synth-ai 0.2.14__py3-none-any.whl → 0.2.16__py3-none-any.whl

synth_ai/demos/math/config.toml

@@ -0,0 +1,76 @@
+type = "rl"
+
+[algorithm]
+type = "online"
+method = "policy_gradient"
+variety = "gspo"
+
+[services]
+task_url = "http://localhost:8101"
+
+[model]
+base = "Qwen/Qwen3-1.7B"
+
+[policy]
+model = "Qwen/Qwen3-1.7B"
+inference_url = "http://localhost:8000/api/inference"
+max_tokens = 1028
+temperature = 0.2
+
+[data]
+split = "train"
+seed_start = 0
+episodes_per_iteration = 1280  # 8 per group * 4 groups per batch * 2 batches per step * 20 steps
+evaluation_split = "validation"
+evaluation_episodes = 50
+
+[training]
+max_turns = 1
+ops = ["agent", "env"]
+batch_size = 2
+group_size = 16
+reward_positive = 1.0
+reward_negative_no_tool = -1.0
+reward_negative_no_answer = -0.5
+learning_rate = 5e-6
+log_interval = 1
+weight_sync_interval = 1
+
+[training.weight_sync]
+enable = true
+targets = ["policy"]
+
+[compute]
+gpu_type = "H100"
+gpu_count = 4
+
+[topology]
+type = "single_node_split"
+gpus_for_vllm = 2
+gpus_for_training = 1
+gpus_for_ref = 1
+tensor_parallel = 1
+
+[vllm]
+tensor_parallel_size = 1
+max_model_len = 4096
+
+[reference]
+placement = "dedicated"
+port = 8002
+tp = 1
+health_max_wait_s = 180
+health_interval_ms = 300
+
+[rollout]
+policy_name = "math-single-step"
+max_turns = 1
+episodes_per_batch = 32  # group_size * batch_size
+
+[evaluation]
+instances = 32
+every_n_iters = 10
+seeds = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]
+
+[tags]
+experiment = "math_single_step_qwen17"

synth_ai/demos/math/deploy_modal.py

@@ -0,0 +1,54 @@
+from __future__ import annotations
+
+import os
+import subprocess
+
+
+def _parse_public_url_from_log(log_path: str) -> str | None:
+    try:
+        with open(log_path) as fh:
+            for line in fh:
+                if "modal.run" in line:
+                    return line.strip().split()[-1].rstrip("/")
+    except Exception:
+        return None
+    return None
+
+
+def deploy(script_path: str | None = None, *, env_api_key: str | None = None) -> str:
+    """
+    Deploy the Math Task App to Modal and return the public URL.
+
+    - If script_path is provided, run it (bash) and parse .last_deploy*.log for URL.
+    - Otherwise, try to call a built-in deploy() in examples.rl.task_app if available.
+    """
+    envp = os.environ.copy()
+    if env_api_key:
+        envp["ENVIRONMENT_API_KEY"] = env_api_key
+
+    if script_path:
+        script_path = os.path.abspath(script_path)
+        if not os.path.isfile(script_path):
+            raise FileNotFoundError(f"Deploy script not found: {script_path}")
+        subprocess.check_call(["bash", script_path], cwd=os.path.dirname(script_path), env=envp)
+        for name in (".last_deploy.log", ".last_deploy.dev.log", ".last_deploy.manual.log"):
+            url = _parse_public_url_from_log(os.path.join(os.path.dirname(script_path), name))
+            if url:
+                return url
+        raise RuntimeError("Deployed, but failed to extract Modal public URL from deploy logs.")
+
+    try:
+        import importlib
+
+        mod = importlib.import_module("examples.rl.task_app")
+        if hasattr(mod, "deploy"):
+            url = mod.deploy(env_api_key=env_api_key)
+            if not url:
+                raise RuntimeError("examples.rl.task_app.deploy() returned empty URL")
+            return str(url).rstrip("/")
+        raise RuntimeError("examples.rl.task_app.deploy() not found")
+    except Exception as exc:
+        raise RuntimeError(
+            f"No deploy script provided and Python-based deploy failed: {exc}. "
+            "Pass --script /path/to/deploy_task_app.sh to demo.deploy."
+        ) from exc