jcbtc/qwopus3.6-27b-v2-chadrock-rocmfp4-mtp

Qwopus3.6 27B v2 Chadrock ROCmFP4 MTP

Qwopus3.6 Chadrock is a ROCmFP4/MTP GGUF build of Jackrong/Qwopus3.6-27B-v2, tuned for AMD Ryzen AI Max+ 395 / Strix Halo systems.

This model keeps the Qwopus3.6 27B v2 behavior people like, then puts it through Charlie's AMD-focused ROCmFP4 + MTP runtime path. The result is a compact 14 GB GGUF that runs with native draft-MTP on Strix Halo and posts a better HumanEval result than the original local Qwopus3.6 27B v2 Q5 row.

This GGUF will not run correctly with stock llama.cpp. You need the custom charlie12345/rocmfp4-llama build because this file uses ROCmFP4 tensor types that upstream llama.cpp does not currently understand.

The model file is already provided here. You do not need to rebuild or quantize the model. You only need to build the custom llama server once.

Why This Mix

Qwopus3.6 27B v2 is already a strong dense local model with vision and tool-use capability. Chadrock adds the AMD runtime piece:

• Qwopus3.6 27B v2 behavior from Jackrong
• native MTP serving
• ROCmFP4 Strix Lean tensor recipe
• AMD ROCm/HIP backend path
• 262K context target
• one-slot draft-MTP profile for real local use

This is not just a smaller file. It is a model/runtime pairing built for Strix Halo.

Model Lineage

This card follows the upstream tree from Jackrong's Qwopus card, then adds this ROCmFP4 / Chadrock build at the end:

Qwen/Qwen3.6-27B
  -> Jackrong/Qwopus3.6-27B-v2
       datasets:
         - Jackrong/Claude-opus-4.6-TraceInversion-9000x
         - Jackrong/Claude-opus-4.7-TraceInversion-5000x
  -> Jackrong/Qwopus3.6-27B-v2-MTP-GGUF
  -> jcbtc/qwopus3.6-27b-v2-chadrock-rocmfp4-mtp

In plain terms: Qwen provides the dense 27B foundation, Jackrong's Qwopus v2 adds the Trace Inversion / curriculum SFT behavior, Jackrong's MTP GGUF provides the MTP source path, and this release converts that line into the Strix-focused ROCmFP4 Chadrock format.

Technical Metadata

Field	Value
model size	27B dense
architecture	qwen35
GGUF size label	27B
direct upstream behavior model	Jackrong/Qwopus3.6-27B-v2
direct upstream MTP GGUF	Jackrong/Qwopus3.6-27B-v2-MTP-GGUF
base family	Qwen/Qwen3.6-27B
local runtime format	ROCmFP4 Chadrock GGUF

Headline Benchmarks

All local numbers below were measured on AMD Ryzen AI Max+ 395 / Strix Halo.

HumanEval

Model / row	HumanEval base	HumanEval+
Qwopus3.6 27B v2 Chadrock ROCmFP4 MTP	159/164 = 96.95%	155/164 = 94.51%
Original Qwopus3.6 27B v2 Q5_K_M	151/164 = 92.07%	147/164 = 89.63%

That is an +8 task improvement on base HumanEval and an +8 task improvement on HumanEval+ versus the recorded original Qwopus3.6 27B v2 Q5_K_M row.

HumanEval Speed

Metric	Qwopus3.6 Chadrock
HumanEval tasks	164
completion tokens generated	45,033
cumulative request latency	1346.8s
mean total-token request speed	59.08 tok/s
median total-token request speed	60.04 tok/s
completion-only llama.cpp eval speed	~33.44 tok/s
peak active completion speed	~37.14 tok/s

The stored original Qwopus3.6 27B v2 Q5_K_M HumanEval run recorded 3834s generation time. This Chadrock run completed the same 164-task HumanEval codegen workload with about 2.8x lower recorded request-generation time while also scoring higher.

