kernelgen-flagos: Скилл генерации и оптимизации GPU ядер

--- name: kernelgen-flagos description: > Unified GPU kernel operator generation and optimization skill. Automatically detects the target repository type (FlagGems, vLLM, or general Python/Triton) and dispatches to the appropriate specialized sub-skill. Includes operator generation, MCP-based iterative optimization, and feedback submission sub-skills. Use this skill when the user wants to generate or optimize a GPU kernel operator, create a Triton kernel, or says things like "generate an operator", "create a kernel for X", "optimize triton kernel", or "/kernelgen-flagos". argument-hint: "<operator_name> [--func-type <type>]" user-invokable: true compatibility: "Python 3.8+, PyTorch with CUDA, Triton" metadata: version: "1.0.0" author: flagos-ai category: gpu-kernel-generation tags: [kernelgen, triton, gpu, mcp, operator-generation, operator-optimization, flaggems, vllm, feedback] allowed-tools: - Bash - Bash(gh:*) - Bash(python:*) - Bash(python3:*) - Bash(command:*) - Read - Write - Edit - Glob - Grep - AskUserQuestion --- # kernelgen-flagos — Unified GPU Operator Generation Skill This is a **unified entry point** that bundles generation and optimization sub-skills into one: | Sub-skill file | Purpose | |---|---| | **Generation** | | | `kernelgen-generate.md` | Generate GPU kernels for **any** Python/Triton repository | | `kernelgen-generate-for-flaggems.md` | Specialized generation for **FlagGems** repositories | | `kernelgen-generate-for-vllm.md` | Specialized generation for **vLLM** repositories | | **Optimization** | | | `kernelgen-optimize.md` | Optimize existing Triton kernels via MCP iterative optimization (general purpose) | | `kernelgen-optimize-for-flaggems.md` | Optimize Triton operators and integrate into **FlagGems** (3 modes: built-in/external/experimental) | | `kernelgen-optimize-for-vllm.md` | Optimize Triton operators and integrate into **vLLM** (with CustomOp registration) | | **Platform Specialization** | | | `kernelgen-specialize.md` | Specialize Triton operators to target platforms (e.g., GPU → Ascend NPU) via MCP `specialize_kernel` | | `kernelgen-specialize-for-flaggems.md` | Platform specialization + **FlagGems** integration (4 modes: vendor-ops/vendor-fused/override-builtin/experimental) | | **Feedback** | | | `kernelgen-submit-feedback.md` | Submit bug reports and feedback via GitHub or email | All sub-skill files are located in the **same directory** as this `SKILL.md` file. --- ## Routing Protocol — Follow This BEFORE Doing Anything Else ### Phase 1: Detect Repository Type Use the Glob tool to check for project identity files in the current working directory: ``` Glob: pyproject.toml Glob: setup.py Glob: setup.cfg ``` Then use the Read tool to read whichever file exists. Determine the **project name** from the file contents (e.g., `name = "flag_gems"` in pyproject.toml, or `name='vllm'` in setup.py). Also use the Glob tool to check for characteristic directory structures: **FlagGems indicators** (match ANY): - `src/flag_gems/` directory exists - Project name is `flag_gems` or `flag-gems` or `FlagGems` - `import flag_gems` appears in test files **vLLM indicators** (match ANY): - `vllm/` directory exists at the repo root (with `vllm/__init__.py`) - Project name is `vllm` - `csrc/` directory exists alongside `vllm/` ### Phase 2: Dispatch to Sub-skill Based on the detection result, use the **Read tool** to read the appropriate sub-skill file from this skill's directory, then **follow the instructions in that file exactly**. **To locate the sub-skill files**: They are in the same directory as this SKILL.md. Use the Glob tool to find the path: ``` Glob: **/skills/kernelgen-flagos/kernelgen-generate.md ``` Then use the Read tool to read the matched path. #### Decision Table **Generation requests** (user wants to create/generate a new operator): | Detection Result | Action | |---|---| | FlagGems repository detected | Read `kernelgen-generate-for-flaggems.md` and follow it | | vLLM repository detected | Read `kernelgen-generate-for-vllm.md` and follow it | | Neither detected (or unknown) | Read `kernelgen-generate.md` and follow it | **Optimization requests** (user wants to optimize an existing operator, mentions "optimize", "speedup", "improve performance"): | Detection Result | Action | |---|---| | FlagGems repository detected | Read `kernelgen-optimize-for-flaggems.md` and follow it | | vLLM repository detected | Read `kernelgen-optimize-for-vllm.md` and follow it | | Neither detected (or unknown) | Read `kernelgen-optimize.md` and follow it | **Specialization requests** (user wants to migrate/specialize an operator to a different platform, mentions "specialize", "migrate to Ascend/NPU", "platform migration"): | Detection Result | Action | |---|---| | FlagGems repository detected | Read `kernelgen-specialize-for-flaggems.md` and follow it | | Neither detected (or unknown) | Read `kernelgen-specialize.md` and follow it | **Feedback requests**: | Detection Result | Action | |---|---| | User reports a bug or requests feedback submission | Read `kernelgen-submit-feedback.md` and follow it | **Important rules:** 1. **Always detect first, dispatch second.** Never skip detection. 2. **Read the entire sub-skill file** before starting execution — do not partially read it. 3. **Follow the sub-skill instructions exactly** as if they were the main SKILL.md. All steps, rules, and protocols in the sub-skill apply fully. 4. **Do not mix sub-skills.** Once you dispatch to a sub-skill, follow it to completion. 5. If the user explicitly requests a specific sub-skill (e.g., "use the FlagGems version"), honor that request regardless of auto-detection results. 6. **CRITICAL — MCP is mandatory**: ALL operator code generation MUST go through the `mcp__kernelgen-mcp__generate_kernel` MCP tool. Optimization uses `mcp__kernelgen-mcp__optimize_kernel`, and platform specialization uses `mcp__kernelgen-mcp__specialize_kernel`. NEVER generate Triton kernels, PyTorch wrappers, or operator implementations yourself. If MCP is not configured, not reachable, or fails after all retries, STOP and report the issue — do NOT fall back to writing code manually. ### Phase 3: Feedback Handling At **any point** during the workflow, if the user reports a bug, says something is broken, or asks to submit feedback about the skill: 1. Use the Read tool to read `kernelgen-submit-feedback.md` from this skill's directory. 2. Follow the feedback submission workflow described in that file. 3. After feedback is submitted, ask the user if they want to continue with the operator generation workflow or stop. --- ## Quick Reference for Users ```bash # === Generation === # Generate a kernel operator (auto-detects repo type) /kernelgen-flagos relu # Generate with explicit function type /kernelgen-flagos rms_norm --func-type normalization # === Optimization === # Optimize an existing Triton kernel (auto-detects repo type) # Just say "optimize the relu kernel" or "improve kernel performance" # The skill will automatically dispatch to the right optimization sub-skill # The skill will automatically: # - Detect if you're in a FlagGems repo → use FlagGems-specific workflow # - Detect if you're in a vLLM repo → use vLLM-specific workflow # - Otherwise → use the general-purpose workflow ``` If you encounter any issues during generation, just say "submit feedback" or "report a bug" and the skill will guide you through the feedback submission process.