synth-ai 0.2.8.dev4__py3-none-any.whl → 0.2.23.dev3__py3-none-any.whl

examples/gepa/multi_stage_gepa_example.toml

@@ -0,0 +1,84 @@
+# Example: Multi-stage GEPA configuration for Banking77 pipeline
+# This demonstrates how to configure GEPA for a 2-stage pipeline
+
+[prompt_learning]
+algorithm = "gepa"
+task_app_url = "http://localhost:8000"
+task_app_api_key = "test-key"
+
+# Pipeline-level initial prompt with module metadata
+[prompt_learning.initial_prompt]
+id = "banking77_pipeline_baseline"
+name = "Banking77 Two-Stage Pipeline"
+
+[prompt_learning.initial_prompt.metadata]
+# Define the stages in the pipeline
+pipeline_modules = ["query_analyzer", "classifier"]
+
+# Policy configuration (applies to all stages)
+[prompt_learning.policy]
+model = "gpt-4o-mini"
+provider = "openai"
+inference_mode = "synth_hosted"
+temperature = 0.0
+max_completion_tokens = 512
+
+# GEPA-specific configuration
+[prompt_learning.gepa]
+env_name = "banking77_pipeline"
+rng_seed = 42
+proposer_type = "dspy"
+
+# Multi-stage module configuration (instruction-only for GEPA)
+[[prompt_learning.gepa.modules]]
+module_id = "query_analyzer"
+max_instruction_slots = 2
+max_tokens = 512
+
+[[prompt_learning.gepa.modules]]
+module_id = "classifier"
+max_instruction_slots = 3
+max_tokens = 1024
+allowed_tools = ["classify", "format_output"]
+
+# Rollout configuration
+[prompt_learning.gepa.rollout]
+budget = 1000
+max_concurrent = 20
+minibatch_size = 8
+
+# Evaluation configuration
+[prompt_learning.gepa.evaluation]
+seeds = [0, 1, 2, 3, 4]
+validation_seeds = [5, 6]
+
+# Mutation configuration (LLM-guided)
+[prompt_learning.gepa.mutation]
+rate = 0.3
+llm_model = "llama3-groq-70b-8192-tool-use-preview"
+llm_provider = "groq"
+
+# Population/evolution configuration
+[prompt_learning.gepa.population]
+initial_size = 20
+num_generations = 10
+children_per_generation = 5
+crossover_rate = 0.5
+selection_pressure = 1.0
+patience_generations = 3
+
+# Archive/Pareto configuration
+[prompt_learning.gepa.archive]
+size = 64
+pareto_set_size = 64
+pareto_eps = 1e-6
+feedback_fraction = 0.5
+
+# Token/budget configuration
+[prompt_learning.gepa.token]
+max_limit = 50000
+counting_model = "gpt-4"
+enforce_pattern_limit = true
+max_spend_usd = 100.0
+
+

