Skip to content

Release/v26.5#90

Merged
yeandy merged 3 commits into
ROCm:release/v26.5from
yeandy:release/v26.5
Jun 18, 2026
Merged

Release/v26.5#90
yeandy merged 3 commits into
ROCm:release/v26.5from
yeandy:release/v26.5

Conversation

@yeandy

@yeandy yeandy commented Jun 18, 2026

Copy link
Copy Markdown
Collaborator

Description

Start with a short description of what the PR does and how this is a change from
the past.

The rest of the description includes relevant details and context, examples:

  • why is this change being made,
  • the problem being solved and any relevant context,
  • why this is a good solution,
  • some information about the specific implementation,
  • shortcomings of the solution and possible future improvements.

If the change fixes a bug or a Github issue, please include a link, e.g.,:
FIXES: b/123456
FIXES: #123456

You can also provide a comma-separated list. If you don't want to close a bug but
simply to reference it, use BUGS, e.g.:
BUGS: b/123456

Notice 1: Once all tests pass, the "pull ready" label will automatically be assigned.
This label is used for administrative purposes. Please do not add it manually.

Notice 2: For external contributions, our settings currently require an approval from a MaxText maintainer to trigger CI tests.

Tests

Please describe how you tested this change, and include any instructions and/or
commands to reproduce.

Checklist

Before submitting this PR, please make sure (put X in square brackets):

  • I have performed a self-review of my code. For an optional AI review, add the gemini-review label.
  • I have necessary comments in my code, particularly in hard-to-understand areas.
  • I have run end-to-end tests tests and provided workload links above if applicable.
  • I have made or will make corresponding changes to the doc if needed, including adding new documentation pages to the relevant Table of Contents (toctree directive) as explained in our documentation.

The dispatch/MLP MoE activations are already expert-sharded via
activation_exp. Since AI-Hypercomputer#4007, their batch dim also maps to
activation_batch_moe, which includes 'expert'. Under single-node expert
parallelism (ici_expert_parallelism=-1) this double-maps two tensor dims
onto the 'expert' mesh axis, so GSPMD falls back from expert-parallel
AllToAll to FSDP-style AllGather+ReduceScatter, regressing throughput for
few-large-expert models (e.g. Mixtral-8x7b: ~7.4k -> ~10.9k tok/s/device
at bs=11 on 8x MI355X).

Add a config flag moe_dispatch_no_expert_sharding (default false) that
selects a new activation_batch_no_exp rule ([data, fsdp, fsdp_transpose],
no 'expert') for the training dispatch/MLP batch axis. Enable it for
mixtral-8x7b. Default-false keeps every other model and all TPU/non-EP
paths byte-identical; the flag only changes sharding when the 'expert'
mesh axis size > 1.
…e-expert geometry as 8x7b, so it benefits from the same expert-parallel MoE dispatch/MLP sharding.
…oe_dispatch_no_expert_sharding the expert dim is sharded by 'expert' and the batch dim is not, guarding the expert-parallel dispatch/MLP sharding.
@yeandy yeandy merged commit 9644edd into ROCm:release/v26.5 Jun 18, 2026
2 of 5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants