freesolo-flash-dev 1.0.78__tar.gz → 1.0.80__tar.gz

{freesolo_flash_dev-1.0.78 → freesolo_flash_dev-1.0.80}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: freesolo-flash-dev
-Version: 1.0.78
+Version: 1.0.80
 Summary: Flash — managed LoRA post-training (SFT/GRPO/OPD) for Freesolo environments, driven by the `flash` CLI
 Project-URL: Homepage, https://github.com/freesolo-co/flash
 Project-URL: Repository, https://github.com/freesolo-co/flash

{freesolo_flash_dev-1.0.78 → freesolo_flash_dev-1.0.80}/flash/cli/envpush.py

@@ -762,9 +762,9 @@ def _iter_env_sidecar_files(
     if not include_full_tree:
         import ast
 
-        # a single-file push carries only the entrypoint, its dataset sidecars, and its own
-        # docs. ship user-authored docs so the synthesized stub readme does not stand in for
-        # real training guidance the user shipped next to the module.
+        # a single-file push uses bounded sidecar selections. root *.toml and recursive
+        # configs/**/*.toml are handled explicitly below rather than widening to neighbouring trees.
+        # ship user-authored docs so the synthesized stub readme does not replace real guidance.
         for doc_name in ("README.md", "TRAINING.md"):
             doc = env_root / doc_name
             if doc.is_file() and not _ignore_env_push_path(
@@ -772,6 +772,43 @@ def _iter_env_sidecar_files(
             ):
                 yielded.add(doc)
                 yield doc, doc.relative_to(env_root)
+        # root config files and the conventional configs tree are runtime inputs, not a request to
+        # widen an exact-file push to unrelated neighbouring trees. apply the shared ignore policy
+        # and yielded set so the limit and copy passes select the identical safe members.
+        for config in sorted(env_root.iterdir()):
+            if (
+                config.is_file()
+                and config.suffix.lower() == ".toml"
+                and config not in yielded
+                and not _ignore_env_push_path(config, env_root=env_root, entrypoint=entrypoint)
+            ):
+                yielded.add(config)
+                yield config, config.relative_to(env_root)
+        configs = env_root / "configs"
+        if configs.is_dir() and not _ignore_env_push_path(
+            configs, env_root=env_root, entrypoint=entrypoint
+        ):
+            for root, dirs, files in os.walk(
+                configs, topdown=True, followlinks=False, onerror=_raise_walk_error
+            ):
+                root_path = Path(root)
+                dirs[:] = sorted(
+                    name
+                    for name in dirs
+                    if not _ignore_env_push_path(
+                        root_path / name, env_root=env_root, entrypoint=entrypoint
+                    )
+                )
+                for name in sorted(files):
+                    config = root_path / name
+                    if (
+                        config.suffix.lower() != ".toml"
+                        or config in yielded
+                        or _ignore_env_push_path(config, env_root=env_root, entrypoint=entrypoint)
+                    ):
+                        continue
+                    yielded.add(config)
+                    yield config, config.relative_to(env_root)
         # the entrypoint's OWN imports ship, exactly as its evaluation sidecar's do. seeding the
         # closure only from evaluations.py published an environment whose entrypoint imported
         # siblings without them: `env push` exits 0 and prints an id, so nothing surfaces until a

{freesolo_flash_dev-1.0.78 → freesolo_flash_dev-1.0.80}/flash/engine/worker/opd_train.py

@@ -1620,6 +1620,8 @@ def build_opd_overrides(config: dict) -> list[str]:
         # that mattered (VERL-123). size the pool to the container instead. this also keeps the
         # storage-unit reservation above satisfiable, since it stays well under the cpu floor.
         f"ray_kwargs.ray_init.num_cpus={ray_num_cpus(config['n_gpus_per_node'])}",
+        # num_gpus is absent from verl's generated ray_init node, so hydra requires add-key syntax.
+        f"+ray_kwargs.ray_init.num_gpus={config['n_gpus_per_node']}",
         "critic.enable=false",
         "reward.reward_model.enable=false",
         # disabling the reward MODEL does not disable reward SCORING. with no custom function the