hud-python 0.4.42__tar.gz → 0.4.43__tar.gz

{hud_python-0.4.42 → hud_python-0.4.43}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hud-python
-Version: 0.4.42
+Version: 0.4.43
 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-0.4.42 → hud_python-0.4.43}/hud/agents/openai_chat_generic.py

@@ -205,7 +205,7 @@ class GenericOpenAIChatAgent(MCPAgent):
         try:
             response = await self._invoke_chat_completion(
                 messages=messages,
-                tools=tools, # type: ignore
+                tools=tools,  # type: ignore
                 extra=extra,
             )
         except Exception as e:

{hud_python-0.4.42 → hud_python-0.4.43}/hud/cli/__init__.py

@@ -1178,6 +1178,11 @@ def rl(
         "--vllm-gpu",
         help="Specific GPU for vLLM server",
     ),
+    vllm_gpu_count: int = typer.Option(
+        1,
+        "--vllm-gpu-count",
+        help="Number of GPUs for vLLM server",
+    ),
     skip_vllm_startup: bool = typer.Option(
         False,
         "--skip-vllm-startup",
@@ -1199,6 +1204,7 @@ def rl(
         no_ddp=no_ddp,
         ddp_gpus=ddp_gpus,
         vllm_gpu=vllm_gpu,
+        vllm_gpu_count=vllm_gpu_count,
         yes=yes,
         skip_vllm_startup=skip_vllm_startup,
     )

{hud_python-0.4.42 → hud_python-0.4.43}/hud/cli/dev.py

@@ -73,6 +73,24 @@ def create_proxy_server(
         "PYTHONUNBUFFERED=1",  # Ensure Python output is not buffered
     ]
 
+    # Check for .env file in the project directory and add env vars
+    env_file = project_path / ".env"
+    loaded_env_vars = {}
+    if env_file.exists():
+        try:
+            from hud.cli.utils.config import parse_env_file
+
+            env_contents = env_file.read_text(encoding="utf-8")
+            loaded_env_vars = parse_env_file(env_contents)
+            for key, value in loaded_env_vars.items():
+                docker_cmd.extend(["-e", f"{key}={value}"])
+            if verbose and loaded_env_vars:
+                hud_console.info(
+                    f"Loaded {len(loaded_env_vars)} environment variable(s) from .env file"
+                )
+        except Exception as e:
+            hud_console.warning(f"Failed to load .env file: {e}")
+
     # Add user-provided Docker arguments
     if docker_args:
         docker_cmd.extend(docker_args)
@@ -112,8 +130,12 @@ def create_proxy_server(
             hud_console.info("The container's CMD determines reload behavior")
         hud_console.command_example(f"docker logs -f {container_name}", "View container logs")
 
-        # Show the full Docker command if there are environment variables
-        if docker_args and any(arg == "-e" or arg.startswith("--env") for arg in docker_args):
+        # Show the full Docker command if there are environment variables (from .env or args)
+        has_env_from_args = docker_args and any(
+            arg == "-e" or arg.startswith("--env") for arg in docker_args
+        )
+        has_env_from_file = bool(loaded_env_vars)
+        if has_env_from_args or has_env_from_file:
             hud_console.info("")
             hud_console.info("Docker command with environment variables:")
             hud_console.info(" ".join(docker_cmd))

{hud_python-0.4.42 → hud_python-0.4.43}/hud/cli/eval.py

@@ -298,16 +298,15 @@ async def run_single_task(
                 agent_config["allowed_tools"] = allowed_tools
 
         # Run with grouping
-        with hud.trace(name=f"{task_prompt} (group_size={group_size})"):
-            stats = await run_tasks_grouped(
-                tasks=[task],
-                agent_class=agent_class,
-                agent_config=agent_config,
-                group_size=group_size,
-                max_parallel_episodes=48,  # Same as RL default
-                max_steps=max_steps,
-                verbose=verbose,
-            )
+        stats = await run_tasks_grouped(
+            tasks=[task],
+            agent_class=agent_class,
+            agent_config=agent_config,
+            group_size=group_size,
+            max_parallel_episodes=48,  # Same as RL default
+            max_steps=max_steps,
+            verbose=verbose,
+        )
 
         # Display results
         display_group_statistics(stats, show_details=True)
@@ -499,7 +498,7 @@ async def run_full_dataset(
             )
 
         # Display results
-        display_group_statistics(stats, show_details=len(stats) <= 20)
+        display_group_statistics(stats, show_details=len(stats) <= 50)
 
