[prototype] N-D distributed parallel support #2947

winglian · 2025-07-19T02:35:20Z

Description

based on @SalmanMohammadi's upstream accelerate PR (https://github.com/huggingface/accelerate/pull/3682/files), I was able to hack together an implementation that needs no upstream changes so we can experiment on what the actual needed changes should be.

One caveat is that accelerate's DataLoaderDispatcher does not support process & iterate on rank0 and dispatch to other ranks. e.g. accelerator_config.dispatch_batches: true. We should also test that this works for various combinations of w/o dispatching as well as split batches, etc.

Combinations to test:

TP support doesn't seem to work properly in transformers w/o Deepspeed AutoTP

related PRs:

coderabbitai · 2025-07-19T02:35:26Z

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

🪧 Tips

Chat

There are 3 ways to chat with CodeRabbit:

Review comments: Directly reply to a review comment made by CodeRabbit. Example:
- I pushed a fix in commit <commit_id>, please review it.
- Explain this complex logic.
- Open a follow-up GitHub issue for this discussion.
Files and specific lines of code (under the "Files changed" tab): Tag @coderabbitai in a new review comment at the desired location with your query. Examples:
- @coderabbitai explain this code block.
- @coderabbitai modularize this function.
PR comments: Tag @coderabbitai in a new PR comment to ask questions about the PR branch. For the best results, please provide a very specific query, as very limited context is provided in this mode. Examples:
- @coderabbitai gather interesting stats about this repository and render them as a table. Additionally, render a pie chart showing the language distribution in the codebase.
- @coderabbitai read src/utils.ts and explain its main purpose.
- @coderabbitai read the files in the src/scheduler package and generate a class diagram using mermaid and a README in the markdown format.
- @coderabbitai help me debug CodeRabbit configuration file.

Support

Need help? Create a ticket on our support page for assistance with any issues or questions.

Note: Be mindful of the bot's finite context window. It's strongly recommended to break down tasks such as reading entire modules into smaller chunks. For a focused discussion, use review comments to chat about specific files and their changes, instead of using the PR comments.

CodeRabbit Commands (Invoked using PR comments)

@coderabbitai pause to pause the reviews on a PR.
@coderabbitai resume to resume the paused reviews.
@coderabbitai review to trigger an incremental review. This is useful when automatic reviews are disabled for the repository.
@coderabbitai full review to do a full review from scratch and review all the files again.
@coderabbitai summary to regenerate the summary of the PR.
@coderabbitai generate docstrings to generate docstrings for this PR.
@coderabbitai generate sequence diagram to generate a sequence diagram of the changes in this PR.
@coderabbitai resolve resolve all the CodeRabbit review comments.
@coderabbitai configuration to show the current CodeRabbit configuration for the repository.
@coderabbitai help to get help.

Other keywords and placeholders

Add @coderabbitai ignore anywhere in the PR description to prevent this PR from being reviewed.
Add @coderabbitai summary to generate the high-level summary at a specific location in the PR description.
Add @coderabbitai anywhere in the PR title to generate the title automatically.

Documentation and Community

Visit our Documentation for detailed information on how to use CodeRabbit.
Join our Discord Community to get help, request features, and share feedback.
Follow us on X/Twitter for updates and announcements.

github-actions · 2025-07-19T02:40:56Z

📖 Documentation Preview: https://6881221bb6eb5599a6a7bb60--resonant-treacle-0fd729.netlify.app

Deployed on Netlify from commit ad70741

codecov · 2025-07-19T04:07:29Z

Codecov Report

❌ Patch coverage is 32.61649% with 188 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/axolotl/core/parallel.py	20.00%	60 Missing ⚠️
src/axolotl/monkeypatch/ring_attn/patch.py	8.82%	31 Missing ⚠️
src/axolotl/monkeypatch/dist_parallel/patch.py	0.00%	22 Missing ⚠️
src/axolotl/utils/schemas/validation.py	45.94%	20 Missing ⚠️
src/axolotl/core/trainers/mixins/dist_parallel.py	50.00%	9 Missing ⚠️
src/axolotl/monkeypatch/accelerate/fsdp2.py	0.00%	9 Missing ⚠️
src/axolotl/core/trainers/grpo/trainer.py	0.00%	8 Missing ⚠️
src/axolotl/loaders/model.py	22.22%	7 Missing ⚠️
src/axolotl/core/builders/base.py	50.00%	4 Missing ⚠️
src/axolotl/core/trainers/grpo/sampler.py	0.00%	3 Missing ⚠️
... and 8 more

📢 Thoughts on this report? Let us know!

src/axolotl/loaders/tokenizer.py

djsaunde

Just leaving a few comments after a quick glance.