examples/gepa/run_gepa_banking77_pipeline.sh

@@ -0,0 +1,157 @@
+#!/bin/bash
+# Run multi-stage GEPA optimization for Banking77 pipeline
+
+set -euo pipefail
+
+echo "🧬 Running GEPA on Banking77 Pipeline"
+echo "======================================"
+
+REPO_ROOT="$(cd "$(dirname "$0")/../.." && pwd)"
+cd "$REPO_ROOT"
+
+SAVED_BACKEND_BASE_URL="${BACKEND_BASE_URL:-}"
+SAVED_SYNTH_BASE_URL="${SYNTH_BASE_URL:-}"
+
+_load_env_file() {
+    local env_file="$1"
+    if [ -f "$env_file" ]; then
+        echo "📝 Loading environment variables from $env_file..."
+        while IFS= read -r line || [ -n "$line" ]; do
+            # Skip comments and empty lines
+            [[ "$line" =~ ^[[:space:]]*# ]] && continue
+            [[ -z "${line// }" ]] && continue
+            # Only process lines with '=' character
+            [[ ! "$line" =~ = ]] && continue
+            
+            if [[ "$line" =~ ^[[:space:]]*BACKEND_BASE_URL= ]]; then
+                if [ -z "$SAVED_BACKEND_BASE_URL" ]; then
+                    export "$line" 2>/dev/null || true
+                else
+                    echo "   ⚠️  Skipping BACKEND_BASE_URL from .env (using environment value)"
+                fi
+            elif [[ "$line" =~ ^[[:space:]]*SYNTH_BASE_URL= ]]; then
+                if [ -z "$SAVED_SYNTH_BASE_URL" ]; then
+                    export "$line" 2>/dev/null || true
+                else
+                    echo "   ⚠️  Skipping SYNTH_BASE_URL from .env (using environment value)"
+                fi
+            elif [[ "$line" =~ ^[[:space:]]*[A-Za-z_][A-Za-z0-9_]*= ]]; then
+                export "$line" 2>/dev/null || true
+            fi
+        done < "$env_file"
+    fi
+}
+
+_load_env_file "$REPO_ROOT/.env"
+_load_env_file "$REPO_ROOT/examples/rl/.env"
+
+if [ -n "$SAVED_BACKEND_BASE_URL" ]; then
+    export BACKEND_BASE_URL="$SAVED_BACKEND_BASE_URL"
+    echo "✅ Using BACKEND_BASE_URL from environment: $BACKEND_BASE_URL"
+fi
+if [ -n "$SAVED_SYNTH_BASE_URL" ]; then
+    export SYNTH_BASE_URL="$SAVED_SYNTH_BASE_URL"
+    echo "✅ Using SYNTH_BASE_URL from environment: $SYNTH_BASE_URL"
+fi
+
+if [ -z "${SYNTH_API_KEY:-}" ]; then
+    echo "❌ Error: SYNTH_API_KEY not set"
+    exit 1
+fi
+
+if [ -z "${ENVIRONMENT_API_KEY:-}" ]; then
+    echo "❌ Error: ENVIRONMENT_API_KEY not set"
+    exit 1
+fi
+
+if [ -z "${GROQ_API_KEY:-}" ]; then
+    echo "⚠️  Warning: GROQ_API_KEY not set (needed for LLM-guided mutations in GEPA)."
+    read -p "Continue anyway? (y/N) " -n 1 -r
+    echo
+    if [[ ! $REPLY =~ ^[Yy]$ ]]; then
+        exit 1
+    fi
+fi
+
+if [ -z "${BACKEND_BASE_URL:-}" ]; then
+    BACKEND_BASE_URL="http://localhost:8000"
+fi
+BACKEND_URL="$BACKEND_BASE_URL"
+BACKEND_URL_NO_API="${BACKEND_URL%/api}"
+
+echo ""
+echo "🔧 Debug Info:"
+echo "   BACKEND_BASE_URL current: $BACKEND_BASE_URL"
+echo "   BACKEND_URL: $BACKEND_URL"
+echo ""
+
+echo "✅ SYNTH_API_KEY: ${SYNTH_API_KEY:0:20}..."
+echo "✅ ENVIRONMENT_API_KEY: ${ENVIRONMENT_API_KEY:0:20}..."
+if [ -n "${GROQ_API_KEY:-}" ]; then
+    echo "✅ GROQ_API_KEY: ${GROQ_API_KEY:0:20}..."
+fi
+echo "✅ Backend URL: $BACKEND_URL"
+echo ""
+
+CONFIG_PATH="examples/gepa/banking77_pipeline_gepa.toml"
+
+# Extract task_app_url from TOML
+if [ -n "${TASK_APP_URL_OVERRIDE:-}" ]; then
+    TASK_APP_URL="$TASK_APP_URL_OVERRIDE"
+    echo "📝 Using OVERRIDE task app URL: $TASK_APP_URL"
+else
+    TASK_APP_URL="$(grep "^task_app_url" "$CONFIG_PATH" | sed -E 's/^[^=]*=[[:space:]]*"([^"]*)".*/\1/' | head -1)"
+    if [ -z "$TASK_APP_URL" ]; then
+        echo "❌ ERROR: task_app_url not found in $CONFIG_PATH" >&2
+        exit 1
+    fi
+    echo "📝 Task app URL from TOML: $TASK_APP_URL"
+fi
+echo ""
+
+echo "🔍 Checking if Banking77 pipeline task app is running on ${TASK_APP_URL}..."
+if ! curl -s -f -H "X-API-Key: $ENVIRONMENT_API_KEY" "$TASK_APP_URL/health" > /dev/null 2>&1; then
+    cat <<EOF
+❌ Error: Banking77 pipeline task app is not running on ${TASK_APP_URL}
+
+Start it with:
+  modal deploy --env dev examples/task_apps/banking77_pipeline/deploy_wrapper.py
+EOF
+    exit 1
+fi
+echo "✅ Pipeline task app is healthy"
+echo ""
+
+echo "🔍 Checking backend connection to $BACKEND_URL..."
+if ! curl -s -f "$BACKEND_URL_NO_API/api/health" > /dev/null 2>&1; then
+    echo "⚠️  Warning: Cannot connect to backend at $BACKEND_URL"
+    read -p "Continue anyway? (y/N) " -n 1 -r
+    echo
+    if [[ ! $REPLY =~ ^[Yy]$ ]]; then
+        exit 1
+    fi
+else
+    echo "✅ Backend is healthy"
+fi
+echo ""
+
+echo "🚀 Starting GEPA training..."
+echo "Config: $CONFIG_PATH"
+echo ""
+echo "GEPA Flow:"
+echo "  1. Initial population: 15 prompts"
+echo "  2. Evolution: 8 generations with LLM-guided mutations"
+echo "  3. Evaluation on banking77 pipeline (classifier + calibrator)"
+echo ""
+
+export BACKEND_BASE_URL="$BACKEND_URL"
+export SYNTH_BASE_URL="$BACKEND_URL"
+
+uvx synth-ai train \
+    --type prompt_learning \
+    --config "$CONFIG_PATH" \
+    --backend "$BACKEND_URL" \
+    --poll
+
+echo ""
+echo "✅ GEPA pipeline training complete!"

