Page-size check prevents mmap for blob-aligned checkpoints
- Dominant language
- C++
- Stars
- 7k
- Forks
- 660
- Avg merge
- 20h 43m
- Merged PRs (30d)
- 33
Description
## Problem
`ChooseMode` disables mmap when the checkpoint size is not divisible by the system page size, even if the file is padded to the blob format's 256-byte alignment.
For example, the 4B checkpoint is 5,401,208,832 bytes: divisible by 256, but with a 2,048-byte remainder on a 4,096-byte page system. It therefore falls back to copying the weights into allocated memory instead of using file-backed mappings.
## Proposed fix
Relax the file-length alignment requirement from the base page size to `kBlobAlign` (256 bytes).
The OS accepts partial-page mappings, but inference kernels can read multiple vectors ahead. Retain the blob-alignment guard and fall back to ordinary reads for lengths that do not meet it. Arbitrary-length mmap support would require separate handling of those reads.
Covered by #1032.
Contributor guide
Research direction
Start at ChooseMode and inspect the checkpoint length check that compares against the system page size, then trace the mmap and ordinary-read fallback paths. Done means blob-aligned lengths such as the 4B checkpoint can use file-backed mappings, while lengths that are not divisible by kBlobAlign still use ordinary reads.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- cpp
- Domain
- machine-learning, performance
- Issue type
- Bug
- Difficulty
- 2/5
- Estimated time
- 1-3 hours
- Activity status
- Active
- Clarity
- Clearly specified
- Newbie friendliness
- 75/100