Add olive mcp server

mcp/.gitignore

@@ -0,0 +1,8 @@
+.venv/
+__pycache__/
+*.pyc
+*.egg-info/
+dist/
+build/
+.olive-mcp/
+.olive-cache/

mcp/.mcp.json.example

@@ -0,0 +1,8 @@
+{
+  "mcpServers": {
+    "olive": {
+      "command": "uv",
+      "args": ["run", "--directory", "/path/to/Olive/mcp", "python", "-m", "olive_mcp"]
+    }
+  }
+}

mcp/README.md

@@ -0,0 +1,131 @@
+# Olive MCP Server
+
+MCP server for Microsoft Olive model optimization. Provides tools for model optimization, quantization, fine-tuning, and benchmarking through the [Model Context Protocol](https://modelcontextprotocol.io/).
+
+## Features
+
+| Tool | Description |
+|------|-------------|
+| `optimize` | End-to-end optimization with automatic pass scheduling |
+| `quantize` | Model quantization (RTN, GPTQ, AWQ, HQQ, and more) |
+| `finetune` | LoRA / QLoRA fine-tuning |
+| `capture_onnx_graph` | Capture ONNX graph via PyTorch Exporter or Model Builder |
+| `benchmark` | Model evaluation using lm-eval tasks |
+| `diffusion_lora` | Train LoRA adapters for diffusion models (SD 1.5, SDXL, Flux) |
+| `recommend` | Preview optimization recommendations instantly without running anything |
+| `explore_passes` | Browse available Olive passes and their parameter schemas |
+| `run_config` | Validate or run custom Olive workflow configs |
+| `detect_hardware` | Auto-detect CPU, RAM, GPU, and disk space for smart defaults |
+| `manage_outputs` | List or delete previous optimization outputs |
+| `get_job_status` | Check progress of a running job with structured phase detection |
+| `cancel_job` | Cancel a running background job |
+
+Each tool runs in an **isolated Python environment** (managed by uv) with the appropriate dependencies, so different onnxruntime variants (CPU, CUDA, DirectML, OpenVINO, etc.) never conflict.
+
+> **Note:** The `olive-config` MCP server has been merged into this server. If you were using `olive-config` separately, you can remove it and use `explore_passes` / `run_config` from this server instead.
+
+## Prerequisites
+
+- Python 3.10+
+- [uv](https://docs.astral.sh/uv/) (recommended) or pip
+
+## Installation
+
+```bash
+git clone https://github.com/microsoft/Olive.git
+cd Olive/mcp
+uv sync
+```
+
+## Configuration
+
+All MCP clients use the same server config — only the config file location differs.
+
+**Server definition:**
+
+```json
+{
+  "command": "uv",
+  "args": ["run", "--directory", "/path/to/Olive/mcp", "python", "-m", "olive_mcp"]
+}
+```
+
+> Replace `/path/to/Olive/mcp` with your actual project path.
+
+| Client | Config file | Key |
+|--------|------------|-----|
+| **VS Code (Copilot)** | `.vscode/mcp.json` | `servers.olive` |
+| **Claude Desktop** | `%APPDATA%\Claude\claude_desktop_config.json` (Win) / `~/Library/Application Support/Claude/claude_desktop_config.json` (Mac) | `mcpServers.olive` |
+| **Claude Code** | `.mcp.json` in project root | `mcpServers.olive` |
+| **Cursor** | `.cursor/mcp.json` | `mcpServers.olive` |
+| **Windsurf** | `~/.codeium/windsurf/mcp_config.json` | `mcpServers.olive` |
+
+<details>
+<summary>VS Code example (.vscode/mcp.json)</summary>
+
+```json
+{
+  "servers": {
+    "olive": {
+      "type": "stdio",
+      "command": "uv",
+      "args": ["run", "--directory", "/path/to/Olive/mcp", "python", "-m", "olive_mcp"]
+    }
+  }
+}
+```
+</details>
+
+<details>
+<summary>Claude Desktop / Claude Code / Cursor / Windsurf example</summary>
+
+```json
+{
+  "mcpServers": {
+    "olive": {
+      "command": "uv",
+      "args": ["run", "--directory", "/path/to/Olive/mcp", "python", "-m", "olive_mcp"]
+    }
+  }
+}
+```
+</details>
+
+## Usage with VS Code Copilot
+
+1. Open **Copilot Chat** panel (`Ctrl+Alt+I`) and switch to **Agent** mode
+2. Click the **Tools** icon to verify the Olive MCP tools are listed
+3. Ask Copilot, for example: *"Optimize microsoft/Phi-3-mini-4k-instruct for CPU with int4"*
+4. Copilot will ask for your confirmation before calling each MCP tool
+
+## Example Prompts
+
+```
+Optimize microsoft/Phi-3-mini-4k-instruct
+
+Quantize microsoft/Phi-3-mini-4k-instruct
+
+Fine-tune microsoft/Phi-3-mini-4k-instruct on nampdn-ai/tiny-codes
+
+Capture ONNX graph from microsoft/Phi-3-mini-4k-instruct
+
+Benchmark microsoft/Phi-3-mini-4k-instruct
+
+Train a LoRA for runwayml/stable-diffusion-v1-5 with dataset linoyts/Tuxemon
+
+What's the best way to optimize Phi-4-mini for my hardware?
+
+What passes are available for int4 quantization?
+
+Help me write a custom Olive config with OnnxQuantization and GraphSurgeries
+```
+
+## Output
+
+All optimization outputs are saved to `~/.olive-mcp/outputs/` with timestamped directories.
+
+Completed jobs include:
+- **Pass summary** — which passes ran and how long each took
+- **File sizes** — output model size (and input model size when available) for before/after comparison
+- **Structured progress** — `get_job_status` returns a `phase` field (e.g. "downloading", "quantizing", "saving") in addition to raw logs
+- **Smart error suggestions** — if a job fails, actionable suggestions are attached (e.g. "Out of GPU memory, try int4" or "CPU does not support fp16")

