synth-ai 0.2.9.dev1__tar.gz → 0.2.9.dev3__tar.gz

{synth_ai-0.2.9.dev1/synth_ai.egg-info → synth_ai-0.2.9.dev3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: synth-ai
-Version: 0.2.9.dev1
+Version: 0.2.9.dev3
 Summary: RL as a service SDK - Core AI functionality and tracing
 Author-email: Synth AI <josh@usesynth.ai>
 License-Expression: MIT

{synth_ai-0.2.9.dev1 → synth_ai-0.2.9.dev3}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "synth-ai"
-version = "0.2.9.dev1"
+version = "0.2.9.dev3"
 description = "RL as a service SDK - Core AI functionality and tracing"
 authors = [{name = "Synth AI", email = "josh@usesynth.ai"}]
 license = "MIT"

{synth_ai-0.2.9.dev1 → synth_ai-0.2.9.dev3}/synth_ai/demo_registry.py

@@ -62,7 +62,7 @@ DEMO_TEMPLATES: tuple[DemoTemplate, ...] = (
                 make_executable=True,
             ),
             CopySpec(
-                "examples/rl/configs/rl_from_base_qwen17.toml",
+                "synth_ai/demos/demo_task_apps/math/config.toml",
                 "configs/rl_from_base_qwen17.toml",
             ),
         ),
@@ -78,7 +78,7 @@ DEMO_TEMPLATES: tuple[DemoTemplate, ...] = (
             "# Optional: set to 'prod' to use production names",
             "ENVIRONMENT=",
         ),
-        config_source="examples/rl/configs/rl_from_base_qwen17.toml",
+        config_source="synth_ai/demos/demo_task_apps/math/config.toml",
         requires_modal=True,
         post_copy=lambda root: _postprocess_math_modal(root),
     ),
@@ -88,19 +88,19 @@ DEMO_TEMPLATES: tuple[DemoTemplate, ...] = (
         description="Lightweight wrapper around synth_ai.task.apps.grpo_crafter for local experimentation.",
         copy_specs=(
             CopySpec(
-                "examples/warming_up_to_rl/task_app/grpo_crafter_task_app.py",
+                "synth_ai/demos/demo_task_apps/crafter/grpo_crafter_task_app.py",
                 "task_app.py",
             ),
             CopySpec(
-                "examples/warming_up_to_rl/task_app/README.md",
+                "synth_ai/demos/demo_task_apps/crafter/README.md",
                 "README.md",
             ),
             CopySpec(
-                "examples/warming_up_to_rl/configs/rl_from_base_qwen4b.toml",
+                "synth_ai/demos/demo_task_apps/crafter/configs/rl_from_base_qwen4b.toml",
                 "configs/rl_from_base_qwen4b.toml",
             ),
             CopySpec(
-                "examples/warming_up_to_rl/configs/crafter_fft_4b.toml",
+                "synth_ai/demos/demo_task_apps/crafter/configs/crafter_fft_4b.toml",
                 "configs/crafter_fft_4b.toml",
             ),
         ),
@@ -112,7 +112,7 @@ DEMO_TEMPLATES: tuple[DemoTemplate, ...] = (
             "# Optional: URL for existing Crafter task app",
             "TASK_APP_BASE_URL=",
         ),
-        config_source="examples/warming_up_to_rl/configs/rl_from_base_qwen4b.toml",
+        config_source="synth_ai/demos/demo_task_apps/crafter/configs/rl_from_base_qwen4b.toml",
         config_destination="demo_config.toml",
         requires_modal=False,
         post_copy=lambda root: _postprocess_crafter_local(root),

synth_ai-0.2.9.dev3/synth_ai/demos/demo_task_apps/crafter/__init__.py

@@ -0,0 +1 @@
+# Crafter demo task app

synth_ai-0.2.9.dev3/synth_ai/demos/demo_task_apps/crafter/configs/crafter_fft_4b.toml

