hud-python 0.6.5__tar.gz → 0.6.7__tar.gz

{hud_python-0.6.5 → hud_python-0.6.7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hud-python
-Version: 0.6.5
+Version: 0.6.7
 Summary: SDK for the HUD platform.
 Project-URL: Homepage, https://github.com/hud-evals/hud-python
 Project-URL: Bug Tracker, https://github.com/hud-evals/hud-python/issues

hud_python-0.6.7/cookbooks/fireworks-rl-training/README.md

@@ -0,0 +1,129 @@
+# Fireworks RL Training
+
+Direct Fireworks Training API loop over the same arithmetic preview task used by
+`cookbooks/rl-training`.
+
+This does **not** use Fireworks native datasets or RFT jobs. It follows the
+Training API service path from the Fireworks docs:
+
+1. `FiretitanServiceClient.from_firetitan_config(...)`
+2. `create_deployment_sampler(...)` for high-parallel rollouts
+3. local grading of HUD-style multiplication tasks
+4. `forward_backward_custom(...)` + `optim_step(...)`
+5. `save_weights_for_sampler(...)` + sampler refresh
+
+References:
+
+- Fireworks Training API introduction: https://docs.fireworks.ai/fine-tuning/training-api/introduction
+- Training and sampling lifecycle: https://docs.fireworks.ai/fine-tuning/training-api/training-and-sampling
+- Loss functions / GRPO reference: https://docs.fireworks.ai/fine-tuning/training-api/loss-functions
+
+## Setup
+
+The repo-level `.env` is loaded automatically. It must contain:
+
+```bash
+FIREWORKS_API_KEY=...
+FIREWORKS_ACCOUNT_ID=...
+```
+
+Install the isolated cookbook environment:
+
+```bash
+uv sync --pre
+```
+
+## Calibrate task difficulty first
+
+What matters for GRPO is **within-group** reward spread: advantages are computed
+within each prompt group, so a group whose rollouts all score the same (all 0 or
+all 1) produces zero advantage and no gradient — even if the *overall* mean looks
+healthy. Calibration reports `within_group_reward_std` for exactly this; treat
+it, not `reward_mean`, as the signal that training has something to learn.
+
+Two backends:
+
+- `--calibration-backend inference` (default): Fireworks' OpenAI-compatible API.
+  Cheap, but samples `gpt-oss-120b` (`--inference-model`), not the training base —
+  the small serverless catalog on the `lorenss` key has no Qwen3 8B. Use it only
+  for a rough task sanity check.
+- `--calibration-backend managed`: provisions the same deployment sampler that
+  training uses and samples the **actual base model** (Qwen3 8B). This is the
+  calibration that counts. It still skips the trainer and `optim_step`.
+
+```bash
+uv run train.py --calibrate-only --calibration-backend managed \
+  --groups-per-step 6 --rollouts-per-prompt 6 --parallelism 18 --debug-samples 4
+```
+
+`--debug-samples N` prints the first N rollouts (reward, output-token count,
+text) so you can see *why* a group scored the way it did. Tune the multiplication
+range until `within_group_reward_std` is clearly above zero:
+
+- Groups all-correct (`within_group_reward_std ~= 0`) → make it harder
+  (`--min-a/--max-a/--min-b/--max-b`).
+- Groups all-wrong → make it easier, or raise `--max-tokens` so the model can
+  finish its working before the budget cuts it off.
+
+The shipped defaults (3-digit × 3-digit, `--max-tokens 512`, thinking disabled)
+calibrate to `reward_mean ~= 0.47`, `within_group_reward_std ~= 0.20` on Qwen3 8B:
+a regime where the same problem is sometimes solved (when the model shows its
+work) and sometimes slipped (when it answers directly) — so RL has a gradient to
+follow.
+
+### Reasoning models and the token budget
+
+Qwen3 is a hybrid reasoning model: by default it opens a `<think>` block and, on
+a tight `--max-tokens`, spends the whole budget reasoning and never emits the
+answer (reward collapses to zero). This cookbook disables thinking by default
+through the chat template so direct rollouts reach the integer. Pass
+`--enable-thinking` to keep the reasoning block — and raise `--max-tokens`
+accordingly so the answer still fits.
+
+## Train
+
+Once calibration has non-trivial rewards:
+
+```bash
+uv run train.py --steps 5 --groups-per-step 8 --rollouts-per-prompt 8 --parallelism 32
+```
+
+This uses the direct Training API managed service path. If you want calibration
+to go through the managed deployment sampler too, pass
+`--calibration-backend managed`; this provisions the same resources as training.
+
+### Preview account constraints
+
+On the `lorenss` preview account today:
+
+- **Trainer creation works** end to end with a provisioned key: rollouts,
+  `forward_backward_custom`, `optim_step`, checkpoint save, and sampler hotload
+  all run, and multi-step training completes. (An earlier `unkey inference api id
+  is not configured` 500 on trainer creation was an account-side provisioning gap,
+  now resolved.)
+- **LoRA is unavailable**: the validated `qwen3-8b-128k` shape only accepts
+  full-parameter training, so `--lora-rank > 0` fails at trainer creation with
+  `no validated training shape exists for ... trainer_mode=LORA_TRAINER`.
+- **Hotloads sync full 8B weights** between steps and occasionally exceed the
+  SDK's 600s hotload budget (`RuntimeError: Hotload failed for sampler snapshot
+  ...`). This is transient preview-infra latency, not a loop bug — re-running the
+  same command generally proceeds. There is no clean knob to extend the timeout
+  on the managed sampler path.
+
+Metrics are written to:
+
+- `runs/fireworks-rl-preview/metrics.jsonl`
+- `runs/fireworks-rl-preview/reward_loss.png` if `matplotlib` is installed
+
+## Notes
+
+- Defaults use Qwen 3 8B full-parameter training:
+  - `accounts/fireworks/models/qwen3-8b`
+  - `Qwen/Qwen3-8B`
+  - `accounts/fireworks/trainingShapes/qwen3-8b-128k`
+- LoRA can be tested with `--lora-rank N`, but the validated Qwen3 8B training
+  shape currently rejects LoRA mode on the `lorenss` preview account.
+- The first checkpoint sync happens after step 0 and subsequent rollouts sample
+  the updated weights through the same deployment.
+- `--keep-trainer` and `--keep-deployment` are available for debugging. By
+  default the trainer is cleaned up and the deployment scales to zero on exit.

