synth-ai 0.2.8.dev2__py3-none-any.whl → 0.4.3__py3-none-any.whl

synth_ai/sdk/task/in_process_runner.py

@@ -0,0 +1,314 @@
+"""SDK helper for running prompt-learning and RL jobs against a tunneled task app.
+
+This module keeps everything in-process:
+1) Spins up a FastAPI task app via InProcessTaskApp
+2) Opens a tunnel (Cloudflare by default, or uses preconfigured URL)
+3) Applies dot-notation overrides (task_app_url, budgets, seeds, models)
+4) Submits jobs to the remote backend using SDK clients
+5) Optionally polls until completion and returns a structured result
+
+Tunnel Modes:
+- "quick" (default): Creates Cloudflare quick tunnel - works for local development
+- "named": Uses Cloudflare managed tunnel (requires setup)
+- "local": No tunnel, uses localhost URL directly
+- "preconfigured": Uses externally-provided URL - for container environments
+  (ngrok, etc.) where Cloudflare tunnels don't work
+
+Environment Variables:
+- SYNTH_TASK_APP_URL: If set, auto-enables preconfigured mode with this URL
+- SYNTH_TUNNEL_MODE: Override tunnel mode (e.g., "preconfigured", "local")
+"""
+
+from __future__ import annotations
+
+import asyncio
+import os
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, Callable, Dict, Literal, Mapping, MutableMapping
+
+from synth_ai.core.env import get_backend_from_env
+from synth_ai.core.telemetry import log_info
+from synth_ai.sdk.api.train.prompt_learning import PromptLearningJob
+from synth_ai.sdk.api.train.rl import RLJob
+from synth_ai.sdk.api.train.local_api import LocalAPIHealth, check_local_api_health
+from synth_ai.sdk.api.train.utils import ensure_api_base
+from synth_ai.sdk.task.in_process import InProcessTaskApp
+
+try:  # CLI deep-update utility (dot-notation aware)
+    from synth_ai.cli.local.experiment_queue.config_utils import _deep_update
+except Exception as exc:  # pragma: no cover - defensive
+    raise RuntimeError("Failed to import config override helper (_deep_update)") from exc
+
+
+BackendMode = Literal["prompt_learning", "rl"]
+
+
+@dataclass
+class InProcessJobResult:
+    """Result of an in-process job submission."""
+
+    job_id: str
+    status: Dict[str, Any]
+    task_app_url: str
+    backend_url: str
+    task_app_health: LocalAPIHealth | None = None
+
+
+def _normalize_base_url(url: str) -> str:
+    """Strip trailing slashes and /api suffix for consistent handling."""
+    base = url.strip().rstrip("/")
+    if base.endswith("/api"):
+        base = base[: -len("/api")]
+    return base
+
+
+def resolve_backend_api_base(override: str | None = None) -> str:
+    """Resolve backend base URL using the documented priority order.
+
+    Priority:
+    1. Explicit override argument
+    2. TARGET_BACKEND_BASE_URL
+    3. BACKEND_OVERRIDE
+    4. SYNTH_BACKEND_URL
+    5. BACKEND_BASE_URL
+    6. NEXT_PUBLIC_API_URL
+    7. Fallback to core get_backend_from_env()
+    """
+
+    env_order = [
+        "TARGET_BACKEND_BASE_URL",
+        "BACKEND_OVERRIDE",
+        "SYNTH_BACKEND_URL",
+        "BACKEND_BASE_URL",
+        "NEXT_PUBLIC_API_URL",
+    ]
+
+    if override and override.strip():
+        candidate = override.strip()
+    else:
+        candidate = ""
+        for key in env_order:
+            value = os.environ.get(key, "").strip()
+            if value:
+                candidate = value
+                break
+        if not candidate:
+            base, _ = get_backend_from_env()
+            candidate = base
+
+    normalized = _normalize_base_url(candidate)
+    return ensure_api_base(normalized)
+
+
+def _require_env(key: str, *, friendly_name: str | None = None) -> str:
+    value = os.environ.get(key, "").strip()
+    if not value:
+        label = friendly_name or key
+        raise ValueError(
+            f"{label} is required. Set {key} in your environment or pass it explicitly."
