unsloth: pin ggml-org#27941, #152 and #154

[danielhanchen]

Overview

Adds three pins: ggml-org#27941, #152 and #154.

pin	commit	note
ggml-org#27941	02eb201d	no longer a draft, MERGEABLE upstream. Replaces the stale 8161d117 that closed PR #148 carried
#152	258345ef	merges onto b10708 on its own
#154	31e432e7	merges onto b10708 on its own

#137 needs no change. It is already pinned at 4e1865e3, which is still its head.

Why ggml-org#27941 matters here

It fixes four qwen4exp correctness defects, and it is the only account of the one Flash-Next reporter the unified-memory work could not explain: edsonmedina, who says "I only use Vulkan", where Studio never sets the allocator variable. Two of the defects need no AMD involvement at all:

• sequence copies lost their indexer keys, so a copied sequence kept the destination stream's stale keys. Silently wrong output, reachable through the OpenAI n parameter.
• blocks were keyed on position alone, correct only for a single sequence, so under --kv-unified a block could be pooled from another sequence's cells.

So Flash-Next gibberish has two independent causes stacked on one model, which is why the workarounds in that thread contradict each other. The AMD pins fix one of them; this fixes the other.

#142 and #144 are deliberately NOT pinned

Both conflict against b10708 on src/llama.cpp and src/llama-model-loader.cpp, and it is a real semantic conflict rather than a textual one. Upstream folded lazy_mode and model_shared into a lazy struct:

b10708:  ml.lazy.mode    = params.lazy_mode;
#142:    ml.tensor_read_lazy = params.tensor_read_lazy;  ml.model_shared = params.model_shared;
#144:    ml.lazy_mode        = params.lazy_mode;         ml.model_shared = params.model_shared;

Both branches predate that refactor. Pinning them as they stand would fail the resolve and take every pin after them down with it, since the preflight stops at the first conflict. They need rebasing onto current upstream first, which is a code change in MTP paths and wants its own review rather than being smuggled through a pin update.

Verification

Each new pin was merged onto b10708 individually before being listed:

#152    ok   9 files changed, 306 insertions(+), 25 deletions(-)
#154    ok   7 files changed, 249 insertions(+), 20 deletions(-)
#27941  ok   6 files changed, 505 insertions(+), 138 deletions(-)

The full chain is left to the preflight, which resolves some overlaps with additive_merge.py and is the authority; a plain git merge replay is not equivalent.

[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 2026-08-31T10:09:47.297728Z	7387fae	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.