examples/multi_step/SFT_README.md

@@ -0,0 +1,147 @@
+# SFT Training for Qwen3-Coder-30B with LoRA
+
+Supervised Fine-Tuning configuration for the same 30B MoE model used in RL training.
+
+## Configuration Overview
+
+**Model:** `Qwen/Qwen3-Coder-30B-A3B-Instruct` (Mixture of Experts)
+
+**Hardware:** 4x H200 GPUs (561GB total VRAM)
+
+**Parallelism Strategy:**
+- **Tensor Parallel (TP)**: 2 GPUs - Splits the model across 2 GPUs for inference/forward pass
+- **Data Parallel (DP)**: 2 GPUs - Splits batches across 2 GPUs for training throughput
+
+**LoRA Configuration:**
+- Rank (r): 16
+- Alpha: 32
+- Dropout: 0.05
+- Target modules: `["all-linear"]` - Applies LoRA to all linear layers
+
+## Memory Breakdown per GPU
+
+With 4x H200 (141GB each):
+
+**Model Split (TP=2):**
+- 2 GPUs hold the base model (70GB each)
+- ~70GB free per GPU for activations and gradients
+
+**Training (DP=2):**
+- 2 GPUs process different batches
+- LoRA adapters: ~5-10GB per GPU
+- Gradients/optimizer states: ~20-30GB per GPU
+- **Total per training GPU: ~50-60GB** ✅
+
+## Quick Start
+
+### 1. Prepare Your Dataset
+
+Your dataset should be in JSONL format with conversation turns:
+
+```jsonl
+{"messages": [{"role": "system", "content": "..."}, {"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}
+{"messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}
+```
+
+### 2. Run Training
+
+```bash
+# Using the helper script
+./examples/multi_step/run_sft_qwen30b.sh path/to/your/dataset.jsonl
+
+# Or directly with synth-ai CLI
+uvx synth-ai train \
+  --type sft \
+  --config examples/multi_step/configs/crafter_sft_qwen30b_lora.toml \
+  --dataset path/to/your/dataset.jsonl \
+  --env-file backend/.env.dev
+```
+
+### 3. Monitor Training
+
+Check the Synth dashboard for:
+- Training loss curve
+- Validation metrics (if validation set provided)
+- GPU utilization
+- Training throughput (tokens/sec)
+
+## Hyperparameters
+
+**Batch Configuration:**
+- Per-device batch size: 1
+- Gradient accumulation: 64 steps
+- **Effective global batch size: 128** (1 × 64 × 2 GPUs)
+
+**Learning Rate:**
+- Initial LR: 5e-6
+- Warmup ratio: 3%
+- Schedule: Linear decay
+
+**Sequence Length:** 4096 tokens
+
+**Training:**
+- Epochs: 1
+- Mixed precision: BF16
+- DeepSpeed: Stage 2 (optimizer state sharding)
+- Activation checkpointing: Enabled
+
+## Configuration File Structure
+
+```toml
+[algorithm]
+type = "offline"        # Supervised (not RL)
+method = "sft"          # Supervised fine-tuning
+variety = "lora"        # Using LoRA adapters
+
+[compute]
+gpu_type = "H200"
+gpu_count = 4
+
+[data.topology]
+tensor_parallel = 2     # Split model across 2 GPUs
+data_parallel = 2       # Split batches across 2 GPUs
+
+[training]
+mode = "lora"
+use_qlora = true        # Quantized LoRA (4-bit base model)
+
+[lora]
+r = 16                  # LoRA rank
+alpha = 32              # LoRA scaling
+dropout = 0.05
+target_modules = ["all-linear"]  # Apply to all linear layers
+```
+
+## Comparison with RL Config
+
+| Aspect | SFT | RL |
+|--------|-----|-----|
+| Purpose | Supervised learning | Reinforcement learning |
+| Data | Labeled examples | Environment interactions |
+| Topology | TP=2, DP=2 | Split: 2 inference + 2 training |
+| Batch size | 128 (effective) | Variable (episode-based) |
+| Training | Standard backprop | Policy gradient (GSPO) |
+
+## Tips
+
+1. **Start Small:** Test with a small dataset first to verify the pipeline
+2. **Validation:** Add a validation set to monitor overfitting
+3. **Checkpointing:** Training saves checkpoints every 100 steps
+4. **Resume:** Can resume from checkpoint if training is interrupted
+5. **Inference:** After training, use the LoRA adapter with the base model
+
+## Output
+
+After training completes, you'll get:
+- LoRA adapter weights (saved to volume)
+- Training metrics and logs
+- Best checkpoint (based on validation loss)
+- Model ready for inference or RL initialization
+
+## Next Steps
+
+1. **Evaluate:** Test your fine-tuned model on held-out data
+2. **RL Training:** Use this as initialization for RL (`init_from_sft = true`)
+3. **Deploy:** Load LoRA adapter for inference
+4. **Iterate:** Adjust hyperparameters based on performance
+

