-
Fix device mismatch in measure_peak_memory: move model to device, guard CUDA peak-stat calls, capture loss
-
Replace mispredicting memory estimate with a measured preflight: real fwd+bwd on the longest example, auto-select max_seq_len
-
Fix CUDA OOM: max_seq_len 2048->1024, enable expandable_segments, add memory preflight calibrated to the observed 13.5GB failure
-
Support transformers 5.x: TrainingArguments adapter converts warmup_ratio to warmup_steps; drop unsupported kwargs with a warning; test covers both
-
Fix: resolved_precision now accepts the (major, minor) tuple from torch.cuda.get_device_capability; add tests module for GPU-only paths
-
Initial: pipeline, training, eval modules
-
initial commit