{hud_python-0.6.5 → hud_python-0.6.7}/hud/cli/deploy.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import asyncio
+import json
 import logging
 import os
 import time
@@ -12,6 +13,7 @@ from typing import Any
 
 import httpx
 import typer
+from pydantic import ValidationError
 
 from hud.cli.utils.build_display import display_build_summary
 from hud.cli.utils.build_logs import poll_build_status, stream_build_logs
@@ -19,6 +21,7 @@ from hud.cli.utils.config import parse_env_file, parse_key_value
 from hud.cli.utils.context import create_build_context_tarball, format_size
 from hud.cli.utils.registry import get_registry_environment
 from hud.cli.utils.source import EnvironmentSource, normalize_environment_name
+from hud.eval.runtime import RuntimeConfig
 from hud.utils.exceptions import HudRequestError
 from hud.utils.hud_console import HUDConsole
 from hud.utils.platform import PlatformClient
@@ -32,6 +35,7 @@ class _DeployPlan:
     name: str
     registry_id: str | None
     runtime: str | None
+    runtime_config: dict[str, Any] | None
     env_vars: dict[str, str]
     build_args: dict[str, str]
     build_secrets: dict[str, str]
@@ -75,6 +79,26 @@ def _normalize_runtime(runtime: str | None, console: HUDConsole) -> str | None:
     raise typer.Exit(1)
 
 
+def _load_runtime_config(path: str | None, console: HUDConsole) -> dict[str, Any] | None:
+    if path is None:
+        return None
+    config_path = Path(path).expanduser()
+    try:
+        raw = json.loads(config_path.read_text(encoding="utf-8"))
+        config = RuntimeConfig.model_validate(raw)
+    except FileNotFoundError:
+        console.error(f"Runtime config file not found: {config_path}")
+        raise typer.Exit(1) from None
+    except json.JSONDecodeError as exc:
+        console.error(f"Invalid runtime config JSON in {config_path}: {exc.msg}")
+        raise typer.Exit(1) from exc
+    except ValidationError as exc:
+        console.error(f"Invalid runtime config in {config_path}: {exc}")
+        raise typer.Exit(1) from exc
+    payload = config.request_payload()
+    return payload or None
+
+
 def _load_env_vars(path: Path, console: HUDConsole, *, warn_missing: bool) -> dict[str, str]:
     if not path.exists():
         if warn_missing:
@@ -322,6 +346,7 @@ def _prepare_deploy_plan(
     build_args: list[str] | None,
     build_secrets: list[str] | None,
     runtime: str | None,
+    runtime_config: str | None,
     verbose: bool,
     platform: PlatformClient,
     console: HUDConsole,
@@ -357,11 +382,13 @@ def _prepare_deploy_plan(
     build_args_dict = _parse_key_value_flags(build_args, option="--build-arg", console=console)
     if build_args_dict and verbose:
         console.info(f"Build arguments: {', '.join(build_args_dict.keys())}")
+    normalized_runtime = _normalize_runtime(runtime, console)
 
