Anyone else getting garbage results fine-tuning on H200 in FP8 mode?

<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1s59rct/rys_part_3_llms_think_in_geometry_not_language/"> <img alt="RYS Part 3: LLMs think in geometry, not

<!-- SC_OFF --><div class="md"><p>Would implementing runtime upcasting from MXFP4 to FP8, performing shard-wise upcasting and storing in FP8, and then conducting LoRA fine-tuning

<!-- SC_OFF --><div class="md"><p>That foodtruck bench post showing deepseek v4 matching gpt-5.2 at 17x cheaper got me thinking. if frontier cloud models are that overpriced for

<!-- SC_OFF --><div class="md"><p>While doing QLoRA fine-tuning (using Unsloth), increasing LoRA rank from 8,16 → 64 often doesn’t improve performance.</p> <p>It feels like it

<!-- SC_OFF --><div class="md"><p>Been building and testing LLM workflows for a bit and started digging into logs more closely.</p> <p>Lo and behold!</p> <p>a pretty large chunk