synth-ai 0.2.4.dev7__py3-none-any.whl → 0.2.4.dev8__py3-none-any.whl

synth_ai/learning/config.py

@@ -0,0 +1,43 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Any, Dict, Optional
+
+
+@dataclass
+class FTJobConfig:
+    model: str
+    training_file_id: str
+    n_epochs: int = 1
+    batch_size: int = 1
+    upload_to_wasabi: bool = True
+
+    def hyperparameters(self) -> Dict[str, Any]:
+        if self.n_epochs < 1:
+            raise ValueError("n_epochs must be >= 1")
+        if self.batch_size < 1:
+            raise ValueError("batch_size must be >= 1")
+        return {"n_epochs": int(self.n_epochs), "batch_size": int(self.batch_size)}
+
+    def metadata(self) -> Dict[str, Any]:  # type: ignore[override]
+        return {"upload_to_wasabi": bool(self.upload_to_wasabi)}
+
+
+@dataclass
+class RLJobConfig:
+    model: str
+    task_app_url: str
+    trainer_id: str
+    batch_size: int = 1
+    group_size: int = 2
+    job_config_id: Optional[str] = None
+    inline_config: Optional[Dict[str, Any]] = None
+
+    def trainer_dict(self) -> Dict[str, Any]:
+        if self.batch_size < 1:
+            raise ValueError("batch_size must be >= 1")
+        if self.group_size < 2:
+            raise ValueError("group_size must be >= 2")
+        return {"batch_size": int(self.batch_size), "group_size": int(self.group_size)}
+
+

synth_ai/learning/constants.py

@@ -0,0 +1,29 @@
+from __future__ import annotations
+
+# Terminal statuses normalized across FT and RL
+TERMINAL_STATUSES = {
+    "succeeded",
+    "failed",
+    "cancelled",
+    "canceled",
+    "error",
+    "completed",
+}
+
+# Terminal event types (success/failure) across FT and RL
+TERMINAL_EVENT_SUCCESS = {
+    "sft.completed",
+    "sft.workflow.completed",
+    "rl.job.completed",
+    "rl.train.completed",
+    "workflow.completed",
+}
+
+TERMINAL_EVENT_FAILURE = {
+    "sft.failed",
+    "sft.workflow.failed",
+    "rl.job.failed",
+    "workflow.failed",
+}
+
+

synth_ai/learning/ft_client.py

@@ -0,0 +1,59 @@
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+from ..http import AsyncHttpClient, HTTPError
+
+
+class FtClient:
+    def __init__(self, base_url: str, api_key: str, *, timeout: float = 30.0) -> None:
+        self._base_url = base_url.rstrip("/")
+        self._api_key = api_key
+        self._timeout = timeout
+
+    async def upload_training_file(self, path: str | Path, *, purpose: str = "fine-tune") -> str:
+        p = Path(path)
+        content = p.read_bytes()
+        async with AsyncHttpClient(self._base_url, self._api_key, timeout=self._timeout) as http:
+            data = {"purpose": purpose}
+            files = {"file": (p.name, content, _infer_content_type(p.name))}
+            js = await http.post_multipart("/api/learning/files", data=data, files=files)
+        if not isinstance(js, dict) or "id" not in js:
+            raise HTTPError(status=500, url="/api/learning/files", message="invalid_upload_response", body_snippet=str(js)[:200])
+        return str(js["id"])
+
+    async def create_sft_job(
+        self,
+        *,
+        model: str,
+        training_file_id: str,
+        hyperparameters: Dict[str, Any],
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> Dict[str, Any]:
+        body = {
+            "training_type": "sft_offline",
+            "model": model,
+            "training_file_id": training_file_id,
+            "hyperparameters": dict(hyperparameters or {}),
+            "metadata": dict(metadata or {}),
+        }
+        async with AsyncHttpClient(self._base_url, self._api_key, timeout=self._timeout) as http:
+            return await http.post_json("/api/learning/jobs", json=body)
+
+    async def start_job(self, job_id: str) -> Dict[str, Any]:
+        async with AsyncHttpClient(self._base_url, self._api_key, timeout=self._timeout) as http:
+            return await http.post_json(f"/api/learning/jobs/{job_id}/start", json={})
+
+
+def _infer_content_type(filename: str) -> str:
+    name = filename.lower()
+    if name.endswith(".jsonl"):
+        return "application/jsonl"
+    if name.endswith(".json"):
+        return "application/json"
+    if name.endswith(".txt"):
+        return "text/plain"
+    return "application/octet-stream"
+
+