Question: can we add support for the HYBRID_SHARD fsdp2 strategy as part of this PR (i.e., replicating across nodes while sharding intra node)? Or maybe it doesn't quite fit here?

djsaunde · 2025-07-19T15:57:26Z

src/axolotl/core/builders/base.py

+        if (
+            self.cfg.tensor_parallel_size > 1
+            or (self.cfg.dp_shard_size and self.cfg.dp_shard_size > 1)
+            or self.cfg.sequence_parallel_degree > 1


aside: we should rename to context_parallel_size to better match the ecosystem and deprecate sequence_parallel_size.

migrated sequence_parallel_degree -> context_parallel_size

djsaunde · 2025-07-19T16:02:05Z

src/axolotl/loaders/model.py

+
+        if self.cfg.tensor_parallel_size and self.cfg.tensor_parallel_size > 1:
+            # from pretrained only needs to know about Tensor Parallelism for sharding weights
+            dist_parallel = DistParallel.build(


I notice we build this three separate times. Is it possible to just have a singleton instance that we build once and refer to thereafter?

I thought about doing a singleton, but then didn't want to deal with the complexity of singletons in CI. Happy to give it a whirl though if we're confident that this shouldn't change ever during runtime

Alternatively, could we have a utility function that accepts a cfg and return the device mesh / distparallel?

Do you actually need this on your side? It feels to me like you can reuse ParallelismConfig from upstream for this and not repeat code that much. If you need extra functionality, open to suggestions.

I think we can swap this out once upstream changes land.

djsaunde · 2025-07-19T16:03:52Z

src/axolotl/monkeypatch/dist_parallel/patch.py

+            try:
+                submesh_dp_fsdp_size = torch_device_mesh["dp_fsdp"].size()
+            except KeyError:
+                pass


Maybe use if "dp_fsdp" in torch_device_mesh here instead

Actually, is this even set anywhere?

thanks, I think I need to go back and fix this whole method.

winglian · 2025-07-19T17:51:59Z

`HYBRID_SHARD` is accomplished using the 2d-mesh now

NanoCode012 · 2025-07-21T03:49:21Z

src/axolotl/utils/schemas/validation.py

+    @model_validator(mode="before")
+    @classmethod
+    def check_dp_shard_size_fsdp(cls, data):
+        if data.get("fsdp_config", {}):
+            dp_shard_size = data.get("dp_shard_size", None)
+            if dp_shard_size and dp_shard_size <= 1:
+                raise ValueError("dp_shard_size must be greater than 1 when using FSDP")
+        if not data.get("fsdp_config", {}) and data.get("dp_shard_size", 1) > 1:
+            raise ValueError("FSDP should be enabled when using dp_shard_size > 1")
+        return data
+


Suggested change

@model_validator(mode="before")

@classmethod

def check_dp_shard_size_fsdp(cls, data):

if data.get("fsdp_config", {}):

dp_shard_size = data.get("dp_shard_size", None)

if dp_shard_size and dp_shard_size <= 1:

raise ValueError("dp_shard_size must be greater than 1 when using FSDP")

if not data.get("fsdp_config", {}) and data.get("dp_shard_size", 1) > 1:

raise ValueError("FSDP should be enabled when using dp_shard_size > 1")

return data

@model_validator(mode="before")

@classmethod

def check_dp_shard_size_fsdp(cls, data):

fsdp_config = data.get("fsdp_config", {})

dp_shard_size = data.get("dp_shard_size", None)

has_fsdp = bool(fsdp_config)

dp_sharding_enabled = dp_shard_size and dp_shard_size > 1

if has_fsdp and dp_shard_size and dp_shard_size <= 1:

raise ValueError("dp_shard_size must be greater than 1 when using FSDP")

if not has_fsdp and dp_sharding_enabled:

raise ValueError("FSDP should be enabled when using dp_shard_size > 1")

return data

Small refactor

src/axolotl/utils/schemas/validation.py

NanoCode012 · 2025-07-21T05:25:49Z

src/axolotl/integrations/liger/args.py

+        if data.get("liger_rms_norm") and data.get("tensor_parallel_size", 1) <= 1:
+            raise ValueError(
+                "`liger_rms_norm` is incompatible with tensor parallelism, "
+                "see https://github.com/linkedin/Liger-Kernel/issues/826"
+            )


Suggested change

if data.get("liger_rms_norm") and data.get("tensor_parallel_size", 1) <= 1:

raise ValueError(

"`liger_rms_norm` is incompatible with tensor parallelism, "

"see https://github.com/linkedin/Liger-Kernel/issues/826"

)

if data.get("liger_rms_norm") and data.get("tensor_parallel_size", 1) > 1:

raise ValueError(

"`liger_rms_norm` is incompatible with tensor parallelism, "

"see https://github.com/linkedin/Liger-Kernel/issues/826"

)

Do you mean this condition?

