Quacken-27B-NVFP4

Quacken-27B-NVFP4 (mixed FP4 + FP8) 🦆

The Rock8 — Got any weights? 💪

A mixed-precision GGUF of Qwen3.6-27B for AMD RDNA4 (gfx1201 — Radeon AI PRO R9700 / RX 9070 / 9070 XT / W-series): NVFP4 (4-bit) MLPs + native FP8 (E4M3) attention/lm_head. Ingested from unsloth/Qwen3.6-27B-NVFP4 with The Rock8's mixed-precision compressed-tensors converter, then run on RDNA4 via native fp8 WMMA + the NVFP4 int8/dp4a path.

Why it's interesting — smaller, more accurate, faster decode than fp8

Benched on one R9700 (gfx1201), vs our own native-fp8 build of the same base:

• Single-card — the FP4 MLPs shrink it to 23 GB, so it runs on one R9700 (the fp8 build needs two).
• Better accuracy — PPL 6.88 beats the fp8 build's 7.14.
• +30% decode — FP4 MLPs move far less data in the memory-bound decode phase.
• Honest tradeoff: prefill ~15% slower than pure fp8. RDNA4 has no native FP4 matmul (that's CDNA4/MI350), so the NVFP4 tensors run the int8/dp4a fallback — the win here is memory & decode bandwidth, not FP4 tensor-core acceleration. Best for interactive / decode-heavy use.

What's inside (verified tensor histogram)

• 168 NVFP4 tensors (8.4 GB) — mlp.{gate,up,down}_proj, layers 0–55 (E2M1, group-16, E4M3 scale)
• 233 F8E4M3 tensors (11.3 GB) — attention q/k/v/o, linear-attn proj, lm_head, layers 56–63 MLP
• F16 remainder (3.4 GB) — SSM/conv1d/norms + vision encoder + MTP head (intentionally unquantized)

This is the faithful mixed quant — the FP8 group is preserved as native F8E4M3, not dequantized to BF16.

Run it (The Rock8 fork, single GPU)

llama-cli   -m Qwen3.6-27B-NVFP4-FP8.gguf -ngl 999 -p "What do you call a dried grape? Answer in one word."
llama-bench -m Qwen3.6-27B-NVFP4-FP8.gguf -ngl 999 -p 512 -n 128

Source & license

Base model Qwen/Qwen3.6-27B (Apache-2.0). NVFP4 quantization by Unsloth; this repo re-packages that checkpoint as a faithful mixed FP4/FP8 GGUF for RDNA4. Apache-2.0.

The Rock8 — RDNA4 fp8

• GitHub: The-Rock8
• Collection: The Rock8 — RDNA4 fp8
• Siblings: Quacken-27B-FP8 · Bonsai-8B-Ternary-RDNA4

🦆 Got any weights?