Remove drive.ini; puffer_drive.yaml is the single config source

MANIFEST.in

@@ -4,7 +4,6 @@ global-include *.h
 global-include *.cpp
 global-include *.cu
 global-include *.py
-recursive-include pufferlib/config *.ini
 recursive-include pufferlib/resources *
 recursive-exclude experiments *
 recursive-exclude wandb *

README.md

@@ -231,9 +231,9 @@ To add a metric to the report: add its key to `TREND_METRICS` /
 `FINALS_METRICS` / `CURVE_METRICS` in `scripts/nightly_report.py`, then run
 `update` followed by `report`.
 
-## Key Configuration (`pufferlib/config/ocean/drive.ini`)
+## Key Configuration (`pufferlib/config/puffer_drive.yaml`)
 
-### `[env]` — Simulation
+### `env` — Simulation
 
 | Parameter | Default | Notes |
 |-----------|---------|-------|
@@ -250,7 +250,7 @@ To add a metric to the report: add its key to `TREND_METRICS` /
 | `reward_conditioning` | `False` | Condition policy on reward weights |
 | `reward_randomization` | `False` | Randomize reward weights each episode |
 
-### `[env]` — Reward Shaping
+### `env` — Reward Shaping
 
 | Parameter | Default | Effect |
 |-----------|---------|--------|
@@ -263,7 +263,7 @@ To add a metric to the report: add its key to `TREND_METRICS` /
 | `reward_stop_line` | `1.0` | Stop line penalty |
 | `reward_overspeed` | `0.05` | Speeding penalty |
 
-### `[train]` — PPO
+### `train` — PPO
 
 | Parameter | Default | Notes |
 |-----------|---------|-------|
@@ -277,7 +277,7 @@ To add a metric to the report: add its key to `TREND_METRICS` /
 | `vtrace_rho_clip` / `vtrace_c_clip` | `1` | V-trace IS ratio clipping |
 | `adv_sampling_prio_alpha` / `adv_sampling_prio_beta0` | `0.85` | Priority sampling exponents |
 
-### `[policy]` — Network
+### `policy` — Network
 
 | Parameter | Default | Notes |
 |-----------|---------|-------|

data_utils/fetch_data.py

@@ -1,7 +1,7 @@
 """Fetch shared PufferDrive datasets from S3.
 
 Datasets are declared in data_utils/datasets.yaml and sync into
-<repo>/data/<name>/ (gitignored), where the drive.ini defaults expect them.
+<repo>/data/<name>/ (gitignored), where the default config expects them.
 Fetching elsewhere (--data-root) means pointing env.map_dir there too.
 
     python data_utils/fetch_data.py                     # the default ~10 GB mini sets
@@ -95,7 +95,7 @@ def main():
     parser.add_argument(
         "--data-root",
         default=DEFAULT_DATA_ROOT,
-        help="destination root (default: <repo>/data, where the drive.ini defaults expect it)",
+        help="destination root (default: <repo>/data, where the default config expects it)",
     )
     parser.add_argument("--dry-run", action="store_true", help="show what would be downloaded without writing")
     args = parser.parse_args()

docs/data_storage.md

@@ -6,7 +6,7 @@ lab's S3 buckets and pulled onto a machine for training and evaluation.
 ## Fetching data
 
