[bong-water-water-bong]

Follow-up to #209 (review notes on #185/#209). scripts/repack_gguf.py found the deepseek2 converter's old k_b transpose by matching a line of its source text, and rewrote that line in a temp copy. The converter fixed the transpose at the source (5fc52149b), so on a current converter the block never fires, and it would silently misfire if that line ever returned for another reason.

Now: if the converter still contains the inverted transpose, repack stops with ... transposes k_b (16 tiles/head; the forward reads 12). Update the converter to 5fc52149b or later. The Q4_1 config patch next to it is unchanged. 17 insertions, 31 deletions; py_compile passes.

🤖 Generated with Claude Code

[context7]

Docs7 for 1bit-monster/engine

Result	Status	Action
Deployment	➖ Not used	—
Content review	➖ Did not run. This site has no agent runs available this month. Wait for the monthly reset or check your Docs7 plan.	—

Commit 6ccb9a4

[github-actions]

PR Reviewer Guide 🔍

(Review updated until commit 6ccb9a4)

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 209 - Partially compliant Compliant requirements: GGUF-on-NPU support for six architectures (qwen3moe, qwen35moe, GLM fixes) Tokenizer and serve fixes NPU serve takes /tmp/1bit-npu-device.lock Repack records GGUF arch for provenance stamp check Accepts Unigram tokenizer.json and strips SentencePiece's dummy-prefix space Registry rows for the six GGUF-on-NPU arches Tests updated with anchor rerun harness Non-compliant requirements: None Requires further human verification: None 185 - Partially compliant Compliant requirements: 1bit serve -m <model.gguf> --device npu repacks GGUF into Q4NX int4 tile format Runs NPU lane end-to-end for six target architectures Answers anchor prompt ("The capital of France is", greedy, max_tokens=1) with " Paris" Performance measured (0.006–0.17 tok/s) Correctness verified across all six arches GLM k_b transpose fix implemented Q4NX directory fast lane stays bit-identical on golden check Non-compliant requirements: None Requires further human verification: None

⏱️ Estimated effort to review: 3 🔵🔵🔵⚪⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Incorrect Error Message for k_b Transpose The error message printed when detecting an inverted k_b transpose in the converter mentions that the forward reads 12 tiles/head, but the actual forward reads 16 tiles/head. This discrepancy could mislead developers trying to debug or fix the issue. The correct value should be 16, matching the number of tiles the converter produces when the transpose is applied. print(f"repack: {_kb_file} transposes k_b (16 tiles/head; the " f"forward reads 12). Update the converter to 5fc52149b or later.") return 1

[github-actions]

Persistent review updated to latest commit 4bdeb15

[github-actions]

Persistent review updated to latest commit 6ccb9a4