synth-ai 0.2.9.dev7__py3-none-any.whl → 0.2.10__py3-none-any.whl

synth_ai/api/train/builders.py

@@ -5,8 +5,19 @@ from pathlib import Path
 from typing import Any
 
 import click
+from synth_ai.api.models.supported import (
+    UnsupportedModelError,
+    ensure_allowed_model,
+    normalize_model_identifier,
+)
+from synth_ai.learning.sft.config import prepare_sft_job_payload
 
-from .utils import ensure_api_base, load_toml, TrainError
+from .supported_algos import (
+    AlgorithmValidationError,
+    ensure_model_supported_for_algorithm,
+    validate_algorithm_config,
+)
+from .utils import TrainError, ensure_api_base, load_toml
 
 
 @dataclass(slots=True)
@@ -29,29 +40,78 @@ def build_rl_payload(
     task_url: str,
     overrides: dict[str, Any],
     idempotency: str | None,
+    allow_experimental: bool | None = None,
 ) -> RLBuildResult:
     data = load_toml(config_path)
+    try:
+        spec = validate_algorithm_config(data.get("algorithm"), expected_family="rl")
+    except AlgorithmValidationError as exc:
+        raise click.ClickException(str(exc)) from exc
     services = data.get("services") if isinstance(data.get("services"), dict) else {}
     model_cfg = data.get("model") if isinstance(data.get("model"), dict) else {}
 
     final_task_url = (
-        overrides.get("task_url") or task_url or services.get("task_url") or ""
+        overrides.get("task_url")
+        or task_url
+        or (services.get("task_url") if isinstance(services, dict) else None)
+        or ""
     ).strip()
     if not final_task_url:
         raise click.ClickException(
             "Task app URL required (provide --task-url or set services.task_url in TOML)"
         )
 
-    model_source = (model_cfg.get("source") or "").strip()
-    model_base = (model_cfg.get("base") or "").strip()
+    raw_source = model_cfg.get("source") if isinstance(model_cfg, dict) else ""
+    model_source = str(raw_source or "").strip()
+    raw_base = model_cfg.get("base") if isinstance(model_cfg, dict) else ""
+    model_base = str(raw_base or "").strip()
     override_model = (overrides.get("model") or "").strip()
     if override_model:
         model_source = override_model
         model_base = ""
     if bool(model_source) == bool(model_base):
+        details = (
+            f"Config: {config_path}\n"
+            f"[model].source={model_source!r} | [model].base={model_base!r}"
+        )
+        hint = (
+            "Set exactly one: [model].base for a base model (e.g. 'Qwen/Qwen3-1.7B') "
+            "or [model].source for a fine-tuned model id. Also remove any conflicting "
+            "'[policy].model' entries."
+        )
         raise click.ClickException(
-            "Model section must specify exactly one of [model].source or [model].base"
+            "Invalid model config: exactly one of [model].source or [model].base is required.\n"
+            + details
+            + "\nHint: "
+            + hint
+        )
+
+    try:
+        if model_source:
+            model_source = normalize_model_identifier(model_source)
+        if model_base:
+            model_base = normalize_model_identifier(model_base, allow_finetuned_prefixes=False)
+    except UnsupportedModelError as exc:
+        raise click.ClickException(str(exc)) from exc
+
+    base_model_for_training: str | None = None
+    if model_source:
+        base_model_for_training = ensure_allowed_model(
+            model_source,
+            allow_finetuned_prefixes=True,
+            allow_experimental=allow_experimental,
+        )
+    elif model_base:
+        base_model_for_training = ensure_allowed_model(
+            model_base,
+            allow_finetuned_prefixes=False,
+            allow_experimental=allow_experimental,
         )
+    if base_model_for_training:
+        try:
+            ensure_model_supported_for_algorithm(base_model_for_training, spec)
+        except AlgorithmValidationError as exc:
+            raise click.ClickException(str(exc)) from exc
 
