Qwen3.6-27B · ROCmFPX

Stock Q6_K quality, ~30% faster prompt-processing on AMD Strix Halo (gfx1151)

ROCmFPX 3→8-bit quants of unsloth/Qwen3.6-27B-MTP-GGUF — the general-purpose base, with MTP speculative decoding + Qwen3-VL vision. Neutral imatrix calibration, so base behavior is preserved (not skewed toward code).

⚠️ Requires the ROCmFPX fork (build main — the FP* types are merged in) — custom AMD quant types (enum IDs 110–115), not upstream-stable. Won't load in stock llama.cpp / LM Studio / Ollama. HF's precision badge is wrong — pick the file by name.

Pick a tier

All filenames prefixed Qwen3.6-27B-STRIX-.

Quick start

# build the fork once — main already has the ROCmFPX quant types
git clone https://github.com/charlie12345/ROCmFPX.git && cd ROCmFPX
JOBS=16 scripts/build-strix-rocmfp4-mtp.sh

# serve the flagship — MTP + vision
HSA_OVERRIDE_GFX_VERSION=11.5.1 build-strix-rocmfp4/bin/llama-server \
  -m Qwen3.6-27B-STRIX-embF16-headQ6-Q6_0_ROCMFPX.gguf \
  -dev ROCm0 -ngl 999 -fa on -c 32768 \
  --spec-type draft-mtp --spec-draft-ngl all --spec-draft-n-max 2 \
  --jinja --mmproj mmproj/mmproj-F32.gguf --host 0.0.0.0 --port 8080

Tool calls: serve with --jinja so the model's own chat template emits them — no special parser flag needed.

All tiers · recipe · benchmarks

All tiers

File suffix	Preset	Enum	Size	Role
embF16-headQ6-Q6_0_ROCMFPX.gguf	Q6_0_ROCMFPX	110	23G	flagship — f16 emb + Q6_K head + imatrix
embF16-Q8_0_ROCMFPX.gguf	Q8_0_ROCMFPX	111	28G	highest fidelity
embF16-Q4_0_ROCMFP4.gguf	Q4_0_ROCMFP4	100	19G	fastest decode (4-bit body)
embF16-Q3_0_ROCMFPX.gguf	Q3_0_ROCMFPX	112	16G	smallest

f16 token embeddings throughout. (HF labels Q4/Q8 but not Q6/Q3 — the latter aren't standard llama.cpp quant names; pick by filename.)

Verification (Strix Halo gfx1151)

Metric	Value
Functional smoke	chat/coding/JSON/tool-call/coherency ✅ (5/5)
PPL vs Q6_K	flagship 5.7030 vs stock Q6_K 5.6618 → Δ +0.73% (general slice, ctx 512)

Performance — prompt-processing throughput (t/s) vs Q6_K

Context	Q6_K	flagship	Δ
pp512	188	242	+29%
pp2048	191	245	+28%
pp10k	180	228	+27%
pp16k	172	217	+26%
pp32k	156	193	+23%
pp64k	133	159	+20%

Clean dedicated-GPU llama-bench (both servers stopped, warm-up pass discarded). The gfx1151-tuned ROCmFPX kernels win the compute-bound prefill — the edge is largest at short context and narrows to +20% at 64k as O(n²) attention takes over. Decode is bandwidth-bound (≈ Q6_K raw), and MTP (--spec-type draft-mtp) ~doubles it in serving. Single-rep llama-bench; treat absolutes as ±a few %.

Credits & license

Apache-2.0 (inherited). Qwen3.6 (base) → charlie12345 / ROCmFPX (quant fork). ROCmFPX quantization by this repo's author.