Draft commit #45453
ArthurZucker
wants to merge
22 commits into
main
from
refactor-tp-loading
init
5c4efd53
up
1ee7fff9
tp loading: CPU staging, per-rank scheduling, slice-first FP8, no_grad
b5dcdf21
tp loading: fix view-keeps-buffer-alive, skip warmup under TP
1fa1c78b
tp loading: use dist.scatter for sharded redistribute
aea39f09
tp loading: batched pre-slice via torch.chunk instead of per-rank loop
fa073e84
tp loading: async pipelined scatter + eliminate per-mapping metadata …
ffac18ae
tp loading: materialize to GPU directly, drop redundant CPU→GPU copy
6a4dea69
tp loading: 3-stage pipeline — async convert + async scatter + finalize
5301a9af
tp loading: batch coalesce scatter/broadcast via _coalescing_manager
84ee0694
tp loading: packed scatter per source rank; drop coalescing_manager
b9d49ae9
up
4e66348e
up
8dc4d753
tp loading: per-batch pinned-memory reads + pipeline overlap
5d9c61e3
tp loading: clean NCCL packed-scatter, remove CUDA IPC experiment
498a7269
tp loading: single-batch packed scatter, document NCCL transfer floor
ffe7a339
tp loading: pipelined batch_isend_irecv with overlap
63d7486e
tp loading: remove dead code, shared thread pool, allocator warmup
e186123d
tp loading: switch to distribute_tensor, drop 200 lines of packing
be9278eb
tp loading: async scatter + broadcast via mesh_scatter(async_op=True)
cdb4a7d9
tp loading: minimal sync distribute_tensor, drop async machinery
c60156f8
update
21ae32d4
ArthurZucker
commented
on 2026-04-20
Reviewers
No reviews
Assignees
No one assigned
Labels
None yet
Milestone
No milestone
Login to write a write a comment.
Login via GitHub