Lemmy: Bestiverse
  • Communities
  • Create Post
  • Create Community
  • heart
    Support Lemmy
  • search
    Search
  • Login
  • Sign Up
RSS BotBM to Hacker NewsEnglish · 1 hour ago

Flattening Every Memory Peak in Long-Context Mixture-of-Experts Training

arxiv.org

external-link
message-square
0
link
fedilink
2
external-link

Flattening Every Memory Peak in Long-Context Mixture-of-Experts Training

arxiv.org

RSS BotBM to Hacker NewsEnglish · 1 hour ago
message-square
0
link
fedilink
Training a Mixture-of-Experts (MoE) model at long context or large batch size fails as soon as any one component's peak allocation exceeds device memory, so the target is every peak at once, not the average footprint. Four are left unbounded by the parallelism plans in common use, and each grows differently: expert dispatch with the routing matrix, the vocabulary projection with tokens times vocabulary, gradient checkpoint boundaries with depth times sequence length, and optimizer state with parameter count. Which one runs out first changes with the model, the context length, and the device count, so lowering the largest only exposes the next. We bound all four with schedules whose GPU working set is fixed at launch: PipelinedLLEP extends least-loaded expert parallelism with a cap on the tokens each source contributes to a dispatch chunk, Ring-DTP circulates activations or weight shards around a ring at the vocabulary projection and folds each block of logits into an online log-sum-exp, Selective checkpoint offload (SCO) keeps the one long-lived tensor of each checkpoint boundary in CPU memory, and OffloadStreamAdamW turns the serial CPU Adam update of optimizer offload into a bucket pipeline. All four change only the order and granularity of computation and data movement, so the loss and gradients stay exact. In matched component tests, they cut the MoE dispatch peak by up to $59.3\%$ without losing throughput, the vocabulary projection peak by $86.6\%$, and the offloaded optimizer step by $2.05\times$ faster. Composed on MoE models from 120B to 667B parameters, they train at 1M context length, $8$--$32\times$ the reach of a tuned FSDP2 baseline, and up to $10.4\times$ its throughput.

Comments

alert-triangle
You must log in or register to comment.

Hacker News

hackernews

Subscribe from Remote Instance

You are not logged in. However you can subscribe from another Fediverse account, for example Lemmy or Mastodon. To do this, paste the following into the search field of your instance: !hackernews@lemmy.bestiver.se
lock
Community locked: only moderators can create posts. You can still comment on posts.

NOTICE: The bot occasionally has downtime, sorry about that. I’m working on adding better monitoring and error recovery so I can get notified faster.

Posts from the RSS Feed of HackerNews.

The feed sometimes contains ads and posts that have been removed by the mod team at HN.

Source of the RSS Bot.

Visibility: Public
globe

This community can be federated to other instances and be posted/commented in by their users.

  • 542 users / day
  • 2.1K users / week
  • 4.14K users / month
  • 8.83K users / 6 months
  • 2 local subscribers
  • 5.8K subscribers
  • 65.5K Posts
  • 36.1K Comments
  • Modlog
  • mods:
  • patrick
  • RSS BotB
  • BE: 0.19.21
  • Modlog
  • Instances
  • Docs
  • Code
  • join-lemmy.org