pugant/Qwen3.8-27B-imatrix

Importance matrix (imatrix) for Qwen3.8-27B

imatrix-qwen38.gguf is an importance matrix for the Qwen3.8-27B dense model (GGUF arch qwen35, the same conversion used by unsloth/BF16 GGUFs), in the standard GGUF imatrix format — usable with any llama.cpp via llama-quantize --imatrix, no fork required.

It is the matrix used to produce pugant/Qwen3.8-27B-MTP-Q4_0_ROCMFP4_STRIX_LEAN (4.34 BPW effective), where it passed our full quality gate.

Calibration data

Generated from a 1.7 MB interleaved composite corpus (131k tokens consumed, --chunks 256), mixed on purpose so that every chunk window sees all three components:

Share	Content	Source
~55%	agentic-coding traces (chat format, <think> traces, tool calls)	ProCreations/grug-think-v3-10k — public 10k-conversation dataset (see its card for license terms), originally used to calibrate the grug finetune quants and reused here as the agentic-coding component
~28%	Italian prose	our own documents
~17%	real source code	our own

Credit where due: the agentic-coding component descends from the grug calibration work — thanks to the grug community and to ProCreations for publishing grug-think-v3-10k. Only aggregate per-block statistics are published here (sum of squares / counts); no calibration text is included in this repo.

Coverage

• 496/496 quantizable tensors covered (importance-collection gate: PASS).
• Not collectable by design (left at defaults by llama-quantize): the MTP draft layers (blk.*.nextn), ssm_conv1d, token_embd, output.weight.

Regenerate it

llama-imatrix -m Qwen3.8-27B-BF16.gguf \
  -f qwen38-calibration.txt \
  -o imatrix-qwen38.gguf \
  --chunks 256 --no-ppl --parse-special \
  --output-frequency 10 --save-frequency 0 \
  -ngl 999

A dense 27B needs the GPU path (-ngl 999); CPU collection would take hours. Note: llama-imatrix saves silently at the end of the run.

File details

• imatrix-qwen38.gguf — 13,642,656 bytes (13.0 MiB)
• SHA256: b5e681d65e726415f899fa8fd4f49e56c3f1d4fd83e3bb8963abce68f881a418

See also

• Quants built with this matrix: pugant/Qwen3.8-27B-MTP-Q4_0_ROCMFP4_STRIX_LEAN (AMD Strix Halo ROCmFP4 preset, includes DFlash2/MTP speculative decoding notes)
• Quantization fork with the ROCmFP4 presets: charlie12345/ROCmFPX
• Calibration dataset: ProCreations/grug-think-v3-10k

---

Software

• Runtime: charlie12345/ROCmFPX (fork of llama.cpp) — our experimental build with per-request drafter routing, reasoning budget and spec-boundary cache salvage: full source included in the lab repo under rocmfpx/
• Lab repo (patches, benchmarks, full replication guide): pugant/strix-halo-llamacpp-lab

The quant built with this matrix and everything linked here is experimental and provided as-is, at your own risk.