synth-ai 0.2.10__tar.gz → 0.2.13.dev1__tar.gz

{synth_ai-0.2.10/synth_ai.egg-info → synth_ai-0.2.13.dev1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: synth-ai
-Version: 0.2.10
+Version: 0.2.13.dev1
 Summary: RL as a service SDK - Core AI functionality and tracing
 Author-email: Synth AI <josh@usesynth.ai>
 License-Expression: MIT
@@ -79,7 +79,7 @@ Dynamic: license-file
 
 [![Python](https://img.shields.io/badge/python-3.11+-blue)](https://www.python.org/)
 [![License](https://img.shields.io/badge/license-MIT-green)](LICENSE)
-[![PyPI](https://img.shields.io/badge/PyPI-0.2.4.dev9-orange)](https://pypi.org/project/synth-ai/)
+[![PyPI](https://img.shields.io/badge/PyPI-0.2.10-orange)](https://pypi.org/project/synth-ai/)
 ![Coverage](https://img.shields.io/badge/coverage-9.09%25-red)
 ![Tests](https://img.shields.io/badge/tests-37%2F38%20passing-brightgreen)
 ![Blacksmith CI](https://img.shields.io/badge/CI-Blacksmith%20Worker-blue)
@@ -88,6 +88,8 @@ Docs: [Synth‑AI Documentation](https://docs.usesynth.ai/welcome/introduction)
 
 Fast and effective reinforcement learning for agents, via an API
 
+> Latest: 0.2.10 published to PyPI (uv publish)
+
 ## Highlights
 
 - Easily scale gpu topologies - train on 3 a10gs or 8 H100s (multi-node available upon request)

{synth_ai-0.2.10 → synth_ai-0.2.13.dev1}/README.md

@@ -2,7 +2,7 @@
 
 [![Python](https://img.shields.io/badge/python-3.11+-blue)](https://www.python.org/)
 [![License](https://img.shields.io/badge/license-MIT-green)](LICENSE)
-[![PyPI](https://img.shields.io/badge/PyPI-0.2.4.dev9-orange)](https://pypi.org/project/synth-ai/)
+[![PyPI](https://img.shields.io/badge/PyPI-0.2.10-orange)](https://pypi.org/project/synth-ai/)
 ![Coverage](https://img.shields.io/badge/coverage-9.09%25-red)
 ![Tests](https://img.shields.io/badge/tests-37%2F38%20passing-brightgreen)
 ![Blacksmith CI](https://img.shields.io/badge/CI-Blacksmith%20Worker-blue)
@@ -11,6 +11,8 @@ Docs: [Synth‑AI Documentation](https://docs.usesynth.ai/welcome/introduction)
 
 Fast and effective reinforcement learning for agents, via an API
 
+> Latest: 0.2.10 published to PyPI (uv publish)
+
 ## Highlights
 
 - Easily scale gpu topologies - train on 3 a10gs or 8 H100s (multi-node available upon request)