@@ -0,0 +1,54 @@
+# FFT job config for Qwen/Qwen3-4B on Crafter SFT dataset
+
+[algorithm]
+type = "offline"
+method = "supervised_finetune"
+variety = "fft"
+
+
+[job]
+model = "Qwen/Qwen3-4B"
+# Limit training to the first 100 conversations (export a 100-row JSONL and point to it here)
+# data = "../ft_data/qwen3_32b_ach_ge3_raw_filtered.head100.jsonl"
+
+[compute]
+# Adjust as needed for your quota
+gpu_type = "H100"
+gpu_count = 1
+nodes = 1
+
+[data]
+# Optional topology metadata (left empty for now)
+topology = {}
+
+# Optional local validation dataset path (JSONL). If set, the client will upload
+# this file and wire up validation so the frontend can display val.loss.
+validation_path = "../ft_data/qwen3_32b_ach_ge3_raw_filtered.tokens_1000000_seed_123.val_2000.jsonl"
+
+[training]
+mode = "sft_offline"
+use_qlora = false
+
+# Validation settings to emit val.loss on the frontend
+[training.validation]
+enabled = true
+evaluation_strategy = "steps"
+eval_steps = 20
+save_best_model_at_end = true
+metric_for_best_model = "val.loss"
+greater_is_better = false
+
+[hyperparameters]
+# Minimal safe defaults; backend can override
+n_epochs = 1
+batch_size = 1
+gradient_accumulation_steps = 64
+sequence_length = 4096
+learning_rate = 5e-6
+warmup_ratio = 0.03
+train_kind = "fft"
+
+# Optional parallelism block example
+#[hyperparameters.parallelism]
+# tensor_parallel_size = 1
+# pipeline_parallel_size = 1

synth_ai-0.2.9.dev3/synth_ai/demos/demo_task_apps/crafter/configs/rl_from_base_qwen4b.toml

@@ -0,0 +1,73 @@
+# RL training starting from base Qwen/Qwen3-4B (TOML-only model selection)
+
+[algorithm]
+type = "online"
+method = "policy_gradient"
+variety = "gspo"
+
+
+[services]
+task_url = "https://synth-laboratories--grpo-crafter-task-app-final-warming--ceb5b2.modal.run"
+
+[compute]
+# Cluster shape for RL pipeline
+gpu_type = "H100"
+gpu_count = 8
+
+[topology]
+# Split GPUs across vLLM, training, and reference
+# Must sum to compute.gpu_count
+type = "single_node_split"
+gpus_for_vllm = 4
+gpus_for_training = 3
+gpus_for_ref = 1
+tensor_parallel = 4
+
+[vllm]
+# Serving tensor parallel size
+tensor_parallel_size = 4
+max_model_len = 8192
+
+[reference]
+# Required by trainer/runtime; ensures dedicated/scoped scoring server config exists
+placement = "dedicated"
+port = 8002
+tp = 1
+health_max_wait_s = 180
+health_interval_ms = 300
+
+[model]
+# Base model start
+base = "Qwen/Qwen3-4B"
+label = "crafter-rl-from-base"
+
+[rollout]
+max_turns = 10
+episodes_per_batch = 64
+policy_name = "crafter"
+
+[evaluation]
+# Run baseline evaluation over the first 100 seeds every 20 training iterations
+instances = 10
+every_n_iters = 10
+seeds = [
+  0, 1, 2, 3, 4, 5, 6, 7, 8, 9,
+]
+
+[training]
+log_interval = 1
+weight_sync_interval = 1
+# Additional RL hyperparameters can go here
+
+# Stepwise rewards (Crafter decision-level)
+step_rewards_enabled = true
+step_rewards_mode = "decision_stepwise"  # "off" | "decision_stepwise" | "env_sparse"
+step_rewards_beta = 0.0
+step_rewards_indicator_lambda = 1.0
+# Optional selector for decision scalar: "unique" | "absolute" (default unique)
+event_rewards_kind = "unique"
+
+[training.weight_sync]
+enable  = true
+targets = ["policy"]
+weight_sync_interval = 1

synth_ai-0.2.9.dev3/synth_ai/demos/demo_task_apps/crafter/grpo_crafter_task_app.py