examples/multi_step/configs/README_verilog_rl.md

@@ -0,0 +1,77 @@
+# Verilog RL with LoRA (Qwen3-0.6B)
+
+## Quick Start
+
+1. **Deploy Verilog Task App**:
+```bash
+cd synth-ai
+uvx synth-ai modal-serve grpo-verilog
+```
+Note the Modal URL and update `task_url` in `verilog_rl_lora.toml`.
+
+2. **Run Training**:
+```bash
+uvx synth-ai rl run --config examples/multi_step/configs/verilog_rl_lora.toml
+```
+
+## Configuration Overview
+
+### **Key Adaptations from Crafter**:
+
+- **Model**: `Qwen/Qwen3-0.6B` (✅ proven in SFT configs)
+- **Environment**: `verilog` instead of `crafter`
+- **Steps**: 15 turns (vs Crafter's 10) for compilation workflows
+- **Rewards**: Adjusted for sparser Verilog rewards (0.5 vs 1.0 indicator_lambda)
+- **Rubrics**: Verilog-specific judging criteria
+
+### **Hardware Requirements** (Standard RL setup):
+- ✅ **2x H100 GPUs** (vLLM inference + LoRA training split)
+- ✅ **No tensor parallelism** needed for 0.6B model
+- ✅ **4x faster inference** than 32B model
+- ✅ **Same compute pattern** as Crafter (just smaller model)
+
+### **Expected Workflow**:
+1. Agent writes Verilog code (`write_file`)
+2. Compiles to check syntax (`compile`)
+3. Simulates to verify behavior (`simulate`)
+4. Submits if tests pass (`submit`)
+5. **Rewards**: +1.0 for compilation success, +10.0 for passing tests
+
+## Rubric Design
+
+### **Event Rewards** (per decision):
+- **Compilation Success**: 70% weight (1.0 for success, 0.0 for errors)
+- **Process Efficiency**: 30% weight (penalizes redundant operations)
+
+### **Outcome Rewards** (final score):
+- **Tests Passed**: 80% weight (full credit when all tests pass)
+- **Design Quality**: 20% weight (code clarity, documentation)
+
+## Troubleshooting
+
+### **If training fails**:
+1. Check Modal URL in `task_url` field
+2. Verify `GROQ_API_KEY` for inference
+3. Ensure `OPENAI_API_KEY` for judging
+
+### **Memory issues** (unlikely with 0.6B):
+- Reduce `batch_size` to 2
+- Set `gradient_accumulation_steps = 2`
+- Verify 2x GPU split is working (vLLM on GPU 0, training on GPU 1)
+
+### **Slow training**:
+- Increase `episodes_per_batch` to 6-8
+- Check network latency to Modal task app
+
+## Expected Results
+
+- **Convergence**: Should learn basic compilation workflow in 1-2 hours
+- **Success Rate**: 20-40% initial test pass rate (improves with training)
+- **Learning**: Agent learns to debug compilation errors and write correct Verilog
+
+## Next Steps
+
+1. **Monitor reward progression** in training logs
+2. **Adjust rubrics** if agent struggles with compilation errors
+3. **Scale to 8B model** once 0.6B baseline works
+4. **Add domain-specific fine-tuning** for Verilog syntax