+        )
+    return value
+
+
+def merge_dot_overrides(
+    base: Mapping[str, Any] | None, extra: Mapping[str, Any] | None
+) -> Dict[str, Any]:
+    """Merge two override dictionaries with dot-notation support."""
+    merged: MutableMapping[str, Any] = {}
+    if base:
+        _deep_update(merged, dict(base))
+    if extra:
+        _deep_update(merged, dict(extra))
+    return dict(merged)
+
+
+async def run_in_process_job(
+    *,
+    job_type: BackendMode,
+    config_path: str | Path,
+    backend_url: str | None = None,
+    api_key: str | None = None,
+    task_app_api_key: str | None = None,
+    allow_experimental: bool | None = None,
+    overrides: Mapping[str, Any] | None = None,
+    poll: bool = True,
+    poll_interval: float = 5.0,
+    timeout: float = 3600.0,
+    on_status: Callable[[Dict[str, Any]], None] | None = None,
+    # InProcessTaskApp args (exactly one of app/config/config_factory/task_app_path required)
+    app: Any | None = None,
+    config: Any | None = None,
+    config_factory: Callable[[], Any] | None = None,
+    task_app_path: str | Path | None = None,
+    tunnel_mode: str = "quick",
+    preconfigured_url: str | None = None,
+    preconfigured_auth_header: str | None = None,
+    preconfigured_auth_token: str | None = None,
+    skip_tunnel_verification: bool = True,  # Default True - verification is unreliable
+    force_new_tunnel: bool = True,  # Default True - always get fresh tunnel
+    host: str = "127.0.0.1",
+    port: int = 8114,
+    auto_find_port: bool = True,
+    health_check_timeout: float = 30.0,
+) -> InProcessJobResult:
+    """Run a prompt-learning or RL job with a tunneled task app."""
+    ctx: Dict[str, Any] = {
+        "job_type": job_type,
+        "config_path": str(config_path),
+        "poll": poll,
+        "tunnel_mode": tunnel_mode,
+    }
+    log_info("run_in_process_job invoked", ctx=ctx)
+    """Run a prompt-learning or RL job with a tunneled task app.
+    
+    Args:
+        job_type: Type of job - "prompt_learning" or "rl"
+        config_path: Path to the TOML config file
+        backend_url: Optional backend URL override
+        api_key: Synth API key for backend auth
+        task_app_api_key: API key for task app auth
+        allow_experimental: Allow experimental features
+        overrides: Config overrides (dot-notation supported)
+        poll: Whether to poll for completion
+        poll_interval: Polling interval in seconds
+        timeout: Maximum time to wait for job completion
+        on_status: Callback for status updates
+        app: FastAPI app instance
+        config: TaskAppConfig object
+        config_factory: Callable that returns TaskAppConfig
+        task_app_path: Path to task app .py file
+        tunnel_mode: Tunnel mode - "quick", "named", "local", or "preconfigured"
+        preconfigured_url: External tunnel URL when tunnel_mode="preconfigured"
+        preconfigured_auth_header: Auth header name for preconfigured URL
+        preconfigured_auth_token: Auth token for preconfigured URL
+        skip_tunnel_verification: Skip HTTP verification of tunnel
+        host: Local host to bind to
+        port: Local port to bind to
+        auto_find_port: Auto-find available port if busy
+        health_check_timeout: Max time to wait for health check
+        
+    Returns:
+        InProcessJobResult with job_id, status, and URLs
+    """
+    backend_api_base = resolve_backend_api_base(backend_url)
+    
+    # Set SYNTH_BACKEND_URL so that tunnel operations (like rotate_tunnel) use the correct backend
+    os.environ["SYNTH_BACKEND_URL"] = backend_api_base
+    
+    resolved_api_key = api_key or _require_env("SYNTH_API_KEY", friendly_name="Backend API key")
+    resolved_task_app_key = task_app_api_key or _require_env(
+        "ENVIRONMENT_API_KEY", friendly_name="Task app API key"
+    )
+
+    config_path = Path(config_path)
+    if not config_path.exists():
+        raise FileNotFoundError(f"Config file not found: {config_path}")