     return _DeployPlan(
         name=resolved_name,
         registry_id=registry_id,
-        runtime=_normalize_runtime(runtime, console),
+        runtime=normalized_runtime,
+        runtime_config=_load_runtime_config(runtime_config, console),
         env_vars=env_vars,
         build_args=build_args_dict,
         build_secrets=_collect_build_secrets(build_secrets, env_dir=env_dir, console=console),
@@ -379,6 +406,7 @@ def deploy_environment(
     build_args: list[str] | None = None,
     build_secrets: list[str] | None = None,
     runtime: str | None = None,
+    runtime_config: str | None = None,
 ) -> None:
     """Deploy one HUD environment to the platform."""
     hud_console = HUDConsole()
@@ -411,6 +439,7 @@ def deploy_environment(
         build_args=build_args,
         build_secrets=build_secrets,
         runtime=runtime,
+        runtime_config=runtime_config,
         verbose=verbose,
         platform=platform,
         console=hud_console,
@@ -485,6 +514,8 @@ async def _trigger_build(
         payload["registry_id"] = plan.registry_id
     if plan.runtime:
         payload["runtime_provider"] = plan.runtime
+    if plan.runtime_config:
+        payload["runtime_config"] = plan.runtime_config
     if plan.env_vars:
         payload["environment_variables"] = plan.env_vars
     if plan.build_args:
@@ -644,6 +675,7 @@ def deploy_all(
     build_args: list[str] | None = None,
     build_secrets: list[str] | None = None,
     runtime: str | None = None,
+    runtime_config: str | None = None,
 ) -> None:
     """Deploy each HUD environment under a parent directory."""
     hud_console = HUDConsole()
@@ -683,6 +715,7 @@ def deploy_all(
                 build_args=build_args,
                 build_secrets=build_secrets,
                 runtime=runtime,
+                runtime_config=runtime_config,
             )
             succeeded.append(env_dir.name)
         except (typer.Exit, SystemExit):
@@ -762,6 +795,11 @@ def deploy_command(
         "--runtime",
         help="Persist Modal as the hosted runtime for this registry",
     ),
+    runtime_config: str | None = typer.Option(
+        None,
+        "--runtime-config",
+        help="Path to a JSON RuntimeConfig for hosted runs",
+    ),
 ) -> None:
     """Deploy HUD environment to the platform.
 
@@ -781,6 +819,7 @@ def deploy_command(
             build_args=build_args,
             build_secrets=secrets,
             runtime=runtime,
+            runtime_config=runtime_config,
         )
         return
 
@@ -795,4 +834,5 @@ def deploy_command(
         build_args=build_args,
         build_secrets=secrets,
         runtime=runtime,
+        runtime_config=runtime_config,
     )

{hud_python-0.6.5 → hud_python-0.6.7}/hud/cli/presets.py

@@ -122,6 +122,14 @@ ENVIRONMENT_PRESETS: tuple[EnvironmentPreset, ...] = (
         "hud-evals",
         "worldsim-template",
     ),
+    EnvironmentPreset(
+        "robot",
+        "🤖",
+        "Robot",
+        "Robotics: run a VLA policy against a containerized robot sim, graded by task success.",
+        "hud-evals",
+        "robot-template",
+    ),
     EnvironmentPreset(
         "videogamebench",
         "🎮",
@@ -130,6 +138,14 @@ ENVIRONMENT_PRESETS: tuple[EnvironmentPreset, ...] = (
         "hud-evals",
         "videogamebench-template",
     ),
+    EnvironmentPreset(
+        "arc-agi-3",
+        "🧩",
+        "ARC-AGI-3",
+        "Interactive reasoning benchmark: agents play ARC-AGI-3 games.",
+        "hud-evals",
+        "ARC-AGI-3",
+    ),
 )
 
