Repository navigation
Conversation
Release Qwen4 load-only shards after placement and model metadata reads, retaining CPU embedding and PLE lookup files, companions and inode aliases. Queue original shard descriptors through every TP load-ready exchange, including side heads and vision. Suppress intermediate hash/pack range advice while loading, and close held descriptors without advice on failure.
…t's pages after every rank is ready
…ayer ahead while it packs, its pages released after
…ad, so TF_LOAD_PREFETCH=0 is main's path
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Part of #2003. Stacked on #1928 (@jschmied) and #1887: the first two commits are #1928's own, and the series sits on #1887's TP page-drop deferral. It lands after both, and #1928's commits drop out then.
#1928's read-ahead (
TF_LOAD_PREFETCH: reader threads read a layer's checkpoint intervals while its loader packs the weights) is ported to the other CUDA loaders. On these families it's opt-in (unset is 0) until each one's GB10 load receipts are in. GLM keeps its existing default, and #1928's own two families are unchanged.TF_LOAD_PREFETCH=0is each family's main path byte for byte: no thread, no read, no change to release order.Reviewed independently three times. The host suite, the Linux CUDA build and lean (424) pass. GB10 load-time cells per family come before any default changes.