القائمة
MaxEvolve Triton/PyTorch to JAX/Pallas porting with GpuToJax 12-step pipeline, interpret=True validation, MAP-Elites evolution, and Colab/SSH hardware profiling.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
Manage Google Colab sessions and execute code on remote Colab VMs via the colab-cli.
Use when you need to generate a synthetic dataset. You can use Afterimage Python package to create, conversational, tool-calling, multi-choice questions, or other types of structured output datasets for SFT, or preference data for DPO.
Schedule deferred or recurring work via the isanagent cron tool—add/remove jobs with correct runtime context fields.
Configure long wall clocks and use execution_run_background + job polling for ML or hours-long runs without blocking the agent loop.
Run numeric or ML code through isanagent execution tools without flooding context—artifacts, incremental runs, sandbox paths.
Turn papers and docs into a concrete training or evaluation recipe (datasets, metrics, hyperparameters, fine-tuning code).
| name | kernel-porting |
| description | MaxEvolve Triton/PyTorch to JAX/Pallas porting with GpuToJax 12-step pipeline, interpret=True validation, MAP-Elites evolution, and Colab/SSH hardware profiling. |
| requires | {"bins":["uv"],"env":[]} |
| always | false |
Native isanagent workflow combining MaxKernel-style multi-agent translation with AlphaEvolve-style MAP-Elites search.
User mentions Triton, Pallas, custom kernels, JAX porting, or MaxEvolve. Delegate to kernel_orchestrator via subagent_spawn(agent="kernel_orchestrator", ...).
BlockSpec + pallas_call orchestration only.jax.custom_vjp for production backward passes — do not rely on bare jax.grad for optimized kernels.interpret=True CPU gate before any hardware run.PrefetchScalarGridSpec (see kernels/reference/Triton To Pallas Conversion.md).kernels/projects/{project_id}/
source/ # input Triton/PyTorch
SIMPLIFICATION_PLAN.md
organized_gpu.py
converted_jax.py
test_correctness.py
profile_script.py
database/map_elites.json
database/lineage.jsonl
candidates/queue.jsonl
artifacts/
REPORT.md
Initialize with kernel_db_init(project_id=..., target_hardware=cpu_interpret|gpu_hopper|tpu_v5e|tpu_v6e).
kernel_db_init, todo_write, copy source into project.subagent_plan_execute using template .agents/kernel-porting/gpu_to_jax_plan.json (each step agent: gpu_to_jax). HITL: ask_user after step 2 (plan approval).test_generator then test_runner; validators via execution_run or exec.evolve_orchestrator + kernel_mutator; kernel_db_sample / kernel_db_insert; background evolve_runner.py.best_kernel.py, write REPORT.md.uv run + PEP 723)All scripts under skills/kernel-porting/scripts/ declare dependencies inline (PEP 723). Run with uv run — the workspace already uses UV via [harness.execution] local_python_runtime = "uv_managed".
Set JAX_PLATFORMS=cpu for local interpret/correctness.
From project directory (execution_run with cwd_relative=kernels/projects/{id}):
uv run ../../skills/kernel-porting/scripts/validators/jax_syntax_check.py converted_jax.py
uv run ../../skills/kernel-porting/scripts/validators/compile_check.py converted_jax.py
uv run ../../skills/kernel-porting/scripts/validators/shape_check.py converted_jax.py
uv run ../../skills/kernel-porting/scripts/validators/correctness_check.py test_correctness.py
From sandbox root:
uv run skills/kernel-porting/scripts/validators/jax_syntax_check.py kernels/projects/{id}/converted_jax.py
uv run skills/kernel-porting/scripts/validators/correctness_check.py kernels/projects/{id}/test_correctness.py
Generated project artifacts (converted_jax.py, test_correctness.py, profile_script.py) should also use PEP 723 metadata so uv run converted_jax.py works standalone.
| Target | Path |
|---|---|
| Correctness / interpret | execution_run local provider + uv run validators |
| GPU/TPU benchmarks | load_skill_instructions(colab-cli) → colab start → colab exec -f profile_script.py → colab stop |
| Persistent remote | [harness.execution.ssh] + execution_run_background |
Profile locally (sandbox root):
uv run skills/kernel-porting/scripts/profile/roofline_mfu.py kernels/projects/{id}/profile_script.py
Windows hosts: local JAX is CPU-only; use Colab or SSH for GPU/TPU MFU.
kernel_db_sample(project_id, top_k=3)candidates/{uuid}/converted_jax.py, append queue line to candidates/queue.jsonluv run skills/kernel-porting/scripts/evolve/evolve_runner.py --project {id} --generations N
kernel_db_insert with fitness_latency_ms (required), optional fitness_mfu, mutation_classMutation classes: tiling, pipelining, layout, sparsity (see scripts/evolve/mutation_spec.json).
kernels/reference/Triton To Pallas Conversion.mdkernels/reference/Implementation_Plan.mdkernels/Test_Prompts.md — copy-paste prompts to exercise MaxEvolveWorkspace [harness.execution] uv_requirements also lists jax, jaxlib, pytest, numpy for the execution REPL. Validator scripts carry their own PEP 723 metadata so uv run path/to/script.py is self-contained.