JackBinary's picture
Upload folder using huggingface_hub
97785d7 verified
|
Raw
History Blame Contribute Delete
2.11 kB
metadata
tags:
  - rocmfpx
  - gguf
  - quantization
  - llama-cpp
  - playbook
license: mit

ROCmFPX Quantize Playbook

A field-tested, agent-ready playbook for producing ROCmFPX hybrid GGUF quantizations (q4_0_rocmfp4_fast / q8_0_rocmfpx) with the ROCmFPX fork of llama.cpp — fully CPU-only, from either pre-quantized GGUF repos (e.g. Unsloth BF16) or raw HF safetensors.

The complete guide is in rocmfpx-quantization-guide.md. Hand this repo (or just that file) to a coding agent together with a Hugging Face link and a recipe, and it can reproduce every step below.

What it covers

  • One-time CPU-only build of llama-quantize / llama-gguf-split + Python conversion deps
  • Workflow A: source repo already has BF16 GGUF shards (e.g. unsloth/*-GGUF)
  • Workflow B: raw safetensors → convert_hf_to_gguf.py → BF16 GGUF
  • How to discover routed-expert tensor names per architecture (regex lookup table)
  • Dry-run verification, shard merging, and output validation
  • VRAM-fit hybrid recipe (q4 bulk + q8 sensitive) derived from Unsloth Dynamic tiers
  • Cheatsheets, timings, and gotchas (regex anchoring, MTP auto-protection, gguf-py version trap…)

Validated runs

Model Recipe Result
Laguna-S-2.1 118B-A10B q4 experts / q8 rest 224 GB → 61.6 GB (4.39 bpw)
Qwen3.8-27B pure q8 and 16 GB hybrid from UD-Q4_K_XL tiers 26.9 GB (8.25 bpw) / 16.4 GB (5.15 bpw)
G4-MeroMero-26B-A4B q4 experts / q8 rest + mmproj 50.5 GB → 14.0 GB (4.64 bpw)

Timings on a 64-core CPU box: 118B MoE ≈ 8 min, 27B dense ≈ 1 min, 26B MoE ≈ 3 min.

Related