         # Return stats for consistency with other modes
         return stats

{hud_python-0.4.42 → hud_python-0.4.43}/hud/cli/flows/tasks.py

@@ -212,17 +212,14 @@ def convert_tasks_to_remote(tasks_file: str) -> str:
     # Check if tasks already have remote URLs
     already_remote = _validate_tasks(tasks)
 
-    # If tasks already reference a remote MCP URL, do not require a local environment
-    # or attempt any image updates. Use the dataset as-is.
-    if already_remote:
-        return str(tasks_path)
-
     # Extract existing images from tasks
     existing_images = _extract_existing_images(tasks)
 
     # Locate environment
     env_dir = find_environment_dir(tasks_path)
     if not env_dir:
+        if already_remote:
+            return str(tasks_path)
         hud_console.error("Could not locate an environment directory (Dockerfile + pyproject.toml)")
         hud_console.hint("Ensure you're in or near your environment folder before running 'hud rl'")
         raise typer.Exit(1)
@@ -373,6 +370,8 @@ def convert_tasks_to_remote(tasks_file: str) -> str:
             item["system_prompt"] = t.system_prompt
         if t.metadata:
             item["metadata"] = t.metadata
+        if t.id is not None:
+            item["id"] = t.id
 
         tasks_payload.append(item)
 

{hud_python-0.4.42 → hud_python-0.4.43}/hud/cli/rl/__init__.py

@@ -78,6 +78,11 @@ def rl_command(
         "-y",
         help="Auto-accept all prompts and use defaults (lazy mode)",
     ),
+    vllm_gpu_count: int = typer.Option(
+        None,
+        "--vllm-gpu-count",
+        help="Number of GPUs for vLLM server",
+    ),
     skip_vllm_startup: bool = typer.Option(
         False,
         "--skip-vllm-startup",
@@ -145,6 +150,7 @@ def rl_command(
                 model=model,
                 config_file=config_file,
                 output_dir=output_dir,
+                vllm_gpu_count=vllm_gpu_count,
                 yes=yes,
             )
             return

{hud_python-0.4.42 → hud_python-0.4.43}/hud/cli/rl/config.py

@@ -84,7 +84,7 @@ def save_config(config: Config, path: Path) -> None:
     """Save configuration to a JSON file."""
     config_dict = config.to_dict()
 
-    with open(path, "w") as f:
+    with open(path, "w", encoding="utf-8") as f:
         json.dump(config_dict, f, indent=2)
         f.write("\n")  # Add newline at end of file
 
@@ -94,7 +94,7 @@ def save_config(config: Config, path: Path) -> None:
 
 def load_config(path: Path) -> Config:
     """Load configuration from a JSON file."""
-    with open(path) as f:
+    with open(path, encoding="utf-8") as f:
         data = json.load(f)
 
     # Use Config.from_dict which handles missing fields gracefully

{hud_python-0.4.42 → hud_python-0.4.43}/hud/cli/rl/gpu_utils.py

@@ -245,10 +245,12 @@ def adjust_config_for_ddp(config: Config, num_gpus: int) -> Config:
     # Apply scaling rule
     if num_gpus == 1:
         # Special case: 2 groups for single GPU
+        groups_per_gpu = 2
         config.training.batch_size = 2 * group_size
     else:
-        # Multi-GPU: each GPU processes 1 group
-        config.training.batch_size = num_gpus * group_size
+        groups_per_gpu = config.training.batch_size // group_size
+        # Multi-GPU: each GPU processes groups_per_gpu groups
+        config.training.batch_size = num_gpus * group_size * groups_per_gpu
 
     # Update max_parallel_episodes to match
     config.actor.max_parallel_episodes = config.training.batch_size
@@ -263,7 +265,7 @@ def adjust_config_for_ddp(config: Config, num_gpus: int) -> Config:
         f"\n[cyan]📊 Adjusted batch_size to {config.training.batch_size} ({config.training.batch_size // group_size} groups)[/cyan]"  # noqa: E501
     )
     console.print(
-        f"[cyan]   Each of the {num_gpus} GPU(s) will process {config.training.batch_size // group_size // num_gpus} group(s) in parallel[/cyan]"  # noqa: E501
+        f"[cyan]   Each of the {num_gpus} GPU(s) will process {groups_per_gpu} group(s) in parallel[/cyan]"  # noqa: E501
     )
 
     return config

{hud_python-0.4.42 → hud_python-0.4.43}/hud/cli/rl/remote_runner.py

@@ -32,7 +32,9 @@ GPU_PRICING = {
 }
 
 
-def ensure_vllm_deployed(model_name: str, gpu_type: str = "A100", timeout: int = 600) -> None:
+def ensure_vllm_deployed(
+    model_name: str, gpu_type: str = "A100", gpu_count: int = 1, timeout: int = 600
+) -> None:
     """Deploy vLLM for a model if needed and wait until it's ready.
 
