license: apache-2.0
base_model: ricursiveAI/Ripple-65B-Preview
base_model_relation: quantized
pipeline_tag: text-generation
library_name: gguf
tags:
- gguf
- ripple
- rocmfpx
- amd
- strix-halo
Ripple-65B-Preview ROCmFPX
ROCmFPX GGUF quants of ricursiveAI/Ripple-65B-Preview, a dense 65B model, for AMD GPUs (Strix Halo) with the ROCmFPX llama.cpp fork.
Files
| File | Size | Mean KLD vs Q8_0 | Same top token | Fits |
|---|---|---|---|---|
Ripple-65B-Preview-Q8_0_ROCMFPX.gguf |
66.8 GB | 0.0043 | 97.6% | 128 GB Strix Halo |
Ripple-65B-Preview-Q6_K_ROCMFPX.gguf |
53.6 GB | 0.0062 | 96.9% | 128 GB Strix Halo, long context |
Ripple-65B-Preview-Q4_0_ROCMFP4_STRIX_A8.gguf |
41.5 GB | 0.022 | 94.0% | 64 GB and up |
stock Q4_K_M, for comparison (not in this repo) |
39.6 GB | 0.038 | 92.1% |
Q4_0_ROCMFP4_STRIX_A8keeps attention and the output layer at 8 bit (ROCmFPX Q8) and the rest at 4 bit (ROCmFP4). Plain 4-bit attention lost too much on this model (KLD 0.086), so we do not ship it. At 2 GB more than stock Q4_K_M it has about 42% less error.- How it was measured:
llama-perplexity, 40 chunks of 512 tokens of wikitext-2, against a stock Q8_0 of the original BF16 weights (BF16 does not fit our 128 GB test machine; stock Q8_0 is very close to BF16). Each file was expanded to Q8_0 and scored on an NVIDIA GB10; this adds about 0.0016 KLD (Q8_0_ROCMFPX scored directly on the CPU: 0.0027, same top token 98.0%). - A chat check on the reference answered correctly and stopped cleanly. Not yet run on a Strix Halo: please report speed and quality in the Community tab.
Run
Ripple's architecture (ricursive) is not in upstream llama.cpp yet. The ROCmFPX fork has it on main (port of llama.cpp PR #31692), together with the ROCmFPX types these files use.
git clone https://github.com/HarshaVardhan3002/ROCmFPX
cd ROCmFPX
# ROCm / HIP on Strix Halo
cmake -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1151
# or Vulkan
# cmake -B build -DGGML_VULKAN=ON
cmake --build build --config Release -j
./build/bin/llama-server -m Ripple-65B-Preview-Q6_K_ROCMFPX.gguf -ngl 99 -fa on -c 32768 --jinja
- Always pass
-c. The model's default context is 2M tokens, and llama.cpp then tries to allocate a KV cache that does not fit. --jinjauses the chat template stored in the file (ChatML with reasoning and tool calls).
Faster with the drafter (DFlash2)
Ripple-65B-Preview-DFlash2-BF16.gguf (8.1 GB) is ricursiveAI's own draft model, ricursiveAI/Ripple-65B-Preview-DFlash2, converted to GGUF without changes (BF16). It guesses up to 16 tokens at a time from Ripple's own hidden states, and Ripple checks the guesses in one pass. Ripple still chooses every token, so the answers stay Ripple's. It needs the fork from 7 November 2026 or later.
./build/bin/llama-server -m Ripple-65B-Preview-Q6_K_ROCMFPX.gguf -md Ripple-65B-Preview-DFlash2-BF16.gguf \
-ngl 99 -ngld 99 -fa on -c 16384 --jinja --spec-type draft-dflash --spec-draft-n-max 7
Measured on an NVIDIA GB10 with a stock Q8_0 of Ripple, greedy decoding, llama-speculative-simple:
| Prompt | Without drafter | --spec-draft-n-max 15 |
--spec-draft-n-max 7 |
|---|---|---|---|
| Python function | 3.4 tokens/s | 14.1 (4.1x) | 13.8 |
| Explain why the sky is blue | 3.4 | 6.2 | 6.9 (2.0x) |
| Short maths question | 3.4 | 8.9 (2.6x) | 7.3 |
| Short story | 3.4 | 4.9 | 5.4 (1.6x) |
- Code gains the most, free prose the least.
7is the safer default; use15for code. - In 6 of 8 runs the text was exactly the same as without the drafter. In the other 2, one word changed where Ripple saw two choices as almost equally likely (Ripple checks several tokens at once, and that rounds very slightly differently), and the answer went on from there.
- The drafter was trained on 16K-token sequences. Past about 16K tokens of context it will guess worse, so the speed gain shrinks; Ripple still checks every token.
- Not yet run on a Strix Halo. The drafter adds about 8 GB, so use it with the Q6 or 4-bit file on a 128 GB machine.
Credits
- Original model: ricursiveAI/Ripple-65B-Preview, Apache-2.0.
- Ripple llama.cpp port: llama.cpp PR #31692.
- Draft model: ricursiveAI/Ripple-65B-Preview-DFlash2, Apache-2.0. DFlash2 runtime from llama.cpp PR #27342.