+
+    # Launch the task app with tunnel (Cloudflare by default, or preconfigured URL)
+    async with InProcessTaskApp(
+        app=app,
+        config=config,
+        config_factory=config_factory,
+        task_app_path=task_app_path,
+        host=host,
+        port=port,
+        tunnel_mode=tunnel_mode,
+        preconfigured_url=preconfigured_url,
+        preconfigured_auth_header=preconfigured_auth_header,
+        preconfigured_auth_token=preconfigured_auth_token,
+        skip_tunnel_verification=skip_tunnel_verification,
+        force_new_tunnel=force_new_tunnel,
+        api_key=resolved_task_app_key,
+        auto_find_port=auto_find_port,
+        health_check_timeout=health_check_timeout,
+    ) as task_app:
+        task_url = task_app.url or f"http://{host}:{task_app.port}"
+
+        # Check if backend verified DNS propagation (so we can skip local health checks)
+        dns_verified_by_backend = getattr(task_app, "_dns_verified_by_backend", False)
+
+        # Skip health check when:
+        # 1. tunnel verification is skipped (DNS may not have propagated locally)
+        # 2. OR backend verified DNS propagation (safe to skip redundant local check)
+        should_skip_health_check = skip_tunnel_verification or dns_verified_by_backend
+        if should_skip_health_check:
+            reason = "tunnel verification disabled" if skip_tunnel_verification else "backend verified DNS"
+            health = LocalAPIHealth(
+                ok=True,
+                health_status=200,
+                task_info_status=200,
+                detail=f"Skipped ({reason})",
+            )
+        else:
+            health = check_local_api_health(task_url, resolved_task_app_key)
+            if not health.ok:
+                raise RuntimeError(f"Task app health check failed for {task_url}: {health.detail}")
+
+        # Common overrides: task URL + API key injected in both dot and flat forms
+        task_overrides = {
+            "task_url": task_url,
+            "task_app_api_key": resolved_task_app_key,
+            "prompt_learning.task_app_url": task_url,
+            "prompt_learning.task_app_api_key": resolved_task_app_key,
+        }
+        merged_overrides = merge_dot_overrides(overrides, task_overrides)
+
+        if job_type == "prompt_learning":
+            job = PromptLearningJob.from_config(
+                config_path=config_path,
+                backend_url=backend_api_base,
+                api_key=resolved_api_key,
+                task_app_api_key=resolved_task_app_key,
+                allow_experimental=allow_experimental,
+                overrides=merged_overrides,
+            )
+        elif job_type == "rl":
+            job = RLJob.from_config(
+                config_path=config_path,
+                backend_url=backend_api_base,
+                api_key=resolved_api_key,
+                task_app_url=task_url,
+                task_app_api_key=resolved_task_app_key,
+                allow_experimental=allow_experimental,
+                overrides=merged_overrides,
+            )
+        else:
+            raise ValueError(f"Unknown job_type: {job_type}")
+
+        # Skip job's health check if we already skipped above (DNS verified by backend or tunnel verification disabled)
+        if should_skip_health_check:
+            job._skip_health_check = True
+
+        job_id = job.submit()
+        if not poll:
+            status = {"status": "submitted", "job_id": job_id}
+        else:
+            status = job.poll_until_complete(
+                timeout=timeout,
+                interval=poll_interval,
+                on_status=on_status,
+            )
+
+    return InProcessJobResult(
+        job_id=job_id,
+        status=status,
+        task_app_url=task_url,
+        backend_url=backend_api_base,
+        task_app_health=health,
+    )
+
+
+def run_in_process_job_sync(**kwargs: Any) -> InProcessJobResult:
+    """Synchronous wrapper for run_in_process_job()."""
