Skip to content

Fix dtype-aware IPC buffers for RL colocated weight updates#1789

Merged
hhaAndroid merged 1 commit into
InternLM:rl_designfrom
CyCle1024:fix_update_weight_dtype_mismatch
May 15, 2026
Merged

Fix dtype-aware IPC buffers for RL colocated weight updates#1789
hhaAndroid merged 1 commit into
InternLM:rl_designfrom
CyCle1024:fix_update_weight_dtype_mismatch

Conversation

@CyCle1024
Copy link
Copy Markdown
Collaborator

@CyCle1024 CyCle1024 commented May 14, 2026

Summary

  • Cache LMDeploy IPC tensors by dtype during RL colocated weight updates.
  • Track the last sent dtype so dtype switches resend the IPC tensor expected by LMDeploy.
  • Track IPC buffer capacity per dtype to avoid reusing undersized cached buffers.

@CyCle1024 CyCle1024 requested a review from hhaAndroid May 14, 2026 13:00
@hhaAndroid hhaAndroid merged commit b8988ea into InternLM:rl_design May 15, 2026
3 checks passed
hhaAndroid pushed a commit that referenced this pull request May 15, 2026
fix(rl): dtype mismatch in RL colocated update weight
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