Repin ggml-org#24423, ggml-org#25731, unslothai#158, unslothai#149 onto b11609, drop unslothai#152

[danielhanchen]

Repins four pins and drops one so the nightly merges again on the current base tag.

The scheduled nightly (run 38462794625) stopped in Resolve tag with a merge conflict in examples/diffusion/diffusion-cli.cpp. No pin moved. The base moved from b11541 to b11609, and three upstream changes in that range break five pins:

pin	upstream change that broke it	fix	new SHA
ggml-org#24423	ggml-org#30389 (b11607) turned common_chat_params::prompt into a common_chat_prompt and edited format_input_text(), which the pin deletes	keep the deletion, port the two .prompt uses to .prompt.str()	6a7388aac74ab896d9d4f0fb187ddac4847c02f0
ggml-org#25731	ggml-org#30389: the text merges, then common/parsers/inkling.cpp:48 fails to compile (operator+= on common_chat_prompt)	data.prompt.append(data.generation_prompt), the same port upstream made in kimi-k2.cpp	607b193a546bf7ebd17cca1c3e355c622fcef605
#158	ggml-org#30327 (b11592) reads the CUDA env vars once	keep upstream's ggml_cuda_pinned_memory_disabled() line, the pin's helper now wraps it	fb9b9462041ad58cde9a274e7984c3c6fb01b279
#149	ggml-org#30327: GGML_CUDA_ENABLE_UNIFIED_MEMORY=0 is parsed upstream now	take upstream's ggml_cuda_unified_memory_enabled() at both sites, keep only the once-per-process HIP warning	30d39d74fd9cad3224c30bdda3a9def6c9aa6719
#152	ggml-org#30405 (b11599) is the same fix (one buffer per file range for the Qwen3.8-Flash-Next per-layer embedding)	dropped, upstream carries it	removed

Each fix is a --no-ff merge of b11609 into the pin's own branch (fast-forward pushes, nothing rewritten). The two ggml-org PRs are mergeable with ggml-org master again. #158 and #149 now target base/upstream-fcac19079, so their PR diffs are the pin's own change again (+24/-1 and +19/-0).

One behaviour change comes from following upstream in #149: GGML_CUDA_ENABLE_UNIFIED_MEMORY=true, yes and on are now read as off, because upstream parses the value with atoi.

Changes

• scripts/unsloth/pr-set.json: repins ggml-org#24423, ggml-org#25731, ggml-cuda: avoid direct ROCm_Host compute on HIP integrated GPUs (port of ggml-org#25863) #158 and HIP: say once that GGML_CUDA_ENABLE_UNIFIED_MEMORY is on #149 to the SHAs above and removes llama: map each contiguous run of a context's tensors, not one span over all of them #152.
• scripts/unsloth/feature-checks.json: removes unslothai#152 from unchecked and rewrites the unslothai#149 reason (the pin is now only the HIP warning).

Checks

check	b11609 (aged)	b11612 (newest)
merge loop + pin_contract + merge_checks	ok, 15 pins, all intact, clean	ok, 15 pins, all intact, clean
audit, 15 pins	15/15 clean, feature-checks coverage ok
CPU gate (llama, test-llama-archs, test-backend-ops, test-mtmd-impl)	ok	ok
full CUDA build (sm80)	ok	ok
test-chat	all passed	all passed
test-backend-ops MUL_MAT / PAD / GET_ROWS	2488/2488, 37/37, 216/216	2488/2488, 37/37, 216/216
test-backend-ops MUL_MAT_ID	940/940 without the dup=1 cases	940/940 without the dup=1 cases
feature_matrix on GPU	all 7 features demonstrated	all 7 features demonstrated

The MUL_MAT_ID dup=1 cases crash the same way on the plain upstream tag (4 f16 n_mats=28 cases, then an illegal memory access in mmq.cuh); they come from ggml-org#29998, so they are not caused by the pins.

Numeric gate for the two model pins (old head A, new head B, new head merged with ggml-org master C):

• DiffusionGemma: all six runs (A, B, C, twice each) give byte-identical output.
• Inkling: B and C give a perplexity of 39.256712, the same as A against itself, and every line of their KLD reports is identical to the A/A report.

Not covered

• ROCm/HIP: ggml-cuda: avoid direct ROCm_Host compute on HIP integrated GPUs (port of ggml-org#25863) #158 and HIP: say once that GGML_CUDA_ENABLE_UNIFIED_MEMORY is on #149 add HIP-only code, and no HIP host was available.
• Only sm80 was built for the CUDA checks; Vulkan, macOS, Windows and ARM run in the nightly itself.
• Tags newer than b11612.

[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 2026-10-14T11:14:21.932515Z	a995c21	PR opened
🔒 Security Review	✅ Completed 2026-10-14T11:14:57.955191Z	a995c21	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.