@@ -0,0 +1,165 @@
+
+"""Compatibility wrapper for the GRPO Crafter task app.
+
+This module now delegates to the shared TaskAppConfig defined in
+`synth_ai.task.apps.grpo_crafter`. It is kept for legacy usage (running the
+file directly or targeting `fastapi_app` from external tooling). Prefer using
+`uvx synth-ai serve grpo-crafter` for local development and testing.
+"""
+
+from __future__ import annotations
+
+import argparse
+from pathlib import Path
+
+from fastapi.exceptions import RequestValidationError
+from fastapi.responses import JSONResponse
+from starlette.requests import Request
+
+from synth_ai.task.apps import ModalDeploymentConfig, registry
+from synth_ai.task.apps.grpo_crafter import build_config
+from synth_ai.task.auth import is_api_key_header_authorized, normalize_environment_api_key
+from synth_ai.task.server import TaskAppConfig, create_task_app, run_task_app
+
+
+APP_ID = "grpo-crafter"
+
+
+_BASE_CONFIG = build_config()
+TASK_APP_CONFIG = TaskAppConfig(
+    app_id="grpo-crafter",
+    name=_BASE_CONFIG.name,
+    description=_BASE_CONFIG.description,
+    base_task_info=_BASE_CONFIG.base_task_info,
+    describe_taskset=_BASE_CONFIG.describe_taskset,
+    provide_task_instances=_BASE_CONFIG.provide_task_instances,
+    rollout=_BASE_CONFIG.rollout,
+    dataset_registry=_BASE_CONFIG.dataset_registry,
+    rubrics=_BASE_CONFIG.rubrics,
+    proxy=_BASE_CONFIG.proxy,
+    routers=_BASE_CONFIG.routers,
+    middleware=_BASE_CONFIG.middleware,
+    app_state=_BASE_CONFIG.app_state,
+    require_api_key=_BASE_CONFIG.require_api_key,
+    expose_debug_env=_BASE_CONFIG.expose_debug_env,
+    cors_origins=_BASE_CONFIG.cors_origins,
+    startup_hooks=_BASE_CONFIG.startup_hooks,
+    shutdown_hooks=_BASE_CONFIG.shutdown_hooks,
+)
+
+try:
+    _REGISTERED_ENTRY = registry.get(APP_ID)
+except Exception:  # pragma: no cover - registry unavailable in some contexts
+    MODAL_DEPLOYMENT: ModalDeploymentConfig | None = None
+    ENV_FILES: tuple[str, ...] = ()
+else:
+    MODAL_DEPLOYMENT = _REGISTERED_ENTRY.modal
+    ENV_FILES = tuple(_REGISTERED_ENTRY.env_files)
+
+
+def build_task_app_config() -> TaskAppConfig:
+    """Return a fresh TaskAppConfig for this wrapper."""
+
+    return TASK_APP_CONFIG.clone()
+
+
+def fastapi_app():
+    """Return the FastAPI application for Modal or other ASGI hosts."""
+
+    app = create_task_app(build_task_app_config())
+
+    # Replace default health endpoints so we can permit soft auth failures and log 422s.
+    filtered_routes = []
+    for route in app.router.routes:
+        path = getattr(route, "path", None)
+        methods = getattr(route, "methods", set()) or set()
+        if path in {"/health", "/health/rollout"} and "GET" in methods:
+            continue
+        filtered_routes.append(route)
+    app.router.routes = filtered_routes
+
+    def _log_env_key_prefix(source: str, env_key: str | None) -> str | None:
+        if not env_key:
+            return None
+        prefix = env_key[: max(1, len(env_key) // 2)]
+        print(f"[{source}] expected ENVIRONMENT_API_KEY prefix: {prefix}")
+        return prefix
+
+    @app.get("/health")
+    async def health(request: Request):
+        env_key = normalize_environment_api_key()
+        if not env_key:
+            return JSONResponse(
+                status_code=503,
+                content={"status": "unhealthy", "detail": "Missing ENVIRONMENT_API_KEY"},
+            )
+        if not is_api_key_header_authorized(request):
+            prefix = _log_env_key_prefix("health", env_key)