synth_ai-0.2.13.dev1/examples/agora_ex/README_MoE.md

@@ -0,0 +1,224 @@
+# Agora EX - Qwen3 MoE Training
+
+## Model: Qwen3-30B-A3B (Small MoE)
+
+**Architecture:**
+- Total Parameters: 30B
+- Activated per Token: 3B (~10% activation)
+- Type: Mixture of Experts (MoE)
+- Context: 4K tokens
+
+**Why MoE for Agora EX?**
+- ✅ Efficient: Only 3B params active → faster inference
+- ✅ Powerful: 30B total capacity → better code generation
+- ✅ Cost-effective: Lower memory footprint than dense 30B
+- ✅ H200 friendly: Fits comfortably on 2x80GB setup
+
+## Hardware: 2xH200
+
+**Configuration:**
+- GPU 0 (H200): vLLM inference server
+- GPU 1 (H200): LoRA training
+- Memory: 80GB per GPU (160GB total)
+- Topology: Single-node split
+
+**Resource Usage:**
+- vLLM (MoE): ~40GB VRAM (3B active + routing)
+- Training: ~50GB VRAM (gradients + optimizer states)
+- Headroom: ~70GB available
+
+## Training Configuration
+
+### File: `configs/rl_lora_qwen3_moe_2xh200.toml`
+
+**Key Parameters:**
+```toml
+[model]
+base = "Qwen/Qwen3-30B-A3B"      # MoE with 3B activation
+trainer_mode = "lora"
+
+[lora]
+r = 16                            # LoRA rank
+target_modules = ["all-linear"]   # Wide coverage for MoE
+
+[rollout]
+episodes_per_batch = 16           # 16 episodes per batch
+max_concurrent_rollouts = 4       # Limited by human judge
+batches_per_step = 2              # 32 episodes per training step
+
+[training]
+num_epochs = 3
+iterations_per_epoch = 4          # 12 total iterations
+batch_size = 2
+group_size = 4                    # GSPO advantage estimation
+learning_rate = 3e-5              # Conservative for MoE
+```
+
+## Usage
+
+### 1. Start Task App (with Human Judge)
+
+```bash
+cd /Users/joshpurtell/Documents/GitHub/synth-ai
+
+# Set environment variables
+export GROQ_API_KEY=gsk_...                              # For inference
+export ENVIRONMENT_API_KEY=sk_env_...                    # For auth
+export EAMES_JUDGE_URL=https://eames-judge-api...        # Human judge
+
+# Serve task app
+uvx synth-ai serve agora-ex --port 8101
+```
+
+### 2. Run RL Training
+
+```bash
+# Train with MoE on 2xH200
+uvx synth-ai train \
+  --type rl \
+  --config examples/agora_ex/configs/rl_lora_qwen3_moe_2xh200.toml \
+  --task-url http://localhost:8101 \
+  --env-file backend/.env.dev
+```
+
+### 3. Monitor Progress
+
+```bash
+# Check logs
+tail -f ~/.synth-ai/logs/train_*.log
+
+# View checkpoints
+ls -lh ~/.synth-ai/checkpoints/agora-ex-qwen3-moe-rl/
+```
+
+## Expected Training Time
+
+**With Human Judge (5-30 min per eval):**
+- 12 iterations × 32 episodes = 384 rollouts
+- At 10 min average: ~64 hours (2.7 days)
+- At 4 concurrent: ~16 hours wall time
+
+**Speedup Options:**
+1. **Use AI Judge:** 10 sec/eval → 2 hours total
+2. **Increase concurrency:** More parallel rollouts
+3. **Reduce episodes:** Fewer samples per iteration
+
+## Training Timeline
+
+```
+Iteration 1:  32 rollouts → ~5 hours
+Iteration 2:  32 rollouts → ~5 hours
+...
+Iteration 12: 32 rollouts → ~5 hours
+────────────────────────────────────
+Total:       384 rollouts → ~60 hours
+
+With 4 concurrent: ~15 hours wall time
+```
+
+## Memory Usage
+
+### vLLM Server (GPU 0)
+
+```
+Model weights (MoE):        ~25GB  (BF16, 3B active + routing)
+KV cache:                   ~10GB  (batch_size=4, context=4K)
+Overhead:                   ~5GB   (vLLM runtime)
+────────────────────────────────────
+Total:                      ~40GB / 80GB
+```
+
+### Training (GPU 1)
+
+```
+LoRA adapters:              ~2GB   (r=16, all-linear)
+Gradients:                  ~10GB  (accumulation)
+Optimizer states:           ~20GB  (AdamW)
+Activations:                ~15GB  (forward pass)
+Overhead:                   ~3GB   (PyTorch)
+────────────────────────────────────
+Total:                      ~50GB / 80GB
+```
+
+## Hyperparameter Tuning
+
+### If overfitting (train reward >> eval reward):
+```toml
+[lora]
+dropout = 0.1              # Increase from 0.05
+
+[training]
+learning_rate = 1e-5       # Decrease from 3e-5
+```
+
+### If underfitting (slow improvement):
+```toml
+[training]
+learning_rate = 5e-5       # Increase from 3e-5
+gradient_accumulation_steps = 16  # More accumulation
+```
+
+### If out of memory:
+```toml
+[training]
+batch_size = 1             # Reduce from 2
+gradient_accumulation_steps = 16  # Compensate with more accumulation
+```
+
+## Comparison: MoE vs Dense
+
+| Metric | Qwen3-30B-A3B (MoE) | Qwen2.5-Coder-7B (Dense) |
+|--------|---------------------|--------------------------|
+| **Total Params** | 30B | 7B |
+| **Active Params** | 3B | 7B |
+| **Inference Speed** | ~30 tok/s | ~50 tok/s |
+| **VRAM (vLLM)** | ~40GB | ~20GB |
+| **Quality** | ⭐⭐⭐⭐⭐ | ⭐⭐⭐⭐ |
+| **Training Time** | Same | Same |
+| **Best For** | Code quality | Speed/cost |
+
+## Next Steps
+
+1. **Baseline:** Train MoE with human judge
+2. **Fast iteration:** Switch to AI judge (10s/eval)
+3. **Scale up:** Move to larger MoE (235B-A22B)
+4. **Deploy:** Export trained adapter for production
+
+## Troubleshooting
+
+### Out of Memory (OOM)
+```bash
+# Reduce batch size
+[training]
+batch_size = 1
+
+# Or reduce context length
+[vllm]
+max_model_len = 3072
+```
+
+### Slow Rollouts
+```bash
+# Switch to AI judge for development
+uvx synth-ai serve agora-ex-ai-judge --port 8102
+
+# Update task_url in training command
+--task-url http://localhost:8102
+```
+
+### Model Not Found
+```bash
+# Ensure model is in permitted list
+python3 -c "
+from backend.app.routes.clustered_training.core.algorithms.gspo.permitted_models import list_permitted_models
+print('\n'.join(list_permitted_models()))
+"
+```
+
+---
+
+**Status:** ✅ Ready for training  
+**Model:** Qwen3-30B-A3B (MoE, 3B active)  
+**Hardware:** 2xH200 (160GB total)  
+**Judge:** Human (Eames) or AI (gpt-oss-120b)
+