+
+    return asyncio.run(run_in_process_job(**kwargs))
+
+
+__all__ = [
+    "InProcessJobResult",
+    "merge_dot_overrides",
+    "resolve_backend_api_base",
+    "run_in_process_job",
+    "run_in_process_job_sync",
+]

synth_ai/sdk/task/inference_api.py

@@ -0,0 +1,299 @@
+from __future__ import annotations
+
+import json
+import os
+import time
+from typing import Any, Dict, List, Optional, Sequence
+
+
+class InferenceAPIClient:
+    """Async client that normalizes chat completion calls across providers."""
+
+    def __init__(self, *, provider: str, inference_url: Optional[str] = None, timeout: float = 30.0) -> None:
+        self.provider = (provider or "").lower()
+        self.inference_url = inference_url.rstrip("/") if inference_url else None
+        self.timeout = timeout
+
+        if self.provider == "openai":
+            try:
+                from openai import AsyncOpenAI
+            except Exception as exc:  # pragma: no cover - import guard
+                raise RuntimeError("openai package not installed") from exc
+
+            api_key = os.getenv("OPENAI_API_KEY")
+            if not api_key:
+                raise RuntimeError("OPENAI_API_KEY must be set for meta-model proposals")
+
+            kwargs: Dict[str, Any] = {"api_key": api_key}
+            if self.inference_url:
+                kwargs["base_url"] = self.inference_url
+            self._client = AsyncOpenAI(**kwargs)
+
+        elif self.provider == "groq":
+            try:
+                from groq import AsyncGroq
+            except Exception as exc:  # pragma: no cover - import guard
+                raise RuntimeError("groq package not installed") from exc
+
+            api_key = os.getenv("GROQ_API_KEY")
+            if not api_key:
+                raise RuntimeError("GROQ_API_KEY must be set for meta-model proposals")
+
+            kwargs: Dict[str, Any] = {"api_key": api_key}
+            if self.inference_url:
+                kwargs["base_url"] = self.inference_url
+            self._client = AsyncGroq(**kwargs)
+
+        elif self.provider == "google":
+            try:
+                import google.genai as genai
+            except Exception as exc:  # pragma: no cover - import guard
+                raise RuntimeError("google-genai package not installed") from exc
+
+            api_key = os.getenv("GEMINI_API_KEY")
+            if not api_key:
+                raise RuntimeError("GEMINI_API_KEY must be set for Google/Gemini models")
+
+            # Set API key for SDK
+            os.environ["GEMINI_API_KEY"] = api_key
+            # Google GenAI SDK doesn't use base_url like OpenAI/Groq
+            self._client = genai.Client()
+
+        else:
+            raise RuntimeError(f"Unsupported meta-model provider: {provider}")
+
+    async def chat_completion(
+        self,
+        *,
+        messages: Sequence[Dict[str, Any]],
+        model: str,
+        temperature: float,
+        max_tokens: int,
+        tools: Sequence[Dict[str, Any]] | None = None,
+        tool_choice: str | Dict[str, Any] | None = None,
+        response_format: Dict[str, Any] | None = None,
+    ) -> Dict[str, Any]:
+        if self.provider == "openai":
+            kwargs: Dict[str, Any] = {
+                "model": model,
+                "messages": self._normalize_messages(messages),
+                "temperature": temperature,
+                "max_tokens": max_tokens,
+                "timeout": self.timeout,
+            }
+            if tools:
+                kwargs["tools"] = list(tools)
+            if tool_choice is not None:
+                kwargs["tool_choice"] = tool_choice
+            if response_format:
+                kwargs["response_format"] = response_format
+            response = await self._client.chat.completions.create(**kwargs)  # type: ignore[attr-defined]
+            return self._model_dump(response)
+
+        if self.provider == "groq":
+            groq_kwargs: Dict[str, Any] = {
+                "model": model,
+                "messages": self._normalize_messages(messages),
+                "temperature": temperature,
+                "max_tokens": max_tokens,
