transformers
Warm the page cache before sharded loading (opt-in via `HF_SHARD_PREFETCH`)
#48227

Open

Warm the page cache before sharded loading (opt-in via `HF_SHARD_PREFETCH`) #48227

qgallouedec wants to merge 6 commits into main from load-shard-prefetch

qgallouedec requested a review from

Cyrilvallez 72 days ago

qgallouedec requested a review from

3outeille 72 days ago

ArthurZucker commented on 2026-09-01

qgallouedec force pushed from 9e8a65d2 to f70364f3 59 days ago

qgallouedec force pushed from f70364f3 to 31a256b5 58 days ago

qgallouedec force pushed from 31a256b5 to a98cde20 58 days ago

qgallouedec force pushed from a98cde20 to 68bc14bf 58 days ago

qgallouedec force pushed from 68bc14bf to a1846bc2 58 days ago

qgallouedec force pushed from a1846bc2 to 6b19d157 57 days ago

qgallouedec force pushed from 6b19d157 to cfdc8bab 56 days ago

ArthurZucker commented on 2026-09-14

qgallouedec force pushed from dcdee520 to 2c302ab3 36 days ago

Warm the page cache before sharded loading (opt-in via HF_SHARD_PREFE…

a07648d4

style

05ca7a73

Move the shard prefetch into distributed/utils.py

f04f7a08

Apply ruff format to the prefetch helper

d17e1335

Bind the prefetch barrier to the accelerator

cd6ad0ec

Prefetch the shards in the background instead of waiting for them

d87e79cc

qgallouedec force pushed from 2c302ab3 to d87e79cc 4 days ago

Reviewers

ArthurZucker

Cyrilvallez

3outeille

Assignees

No one assigned

Labels

None yet

Milestone

No milestone