examples/multi_step/configs/VERILOG_REWARDS.md

@@ -0,0 +1,103 @@
+# Verilog Reward Structure (Normalized to 1.0)
+
+## Overview
+All rewards in the Verilog task app are normalized so the maximum possible reward is **1.0**.
+
+## Reward Components
+
+### 1. Step Penalty: **-0.001** per step
+- Applied to every action taken
+- Encourages efficient solutions
+- Normalized from `-0.01` (original)
+
+### 2. Compile Success: **+0.01**
+- Awarded when `iverilog` compilation succeeds (returncode 0)
+- Validates syntax correctness
+- Normalized from `+0.1` (original)
+
+### 3. Simulation Pass: **+0.1**
+- Awarded when `vvp` simulation passes all tests
+- Validates behavioral correctness
+- Normalized from `+1.0` (original)
+
+### 4. Submit Success: **+1.0** (maximum reward)
+- Awarded when final submission passes all verification tests
+- This is the goal state
+- Normalized from `+10.0` (original)
+
+## Typical Reward Trajectories
+
+### ✅ Optimal Path (3 steps)
+```
+Step 1: write_file        → -0.001
+Step 2: compile (success) → +0.01 - 0.001 = +0.009
+Step 3: simulate (pass)   → +0.1 - 0.001 = +0.099
+Total:                      ~0.107
+```
+
+### ✅ Good Path (4 steps with submit)
+```
+Step 1: write_file        → -0.001
+Step 2: compile (success) → +0.009
+Step 3: simulate (pass)   → +0.099
+Step 4: submit (success)  → +1.0 - 0.001 = +0.999
+Total:                      ~1.106
+```
+*Note: Can exceed 1.0 if intermediate rewards stack with final submit*
+
+### ❌ Failure Path (compilation errors)
+```
+Step 1: write_file        → -0.001
+Step 2: compile (fail)    → -0.001
+Step 3: write_file (fix)  → -0.001
+Step 4: compile (success) → +0.009
+Step 5: simulate (pass)   → +0.099
+Total:                      ~0.105
+```
+
+## Implementation Details
+
+### Location
+- **Reward components**: `synth_ai/environments/examples/verilog/engine.py`
+  - `VerilogCompileSuccessComponent`: +0.01
+  - `VerilogSimulationPassComponent`: +0.1
+  - `VerilogSubmitSuccessComponent`: +1.0
+  - `VerilogStepPenaltyComponent`: -0.001
+
+### Normalization Ratio
+All rewards were divided by **10.0** to normalize:
+- Original max: ~10.0
+- Normalized max: ~1.0
+- Ratio: 10.0
+
+## Why Normalize?
+
+1. **Consistency**: Makes it easier to compare rewards across different task types
+2. **RL Training**: Standard reward scales improve learning stability
+3. **Interpretability**: Rewards as percentages (0.0 to 1.0) are intuitive
+4. **Judge Compatibility**: Rubric scores typically range 0-1, making blending easier
+
+## Testing
+```bash
+# Run eval to verify normalized rewards
+uvx synth-ai eval --config examples/multi_step/configs/verilog_eval_groq_qwen32b.toml
+```
+
+Expected output for successful rollout:
+- `mean_return` ≈ 0.1 (if only compile+simulate)
+- `mean_return` ≈ 1.0+ (if full submit success)
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+