[danielhanchen]

Changes

1. Repin model : DeepSeek-V4.1-Flash-1017 full stack pin (all DeepSeek-V4.1 PRs except #292) #298 (DeepSeek-V4.1-Flash-1017 full stack) from 9d907b7 to d1781c3.
   • The new head carries the model : deepseek41 DSpark speculative draft #296 fix that lets --fit measure the DSpark draft model. Before it, --fit sized the target as if it ran alone, and the draft then failed cudaMalloc.
   • On UD-IQ1_S with the DSpark Q8_0 draft and a 12 GB budget, the old head left the draft 8,218 MiB over budget. The new head measures the draft at 8,332 MiB (7,602 MiB model, 4.5 MiB KV, 725.5 MiB compute), which matches the real allocation, and stays within budget.
   • d1781c3 is 9d907b7 plus the model : deepseek41 DSpark speculative draft #296 fix merged up the stack (+100 lines in 5 files), nothing else.
2. Pin chat : end Beam analysis on <|constrain|> header variants #311 (Reflection Beam: end the analysis message on <|constrain|> header variants in the harmony-xml parser). It has no feature surface of its own; test-chat covers it, so it is listed under unchecked in feature-checks.json.
3. Repin CUDA: faster MoE prompt processing (compact MMQ tiling, device expert sort, GET_ROWS) #264 (CUDA MoE prompt processing) from 5bb6845 to 194170f: adds the cuBLAS MoE fallback with host sort for repeated expert ids, stops the ids counting sort on out of range ids, and merges b11627.
4. Repin llama-mmap: populate the non-lazy ranges when MAP_POPULATE is skipped for lazy tensors #266 (llama-mmap populate for lazy tensors) from 88035b0 to 1eae0b5: clamps the populate budget to the cgroup memory limits, samples it before mapping, and decodes octal escapes in mountinfo cgroup paths.
5. Pin CUDA: add IQ1_M to MMQ #304 (CUDA: add IQ1_M to MMQ), with a new iq1m-mmq feature check owned by CUDA: add IQ1_M to MMQ #304 (the IQ1_M MUL_MAT cases with m = 129 and the MUL_MAT_ID cases with 256/512 experts that CUDA: add IQ1_M to MMQ #304 adds; both match 0 cases on a tree without the pin). This supersedes pr-set: pin CUDA IQ1_M MMQ (#304) #305.
6. Pin Model arch: Flint #314 (Model arch: Flint), with an arch feature check for flint owned by Model arch: Flint #314. Model arch: Flint #314 numbers the Flint pre-tokenizer 63 so it does not collide with Inkling's 62.
7. Repin ggml-org#25731 (Inkling) to e1b89cf. Inkling at 534c321 no longer merges on b11785 or b11796, which is why the nightly has failed since 10-20; e1b89cf is the upstream PR after Fix merge conflicts.

Evidence

llama_pinset.py with this pr-set and Inkling at e1b89cf:

check	b11785	b11796
sim: merge loop	22 pins; Inkling, #70, #298 additive, rest clean	same
pin_contract	all 22 pins intact	all 22 pins intact
merge_checks	clean	clean

• audit: 21/22 pins clean. The remaining row is model : DeepSeek-V4.1-Flash-1017 full stack pin (all DeepSeek-V4.1 PRs except #292) #298, whose base dsv41-pin-base61b carries the earlier DeepSeek-V4.1 stack; this is unchanged from master.
• Without the Inkling repin the same sim refuses Inkling in 8 files (common/parsers, ggml-cuda.cu, llama-arch, llama-model-saver, test-llama-archs, mtmd-audio.h); every other pin merges.
• The Beam pin: test-chat on b11785 fails with the old parser and passes with chat : end Beam analysis on <|constrain|> header variants #311.
• With Model arch: Flint #314 and the Inkling repin: sim on b11786 merges all 23 pins (Model arch: Flint #314 additive), pin_contract intact, merge_checks clean. CUDA verify on b11796 and the Inkling numeric gate (old head vs e1b89cf vs e1b89cf + master) were still running at merge time.

Running now: verify --cuda on b11785 (full CUDA build, test-backend-ops, test-chat, feature matrix including iq1m-mmq). ROCm, Vulkan, macOS, Windows and ARM legs run only in 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 2026-10-21T12:56:15.224419Z	ede7f68	PR opened
🔒 Security Review	⚠️ Failed 2026-10-21T12:53:31.285809Z We couldn’t complete this review. Comment @codex security review to try again.	ede7f68	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.

[chatgpt-codex-connector]

The account paying for this security review has reached its Codex usage limits. The payer can check the Codex usage dashboard. For personal accounts, using credits requires enabling “Use credits for security reviews” in Code review settings. If you do not manage the paying account, contact this repository's admins.