qwen3-8b-tmax-sft-run2
Full supervised fine-tune of Qwen/Qwen3-8B on
vittorino/tmax-sft-cleaned, trained on a single NVIDIA DGX
Spark (GB10) with FP8 (torchao tensorwise) matmuls, torch.compile and adaptive
gradient checkpointing.
Training in progress. A snapshot is uploaded every 6 hours of training and
tagged h<hours>-step<optimizer step>; the finished model is tagged final-step<N>.
Load a specific snapshot with revision=<tag>:
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("vittorino/qwen3-8b-tmax-sft-run2", revision="h06-step...")
Intermediate snapshots are mid-schedule and not evaluated. FP8 training is experimental here; it will be compared against a BF16 run of the same recipe.