unsloth: pin #241 with the CUDA graph cache cap removal, repin Inkling and #61 onto b11529

[danielhanchen]

Pins #241 again in place of #253. The built tree does not change: both point at the same commit, 20dd977.

The 64-entry CUDA graph cache cap that #253 removes came from #241's own commit, ggml-cuda: key the CUDA graph cache by shape. So the fix belongs in #241:

• Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream #241's branch was fast-forwarded to 20dd977. That is Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream #241's previous head 9dd7972 plus the cap removal.
• Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream #241 is now +48/-17 in 4 files against upstream.
• ggml-cuda: drop the CUDA graph cache count cap, as upstream does #253's commit is now part of Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream #241, so ggml-cuda: drop the CUDA graph cache count cap, as upstream does #253 shows as merged.

feature-checks.json moves the unchecked entry back to #241, with the cap noted in its reason.

Verification

• Merge loop plus pin_contract.py on b11505 (tonight's base): all 15 pins merge, 2 through the additive merge script. All 15 are intact, and Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream #241 accounts for its 22 lines.
• The feature-checks.json coverage lint passes: 15 pins, 7 with a feature check, 8 knowingly unchecked.

Inkling repin for b11515

Today's base, b11515, includes ggml-org#30147. That PR loops the CUDA PAD kernel over rows and slices past 65535 and rewrote the line Inkling's int64 dst_idx fix touches, so ggml-org#25731 stopped merging. Its branch now merges b11515 and keeps the casts inside the new loop. Upstream's version still computes the index in int.

• test-backend-ops -o PAD: 37/37 on CUDA.
• All 15 pins merge onto b11515, Inkling and kimi-k3 : the MoonViT-3d vision tower and full-size loading fixes #70 through the additive merge script.
• pin_contract.py: all 15 pins intact. merge_checks.py: clean.

#61 repin for b11517

b11517 brings ggml-org#30168. It makes the src1 precision an explicit template argument of the CUDA MMQ helpers and drops its GGML_PREC_Q8 default. #61 broke in three places:

• the mmq.cuh switch cases, a text conflict;
• the three IQ1 tile loaders, which merged cleanly but no longer compiled;
• mul_mat_q_case in mmq.cu, which also merged cleanly but no longer compiled.

The CPU-only dry run cannot see the last two. #61 now merges b11517 and passes GGML_PREC_Q8 in all three places, the same value the default had and the same thing upstream does for IQ1_S. Its PR base moves to b11517 as well.

• CUDA test-backend-ops, all passing:
  • MUL_MAT: 2483/2483 (39 iq1_x* cases, batch sizes up to 64);
  • MUL_MAT_ID: 940/940 (9 iq1_x* cases);
  • GET_ROWS: 216/216 (24 iq1_x* cases).
• On b11517:
  • all 15 pins merge;
  • pin_contract.py: all intact;
  • merge_checks.py: clean;
  • the CPU gate build passes;
  • feature_matrix.py: 7/7 features.

Inkling repin for b11529

b11529 brings ggml-org#28636, which moves the ModelOpt FP8 dequant loop in conversion/base.py into a dequant_fp8() helper. Inkling's stacked-expert NVFP4 rejection sat in front of that loop and now sits in front of the call. Inkling's copy of the old loop was identical to the old upstream copy before the merge, so no behaviour is lost.

Results on b11529:

• Merge, pin_contract.py and merge_checks.py: all 15 pins merge, all are intact, and the merged tree is clean. The same holds on b11517.
• Converter: a tiny Qwen2 converts to a byte-identical GGUF on upstream b11529 and on the merged Inkling branch.
• Build: the CPU gate and the full CUDA build both pass.
• CUDA test-backend-ops: MUL_MAT 2483/2483, MUL_MAT_ID 940/940, PAD 37/37 and GET_ROWS 216/216.
• feature_matrix.py --gpu on the CUDA build: all 7 features pass on GPU, including the checks CI skips (iq1 grids 3 x 13/13, glm5next 192/192, Inkling 13/13).

[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 2026-10-09T12:15:14.549697Z	40ac0cb	PR opened
🔒 Security Review	✅ Completed 2026-10-09T12:16:50.271808Z	40ac0cb	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.