synth_ai/learning/health.py

@@ -0,0 +1,43 @@
+from __future__ import annotations
+
+from typing import Any, Dict, Optional
+import aiohttp
+
+from ..http import AsyncHttpClient
+
+
+def _api_base(b: str) -> str:
+    b = (b or "").rstrip("/")
+    return b if b.endswith("/api") else f"{b}/api"
+
+
+async def backend_health(base_url: str, api_key: str) -> Dict[str, Any]:
+    async with AsyncHttpClient(base_url, api_key, timeout=15.0) as http:
+        js = await http.get(f"{_api_base(base_url)}/health")
+    return {"ok": True, "raw": js}
+
+
+async def task_app_health(task_app_url: str) -> Dict[str, Any]:
+    # Delegate to central task module for consistency
+    from synth_ai.task.health import task_app_health as _th
+
+    return await _th(task_app_url)
+
+
+async def pricing_preflight(base_url: str, api_key: str, *, job_type: str, gpu_type: str, estimated_seconds: float, container_count: int) -> Dict[str, Any]:
+    body = {
+        "job_type": job_type,
+        "gpu_type": gpu_type,
+        "estimated_seconds": float(estimated_seconds or 0.0),
+        "container_count": int(container_count or 1),
+    }
+    async with AsyncHttpClient(base_url, api_key, timeout=30.0) as http:
+        js = await http.post_json(f"{_api_base(base_url)}/v1/pricing/preflight", json=body)
+    return js if isinstance(js, dict) else {"raw": js}
+
+
+async def balance_autumn_normalized(base_url: str, api_key: str) -> Dict[str, Any]:
+    async with AsyncHttpClient(base_url, api_key, timeout=30.0) as http:
+        js = await http.get(f"{_api_base(base_url)}/v1/balance/autumn-normalized")
+    return js if isinstance(js, dict) else {"raw": js}
+

