synth-ai 0.2.14__py3-none-any.whl → 0.2.17__py3-none-any.whl

synth_ai/__init__.py

@@ -1,12 +1,22 @@
-"""
-Synth AI - Software for aiding the best and multiplying the will.
-"""
-
 from __future__ import annotations
 
+import importlib
 from importlib import metadata as _metadata
 from importlib.metadata import PackageNotFoundError
 from pathlib import Path
+from typing import Any, cast
+
+import synth_ai.environments as environments  # expose module name for __all__
+from synth_ai.environments import *  # noqa
+from synth_ai.judge_schemas import (
+    CriterionScorePayload,
+    JudgeOptions,
+    JudgeScoreRequest,
+    JudgeScoreResponse,
+    JudgeTaskApp,
+    JudgeTracePayload,
+    ReviewPayload,
+)
 
 try:  # Prefer the installed package metadata when available
     __version__ = _metadata.version("synth-ai")
@@ -24,35 +34,39 @@ except PackageNotFoundError:  # Fallback to pyproject version for editable insta
     except Exception:
         __version__ = "0.0.0.dev0"
 
-# Environment exports - moved from synth-env
-from synth_ai.environments import *  # noqa
-import synth_ai.environments as environments  # expose module name for __all__
+def _optional_import(module_path: str) -> Any | None:
+    try:
+        return importlib.import_module(module_path)
+    except Exception:
+        return None
 
-try:
-    from synth_ai.lm.core.main import LM  # Moved from zyk to lm for better organization
-except Exception:  # allow minimal imports (e.g., tracing) without LM stack
-    LM = None  # type: ignore
-try:
-    from synth_ai.lm.provider_support.anthropic import Anthropic, AsyncAnthropic
-except Exception:  # optional in minimal environments
-    Anthropic = AsyncAnthropic = None  # type: ignore
 
-# Provider support exports - moved from synth-sdk to synth_ai/lm
-try:
-    from synth_ai.lm.provider_support.openai import AsyncOpenAI, OpenAI
-except Exception:
-    AsyncOpenAI = OpenAI = None  # type: ignore
+_lm_module = _optional_import("synth_ai.lm.core.main")
+LM = cast(Any, _lm_module).LM if _lm_module and hasattr(_lm_module, "LM") else None  # type: ignore[attr-defined]
 
-# Judge API contract schemas
-from synth_ai.judge_schemas import (
-    CriterionScorePayload,
-    JudgeOptions,
-    JudgeScoreRequest,
-    JudgeScoreResponse,
-    JudgeTaskApp,
-    JudgeTracePayload,
-    ReviewPayload,
-)
+_anthropic_module = _optional_import("synth_ai.lm.provider_support.anthropic")
+Anthropic = (
+    cast(Any, _anthropic_module).Anthropic
+    if _anthropic_module and hasattr(_anthropic_module, "Anthropic")
+    else None
+)  # type: ignore[attr-defined]
+AsyncAnthropic = (
+    cast(Any, _anthropic_module).AsyncAnthropic
+    if _anthropic_module and hasattr(_anthropic_module, "AsyncAnthropic")
+    else None
+)  # type: ignore[attr-defined]
+
+_openai_module = _optional_import("synth_ai.lm.provider_support.openai")
+AsyncOpenAI = (
+    cast(Any, _openai_module).AsyncOpenAI
+    if _openai_module and hasattr(_openai_module, "AsyncOpenAI")
+    else None
+)  # type: ignore[attr-defined]
+OpenAI = (
+    cast(Any, _openai_module).OpenAI
+    if _openai_module and hasattr(_openai_module, "OpenAI")
+    else None
+)  # type: ignore[attr-defined]
 
