hud-python 0.4.31__tar.gz → 0.4.32__tar.gz

{hud_python-0.4.31 → hud_python-0.4.32}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hud-python
-Version: 0.4.31
+Version: 0.4.32
 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-0.4.31 → hud_python-0.4.32}/hud/cli/flows/tasks.py

@@ -27,9 +27,27 @@ def _is_remote_url(url: str) -> bool:
 
 
 def _validate_tasks(tasks: list[Task]) -> bool:
-    """Validate the tasks file."""
+    """Validate the tasks file: return True if tasks already reference a remote MCP URL.
+
+    A task is considered remote if any "url" field anywhere inside mcp_config
+    is a valid remote URL (e.g., https://mcp.hud.so/v3/mcp).
+    """
+    def _has_remote_url(obj: Any) -> bool:
+        if isinstance(obj, dict):
+            for k, v in obj.items():
+                if k == "url" and isinstance(v, str) and _is_remote_url(v):
+                    return True
+                if _has_remote_url(v):
+                    return True
+        elif isinstance(obj, list):
+            for item in obj:
+                if _has_remote_url(item):
+                    return True
+        return False
+
     for task in tasks:
-        if not task.mcp_config or (not _is_remote_url(task.mcp_config.get("url", ""))):
+        cfg = task.mcp_config or {}
+        if not _has_remote_url(cfg):
             return False
     return True
 
@@ -100,7 +118,7 @@ def _ensure_pushed(env_dir: Path, lock_data: dict[str, Any]) -> dict[str, Any]:
         require_docker_running()
 
         # If Docker or login is not configured, the push function will fail and halt.
-        push_environment(str(env_dir))
+        push_environment(str(env_dir), yes=True)
 
         # Reload lock after push
         lock_path = env_dir / "hud.lock.yaml"
@@ -111,7 +129,21 @@ def _ensure_pushed(env_dir: Path, lock_data: dict[str, Any]) -> dict[str, Any]:
 
 
 def _derive_remote_image(lock_data: dict[str, Any]) -> str:
-    """Derive org/name:tag from lock file image field for MCP header."""
+    """Derive org/name:tag from lock file for MCP header.
+
+    Preference order:
+    1) lock_data["push"]["image_with_tag"] if present
+    2) Derive from lock_data["image"] (may be a digest; falls back to latest)
+    """
+    push_info = lock_data.get("push", {}) if isinstance(lock_data, dict) else {}
+
+    # 1) Exact image_with_tag if present
+    pushed_with_tag = str(push_info.get("image_with_tag", "")).strip()
+    if pushed_with_tag:
+        name, tag = extract_name_and_tag(pushed_with_tag)
+        return f"{name}:{tag}"
+
+    # Base name always comes from lock_data.image to preserve org/repo
     image_ref = str(lock_data.get("image", "")).strip()
     if not image_ref:
         raise typer.Exit("Lock file missing image reference")
@@ -157,19 +189,55 @@ def convert_tasks_to_remote(tasks_file: str) -> str:
     # Derive remote image name org/name:tag
     remote_image = _derive_remote_image(lock_data)
 
+    # Helper to strip extra fields from tool calls
+    def _simplify_tool_call(tool: Any) -> Any:
+        def _one(x: Any) -> dict[str, Any]:
+            try:
+                data = x.model_dump() if hasattr(x, "model_dump") else dict(x)
+            except Exception:
+                try:
+                    data = dict(x)
+                except Exception:
+                    return {}
+            # Keep only name and arguments
+            name = data.get("name")
+            arguments = data.get("arguments", {})
+            return {"name": name, "arguments": arguments}
+
+        if tool is None:
+            return None
+        if isinstance(tool, list):
+            return [_one(x) for x in tool]
+        return _one(tool)
+
     # Convert to list[dict]
     tasks_payload: list[dict[str, Any]] = []
     for t in tasks:
-        item = t.model_dump()
-        item["mcp_config"] = {
-            "hud": {
-                "url": "https://mcp.hud.so/v3/mcp",
-                "headers": {
-                    "Authorization": "Bearer ${HUD_API_KEY}",
-                    "Mcp-Image": remote_image,
-                },
-            }
+        item: dict[str, Any] = {
+            "prompt": t.prompt,
+            "mcp_config": {
+                "hud": {
+                    "url": "https://mcp.hud.so/v3/mcp",
+                    "headers": {
+                        "Authorization": "Bearer ${HUD_API_KEY}",
+                        "Mcp-Image": remote_image,
+                    },
+                }
+            },
         }
+
+        # Optional fields, omit Nones
+        if t.setup_tool is not None:
+            item["setup_tool"] = _simplify_tool_call(t.setup_tool)
+        if t.evaluate_tool is not None:
+            item["evaluate_tool"] = _simplify_tool_call(t.evaluate_tool)
+        if t.agent_tools is not None:
+            item["agent_tools"] = t.agent_tools
+        if t.system_prompt is not None:
+            item["system_prompt"] = t.system_prompt
+        if t.metadata:
+            item["metadata"] = t.metadata
+
         tasks_payload.append(item)
 