 Datasets are registered in `data_utils/datasets.yaml` and land under
-`<repo>/data/` (gitignored), where the `drive.ini` defaults expect them.
+`<repo>/data/` (gitignored), where the default config expects them.
 Fetching elsewhere (`--data-root`) means pointing `env.map_dir` there too.
 
 ```bash

docs/evaluation.md

@@ -8,10 +8,10 @@ training and standalone from the CLI — there is no second eval path.
 
 ## Concepts
 
-- **Evaluator** — one evaluation, defined by an `[eval.<name>]` section in
-  `pufferlib/config/ocean/drive.ini`. It owns an env config, a rollout, a set
+- **Evaluator** — one evaluation, defined by an `eval.<name>` section in
+  `pufferlib/config/puffer_drive.yaml`. It owns an env config, a rollout, a set
   of metrics, and optional rendering.
-- **EvalManager** — discovers every `[eval.<name>]` section, instantiates the
+- **EvalManager** — discovers every `eval.<name>` section, instantiates the
   evaluators, and runs the ones whose `interval` is due (inline) or that you
   name (standalone). One evaluator failing doesn't stop the rest.
 - **Evaluator types** (the `type` field): `multi_scenario` (sweep a scenario
@@ -20,20 +20,24 @@ training and standalone from the CLI — there is no second eval path.
 
 ## Config schema
 
-```ini
-[eval.<name>]
-type          = "multi_scenario"   ; registered evaluator class (omit for a template)
-enabled       = true               ; skip when false
-interval      = 250                ; run every N epochs inline (0 disables inline)
-mode          = "inline"           ; "inline" (block training) | "subprocess"
-inherits      = "<other_section>"  ; optional: pull defaults from another section
-clean         = true               ; zero perturbations/dropout + enforce red lights
-render        = true               ; capture renders during the rollout
-render_views  = ["sim_state","bev"]; camera views for the egl backend
-render_backend = "egl"             ; egl | triage_html | obs_html (see Render backends)
-env.<key>     = <value>            ; any [env] override (dotted)
-eval.<key>    = <value>            ; evaluator-specific knob (see below)
-vec.<key>     = <value>            ; any [vec] override
+```yaml
+eval:
+  <name>:
+    type: multi_scenario        # registered evaluator class (omit for a template)
+    enabled: 'true'             # skip when false
+    interval: 250               # run every N epochs inline (0 disables inline)
+    mode: inline                # inline (block training) | subprocess
+    inherits: <other_section>   # optional: pull defaults from another section
+    clean: 'true'               # zero perturbations/dropout + enforce red lights
+    render: 'true'              # capture renders during the rollout
+    render_views: [sim_state, bev]  # camera views for the egl backend
+    render_backend: egl         # egl | triage_html | obs_html (see Render backends)
+    env:                        # any env override
+      <key>: <value>
+    eval:                       # evaluator-specific knobs (see below)
+      <key>: <value>
+    vec:                        # any vec override
+      <key>: <value>
 ```
 
 A section **without** a `type` is a *template*: it is never instantiated, only
@@ -70,23 +74,23 @@ puffer eval puffer_drive --evaluator validation_gigaflow \
     --load-model-path experiments/puffer_drive_xxxx/models/model_000500.pt
 ```
 
-Runs that one evaluator with its `[eval.validation_gigaflow]` config. The
+Runs that one evaluator with its `eval.validation_gigaflow` config. The
 checkpoint's network architecture is read from the sibling `config.yaml` (next
 to `models/`), so a checkpoint loads even if its policy/rnn dims differ from
-`drive.ini`. With no `--load-model-path`, a fresh (random) policy is used —
+`puffer_drive.yaml`. With no `--load-model-path`, a fresh (random) policy is used —
 useful for smoke-testing the eval path itself.
 
 ### Standalone, ad-hoc
 
 Same as by-name, except instead of naming an evaluator you select one of the two
 built-in `validation_*` evaluators by simulation and override its config from the
-CLI — no `drive.ini` edit needed:
+CLI — no config edit needed:
 
 - `--eval_simulation gigaflow` → runs the `validation_gigaflow` section
 - `--eval_simulation replay` → runs the `validation_replay` section
 
 The flags below override that evaluator's config for this run, and each applies
-**only when passed** — omit one and the evaluator's own `[eval.*]` value stands:
+**only when passed** — omit one and the evaluator's own `eval.*` value stands:
 
 ```bash
 puffer eval puffer_drive --eval_simulation gigaflow \
@@ -162,7 +166,7 @@ weights and clean-eval knobs from the `validation_defaults` template.
    `evaluators/__init__.py`. Most subclasses only override `env_overrides`,
    `_should_stop`, and optionally `_render_env_overrides`; the base `rollout`
    handles the step loop, metric aggregation, CSV, and coverage.
-2. Add an `[eval.<name>]` section with `type = "<your_type_name>"`.
+2. Add an `eval.<name>` section with `type: <your_type_name>`.
 
 ## Scripts