 # Legacy tracing v1 is not required for v3 usage and can be unavailable in minimal envs.
 tracing = None  # type: ignore

synth_ai/_utils/__init__.py

@@ -0,0 +1,47 @@
+"""
+Compatibility shims for legacy `synth_ai._utils.*` imports.
+
+The modern codebase exposes these helpers under ``synth_ai.utils``.  These
+modules re-export the public symbols so existing downstream code (and our own
+older examples/tests) continue to work without modification.
+"""
+
+from __future__ import annotations
+
+from importlib import import_module
+from types import ModuleType
+
+_MAPPING = {
+    "base_url": "synth_ai.utils.base_url",
+    "http": "synth_ai.utils.http",
+    "prompts": "synth_ai.utils.prompts",
+    "task_app_state": "synth_ai.utils.task_app_state",
+    "user_config": "synth_ai.utils.user_config",
+}
+
+__all__ = sorted(_MAPPING.keys())
+
+
+def __getattr__(name: str) -> ModuleType:
+    target = _MAPPING.get(name)
+    if not target:
+        raise AttributeError(f"module 'synth_ai._utils' has no attribute '{name}'")
+    module = import_module(target)
+    globals()[name] = module
+    return module
+
+
+def __dir__() -> list[str]:
+    return sorted(set(globals()) | set(__all__))
+
+
+def _export(module_name: str) -> None:
+    module = import_module(_MAPPING[module_name])
+    globals().setdefault(module_name, module)
+    if hasattr(module, "__all__"):
+        for attr in module.__all__:  # type: ignore[attr-defined]
+            globals().setdefault(attr, getattr(module, attr))
+
+
+for _name in __all__:
+    _export(_name)

synth_ai/_utils/base_url.py

@@ -0,0 +1,10 @@
+from __future__ import annotations
+
+from synth_ai.utils.base_url import *  # noqa: F401,F403
+
+try:
+    from synth_ai.utils.base_url import __all__ as __wrapped_all__  # type: ignore[attr-defined]
+except ImportError:  # pragma: no cover - defensive
+    __wrapped_all__ = []
+
+__all__ = list(__wrapped_all__)

synth_ai/_utils/http.py

@@ -0,0 +1,10 @@
+from __future__ import annotations
+
+from synth_ai.utils.http import *  # noqa: F401,F403
+
+try:
+    from synth_ai.utils.http import __all__ as __wrapped_all__  # type: ignore[attr-defined]
+except ImportError:  # pragma: no cover - defensive
+    __wrapped_all__ = []
+
+__all__ = list(__wrapped_all__)

synth_ai/_utils/prompts.py

@@ -0,0 +1,10 @@
+from __future__ import annotations
+
+from synth_ai.utils.prompts import *  # noqa: F401,F403
+
+try:
+    from synth_ai.utils.prompts import __all__ as __wrapped_all__  # type: ignore[attr-defined]
+except ImportError:  # pragma: no cover - defensive
+    __wrapped_all__ = []
+
+__all__ = list(__wrapped_all__)

synth_ai/_utils/task_app_state.py

@@ -0,0 +1,12 @@
+from __future__ import annotations
+
+from synth_ai.utils.task_app_state import *  # noqa: F401,F403
+
+try:
+    from synth_ai.utils.task_app_state import (
+        __all__ as __wrapped_all__,  # type: ignore[attr-defined]
+    )
+except ImportError:  # pragma: no cover - defensive
+    __wrapped_all__ = []
+
+__all__ = list(__wrapped_all__)

synth_ai/_utils/user_config.py

@@ -0,0 +1,10 @@
+from __future__ import annotations
+
+from synth_ai.utils.user_config import *  # noqa: F401,F403
+
+try:
+    from synth_ai.utils.user_config import __all__ as __wrapped_all__  # type: ignore[attr-defined]
+except ImportError:  # pragma: no cover - defensive
+    __wrapped_all__ = []
+
+__all__ = list(__wrapped_all__)

synth_ai/api/models/supported.py

@@ -20,35 +20,106 @@ QWEN3_MODELS: list[str] = [
     "Qwen/Qwen3-14B",
     "Qwen/Qwen3-30B-A3B",
     "Qwen/Qwen3-32B",
-    # Include 4B-2507 and Thinking variants used in RL
+    # 2507 baseline models
+    "Qwen/Qwen3-4B-2507",
+    # Instruct variants (no <think> tags)
+    "Qwen/Qwen3-4B-Instruct-2507",
+    "Qwen/Qwen3-4B-Instruct-2507-FP8",
+    "Qwen/Qwen3-30B-A3B-Instruct-2507",
+    "Qwen/Qwen3-30B-A3B-Instruct-2507-FP8",
+    "Qwen/Qwen3-235B-A22B-Instruct-2507",
+    "Qwen/Qwen3-235B-A22B-Instruct-2507-FP8",
+    # Thinking variants (with <think> tags)
     "Qwen/Qwen3-4B-Thinking-2507",
+    "Qwen/Qwen3-4B-Thinking-2507-FP8",
     "Qwen/Qwen3-30B-A3B-Thinking-2507",
+    "Qwen/Qwen3-30B-A3B-Thinking-2507-FP8",
     "Qwen/Qwen3-235B-A22B-Thinking-2507",
+    "Qwen/Qwen3-235B-A22B-Thinking-2507-FP8",
 ]
 
