[danielhanchen]

Repins three pins onto merges of b11871 so that the scheduled prebuilt builds again on the newest aged tags. Nightlies have failed since 10-22, and the #320 release currently needs a fixed -f tag=b11849 dispatch.

pin	upstream change that broke it	first tag	fix on the pin branch	old -> new
ggml-org#25731 Inkling	ggml-org#28607 (Soprano) takes LLAMA_VOCAB_PRE_TYPE 65; ggml-org#30656 reworks the Vulkan FA accumulator precision	b11850	merge b11871; Inkling pre-type becomes 66; take upstream's FA line and add GGML_PREC_F32_PEDANTIC to the f32-accumulator cases of ggml_vk_prec_allows_f16acc	fb9ddb92d -> b49c8baf4
#320 MiniMax-M3.1	ggml-org#31097 forwards the MiniMax-M3 expert quant scales to the graph	b11861	merge b11871; upstream's scale lines kept verbatim (same lines as ours); DSparkMiniMaxDraftModel registry line additive	ef8f94ccb -> 822be707b
#298 DeepSeek-V4.1	ggml-org#31086 adds DeepSeek-V4.1 Flash under the same deepseek41 arch name	b11870	merge b11871; our deepseek41 model, converter and DSV4 KV cache kept; build_hc_mixes takes upstream's new (..., bool f32_src, int il) signature, our two call sites ported; deepseek4.cpp and dflash.cpp take upstream	be8365722 -> 14539304e

#298 cannot simply be dropped in favour of upstream's deepseek41: plain b11871 aborts while loading the published DeepSeek-V4.1-Flash UD-IQ1_M (GGML_ASSERT(*cur_backend_id != -1)).

Changes

• scripts/unsloth/pr-set.json: the three pin URLs above move to the new commits. Each new commit is a fast-forward of the old one. Order, count (23) and feature-checks.json are unchanged.

Evidence

sim --tags b11871,b11873,b11874 (window): ok on all three tags, 23 pins, pin_contract intact, merge_checks clean, ggml-org#25731 and #70 additive.

tag	CPU gate	CUDA build	test-chat	MUL_MAT	MUL_MAT_ID	PAD	GET_ROWS	features
b11871 (aged)	ok	ok	ok	2582/2582 on rerun (see note)	1027/1027	37/37	349/349	16/16 on CPU and GPU
b11874 (newest)	ok	ok	ok	2582/2582	1027/1027	37/37	349/349	16/16 on CPU and GPU

At b11871 the first MUL_MAT pass flagged one case, q4_1 x f32, m=16, n=1, k=32. That case passed 5/5 targeted reruns, and the whole op then passed 2582/2582. It is a 16x32 q4_1 matmul that none of the three repinned branches touch.

Numeric checks on the published GGUFs (wikitext-2, -c 512):

model (pin)	base build	candidate builds	PPL	mean KLD	same top-1
Inkling-Small UD-IQ1_S (ggml-org#25731), 20 chunks	current pin set @ b11849	pin set @ b11871, @ b11874	40.027352 on all three	0	100%
MiniMax-M3.1-Flash UD-IQ2_M (#320), 10 chunks	current pin set @ b11849	pin set @ b11871, @ b11874	142.527956 on all three	0	100%
DeepSeek-V4.1-Flash UD-IQ1_M (#298), 8 chunks	old #298 in the b11849 pin set	pin set @ b11871	2.097382 on both	0	100%

Against Inkling's own pre-merge head (fb9ddb92d alone), b49c8baf4 moves by KLD 0.208. All of that comes from ggml-org#31086's softplus change, which the current prebuilt already ships. fb9ddb92d plus only that change reproduces the new numbers exactly, so the released prebuilt does not move.

Not covered: ROCm, Vulkan, Metal, macOS, Windows and ARM legs (these run on the nightly).

[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	✅ Completed 2026-10-25T14:32:39.604908Z	ab53af2	PR opened
🔒 Security Review	✅ Completed 2026-10-25T14:34:14.866848Z	ab53af2	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.