Heads-up: V4.1 layers 0 and 1 onto the cards (fused layer with a Q5_K down)
Maintainers usually reply within 1 day
Nobody has claimed this yet.
Assessment
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Newbie friendliness
- 10/100
- Issue type
- Feature
- Clarity
- Clearly specified
- Activity status
- Active
- Tech stack
- rust
- Domain
- backend, performance
Research direction
This issue announces a change already landing and explicitly says there is nothing to do now. Read crates/placement/src/accepts.rs, crates/gpu-deepseek41/src/chain/ffn.rs, chain/ffn/batch.rs, chain/ffn/tier.rs, and crates/gpu/src/host/swap_source.rs to understand the affected paths; gate-fixture-budget identifies the fixture that needs regenerating. The reported completion evidence is identical scan bins and the stated A6000 benchmark results.
Written by the indexing model from the issue text.
Description
@juliankang4 a heads-up before a landing that touches files you changed this week. Nothing for you to do now.
What: V4.1's layers 0 and 1 (Q3_K gate and up, Q5_K down) get card experts. The fused layer reads its down by type: Q4_K as before, or Q5_K through the existing q5k_gemv_sel and q5k_gemv_tiles entries. No device code changes; narrow-scan shows every scan bin identical.
Files: crates/placement/src/accepts.rs (ds41_pair admits a Q5_K down on a stage card; a tier keeps the Q4_K triple), crates/gpu-deepseek41/src/chain/ffn.rs, chain/ffn/batch.rs, chain/ffn/tier.rs, and crates/gpu/src/host/swap_source.rs.
One common change: Convert gains converts(part), with a default of every part. one_layout now compares only the parts a convert step rewrites, so V4.1's r8 unpack (gate and up) accepts a layer whose down differs. Any other Convert impl keeps the old rule.
One effect on your machine: after you pull, your V4.1 fixture's recorded card budget was chosen under the old rule, so gate-fixture-budget will name it. Regenerating the V4.1 fixture fixes it. I am happy to send you the new budget line or do anything else that helps.
Measured: A6000, plan (a), depth 512, 6 rounds, same lease: decode prose +2.4 % ± 2.9, code +4.3 % ± 3.1.
If this collides with something you have locally, tell me and I will rebase my side.
(Written by an AI coding agent.)
- Dominant language
- Rust
- Stars
- 21
- Forks
- 4
- Avg merge
- 2h 34m
- Merged PRs (30d)
- 836
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
midagedev/bloomery#863 · 4 comments ·
Maintainers usually reply within 1 day
-
Difficulty 5/5 Over a week Newbie friendliness 25/100
midagedev/bloomery#1163 · 2 comments ·
Maintainers usually reply within 1 day
-
Difficulty 5/5 Over a week Newbie friendliness 28/100
midagedev/bloomery#1162 · 8 comments ·
Maintainers usually reply within 1 day
-
Difficulty 4/5 1-2 days Newbie friendliness 25/100
Maintainers usually reply within 1 day
All issues in midagedev/bloomery
Similar issues
-
Difficulty 2/5 1-3 hours Newbie friendliness 68/100
Maintainers usually reply within 1 day
-
bug
Difficulty 2/5 1-3 hours Newbie friendliness 66/100
-
Difficulty 2/5 1-3 hours Newbie friendliness 82/100
aosama/astronomical#1973 ·
Maintainers usually reply within 1 day
-
bug triage
Difficulty 2/5 1-3 hours Newbie friendliness 84/100
Maintainers usually reply within 6 days
-
Difficulty 2/5 1-3 hours Newbie friendliness 76/100
voidzero-dev/oxc-angular-compiler#686 ·
Maintainers usually reply within 1 day