Quacken-27B-FP8

The Rock8 - Got any weights? 💪🦆

Native fp8 E4M3 GGUF of Qwen3.6-27B (a GatedDeltaNet hybrid) for AMD RDNA4 (gfx1201 - Radeon AI PRO R9700 / RX 9070 / 9070 XT / W-series), quantized with AMD Quark from the full-precision BF16 weights by The Rock8.

The Rock8's llama.cpp fork runs this fp8 on RDNA4's native WMMA fp8 tensor cores (prefill) and v_dot4_f32_fp8_fp8 (decode) - not a dequant-to-f16 fallback. At 27B this is a 2-GPU model (tensor-split across two 32 GB R9700s).

What it is

• Format: fp8 E4M3 (F8E4M3), block-scaled, produced by AMD Quark from BF16.
• Target: AMD RDNA4 / gfx1201; runs across two 32 GB cards (Radeon AI PRO R9700).
• Runtime: The Rock8 (llama.cpp fork with native RDNA4 fp8 kernels) on TheRock ROCm 7.13.
• File: Qwen3.6-27B-Quark-F8E4M3.gguf (29.3 GiB).

Source model + license

• Source: Qwen3.6-27B (Qwen).
• License: Apache-2.0 (the source repo bundles an Apache-2.0 LICENSE; redistribution of this quantized derivative is permitted with attribution). This is a derivative work.

Validation (real gfx1201 hardware)

Metric	Value
Perplexity (wikitext, 20 chunks, n_ctx=512)	7.14
Prefill pp512 (2-GPU)	1251.6 t/s
Decode tg128 (2-GPU)	18.52 t/s
Coherence check ("dried grape" -> "raisin")	Pass

Benched tensor-split across two R9700s (gfx1201). This is the authentic Quark build validated to load and generate coherently.

Run it

llama.cpp (The Rock8 fork) - 2-GPU

# -ngl 999 lets llama.cpp see and split across both R9700s
llama-cli   -m Qwen3.6-27B-Quark-F8E4M3.gguf -ngl 999 -p "What do you call a dried grape? Answer in one word."
llama-bench -m Qwen3.6-27B-Quark-F8E4M3.gguf -ngl 999 -p 512 -n 128

Lemonade appliance (container)

podman run -d --rm --runtime crun --name lemonade \
  --device /dev/kfd --device /dev/dri \
  --group-add keep-groups --security-opt seccomp=unconfined \
  -v /path/to/quacken-27b:/models:ro \
  -e MODEL=/models/Qwen3.6-27B-Quark-F8E4M3.gguf -e MODEL_NAME=Quacken-27B-FP8 \
  -p 13305:13305 \
  ghcr.io/the-monk/the-rock8:rdna4-tr713 serve
# note: 27B needs both GPUs - do NOT pin HIP_VISIBLE_DEVICES to a single card

Container (same image on each registry; --runtime crun is required for GPU): ghcr.io/the-monk/the-rock8:rdna4-tr713 - docker.io/gorilla4x/the-rock8:rdna4-tr713 - quay.io/the-monk/the-rock8:rdna4-tr713 (images may not be pushed to every registry yet).

Omni mode - MTP self-speculative decode (Lemonade config, 2.43x decode)

Omni is not a separate model - it's a Lemonade configuration on this same fp8 checkpoint. Qwen3.6-27B ships a built-in MTP (multi-token-prediction) head, so it can speculate against itself - no draft model, no extra VRAM. The Rock8 verifies the speculated tokens on the fp8 WMMA tensor cores. Result on gfx1201 (single-stream, greedy):

Config	Decode t/s	vs raw
Raw fp8 decode	18.67	1.00x
Omni: fp8 + MTP self-spec (draft-n=8)	45.45	2.43x

Byte-identical to greedy at n=8 (the ship setting). This is the single-GPU latency champion.

Lemonade recipe_options (drop into user_models.json)

"Quacken-27B-FP8-Omni": {
  "checkpoint": "Gorilla4X/Quacken-27B-FP8",
  "recipe": "llamacpp",
  "recipe_options": {
    "llamacpp_backend": "rocm",
    "llamacpp_args": "-ngl 999 --spec-type draft-mtp --spec-draft-n-max 8"
  }
}

The --spec-type draft-mtp flag routes decoding through the model's own MTP head; Lemonade serves it on :13305 like any other model. (Raw llama.cpp equivalent: llama-server -m Qwen3.6-27B-Quark-F8E4M3.gguf -ngl 999 --spec-type draft-mtp --spec-draft-n-max 8.)

The async 2-GPU pipeline (LLAMA_SPEC_ASYNC=2) does not compose with MTP on this hybrid-SSM target - see Bonsai-8B-Ternary-RDNA4 for the async lever, which needs a dense target + a cheap ternary draft.

The Rock8 - RDNA4 fp8 (links)

• GitHub: The-Rock8 - kernels, patch series, appliance recipe, full feature doc.
• Collection: The Rock8 - RDNA4 fp8.
• Sibling models: Quacken-8B-FP8 - Quacken-R1-14B-FP8 - Quacken-35B-A3B-FP8 (MoE) - Quacken-Ornith-35B-FP8 - Bonsai-8B-Ternary-RDNA4 (async spec-decode showcase).

Every artifact links to the others - land on any one, reach them all.