DiffusionGemma 26B-A4B · file← DiffusionGemma 26B-A4B
GGUF Q4_K_M
A GGUF that llama.cpp couldn't load. Downloaded twice, never run, deleted on 11 September. Unsloth.
- Size
- 16.8 GB
- Status
- Deleted 11 Sep 2026
- Scored runs
- 0
This file never completed a scored run. Its record below says why.
Rebuilt from the lab's records by Claude, an AI model. The results above come straight from the automated runs. How these pages are made.
| File | models/diffusiongemma-26b/diffusiongemma-26B-A4B-it-Q4_K_M.gguf |
| Format | GGUF Q4_K_M, 16,806,810,208 B (local ls) |
| Where from | downloaded from unsloth/diffusiongemma-26B-A4B-it-GGUF: first attempt 2026-09-10 23:50Z, abandoned 2026-09-11 00:10Z; second, complete download 07:59Z–08:25Z |
| Status | deleted 2026-09-11 15:14Z, unscored |
| Hugging Face | https://huggingface.co/unsloth/diffusiongemma-26B-A4B-it-GGUF |
Where it came from
- Repo / file:
unsloth/diffusiongemma-26B-A4B-it-GGUF,diffusiongemma-26B-A4B-it-Q4_K_M.gguf(HF listing 15.7 GiB); revision not pinned. - Command: a one-off script, a 6-range parallel downloader, run with uv's CPython 3.12. Its source text is not on record.
- Attempts: started 2026-09-10 23:50Z behind the REAP safetensors download (Alan: "Can you pause reap and start diffusion?"); stopped 23:53Z at 5.5 of 15.7 GB when Alan said "Stop then" to an unrelated conversion; restarted 23:58Z, which re-downloaded from scratch because the script only checked for the finished file; killed and the partial removed 2026-09-11 00:10Z on the switch to MLX. Re-started 07:59Z as "the third arm" and complete by 08:25Z.
- Size: 16,806,810,208 B (
ls -la). sha256 not on record.
How it was served
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / reasoning | harness profile | why this label exists |
|---|---|---|---|---|---|
| none (no run-log label) | never served | none; no binary loaded it | not applicable | none | the llama.cpp arm of the diffusion experiment never reached a server |
No run-log label exists. It was never loaded: at 10:14Z the header read general.architecture = diffusion-gemma, and llama.cpp/src/llama-arch.cpp (then the fc/diffusion-confidence build, b10893 base) knew only dream, llada, llada-moe. The assistant: "DiffusionGemma GGUF can't run here at all"; it "needs PR #24427 or the unsloth fork". No seat.
Decisions
- Chosen first (23:44Z) for runtime comparability with the board.
- Abandoned (00:09Z) on unsloth #6336, a Metal garbled-output report for this exact file on an M5 Pro / llama.cpp b9632.
- Re-fetched (07:59Z) after issue #24529 showed coherent Metal output on an M3 Max.
- Declared unloadable (10:14Z) and listed as Tier 1, "no judgement needed", at the 13:50Z disk triage.
- Deleted 2026-09-11 15:14Z:
rm 16G diffusiongemma-26b -- arch diffusion-gemma, no engine here loads it, under Alan's "kill what you can disk wise, keep any q4_0 and drafters...".
What went wrong
- Two wasted downloads (~5.5 GB, then a from-scratch restart), from a resume check that only looked for the finished file.
- A bug report treated as decisive. #6336 was a different chip and build; the file later proved unloadable for a different reason, so the bug never applied here.
- No RESTORE entry (
RESTORE-deleted-20260911.shhas none; 2026-10-23).
What happened to it
Download attempted 2026-09-10 23:50Z → abandoned 2026-09-11 00:10Z → re-downloaded 07:59Z–08:25Z → found unloadable 10:14Z → deleted 15:14Z, unscored, no RESTORE entry → re-downloadable from unsloth/diffusiongemma-26B-A4B-it-GGUF (needs a llama.cpp build with diffusion-gemma support to be of any use).