 PRESETS_BY_ID: dict[str, EnvironmentPreset] = {p.id: p for p in ENVIRONMENT_PRESETS}

{hud_python-0.6.5 → hud_python-0.6.7}/hud/cli/tests/test_deploy.py

@@ -179,6 +179,47 @@ class TestCollectEnvironmentVariables:
         assert "INVALID_FORMAT" not in result
 
 
+class TestRuntimeConfigFile:
+    def test_load_runtime_config_uses_sdk_shape(self, tmp_path: Path) -> None:
+        from hud.cli.deploy import _load_runtime_config
+        from hud.utils.hud_console import HUDConsole
+
+        config_path = tmp_path / "runtime.json"
+        config_path.write_text(
+            json.dumps(
+                {
+                    "resources": {"gpu": {"type": "A10G", "count": 2}},
+                    "limits": {"startup_timeout_s": 300},
+                }
+            ),
+            encoding="utf-8",
+        )
+
+        assert _load_runtime_config(str(config_path), HUDConsole()) == {
+            "resources": {"gpu": {"type": "A10G", "count": 2}},
+            "limits": {"startup_timeout_s": 300},
+        }
+
+    def test_load_runtime_config_preserves_null_override(self, tmp_path: Path) -> None:
+        from hud.cli.deploy import _load_runtime_config
+        from hud.utils.hud_console import HUDConsole
+
+        config_path = tmp_path / "runtime.json"
+        config_path.write_text(json.dumps({"resources": None}), encoding="utf-8")
+
+        assert _load_runtime_config(str(config_path), HUDConsole()) == {"resources": None}
+
+    def test_load_runtime_config_rejects_unknown_fields(self, tmp_path: Path) -> None:
+        from hud.cli.deploy import _load_runtime_config
+        from hud.utils.hud_console import HUDConsole
+
+        config_path = tmp_path / "runtime.json"
+        config_path.write_text(json.dumps({"provider_config": {}}), encoding="utf-8")
+
+        with pytest.raises(typer.Exit):
+            _load_runtime_config(str(config_path), HUDConsole())
+
+
 class TestDeployEnvironment:
     """Tests for deploy_environment function."""
 
@@ -262,6 +303,7 @@ class TestDeployAsync:
                     name="test-env",
                     registry_id=None,
                     runtime=None,
+                    runtime_config=None,
                     env_vars={},
                     build_args={},
                     build_secrets={},
@@ -292,6 +334,7 @@ class TestDeployAsync:
                     name="test-env",
                     registry_id=None,
                     runtime=None,
+                    runtime_config=None,
                     env_vars={},
                     build_args={},
                     build_secrets={},
@@ -331,6 +374,7 @@ class TestDeployAsync:
                 name="test-env",
                 registry_id=None,
                 runtime="modal",
+                runtime_config=None,
                 env_vars={},
                 build_args={},
                 build_secrets={},
@@ -343,6 +387,48 @@ class TestDeployAsync:
         assert platform.payload is not None
         assert platform.payload["runtime_provider"] == "modal"
 
+    @pytest.mark.asyncio
+    async def test_trigger_build_sends_runtime_config(self) -> None:
+        from hud.cli.deploy import _DeployPlan, _trigger_build
+        from hud.utils.hud_console import HUDConsole
+        from hud.utils.platform import PlatformClient
+
+        class FakePlatform(PlatformClient):
+            payload: dict[str, object] | None = None
+
+            async def apost(
+                self,
+                path: str,
+                *,
+                json: object | None = None,
+            ) -> dict[str, object]:
+                assert path == "/builds/trigger"
+                assert isinstance(json, dict)
+                object.__setattr__(self, "payload", json)
+                return {"id": "build-1", "registry_id": "registry-1"}
+
+        runtime_config = {"resources": {"gpu": {"type": "A10G", "count": 1}}}
+        platform = FakePlatform("https://api.example", "key")
+        result = await _trigger_build(
+            platform,
+            build_id="build-1",
+            plan=_DeployPlan(
+                name="test-env",
+                registry_id=None,
+                runtime="modal",
+                runtime_config=runtime_config,
+                env_vars={},
+                build_args={},
+                build_secrets={},
+            ),
+            no_cache=False,
+            console=HUDConsole(),
+        )
+
+        assert result == {"id": "build-1", "registry_id": "registry-1"}
+        assert platform.payload is not None
+        assert platform.payload["runtime_config"] == runtime_config
+
 
 class TestSaveDeployLink:
     """Tests for _save_deploy_link function."""

