Prefetch mmap'd weight blobs to eliminate page fault bottleneck#18356
Merged
manuelcandales merged 1 commit intoMar 20, 2026
Conversation
Weight loading via update_constants_from_blob was achieving only 0.3-0.4 GB/s (vs 8 GB/s hardware capability) because memcpy from mmap'd pages triggers synchronous page faults — each 16K page traps into the kernel for NVMe I/O. Call madvise(MADV_WILLNEED) on the weights blob early in Metal backend init, before writing/dlopen'ing the .so file. The kernel prefaults pages asynchronously during the ~200ms of other init work. By the time memcpy runs, pages are already resident and throughput reaches 5-8 GB/s. Metal init time: ~25s -> ~9s (2.7x faster) on int4 Voxtral model. (cherry picked from commit b7ca1a4)
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/18356
Note: Links to docs will display an error until the docs builds have been completed. ⏳ No Failures, 122 PendingAs of commit e3f3328 with merge base 8c0a60b ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
manuelcandales
approved these changes
Mar 20, 2026
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
3 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Weight loading via update_constants_from_blob was achieving only
0.3-0.4 GB/s (vs 8 GB/s hardware capability) because memcpy from
mmap'd pages triggers synchronous page faults — each 16K page traps
into the kernel for NVMe I/O.
Call madvise(MADV_WILLNEED) on the weights blob
early in Metal backend init, before writing/dlopen'ing the .so file.
The kernel prefaults pages asynchronously during the ~200ms of other
init work. By the time memcpy runs, pages are already resident and
throughput reaches 5-8 GB/s.
Metal init time: ~25s -> ~9s (2.7x faster) on int4 Voxtral model.