feat(models): Nail 35B-A3B at 262k with CPU vision; trim roster; acceptance harness - #19
Open
Hrt-Htk wants to merge 1 commit into
Open
feat(models): Nail 35B-A3B at 262k with CPU vision; trim roster; acceptance harness#19Hrt-Htk wants to merge 1 commit into
Hrt-Htk wants to merge 1 commit into
Conversation
…acceptance harness Add Nail-Qwen3.6-35B-A3B-UD-Q4_K_XL as `nail-35b-a3b-q4-256k` — full 262k context on the 3090 Ti with the projector on CPU (mmproj-offload) and its froggeric chat template. - ModelChoice gains mmproj_offload, chat_template_file, and per-model ctx_choices; chat templates now emitted per-preset. A router-global --chat-template-file was silently overriding each preset's template, denying Nail its froggeric template. - Roster trimmed to two 262k models: nail-35b-a3b-q4-256k and qwen3.6-27b-q4-mtp-224k (MTP capped at 224k; draft buffers leave only ~300 MiB headroom at 262k). Retired presets' weights remain on disk. - Add reusable live acceptance harness (tests/model_acceptance.py) + plan (docs/dev/model-acceptance.md); validated on 35B/27B/Nail. Logs pp/tg from llama-server timings (prompt_per_second / predicted_per_second). #18 Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #18.
What
Add Nail-Qwen3.6-35B-A3B (
nail-35b-a3b-q4-256k) as a production model — full 262k context on the 3090 Ti with the vision projector on CPU (mmproj-offload) and its froggeric chat template. Trim the roster to two 262k models and add a reusable model-acceptance harness.Changes
ModelChoicegainsmmproj_offload,chat_template_file, and per-modelctx_choices. Chat templates are now emitted per-preset — a router-global--chat-template-filewas silently overriding each preset's template, denying Nail its froggeric template (caught via/apply-template).nail-35b-a3b-q4-256k— 262k, CPU vision, froggeric template (default)qwen3.6-27b-q4-mtp-224k— MTP capped at 224k; draft buffers leave only ~300 MiB headroom at 262ktests/model_acceptance.py) + plan (docs/dev/model-acceptance.md): live gate covering load, VRAM fit, streaming/think-integrity (issue SSEChunkLogger rescue state machine splits reasoning_content on literal </thinking> text #8), tool-calling, long-context needle, vision (CPU mmproj), and pp/tg from llama-servertimings. Validated on 35B/27B/Nail.Verification
/apply-template(froggeric markers present for Nail, unchanged for others).🤖 Generated with Claude Code