#1
Distributed Training Infrastructure and Model Compatibility Fixes
41 issues · 7.2 demand · Performance
Users are experiencing multiple training infrastructure issues when fine-tuning large language and vision-language models, including distributed training errors (FSDP2, Ray, NCCL), data processing bugs, and model-specific compatibility problems with Qwen3.5 and Gemma4. These issues prevent reliable training across multi-GPU and specialized hardware (Apple Silicon MPS, Ascend NPUs) setups, requiring fixes to resource allocation, parameter offloading, and model loading to ensure stable fine-tuning workflows.