+                "timeout": self.timeout,
+            }
+            if tools:
+                groq_kwargs["tools"] = list(tools)
+            if tool_choice is not None:
+                groq_kwargs["tool_choice"] = tool_choice
+            if response_format:
+                # Groq response_format normalization (json_schema -> json_object for older models)
+                from .proxy import normalize_response_format_for_groq
+                normalized_format = dict(response_format)
+                normalize_response_format_for_groq(model, {"response_format": normalized_format})
+                groq_kwargs["response_format"] = normalized_format
+            response = await self._client.chat.completions.create(**groq_kwargs)  # type: ignore[attr-defined]
+            return self._model_dump(response)
+
+        if self.provider == "google":
+            return await self._google_chat_completion(
+                messages=messages,
+                model=model,
+                temperature=temperature,
+                max_tokens=max_tokens,
+                tools=tools,
+                tool_choice=tool_choice,
+                response_format=response_format,
+            )
+
+        raise RuntimeError(f"Unsupported meta-model provider: {self.provider}")
+
+    async def _google_chat_completion(
+        self,
+        *,
+        messages: Sequence[Dict[str, Any]],
+        model: str,
+        temperature: float,
+        max_tokens: int,
+        tools: Sequence[Dict[str, Any]] | None = None,
+        tool_choice: str | Dict[str, Any] | None = None,
+        response_format: Dict[str, Any] | None = None,
+    ) -> Dict[str, Any]:
+        """Handle Google Gemini chat completion."""
+        from google.genai import types
+
+        client = self._client
+
+        # Convert messages format (OpenAI -> Gemini)
+        contents = []
+        system_instruction = None
+        for msg in messages:
+            role = msg.get("role", "").lower()
+            content = msg.get("content", "")
+            if role == "system":
+                system_instruction = content
+            elif role in ["user", "assistant"]:
+                # Gemini uses "user" and "model" roles (not "assistant")
+                gemini_role = "user" if role == "user" else "model"
+                contents.append(types.Content(role=gemini_role, parts=[types.Part.from_text(text=str(content))]))
+
+        # Convert tools format (OpenAI -> Gemini)
+        gemini_tools = None
+        tool_config = None
+        if tools:
+            function_declarations = []
+            for tool in tools:
+                if isinstance(tool, dict) and tool.get("type") == "function":
+                    func = tool.get("function", {})
+                    func_decl = types.FunctionDeclaration(
+                        name=func.get("name", ""),
+                        description=func.get("description", ""),
+                        parameters=func.get("parameters", {}),
+                    )
+                    function_declarations.append(func_decl)
+
+            if function_declarations:
+                gemini_tools = [types.Tool(function_declarations=function_declarations)]
+
+                # Handle tool_choice - Gemini uses tool_config
+                if tool_choice and isinstance(tool_choice, dict) and tool_choice.get("type") == "function":
+                    func_name = tool_choice.get("function", {}).get("name")
+                    if func_name:
+                        mode_enum = types.FunctionCallingConfigMode.ANY
+                        tool_config = types.ToolConfig(
+                            function_calling_config=types.FunctionCallingConfig(
+                                mode=mode_enum,
+                                allowed_function_names=[func_name],
+                            )
+                        )
+
+        # Build generation config
+        cfg_kwargs: Dict[str, Any] = {"temperature": temperature}
+        if system_instruction:
+            cfg_kwargs["system_instruction"] = system_instruction
+        if max_tokens:
+            cfg_kwargs["max_output_tokens"] = max_tokens
+
