hud-python 0.4.45__tar.gz → 0.4.47__tar.gz

{hud_python-0.4.45 → hud_python-0.4.47}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hud-python
-Version: 0.4.45
+Version: 0.4.47
 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-0.4.45 → hud_python-0.4.47}/hud/rl/config.py

@@ -57,6 +57,7 @@ class ModelConfig:
     attn_implementation: str = "flash_attention_2"
     use_liger: bool = True
     gradient_checkpointing: bool = True
+    adapter_path: str | None = None  # Path to existing LoRA adapter to load as baseline
 
 
 @dataclass

{hud_python-0.4.45 → hud_python-0.4.47}/hud/rl/distributed.py

@@ -81,7 +81,7 @@ def broadcast_object(obj: Any, src: int = 0) -> Any:
         return obj
 
     obj_list = [obj] if dist.get_rank() == src else [None]
-    dist.broadcast_object_list(obj_list, src=src, device=torch.device("cpu"))
+    dist.broadcast_object_list(obj_list, src=src) #, device=torch.device("cpu"))
     return obj_list[0]
 
 

{hud_python-0.4.45 → hud_python-0.4.47}/hud/rl/learner.py

@@ -146,17 +146,27 @@ class GRPOLearner:
             policy.gradient_checkpointing_enable()
             self.log("Gradient checkpointing enabled for memory efficiency")
 
-        # Add LoRA adapters
-        lora_config = LoraConfig(
-            r=model_cfg.lora_r,
-            lora_alpha=model_cfg.lora_alpha,
-            lora_dropout=model_cfg.lora_dropout,
-            task_type="CAUSAL_LM",
-            bias="none",
-            target_modules=list(model_cfg.target_modules),
-        )
+        # Add LoRA adapters or load existing adapter
         policy.config.use_cache = False
-        policy = get_peft_model(policy, lora_config)
+        
+        if model_cfg.adapter_path:
+            # Load existing adapter as baseline
+            self.log(f"Loading existing LoRA adapter from: {model_cfg.adapter_path}")
+            from peft import PeftModel
+            policy = PeftModel.from_pretrained(policy, model_cfg.adapter_path)
+            # Enable adapter training
+            policy.train()
+        else:
+            # Create new LoRA adapter
+            lora_config = LoraConfig(
+                r=model_cfg.lora_r,
+                lora_alpha=model_cfg.lora_alpha,
+                lora_dropout=model_cfg.lora_dropout,
+                task_type="CAUSAL_LM",
+                bias="none",
+                target_modules=list(model_cfg.target_modules),
+            )
+            policy = get_peft_model(policy, lora_config)
 
         # Wrap with DDP if in distributed mode
         if self.world_size > 1:

{hud_python-0.4.45 → hud_python-0.4.47}/hud/rl/train.py

@@ -95,6 +95,18 @@ async def train(config: Config, tasks: list[Task]) -> None:
         if is_main_process()
         else None
     )
+    
+    # Load initial adapter if provided
+    if is_main_process() and config.model.adapter_path and vllm:
+        hud_console.info(f"Loading baseline adapter from: {config.model.adapter_path}")
+        success = vllm.load_adapter(config.model.base_model, config.model.adapter_path)
+        if success and actor is not None:
+            hud_console.info("Successfully loaded baseline adapter as 'base_model'")
+            # Update actor to use the loaded adapter
+            actor.update_adapter(config.model.base_model)
+        else:
+            hud_console.error("Failed to load baseline adapter")
+            exit(1)
 
     # Training state
     step = 0

{hud_python-0.4.45 → hud_python-0.4.47}/hud/tools/base.py

@@ -1,7 +1,7 @@
 from __future__ import annotations
 
 from abc import ABC, abstractmethod
-from typing import TYPE_CHECKING, Any, cast
+from typing import TYPE_CHECKING, Any, cast, Awaitable
 
 from fastmcp import FastMCP
 
@@ -16,6 +16,8 @@ if TYPE_CHECKING:
 # Basic result types for tools
 BaseResult = list[ContentBlock] | EvaluationResult
 
+import logging
+logger = logging.getLogger(__name__)
 
 class BaseTool(ABC):
     """
@@ -58,6 +60,10 @@ class BaseTool(ABC):
         self.title = title or self.__class__.__name__.replace("Tool", "").replace("_", " ").title()
         self.description = description or (self.__doc__.strip() if self.__doc__ else None)
         self.meta = meta
+        self._callbacks: dict[
+            str,
+            list[Callable[..., Awaitable[Any]]],
+        ] = {}  # {"event_name": [callback_functions]}
 
         # Expose attributes FastMCP expects when registering an instance directly
         self.__name__ = self.name  # FastMCP uses fn.__name__ if name param omitted
@@ -100,6 +106,36 @@ class BaseTool(ABC):
             )
         return self._mcp_tool
 
+    def add_callback(self, event_type: str, callback: Callable[..., Awaitable[Any]]):
+        """Register a callback function for specific event
+        
+        Args:
+            event_type: (Required) Specific event name to trigger callback
+                        e.g. "after_click", "before_navigate"
+            callback: (Required) Async function to call. Must be defined by `async def f(...)`
+        """
+        if event_type not in self._callbacks:
+            self._callbacks[event_type] = []
+        self._callbacks[event_type].append(callback)
+
+    def remove_callback(self, event_type: str, callback: Callable[..., Awaitable[Any]]):
+        """Remove a registered callback
+        Args:
+            event_type: (Required) Specific event name to trigger callback
+                        e.g. "after_click", "before_navigate"
+            callback: (Required) Function to remove from callback list.
+        """
+        if (event_type in self._callbacks) and (callback in self._callbacks[event_type]):
+            self._callbacks[event_type].remove(callback)
+    
+    async def _trigger_callbacks(self, event_type: str, **kwargs):
+        """Trigger all registered callback functions of an event type"""
+        callback_list = self._callbacks.get(event_type, [])
+        for callback in callback_list:
+            try:
+                await callback(**kwargs)
+            except Exception as e:
+                logger.warning(f"Callback failed for {event_type}: {e}")
 
 # Prefix for internal tool names
 _INTERNAL_PREFIX = "int_"

{hud_python-0.4.45 → hud_python-0.4.47}/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__ == "0.4.45"
+    assert hud.__version__ == "0.4.47"

{hud_python-0.4.45 → hud_python-0.4.47}/hud/version.py

@@ -4,4 +4,4 @@ Version information for the HUD SDK.
 
 from __future__ import annotations
 
-__version__ = "0.4.45"
+__version__ = "0.4.47"

{hud_python-0.4.45 → hud_python-0.4.47}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "hud-python"
-version = "0.4.45"
+version = "0.4.47"
 description = "SDK for the HUD platform."
 readme = "README.md"
 requires-python = ">=3.11, <3.13"