     Args:
@@ -47,7 +49,7 @@ def ensure_vllm_deployed(model_name: str, gpu_type: str = "A100", timeout: int =
         return
 
     hud_console.info(f"Deploying vLLM server for {model_name}...")
-    rl_api.deploy_vllm(model_name, gpu_type=gpu_type)
+    rl_api.deploy_vllm(model_name, gpu_type=gpu_type, gpu_count=gpu_count)
     hud_console.success("vLLM deployment started")
 
     hud_console.info("Waiting for vLLM server to be ready...")
@@ -72,6 +74,7 @@ def run_remote_training(
     model: str | None,
     config_file: Path | None,
     output_dir: str,
+    vllm_gpu_count: int = 1,
     yes: bool = False,
 ) -> None:
     """Run RL training remotely via the API server following the new interactive flow."""
@@ -183,14 +186,18 @@ def run_remote_training(
 
             # Ask for model type
             if yes:
-                model_type = "Qwen/Qwen2.5-VL-3B-Instruct"  # Default model in yes mode
+                if config_file:
+                    config = load_config(config_file)
+                    model_type = config.model.base_model
+                else:
+                    model_type = "Qwen/Qwen2.5-VL-3B-Instruct"
                 hud_console.info(f"Auto-selecting base model: {model_type} (--yes mode)")
             else:
                 model_type = hud_console.select(
                     "Select base model type:",
                     choices=[
                         {"name": "Qwen2.5-VL-3B-Instruct", "value": "Qwen/Qwen2.5-VL-3B-Instruct"},
-                        # {"name": "Qwen2.5-VL-7B-Instruct", "value": "Qwen/Qwen2.5-VL-7B-Instruct"}, # noqa: E501
+                        {"name": "Qwen2.5-3B-Instruct", "value": "Qwen/Qwen2.5-3B-Instruct"},
                     ],
                     default=0,
                 )
@@ -218,7 +225,7 @@ def run_remote_training(
             try:
                 rl_api.create_model(model_name, model_type)
                 hud_console.success(f"Created model: {model_name}")
-                ensure_vllm_deployed(model_name, gpu_type="A100")
+                ensure_vllm_deployed(model_name, gpu_type="A100", gpu_count=vllm_gpu_count)
 
             except Exception as e:
                 # If the name already exists, suggest a new name and prompt once
@@ -247,7 +254,7 @@ def run_remote_training(
                         rl_api.create_model(chosen, model_type)
                         hud_console.success(f"Created model: {chosen}")
                         model_name = chosen
-                        ensure_vllm_deployed(model_name, gpu_type="A100")
+                        ensure_vllm_deployed(model_name, gpu_type="A100", gpu_count=vllm_gpu_count)
                     except Exception as e2:
                         hud_console.error(f"Failed to create model: {e2}")
                         raise
@@ -281,7 +288,7 @@ def run_remote_training(
                     return
 
             # Ensure vLLM is deployed
-            ensure_vllm_deployed(model_name, gpu_type="A100")
+            ensure_vllm_deployed(model_name, gpu_type="A100", gpu_count=vllm_gpu_count)
     except KeyboardInterrupt:
         hud_console.dim_info("Training cancelled", "")
         return
@@ -323,7 +330,7 @@ def run_remote_training(
             )
 
         if yes:
-            num_gpus = 2 # Default to 2 GPUs in yes mode
+            num_gpus = 2  # Default to 2 GPUs in yes mode
             hud_console.info(f"Auto-selecting {num_gpus} GPU(s) (--yes mode)")
         else:
             num_gpus = hud_console.select(
@@ -425,10 +432,12 @@ def run_remote_training(
         # Load provided config
         hud_console.info(f"Loading configuration from: {config_file}")
         config = load_config(config_file)
-        config_dict = config.to_dict()
         gpu_choice = config.training.gpu_type
         num_gpus = config.training.num_gpus
 
+        config = adjust_config_for_ddp(config, int(num_gpus))
+        config_dict = config.to_dict()
+
     # Launch training
     try:
         # Little celebration before launching

{hud_python-0.4.42 → hud_python-0.4.43}/hud/cli/rl/rl_api.py

@@ -61,12 +61,12 @@ def list_models() -> list[RLModelInfo]:
     ]
 