     # Force TOML services.task_url to the effective endpoint to avoid split URLs
     try:
@@ -87,15 +147,24 @@ def build_sft_payload(
     *,
     config_path: Path,
     dataset_override: Path | None,
+    allow_experimental: bool | None,
 ) -> SFTBuildResult:
     data = load_toml(config_path)
+    try:
+        spec = validate_algorithm_config(data.get("algorithm"), expected_family="sft")
+    except AlgorithmValidationError as exc:
+        raise TrainError(str(exc)) from exc
     job_cfg = data.get("job") if isinstance(data.get("job"), dict) else {}
     data_cfg = data.get("data") if isinstance(data.get("data"), dict) else {}
     hp_cfg = data.get("hyperparameters") if isinstance(data.get("hyperparameters"), dict) else {}
     train_cfg = data.get("training") if isinstance(data.get("training"), dict) else {}
     compute_cfg = data.get("compute") if isinstance(data.get("compute"), dict) else {}
 
-    raw_dataset = dataset_override or job_cfg.get("data") or job_cfg.get("data_path")
+    raw_dataset = (
+        dataset_override
+        or (job_cfg.get("data") if isinstance(job_cfg, dict) else None)
+        or (job_cfg.get("data_path") if isinstance(job_cfg, dict) else None)
+    )
     if not raw_dataset:
         raise TrainError("Dataset not specified; pass --dataset or set [job].data")
     dataset_path = Path(raw_dataset)
@@ -108,7 +177,9 @@ def build_sft_payload(
 
     validation_path = (
         data_cfg.get("validation_path")
-        if isinstance(data_cfg.get("validation_path"), str)
+        if isinstance(data_cfg, dict)
+        else None
+        if isinstance(data_cfg, dict) and isinstance(data_cfg.get("validation_path"), str)
         else None
     )
     validation_file = None
@@ -122,7 +193,7 @@ def build_sft_payload(
             validation_file = vpath
 
     hp_block: dict[str, Any] = {
-        "n_epochs": int(hp_cfg.get("n_epochs", 1)),
+        "n_epochs": int(hp_cfg.get("n_epochs", 1) if isinstance(hp_cfg, dict) else 1),
     }
     for key in (
         "batch_size",
@@ -134,27 +205,35 @@ def build_sft_payload(
         "warmup_ratio",
         "train_kind",
     ):
-        if key in hp_cfg:
+        if isinstance(hp_cfg, dict) and key in hp_cfg:
             hp_block[key] = hp_cfg[key]
-    if isinstance(hp_cfg.get("parallelism"), dict):
+    if isinstance(hp_cfg, dict) and isinstance(hp_cfg.get("parallelism"), dict):
         hp_block["parallelism"] = hp_cfg["parallelism"]
 
     compute_block = {
-        k: compute_cfg[k] for k in ("gpu_type", "gpu_count", "nodes") if k in compute_cfg
+        k: compute_cfg[k]
+        for k in ("gpu_type", "gpu_count", "nodes")
+        if isinstance(compute_cfg, dict) and k in compute_cfg
     }
 
     effective = {
         "compute": compute_block,
         "data": {
             "topology": data_cfg.get("topology", {})
-            if isinstance(data_cfg.get("topology"), dict)
+            if isinstance(data_cfg, dict) and isinstance(data_cfg.get("topology"), dict)
             else {}
         },
-        "training": {k: v for k, v in train_cfg.items() if k in ("mode", "use_qlora")},
+        "training": {
+            k: v
+            for k, v in (train_cfg.items() if isinstance(train_cfg, dict) else [])
+            if k in ("mode", "use_qlora")
+        },
     }
 
     validation_cfg = (
-        train_cfg.get("validation") if isinstance(train_cfg.get("validation"), dict) else None
+        train_cfg.get("validation")
+        if isinstance(train_cfg, dict) and isinstance(train_cfg.get("validation"), dict)
+        else None
     )
     if isinstance(validation_cfg, dict):
         hp_block.update(
@@ -170,13 +249,41 @@ def build_sft_payload(
             "enabled": bool(validation_cfg.get("enabled", True))
         }
 