     # Write new file: remote_<name>.json (always JSON array)

{hud_python-0.4.31 → hud_python-0.4.32}/hud/cli/push.py

@@ -332,6 +332,7 @@ def push_environment(
         "source": local_image,
         "pushedAt": datetime.now(UTC).isoformat().replace("+00:00", "Z"),
         "registry": pushed_digest.split("/")[0] if "/" in pushed_digest else "docker.io",
+        "image_with_tag": image,
     }
 
     # Save updated lock file

{hud_python-0.4.31 → hud_python-0.4.32}/hud/cli/rl/remote_runner.py

@@ -10,6 +10,7 @@ import os
 import subprocess
 import time
 from pathlib import Path
+import uuid
 
 from rich.console import Console
 
@@ -29,6 +30,43 @@ GPU_PRICING = {
 }
 
 
+def ensure_vllm_deployed(model_name: str, gpu_type: str = "A100", timeout: int = 600) -> None:
+    """Deploy vLLM for a model if needed and wait until it's ready.
+
+    Args:
+        model_name: The name of the model to deploy vLLM for
+        gpu_type: GPU type to use for deployment (e.g., A100, H100)
+        timeout: Max seconds to wait for vLLM to be ready
+    """
+    # Check current model status
+    info = rl_api.get_model(model_name)
+    if info.vllm_url:
+        hud_console.success("vLLM server already running")
+        return
+
+    hud_console.info(f"Deploying vLLM server for {model_name}...")
+    rl_api.deploy_vllm(model_name, gpu_type=gpu_type)
+    hud_console.success("vLLM deployment started")
+
+    hud_console.info("Waiting for vLLM server to be ready...")
+    start_time = time.time()
+    with hud_console.progress() as progress:
+        progress.update(
+            "Checking deployment status (see live status on https://app.hud.so/models)"
+        )
+        while True:
+            if time.time() - start_time > timeout:
+                hud_console.error("Timeout waiting for vLLM deployment")
+                raise ValueError("vLLM deployment timeout")
+            info = rl_api.get_model(model_name)
+            if info.vllm_url or info.status == "ready":
+                hud_console.success(
+                    f"vLLM server ready at http://rl.hud.so/v1/models/{model_name}/vllm"
+                )
+                break
+            time.sleep(5)
+
+
 def run_remote_training(
     tasks_file: str | None,
     model: str | None,
@@ -128,49 +166,55 @@ def run_remote_training(
             from rich.prompt import Prompt
 
             # Ask for model name
-            default_name = model_type.split("/")[-1].lower()
+            base_default = model_type.split("/")[-1].lower()
+            default_name = base_default
+            existing_names = {m.name for m in active_models}
+            suffix = 1
+            while default_name in existing_names:
+                default_name = f"{base_default}-{suffix}"
+                suffix += 1
+
             hud_console.info(f"Enter model name (default: {default_name}):")
             model_name = Prompt.ask("Model name", default=default_name)
             model_name = model_name.replace("/", "-").lower()
 
-            # Create the model
+            # Create the model with retry on name conflict
             hud_console.info(f"Creating model: {model_name}")
             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")
 
