Modelpublic

jcbtc/Chadrockv2-Qwen3.6-27B-ROCmFP6-STRIX-QUALITY

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

5likes54downloads

Model Card

[image]

Chadrockv2 Qwen3.6 27B ROCmFP6 STRIX QUALITY

Chadrockv2 Qwen3.6 27B ROCmFP6 STRIX QUALITY is an AMD-tuned GGUF release of the Unsloth Qwen3.6 27B MTP line. It uses a new ROCmFP6 Strix Quality recipe designed to recover Q6-class agent behavior while keeping the ROCmFPX served-speed advantages on AMD Ryzen AI Max+ 395 / Strix Halo systems.

This is a model/runtime pairing, not a generic GGUF quant. The file uses custom ROCmFPX tensor types and will not run correctly with stock upstream llama.cpp. Use the ROCmFPX branch and launch profile documented below.

Full research report:

https://llm.ciru.ai/reports/rocmfp6-quality-research-report-20260624/

Why This Build Exists

The earlier Strix speed ROCmFP6 recipe was too small for agent quality. It measured about 4.82 BPW and scored clearly below the downloaded Unsloth Q6 baseline on HermesAgent-20. This STRIX QUALITY recipe moves closer to a real Q6-class file by keeping the bulk of tensors in Q6_0_ROCMFPX and promoting high-impact tensors to Q8_0_ROCMFPX.

The result is larger than the old speed recipe but materially better on agent quality:

Model	HermesAgent-20 score	Base pass	Plus pass	HumanEval+ plus	PPL
Chadrockv2 ROCmFP6 STRIX QUALITY	`0.78`	`14/20`	`11/20`	`155/164 = 94.51%`	`6.5543 +/- 0.0941`
Unsloth Q6 baseline	`0.76`	`13/20`	`11/20`	`153/164 = 93.29%`	`6.5296 +/- 0.0934`
Old ROCmFP6 Strix Speed	`0.60`	`10/20`	`9/20`	`152/164 = 92.68%`	`6.4077 +/- 0.0902`

The important lesson from the tuning run is that perplexity alone was not enough. The old small FP6 recipe looked acceptable by PPL, but failed agent scenarios. HermesAgent-20 and EvalPlus showed that the quality recipe recovered the behavior we needed.

Lineage

text

Qwen/Qwen3.6-27B
  -> unsloth/Qwen3.6-27B
  -> unsloth/Qwen3.6-27B-MTP-GGUF
  -> Chadrockv2 Qwen3.6 27B ROCmFP6 STRIX QUALITY

The public release name and artifact names are Chadrock names. The source lineage remains explicit in metadata, benchmark notes, and credits.

Files

File	Size	SHA256
`Chadrockv2-Qwen3.6-27B-ROCmFP6-STRIX-QUALITY.gguf`	`25,196,024,736 bytes`	`144062b43fade17c15217acf0b4974041f6135d73945bc13e7c13b1d18946b84`
`Chadrockv2-Qwen3.6-27B-ROCmFP6-STRIX-QUALITY.gguf.sha256`	checksum	same hash as above
`profiles/unsloth-qwen36-27b-mtp-rocmfp6-strix-quality-cap6-q8kv-rocm-hermes64k.env`	launch profile	AMD Strix Halo ROCm profile

Recipe

Recipe	Estimated size	BPW	Tensor mix
STRIX QUALITY	`24018.32 MiB`	`7.37`	`312` Q6 tensors, `194` Q8 tensors
Straight Q6 ROCmFPX	local dry-run	`6.59`	`486` Q6 tensors, `20` Q8 tensors
Old Strix Speed	local dry-run	`4.82`	`388` FP4-fast tensors, `118` Q6 tensors
Q6 ROCmFPX Agent	local dry-run	`7.40`	`340` Q6 tensors, `166` Q8 tensors

STRIX QUALITY keeps the default tensor type at Q6_0_ROCMFPX, then promotes:

—token embedding and output tensors
—attention Q, K, V, O, and fused QKV tensors
—selected FFN down/gate tensor bands
—llama.cpp tensors marked by the use_more_bits heuristic

The recipe is implemented as:

text

LLAMA_FTYPE_MOSTLY_Q6_0_ROCMFPX_STRIX_QUALITY = 118
scripts/quantize-rocmfpx-agent.sh --profile strix-quality

Quality Results

HermesAgent-20 is the deciding quality test for this release because it exposes scenario-level failures that aggregate PPL missed.

Model	Score	Base pass	Plus pass	Generation time
Chadrockv2 ROCmFP6 STRIX QUALITY	`0.78`	`14/20`	`11/20`	`1541.503 s`
Unsloth Q6 baseline	`0.76`	`13/20`	`11/20`	`1037.491 s`
Old ROCmFP6 Strix Speed	`0.60`	`10/20`	`9/20`	`791.457 s`

EvalPlus confirms that the quality recipe did not trade away coding correctness:

Model	HumanEval base	HumanEval+
Chadrockv2 ROCmFP6 STRIX QUALITY	`161/164`	`155/164 = 94.51%`
Unsloth Q6 baseline	`160/164`	`153/164 = 93.29%`
Old ROCmFP6 Strix Speed	`159/164`	`152/164 = 92.68%`

Speed Results

All rows were measured locally on AMD Ryzen AI Max+ 395 / Strix Halo, one-slot served MTP, q8_0 target KV, f16 draft KV, b2048/u512, temperature=0, 512 generated tokens, and no prompt cache reuse.

