supported wheel the notebook is not Python 3.13, or Internet is OFF, or an old install line with the retired index address is used — use the pinned command in Getting started
THESE PACKAGES DO NOT MATCH THE HASHES the downloaded file is not the one the command names; do not use it. Retry; if it persists, report it with the full log
pip wants to change torch / transformers should never happen; report it with the full pip install log
Engine start fails, no GPU Accelerator is not GPU T4 × 2
ProhibitedBackendError (an ImportError raised by the bundled vLLM) or ConfigurationError mentioning Flash a Flash backend was requested; only TRITON_ATTN exists
bfloat16 error use dtype="half"
Out of memory at start-up lower max_model_len (e.g. 2048), lower gpu_memory_utilization, or use tensor_parallel_size=2
Out of memory after an earlier experiment the previous engine is still alive: close() it, gc.collect(), check torch.cuda.mem_get_info(i)
Engine code hangs or re-runs in a script add if __name__ == "__main__": (vLLM spawns processes that re-import the script)
Triton kernel JIT compilation during inference one-time warm-up warning on the first request; harmless
unauthenticated requests to the HF Hub informational; add an HF_TOKEN Kaggle secret for faster downloads
ProcessGroupNCCL … unbatched P2P op (PP=2) harmless warning
UNKNOWN_REQUIRES_EVIDENCE from check_model that exact model id was never run on Kaggle (or its architecture could not be determined: pass architectures=[...]). It is not an error — see Checks
Process manager: force killing remaining processes at shutdown seen after every TP=2/PP=2 engine in the Kaggle runs; the GPUs were fully released each time (see The LLM engine)
StructuredOutputError: truncated bound the schema fields and/or raise max_tokens
ServerStartupError read e.log_excerpt and e.log_path; a second server on the same port also causes this
Slow with TP=2 / PP=2 expected for a model that fits one T4 (PCIe, no NVLink); use TP=1 or ReplicaPool
Disk full on /kaggle/working quota is 20 GiB; check stokastisk.working_dir_status()

To start clean: llm.close(), then restart the Kaggle session.

Report problems at https://github.com/stokastisk-ai/stokastisk/issues with the output of stokastisk.report().