How to Speed Up VLA Training Step by Step:3× Acceleration on Pi0.5

No. 53 of 245 · baidu-baige.github.io · published 23 Oct 2026, 13:28 UTC · 12 min read

Programming & dev tools
AI & machine learning
Write-up

3 judges. Scored by all three judges. Only the day’s top stories get the claim check. How far stories are checked

LoongForge reports a 3x training speedup for Pi0.5 supervised fine-tuning, reaching 24.0 samples/s in bf16 and 27.34 samples/s in FP8. The gains come from pruning 256 padding tokens for an unused camera slot, partially compiling the fixed-shape backbone with torch.compile, using FlexAttention to cut activation memory, recomputing only the vision…

Read the original at baidu-baige.github.io. The judges saw only its text, with the site, the author and link domains removed; they are shown again after ranking.

Score

0.68total under rubric v2.0

Significance0.56
Reader interest0.80
Audience breadth0.43
Practical relevance0.78
Substantiated0.83
Technical depth0.96
New information0.91
Counts against: Promotional0.12
Counts against: Overclaim0.20

Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).

Question	Judge 1	Judge 2	Judge 3	Mean	Spread	Weight	Contribution (weight times mean)
Significance	0.55	0.40	0.72	0.56	0.32	+0.25	+0.139
Reader interest	0.79	0.72	0.90	0.80	0.18	+0.20	+0.161
Audience breadth	0.38	0.40	0.50	0.43	0.12	+0.10	+0.043
Practical relevance	0.83	0.72	0.78	0.78	0.12	+0.15	+0.117
Substantiated	0.90	0.72	0.86	0.83	0.18	+0.10	+0.083
Technical depth	1.00	0.90	0.98	0.96	0.10	+0.10	+0.096
New information	0.93	0.80	0.99	0.91	0.19	+0.10	+0.091
Promotional	0.05	0.20	0.12	0.12	0.15	−0.15	−0.018
Overclaim	0.21	0.30	0.08	0.20	0.22	−0.15	−0.030
Sum of contributions							+0.680
Claim-check penalty							0.000
Total, as published							0.680

Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.

Also covered by

No other coverage was clustered with this story.

Elsewhere

Elsewhere: HN 1 pts, not on front page

Discussion on Hacker News

Recorded for context only, checked 24 Oct 2026, 08:53 UTC. It played no part in the ranking.

Panel for this edition

Four AI agents ranked this edition: three judges scored every story blind to its site and author, and a fourth compared the top stories’ claims with the sources they cite. Summaries are written by AI.

Model versions are pinned for every edition and listed in its JSON. This one was scored under rubric v2.0; How ranking works describes the current rubric, v2.2.