synth_ai/learning/jobs.py

@@ -0,0 +1,205 @@
+from __future__ import annotations
+
+from typing import Any, Callable, Dict, List, Optional
+import time
+
+from .constants import TERMINAL_EVENT_FAILURE, TERMINAL_EVENT_SUCCESS, TERMINAL_STATUSES
+from ..http import AsyncHttpClient, sleep
+
+
+def _api_base(b: str) -> str:
+    b = (b or "").rstrip("/")
+    return b if b.endswith("/api") else f"{b}/api"
+
+
+class JobsApiResolver:
+    def __init__(self, base_url: str, *, strict: bool) -> None:
+        self._base = _api_base(base_url)
+        self._strict = strict
+
+    def status_urls(self, job_id: str) -> List[str]:
+        if self._strict:
+            return [f"{self._base}/learning/jobs/{job_id}"]
+        return [
+            f"{self._base}/learning/jobs/{job_id}",
+            f"{self._base}/rl/jobs/{job_id}",
+            f"{self._base}/orchestration/jobs/{job_id}",
+        ]
+
+    def events_urls(self, job_id: str, since: int) -> List[str]:
+        if self._strict:
+            return [f"{self._base}/learning/jobs/{job_id}/events?since_seq={since}&limit=200"]
+        return [
+            f"{self._base}/learning/jobs/{job_id}/events?since_seq={since}&limit=200",
+            f"{self._base}/orchestration/jobs/{job_id}/events?since_seq={since}&limit=200",
+            # RL /jobs/{id}/events is SSE in backend; avoid in JSON poller
+        ]
+
+    def metrics_url(self, job_id: str, after_step: int) -> str:
+        return f"{self._base}/learning/jobs/{job_id}/metrics?after_step={after_step}&limit=200"
+
+
+class JobHandle:
+    def __init__(self, base_url: str, api_key: str, job_id: str, *, strict: bool = True, timeout: float = 600.0) -> None:
+        self.base_url = base_url.rstrip("/")
+        self.api_key = api_key
+        self.job_id = job_id
+        self.strict = strict
+        self.timeout = timeout
+
+    async def poll_until_terminal(
+        self,
+        *,
+        interval_seconds: float = 2.0,
+        max_seconds: float | None = None,
+        empty_polls_threshold: int = 5,
+        startup_deadline_s: int = 45,
+        on_event: Optional[Callable[[Dict[str, Any]], None]] = None,
+        on_metric: Optional[Callable[[Dict[str, Any]], None]] = None,
+    ) -> Dict[str, Any]:
+        last_seq_by_stream: Dict[str, int] = {}
+        events_job_id: Optional[str] = None
+        last_status: Optional[str] = None
+        last_step_by_name: Dict[str, int] = {}
+        empty_polls = 0
+        saw_any_event = False
+        start_t = time.time()
+        resolver = JobsApiResolver(self.base_url, strict=self.strict)
+        detected_fine_tuned_model: Optional[str] = None
+
+        async with AsyncHttpClient(self.base_url, self.api_key, timeout=self.timeout) as http:
+            while True:
+                # Status
+                status_data: Optional[Dict[str, Any]] = None
+                for su in resolver.status_urls(self.job_id):
+                    try:
+                        status_data = await http.get(su)
+                        if isinstance(status_data, dict):
+                            break
+                    except Exception:
+                        continue
+                status = str((status_data or {}).get("status") or "").lower()
+                if status_data:
+                    linked = status_data.get("linked_job_id")
+                    if isinstance(linked, str) and linked and linked != events_job_id:
+                        events_job_id = linked
+                    # Capture fine_tuned_model if already present on status
+                    if not detected_fine_tuned_model:
+                        ftm = status_data.get("fine_tuned_model")
+                        if isinstance(ftm, str) and ftm:
+                            detected_fine_tuned_model = ftm
+                if status and status != last_status:
+                    last_status = status
+                    if on_event:
+                        try:
+                            on_event({"type": "job.status", "message": status})
+                        except Exception:
+                            pass
+
+                # Events
+                stream_ids = [self.job_id]
+                if events_job_id and events_job_id not in stream_ids:
+                    stream_ids.append(events_job_id)
+                total_events_this_cycle = 0
+                terminal_event_seen = False
+                terminal_event_status: Optional[str] = None
+                for ev_id in stream_ids:
+                    since = last_seq_by_stream.get(ev_id, 0)
+                    for eu in resolver.events_urls(ev_id, since):
+                        try:
+                            ev_js = await http.get(eu)
+                        except Exception:
+                            continue
+                        try:
+                            events = (ev_js or {}).get("events") or (ev_js or {}).get("data") or []
+                            if not isinstance(events, list):