BFCL Tool Calling

Qwopus3.6 Chadrock also did well on BFCL v4 non-live tool-calling rows:

BFCL v4 row	Score
non-live overall	85.88%
simple Python AST	94.50%
multiple-call AST	96.00%
parallel-call AST	86.50%
parallel multiple-call AST	85.50%
irrelevance detection	81.67%

This is the profile to try if you want a local Strix Halo model that feels fast while still staying sharp on coding and tool-use formats.

Run With llama-server

Build Charlie's custom llama.cpp once, download this GGUF, then run:

HSA_OVERRIDE_GFX_VERSION=11.5.1 \
GGML_HIP_ENABLE_UNIFIED_MEMORY=1 \
/path/to/rocmfp4-llama/build-strix-rocmfp4/bin/llama-server \
  -m Qwopus3.6-27B-v2-MTP-BF16-to-ROCmFP4-STRIX_LEAN.gguf \
  --mmproj mmproj-F32.mmproj \
  --alias qwopus3.6-27b-v2-chadrock \
  --host 127.0.0.1 \
  --port 8080 \
  --jinja \
  -c 262144 \
  -ngl 999 \
  -fa on \
  -dev ROCm0 \
  -b 512 \
  -ub 512 \
  -t 16 \
  -tb 32 \
  -ctk q4_0 \
  -ctv q4_0 \
  --spec-type draft-mtp \
  --spec-draft-device ROCm0 \
  --spec-draft-ngl all \
  --spec-draft-type-k q4_0 \
  --spec-draft-type-v q4_0 \
  --spec-draft-n-max 4 \
  --spec-draft-n-min 0 \
  --spec-draft-p-min 0.0 \
  --spec-draft-p-split 0.10 \
  --parallel 1 \
  --metrics \
  --no-mmap

Use --parallel 1 for MTP. Multi-slot serving changes the draft-MTP behavior and is not the intended profile.

For text-only use, you may omit --mmproj.

For vision use, keep mmproj-F32.mmproj beside the main GGUF, but run with MTP off. In practice, that means using the vision projector and removing the --spec-* draft-MTP flags from the command.

The projector is a GGUF-format projector file with a .mmproj repo extension so Hugging Face's GGUF metadata badge tracks the 27B language model rather than the smaller CLIP projector.

Build The Required llama.cpp

The GGUF is already provided. You only need to build the custom llama.cpp server once:

git clone https://github.com/charlie12345/rocmfp4-llama.git
cd rocmfp4-llama
git checkout mtp-rocmfp4-strix
env JOBS=16 scripts/build-strix-rocmfp4-mtp.sh

The server binary will be here:

build-strix-rocmfp4/bin/llama-server

About ROCmFP4 / Chadrock

Charlie's ROCmFP4 method adds AMD-focused GGUF tensor formats and backend paths to llama.cpp.

ROCmFP4 is not stock Q4, MXFP4, or NVFP4. It uses custom 4-bit tensor layouts, Codebook10 values, finite unsigned E4M3 scale semantics, tensor-aware Strix presets, ROCm/HIP kernels, Vulkan support, and MTP regression guards.

Why it matters: Strix Halo has a large unified-memory pool, but good local serving still depends on memory bandwidth, tensor layout, KV traffic, and draft-token acceptance. Chadrock is built for that exact hardware shape.

Files

File	Size	SHA256
Qwopus3.6-27B-v2-MTP-BF16-to-ROCmFP4-STRIX_LEAN.gguf	14 GB	1f1c0a9d63b9b38b06feb4f460f9cb6ed85f001331be713f09e6c2aaff5367e4
mmproj-F32.mmproj	889 MB	bf51f62572c6e513659d3fa4989ac846e58fa8c30fb60ecb8112aebb1f3128a3

Credits

• Qwen: Qwen/Qwen3.6-27B base model family.
• Jackrong: Qwopus3.6-27B-v2, the Trace Inversion datasets, and the MTP GGUF source.
• charlie12345 / @Italianclownz: ROCmFP4 llama.cpp fork, Strix Halo build path, and AMD-focused MTP runtime work.

Notes

This is an experimental AMD ROCmFP4/MTP build. Performance depends on driver version, clocks, prompt shape, MTP acceptance, and serving flags. The numbers above are local reproducible measurements, not universal llama.cpp claims.