 
-def deploy_vllm(model_name: str, gpu_type: str = "A100") -> dict[str, Any]:
+def deploy_vllm(model_name: str, gpu_type: str = "A100", gpu_count: int = 1) -> dict[str, Any]:
     """Deploy a vLLM server for a model."""
     return make_request_sync(
         method="POST",
         url=f"{settings.hud_rl_url}/models/{model_name}/deploy",
-        json={"gpu_type": gpu_type},
+        json={"gpu_type": gpu_type, "gpu_count": gpu_count},
         api_key=settings.api_key,
     )
 

{hud_python-0.4.42 → hud_python-0.4.43}/hud/cli/utils/environment.py

@@ -127,8 +127,4 @@ def is_environment_directory(path: str | Path) -> bool:
         return False
 
     # Must have pyproject.toml
-    if not (dir_path / "pyproject.toml").exists():
-        hud_console.error("pyproject.toml not found")
-        return False
-
-    return True
+    return (dir_path / "pyproject.toml").exists()

{hud_python-0.4.42 → hud_python-0.4.43}/hud/rl/config.py

@@ -13,6 +13,7 @@ SUPPORTED_MODELS = [
     "Qwen/Qwen2.5-VL-32B-Instruct",
     "Qwen/Qwen2.5-VL-72B-Instruct",
     "Qwen/Qwen2.5-7B-Instruct",
+    "Qwen/Qwen2.5-3B-Instruct",
 ]
 
 
@@ -39,9 +40,9 @@ class ModelConfig:
     """Model and LoRA configuration."""
 
     base_model: str = "Qwen/Qwen2.5-VL-3B-Instruct"
-    lora_r: int = 8
-    lora_alpha: int = 16
-    lora_dropout: float = 0.05
+    lora_r: int = 16
+    lora_alpha: int = 32
+    lora_dropout: float = 0.1
     target_modules: tuple[str, ...] = (
         "q_proj",
         "k_proj",
@@ -61,6 +62,7 @@ class ModelConfig:
 @dataclass
 class TrainingConfig:
     """Training hyperparameters."""
+
     # GPU parameters
     gpu_type: str = "A100"
     num_gpus: int = 2
@@ -71,9 +73,9 @@ class TrainingConfig:
     save_every_batches: int = 1
 
     # Batching parameters
-    epochs: int = 2
-    batch_size: int = 24
-    group_size: int = 4
+    epochs: int = 1
+    batch_size: int = 16
+    group_size: int = 8
     mini_batch_size: int = 1
     update_after_group: bool = True  # Whether to update the policy after each task group
     accumulate_over_minibatches: bool = False  # Whether to accumulate over minibatches
@@ -84,7 +86,7 @@ class TrainingConfig:
     leave_one_out: bool = True
 
     # Replay buffer parameters
-    buffer_steps: int = 4
+    buffer_steps: int = 8
     select_strategy: Literal["recent", "variance", "random"] = "variance"
 
     # Aggregation parameters
@@ -92,8 +94,8 @@ class TrainingConfig:
     token_agg: Literal["mean", "sum"] = "mean"  # noqa: S105
 
     # Regularization parameters
-    kl_beta: float = 0.0
-    entropy_beta: float = 0.0
+    kl_beta: float = 0.001
+    entropy_beta: float = 0.001
     top_eps: float = 0.2
     bottom_eps: float = 0.1
 
@@ -143,6 +145,7 @@ class Config:
     job_id: str | None = None  # Use existing job ID if provided
     stats_interval: int = 1
     verbose: bool = False
+    very_verbose: bool = False
 
     # Paths
     out_dir: str = "./checkpoints"
@@ -166,6 +169,7 @@ class Config:
             job_id=d.get("job_id"),
             stats_interval=d.get("stats_interval", 1),
             verbose=d.get("verbose", False),
+            very_verbose=d.get("very_verbose", False),
             out_dir=d.get("out_dir", "./checkpoints"),
             adapter_prefix=d.get("adapter_prefix", "cua-grpo-step"),
             seed=d.get("seed", 1234),
@@ -181,6 +185,7 @@ class Config:
             "job_id": self.job_id,
             "stats_interval": self.stats_interval,
             "verbose": self.verbose,
+            "very_verbose": self.very_verbose,
             "out_dir": self.out_dir,
             "adapter_prefix": self.adapter_prefix,
             "seed": self.seed,

