Adopt the best-launch training recipe as the default config; nightly derives from it#537
Open
eugenevinitsky wants to merge 10 commits into
Open
Adopt the best-launch training recipe as the default config; nightly derives from it#537eugenevinitsky wants to merge 10 commits into
eugenevinitsky wants to merge 10 commits into
Conversation
eugenevinitsky
commented
Jul 24, 2026
eugenevinitsky
force-pushed
the
ev/mimolette-defaults
branch
from
July 24, 2026 02:52
13ec391 to
de369c7
Compare
Collaborator
|
In my opinion the default config should run on premise on (almost) any machine to debug/dev purposes and we add the scale run to a different .yaml |
vcharraut
reviewed
Jul 24, 2026
| precision: float32 | ||
| total_timesteps: 10000000000 | ||
| precision: bfloat16 | ||
| total_timesteps: 62500000000 |
Author
There was a problem hiding this comment.
Done — default train.total_timesteps is now 500_000_000_000 (nightly still overrides to a night-sized 10B).
puffer_drive.yaml defaults now match weights/mimolette/config.yaml (our best model) for the full training recipe: env physics and scale (dt 0.3, 3200 agents, 120 max per env, scenario_length 2560), map goal source with lane-distance obs, mimolette observation shaping and dropouts, partner-blindness/phantom-braking perturbations, reward conditioning + randomization, the 3x1024 split actor/critic policy, and the optimizer recipe (256k minibatch with 64k accumulation chunks, 3 update epochs, compiled bfloat16, no reward normalization, 62.5B-step schedule), plus vec.num_envs 40. Not adopted (infra/bookkeeping, not recipe): map_dir/num_maps (local CARLA instead of the 72-map GCS set), data_dir, checkpoint_interval, and keys absent from the 3.0 schema (traffic_control_scope). nightly_best.yaml shrinks to run-scale and logging only — night-sized 10B timesteps, checkpoint interval, eval toggle, wandb — so the nightly always trains the current default recipe. single_agent_speed_run.yaml pins vec.num_envs=20, the one default it silently inherited. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The embedding-dim heatmap stacked pooled embeddings across encoders, which requires every encoder to share one width. The mimolette policy uses per-encoder widths (128/256), so NaN-pad rows to the widest instead; imshow leaves the padding blank. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…te-golden] Per review: the default config should run on-premise on (almost) any machine. Defaults keep the mimolette recipe (env physics, goals, observations, perturbations, rewards, policy architecture, precision) but stay at dev scale: 1024 agents, 20 envs, 64k minibatch, compile off. The scale knobs for a real training run (3200 agents, 40 envs, 256k minibatch with 64k accumulation, compile) move to nightly_best.yaml. total_timesteps default is 500B per review. The single_agent vec.num_envs pin is dropped — it existed only because the default briefly moved to 40. The [update-golden] tag regenerates the smoke goldens: the smoke config pins its own scale but inherits the recipe defaults (dt 0.3, map goal source, obs shaping), so the golden metrics legitimately shift. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
eugenevinitsky
force-pushed
the
ev/mimolette-defaults
branch
from
July 26, 2026 21:52
772b97a to
a91f276
Compare
Author
|
Restructured per this thread: the defaults keep the mimolette recipe (env physics, goals, observations, perturbations, rewards, policy architecture) but stay dev-sized and the full values move to nightly best |
The workflow only refreshed drive_smoke_golden.json; the rollout golden (drive_rollout_golden.json, added with test_drive_rollout) had no regeneration path, so any legitimate metric shift left the rollout smoke test permanently red. Run both golden tests in the regen container, and commit/upload whichever files changed. The commit message also stops naming the feature from the workflow's first use. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
eugenevinitsky
changed the title
Drop the 68 override keys that now just repeat puffer_drive.yaml defaults; only genuine single-agent divergences remain (single Town10HD map, 1 agent/env, single close goal, no partner/traffic obs, speed-run reward weights, smaller shared-backbone policy, short schedule). Verified the pruned overrides compose to an identical effective config, modulo two trigger-prob knobs that are dead with their enable probabilities at 0. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
eugenevinitsky
force-pushed
the
ev/mimolette-defaults
branch
from
July 26, 2026 23:22
a182648 to
29bc59e
Compare
Port of the 2.0-lineage change: submit_cluster.py gains --wandb-base-url/--wandb-entity flags exported as env vars in the container command (not --args, whose ':' sweep splitting breaks URLs), and both nightly launchers forward WANDB_BASE_URL/WANDB_ENTITY from the launch environment. Lets a launch pin the wandb server + org (e.g. api.wandb.ai / emerge_) over the container's self-hosted default. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
All evaluator sections in puffer_drive.yaml (validation_replay, validation_gigaflow, dnf_triage) and the code fallback in Evaluator._render_pass switch from egl/triage_html to obs_html. validation_gigaflow drops its egl-only render_views and no longer spawns the eight-way ffmpeg/EGL spike at eval time; the single-agent launcher's obs_html override becomes the default and is removed. The egl mp4-naming test pins render_backend=egl explicitly. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
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.
What
pufferlib/config/puffer_drive.yamldefaults now carry the best-launch training recipe:goal_source: mapwith lane-distance obs, goal_speed 3.0, spacing up to 200 m; updated obs slots/ranges/norms, lane dropout 0.3 / boundary 0.4, lane stride 2; partner-blindness and phantom-braking perturbations on; reward conditioning + randomization on, lane_center 0.001.normalize_rewards: false, 500B-step schedule.compile: false. The full-scale knobs (3200 agents, 40 envs, 256k minibatch with 64k accumulation chunks, compile on) live inscripts/cluster_configs/nightly_best.yaml, which shrinks from ~110 override lines to scale + run-length + logging — the nightly now always trains whatever the default recipe is.05_inference.pynotebook: the embedding heatmap cell now handles per-encoder embedding widths (the new encoders are heterogeneous; the old uniform-64 defaults were the only reasonnp.stackworked).update-smoke-golden.ymlnow regenerates both goldens —drive_rollout_golden.jsonpreviously had no regeneration path, so any legitimate metric shift left the rollout smoke test permanently red.single_agent_speed_run.yamlalready pins its full recipe explicitly, so that launcher's behavior is unchanged.Why
The best-known recipe lived only in a checkpoint's
config.yamland, approximately, in a 110-line nightly override file that had drifted from it (num_agents 4096 vs 3200, goal_speed 1000 vs 3, reward_goal 0.5 vs 1.0, blindness probs, minibatch sizing). Making it the default meanspuffer train puffer_driveis the best-launch recipe at dev scale, the nightly can't drift from it again, and future recipe improvements are one edit in one place.Notes
map_dir/num_maps(defaults stay on the 8 local CARLA maps),data_dir,checkpoint_interval.[update-golden]— the recipe defaults (dt 0.3, map goal source, obs shaping) legitimately shift the golden metrics.launch_nightly_best.sh's defaultMEM=192gblikely still holds but the first night is worth watching.🤖 Generated with Claude Code