TileGym Transformers Kernel Integration
Integrate TileGym kernels into Hugging Face Transformers models without editing Transformers source.
The documentation requires user confirmation before experimentation, constrains GPU selection by UUID, mandates Docker, limits editable paths, and requires baseline comparison, correctness checks, and rollback attempts. However, least-privilege scope, data-flow disclosure, recovery boundaries, and the implications of branch/reset, GPU execution, web lookup, and autonomous code generation remain incomplete, so 9 points are deducted.
The workflow specifies environment constraints, timeouts, baselines, metrics, and failure branches, including abandoning invalid candidates. However, no reproducible implementation tests are included for this skill path, dependency availability and diagnostic failure messages are only partly covered, and static calibration caps the score at 10; 13 points are deducted.
The audience, trigger semantics, primary input (model ID), and GPU/Docker environment are fairly clear, with supported GPU architectures stated. Non-fit model ranges, no-GPU alternatives, Chinese interaction, and mainland-China network reachability are not adequately addressed, so 6 points are deducted.
The material is layered into workflow, environment, integration, auto-kernelization, and inventory-schema references, and includes version, author, license, dependencies, and risks. The recommended Examples section is missing, while changelog, maintenance ownership, and update path are unclear; 5 points are deducted.
The core goal, experiment loop, coverage/throughput constraint, and results-record format are defined and could guide integration. However, there is no directly usable implementation example for this skill, model adaptation is costly, and performance/correctness evidence is reported rather than statically reproducible, so 9 points are deducted.
The skill defines citation requirements and repository path constraints and includes a benchmark report and evaluation tasks. The cited material was not verified within the supplied evidence, and there is no committed test suite or CI covering key paths; benchmark claims do not provide independent reproduction, so 7 points are deducted.
- The skill guides branch operations, Docker/GPU experiments, autonomous code generation, and commits; confirm the worktree is clean and define acceptable changes, resource use, and rollback scope before execution.
- Performance, coverage, and correctness results are presented as report material; no experiments were executed during this static review, so independently validate them on the target model, CUDA/cuTile versions, and GPU.
- The workflow depends on an NVIDIA GPU, Docker, CUDA tooling, Transformers/Hugging Face, and potentially overseas documentation or model services; no mainland-China fallback is specified.
What does this skill do, and when should you use it?
This skill guides an AI agent in integrating TileGym kernels into Hugging Face transformers models. It uses non-intrusive monkey patching to replace selected modules, classes, and initialization, forward, and weight-loading methods without directly modifying the transformers source. The workflow also creates new cuTile kernels for uncovered PyTorch code and evaluates end-to-end correctness and performance. The skill is specifically intended for TileGym and transformers integration, and is stated to be verified with specified Claude Code, Codex, and Cursor model environments.
Guides the agent through experiment-environment preparation, existing-kernel integration, and autonomous kernelization; replaces selected transformers modules, classes, and methods for a target Hugging Face model; applies patches before model instantiation so runtime execution uses TileGym implementations; creates new cuTile kernels for uncovered PyTorch code; records reusable transformer-local kernels using FlashInfer-Bench-style Definition and Solution metadata; and summarizes the resulting work.
- An engineer validating TileGym kernel correctness in a specified Hugging Face language model.
- A developer who needs to replace model components without modifying the transformers source tree.
- A performance engineer measuring end-to-end throughput effects of TileGym kernels during LLM training or inference.
- A researcher creating cuTile kernels for PyTorch operations not yet covered in a target model.
What are this skill's strengths and limitations?
- Uses a non-intrusive monkey-patch approach instead of directly editing transformers source.
- Covers both integration of existing TileGym kernels and creation of new cuTile kernels.
- Addresses end-to-end functional correctness, kernel coverage, and throughput improvement.
- Requires structured Definition and Solution metadata for reusable kernels.
- The supplied material does not identify supported models or transformers versions.
- The detailed environment, integration, and auto-kernelization procedures are in references files that were not provided.
- No concrete performance figures or test results are included.
- Monkey patching depends on the target model's internal transformers modules and methods, so implementation changes may affect integration.
How do you install this skill?
Install the skill from NVIDIA/skills with the skills CLI:
npx skills add nvidia/skills --skill tilegym-monkey-patch-kernels-to-transformers --yes
You can target Codex explicitly:
npx skills add nvidia/skills --skill tilegym-monkey-patch-kernels-to-transformers --agent codex
The supplied source does not document a manual-copy destination.
How do you use this skill?
Give a loaded agent the skill name and target model ID, for example:
Hi, please /monkey-patch-kernels-to-transformers Qwen/Qwen3.5-0.8B.
The agent may ask clarification questions before proceeding. The supplied SKILL.md does not include the detailed environment commands, supported-model list, or failure-recovery procedures.
How does this skill compare with similar options?
The source identifies direct modification of the transformers source as an alternative; this skill instead uses non-intrusive monkey patching before model instantiation.