 # Qwen3 Coder family (backend-supported); text-only, SFT/inference
 QWEN3_CODER_MODELS: list[str] = [
-    # Instruct variants used for coding tasks
+    # Instruct variants used for coding tasks (no <think> tags)
     "Qwen/Qwen3-Coder-30B-A3B-Instruct",
+    "Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8",
     "Qwen/Qwen3-Coder-480B-A35B-Instruct",
+    "Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8",
+]
+
+# Qwen3-VL family (vision-language models); multimodal, SFT/inference
+QWEN3_VL_MODELS: list[str] = [
+    # Vision-Language Models (Qwen3-VL)
+    "Qwen/Qwen3-VL-2B-Instruct",
+    "Qwen/Qwen3-VL-2B-Thinking",
+    "Qwen/Qwen3-VL-4B-Instruct",
+    "Qwen/Qwen3-VL-4B-Thinking",
+    "Qwen/Qwen3-VL-8B-Instruct",
+    "Qwen/Qwen3-VL-8B-Thinking",
+    "Qwen/Qwen3-VL-30B-A3B-Instruct",
+    "Qwen/Qwen3-VL-30B-A3B-Thinking",
+    "Qwen/Qwen3-VL-32B-Instruct",
+    "Qwen/Qwen3-VL-32B-Thinking",
+    "Qwen/Qwen3-VL-235B-A22B-Instruct",
+    "Qwen/Qwen3-VL-235B-A22B-Thinking",
 ]
 
 # Training support sets
 RL_SUPPORTED_MODELS: frozenset[str] = frozenset(
     {
+        # Legacy base models
         "Qwen/Qwen3-0.6B",
         "Qwen/Qwen3-1.7B",
         "Qwen/Qwen3-4B",
-        "Qwen/Qwen3-4B-Thinking-2507",
         "Qwen/Qwen3-8B",
         "Qwen/Qwen3-14B",
         "Qwen/Qwen3-30B-A3B",
+        # 2507 models - base
+        "Qwen/Qwen3-4B-2507",
+        # 2507 models - instruct (no <think> tags)
+        "Qwen/Qwen3-4B-Instruct-2507",
+        "Qwen/Qwen3-4B-Instruct-2507-FP8",
+        "Qwen/Qwen3-30B-A3B-Instruct-2507",
+        "Qwen/Qwen3-30B-A3B-Instruct-2507-FP8",
+        # 2507 models - thinking (with <think> tags)
+        "Qwen/Qwen3-4B-Thinking-2507",
+        "Qwen/Qwen3-4B-Thinking-2507-FP8",
         "Qwen/Qwen3-30B-A3B-Thinking-2507",
+        "Qwen/Qwen3-30B-A3B-Thinking-2507-FP8",
+        # Coder instruct models
+        "Qwen/Qwen3-Coder-30B-A3B-Instruct",
+        "Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8",
+        # Vision-Language models (Qwen3-VL)
+        "Qwen/Qwen3-VL-2B-Instruct",
+        "Qwen/Qwen3-VL-2B-Thinking",
+        "Qwen/Qwen3-VL-4B-Instruct",
+        "Qwen/Qwen3-VL-4B-Thinking",
+        "Qwen/Qwen3-VL-8B-Instruct",
+        "Qwen/Qwen3-VL-8B-Thinking",
     }
 )
 
