Skip to content

qwen3 0.1.0rc1: one package for every Qwen3.x LLM, on abliterated checkpoints (se-145) - #954

Draft
PaulFidika wants to merge 3 commits into
rc/0.26.0from
claude/se-145-qwen3
Draft

PaulFidika wants to merge 3 commits into
rc/0.26.0from
claude/se-145-qwen3

Conversation

@PaulFidika

Copy link
Copy Markdown
Contributor

se-145. Paul, 11-16: "can't we just have like a qwen 3.x package, and then its default is the newest qwen 3.x model we have?" "but the user can override the checkpoint used? ideally use only abliterated checkpoints of it, not the original."

For the next packages RC, not rc/0.26.0 (frozen). Retarget when the coordinator opens it; it goes in with its own packages gate (--only qwen3). Its final and the mirrors publish at promotion, by the release session.

  • qwen3 0.1.0rc1: ask, caption and the canvas write (Write, text_N), as qwen3.6 had them, on Runtime's cozy_runtime.vlm. No package-local fields.
  • Default heretic-org/qwen3.8-27b-heretic-ara@0.1.0: Heretic's ARA of Qwen3.8-27B, 0/100 refusals, KL 0.054. A request or box picks another checkpoint with model.
  • Qwen3.8-27B is qwen3_5 / Qwen3_5ForConditionalGeneration: the same 1199 tensor names and sizes as Qwen3.6-27B. Its config differs only in transformers_version. Its chat template adds reasoning_effort, which defaults to xhigh when thinking is on. The Runtime's enable_thinking switch is unchanged, so no Runtime change is needed.
  • The int8, nvfp4 and w4a8 recipes (qwen3.6's) cover dense, MoE and multi-token-prediction layers.
  • qwen3.5 and qwen3.6 are removed from the repo. Their published versions stay, get no new ones, and retire after 0.27.0. These references move to qwen3:
    • the workspace lists
    • the root README
    • qwen-image-2's enhancer --package
    • scripts/qwen3-model-proof.py
    • examples/client-scripts/qwen3-routing.py
  • Model refs stay as they are (pan2's reasoner qwen/qwen3.5-9b@0.2.0, Klein's drafter qwen/qwen3.5-9b-mtp@0.2.0, qwen-image-2's enhancers).

Cells: cozy-creator draft PR on claude/se-145-qwen3.

Proof

Pending:

  • The GPU proof is model-ingest (held mirrors with int8, nvfp4 and w4a8 lanes, plus one ./qwen3/ask call on each int8 lane) on one rtx-a6000. It waits on the rental catalog.
  • The llm-tools rows run once the cell is in the next RC.

Removed

None: scripts/removed-features.py origin/rc/0.26.0 HEAD finds 0 definitions deleted.

…ckpoints; qwen3.5 and qwen3.6 removed (se-145)

ask, caption and the canvas write, as qwen3.6 had them, over any Qwen3.5-family checkpoint (Qwen3.5, 3.6, 3.8; dense or
MoE). The default is heretic-org/qwen3.8-27b-heretic-ara@0.1.0 at int8; a request or a box picks another by model. The
int8, nvfp4 and w4a8 recipes cover dense, MoE and multi-token-prediction layers. Every package-name reference moves to
qwen3 (workspace lists, README, qwen-image-2's enhancer quantize, the proof script, the routing example).
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant