GnosisA personal journal of research, code & thought.https://gnosis-ay9.pages.dev/zh-CNThree-Layer Progressive Optimization: Systematic Limit Approximation for Consumer-GPU GRPO Post-Traininghttps://gnosis-ay9.pages.dev/writing/2026-08-15-rdna3-three-layer-grpo/https://gnosis-ay9.pages.dev/writing/2026-08-15-rdna3-three-layer-grpo/DGBB, ZSBR, and SFOC on one RX 7900 XTX: 795.5 tok/s (97.6% of the wall argmax), +7.8 pp from ZSBR, and 23.2% → 45.6% held-out in 8.5 hours of pure RL.Sat, 15 Aug 2026 00:00:00 GMTCharacterizing and Optimizing LLM Post-Training Efficiency on Consumer AMD RDNA3 GPUshttps://gnosis-ay9.pages.dev/writing/2026-08-14-rdna3-dual-peak-mfu/https://gnosis-ay9.pages.dev/writing/2026-08-14-rdna3-dual-peak-mfu/Dual-peak MFU, gap attribution, and a tuning recipe for GRPO on a single RX 7900 XTX. Unsloth +5.2% vs HuggingFace; 4-bit −40.7%; 97.7% of the GEMM-to-GRPO gap is generation.Fri, 14 Aug 2026 00:00:00 GMT