unsloth: pin #253 (no CUDA graph cache count cap) in place of #241

[danielhanchen]

Pins #253 (no count cap on the CUDA graph cache) in place of #241.

#253 is #241 plus one commit. That commit removes the 64 entry LRU cap that the shape keyed graph cache from #241 added. With the cap, tensor split re-captures graphs on every token, because it needs about 2 * n_layers + 1 graphs per device. The result is 2.4x to 4.3x slower decode than the official build (unslothai/unsloth#12468). Upstream dropped the same cap in review of ggml-org#21611.

Why replace #241 rather than add #253 after it

• ggml-cuda: drop the CUDA graph cache count cap, as upstream does #253's branch was behind Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream #241. It was stacked on Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream #241's old head b96a713, which still had the NextN graph that upstream replaced with ggml-org#29928. I merged Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream #241's current head 9dd7972 into it. ggml-cuda: drop the CUDA graph cache count cap, as upstream does #253 is now 20dd977: Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream #241 plus the cap removal (+5/-14 in common.cuh), nothing else.
• Pinned after Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream #241, it merges cleanly but fails pin_contract. ggml-cuda: drop the CUDA graph cache count cap, as upstream does #253 deletes the cap lines Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream #241 adds, so the contract reports Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream #241 as not intact (common.cuh kept 4/10). Pinning ggml-cuda: drop the CUDA graph cache count cap, as upstream does #253 alone carries every Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream #241 item and the fix, and the contract passes.
• feature-checks.json: Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream #241's unchecked entry moves to ggml-cuda: drop the CUDA graph cache count cap, as upstream does #253, with the cap removal added to the reason. It needs a multi-GPU CUDA host, and no runner has one.

Verification

Merge loop (diff3, then additive_merge.py, as unsloth-prebuilt.yml runs it) plus pin_contract.py:

Base	Merge	pin_contract
b11493	15/15, 2 additive (Inkling, #70)	all 15 intact
b11500	15/15, 2 additive	all 15 intact
upstream master c811cb8f0	15/15, 2 additive	all 15 intact

The feature-checks.json coverage lint passes: 15 pins, 7 with a feature check, 8 knowingly unchecked.

[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 2026-10-08T15:46:56.932164Z	3d974a0	PR opened
🔒 Security Review	✅ Completed 2026-10-08T15:49:57.889754Z	3d974a0	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.