+            content = {"status": "healthy", "authorized": False}
+            if prefix:
+                content["expected_api_key_prefix"] = prefix
+            return JSONResponse(status_code=200, content=content)
+        return {"status": "healthy", "authorized": True}
+
+    @app.get("/health/rollout")
+    async def health_rollout(request: Request):
+        env_key = normalize_environment_api_key()
+        if not env_key:
+            return JSONResponse(
+                status_code=503,
+                content={"status": "unhealthy", "detail": "Missing ENVIRONMENT_API_KEY"},
+            )
+        if not is_api_key_header_authorized(request):
+            prefix = _log_env_key_prefix("health/rollout", env_key)
+            content = {"status": "healthy", "authorized": False}
+            if prefix:
+                content["expected_api_key_prefix"] = prefix
+            return JSONResponse(status_code=200, content=content)
+        return {"ok": True, "authorized": True}
+
+    @app.exception_handler(RequestValidationError)
+    async def _on_validation_error(request: Request, exc: RequestValidationError):
+        try:
+            hdr = request.headers
+            snapshot = {
+                "path": str(getattr(request, "url").path),
+                "have_x_api_key": bool(hdr.get("x-api-key")),
+                "have_x_api_keys": bool(hdr.get("x-api-keys")),
+                "have_authorization": bool(hdr.get("authorization")),
+                "errors": exc.errors()[:5],
+            }
+            print("[422] validation", snapshot, flush=True)
+        except Exception:
+            pass
+        return JSONResponse(
+            status_code=422,
+            content={"status": "invalid", "detail": exc.errors()[:5]},
+        )
+
+    return app
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Run the Crafter task app locally")
+    parser.add_argument("--host", default="0.0.0.0")
+    parser.add_argument("--port", type=int, default=8001)
+    parser.add_argument("--reload", action="store_true", help="Enable uvicorn autoreload")
+    parser.add_argument(
+        "--env-file",
+        action="append",
+        default=[],
+        help="Additional .env files to load before startup",
+    )
+    args = parser.parse_args()
+
+    default_env = Path(__file__).resolve().parents[4] / "backend" / ".env.dev"
+    env_files = [str(default_env)] if default_env.exists() else []
+    env_files.extend(args.env_file or [])
+
+    run_task_app(
+        build_task_app_config,
+        host=args.host,
+        port=args.port,
+        reload=args.reload,
+        env_files=env_files,
+    )

{synth_ai-0.2.9.dev1 → synth_ai-0.2.9.dev3/synth_ai.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: synth-ai
-Version: 0.2.9.dev1
+Version: 0.2.9.dev3
 Summary: RL as a service SDK - Core AI functionality and tracing
 Author-email: Synth AI <josh@usesynth.ai>
 License-Expression: MIT

{synth_ai-0.2.9.dev1 → synth_ai-0.2.9.dev3}/synth_ai.egg-info/SOURCES.txt

@@ -44,6 +44,10 @@ synth_ai/demos/core/__init__.py
 synth_ai/demos/core/cli.py
 synth_ai/demos/demo_task_apps/__init__.py
 synth_ai/demos/demo_task_apps/core.py
+synth_ai/demos/demo_task_apps/crafter/__init__.py
+synth_ai/demos/demo_task_apps/crafter/grpo_crafter_task_app.py
+synth_ai/demos/demo_task_apps/crafter/configs/crafter_fft_4b.toml
+synth_ai/demos/demo_task_apps/crafter/configs/rl_from_base_qwen4b.toml
 synth_ai/demos/demo_task_apps/math/__init__.py
 synth_ai/demos/demo_task_apps/math/_common.py
 synth_ai/demos/demo_task_apps/math/app.py