Repin the Qwen3.8-Flash-Next MTP carry onto its fixed head

[danielhanchen]

scripts/unsloth/pr-set.json pins the Qwen3.8-Flash-Next MTP carry (#144) at f40f64a81. That commit aborts at load whenever the MTP head is attached:

ggml.c:2293: GGML_ASSERT(ggml_can_repeat(b, a)) failed
  libllama.so.0(llama_model_qwen4exp::graph::build_hc_mix(...))
  libllama.so.0(llama_model_qwen4exp::graph_mtp::graph_mtp(...))
  libllama.so.0(llama_context::graph_reserve(...))
  libllama-common.so.0(common_fit_params(...))

This moves the pin forward to ca1426903, #144's current head, which now carries two fixes that landed on its branch through #219.

Why it broke

ggml-org#28896 (41abbfd599, 2026-09-14) moved the qwen4exp hyper-connection gammas to { n_embd, hc } with TENSOR_ALLOW_RESHAPE. Upstream did nothing wrong: it changed code it owns. The MTP head's gamma is fork-only, lives 110 lines further down in the same file, and did not conflict, so the rebase in f40f64a81 ("Fix merge conflicts") converted the three trunk gammas and left layer.nextn.hc_head_norm at { hc_dim }.

git blame src/models/qwen4exp.cpp on the old head:

41abbfd599 (Aman Gupta     2026-09-14 :168) // the gammas load as [n_embd, hc] so the grouped norm multiplies them without a graph reshape
f40f64a81c (danielhanchen  2026-09-16 :170) hc_head_norm = create_tensor(..., { n_embd, hc }, trunk_flags | TENSOR_ALLOW_RESHAPE);
e52b4b00c2 (Ryan Monsurate 2026-08-27 :280) layer.nextn.hc_head_norm = create_tensor(..., { hc_dim }, flags);

build_hc_mix then multiplies a [2560, 4, n_tokens] stream by a [10240] weight. Nothing about the published GGUFs changed; only the expected shape moved.

The second fix addresses a standing bug rather than this regression. common/fit.cpp opens the draft model on its own to price it, and a borrowing shared- head has no token_embd.weight, so borrow_shared_tensor threw, the fit warned fitting without it and budgeted nothing for the draft, and the real load ran out of device memory by roughly the head's size.

Measured

Built from this repo with identical flags (-DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=100, Release), B200 held at 30.7 GiB free so the load matches a 32 GB consumer card. Target unsloth/Qwen3.8-Flash-Next-GGUF UD-Q4_K_XL, --fit on --flash-attn on -c 8192 --spec-type draft-mtp --spec-draft-n-max 2, 512-token prompt, 128 generated, median of 5 samples after a warmup.

build	self-contained Q8_0 head	recommended shared-Q8_0 head
f40f64a81 (current pin)	aborts at load	aborts at load
ca1426903 (this pin)	61.2 tok/s	61.6 to 63.6 tok/s

ca1426903 is a squash of the commits measured above and produces a byte-identical tree, so those numbers carry over unchanged.

Against the published prebuilts on the same harness and card, without a drafter:

prebuilt	MTP	no MTP
b10909-mix-bea84f7 (pin a9e9c3c)	60.7 tok/s	43.8 tok/s
b10995-mix-3e83366	aborts
b11007-mix-3e83366	aborts
b11027-mix-3e83366	aborts	45.7 tok/s

So b10995-mix-3e83366 (2026-09-16) is the first affected build, the fix restores decode to the pre-break figure, and losing the drafter costs 28 percent. Without a drafter the newer builds are faster than b10909, by 4.4 percent offloaded and 8.4 percent fully resident, so nothing else regressed.

Reported in unslothai/unsloth#11143, unslothai/unsloth#11219 and unslothai/unsloth#11221.

[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 2026-09-18T09:33:21.721543Z	55d46a4	PR opened
🔒 Security Review	✅ Completed 2026-09-18T09:36:07.541238Z	55d46a4	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.