Skip to content

feat(dynamo): add native inference backend#2964

Open
biswapanda wants to merge 18 commits into
PrimeIntellect-ai:mainfrom
biswapanda:bis/dynamo-core
Open

feat(dynamo): add native inference backend#2964
biswapanda wants to merge 18 commits into
PrimeIntellect-ai:mainfrom
biswapanda:bis/dynamo-core

Conversation

@biswapanda

@biswapanda biswapanda commented Jul 6, 2026

Copy link
Copy Markdown

Summary

Adds Dynamo as a first-class Prime RL inference provider and routes rollout generation through the provider-neutral generate request/response path.

This branch is rebased directly on canonical Prime RL main at 84b413fd73a5b5f85f772e42bf1beb52fe0a265e; the current PR head is f077fbf9620f144ec2cd5d74c4dc0eb147f5ceb2.

What changed

  • adds Dynamo provider configuration, launch, discovery, and admin integration;
  • carries policy-version metadata through generation for strict off-policy control;
  • coordinates pause/resume and distributed NCCL weight updates across discovered workers;
  • unifies generation behind the common request/response model;
  • updates policy-marker handling for the current verifiers TraceTask API;
  • expands focused configuration, barrier, and inference tests.

Why it is needed

Prime RL previously assumed direct ownership of native vLLM workers. Dynamo introduces a frontend/router and independently discovered prefill/decode workers, so rollout traffic and administrative weight-update traffic must be separated while retaining stable worker identity, policy-version accounting, and synchronized weight publication.

Validation

  • 17/17 orchestrator/barrier tests passed;
  • 11/11 Dynamo configuration tests passed;
  • 181/181 stacked Prime inference/Kubernetes tests passed;
  • a broader targeted run passed 304 tests; 17 collection cases were unavailable only because optional external task-set plugins were not installed;
  • controlled Qwen3-30B-A3B-Thinking-2507 Hendrycks Math E2E passed in native-Dynamo and OpenEngine modes using the same digest-pinned image:
    • five orchestrator and five four-rank trainer steps in each mode;
    • maximum off-policy distance remained 0 for every step;
    • all four workers reached final weight version 3;
    • post-training generation returned 16 tokens in each mode;
    • trainer/GPU cleanup, release deletion, and artifact-manifest verification passed.

Accepted evidence:

  • baseline: /data/openengine-e2e/runs/q30-baseline-final-20260711T111946Z
  • OpenEngine: /data/openengine-e2e/runs/q30-openengine-final-20260711T114759Z
  • image digest: sha256:9c0f57ca6634a569e1df53089cecf6f5d6b622e9e49bb58ebc609f976157e1e6

@biswapanda biswapanda changed the title [DYNAMO] add native inference backend feat(dynamo): add native inference backend Jul 7, 2026
@biswapanda biswapanda marked this pull request as ready for review July 7, 2026 00:39
Comment thread src/prime_rl/inference/dynamo.py Outdated
Comment thread src/prime_rl/entrypoints/inference.py
Comment thread src/prime_rl/utils/client.py Outdated
Comment thread src/prime_rl/orchestrator/dispatcher.py

@cursor cursor Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

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

Cursor Bugbot has reviewed your changes and found 1 potential issue.

Fix All in Cursor

❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.

Reviewed by Cursor Bugbot for commit b653a77. Configure here.

Comment thread src/prime_rl/inference/dynamo_admin.py Outdated
Signed-off-by: Biswa Panda <biswa.panda@gmail.com>
Signed-off-by: Biswa Panda <biswa.panda@gmail.com>
Signed-off-by: Biswa Panda <biswa.panda@gmail.com>
Signed-off-by: Biswa Panda <biswa.panda@gmail.com>
Signed-off-by: Biswa Panda <biswa.panda@gmail.com>
Signed-off-by: Biswa Panda <biswa.panda@gmail.com>
Signed-off-by: Biswa Panda <biswa.panda@gmail.com>
Signed-off-by: Biswa Panda <biswa.panda@gmail.com>
Signed-off-by: Biswa Panda <biswa.panda@gmail.com>
Signed-off-by: Biswa Panda <biswa.panda@gmail.com>
Signed-off-by: Biswa Panda <biswa.panda@gmail.com>
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.

1 participant