DaisyRose-27B ROCmFPX (AMD GPUs and Strix Halo)

Hardware-aware GGUF quants of DaisyRose-27B by RadixArk, built with the ROCmFPX weight formats for AMD GPUs (Vulkan and ROCm).

DaisyRose-27B is RadixArk's reinforcement-learning fine-tune of Qwen3.8-27B (dense, 64 layers, gated delta net with full attention every fourth layer, image input). The multi-token prediction layer of the original is kept in every file.

Not an official release. All credit for the model goes to RadixArk and the Qwen team.

Files

File Size Weights Fits
DaisyRose-27B-Q8_0_ROCMFPX.gguf 28.3 GB 8-bit ROCmFPX, token embedding and output Q8_0 Strix Halo, 32 GB AMD GPUs
DaisyRose-27B-Q6_K_ROCMFPX.gguf 22.5 GB 6-bit ROCmFPX, token embedding and output Q6_K Strix Halo, 32 GB AMD GPUs, 24 GB with short context
DaisyRose-27B-Q4_0_ROCMFP4_STRIX_O8.gguf 19.2 GB ROCmFP4, attention, delta net output and output layer 8-bit ROCmFPX 24 GB AMD GPUs, Strix Halo
DaisyRose-27B-IQ2_XXS-8GB.gguf 7.7 GB IQ2_XXS, output Q3_K, token embedding Q4_K 8 GB GPUs (7.0 GB of it on the GPU)
mmproj-DaisyRose-27B-F16.gguf 0.9 GB vision encoder, F16 add for image input

The ROCmFPX names start with a stock type name (Q8_0, Q6_K, Q4_0) only so the Hugging Face file list sorts them by bit width. Those weights are ROCmFPX formats and need the ROCmFPX fork; stock llama.cpp cannot load them.

About the 8 GB file. A 27B model has to come down to about 2 bits per weight to fit in 8 GB, and that costs a lot of quality (see the table: it is clearly weaker than the 4-bit file). It is for people who only have an 8 GB card. At this size our 2-bit ROCmFPX format (2.5 bits per weight) is too large, and mixes of it scored worse than plain IQ2_XXS at the same size, so this file is IQ2_XXS (only the multi-token prediction layer is 2-bit ROCmFPX). The token embedding stays in system memory, so about 7.0 GB goes to the GPU, which leaves room for about 8K tokens of context.

Quality

KL divergence against the BF16 weights, 40 chunks of 512 tokens of English Wikipedia text (wikitext-2). Each file was expanded to F16 and scored on the same GPU backend as the BF16 reference, so the numbers show the weight format only. Lower is better. "Same top token" is how often the file's most likely next token matches BF16.

Build Size Perplexity KL to BF16 Same top token
BF16 (reference, not published) 54.6 GB 5.941
Q8_0_ROCMFPX 28.3 GB 5.946 0.00025 99.3%
Q6_K_ROCMFPX 22.5 GB 5.947 0.0020 98.0%
Q4_0_ROCMFP4_STRIX_O8 19.2 GB 5.998 0.019 94.0%
IQ2_XXS-8GB 7.7 GB 8.409 0.42 72.6%
stock Q8_0 (comparison, not uploaded) 29.0 GB 5.946 0.00037 99.2%
stock Q6_K (comparison, not uploaded) 22.4 GB 5.940 0.0018 97.9%
stock Q4_K_M (comparison, not uploaded) 16.8 GB 5.942 0.014 94.6%

At 8 bits the ROCmFPX file is smaller than stock Q8_0 and closer to BF16. At 6 bits the two are about even. At 4 bits stock Q4_K_M is smaller and closer to BF16 than this ROCmFP4 file; the ROCmFP4 file is the one laid out for the AMD ROCmFP4 kernels, so compare the speed on your card before choosing.

Every file also passed a greedy chat check before upload ("What is the capital of France?" gives "Paris").

Running it

Qwen3.8 support with ROCmFPX formats and tensor scales is in the ROCmFPX fork of llama.cpp (main). Build it with Vulkan or ROCm/HIP, then for example:

llama-server -m DaisyRose-27B-Q6_K_ROCMFPX.gguf --mmproj mmproj-DaisyRose-27B-F16.gguf -ngl 99 -c 32768 --jinja

On an 8 GB card, keep the context short and the vision encoder on the CPU:

llama-server -m DaisyRose-27B-IQ2_XXS-8GB.gguf -ngl 99 -c 8192 -fa on --jinja
llama-server -m DaisyRose-27B-IQ2_XXS-8GB.gguf --mmproj mmproj-DaisyRose-27B-F16.gguf --no-mmproj-offload -ngl 99 -c 4096 -fa on --jinja

The chat template is the one from Qwen3.8 (thinking on by default).

How it was made

  • convert_hf_to_gguf.py to BF16, then llama-quantize with ROCmFPX tensor scales on (a power-of-two scale per tensor, stored as .scale).
  • Importance matrix from calibration_datav3 on the BF16 weights.
  • Checked on the CPU only so far; Vulkan and ROCm have not been tried with this model yet. Please report in the Community tab how it runs.

Credits and license

Apache-2.0. See LICENSE.

Downloads last month
17
GGUF
Model size
27B params
Architecture
qwen35
Hardware compatibility
Log In to add your hardware

2-bit

4-bit

6-bit

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Harsha001/DaisyRose-27B-ROCmFPX

Base model

Qwen/Qwen3.8-27B
Quantized
(41)
this model