unsloth: pin #253 in place of #241 to drop the CUDA graph cache cap (stacked on #252)

[danielhanchen]

Stacked on #252. Takes #252's pin set and puts #253 in #241's slot, so the nightly ships #241 without the 64 entry CUDA graph cache cap behind the --split-mode tensor decode regression in unslothai/unsloth#12468 (2.4x to 4.3x slower than the official ggml-org build on 2x RTX 5070 Ti, 1.57x on 2x B200). Against #252 this is a 2 line change.

Why #253 replaces #241 instead of following it

#253's head (20dd977ea) is #241's head 9dd797259, which #252 pins, merged with the one commit that removes the cap. Its tree is #241's with only ggml/src/ggml-cuda/common.cuh changed (+5 -14), so it carries everything #241's pin does.

Listing #253 after #241 cannot work: pin_contract.py checks that every line a pin adds is still in the merged tree, so #241 would fail on the removed max_cuda_graphs line, and unsloth-prebuilt.yml exits on that check. Taking #241's slot keeps the order and the composition, minus the cap.

feature-checks.json moves #241's unchecked reason to #253, noting that the cap only shows with --split-mode tensor on two GPUs.

Pin set merges additively

Replayed locally the way unsloth-prebuilt.yml does it (diff3 merge per pin, then scripts/unsloth/additive_merge.py on a conflict), onto b11491, the base the preflight picked today:

	#252	this PR
pins that merge	15 of 15	15 of 15
resolved by additive_merge.py	ggml-org#25731, #70	ggml-org#25731, #70
merge_checks.py	clean	clean
pin_contract.py	all 15 intact	all 15 intact
max_cuda_graphs lines in the merged common.cuh	2	0
shape keyed ggml_cuda_graph_get_key	present	present

• The merged tree passes the CPU compile gate (llama and mtmd) and builds llama-bench with CUDA (sm_100).
• On the merged tree's own llama-bench (Qwen3-4B Q4_K_M, -sm tensor, 2x B200, shared GPUs on a loaded host, so this shows engagement rather than a speed figure), CUDA graphs beat GGML_CUDA_DISABLE_GRAPHS=1 in all 3 rounds: 150 vs 85, 194 vs 30, 93 vs 50 tok/s. With the cap the two were equal.
• The pin coverage lint passes (15 pins, 7 checked, 8 knowingly unchecked).
• The replay was checked against a known result first: the pin set of the b11443-mix-d65395f release, replayed onto b11443, merges all 15 pins (Add TML Inkling architecture ggml-org/llama.cpp#25731, kimi-k3 : the MoonViT-3d vision tower and full-size loading fixes #70 and EmbeddingGemma-2 support #247 additively), as that nightly did.

The preflight run on this branch before it was stacked failed at ggml-org#25731 on b11491. That is the Inkling conflict #252 fixes, and it never reached #253.

Order

Merge #252 first, then this. If #252 changes again, this needs only the #241 line re-swapped to a #253 head that contains #241's new head.

[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 2026-10-08T11:43:45.867573Z	1654d1f	PR opened
🔒 Security Review	✅ Completed 2026-10-08T11:45:39.241135Z	1654d1f	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.