mlcommons / mlcommons/inference
GPT-OSS OOM errors despite having plenty memory
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 1.6k
- Forks
- 650
- Avg merge
- 1d 22h
- Merged PRs (30d)
- 6
Description
When trying to set up the servers for GPT-OSS using run_scaleout.sh script, the following error occurs:
RuntimeError: Executor creation failed due to insufficient GPU memory.
The following component could not be created: KV cache
Total GPU memory (GiB): 276.50
Free GPU memory before component creation attempt (GiB): 200.47
Previously created components and free GPU memory before/after creation (GiB):
model: 274.64 / 208.12
sampler: 208.12 / 204.34
_no_capture_init_kv_cache: 204.27 / 202.40
spec_resource_manager: 202.40 / 202.40
drafter: 202.40 / 202.40
_no_capture_init_extra_resources: 202.40 / 197.99
model_extra: 197.99 / 200.47
We reduced the kvcache_free_gpu_mem_frac flag in the config file all the way down to .10 and still see this error. There should be plenty of free memory to use, yet we still receive OOM errors no matter what.
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start by reproducing the failure while running run_scaleout.sh and reviewing the configuration value for kvcache_free_gpu_mem_frac. Compare the reported GPU memory before KV-cache creation with the executor's allocation requirements and trace where the configuration is consumed. Done means the GPT-OSS servers start successfully without the misleading insufficient-memory failure.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 35/100