v3.0-alpha Week 5: Deployment tools

- Add sparseflow-audit (cost/ROI analyzer)
- Add sparseflow-convert (model conversion)
- Add sparseflow-benchmark (performance validation)
- Add deployment documentation

Tools provide:
- GPU requirement calculation
- Annual cost projections
- Carbon footprint analysis
- Accuracy impact reporting
- Performance validation
- ROI analysis

Ready for enterprise evaluation.

Author: Gourav Kumar

docs/DEPLOYMENT.md

@@ -0,0 +1,128 @@
+# SparseFlow Deployment Guide
+
+## Quick Start
+
+### 1. Check GPU Compatibility
+```bash
+python3 -c "import sparseflow; print(sparseflow.check_sparse_support())"
+```
+
+Requirements:
+- NVIDIA GPU with compute capability ≥ 8.0 (Ampere or newer)
+- CUDA 11.8+
+- PyTorch 2.0+
+
+### 2. Analyze Your Deployment
+```bash
+sparseflow-audit --model llama-7b --qps 1000
+```
+
+This shows:
+- GPU requirements (dense vs sparse)
+- Annual cost savings
+- Carbon footprint reduction
+- ROI timeline
+
+### 3. Convert Your Model
+```bash
+sparseflow-convert \
+  --input model.pt \
+  --output model_sparse.sf \
+  --validate
+```
+
+Validates 2:4 patterns and reports accuracy impact.
+
+### 4. Benchmark Performance
+```bash
+sparseflow-benchmark --size 4096x4096 --iterations 100
+```
+
+Measures actual speedup on your hardware.
+
+## Production Deployment
+
+### Step 1: Model Conversion
+```python
+import torch
+from torch import nn
+import sparseflow as sf
+
+# Load your model
+model = torch.load("model.pt")
+
+# Convert Linear layers
+for name, module in model.named_modules():
+    if isinstance(module, nn.Linear):
+        sparse_layer, diff = sf.SparseLinear.from_dense(
+            module,
+            method="magnitude",
+            return_diff=True
+        )
+        print(f"{name}: {diff['max_error']:.6f} max error")
+        
+        # Replace in model
+        # parent.layer = sparse_layer
+```
+
+### Step 2: Validate Accuracy
+```python
+# Test on validation set
+dense_accuracy = evaluate(dense_model, val_loader)
+sparse_accuracy = evaluate(sparse_model, val_loader)
+
+print(f"Accuracy delta: {sparse_accuracy - dense_accuracy:.4f}")
+```
+
+Typical accuracy impact: < 0.5% on most tasks
+
+### Step 3: Deploy
+```python
+# Inference
+x = torch.randn(1, 4096, device='cuda', dtype=torch.float16)
+y = sparse_model(x)  # 2× faster
+```
+
+## Cost Analysis
+
+### Example: LLaMA 7B @ 1000 QPS
+
+**Dense (baseline):**
+- GPUs: 16× A100-80GB
+- Annual GPU cost: $515K
+- Annual power cost: $67K
+- Total: $582K/year
+
+**SparseFlow:**
+- GPUs: 8× A100-80GB
+- Annual GPU cost: $258K
+- Annual power cost: $34K
+- Total: $292K/year
+
+**Savings:**
+- $290K/year (50% reduction)
+- 14 tons CO₂/year
+- ROI: Immediate
+
+## Troubleshooting
+
+### "CUDA not available"
+
+- Check: `nvidia-smi`
+- Install: CUDA Toolkit 11.8+
+
+### "2:4 sparse not supported"
+
+- Requires: Ampere (SM80) or newer
+- Check: `torch.cuda.get_device_capability()`
+
+### "Slower than dense"
+
+- Check batch size (need ≥32 for speedup)
+- Check GPU utilization
+- Try different tile sizes
+
+## Support
+
+- GitHub Issues: https://github.com/MapleSilicon/SparseFlow/issues
+- Email: engineering@maplesilicon.com