Repository navigation
Conversation
3outeille
changed the title
[distributed] Count EP-sharded experts in the allocator warmup, and keep the model's plans resolved
3outeille
changed the title
[distributed] Count EP-sharded experts in the allocator warmup, and keep the model's plans resolved
3outeille
force-pushed
the
fix-warmup-byte-count-ep-plan
branch
from
October 7, 2026 10:25
9c05508 to
80d83d1
Compare
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
…eep the model's plans resolved `get_total_byte_count` divided a parameter's size only when it matched `model.tp_plan`, always by the world size. Since the TP and EP plans were decoupled (#48859), EP-sharded experts no longer show up there unless the TP plan happens to list them too, so an EP-only layout (`tp_plan={}` or a model without a `base_model_tp_plan`, e.g. DeepSeek V4) counted every expert at full size on every rank and `caching_allocator_warmup` tried to reserve the whole model per GPU. `resolve_parallel_plans` now writes the resolved plans back to `model.tp_plan` and `model.ep_plan`: after loading they hold the rules that were applied, each plan is empty when its parallel size is 1, and no parameter is named by both. The warmup then divides a parameter by `tp_size` when the TP plan names it and by `ep_size` when the EP plan does, rather than by the world size, which also makes the estimate right on 2-D meshes and under token dispatch with `tp_size=1`.
3outeille
force-pushed
the
fix-warmup-byte-count-ep-plan
branch
from
October 7, 2026 10:26
80d83d1 to
87f176f
Compare
vasqu
reviewed
Oct 7, 2026
vasqu
left a comment
vasqu
left a comment
Collaborator
There was a problem hiding this comment.
Some small initial questions, not too familiar with the new system so wanna make sure I understand
3outeille
marked this pull request as draft
October 7, 2026 13:25
3outeille
changed the title
3outeille
marked this pull request as ready for review
October 8, 2026 00:56
Member
Author
|
@vasqu decided to rely on dtensor which is cleaner |
vasqu
approved these changes
Oct 8, 2026
vasqu
left a comment
vasqu
left a comment
Collaborator
There was a problem hiding this comment.
Much cleaner now thanks 🫡
| total_byte_count[device] += param_byte_count | ||
| # Parallelism is applied before loading, so a sharded parameter is already a DTensor placeholder | ||
| # Therefore, we can use the local tensor size to determine the total byte count. | ||
| numel = param._local_tensor.numel() if is_dtensor(param) else param.numel() |
Collaborator
There was a problem hiding this comment.
Oh this is really neat, just to be sure it's safe for torch 2.6+? Got bitten too many times recently
…e/transformers into fix-warmup-byte-count-ep-plan
Contributor
CI recapDashboard: View test results in Grafana |
ArthurZucker
approved these changes
Oct 9, 2026
ArthurZucker
left a comment
ArthurZucker
left a comment
Collaborator
There was a problem hiding this comment.
ty, and yeah thanks for pushing for DTensor its a great design for this!
Comment on lines
+482
to
+483
| # Mixtral's default plans would also work, since EP rules take precedence over TP rules on shared keys, but the | ||
| # plans are spelled out so that readers can easily understand the intent. |
Collaborator
There was a problem hiding this comment.
Suggested change
| # Mixtral's default plans would also work, since EP rules take precedence over TP rules on shared keys, but the | |
| # plans are spelled out so that readers can easily understand the intent. | |
| # set this to allows default plan to change |
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
the
get_total_byte_countfunction is used to calculate the total bytes count needed to load the model on each device by reading thetp_plan. This is useful forcaching_allocator_warmupas we want to know how much cache we need to pre-allocate.Since we support more parallelism now, we can rely on dtensor to get the local tensor and thus its local shape after being distributed.