diffusers
[feat] Tensor parallelism for MiniMax-H3
#14609
Merged

[feat] Tensor parallelism for MiniMax-H3 #14609

JingyaHuang
github-actions github-actions added size/L
github-actions github-actions removed size/L
github-actions github-actions added documentation
github-actions github-actions added lora
github-actions github-actions added models
github-actions github-actions added tests
github-actions github-actions added modular-pipelines
github-actions github-actions added utils
github-actions github-actions added pipelines
github-actions github-actions added hooks
HuggingFaceDocBuilderDev
github-actions github-actions added size/L
whn09
whn09
github-actions github-actions added quantization
github-actions github-actions added CI
github-actions github-actions added single-file
whn09
JingyaHuang JingyaHuang force pushed from 7e4a7eff to e4ae9de8 32 days ago
github-actions github-actions removed documentation
github-actions github-actions removed lora
github-actions github-actions removed quantization
github-actions github-actions removed utils
github-actions github-actions removed pipelines
github-actions github-actions removed CI
github-actions github-actions removed single-file
JingyaHuang JingyaHuang marked this pull request as ready for review 32 days ago
JingyaHuang Add tensor-parallel support for MiniMax-H3
764e7c8a
JingyaHuang Shard MiniMax-H3's adaln_proj to fit tensor parallelism on one device
58d9b513
JingyaHuang Build MiniMax-H3's row timestep plan on CPU, as its caller expects
c766b888
JingyaHuang feat:combine tp+cp
41627842
whn09 Allow tensor parallelism and context parallelism in one ParallelConfig
7f649f94
JingyaHuang feat: validate the support on TPU
673a3f10
JingyaHuang fix: put tp first in the combined TP+CP mesh
3db653d8
JingyaHuang fix: style
9c0f7aaf
JingyaHuang JingyaHuang force pushed from 3e2576fb to 9c0f7aaf 31 days ago
JingyaHuang fix: apply context parallelism on the sharded-load path
bbc55a76
JingyaHuang revert: attention head padding
6ce34c77
JingyaHuang revert: drop the Neuron all-gather padding and restore the TPU pre-sh…
986ea9a4
JingyaHuang refactor: share the CPU pre-shard path between Neuron and TPU tensor …
6845f333
whn09
JingyaHuang Merge branch 'main' into add-h3-tp-support
e8a38d37
JingyaHuang fix: new device count API for TPU since deprecation
2590829f
JingyaHuang fix: pad uneven all-gathers on Neuron and TPU in _all_gather_anything
ce4198ac
JingyaHuang JingyaHuang requested a review from sayakpaul sayakpaul 27 days ago
JingyaHuang JingyaHuang requested a review from DN6 DN6 27 days ago
JingyaHuang
JingyaHuang Merge branch 'main' into add-h3-tp-support
d2c362ed
sayakpaul
sayakpaul commented on 2026-10-14
JingyaHuang Merge branch 'main' into add-h3-tp-support
3cd26a95
JingyaHuang Merge branch 'main' into add-h3-tp-support
6b3b3e89
JingyaHuang fix: fall back to one device when no accelerator is present
cd611662
JingyaHuang Move TP+CP composition to #14725
cc851100
JingyaHuang Merge branch 'main' into add-h3-tp-support
b1076e6f
JingyaHuang review: use `config._device` for the Neuron TP device
8526d942
JingyaHuang review: replace ReplicatedInputRowwiseParallel with a "rowwise_split_…
ed452ef8
github-actions github-actions added documentation
JingyaHuang review: explain when to use "rowwise_split_input"
bc35ec2d
JingyaHuang Merge branch 'main' into add-h3-tp-support
3ea9cc59
JingyaHuang feat:shard tensor-parallel weights locally on every device
660bb6da
JingyaHuang Merge branch 'main' into add-h3-tp-support
82e45d41
JingyaHuang JingyaHuang requested a review from sayakpaul sayakpaul 13 days ago
sayakpaul
sayakpaul commented on 2026-10-27
sayakpaul Merge branch 'main' into add-h3-tp-support
dd54cfb8
sayakpaul sayakpaul added this to the Release 0.43.0 milestone 13 days ago
JingyaHuang doc: improve the explanation of rowwise_split_input
9ec46667
JingyaHuang review: remove useless zero division prevention
d13196bb
JingyaHuang review: simplify the docstring
ff24da56
JingyaHuang revert: Restore standard TP sharding and isolate the Neuron workaround
63732ec3
JingyaHuang revert: no refactoring of in-memory sharding
8e2fa1d5
JingyaHuang Merge branch 'main' into add-h3-tp-support
669f503d
JingyaHuang Merge branch 'main' into add-h3-tp-support
4d136666
JingyaHuang revert: keep the in-memory TP path as on main
b680a255
JingyaHuang style: pika
39bcfef6
JingyaHuang Merge branch 'add-h3-tp-support' of github.com:JingyaHuang/diffusers …
8b78dfe0
github-actions github-actions removed size/L
github-actions github-actions added size/M
JingyaHuang style: make fix-copies
8c74bb17
JingyaHuang JingyaHuang requested a review from sayakpaul sayakpaul 5 days ago
sayakpaul Merge branch 'main' into add-h3-tp-support
3dc01948
sayakpaul
sayakpaul commented on 2026-11-05
sayakpaul Merge branch 'main' into add-h3-tp-support
510487e1
sayakpaul
sayakpaul approved these changes on 2026-11-05
sayakpaul sayakpaul merged 7e46c94c into main 4 days ago

Login to write a write a comment.

Login via GitHub

Reviewers
Assignees
No one assigned
Labels
Milestone