ROCmFP6 STRIX QUALITY vs Unsloth Q6 Baseline

Prompt tokens	FP6 ROCm PP tok/s	FP6 ROCm TG tok/s	FP6 total	Q6 ROCm PP tok/s	Q6 ROCm TG tok/s	Q6 total
`512`	`177.98`	`29.52`	`20.1 s`	`200.84`	`22.10`	`25.6 s`
`2048`	`188.44`	`20.64`	`34.7 s`	`208.53`	`17.38`	`38.4 s`
`4096`	`213.53`	`30.73`	`33.5 s`	`227.13`	`27.75`	`34.3 s`
`16384`	`223.76`	`30.03`	`85.9 s`	`218.75`	`25.76`	`90.3 s`
`65536`	`171.08`	`15.72`	`388.4 s`	`166.15`	`10.81`	`413.7 s`

ROCm vs Vulkan for This FP6 File

Prompt tokens	ROCm TG tok/s	ROCm total	Vulkan TG tok/s	Vulkan total
`512`	`29.52`	`20.1 s`	`19.58`	`28.9 s`
`2048`	`20.64`	`34.7 s`	`19.45`	`36.3 s`
`4096`	`30.73`	`33.5 s`	`13.10`	`57.6 s`
`16384`	`30.03`	`85.9 s`	`13.41`	`120.6 s`
`65536`	`15.72`	`388.4 s`	`9.19`	`471.6 s`

ROCm0 is the recommended backend for this release. Vulkan remains useful as a portability path, but it was slower across this Strix Quality speed matrix.

Run With ROCmFPX

Build the ROCmFPX runner branch containing this ftype and recipe:

bash

git clone https://github.com/ciru-ai/ROCmFPX.git
cd ROCmFPX
git checkout rocmfp6-strix-quality
cmake -S . -B build-strix-rocmfp6-quality-hip \
  -DGGML_HIP=ON \
  -DGGML_VULKAN=ON \
  -DCMAKE_BUILD_TYPE=Release
cmake --build build-strix-rocmfp6-quality-hip -j

Launch the validated AMD Strix Halo profile:

bash

HSA_OVERRIDE_GFX_VERSION=11.5.1 \
GGML_HIP_ENABLE_UNIFIED_MEMORY=1 \
./build-strix-rocmfp6-quality-hip/bin/llama-server \
  -m /path/to/Chadrockv2-Qwen3.6-27B-ROCmFP6-STRIX-QUALITY.gguf \
  --alias chadrockv2-qwen36-27b-rocmfp6-strix-quality \
  --host 127.0.0.1 \
  --port 8080 \
  --jinja \
  -c 65536 \
  -ngl 999 \
  -fa on \
  -dev ROCm0 \
  -sm none \
  -b 2048 \
  -ub 512 \
  -t 16 \
  -tb 32 \
  -ctk q8_0 \
  -ctv q8_0 \
  --ctx-checkpoints 0 \
  --checkpoint-every-n-tokens -1 \
  --spec-type draft-mtp \
  --spec-draft-device ROCm0 \
  --spec-draft-ngl all \
  --spec-draft-type-k f16 \
  --spec-draft-type-v f16 \
  --spec-draft-n-max 6 \
  --spec-draft-n-min 0 \
  --spec-draft-p-min 0.0 \
  --spec-draft-p-split 0.20 \
  --parallel 1 \
  --metrics \
  --no-mmproj \
  --no-context-shift \
  --reasoning off \
  --reasoning-format none \
  --reasoning-budget 0 \
  --temp 0 \
  --top-p 0.95 \
  --top-k 20 \
  --repeat-penalty 1.0 \
  --seed 123

The profile in this repository is the exact env profile used for the HermesAgent-20 lane:

text

profiles/unsloth-qwen36-27b-mtp-rocmfp6-strix-quality-cap6-q8kv-rocm-hermes64k.env

Provenance

Item	Value
quant format	`Q6_0_ROCMFPX_STRIX_QUALITY`
ROCmFPX branch	`rocmfp6-strix-quality`
ROCmFPX commit	`7026d4ea51acb6e314526506eccdccdc31987855`
public report	`https://llm.ciru.ai/reports/rocmfp6-quality-research-report-20260624/`
local source filename	`Qwen3.6-27B-MTP-BF16-to-Q6_0_ROCMFPX_STRIX_QUALITY.gguf`
public filename	`Chadrockv2-Qwen3.6-27B-ROCmFP6-STRIX-QUALITY.gguf`

The local source filename is intentionally not used as the public artifact name. The uploaded GGUF uses the clean Chadrockv2 release filename shown above.

Limitations

—This is specifically AMD tuned, with Strix Halo as the measured target.
—The GGUF requires a ROCmFPX-aware llama.cpp runner.
—The recipe prioritizes agent quality and served decode speed, not smallest file size.
—Benchmark numbers are local Strix Halo measurements and depend on driver version, clocks, prompt shape, KV cache settings, and draft-token acceptance.

Credits

—[Qwen](https://huggingface.co/Qwen): Qwen3.6 27B base model family.
—[Unsloth](https://huggingface.co/unsloth): Qwen3.6 27B MTP GGUF source lineage and Q6 baseline used for same-source comparison.
—[Charlie / ROCmFPX](https://github.com/charlie12345/ROCmFPX): ROCmFPX tensor formats and llama.cpp runtime work.
—[Ciru Inference Lab](https://llm.ciru.ai/): AMD Strix Halo recipe tuning, quality evaluation, speed testing, and report publishing.