-# SFT allowlist includes core Qwen3 plus Coder family
-SFT_SUPPORTED_MODELS: frozenset[str] = frozenset([*QWEN3_MODELS, *QWEN3_CODER_MODELS])
+# SFT allowlist includes core Qwen3 plus Coder and VL families
+SFT_SUPPORTED_MODELS: frozenset[str] = frozenset([*QWEN3_MODELS, *QWEN3_CODER_MODELS, *QWEN3_VL_MODELS])
+
+# Models that support <think> reasoning tags
+THINKING_MODELS: frozenset[str] = frozenset(
+    {
+        "Qwen/Qwen3-4B-Thinking-2507",
+        "Qwen/Qwen3-4B-Thinking-2507-FP8",
+        "Qwen/Qwen3-30B-A3B-Thinking-2507",
+        "Qwen/Qwen3-30B-A3B-Thinking-2507-FP8",
+        "Qwen/Qwen3-235B-A22B-Thinking-2507",
+        "Qwen/Qwen3-235B-A22B-Thinking-2507-FP8",
+        # Vision-Language Thinking models
+        "Qwen/Qwen3-VL-2B-Thinking",
+        "Qwen/Qwen3-VL-4B-Thinking",
+        "Qwen/Qwen3-VL-8B-Thinking",
+        "Qwen/Qwen3-VL-30B-A3B-Thinking",
+        "Qwen/Qwen3-VL-32B-Thinking",
+        "Qwen/Qwen3-VL-235B-A22B-Thinking",
+    }
+)
 
 # ------------------------------------------------------------------------------
 # Lifecycle classification (core vs experimental)
@@ -58,11 +129,17 @@ SFT_SUPPORTED_MODELS: frozenset[str] = frozenset([*QWEN3_MODELS, *QWEN3_CODER_MO
 _EXPERIMENTAL_DEFAULTS: frozenset[str] = frozenset(
     {
         # Larger (>= 64B) or bleeding-edge variants are experimental by default.
+        "Qwen/Qwen3-235B-A22B-Instruct-2507",
+        "Qwen/Qwen3-235B-A22B-Instruct-2507-FP8",
         "Qwen/Qwen3-235B-A22B-Thinking-2507",
+        "Qwen/Qwen3-235B-A22B-Thinking-2507-FP8",
         "Qwen/Qwen3-Coder-480B-A35B-Instruct",
+        "Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8",
         # Thinking variants can fluctuate more rapidly.
         "Qwen/Qwen3-30B-A3B-Thinking-2507",
+        "Qwen/Qwen3-30B-A3B-Thinking-2507-FP8",
         "Qwen/Qwen3-4B-Thinking-2507",
+        "Qwen/Qwen3-4B-Thinking-2507-FP8",
     }
 )
 
@@ -77,8 +154,8 @@ def _parse_experimental_env() -> frozenset[str]:
 # Final experimental set (defaults ∪ optional env override)
 EXPERIMENTAL_MODELS: frozenset[str] = frozenset(_EXPERIMENTAL_DEFAULTS | _parse_experimental_env())
 
-# Build catalog entries for both core and coder families under unified "Qwen3"
-_ALL_QWEN3_IDS: list[str] = [*QWEN3_MODELS, *QWEN3_CODER_MODELS]
+# Build catalog entries for core, coder, and VL families under unified "Qwen3"
+_ALL_QWEN3_IDS: list[str] = [*QWEN3_MODELS, *QWEN3_CODER_MODELS, *QWEN3_VL_MODELS]
 
 CORE_MODELS: frozenset[str] = frozenset(m for m in _ALL_QWEN3_IDS if m not in EXPERIMENTAL_MODELS)
 
@@ -120,6 +197,7 @@ class SupportedModel:
     modalities: tuple[str, ...] = ()
     training_modes: tuple[str, ...] = ()
     lifecycle: str = "core"  # "core" | "experimental"
+    supports_thinking: bool = False  # Whether model supports <think> reasoning tags
 
     def as_dict(self) -> dict[str, object]:
         data: dict[str, object] = {
@@ -127,6 +205,7 @@ class SupportedModel:
             "family": self.family,
             "provider": self.provider,
             "lifecycle": self.lifecycle,
+            "supports_thinking": self.supports_thinking,
         }
         if self.modalities:
             data["modalities"] = list(self.modalities)
@@ -150,6 +229,7 @@ SUPPORTED_MODELS: tuple[SupportedModel, ...] = tuple(
             )
         ),
         lifecycle=("experimental" if model in EXPERIMENTAL_MODELS else "core"),
+        supports_thinking=(model in THINKING_MODELS),
     )
     for model in _ALL_QWEN3_IDS
 )
@@ -347,11 +427,66 @@ def training_modes_for_model(model_id: str) -> tuple[str, ...]:
     return model.training_modes
 
