Repin qwen4exp (ggml-org#27742) to pick up the nextn draft head

[danielhanchen]

Second qwen4exp repin today. #122 pinned c4ddc4805; the carry has moved on four commits since, and the upstream PR has gone from 27 to 30. The base tag is unchanged at b10639, so unlike #122 this is purely a content refresh.

scripts/unsloth/pr-set.json, one line: #114 moves from c4ddc4805 to 92ee7a5ce.

What is in it

Three of the four new carry commits correspond one to one with the three new upstream commits:

• qwen4exp: include llama-impl.h explicitly for llama_mul_mat_hadamard
• llama: fix the qwen4exp PLE history seq_rm(-1) iterator invalidation and the fatal-warning build
• llama: segment the qwen4exp fused QKV for tensor split

The fourth, 92ee7a5ce qwen4exp: multi-token prediction (nextn) draft head, is not in the upstream PR and has no counterpart there. It is 337 lines across 9 files, mostly src/models/qwen4exp.cpp (+235), and it adds two GGUF tensors to the existing NEXTN family: blk.{bid}.nextn.shared_head_down and ..._up.

I checked this rather than assuming, because it changes what a pin means. Below the nextn commit the carry is byte-identical to upstream 0b19188e9 on every file the PR touches except src/llama-model.cpp, and that difference is entirely the filter_attn / filter_recr reorder below filter_idx that lets #118 add its own without colliding. So: upstream-equivalent plus one reviewed-only-here feature.

Verification

• All seven pins replay on b10639 in order: #107, #108, #70 (additive), #91, #95, #114, #118.
• The composed tree built with -DLLAMA_FATAL_WARNINGS=ON -DGGML_RPC=ON: 573/573, exit 0, zero warnings. test-llama-archs exits 0 with qwen4exp, glm5next and inkling registered. gguf-py imports and exposes the new tensor enums.
• ca5d0a1dd1 is still an ancestor, so #118 composes unchanged, and it does.
• 92ee7a5ce is in #114's commit list and mirrored to refs/pins/92ee7a5ce.

One thing to decide

This probe used 2c5b0007bc for pin #108, which is what #123 proposes and master does not have yet. Merge #123 first, or the nightly still stops at #108 on b10639 regardless of this PR.

Worth being explicit that pinning 92ee7a5ce ships the nextn draft head in the prebuilt binaries with no upstream review behind it, and that a qwen4exp GGUF converted with those two tensors present will not load in a stock build once #27742 lands there. That is a call for whoever owns the feature, not a blocker for the pin; drop the last carry commit and repin at dbdcd49ea if the intent was upstream parity only.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f1a0783f6b

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".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Include the required #108 repin

When this commit lands before #123, the manifest still uses 27278df700... for #108 even though the stated replay verification substituted 2c5b0007bc. The nightly merges pins sequentially and aborts on the first unresolved merge (.github/workflows/unsloth-prebuilt.yml:300-331), so on the current b10639 base it stops at #108 and never includes this new #114 head. Base this change on #123 or update #108 in the same manifest so the committed pin set is the one that was actually verified.

Useful? React with 👍 / 👎.