Releases·Tool Release
Developer releases R9V custom RDNA4 kernels, tripling Qwen3.8-Flash-Next speed on dual R9700s
A developer released R9V, a set of hand-tuned kernels for AMD RDNA4 (R9700) GPUs, plus quantized model packages on Hugging Face. Applied to vLLM-Radiance, the kernels push Qwen3.8-Flash-Next to 78 tok/s TG256 on two R9700s, roughly 3x faster than before. The release targets local-inference practitioners on AMD hardware and includes an early-stage inference engine for dense models.
49 engagement1 sourceMon, Aug 31, 2026, 01:13 AM
AI Overview
- R9V kernels target RDNA4 and plug into vLLM-Radiance; repo at github.com/Dyluhn/R9V
- Qwen3.8-Flash-Next Unsloth IQ4_XS package runs with tensor parallelism on 2 R9700s, MTP, SSD n-gram, 128k context, and vision
- Measured TG256 of 78 tok/s (~3x increase) and PP81 on dual R9700s
- Aimed at dual R9700 setups with 48GB RAM or higher, but tinkerers welcome
- A separate R9V package serves Muse Glimmer 30B (Q8/Q4) on a single R9700
- The author's own inference engine for dense models is in early stages
Related
Social·Community
zzpanic ships final vllm-radiance build for single R9700 GPU running qwen3.8-27b as StillDeadcode's new radiance engine supersedes old builds
A community developer (zzpanic) published the final revision of a personal vllm-radiance inference build targeting a single AMD R9700 GPU, as StillDeadcode's standalone radiance inference engine supersedes old vllm-radiance builds. The build adds startup caching and a KV-offload feature enabling long-context multi-agent work on one card.
A community developer (zzpanic) published the final revision of a personal vllm-radiance inference build targeting a single AMD R9700 GPU, as StillDeadcode's standalone radiance inference engine supersedes old vllm-radiance builds. The build adds startup caching and a KV-offload feature enabling long-context multi-agent work on one card.
15 engagement1 sourceredditSun, Oct 11, 2026, 08:47 AM
Releases·Tool Release
Dev's vLLM MXFP4 build runs NVFP4 Qwen3.8-27B on AMD R9700 at 5,809 tok/s prefill, 276 tok/s decode
A developer extended their MXFP4 vLLM build on the Radiance image with online NVFP4-to-MXFP4 conversion, letting NVFP4-quantized models use the MXFP4 fast path on AMD R9700 GPUs. Running Unsloth's Qwen3.8-27B NVFP4 quant, they measured 5,809 tok/s prefill and 276 tok/s decode via their BetterBench harness.
A developer extended their MXFP4 vLLM build on the Radiance image with online NVFP4-to-MXFP4 conversion, letting NVFP4-quantized models use the MXFP4 fast path on AMD R9700 GPUs. Running Unsloth's Qwen3.8-27B NVFP4 quant, they measured 5,809 tok/s prefill and 276 tok/s decode via their BetterBench harness.
69 engagement1 sourceredditTue, Sep 15, 2026, 08:51 PM
Social·Community
Reddit user benchmarks Qwen 3.8 Flash Next IQ4_XS vs Qwen3.8 27B FP8 on MMLU-Pro, GPQA, GSM8K
A Reddit user ran Codex-driven evals comparing Qwen 3.8 Flash Next IQ4_XS (served via vllm and r9v) against Qwen 3.8 27B FP8 (vllm and radiance). Flash-Next won on MMLU-Pro, GPQA Diamond, GSM8K, and German MGSM, with IFEval tied. The 27B FP8 model was roughly 3.5x faster to run the trial and scaled better with concurrency, though the poster still prefers the low quant for single-user use.
A Reddit user ran Codex-driven evals comparing Qwen 3.8 Flash Next IQ4_XS (served via vllm and r9v) against Qwen 3.8 27B FP8 (vllm and radiance). Flash-Next won on MMLU-Pro, GPQA Diamond, GSM8K, and German MGSM, with IFEval tied. The 27B FP8 model was roughly 3.5x faster to run the trial and scaled better with concurrency, though the poster still prefers the low quant for single-user use.
38 engagement1 sourceredditFri, Sep 25, 2026, 06:32 AM
Releases·Tool Release
R9V Update: ~100 tok/s TG on Qwen3.8 Flash Next IQ4_XS, dual R9700 + 128GB RAM; crash fixes via n-gram SSD streaming, Q4_K_XL support
The R9V developer released an update improving RDNA4 inference performance and stability, reporting ~100 tok/s in llama.cpp text generation on Qwen3.8 Flash Next IQ4_XS across two R9700 GPUs with 128GB RAM. The update fixes crashes via n-gram SSD streaming and adds Q4_K_XL support. The developer says further tuning will pause while they ship a separate inference engine and a research/site-builder app.
The R9V developer released an update improving RDNA4 inference performance and stability, reporting ~100 tok/s in llama.cpp text generation on Qwen3.8 Flash Next IQ4_XS across two R9700 GPUs with 128GB RAM. The update fixes crashes via n-gram SSD streaming and adds Q4_K_XL support. The developer says further tuning will pause while they ship a separate inference engine and a research/site-builder app.
76 engagement1 sourceredditMon, Sep 14, 2026, 03:01 AM
Releases·Tool Releasefeatured
Open-source kernel hits 78,498 output tok/s for Qwen3.6-35B-A3B on 8 AMD MI350X GPUs
A developer team optimized AMD's software stack down to the kernel level for Qwen3.6-35B-A3B on MI350X GPUs, achieving 78,498 output tokens per second across 8 GPUs. The work aims to close the software gap with NVIDIA's CUDA ecosystem, leveraging MI350X's raw compute advantage over B200.
A developer team optimized AMD's software stack down to the kernel level for Qwen3.6-35B-A3B on MI350X GPUs, achieving 78,498 output tokens per second across 8 GPUs. The work aims to close the software gap with NVIDIA's CUDA ecosystem, leveraging MI350X's raw compute advantage over B200.
80 engagement1 sourceredditWed, Aug 26, 2026, 04:12 AM