+                                events = []
+                        except Exception:
+                            events = []
+                        total_events_this_cycle += len(events)
+                        if events:
+                            saw_any_event = True
+                        for e in events:
+                            seq_val = int(e.get("seq") or 0)
+                            if seq_val <= last_seq_by_stream.get(ev_id, 0):
+                                continue
+                            last_seq_by_stream[ev_id] = seq_val
+                            if on_event:
+                                try:
+                                    on_event(e)
+                                except Exception:
+                                    pass
+                            et = str(e.get("type") or e.get("event_type") or "").lower()
+                            # Capture fine_tuned_model from event data when available
+                            if not detected_fine_tuned_model:
+                                try:
+                                    data_obj = e.get("data") or {}
+                                    ftm = data_obj.get("fine_tuned_model") if isinstance(data_obj, dict) else None
+                                    if isinstance(ftm, str) and ftm:
+                                        detected_fine_tuned_model = ftm
+                                except Exception:
+                                    pass
+                            if et in TERMINAL_EVENT_SUCCESS:
+                                terminal_event_seen = True
+                                terminal_event_status = "succeeded"
+                            elif et in TERMINAL_EVENT_FAILURE:
+                                terminal_event_seen = True
+                                terminal_event_status = "failed"
+
+                # Metrics
+                try:
+                    after = max(last_step_by_name.values()) if last_step_by_name else -1
+                    mu = resolver.metrics_url(self.job_id, after)
+                    md = await http.get(mu)
+                    for p in (md or {}).get("points", []):
+                        name = str(p.get("name") or "")
+                        step = int(p.get("step") or -1)
+                        if step <= last_step_by_name.get(name, -1):
+                            continue
+                        last_step_by_name[name] = step
+                        if on_metric:
+                            try:
+                                on_metric(p)
+                            except Exception:
+                                pass
+                except Exception:
+                    pass
+
+                # Terminal decisions
+                if terminal_event_seen or (status and status in TERMINAL_STATUSES):
+                    # Best-effort enrichment of final result with fine_tuned_model
+                    result_status = terminal_event_status or status or "completed"
+                    final_res: Dict[str, Any] = {"status": result_status, "job_id": self.job_id}
+                    if not detected_fine_tuned_model:
+                        # Briefly try to re-fetch status to see if fine_tuned_model is persisted
+                        try:
+                            for su in resolver.status_urls(self.job_id):
+                                try:
+                                    final_status = await http.get(su)
+                                    if isinstance(final_status, dict):
+                                        ftm2 = final_status.get("fine_tuned_model")
+                                        if isinstance(ftm2, str) and ftm2:
+                                            detected_fine_tuned_model = ftm2
+                                            break
+                                except Exception:
+                                    continue
+                        except Exception:
+                            pass
+                    if detected_fine_tuned_model:
+                        final_res["fine_tuned_model"] = detected_fine_tuned_model
+                    return final_res
+
+                # Guards (relaxed): do not abort on consecutive empty polls
+                if total_events_this_cycle == 0:
+                    empty_polls += 1
+                else:
+                    empty_polls = 0
+                if not saw_any_event and (time.time() - start_t) > int(startup_deadline_s):
+                    raise AssertionError(
+                        f"No events observed within startup window ({startup_deadline_s}s). Investigate event streaming."
+                    )
+                await sleep(interval_seconds)
+                if max_seconds is not None and (time.time() - start_t) >= max_seconds:
+                    raise TimeoutError(f"Polling timed out after {max_seconds}s for job {self.job_id}")
+
+