-                # Deploy vLLM automatically
-                hud_console.info(f"Deploying vLLM server for {model_name}...")
-                rl_api.deploy_vllm(model_name, gpu_type="A100")
-                hud_console.success("vLLM deployment started")
-
-                # Wait for deployment
-                hud_console.info("Waiting for vLLM server to be ready...")
-                max_wait = 600  # 10 minutes
-                start_time = time.time()
-
-                with hud_console.progress() as progress:
-                    progress.update(
-                        "Checking deployment status (see live status on https://app.hud.so/models)"
-                    )
-
+            except Exception as e:
+                # If the name already exists, suggest a new name and prompt once
+                message = str(e)
+                if "already exists" in message or "409" in message:
+                    alt_name = f"{model_name}-1"
+                    i = 1
                     while True:
-                        if time.time() - start_time > max_wait:
-                            hud_console.error("Timeout waiting for vLLM deployment")
-                            raise ValueError("vLLM deployment timeout")
-
-                        model_info = rl_api.get_model(model_name)
-                        if model_info.status == "ready":
-                            hud_console.success(
-                                f"vLLM server ready at http://rl.hud.so/v1/models/{model_name}/vllm"
-                            )
+                        candidate = f"{model_name}-{str(uuid.uuid4())[:4]}"
+                        if candidate not in existing_names:
+                            alt_name = candidate
                             break
-
-                        time.sleep(5)
-
-            except Exception as e:
-                hud_console.error(f"Failed to create model: {e}")
-                raise
+                        i += 1
+                    hud_console.warning(
+                        f"Model '{model_name}' exists. Suggesting '{alt_name}' instead."
+                    )
+                    try:
+                        from rich.prompt import Prompt as _Prompt
+
+                        chosen = _Prompt.ask("Use different name", default=alt_name)
+                        chosen = chosen.replace("/", "-").lower()
+                        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")
+                    except Exception as e2:
+                        hud_console.error(f"Failed to create model: {e2}")
+                        raise
+                else:
+                    hud_console.error(f"Failed to create model: {e}")
+                    raise
 
         else:
             # Existing model selected
@@ -194,36 +238,7 @@ def run_remote_training(
                     return
 
             # Ensure vLLM is deployed
-            if not model_info.vllm_url:
-                hud_console.info(f"Deploying vLLM server for {model_name}...")
-                rl_api.deploy_vllm(model_name, gpu_type="A100")
-                hud_console.success("vLLM deployment started")
-
-                # Wait for deployment
-                hud_console.info("Waiting for vLLM server to be ready...")
-                max_wait = 600  # 10 minutes
-                start_time = time.time()
-
-                with hud_console.progress() as progress:
-                    progress.update(
-                        "Checking deployment status (see live status on https://app.hud.so/models)"
-                    )
-
-                    while True:
-                        if time.time() - start_time > max_wait:
-                            hud_console.error("Timeout waiting for vLLM deployment")
-                            raise ValueError("vLLM deployment timeout")
-
-                        model_info = rl_api.get_model(model_name)
-                        if model_info.vllm_url:
-                            hud_console.success(
-                                f"vLLM server ready at http://rl.hud.so/v1/models/{model_name}/vllm"
-                            )
-                            break
-
-                        time.sleep(5)
-            else:
-                hud_console.success("vLLM server already running")
+            ensure_vllm_deployed(model_name, gpu_type="A100")
     except KeyboardInterrupt:
         hud_console.dim_info("Training cancelled", "")
         return

{hud_python-0.4.31 → hud_python-0.4.32}/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__ == "0.4.31"
+    assert hud.__version__ == "0.4.32"

{hud_python-0.4.31 → hud_python-0.4.32}/hud/version.py

@@ -4,4 +4,4 @@ Version information for the HUD SDK.
 
 from __future__ import annotations
 
-__version__ = "0.4.31"
+__version__ = "0.4.32"

{hud_python-0.4.31 → hud_python-0.4.32}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "hud-python"
-version = "0.4.31"
+version = "0.4.32"
 description = "SDK for the HUD platform."
 readme = "README.md"
 requires-python = ">=3.11, <3.13"