{hud_python-0.6.5 → hud_python-0.6.7}/hud/eval/runtime.py

@@ -108,6 +108,9 @@ class RuntimeConfig(BaseModel):
             self.model_dump() | override.model_dump(exclude_unset=True)
         )
 
+    def request_payload(self) -> dict[str, Any]:
+        return self.model_dump(mode="json", exclude_unset=True)
+
 
 class Provider(Protocol):
     """Server placement: called with the task row being placed, acquire one
@@ -925,7 +928,7 @@ class HostedRuntime:
         if group_id is not None:
             payload["group_id"] = group_id
         if task.runtime_config is not None:
-            runtime_config = task.runtime_config.model_dump(mode="json", exclude_none=True)
+            runtime_config = task.runtime_config.request_payload()
             if runtime_config:
                 payload["runtime_config"] = runtime_config
         await platform.apost("/rollouts/submit", json=payload)

{hud_python-0.6.5 → hud_python-0.6.7}/hud/eval/sync.py

@@ -163,7 +163,7 @@ def task_upload_payload(task: Task) -> dict[str, Any]:
     if task.columns:
         payload["columns"] = task.columns
     if task.runtime_config is not None:
-        payload["runtime_config"] = task.runtime_config.model_dump(exclude_none=True)
+        payload["runtime_config"] = task.runtime_config.request_payload()
     return payload
 
 
@@ -176,7 +176,7 @@ def _task_signature(task: Task) -> str:
     if task.columns:
         sig_data["columns"] = task.columns
     if task.runtime_config is not None:
-        sig_data["runtime_config"] = task.runtime_config.model_dump(exclude_none=True)
+        sig_data["runtime_config"] = task.runtime_config.request_payload()
     return f"{task.id}|" + json.dumps(
         sig_data,
         sort_keys=True,

{hud_python-0.6.5 → hud_python-0.6.7}/hud/eval/tests/test_hosted.py

@@ -164,6 +164,25 @@ async def test_run_submits_and_polls_to_terminal(monkeypatch: pytest.MonkeyPatch
     assert payload["agent"]["config"]["model"] == "test-model"
 
 
+@pytest.mark.asyncio
+async def test_run_preserves_runtime_config_null_override(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    platform = _FakePlatform([{"status": "completed", "reward": 0.5}])
+    monkeypatch.setattr(
+        "hud.eval.runtime.PlatformClient.from_settings", classmethod(lambda cls: platform)
+    )
+
+    await HostedRuntime(poll_interval=0.0).run(
+        Task(env="sums", id="add", runtime_config=RuntimeConfig(resources=None)),
+        _agent(),
+        job_id=uuid.uuid4().hex,
+        trace_id=uuid.uuid4().hex,
+    )
+
+    assert platform.posts[0][1]["runtime_config"] == {"resources": None}
+
+
 @pytest.mark.asyncio
 async def test_run_timeout_requests_platform_cancel(monkeypatch: pytest.MonkeyPatch) -> None:
     platform = _FakePlatform([{"status": "running"}])

{hud_python-0.6.5 → hud_python-0.6.7}/hud/eval/tests/test_sync.py

@@ -148,3 +148,15 @@ def test_task_upload_payload_includes_runtime_config() -> None:
     payload = task_upload_payload(task)
 
     assert payload["runtime_config"] == {"image": "img:tag"}
+
+
+def test_task_upload_payload_preserves_runtime_config_null_override() -> None:
+    task = Task(
+        env="e",
+        id="solve",
+        runtime_config=RuntimeConfig(resources=None),
+    )
+
+    payload = task_upload_payload(task)
+
+    assert payload["runtime_config"] == {"resources": None}

{hud_python-0.6.5 → hud_python-0.6.7}/hud/version.py

@@ -4,4 +4,4 @@ Version information for the HUD SDK.
 
 from __future__ import annotations
 
-__version__ = "0.6.5"
+__version__ = "0.6.7"

