A user is asking how to replicate Hugging Face SFTTrainer's auto_find_batch_size=True behavior, which automatically reduces batch size after a CUDA out-of-memory error, when training across multiple GPUs on a single node with Accelerate and FSDP2. The question notes the single-GPU workflow works but seeks equivalent automatic batch-size discovery for the multi-GPU FSDP2 setup.
Hi, For single-GPU training, I’m using Hugging Face SFTTrainer with auto_find_batch_size=True, which automatically reduces the batch size after a CUDA OOM until it finds a batch size that works. I would like to have similar behavior when training on multiple GPUs on a single node using accelerate la
EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.