Modelpublic

philtheriver/Qwopus3.6-27B-v2-MTP-ROCmFPX

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

2likes478downloads

Model Card

Qwopus3.6-27B-v2 · ROCmFPX

Stock `Q6_K` quality, ~25% faster prompt-processing on AMD Strix Halo (`gfx1151`)

ROCmFPX 3→8-bit quants of `Jackrong/Qwopus3.6-27B-v2-MTP-GGUF` — the general-purpose 27B Qwopus, MTP speculative decoding + Qwen3-VL vision, agent/tool-use tuned.


Quality	≈ stock `Q6_K` — PPL +0.28% (within error)
Prompt processing	+27% vs Q6_K (short ctx) → +16% at 64k
Decode	\~18 tok/s with MTP (~9 raw)
Vision	Qwen3-VL — bundled `mmproj/`

⚠️ Requires the [ROCmFPX fork](https://github.com/charlie12345/ROCmFPX) (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

File suffix	Size	Best for
`…embF16-headQ6-Q6_0_ROCMFPX_AGENT.gguf` ★	26 GB	best overall — the flagship
`…embF16-Q8_0_ROCMFPX.gguf`	29 GB	maximum fidelity
`…embF16-Q4_0_ROCMFP4.gguf`	19 GB	fastest decode (4-bit)
`…embF16-Q3_0_ROCMFPX.gguf`	17 GB	smallest

Agent-routed _AGENT tiers + the full enum/bpw table are in the details below and the Files tab. All filenames prefixed Qwopus3.6-27B-v2-MTP-STRIX-.

Quick start

bash

# 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 Qwopus3.6-27B-v2-MTP-STRIX-embF16-headQ6-Q6_0_ROCMFPX_AGENT.gguf \
  -dev ROCm0 -ngl 999 -fa on -c 131072 \
  --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: point your client at the `qwen3_coder` parser, or the model narrates instead of emitting structured calls.

<details> <summary><b>All tiers · recipe · benchmarks</b></summary>

All tiers

File suffix	Preset	Enum	Size	Role
`embF16-headQ6-Q6_0_ROCMFPX_AGENT.gguf`	`Q6_0_ROCMFPX_AGENT`	114	26 GB	flagship — f16 emb + Q6_K head + imatrix
`embF16-Q8_0_ROCMFPX_AGENT.gguf`	`Q8_0_ROCMFPX_AGENT`	115	30 GB	highest-fidelity agent
`embF16-Q8_0_ROCMFPX.gguf`	`Q8_0_ROCMFPX`	111	29 GB	highest fidelity
`embF16-Q6_0_ROCMFPX.gguf`	`Q6_0_ROCMFPX`	110	24 GB	balanced
`embF16-Q3_0_ROCMFPX_AGENT.gguf`	`Q3_0_ROCMFPX_AGENT`	113	21 GB	smallest agent
`embF16-Q3_0_ROCMFPX.gguf`	`Q3_0_ROCMFPX`	112	17 GB	smallest
`embF16-Q4_0_ROCMFP4.gguf`	`Q4_0_ROCMFP4`	100	19 GB	fastest decode (4-bit body)

f16 token embeddings throughout; _AGENT presets keep attention/FFN routing at higher precision for tool-call/JSON coherence. (HF labels Q4/Q8 but not Q6/Q3 — the latter aren't standard llama.cpp quant names.)

Verification (Strix Halo gfx1151)

Metric	Value
Functional smoke	chat/coding/JSON/tool-call/coherency ✅ (5/5)
PPL vs `Q6_K` (code corpus)	flagship 2.962 vs Q6_K 2.954 → +0.28% (within ±0.04)

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

Context	`Q6_K`	flagship	Δ
pp512	195	248	+27%
pp2048	189	242	+28%
pp10k	179	222	+24%
pp16k	172	208	+21%
pp32k	157	186	+19%
pp64k	133	155	+16%

The gfx1151-tuned kernels win the compute-bound prefill; the edge is largest at short context and narrows toward +16% at 64k as O(n²) attention takes over. Decode is bandwidth-bound (≈ Q6K raw), and **MTP (`--spec-type draft-mtp`) ~doubles it** in serving. `Q40_ROCMFP4 is the decode king (~13 tok/s raw). *Single-rep llama-bench`, ran warm — absolutes run a touch conservative vs the Coder sibling.*

</details>

Credits & license

Apache-2.0 (inherited). Jackrong + Kyle Hessling (fine-tune) → Qwen3.6-27B (base) → charlie12345 / ROCmFPX (quant fork). ROCmFPX quantization by this repo's author.