transformers
Prefetch only the byte spans each rank will read
#48604
Open

Prefetch only the byte spans each rank will read #48604

qgallouedec wants to merge 49 commits into load-shard-prefetch from prefetch-rank-spans
qgallouedec
qgallouedec Warm the page cache before sharded loading (opt-in via HF_SHARD_PREFE…
688b4262
qgallouedec style
c9715177
qgallouedec Merge branch 'main' into load-shard-prefetch
b483d0c8
qgallouedec Merge branch 'main' into load-shard-prefetch
2fcdf17a
qgallouedec Merge branch 'main' into load-shard-prefetch
d706b11e
qgallouedec Merge branch 'main' into load-shard-prefetch
002a2758
qgallouedec Move the shard prefetch into distributed/utils.py
686607e0
qgallouedec Apply ruff format to the prefetch helper
69bb33be
qgallouedec Merge branch 'main' into load-shard-prefetch
f783ff39
qgallouedec Merge branch 'main' into load-shard-prefetch
1bd524ae
qgallouedec Merge branch 'main' into load-shard-prefetch
14f0a81c
qgallouedec Merge branch 'main' into load-shard-prefetch
07eb49e4
qgallouedec Merge branch 'main' into load-shard-prefetch
9059c09e
qgallouedec Prefetch only the byte spans each rank will read
b01da96b
HuggingFaceDocBuilderDev
qgallouedec Cover per-expert checkpoints, and compare against a fair share of the…
69e08c23
qgallouedec Warm short ranges with kernel readahead, and bridge cheap gaps
862086dc
qgallouedec Take the meta state dict again, not named_parameters
cca1d864
VI-Arthur
VI-Arthur commented on 2026-09-08
VI-Arthur
VI-Arthur
qgallouedec Decide readahead on the total a rank warms, not the size of each range
f2c47436
qgallouedec Test the byte spans a rank warms
e21db3dc
qgallouedec Make local ranks pick the same prefetch plan
20896a7b
qgallouedec Reduce over the accelerator the file already binds to
b191be0e
qgallouedec Merge branch 'main' into load-shard-prefetch
19f7ae1d
qgallouedec Merge branch 'main' into load-shard-prefetch
9f432f54
qgallouedec Bind the prefetch barrier to the accelerator
9e8a65d2
qgallouedec Merge branch 'load-shard-prefetch' into prefetch-rank-spans
82a4f5ba
qgallouedec Warm the page cache before sharded loading (opt-in via HF_SHARD_PREFE…
153a3994
qgallouedec style
cf674aa3
qgallouedec Move the shard prefetch into distributed/utils.py
77ad02d0
qgallouedec Apply ruff format to the prefetch helper
70c6ac4b
qgallouedec Bind the prefetch barrier to the accelerator
f70364f3
qgallouedec qgallouedec force-pushed the load-shard-prefetch branch from 9e8a65d2 to f70364f3 53 days ago
qgallouedec Merge remote-tracking branch 'upstream/load-shard-prefetch' into pref…
a687e4da
qgallouedec qgallouedec force-pushed the load-shard-prefetch branch from f70364f3 to 31a256b5 52 days ago
qgallouedec Warm the page cache before sharded loading (opt-in via HF_SHARD_PREFE…
6e40b0b1
qgallouedec style
9a61e7b7
qgallouedec Move the shard prefetch into distributed/utils.py
e8b13c27
qgallouedec Apply ruff format to the prefetch helper
eed24d1e
qgallouedec Bind the prefetch barrier to the accelerator
a98cde20
qgallouedec qgallouedec force-pushed the load-shard-prefetch branch from 31a256b5 to a98cde20 52 days ago
qgallouedec Merge remote-tracking branch 'upstream/load-shard-prefetch' into pref…
62e85eab
qgallouedec Warm the page cache before sharded loading (opt-in via HF_SHARD_PREFE…
72548d08
qgallouedec style
dd8804bd
qgallouedec Move the shard prefetch into distributed/utils.py
acae96cf
qgallouedec Apply ruff format to the prefetch helper
7db2467b
qgallouedec Bind the prefetch barrier to the accelerator
68bc14bf
qgallouedec qgallouedec force-pushed the load-shard-prefetch branch from a98cde20 to 68bc14bf 52 days ago
qgallouedec Merge remote-tracking branch 'upstream/load-shard-prefetch' into pref…
e514c7fd
qgallouedec Warm the page cache before sharded loading (opt-in via HF_SHARD_PREFE…
5489cde1
qgallouedec style
0b5d4402
qgallouedec Move the shard prefetch into distributed/utils.py
170161c5
qgallouedec Apply ruff format to the prefetch helper
2f54f17c
qgallouedec Bind the prefetch barrier to the accelerator
a1846bc2
qgallouedec qgallouedec force-pushed the load-shard-prefetch branch from 68bc14bf to a1846bc2 52 days ago
qgallouedec Merge remote-tracking branch 'upstream/load-shard-prefetch' into pref…
37c8ee6f
qgallouedec qgallouedec force-pushed the load-shard-prefetch branch from a1846bc2 to 6b19d157 51 days ago
github-actions
qgallouedec qgallouedec force-pushed the load-shard-prefetch branch from 6b19d157 to cfdc8bab 50 days ago
qgallouedec qgallouedec force-pushed the load-shard-prefetch branch from dcdee520 to 2c302ab3 30 days ago

Login to write a write a comment.

Login via GitHub

Reviewers
Assignees
No one assigned
Labels
Milestone