synth_ai/learning/rl_client.py

@@ -0,0 +1,256 @@
+from __future__ import annotations
+
+from typing import Any, Dict, List, Optional, Callable
+import os
+import time
+
+from ..http import AsyncHttpClient, HTTPError, sleep
+
+
+def _api_base(b: str) -> str:
+    b = (b or "").rstrip("/")
+    return b if b.endswith("/api") else f"{b}/api"
+
+
+class RlClient:
+    """Lightweight RL client for provider-agnostic job control.
+
+    Notes:
+    - Uses learning/* for status/events/metrics and rl/* for creation/start.
+    - Trainer endpoints are resolved server-side via trainer_id.
+    """
+
+    def __init__(self, base_url: str, api_key: str, *, timeout: float = 600.0) -> None:
+        self._base_url = base_url.rstrip("/")
+        self._api_key = api_key
+        self._timeout = timeout
+
+    async def resolve_trainer_start_url(self, trainer_id: str) -> str:
+        """GET /api/rl/services/{id} → { training_start_url }"""
+        path = f"/api/rl/services/{trainer_id}"
+        async with AsyncHttpClient(self._base_url, self._api_key, timeout=30.0) as http:
+            js = await http.get(path)
+        if not isinstance(js, dict):
+            raise HTTPError(status=500, url=path, message="invalid_service_response", body_snippet=str(js)[:200])
+        start_url = js.get("training_start_url")
+        if not isinstance(start_url, str) or not start_url:
+            raise HTTPError(status=500, url=path, message="missing_training_start_url", body_snippet=str(js)[:200])
+        return start_url
+
+    async def create_job(
+        self,
+        *,
+        model: str,
+        task_app_url: str,
+        trainer: Dict[str, Any],
+        trainer_id: Optional[str] = None,
+        job_config_id: Optional[str] = None,
+        inline_config: Optional[Dict[str, Any]] = None,
+    ) -> Dict[str, Any]:
+        body = {
+            "job_type": "rl",
+            "data": {
+                "model": model,
+                "endpoint_base_url": task_app_url,
+                **({"job_config_id": job_config_id} if job_config_id else {}),
+                **({"config": inline_config} if inline_config else {}),
+                "trainer": {
+                    "batch_size": int(trainer.get("batch_size", 1)),
+                    "group_size": max(2, int(trainer.get("group_size", 2))),
+                },
+            },
+        }
+        async with AsyncHttpClient(self._base_url, self._api_key, timeout=self._timeout) as http:
+            js = await http.post_json(f"{_api_base(self._base_url)}/rl/jobs", json=body)
+        if not isinstance(js, dict):
+            raise HTTPError(status=500, url="/api/rl/jobs", message="invalid_create_response", body_snippet=str(js)[:200])
+        return js
+
+    async def start_job_if_supported(self, job_id: str) -> Optional[Dict[str, Any]]:
+        path = f"{_api_base(self._base_url)}/rl/jobs/{job_id}/start"
+        try:
+            async with AsyncHttpClient(self._base_url, self._api_key, timeout=30.0) as http:
+                return await http.post_json(path, json={})
+        except HTTPError as he:  # noqa: PERF203
+            if he.status == 404:
+                return None
+            raise
+
+    async def get_job(self, job_id: str) -> Dict[str, Any]:
+        async with AsyncHttpClient(self._base_url, self._api_key, timeout=30.0) as http:
+            return await http.get(f"{_api_base(self._base_url)}/learning/jobs/{job_id}")
+
+    async def get_events(self, job_id: str, *, since_seq: int = 0, limit: int = 200) -> List[Dict[str, Any]]:
+        params = {"since_seq": since_seq, "limit": limit}
+        async with AsyncHttpClient(self._base_url, self._api_key, timeout=30.0) as http:
+            try:
+                js = await http.get(f"{_api_base(self._base_url)}/learning/jobs/{job_id}/events", params=params)
+            except HTTPError as he:
+                try:
+                    print(
+                        f"[poll] events HTTPError status={he.status} url={he.url} since_seq={since_seq} body={(he.body_snippet or '')[:200]}"
+                    )
+                except Exception:
+                    pass
+                raise
+        if isinstance(js, dict):
+            evs = js.get("events") or js.get("data")
+            if isinstance(evs, list):
+                return evs
+        return []
+
+    async def get_metrics(self, job_id: str, *, after_step: int = -1, limit: int = 200) -> List[Dict[str, Any]]:
+        params = {"after_step": after_step, "limit": limit}
+        async with AsyncHttpClient(self._base_url, self._api_key, timeout=30.0) as http:
+            js = await http.get(f"{_api_base(self._base_url)}/learning/jobs/{job_id}/metrics", params=params)
+        if isinstance(js, dict) and isinstance(js.get("points"), list):
+            return js["points"]
+        return []
+
+    async def poll_until_terminal(
+        self,
+        job_id: str,
+        *,
+        interval_seconds: float = 2.0,
+        max_seconds: float | None = None,
+        empty_polls_threshold: int = 5,
+        startup_deadline_s: int = 45,
+        on_event: Optional[Callable[[Dict[str, Any]], None]] = None,
+        on_metric: Optional[Callable[[Dict[str, Any]], None]] = None,
+    ) -> Dict[str, Any]:
+        last_seq_by_stream: Dict[str, int] = {}
+        events_job_id: Optional[str] = None
+        last_status: Optional[str] = None
+        last_step_by_name: Dict[str, int] = {}
+        empty_polls = 0
+        saw_any_event = False
+        start_t = time.time()
+        terminal = {"succeeded", "failed", "cancelled", "canceled", "error", "completed"}
+
+        while True:
+            status_data: Optional[Dict[str, Any]] = None
+            try:
+                status_data = await self.get_job(job_id)
+            except Exception:
