Repin DiffusionGemma to the commit that merges cleanly

[danielhanchen]

Why

The nightly stopped on the merge step in run 34907376192, before any compiling:

CONFLICT (content): Merge conflict in src/llama-model-saver.cpp
CONFLICT (content): Merge conflict in tests/test-llama-archs.cpp
refused  tests/test-llama-archs.cpp: merge base is not empty, so at least one side edited existing text
ggml-org/llama.cpp#24423 (168b21ad) does not merge cleanly onto b10964 + the PRs listed before it

Upstream landed LLM_ARCH_MAPLE in exactly the two places DiffusionGemma extends: the unsupported-arch list in llama_model_saver_supports_arch, and the alternating SWA-pattern branch in test-llama-archs.cpp. Both are add/add in meaning, but in the test the closing ) { moved, so both sides edited one line and additive_merge.py refused it. That refusal is correct behaviour, not a bug.

What changed

12e0a962 is ggml-org#24423 merged with current upstream master, with both conflicts resolved by keeping both arches:

        case LLM_ARCH_MAPLE:
        case LLM_ARCH_DIFFUSION_GEMMA: // same: a per-layer SWA pattern add_kv_from_model cannot write back
            return false;

Only the pin line moves here; no build logic is touched.

Verified

Replayed the resolve merge loop from unsloth-prebuilt.yml locally against b10988, the tag tonight's run would select:

• ggml-org#24423 at the new pin merges clean
• the additive path still works where it should: ggml-org#25731 (4 files) and kimi-k3 : the MoonViT-3d vision tower and full-size loading fixes #70 (1 file) both hit conflicts and additive_merge.py resolved them, keeping both sides
• scripts/unsloth/test_additive_merge.py: 0 failures
• the new pin is in the PR's commit list, so unsloth-pr-set-lint.yml passes
• on the PR branch itself: full CPU build exit 0 including all four llama-diffusion-gemma-* targets, test-llama-archs -a diffusion-gemma OK at NMSE 0.00e+00, -a maple OK, and the full 312 row suite green

This unblocks the merge step but not the whole release

The same replay found two further blockers behind this one, which the failed run never reached because it stopped on the first pin. Neither is caused by this change and neither is fixed here:

1. ggml-org#27754 (GLM-5-Next) no longer merges. Upstream hoisted the src/CMakeLists.txt source list into set(LLAMA_CORE_SOURCES ...), and the pin still adds llama-kv-cache-kpool.cpp to the inline list that no longer exists. There is also a one-line llama-graph.cpp conflict, MAPLE against GLM5NEXT in the same condition. GitHub reports the PR as mergeable: false against upstream master too, and it is pinned at its current head, so there is no newer commit to repin to. It needs a rebase upstream, or the pin dropped.
2. llama: batched readahead for lazily read gather tables #137 and MTP for Qwen3.8-Flash-Next #144 disagree on src/models/qwen4exp.cpp. Three conflicting hunks over the hyper-connection tensors: llama: batched readahead for lazily read gather tables #137 has hc_head_norm as {n_embd, hc} with TENSOR_ALLOW_RESHAPE, MTP for Qwen3.8-Flash-Next #144 has {hc_dim} with trunk_flags. That is a real disagreement about tensor shape, not an add/add, so nothing should auto-resolve it. One of the two needs to be updated to match the other.

[chatgpt-codex-connector]

You have reached your Codex usage limits for security reviews. Please try again later.

[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 2026-09-16T01:29:33.765967Z	b884607	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.