2026-11-13 · America/Los_Angeles · 社区动态 · #8
I let qwen3.8-27B build a game overnight on an RTX 5060 Ti 16GB: ASCII-pruned Unsloth Q3_K_XL + DFlash; 20-40%+ decode speed on larger prompts with 100k+ context
I've been experimenting with qwen3.8 27B on an RTX 5060 Ti 16GB, trying to find a practical balance between context length, decode speed, and vram usage for local agentic workloads. one optimization I found was ascii vocabulary pruning, which reduced the vocabulary from 248,320 to 129,006 tokens while retaining byte fallback for non ascii text. This freed up enough vram to fit a compatible Q2_K DFlash draft alongside the Unsloth Q3_K_XL target. Some benchmarks on a few set of sample prompts: Full vocabulary + MTP2: Highest working context allocation tested: ~104K Decode: 34.90 tok/s at ~50K…
热度 46.0 / 100;排名与评分保留该期记录。