{ "updated_at": "2026-11-05T11:29:44.888395+00:00", "baseline_seconds_per_update": 96.92361264521189, "optimized_seconds_per_update": 57.35987377807032, "observed_speedup": 1.6897459192503919, "measurement_note": "Early post-startup updates; example order and sequence lengths differ. Long-run throughput may vary.", "runtime_configuration": { "world_size": 2, "effective_batch_size": 32, "train_examples": 23122, "validation_examples": 2892, "packing": false, "sampling": "group_by_length", "average_tokens_across_devices": true, "pad_to_multiple_of": 8, "content_max_length": 1730, "fused_convolution_layers": 48 }, "kernel_checks_passed": 48, "kernel_max_relative_l2_error": { "torch.float32": 2.3887807287792384e-07, "torch.bfloat16": 0.0029445830732584 }, "distributed_gradient_max_absolute_error": 5.551115123125783e-17, "training_pid": 401763, "launcher_pid": 401751, "current_step": 4, "restarted_from_base": true, "baseline_log": "training.single_gpu.log", "active_log": "training.log", "unchanged": [ "eligible training and validation data", "effective batch size of 32", "content limit of 1730 tokens", "five epochs", "learning rate and scheduler", "LoRA rank, alpha and dropout", "response-only labels", "evaluation and checkpoint frequency", "early stopping" ], "numerical_note": "Distributed arithmetic, batch order and normal bfloat16 rounding can change the exact trajectory; bitwise equivalence is not claimed." }