 
+def supports_thinking(model_id: str) -> bool:
+    """Return True if the model supports <think> reasoning tags.
+    
+    Thinking models use structured <think>...</think> tags for reasoning.
+    Instruct models do not have these tags and should not use thinking-specific logic.
+    
+    Args:
+        model_id: Model identifier (can include prefixes like 'rl:', 'fft:', etc.)
+        
+    Returns:
+        True if the model supports thinking tags, False otherwise.
+        Returns False for unsupported models.
+        
+    Example:
+        >>> supports_thinking("Qwen/Qwen3-4B-Thinking-2507")
+        True
+        >>> supports_thinking("Qwen/Qwen3-4B-Instruct-2507")
+        False
+        >>> supports_thinking("rl:Qwen/Qwen3-4B-Thinking-2507")
+        True
+    """
+    try:
+        canonical = ensure_supported_model(model_id, allow_finetuned_prefixes=True)
+    except UnsupportedModelError:
+        return False
+    model = _MODEL_BY_ID.get(canonical)
+    if not model:
+        return False
+    return model.supports_thinking
+
+
+def get_model_metadata(model_id: str) -> SupportedModel | None:
+    """Return the full metadata for a supported model, or None if not supported.
+    
+    Args:
+        model_id: Model identifier (can include prefixes like 'rl:', 'fft:', etc.)
+        
+    Returns:
+        SupportedModel instance with full metadata, or None if model is not supported.
+        
+    Example:
+        >>> meta = get_model_metadata("Qwen/Qwen3-4B-Instruct-2507")
+        >>> meta.supports_thinking
+        False
+        >>> meta.training_modes
+        ('rl', 'sft')
+    """
+    try:
+        canonical = ensure_supported_model(model_id, allow_finetuned_prefixes=True)
+    except UnsupportedModelError:
+        return None
+    return _MODEL_BY_ID.get(canonical)
+
+
 __all__ = [
     "QWEN3_MODELS",
     "QWEN3_CODER_MODELS",
     "RL_SUPPORTED_MODELS",
     "SFT_SUPPORTED_MODELS",
+    "THINKING_MODELS",
     "EXPERIMENTAL_MODELS",
     "CORE_MODELS",
     "ExperimentalWarning",
@@ -373,5 +508,7 @@ __all__ = [
     "core_model_ids",
     "format_supported_models",
     "training_modes_for_model",
+    "supports_thinking",
+    "get_model_metadata",
 ]
 

synth_ai/api/train/__init__.py

@@ -1,5 +1,17 @@
 from __future__ import annotations
 
-from .cli import register, train_command
+from typing import Any
 
 __all__ = ["register", "train_command"]
+
+
+def register(cli: Any) -> None:
+    from synth_ai.cli.train import register as _register  # local import avoids circular dependency
+
+    _register(cli)
+
+
+def train_command(*args: Any, **kwargs: Any) -> Any:
+    from synth_ai.cli.train import train_command as _train_command  # local import avoids cycle
+
+    return _train_command(*args, **kwargs)

synth_ai/api/train/builders.py

@@ -74,8 +74,14 @@ def build_rl_payload(
     idempotency: str | None,
     allow_experimental: bool | None = None,
 ) -> RLBuildResult:
+    # Load and validate config with SDK-level checks
+    from synth_ai.api.train.utils import load_toml
+    from synth_ai.cli.commands.train.validation import validate_rl_config
+    
     try:
-        rl_cfg = RLConfig.from_path(config_path)
+        raw_config = load_toml(config_path)
+        validated_config = validate_rl_config(raw_config)  # Adds defaults & validates
+        rl_cfg = RLConfig.from_mapping(validated_config)
     except ValidationError as exc:
         raise click.ClickException(_format_validation_error(config_path, exc)) from exc
 
@@ -110,8 +116,8 @@ def build_rl_payload(
             "Task app URL required (provide --task-url or set services.task_url in TOML)"
         )
 
-    model_source = (model_cfg.source or "").strip()
-    model_base = (model_cfg.base or "").strip()
+    model_source = (model_cfg.source or "").strip() if model_cfg else ""
+    model_base = (model_cfg.base or "").strip() if model_cfg else ""
     override_model = (overrides.get("model") or "").strip()
     if override_model:
         model_source = override_model