{hud_python-0.4.42 → hud_python-0.4.43}/hud/rl/train.py

@@ -56,6 +56,10 @@ async def train(config: Config, tasks: list[Task]) -> None:
         logging.basicConfig(level=logging.INFO)
         # Remove httpx logger
         logging.getLogger("httpx").setLevel(logging.WARNING)
+    if config.very_verbose:
+        logging.basicConfig(level=logging.DEBUG)
+        # Remove httpx logger
+        logging.getLogger("httpx").setLevel(logging.INFO)
 
     if is_main_process():
         hud_console.header("Starting GRPO Training")
@@ -103,10 +107,9 @@ async def train(config: Config, tasks: list[Task]) -> None:
     if is_main_process():
         hud_console.info(f"Creating job with config.job_id: {config.job_id}")
         job_obj = hud.create_job(
-            job_id=config.job_id, name=config.job_name, metadata={
-                "config": config.to_dict(),
-                "agent_class": config.model.base_model
-            }
+            job_id=config.job_id,
+            name=config.job_name,
+            metadata={"config": config.to_dict(), "agent_class": config.model.base_model},
         )
         hud_console.info(f"Created job with job_obj.id: {job_obj.id}")
         job_obj.update_status_sync("running")
@@ -299,7 +302,7 @@ async def main() -> None:
 
     # Load config
     if args.config:
-        with open(args.config) as f:  # noqa: ASYNC230
+        with open(args.config, encoding="utf-8") as f:  # noqa: ASYNC230
             config_dict = json.load(f)
         config = Config.from_dict(config_dict)
     else:
@@ -337,7 +340,7 @@ async def main() -> None:
     # Load tasks
     if args.tasks_json:
         # Tasks provided as JSON list via command line
-        tasks = load_tasks(args.tasks_jso)
+        tasks = load_tasks(args.tasks_json)
     elif args.tasks:
         # Tasks provided as file path or HuggingFace dataset
         tasks = load_tasks(args.tasks)

{hud_python-0.4.42 → hud_python-0.4.43}/hud/utils/group_eval.py

@@ -189,7 +189,7 @@ def display_group_statistics(stats: list[dict[str, Any]], show_details: bool = T
     hud_console.info(f"Overall mean reward: {overall_mean:.3f} ± {overall_std:.3f}")
 
     # Detailed table
-    if show_details and len(stats) <= 20:  # Only show for reasonable dataset sizes
+    if show_details and len(stats) <= 50:  # Only show for reasonable dataset sizes
         table = Table(title="\nPer-Task Performance Distribution")
         table.add_column("Task", style="cyan", no_wrap=True)
         table.add_column("Mean±Std", justify="right", style="green")
@@ -216,7 +216,7 @@ def display_group_statistics(stats: list[dict[str, Any]], show_details: bool = T
     # High variance tasks
     high_variance_tasks = [s for s in stats if s["std_reward"] > 0.3 and s["group_size"] > 1]
     if high_variance_tasks:
-        hud_console.warning(f"\n⚠️  {len(high_variance_tasks)} tasks show high variance (std > 0.3)")
+        hud_console.warning(f"\n{len(high_variance_tasks)} tasks show high variance (std > 0.3)")
         for task in high_variance_tasks[:3]:
             hud_console.info(
                 f"  • {task['task_id']}: μ={task['mean_reward']:.3f}, σ={task['std_reward']:.3f}"  # noqa: RUF001

{hud_python-0.4.42 → hud_python-0.4.43}/hud/utils/tasks.py

@@ -40,7 +40,7 @@ def load_tasks(tasks_input: str | list[dict], *, raw: bool = False) -> list[Task
         if Path(tasks_input).exists():
             file_path = Path(tasks_input)
 
-            with open(file_path) as f:
+            with open(file_path, encoding="utf-8") as f:
                 # Handle JSON files (array of tasks)
                 if file_path.suffix.lower() == ".json":
                     data = json.load(f)

{hud_python-0.4.42 → hud_python-0.4.43}/hud/utils/tests/test_version.py

@@ -5,4 +5,4 @@ def test_import():
     """Test that the package can be imported."""
     import hud
 
-    assert hud.__version__ == "0.4.42"
+    assert hud.__version__ == "0.4.43"

{hud_python-0.4.42 → hud_python-0.4.43}/hud/version.py

@@ -4,4 +4,4 @@ Version information for the HUD SDK.
 
 from __future__ import annotations
 
-__version__ = "0.4.42"
+__version__ = "0.4.43"

{hud_python-0.4.42 → hud_python-0.4.43}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "hud-python"
-version = "0.4.42"
+version = "0.4.43"
 description = "SDK for the HUD platform."
 readme = "README.md"
 requires-python = ">=3.11, <3.13"