mcp/pyproject.toml

@@ -0,0 +1,18 @@
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[project]
+name = "olive-mcp"
+version = "0.1.0"
+description = "MCP server for Microsoft Olive model optimization"
+requires-python = ">=3.10"
+dependencies = [
+  "mcp[cli]"
+]
+
+[project.scripts]
+olive-mcp = "olive_mcp:main"
+
+[tool.hatch.build.targets.wheel]
+packages = ["src/olive_mcp"]

mcp/src/olive_mcp/__init__.py

@@ -0,0 +1,10 @@
+# -------------------------------------------------------------------------
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT License.
+# --------------------------------------------------------------------------
+import olive_mcp.tools  # noqa: F401 — registers @mcp.tool() and @mcp.prompt() on import
+from olive_mcp.server import mcp
+
+
+def main():
+    mcp.run()

mcp/src/olive_mcp/__main__.py

@@ -0,0 +1,7 @@
+# -------------------------------------------------------------------------
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT License.
+# --------------------------------------------------------------------------
+from olive_mcp import main
+
+main()

mcp/src/olive_mcp/constants.py

@@ -0,0 +1,81 @@
+# -------------------------------------------------------------------------
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT License.
+# --------------------------------------------------------------------------
+from pathlib import Path
+
+# ---------------------------------------------------------------------------
+# Paths
+# ---------------------------------------------------------------------------
+
+VENV_BASE = Path.home() / ".olive-mcp" / "venvs"
+OUTPUT_BASE = Path.home() / ".olive-mcp" / "outputs"
+WORKER_PATH = Path(__file__).parent / "worker.py"
+
+# Auto-purge venvs not used within this many days.
+_VENV_MAX_AGE_DAYS = 14
+
+# ---------------------------------------------------------------------------
+# Command names
+# ---------------------------------------------------------------------------
+
+CMD_OPTIMIZE = "optimize"
+CMD_QUANTIZE = "quantize"
+CMD_FINETUNE = "finetune"
+CMD_CAPTURE_ONNX_GRAPH = "capture_onnx_graph"
+CMD_BENCHMARK = "benchmark"
+CMD_DIFFUSION_LORA = "diffusion_lora"
+CMD_EXPLORE_PASSES = "explore_passes"
+CMD_VALIDATE_CONFIG = "validate_config"
+CMD_RUN_CONFIG = "run_config"
+
+# ---------------------------------------------------------------------------
+# Constants
+# ---------------------------------------------------------------------------
+
+SUPPORTED_PROVIDERS = [
+    "CPUExecutionProvider",
+    "CUDAExecutionProvider",
+    "DmlExecutionProvider",
+    "OpenVINOExecutionProvider",
+    "TensorrtExecutionProvider",
+    "ROCMExecutionProvider",
+    "QNNExecutionProvider",
+    "VitisAIExecutionProvider",
+    "WebGpuExecutionProvider",
+    "NvTensorRTRTXExecutionProvider",
+]
+
+SUPPORTED_PRECISIONS = [
+    "fp32",
+    "fp16",
+    "bf16",
+    "int4",
+    "int8",
+    "int16",
+    "int32",
+    "uint4",
+    "uint8",
+    "uint16",
+    "uint32",
+]
+
+SUPPORTED_QUANT_ALGORITHMS = ["rtn", "gptq", "awq", "hqq"]
+
+# Maps provider → olive-ai extras key for onnxruntime variant
+PROVIDER_TO_EXTRAS = {
+    "CPUExecutionProvider": "cpu",
+    "CUDAExecutionProvider": "gpu",
+    "TensorrtExecutionProvider": "gpu",
+    "ROCMExecutionProvider": "gpu",
+    "OpenVINOExecutionProvider": "openvino",
+    "DmlExecutionProvider": "directml",
+    "QNNExecutionProvider": "qnn",
+}
+
+# Maps provider → onnxruntime-genai variant (for ModelBuilder pass)
+PROVIDER_TO_GENAI = {
+    "CPUExecutionProvider": "onnxruntime-genai",
+    "CUDAExecutionProvider": "onnxruntime-genai-cuda",
+    "DmlExecutionProvider": "onnxruntime-genai-directml",
+}