# Reproducible provenance

Generated: `2026-08-18T00:16:18Z`

## Source

- Repository: `unsloth/Qwen3.8-27B-GGUF`
- Revision: `f1bfb127c64f7072bdd2cad55f258b9c8b2910fe`
- Conversion source: pinned two-part BF16 GGUF, not an intermediate Q8 or another lossy quantization

- `BF16/Qwen3.8-27B-BF16-00001-of-00002.gguf`: `b9966e82b7a4d87028b5eae061d578ee826305ebf8baea5bfc6e09bad0ba191f` (49986159616 bytes)
- `BF16/Qwen3.8-27B-BF16-00002-of-00002.gguf`: `92e3943c4f9bd6292a7bef82369f65fed9bfed088b9df0fb2fa2ce17c9edfa02` (4671576000 bytes)

## Quantizer

- llama.cpp base commit: `62061f91088281e65071cc38c5f69ee95c39f14e`
- Official upstream MTP fix: PR `#24986`, merge commit `b3ce5cedf4c007b78a45befe839fa3abada03c0b`
- Patched quantizer SHA256: `5d3a8456974b28569322dea7ee33941e3c9f09750e100dfce358326110b29187`
- MTP patch SHA256: `01a53c23afd4ed81b79ffc697b1c4a4b83443253b6095732eee3f8fc58e96b40`
- Override file SHA256: `ee2b07a31d8c8811a852f71a748c29b54e3dc599e7cf0d7cdb5e217723e12ec2`
- Run script SHA256: `9c44f21e3d6317dad6160b79a31400b8159785cb7b5813c386655736bc321daf`

The exact executable and all six loaded llama/ggml shared libraries are recorded in `QUANTIZER_TOOLCHAIN.json`.

## Commands

Dry run:

```text
/mnt/geth-vol1/qwen38_tools/llama.cpp-62061f910-mtpfix/build/bin/llama-quantize --imatrix /mnt/geth-vol1/qwen38_assets/Qwen3.8-27B-agentic.imatrix --tensor-type-file /mnt/geth-vol1/qwen38_assets/iq1m_mtp_q4k_overrides.txt --dry-run /root/qwen38_bf16/BF16/Qwen3.8-27B-BF16-00001-of-00002.gguf /mnt/geth-vol1/qwen38_out/Qwen3.8-27B-IQ1_M.gguf.part IQ1_M 16
```

Quantization:

```text
nice -n 5 /mnt/geth-vol1/qwen38_tools/llama.cpp-62061f910-mtpfix/build/bin/llama-quantize --imatrix /mnt/geth-vol1/qwen38_assets/Qwen3.8-27B-agentic.imatrix --tensor-type-file /mnt/geth-vol1/qwen38_assets/iq1m_mtp_q4k_overrides.txt /root/qwen38_bf16/BF16/Qwen3.8-27B-BF16-00001-of-00002.gguf /mnt/geth-vol1/qwen38_out/Qwen3.8-27B-IQ1_M.gguf.part IQ1_M 16
```

## MTP exception

The pinned importance matrix has no calibration entries for the auxiliary `blk.64` prediction block. The official `n_layer_all` accounting fix makes that extra block visible to the quantizer. Exactly the following eight two-dimensional tensors are kept at Q4_K:

- `blk.64.attn_k.weight`
- `blk.64.attn_output.weight`
- `blk.64.attn_q.weight`
- `blk.64.attn_v.weight`
- `blk.64.ffn_down.weight`
- `blk.64.ffn_gate.weight`
- `blk.64.ffn_up.weight`
- `blk.64.nextn.eh_proj.weight`

Together they contain 424,673,280 parameters (1.5544% of the model). `MTP_Q4K_AUDIT.json` proves the dry run saw all 15 block-64 rows and assigned Q4_K to exactly these eight BF16 matrices.

## Validation gates

- GGUF magic and `qwen35` architecture: hard gate
- IQ1_M present and all eight manual block-64 Q4_K override tensors present: hard gate
- Whole-file BPW range 2.25–2.35: hard gate
- Projector size and SHA256: hard gate
- Text runtime probe: completed non-zero exit is a hard failure; timeout is recorded as a soft warning
- Vision runtime probe with the pinned projector: completed non-zero exit is a hard failure; timeout is recorded as a soft warning
- Remote Hugging Face LFS size/SHA and Range GGUF magic: hard gates performed by `upload_verify.py`