Matched global sample indices, token/frame counts, and training settings: PASS

Measured optimizer updates; excludes warmup, checkpoint saving, and periodic logging.
Run | BS | max-rank wall s/update | samples/s | padding % | VGGT fallback %
run1_bs4 | 4 | 14.674 | 4.361 | 7.72 | 40.00
run2_bs8 | 8 | 15.416 | 4.152 | 9.01 | 73.75
run3_bs8 | 8 | 15.348 | 4.170 | 9.01 | 73.75
run4_bs4 | 4 | 14.716 | 4.349 | 7.72 | 40.00

Stage timings: CUDA stream elapsed seconds/update, from the slowest wall-time rank.
Nested rows overlap: vision/geometry/language are within forward; VGGT is within geometry;
DeepSpeed backward/step are within backward_and_engine_step. Do NOT sum all rows.
Stage | run1_bs4 | run2_bs8 | run3_bs8 | run4_bs4
batch_fetch_and_count | 0.0500 | 0.0496 | 0.0497 | 0.0494
prepare_inputs | 0.0010 | 0.0009 | 0.0009 | 0.0010
vision_forward | 0.5875 | 0.5741 | 0.5768 | 0.5832
geometry_total | 4.6608 | 5.4370 | 5.3886 | 4.7035
vggt_single | 2.3606 | 4.4515 | 4.3149 | 2.4984
vggt_batched | 2.2991 | 0.9838 | 1.0720 | 2.2039
language_forward | 2.7816 | 2.7305 | 2.7308 | 2.7825
forward_other | 0.0108 | 0.0086 | 0.0087 | 0.0113
forward_and_loss | 8.0407 | 8.7501 | 8.7048 | 8.0805
backward_and_communication | 6.4247 | 6.4881 | 6.4649 | 6.4265
deepspeed_step | 0.0875 | 0.0877 | 0.0881 | 0.0875
backward_and_engine_step | 6.5123 | 6.5759 | 6.5531 | 6.5142
gradient_clipping | 0.0001 | 0.0001 | 0.0001 | 0.0001
optimizer | 0.0000 | 0.0000 | 0.0000 | 0.0000

CPU batch-fetch/count seconds/update (includes any Accelerate transfer/counting): run1_bs4=0.0500, run2_bs8=0.0497, run3_bs8=0.0497, run4_bs4=0.0495
run1_bs4: per-rank wall seconds/update={'0': 14.673566439468413, '1': 14.673353948444127, '2': 14.673329481016845, '3': 14.67335846312344}; frames={'9': 1152, '1': 22, '3': 17, '7': 18, '5': 13, '4': 14, '2': 13, '8': 17, '6': 14}
run2_bs8: per-rank wall seconds/update={'0': 15.414719466585666, '1': 15.41470687771216, '2': 15.41471011918038, '3': 15.414744078554213}; frames={'9': 1152, '7': 18, '4': 14, '5': 13, '3': 17, '1': 22, '8': 17, '2': 13, '6': 14}
run3_bs8: per-rank wall seconds/update={'0': 15.347052206285298, '1': 15.347155910357833, '2': 15.347095459420235, '3': 15.347046211827546}; frames={'9': 1152, '7': 18, '4': 14, '5': 13, '3': 17, '1': 22, '8': 17, '2': 13, '6': 14}
run4_bs4: per-rank wall seconds/update={'0': 14.715134540852159, '1': 14.715142120141536, '2': 14.715132261626422, '3': 14.715165278408676}; frames={'9': 1152, '1': 22, '3': 17, '7': 18, '5': 13, '4': 14, '2': 13, '8': 17, '6': 14}

Stream timing includes stalls and rank waits; it is not isolated kernel compute time.
Compare phase deltas across repetitions before naming the cause; fallback/padding counts alone do not prove a time bottleneck.