+        # Handle response_format -> Gemini responseSchema conversion
+        # OpenAI: {"type": "json_schema", "json_schema": {"name": "...", "schema": {...}}}
+        # Gemini: response_mime_type="application/json", response_schema={...}
+        if response_format and isinstance(response_format, dict):
+            format_type = response_format.get("type")
+            if format_type == "json_schema":
+                json_schema_obj = response_format.get("json_schema", {})
+                schema = json_schema_obj.get("schema", {})
+                if schema:
+                    cfg_kwargs["response_mime_type"] = "application/json"
+                    cfg_kwargs["response_schema"] = schema
+            elif format_type == "json_object":
+                # json_object mode - just request JSON without schema
+                cfg_kwargs["response_mime_type"] = "application/json"
+
+        # Add tool config if present
+        if tool_config:
+            cfg_kwargs["tool_config"] = tool_config
+        if gemini_tools:
+            cfg_kwargs["tools"] = gemini_tools
+
+        generation_config = types.GenerateContentConfig(**cfg_kwargs)
+
+        # Make the API call
+        resp = await client.aio.models.generate_content(  # type: ignore[attr-defined]
+            model=model,
+            contents=contents,
+            config=generation_config,
+        )
+
+        # Convert response format (Gemini -> OpenAI)
+        # Extract text
+        try:
+            text = resp.text
+            if text is None:
+                text = ""
+        except (ValueError, AttributeError):
+            text = ""
+
+        # Extract tool calls (if any)
+        tool_calls = []
+        if resp.candidates and resp.candidates[0].content:
+            for part in resp.candidates[0].content.parts:
+                if part.function_call:
+                    tool_calls.append({
+                        "id": f"call_{len(tool_calls) + 1}",
+                        "type": "function",
+                        "function": {
+                            "name": part.function_call.name,
+                            "arguments": json.dumps(dict(part.function_call.args)),
+                        },
+                    })
+
+        # Build OpenAI-compatible response
+        message = {"role": "assistant", "content": text or ""}
+        if tool_calls:
+            message["tool_calls"] = tool_calls
+
+        # Estimate token usage (Gemini doesn't always provide this in the same format)
+        # Simple approximation: 1 token per 4 characters
+        prompt_text = "\n".join(str(m.get("content", "")) for m in messages)
+        prompt_tokens = max(1, len(prompt_text) // 4)
+        completion_tokens = max(1, len(text or "") // 4)
+
+        response_data = {
+            "id": f"chatcmpl-{int(time.time())}",
+            "object": "chat.completion",
+            "created": int(time.time()),
+            "model": model,
+            "choices": [{
+                "index": 0,
+                "message": message,
+                "finish_reason": "tool_calls" if tool_calls else "stop",
+            }],
+            "usage": {
+                "prompt_tokens": prompt_tokens,
+                "completion_tokens": completion_tokens,
+                "total_tokens": prompt_tokens + completion_tokens,
+            },
+        }
+
+        return response_data
+
+    @staticmethod
+    def _normalize_messages(messages: Sequence[Dict[str, Any]]) -> List[Dict[str, Any]]:
+        return [dict(message) for message in messages]
+
+    @staticmethod
+    def _model_dump(obj: Any) -> Dict[str, Any]:
+        if hasattr(obj, "model_dump"):
+            data = obj.model_dump()
+        elif hasattr(obj, "dict"):
+            data = obj.dict()
+        else:
+            data = dict(obj)  # type: ignore[arg-type]
+
+        usage = getattr(obj, "usage", None)
+        if usage and "usage" not in data:
+            if hasattr(usage, "model_dump"):
+                data["usage"] = usage.model_dump()
+            elif isinstance(usage, dict):
+                data["usage"] = usage
+        return data
+
+
+__all__ = ["InferenceAPIClient"]
+