Releases·Tool Release
dual 7900 xtx - some guy made a pretty optimized fork of lamacpp optimized for this setup Qwen 3.8 Q8 at 82 tokens / seconds decode
A community-optimized llama.cpp fork targeting dual AMD Radeon 7900 XTX setups boosts Qwen 3 8B Q8 decode speed from ~28 tok/s (vanilla llama.cpp with Vulkan) to 82 tok/s. A Reddit user shared the GitHub repo, reporting stable results at 60k context, which could interest practitioners running local inference on consumer AMD hardware.
1 engagement1 sourceFri, Sep 18, 2026, 02:00 AM
AI Overview
- Fork nasone32/llama.cpp-RDNA3-7900xtx-opt boosts Qwen 3.8 Q8 decode from 28 tok/s (vanilla llama.cpp + Vulkan) to 82 tok/s at 60k context
- Setup uses two Radeon 7900 XTX GPUs on a consumer PC running Linux
Related
Products·launched
Custom llama.cpp build optimized for AMD 7900 XTX (single or dual) targeting Qwen 3.8 Next and 27B models
A custom-compiled llama.cpp build tuned for one or two AMD Radeon 7900 XTX GPUs, with optimizations for PCIe x4 bandwidth constraints and tensor parallelism across two cards. It improves prompt processing and generation speeds for Qwen 3.8 Next and 27B models, serving local LLM hobbyists running inference on consumer AMD hardware.
A custom-compiled llama.cpp build tuned for one or two AMD Radeon 7900 XTX GPUs, with optimizations for PCIe x4 bandwidth constraints and tensor parallelism across two cards. It improves prompt processing and generation speeds for Qwen 3.8 Next and 27B models, serving local LLM hobbyists running inference on consumer AMD hardware.
92 engagement1 sourceredditTue, Sep 8, 2026, 08:22 AM
Social·Community
Qwen3.8 27B on RX 7900 XTX: Ollama ROCm vs llama.cpp Vulkan results
Practitioners shared local-inference benchmarks for Qwen3.8 27B on AMD consumer GPUs: one ran Ollama/ROCm vs llama.cpp Vulkan on an RX 7900 XTX, and another weighs upgrading from dual RTX 2060 12GB to an RX 6800 + 6800 XT pair. The posts serve as practical reference points for AMD multi-GPU llama.cpp performance.
Practitioners shared local-inference benchmarks for Qwen3.8 27B on AMD consumer GPUs: one ran Ollama/ROCm vs llama.cpp Vulkan on an RX 7900 XTX, and another weighs upgrading from dual RTX 2060 12GB to an RX 6800 + 6800 XT pair. The posts serve as practical reference points for AMD multi-GPU llama.cpp performance.
Updated Tue, Sep 8, 2026, 10:47 PM — A follow-up post weighs RX 6800 + 6800 XT (32GB) as an upgrade over 2x RTX 2060 12GB for llama.cpp.
90 engagement2 sourcesredditFri, Sep 4, 2026, 08:16 PM
Releases·Tool Release
Developer updates gfx906 llama.cpp fork for Radeon VII / MI50 / MI60 with +14% PP and adaptive Flash Attention
A developer updated their gfx906 llama.cpp fork for AMD Radeon VII / MI50 / MI60 GPUs after earlier criticism over insufficient benchmarks. Reworking the small-Q Flash Attention path and adding adaptive native/convert selection yielded +14.1% first-batch PP and +9% long-context fill vs upstream llama.cpp mainline. The update matters to practitioners running local inference on older AMD CDNA/GFX9 hardware, where DFlash2 adoption on a Qwen 3.8 27B production setup had made some prior optimizations regress.
A developer updated their gfx906 llama.cpp fork for AMD Radeon VII / MI50 / MI60 GPUs after earlier criticism over insufficient benchmarks. Reworking the small-Q Flash Attention path and adding adaptive native/convert selection yielded +14.1% first-batch PP and +9% long-context fill vs upstream llama.cpp mainline. The update matters to practitioners running local inference on older AMD CDNA/GFX9 hardware, where DFlash2 adoption on a Qwen 3.8 27B production setup had made some prior optimizations regress.
68 engagement2 sourcesredditTue, Sep 1, 2026, 02:11 PM
Social·Community
Users report Qwen 3.8 27B slower token speeds but better results on RTX 3090
Qwen 3.8 27B users on RTX 3090 report lower token generation speeds compared to Qwen 3.6, with one user seeing 34-36 tokens per second versus 50-60 before. However, some find the results better and faster overall, and others note identical performance with the same architecture and quant.
Qwen 3.8 27B users on RTX 3090 report lower token generation speeds compared to Qwen 3.6, with one user seeing 34-36 tokens per second versus 50-60 before. However, some find the results better and faster overall, and others note identical performance with the same architecture and quant.
110546 engagement4 sourcesredditFri, Aug 14, 2026, 06:19 PM
Social·Community
Qwen 3.8 27B runs on AMD NPUs via FastFlowLM at 1 token/s decode
Users run Qwen 3.8 27B on AMD NPUs via FastFlowLM at 1 token/s decode
Users run Qwen 3.8 27B on AMD NPUs via FastFlowLM at 1 token/s decode
51 engagement1 sourceredditWed, Sep 30, 2026, 06:45 PM