gsrunion/gpt-oss-120b-ROCmFP4-STRIX_LEAN-GGUF

GPT-OSS-120B — ROCmFP4 STRIX_LEAN for AMD Strix Halo

ROCmFP4 GGUF requantization of openai/gpt-oss-120b (117B MoE, 5.1B active), produced for AMD Strix Halo (Ryzen AI Max, gfx1151).

To our knowledge the first ROCmFP4-family quant of gpt-oss. +21% decode over the native MXFP4 release on Strix Halo at essentially identical size — the gain is pure kernel efficiency (the ROCmFP4 layout streams at near-peak bandwidth on gfx1151).

⚠️ Stock llama.cpp will reject this file (invalid ggml type 106). It runs on: - the kyuz0/amd-strix-halo-toolboxes rocm-7.2.4-rocmfp4 container image, or - the Hal0ai/Hal0_ROCmFPX llama.cpp fork.

Companion repos: Qwen3.6-35B-A3B · Nemotron-3-Nano-30B · Ornith-1.0-35B · Ornith-1.0-9B

Files

Measured performance

AMD Ryzen AI Max+ 395 (Strix Halo, 128 GB unified LPDDR5X, ROCm, llama-bench -fa 1 --mmap 0):

Provenance & quality disclosure

Requantized from the native MXFP4 release (no BF16 master exists for gpt-oss) with llama-quantize --allow-requantize, calibrated with bartowski's imatrix. FP4→FP4 regridding adds rounding noise; on an 8-prompt quality battery vs the native release (temp 0), 7/8 answers were byte-equivalent-or-equal-quality and one contained a minor arithmetic slip in an illustrative example (the substantive answer remained correct). Judge that trade for your workload — the native MXFP4 remains available from ggml-org.

Serving

llama-server -m gpt-oss-120b-Q4_0_ROCMFP4_STRIX_LEAN.gguf \
  -ngl 999 -fa on --jinja -c 65536

Credits

• —Base model: OpenAI — gpt-oss-120b (Apache-2.0)
• —imatrix: bartowski
• —ROCmFP4 quant formats: Hal0ai; GPU quantizer build: kyuz0