+                status_data = None
+            if status_data is None:
+                try:
+                    print(f"[poll] get_job returned None base={self._base_url} job_id={job_id}")
+                except Exception:
+                    pass
+            status = str((status_data or {}).get("status") or "").lower()
+            if status_data:
+                linked = status_data.get("linked_job_id")
+                if isinstance(linked, str) and linked and linked != events_job_id:
+                    events_job_id = linked
+                    try:
+                        print(f"[poll] discovered linked_job_id stream={events_job_id}")
+                    except Exception:
+                        pass
+            if status and status != last_status:
+                last_status = status
+                # Status transitions only to avoid log spam
+                if on_event:
+                    try:
+                        on_event({"type": "rl.status", "message": status})
+                    except Exception:
+                        pass
+
+            # Events
+            stream_ids = [job_id]
+            if events_job_id and events_job_id not in stream_ids:
+                stream_ids.append(events_job_id)
+            try:
+                print(f"[poll] streams={stream_ids} intervals={interval_seconds}s since_map={last_seq_by_stream} empty_polls={empty_polls}")
+            except Exception:
+                pass
+            total_events_this_cycle = 0
+            terminal_event_seen = False
+            terminal_event_status: Optional[str] = None
+            for ev_id in stream_ids:
+                since = last_seq_by_stream.get(ev_id, 0)
+                try:
+                    events = await self.get_events(ev_id, since_seq=since, limit=200)
+                except HTTPError as he:
+                    try:
+                        print(f"[poll] get_events error status={he.status} url={he.url} since={since} body={(he.body_snippet or '')[:200]}")
+                    except Exception:
+                        pass
+                    events = []
+                except Exception as e:
+                    try:
+                        print(f"[poll] get_events unexpected error ev_id={ev_id} since={since} err={type(e).__name__}: {e}")
+                    except Exception:
+                        pass
+                    events = []
+                total_events_this_cycle += len(events)
+                if events:
+                    saw_any_event = True
+                for e in events:
+                    seq_val = int(e.get("seq") or 0)
+                    if seq_val <= last_seq_by_stream.get(ev_id, 0):
+                        continue
+                    last_seq_by_stream[ev_id] = seq_val
+                    if on_event:
+                        try:
+                            on_event(e)
+                        except Exception:
+                            pass
+                    et = str(e.get("type") or e.get("event_type") or "").lower()
+                    if et in ("rl.job.completed", "workflow.completed", "rl.train.completed"):
+                        terminal_event_seen = True
+                        terminal_event_status = "succeeded"
+                    elif et in ("rl.job.failed", "workflow.failed"):
+                        terminal_event_seen = True
+                        terminal_event_status = "failed"
+
+            # Metrics
+            try:
+                after = max(last_step_by_name.values()) if last_step_by_name else -1
+                points = await self.get_metrics(job_id, after_step=after, limit=200)
+                for p in points:
+                    name = str(p.get("name") or "")
+                    step = int(p.get("step") or -1)
+                    if step <= last_step_by_name.get(name, -1):
+                        continue
+                    last_step_by_name[name] = step
+                    if on_metric:
+                        try:
+                            on_metric(p)
+                        except Exception:
+                            pass
+            except Exception:
+                pass
+
+            if terminal_event_seen:
+                return {"status": terminal_event_status or status or "completed", "job_id": job_id}
+            if status and status in terminal:
+                return {"status": status, "job_id": job_id}
+
+            if total_events_this_cycle == 0:
+                empty_polls += 1
+            else:
+                empty_polls = 0
+            if empty_polls >= max(1, int(empty_polls_threshold)):
+                try:
+                    print(
+                        f"[poll] threshold hit: empty_polls={empty_polls} >= {empty_polls_threshold} streams={stream_ids} last_seq_map={last_seq_by_stream}"
+                    )
+                except Exception:
+                    pass
+                raise AssertionError(f"No new events detected for {empty_polls_threshold} consecutive polls. Check event ingestion.")
+
+            if not saw_any_event and (time.time() - start_t) > int(startup_deadline_s):
+                try:
+                    print(
+                        f"[poll] startup window exceeded: {startup_deadline_s}s base={self._base_url} job={job_id} streams={stream_ids} last_seq_map={last_seq_by_stream}"
+                    )
+                except Exception:
+                    pass
+                raise AssertionError(f"No events observed within startup window ({startup_deadline_s}s). Investigate event streaming.")
+
+            await sleep(interval_seconds)
+            if max_seconds is not None and (time.time() - start_t) >= max_seconds:
+                raise TimeoutError(f"Polling timed out after {max_seconds}s for job {job_id}")
+
+