Exploring Runtime Upcasting from MXFP4 to FP8 for Efficient LoRA Fine-Tuning wit

<!-- SC_OFF --><div class="md"><p>Running Llama-3-70B with long contexts and hitting HBM limits constantly. 128K context burns approx 40GB just for KV cache — half the H100 gone

<!-- SC_OFF --><div class="md"><p>The modern ML (LLM) compiler stack is brutal. TVM is 500K+ lines of C++. PyTorch piles Dynamo, Inductor, and Triton on top of each other. I built

<!-- SC_OFF --><div class="md"><p>Every output token the model generates requires reading the entire KV cache from High Bandwidth Memory.</p> <p>Here's what that actually looks

<!-- SC_OFF --><div class="md"><p>Been losing my mind trying to figure out why my H200 fine-tunes look fine during training then perform terribly. Validation loss looks