Q8_0 Qwen 27B decode is ~85x below isolated GEMV bandwidth
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 33.6k
- Forks
- 4.3k
- Avg merge
- 3h 11m
- Merged PRs (30d)
- 491
Description
The tinygrad engine decodes tokens from Qwen3.8-27B at 1 < tok/s bs=1 (19.4 GB/s streaming).
Measured bandwidth references on the same idle GPU (GH200) is around 2.7 TB/s for a plain f16 stream-add, and 1.6 TB/s for an M=1 GEMV against the biggest layer (lm_head, 5120×248320).
The integrated decode is around 85× below what the GEMV kernels achieve in isolation, with the Q8_0→f16 expansion redone inside every kernel launch of every step.
Reference ~50/s on a (probably very unoptimized) ollama setup
Contributor guide
No contributing guide indexed for this repository
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
No source file or test is named. Start by reproducing the Qwen3.8-27B bs=1 decode measurement on the GH200, then trace the integrated decode kernels and the repeated Q8_0→f16 expansion described in the report. Done means substantially narrowing the gap between integrated decode bandwidth and the isolated GEMV reference.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- performance
- Issue type
- Bug
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Active
- Clarity
- Mostly clear
- Newbie friendliness
- 38/100