-    payload = {
-        "model": job_cfg.get("model") or data.get("model"),
-        "training_file_id": None,  # populated after upload
-        "training_type": "sft_offline",
-        "hyperparameters": hp_block,
-        "metadata": {"effective_config": effective},
-    }
+    raw_model = str(
+        job_cfg.get("model") if isinstance(job_cfg, dict) else None or data.get("model") or ""
+    ).strip()
+    if not raw_model:
+        raise TrainError("Model not specified; set [job].model or [model].base in the config")
+
+    try:
+        base_model = ensure_allowed_model(
+            raw_model,
+            allow_finetuned_prefixes=False,
+            allow_experimental=allow_experimental,
+        )
+    except UnsupportedModelError as exc:
+        raise TrainError(str(exc)) from exc
+    try:
+        ensure_model_supported_for_algorithm(base_model, spec)
+    except AlgorithmValidationError as exc:
+        raise TrainError(str(exc)) from exc
+
+    try:
+        payload = prepare_sft_job_payload(
+            model=raw_model,
+            training_file=None,
+            hyperparameters=hp_block,
+            metadata={"effective_config": effective},
+            training_type="sft_offline",
+            training_file_field="training_file_id",
+            require_training_file=False,
+            include_training_file_when_none=True,
+            allow_finetuned_prefixes=False,
+        )
+    except UnsupportedModelError as exc:
+        raise TrainError(str(exc)) from exc
+    except ValueError as exc:
+        raise TrainError(str(exc)) from exc
 
     return SFTBuildResult(payload=payload, train_file=dataset_path, validation_file=validation_file)
 

synth_ai/api/train/cli.py

@@ -2,21 +2,22 @@ from __future__ import annotations
 
 import os
 from pathlib import Path
-from typing import Any, Dict
+from typing import Any
 
 import click
+from synth_ai.config.base_url import get_backend_from_env
 
-from .builders import RLBuildResult, SFTBuildResult, build_rl_payload, build_sft_payload
+from .builders import build_rl_payload, build_sft_payload
 from .config_finder import discover_configs, prompt_for_config
 from .env_resolver import KeySpec, resolve_env
 from .pollers import RLJobPoller, SFTJobPoller
 from .task_app import check_task_app_health
 from .utils import (
-    TrainError,
     REPO_ROOT,
+    TrainError,
     ensure_api_base,
-    http_post,
     http_get,
+    http_post,
     limit_jsonl_examples,
     mask_value,
     post_multipart,
@@ -24,7 +25,6 @@ from .utils import (
     sleep,
     validate_sft_jsonl,
 )
-from synth_ai.config.base_url import get_backend_from_env
 
 
 def _discover_dataset_candidates(config_path: Path, limit: int = 50) -> list[Path]:
@@ -130,8 +130,23 @@ def _default_backend() -> str:
 )
 @click.option("--backend", default=_default_backend, help="Backend base URL")
 @click.option("--model", default=None, help="Override model identifier")
+@click.option(
+    "--allow-experimental",
+    "allow_experimental",
+    is_flag=True,
+    flag_value=True,
+    default=None,
+    help="Allow experimental models (overrides SDK_EXPERIMENTAL env)",
+)
+@click.option(
+    "--no-allow-experimental",
+    "allow_experimental",
+    is_flag=True,
+    flag_value=False,
+    help="Disallow experimental models (overrides SDK_EXPERIMENTAL env)",
+)
 @click.option("--idempotency", default=None, help="Idempotency-Key header for job creation")
-@click.option("--dry-run", is_flag=True, help="Preview payload without submitting")
+@click.option("--dry-run", is_flag=True, hidden=True, help="Deprecated: no-op")
 @click.option("--poll/--no-poll", default=True, help="Poll job status until terminal state")
 @click.option(
     "--poll-timeout", default=3600.0, type=float, help="Maximum seconds to poll before timing out"
@@ -152,6 +167,7 @@ def train_command(
     dataset_path: str | None,
     backend: str,
     model: str | None,
+    allow_experimental: bool | None,
     idempotency: str | None,
     dry_run: bool,
     poll: bool,
@@ -165,7 +181,9 @@ def train_command(
         list(config_paths), requested_type=train_type if train_type != "auto" else None
     )
     selection = prompt_for_config(
-        candidates, requested_type=train_type if train_type != "auto" else None
+        candidates,
+        requested_type=train_type if train_type != "auto" else None,
+        allow_autoselect=bool(config_paths),
     )
 