{hud_python-0.6.5 → hud_python-0.6.7}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "hud-python"
-version = "0.6.5"
+version = "0.6.7"
 description = "SDK for the HUD platform."
 readme = "README.md"
 requires-python = ">=3.11, <3.13"

hud_python-0.6.5/cookbooks/fireworks-rl-training/README.md

@@ -1,114 +0,0 @@
-# Fireworks RL Training
-
-Direct Fireworks Training API loop over the same arithmetic preview task used by
-`cookbooks/rl-training`.
-
-This does **not** use Fireworks native datasets or RFT jobs. It follows the
-Training API service path from the Fireworks docs:
-
-1. `FiretitanServiceClient.from_firetitan_config(...)`
-2. `create_deployment_sampler(...)` for high-parallel rollouts
-3. local grading of HUD-style multiplication tasks
-4. `forward_backward_custom(...)` + `optim_step(...)`
-5. `save_weights_for_sampler(...)` + sampler refresh
-
-References:
-
-- Fireworks Training API introduction: https://docs.fireworks.ai/fine-tuning/training-api/introduction
-- Training and sampling lifecycle: https://docs.fireworks.ai/fine-tuning/training-api/training-and-sampling
-- Loss functions / GRPO reference: https://docs.fireworks.ai/fine-tuning/training-api/loss-functions
-
-## Setup
-
-The repo-level `.env` is loaded automatically. It must contain:
-
-```bash
-FIREWORKS_API_KEY=...
-FIREWORKS_ACCOUNT_ID=...
-```
-
-Install the isolated cookbook environment:
-
-```bash
-uv sync --pre
-```
-
-## Calibrate task difficulty first
-
-Calibration defaults to Fireworks' OpenAI-compatible inference API, so it does
-**not** create a trainer, provision a Training API deployment, or call
-`optim_step`. This is the cheap way to tune task difficulty before paying for a
-Training API run.
-
-The calibration model is separate from the training base model because the
-`lorenss` key currently exposes only a small serverless inference catalog (no
-Qwen3 8B deployment). Override it with `--inference-model` if you have a closer
-deployed model.
-
-```bash
-uv run train.py --calibrate-only --groups-per-step 8 --rollouts-per-prompt 8 --parallelism 32
-```
-
-The goal is a reward distribution with variance. If reward is all zero, make the
-task easier:
-
-```bash
-uv run train.py --calibrate-only --min-a 10 --max-a 99 --min-b 2 --max-b 9
-```
-
-If reward is all one, make the task harder:
-
-```bash
-uv run train.py --calibrate-only --min-a 1000 --max-a 9999 --min-b 11 --max-b 99
-```
-
-The current defaults are calibrated for the visible `gpt-oss-120b` inference
-model on the `lorenss` key: 2-digit by 1-digit multiplication with a direct
-"reply only with the integer" prompt. A 32-rollout calibration gave a non-trivial
-baseline (`reward_mean ~= 0.22`, `reward_std ~= 0.42`), while the original
-3-digit by 2-digit range was all-zero.
-
-## Train
-
-Once calibration has non-trivial rewards:
-
-```bash
-uv run train.py --steps 5 --groups-per-step 8 --rollouts-per-prompt 8 --parallelism 32
-```
-
-This uses the direct Training API managed service path. If you want calibration
-to go through the managed deployment sampler too, pass
-`--calibration-backend managed`; this provisions the same resources as training.
-
-### Current Fireworks preview account blocker
-
-On the `lorenss` preview account, trainer creation currently fails before the
-first train step with:
-
-```text
-failed to ensure FIREWORKS_API_KEY secret: unkey inference api id is not configured
-```
-
-This happens even with `create_deployment=False`, so it is an account/control
-plane provisioning issue rather than a problem in the rollout or loss code. Once
-Fireworks enables the missing Unkey inference API config for the account, the
-same `uv run train.py ...` command should proceed to trainer startup and the
-first `forward_backward_custom(...)` call.
-
-Metrics are written to:
-
-- `runs/fireworks-rl-preview/metrics.jsonl`
-- `runs/fireworks-rl-preview/reward_loss.png` if `matplotlib` is installed
-
-## Notes
-
-- Defaults use Qwen 3 8B full-parameter training:
-  - `accounts/fireworks/models/qwen3-8b`
-  - `Qwen/Qwen3-8B`
-  - `accounts/fireworks/trainingShapes/qwen3-8b-128k`
-- LoRA can be tested with `--lora-rank N`, but the validated Qwen3 8B training
-  shape currently rejects LoRA mode on the `lorenss` preview account.
-- The first checkpoint sync happens after step 0 and subsequent rollouts sample
-  the updated weights through the same deployment.
-- `--keep-trainer` and `--keep-deployment` are available for debugging. By
-  default the trainer is cleaned up and the deployment scales to zero on exit.