| name | makora-skill |
| description | Format, evaluate, and profile GPU kernel code using the Makora API |
| allowed-tools | Bash(*), Read(*), Glob(*) |
| user-invocable | true |
Makora GPU Kernel Skill
When This Skill Applies
Use this skill when the user mentions:
- Evaluating or benchmarking GPU code
- Profiling GPU code
- Running expert-generate on an existing kernel file
- Searching Makora indexed documents with document-search
- Converting code into Makora file format
- Makora or the Makora API/CLI
Important Defaults
- Always use
torch.float32 dtype unless the problem explicitly specifies otherwise (e.g., torch.float16 or torch.bfloat16)
Workflow
- ALWAYS read FORMATTING.md first when creating or converting files - this defines the exact required structure
- Identify reference (problem definition) and candidate (solution) implementation files
- If files need conversion to the Makora format, follow the structure in FORMATTING.md exactly
- Refer to EVALUATION.md for command usage and examples
- Refer to GPUS.md for available devices
- Run the evaluation command
- Refer to PROFILING.md for profiling usage and output details
- Refer to EXPERT_GENERATE.md when running
makora expert-generate
- Refer to DOCUMENT_SEARCH.md when running
makora document-search
Reference Files
- FORMATTING.md: Required file formats for problem and solution files
- EVALUATION.md: Command usage and CLI examples
- PROFILING.md: How to profile kernels for performance bottlenecks
- EXPERT_GENERATE.md: Command usage for
makora expert-generate
- DOCUMENT_SEARCH.md: Command usage for
makora document-search
- GPUS.md: Available GPU devices
- hardware/NVIDIA.md: Device-selection guidance and valid device names
- languages/TRITON.md: Required Triton solution file structure
- languages/CUDA.md: Required CUDA solution file structure