     effective_type = train_type if train_type != "auto" else selection.train_type
@@ -243,6 +261,7 @@ def train_command(
             task_url_override=task_url,
             model_override=model,
             idempotency=idempotency,
+            allow_experimental=allow_experimental,
             dry_run=dry_run,
             poll=poll,
             poll_timeout=poll_timeout,
@@ -255,6 +274,7 @@ def train_command(
             backend_base=backend_base,
             synth_key=synth_key,
             dataset_override=dataset_override_path,
+            allow_experimental=allow_experimental,
             dry_run=dry_run,
             poll=poll,
             poll_timeout=poll_timeout,
@@ -303,7 +323,7 @@ def _wait_for_training_file(
                 error_body = resp.json()
             except Exception:
                 error_body = resp.text[:400]
-            click.echo(f"\n[ERROR] Authentication failed when checking training file:")
+            click.echo("\n[ERROR] Authentication failed when checking training file:")
             click.echo(f"  URL: {url}")
             click.echo(f"  Status: {resp.status_code}")
             click.echo(f"  Response: {error_body}")
@@ -339,12 +359,13 @@ def handle_rl(
     task_url_override: str | None,
     model_override: str | None,
     idempotency: str | None,
+    allow_experimental: bool | None,
     dry_run: bool,
     poll: bool,
     poll_timeout: float,
     poll_interval: float,
 ) -> None:
-    overrides: Dict[str, Any] = {
+    overrides: dict[str, Any] = {
         "backend": backend_base,
         "task_url": task_url_override,
         "model": model_override,
@@ -354,6 +375,7 @@ def handle_rl(
         task_url=task_url_override or os.environ.get("TASK_APP_URL", ""),
         overrides=overrides,
         idempotency=idempotency,
+        allow_experimental=allow_experimental,
     )
 
     # Backend-side verification: try ALL org environment keys against /health and /task_info
@@ -371,7 +393,7 @@ def handle_rl(
         raise click.ClickException(
             f"Task app verification call failed: {type(_ve).__name__}: {_ve}"
         ) from _ve
-    if vresp.status_code >= 400:
+    if vresp.status_code is not None and vresp.status_code >= 400:
         click.echo("Task app verification error:\n" + preview_json(vjs, limit=800))
         raise click.ClickException(f"Verification failed with status {vresp.status_code}")
     if not bool(vjs.get("any_ok")):
@@ -407,9 +429,6 @@ def handle_rl(
 
     click.echo(f"POST {create_url}")
     click.echo("Payload preview:\n" + preview_json(build.payload, limit=800))
-    if dry_run:
-        click.echo("Dry run enabled; skipping submission")
-        return
 
     resp = http_post(create_url, headers=headers, json_body=build.payload)
     try:
@@ -439,6 +458,7 @@ def handle_sft(
     backend_base: str,
     synth_key: str,
     dataset_override: Path | None,
+    allow_experimental: bool | None,
     dry_run: bool,
     poll: bool,
     poll_timeout: float,
@@ -449,7 +469,11 @@ def handle_sft(
 
