WIP: Greene nightly multi-agent launcher + date-scoped wandb groups (do not merge)#497
Draft
eugenevinitsky wants to merge 2 commits into
Draft
WIP: Greene nightly multi-agent launcher + date-scoped wandb groups (do not merge)#497eugenevinitsky wants to merge 2 commits into
eugenevinitsky wants to merge 2 commits into
Conversation
Mirror the existing single-agent nightly launcher with a multi-agent
counterpart so both flavors can run on Greene with the same shape.
- scripts/cluster_configs/nightly_best.yaml — multi-agent (oignons2)
config (720k agents, 8 maps, 10B steps, gigaflow encoder + split
network).
- scripts/launch_nightly_best.sh — Greene launcher; same env-overridable
shape as launch_single_agent.sh but defaults TIME=1800 / MEM=192gb to
match the multi-agent profile.
- Both launchers add wandb_group=${DATE_STAMP} to --args so a night's 3
seeds cluster under one date-named wandb group.
- Yamls now point at split wandb projects:
single_agent_speed_run.yaml -> nightly-single
nightly_best.yaml -> nightly-multi
Date-scoped grouping happens within each project.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This was referenced Jun 28, 2026
Comments described the file as "derived from oignons2 weights/oignons2/config.yaml" — code-history detail that doesn't help a future reader and breaks once the source branch is gone. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
vcharraut
reviewed
Jun 29, 2026
Comment on lines
+94
to
+95
| train.minibatch_size: 153600 | ||
| train.max_minibatch_size: 153600 |
Collaborator
There was a problem hiding this comment.
imo the bath size can be simplified to power of 2, with adv filtering and masking you never get perfect bs anyway
vcharraut
reviewed
Jun 29, 2026
Comment on lines
+80
to
+89
| policy.input_size: 256 | ||
| policy.backbone_hidden_size: 1024 | ||
| policy.backbone_num_layers: 3 | ||
| policy.actor_hidden_size: 1024 | ||
| policy.actor_num_layers: 0 | ||
| policy.critic_hidden_size: 1024 | ||
| policy.critic_num_layers: 0 | ||
| policy.split_network: true | ||
| policy.encoder_gigaflow: true | ||
| policy.dropout: 0.0 |
Collaborator
There was a problem hiding this comment.
this should be deprecated, especially the policy.encoder_gigaflow
Comment on lines
+38
to
+39
| env.obs_slots_lane_n: 80 | ||
| env.obs_slots_boundary_n: 80 |
Collaborator
There was a problem hiding this comment.
I think the env.obs_slots_boundary_n: 80 can be reduced, since it is not similar to gigaflow implen
| env.simulation_mode: gigaflow | ||
| env.map_dir: pufferlib/resources/drive/binaries/carla | ||
| env.num_maps: 8 | ||
| env.num_agents: 720000 |
Collaborator
There was a problem hiding this comment.
this will break the memory i think?
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
2 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.
Summary
WIP — adds a Greene nightly launcher for the multi-agent (oignons2-derived) config, mirroring the existing single-agent launcher, plus date-scoped wandb grouping for both.