[danielhanchen]

Adds #298 (DeepSeek-V4.1-Flash-1017: model, MXFP8 type, runtime, chat parser, perf, DSpark draft, vision; everything from #289 to #297 except #292) to the nightly pin set, pinned at 9d907b7.

How the prebuilt uses this: release = an aged upstream ggml-org b#### tag plus the pins in scripts/unsloth/pr-set.json, merged in order (git merge, then additive_merge.py for pure add/add conflicts only). Nothing is carried in fork master; once upstream takes the model, the pin is deleted.

Ordering and the #61 base

#61 (IQ1_XS / IQ1_XXS / IQ1_XXXS, types 64..66) and #298 (MXFP8, type 43) both add a ggml_type and both edit GGML_TYPE_COUNT = 43,, which additive_merge.py correctly refuses. #298 is therefore based on dsv41-pin-base61b (the pinned #61 commit 05a6693 plus upstream 797de63b2) and carries the combined enum (MXFP8 = 43, IQ1_XS..IQ1_XXXS = 64..66, GGML_TYPE_COUNT = 67). It must stay after #61 in this list. MXFP8 stays at 43 because the published DeepSeek-V4.1 GGUFs store the n-gram table as type 43.

llama_pinset.py audit flags this as a FIX row (base carries #61's source files, so the pin's own change counts #61 as well). That is the stacking above, not a stale base. If #61 is ever re-pinned or dropped, #298 has to be updated with it.

Feature checks

New deepseek41 entry owned by unslothai#298:

• arch deepseek41 (test-llama-archs)
• backend-op GET_ROWS with type=mxfp8 (MXFP8 is gather-only)
• mtmd projector deepseek41v

Evidence

llama_pinset.py sim with this pr-set.json, window of tags:

tag	#298	other pins	pin_contract	merge_checks
b11724	additive	unchanged (ggml-org#25731, #70 additive; rest clean)	all 20 intact	clean
b11727	additive	unchanged	all 20 intact	clean
b11729	additive	unchanged	all 20 intact	clean

pin_contract for #298: 46 files with new symbols, 4462 lines accounted for, on every tag.

audit: feature-checks coverage ok. Pre-existing FIX rows for #264 and #266 (pin != PR head) are unrelated to this change.

verify --tag aged --cuda (b11724) and verify --tag newest --cuda (b11729), same result on both:

• CPU gate build (llama, test-llama-archs, test-backend-ops, test-mtmd-impl): ok
• full CUDA build: ok
• test-chat: all tests passed
• test-backend-ops: MUL_MAT 2484/2484, MUL_MAT_ID 993/993, PAD 37/37, GET_ROWS 349/349

feature_matrix --only deepseek41 --gpu on the b11729 merged tree: 3/3 device rows decoded and matched CPU, 9/9 MXFP8 GET_ROWS cases matched the CPU reference, projector registry intact over 221 assertions.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	⚠️ Failed 2026-10-19T12:48:42.107746Z	27f21f7	PR opened
🔒 Security Review	✅ Completed 2026-10-19T12:45:33.259062Z	27f21f7	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.