gate-gpu-qwen4exp-e2e q38dense red since b825b640c (BF16 residual streams)
Maintainers usually reply within 1 day
Nobody has claimed this yet.
- #951 by @juliankang4 — closed without merging
Assessment
- Difficulty
- 4/5
- Estimated time
- 1-2 days
- Newbie friendliness
- 48/100
- Issue type
- Bug
- Clarity
- Mostly clear
- Activity status
- Active
- Tech stack
- rust
- Domain
- machine-learning, testing-qa
Research direction
Start with commit b825b640c and the gate-gpu-qwen4exp-e2e test for BLOOMERY_FIXTURE_VARIANT=q38dense on an RTX A6000; compare its behavior with the green commits 4be577d34 and bd9afe0c8. Inspect the Qwen3.8 residual stream precision change and the gemm arm's flip frontier. Done means the gate passes with a justified BF16 frontier update or streams retained as F32.
Written by the indexing model from the issue text.
Description
gate-gpu-qwen4exp-e2e at BLOOMERY_FIXTURE_VARIANT=q38dense (fixture tier, RTX A6000) is red on main since b825b640c ("a Qwen3.8 ubatch walk carries its residual streams as BF16 values").
Bisect on the box, same command at each commit:
- 4be577d34: green
- bd9afe0c8: green
- b825b640c: red (5 FAIL lines)
- e375cc6d5, fc1930a1f, 8208db8ab, 645c6af98, 62c507523: red, the same 5 lines (md5 b4651fd2)
The failing check is the gemm arm's flip frontier:
gemm free: last argmax ours=34727 (returned 34727) ik=not compared in this tier (the pass's [34727]) PASS
gemm free: route taps of 5 tokens, those not the top 10 of their own logits: 0 [] PASS
gemm flip layer=1 token=0: ik margin 1.707e-2; 377<-218 gap 2.996e-2 err 1.550e-1, 377<-390 gap 1.707e-2 err 5.488e-2, 132<-218 gap 5.963e-2 err 1.624e-1, 132<-390 gap 4.674e-2 err 6.229e-2 (cap 0.11)
The BF16 streams look unconditional (not only on compute capability 12), so the A6000 runs them too. Either the gate's frontier for that arm needs a derived re-pin for BF16 streams, or the streams should stay F32 where the gate expects them. The landing did not run this variant.
🤖 Generated with Claude Code
- Dominant language
- Rust
- Stars
- 21
- Forks
- 4
- Avg merge
- 2h 42m
- Merged PRs (30d)
- 692
Getting set up
- No Dockerfile or Docker Compose file
- Has a pull request template
- Read the contributing 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.
More from midagedev/bloomery
-
good first issue hardware help wanted no-box serve
Difficulty 2/5 1-3 hours Newbie friendliness 70/100
Maintainers usually reply within 1 day
-
good first issue hardware help wanted no-box
Difficulty 2/5 1-3 hours Newbie friendliness 73/100
Maintainers usually reply within 1 day
-
model new-model
Difficulty 5/5 Over a week Newbie friendliness 20/100
Maintainers usually reply within 1 day
-
gpu-evidence hardware help wanted needs-gpu
Difficulty 2/5 1-3 hours Newbie friendliness 58/100
midagedev/bloomery#926 · 2 comments ·
Maintainers usually reply within 1 day
-
Difficulty 1/5 Under an hour Newbie friendliness 10/100
midagedev/bloomery#900 · 1 comment ·
Maintainers usually reply within 1 day
All issues in midagedev/bloomery
Similar issues
-
good first issue help wanted r-docs
Difficulty 2/5 1-3 hours Newbie friendliness 72/100
servo/rust-smallvec#824 ·
Maintainers usually reply within 1 day
-
Difficulty 2/5 1-3 hours Newbie friendliness 76/100
-
Difficulty 2/5 1-3 hours Newbie friendliness 76/100
superradcompany/microsandbox#1974 ·
Maintainers usually reply within 1 day
-
Difficulty 2/5 1-3 hours Newbie friendliness 83/100
Maintainers usually reply within 1 day
-
priority middle
Difficulty 1/5 Under an hour Newbie friendliness 62/100
KATO-Hiro/AtCoderClans#12940 ·
Maintainers usually reply within 1 day