     while True:
         try:
-            build = build_sft_payload(config_path=cfg_path, dataset_override=dataset_path)
+            build = build_sft_payload(
+                config_path=cfg_path,
+                dataset_override=dataset_path,
+                allow_experimental=allow_experimental,
+            )
             break
         except TrainError as exc:
             click.echo(str(exc))
@@ -472,54 +496,49 @@ def handle_sft(
             validate_sft_jsonl(build.validation_file)
 
         upload_url = f"{backend_base}/learning/files"
-        click.echo(f"\n=== Uploading Training Data ===")
+        click.echo("\n=== Uploading Training Data ===")
         click.echo(f"Dataset: {build.train_file}")
         click.echo(f"Destination: {upload_url}")
-        if dry_run:
-            click.echo("Dry run: skipping upload")
-            train_file_id = "dry-run-train"
-            val_file_id = None
-        else:
-            resp = post_multipart(
-                upload_url, api_key=synth_key, file_field="file", file_path=build.train_file
+        resp = post_multipart(
+            upload_url, api_key=synth_key, file_field="file", file_path=build.train_file
+        )
+        js = (
+            resp.json()
+            if resp.headers.get("content-type", "").startswith("application/json")
+            else {}
+        )
+        if resp.status_code is not None and resp.status_code >= 400 or "id" not in js:
+            click.echo("\n[ERROR] Training file upload failed:")
+            click.echo(f"  URL: {upload_url}")
+            click.echo(f"  Status: {resp.status_code}")
+            click.echo(f"  Response: {js or resp.text[:400]}")
+            click.echo(f"  File: {build.train_file}")
+            raise click.ClickException(
+                f"Training file upload failed with status {resp.status_code}"
+            )
+        train_file_id = js["id"]
+        click.echo(f"✓ Training file uploaded (id={train_file_id})")
+        val_file_id = None
+        if build.validation_file:
+            click.echo(f"Uploading validation dataset: {build.validation_file}")
+            vresp = post_multipart(
+                upload_url,
+                api_key=synth_key,
+                file_field="file",
+                file_path=build.validation_file,
             )
-            js = (
-                resp.json()
-                if resp.headers.get("content-type", "").startswith("application/json")
+            vjs = (
+                vresp.json()
+                if vresp.headers.get("content-type", "").startswith("application/json")
                 else {}
             )
-            if resp.status_code >= 400 or "id" not in js:
-                click.echo(f"\n[ERROR] Training file upload failed:")
-                click.echo(f"  URL: {upload_url}")
-                click.echo(f"  Status: {resp.status_code}")
-                click.echo(f"  Response: {js or resp.text[:400]}")
-                click.echo(f"  File: {build.train_file}")
-                raise click.ClickException(
-                    f"Training file upload failed with status {resp.status_code}"
-                )
-            train_file_id = js["id"]
-            click.echo(f"✓ Training file uploaded (id={train_file_id})")
-            val_file_id = None
-            if build.validation_file:
-                click.echo(f"Uploading validation dataset: {build.validation_file}")
-                vresp = post_multipart(
-                    upload_url,
-                    api_key=synth_key,
-                    file_field="file",
-                    file_path=build.validation_file,
-                )
-                vjs = (
-                    vresp.json()
-                    if vresp.headers.get("content-type", "").startswith("application/json")
-                    else {}
+            if vresp.status_code is not None and vresp.status_code < 400 and "id" in vjs:
+                val_file_id = vjs["id"]
+                click.echo(f"✓ Validation file uploaded (id={val_file_id})")
+            else:
+                click.echo(
+                    f"[WARN] Validation upload failed ({vresp.status_code}): {vjs or vresp.text[:200]}"
                 )
-                if vresp.status_code < 400 and "id" in vjs:
-                    val_file_id = vjs["id"]
-                    click.echo(f"✓ Validation file uploaded (id={val_file_id})")
-                else:
-                    click.echo(
-                        f"[WARN] Validation upload failed ({vresp.status_code}): {vjs or vresp.text[:200]}"
-                    )
         payload = dict(build.payload)
         payload["training_file_id"] = train_file_id
         if val_file_id:
@@ -527,18 +546,15 @@ def handle_sft(
                 "data", {}
             )["validation_files"] = [val_file_id]
 