synth_ai-0.2.13.dev1/examples/agora_ex/__init__.py

@@ -0,0 +1,7 @@
+"""Agora EX landing page generation task app."""
+
+# Import to trigger registration
+from .task_app import agora_ex_task_app  # noqa: F401
+
+__all__ = ["agora_ex_task_app"]
+

synth_ai-0.2.13.dev1/examples/agora_ex/agora_ex.py

@@ -0,0 +1,65 @@
+"""Task App configuration for Agora EX landing page generation."""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+from synth_ai.task.apps import ModalDeploymentConfig, TaskAppEntry, register_task_app
+
+# Add this directory to path to import the task app module
+_HERE = Path(__file__).resolve().parent
+if str(_HERE) not in sys.path:
+    sys.path.insert(0, str(_HERE))
+
+from agora_ex_task_app import APP_DESCRIPTION, APP_ID, build_config
+
+# Resolve repo root for Modal mounts
+def _resolve_repo_root() -> Path:
+    """Find repo root from this file's location."""
+    candidates = [_HERE.parent.parent.parent]  # examples/agora_ex -> synth-ai
+    for candidate in candidates:
+        if (candidate / "pyproject.toml").exists():
+            return candidate
+    return _HERE.parent.parent.parent
+
+REPO_ROOT = _resolve_repo_root()
+
+# Register at module level
+register_task_app(
+    entry=TaskAppEntry(
+        app_id=APP_ID,
+        description=APP_DESCRIPTION,
+        config_factory=build_config,
+        aliases=(APP_ID, "agora-ex", "agora_ex"),
+        env_files=(),
+        modal=ModalDeploymentConfig(
+            app_name="agora-ex-task-app",
+            python_version="3.11",
+            pip_packages=(
+                "fastapi>=0.100.0",
+                "uvicorn>=0.23.0",
+                "pydantic>=2.0.0",
+                "httpx>=0.24.0",
+                "python-dotenv>=1.0.1",
+                # Tracing/DB runtime deps
+                "sqlalchemy>=2.0.42",
+                "aiosqlite>=0.21.0",
+                "greenlet>=3.2.3",
+            ),
+            extra_local_dirs=(
+                # Mount repo root so local modules resolve when deployed on Modal
+                (str(REPO_ROOT), "/opt/synth_ai_repo"),
+                (str(REPO_ROOT / "synth_ai"), "/opt/synth_ai_repo/synth_ai"),
+                (str(_HERE), "/opt/synth_ai_repo/examples/agora_ex"),
+            ),
+            secret_names=("groq-api-key", "openai-api-key"),
+            memory=8192,   # 8GB memory for inference + judge calls
+            cpu=2.0,       # 2 CPUs
+            max_containers=10,
+        ),
+    )
+)
+
+__all__ = ["build_config"]
+