Skip to content

update: support data parallel without expert parallel#376

Open
rebel-wonsubkim wants to merge 8 commits intodev-0.12-rebasefrom
dev-0.12-rebase_dp_wo_ep
Open

update: support data parallel without expert parallel#376
rebel-wonsubkim wants to merge 8 commits intodev-0.12-rebasefrom
dev-0.12-rebase_dp_wo_ep

Conversation

@rebel-wonsubkim
Copy link
Copy Markdown
Contributor

@rebel-wonsubkim rebel-wonsubkim commented Feb 13, 2026

  • support data parallel without expert parallel for MoE model
  • fix up sampler warm-up for pipeline parallel
  • disable dp metadata if ep is disabled

+ DO NOT flatten tp across dp if expert_parallel is disabled in RBLN

Signed-off-by: wonsub kim <subang0@rebellions.ai>
Signed-off-by: wonsub kim <subang0@rebellions.ai>
Signed-off-by: wonsub kim <subang0@rebellions.ai>
"""
dp_size = self.vllm_config.parallel_config.data_parallel_size
dp_size = self.vllm_config.parallel_config.data_parallel_size \
and self.vllm_config.parallel_config.enable_expert_parallel
Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think it needs to be fixed like below.
dp_ep_enable = self.vllm_config.parallel_config.data_parallel_size > 1
and self.vllm_config.parallel_config.enable_expert_parallel

dp_size > 1 --> dp_ep_enable

rebel-ykchoi and others added 5 commits March 4, 2026 15:21
- add ManualBucketingManager
  - builds decode buckets from a user-provided bucket list
- add rbln_envs
  - add "manual" to VLLM_RBLN_DECODE_BATCH_BUCKET_STRATEGY options
  - add VLLM_RBLN_DECODE_BATCH_BUCKET_MANUAL_BUCKETS env var
    - comma-separated integers
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.

3 participants