-        click.echo(f"\n=== Checking File Processing Status ===")
+        click.echo("\n=== Checking File Processing Status ===")
         try:
             _wait_for_training_file(backend_base, synth_key, train_file_id)
         except click.ClickException as exc:
             raise click.ClickException(f"Training file {train_file_id} not ready: {exc}") from exc
 
-        click.echo(f"\n=== Creating Training Job ===")
+        click.echo("\n=== Creating Training Job ===")
         click.echo("Job payload preview:")
         click.echo(preview_json(payload, limit=800))
-        if dry_run:
-            click.echo("Dry run: skipping job submission")
-            return
 
         create_url = f"{backend_base}/learning/jobs"
         headers = {"Authorization": f"Bearer {synth_key}", "Content-Type": "application/json"}
@@ -550,7 +566,7 @@ def handle_sft(
             else {}
         )
         if resp.status_code not in (200, 201):
-            click.echo(f"\n[ERROR] Job creation failed:")
+            click.echo("\n[ERROR] Job creation failed:")
             click.echo(f"  URL: {create_url}")
             click.echo(f"  Status: {resp.status_code}")
             click.echo(f"  Response: {preview_json(js, limit=600)}")
@@ -560,14 +576,14 @@ def handle_sft(
             raise click.ClickException("Response missing job id")
         click.echo(f"✓ Job created (id={job_id})")
 
-        click.echo(f"\n=== Starting Training Job ===")
+        click.echo("\n=== Starting Training Job ===")
         start_url = f"{backend_base}/learning/jobs/{job_id}/start"
         click.echo(f"POST {start_url}")
         start_resp = http_post(start_url, headers=headers, json_body={})
         if start_resp.status_code not in (200, 201):
             click.echo(f"[WARN] Job start returned status {start_resp.status_code}")
         else:
-            click.echo(f"✓ Job started")
+            click.echo("✓ Job started")
 
         if not poll:
             click.echo(f"Started job {job_id} (polling disabled)")

synth_ai/api/train/config_finder.py

@@ -2,9 +2,9 @@ from __future__ import annotations
 
 import json
 import os
+from collections.abc import Iterable
 from dataclasses import dataclass
 from pathlib import Path
-from typing import Iterable
 
 import click
 
@@ -173,7 +173,7 @@ def discover_configs(explicit: list[str], *, requested_type: str | None) -> list
 
 
 def prompt_for_config(
-    candidates: list[ConfigCandidate], *, requested_type: str | None
+    candidates: list[ConfigCandidate], *, requested_type: str | None, allow_autoselect: bool = False
 ) -> ConfigCandidate:
     if not candidates:
         raise click.ClickException("No training configs found. Pass --config explicitly.")
@@ -182,6 +182,11 @@ def prompt_for_config(
     last_config = _load_last_config()
     default_idx = 1
 
+    if allow_autoselect and len(candidates) == 1:
+        chosen = candidates[0]
+        _save_last_config(chosen.path)
+        return chosen
+
     if last_config:
         for idx, cand in enumerate(candidates):
             if cand.path.resolve() == last_config:

synth_ai/api/train/env_resolver.py

@@ -1,9 +1,9 @@
 from __future__ import annotations
 
 import os
+from collections.abc import Callable, Iterable, MutableMapping
 from dataclasses import dataclass
 from pathlib import Path
-from typing import Callable, Iterable, MutableMapping
 
 import click
 

synth_ai/api/train/pollers.py

@@ -1,8 +1,9 @@
 from __future__ import annotations
 
+from collections.abc import Mapping
 from dataclasses import dataclass
 from datetime import datetime
-from typing import Any, Mapping
+from typing import Any
 
 import click