NanoCode012 · 2025-07-21T05:26:43Z

src/axolotl/loaders/model.py

+
+        if self.cfg.tensor_parallel_size and self.cfg.tensor_parallel_size > 1:
+            # from pretrained only needs to know about Tensor Parallelism for sharding weights
+            dist_parallel = DistParallel.build(


Alternatively, could we have a utility function that accepts a cfg and return the device mesh / distparallel?

NanoCode012 · 2025-07-21T05:27:57Z

src/axolotl/monkeypatch/accelerate/axolotl.py

+        if (
+            not (
+                self.distributed_type
+                == accelerate.accelerator.DistributedType.DEEPSPEED
+                and hasattr(self.state, "ds_device_mesh")
+            )
+            and self.world_device_mesh is not None
+        ):
+            return self.world_device_mesh


Would it be better to check, if distributed_type == fsdp, to return world_device_mesh?

SalmanMohammadi · 2025-07-21T08:58:31Z

src/axolotl/loaders/model.py

+            device_mesh = dist_parallel.get_device_mesh()
+            self.model_kwargs["tp_size"] = self.cfg.tensor_parallel_size
+            self.model_kwargs["tp_plan"] = "auto"
+            self.model_kwargs["device_mesh"] = device_mesh[("tp",)]


winglian requested review from djsaunde and SalmanMohammadi July 19, 2025 02:35

NanoCode012 reviewed Jul 19, 2025

View reviewed changes

src/axolotl/loaders/tokenizer.py Outdated Show resolved Hide resolved

djsaunde reviewed Jul 19, 2025

View reviewed changes

winglian force-pushed the nd-dist-parallel branch 2 times, most recently from 71d1133 to 2d024be Compare July 21, 2025 00:50

NanoCode012 reviewed Jul 21, 2025

View reviewed changes

SalmanMohammadi reviewed Jul 21, 2025

View reviewed changes

winglian added 18 commits July 23, 2025 13:49

ND parallelism for dp replicas

85397f4

more nd-parallel wip

ffe3bd1

remove module imporT

aa77063

Fix indent

7534d2b

use default device when using device mesh

ad668b2

handle dist parallel training args

3292c41

debugging and shard model only over tp dim on model load

7e436ab

more fixes

21a2a27

improve handling of empty dp_shard_size

92deaf1

validation around shard size

7db2862

more fixes for TP

a12564e

tensor parallel doesn't work with liger_rms_norm

b62399e

fix ring attn to use device mesh

d0242b2

pass device mesh to sp context manager

83ca223

fix sequence_parallel_degree input

850917a

chore: lint

4859b19

rename sequence_parallel_degree to sequence_parallel_size

5b1a4e6

rename sequence_parallel_size to context_parallel_size

80f399c

winglian added 2 commits July 23, 2025 13:49

handle dp_shard_size unset

b1ea824

remove mesh name

ad70741

winglian force-pushed the nd-dist-parallel branch from 9fabce8 to ad70741 Compare July 23, 2025 17:50

winglian changed the title ~~N-D distributed parallel support~~ [prototype] N-D distributed parallel support Jul 24, 2025

winglian mentioned this pull request Jul 24, 2025

Distributed/ND-Parallel #2977

Merged

winglian closed this Jul 31, 2025

winglian deleted the nd-dist-parallel branch August 2, 2025 14:51

Uh oh!

[prototype] N-D distributed parallel support #2947

[prototype] N-D distributed parallel support #2947

Uh oh!

Conversation

winglian commented Jul 19, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Uh oh!

coderabbitai bot commented Jul 19, 2025

Review skipped

Chat

Support

CodeRabbit Commands (Invoked using PR comments)

Other keywords and placeholders

Documentation and Community

Uh oh!

github-actions bot commented Jul 19, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

codecov bot commented Jul 19, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

Uh oh!

djsaunde left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

winglian Jul 19, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

winglian commented Jul 19, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

winglian commented Jul 19, 2025 •

edited

Loading

github-actions bot commented Jul 19, 2025 •

edited

Loading

codecov bot commented Jul 19, 2025 •

edited

Loading

winglian Jul 19, 2025 •

edited

Loading