The current record (#1, 6m05s) runs eager-mode PyTorch. torch.compile on the model should cut per-step time with zero effect on the math, so it's close to a free speedup on top of any existing technique.
Starting point: copy submissions/001-saivineeth147/, add model = torch.compile(model) after get_peft_model, watch the first-step compile cost vs. the per-step savings (short runs may not amortize it — that trade-off is the interesting part).
Bar: beat 6m05s, clear 57% on 3 fresh seeds. First verified record holds this lane permanently. Free to run on Modal — see CONTRIBUTING.md.
The current record (#1, 6m05s) runs eager-mode PyTorch.
torch.compileon the model should cut per-step time with zero effect on the math, so it's close to a free speedup on top of any existing technique.Starting point: copy
submissions/001-saivineeth147/, addmodel = torch.compile(model)afterget_peft_model, watch the first-step compile cost vs. the per-step savings (short runs may not amortize it — that trade-off is the interesting part).Bar: beat 6m05s, clear 57% on 3 fresh seeds. First verified record holds this lane permanently. Free to run on Modal — see CONTRIBUTING.md.