🚀 You don't need a bigger GPU cluster to make long Qwen prompts feel fast. Ten… · @QwenAIFans · TwKit

@QwenAIFans

🚀 You don't need a bigger GPU cluster to make long Qwen prompts feel fast. TensorFold 1.1.4 just rewrote the default Qwen3.8-27B CUDA recipe (1.4k⭐). On one DGX Spark (GB10), cold prefill: • 2k: 834 → 1,281 tok/s (+54%) • 8k: 612 → 1,266 (+107%) • 32k: 326 → 1,165 (+258%) • 64k: 203 → 1,045 (+416% ≈ 5×) Decode only ~+2%. Quality gate: 27/40 exact matches vs the old recipe. Caveat: bit-level replies can differ from 1.1.3; numbers are GB10 CUDA only. via GitHub release 👇 https://github.com/ashhart/TensorFold/releases/tag/v1.1.4

曝光 779 · 评论 0 · 点赞 12 · 书签 3