tonno

Autotuning for JAX/Pallas kernels — a lightweight @autotune decorator that sweeps block-size configs, times them on the current device, and caches the winner so the sweep only runs once per (hardware, problem shape) pair.

Inspired by Triton's @triton.autotune. Fills the same gap for Pallas (see jax-ml/jax#24340).

Install

pip install tonno          # or: uv add tonno

Usage

Stack @autotune on top of @jax.jit. The tunable parameters (keys of each config dict) must be declared as static_argnames in the jax.jit call and have defaults in the function signature — autotune injects the best values at call time.

import jax
from tonno import autotune

@autotune(configs=[
    {"BM": 32, "BN": 64},
    {"BM": 64, "BN": 128},
])
@jax.jit(static_argnames=["BM", "BN"])
def matmul(a: jax.Array, b: jax.Array, BM: int = 32, BN: int = 64) -> jax.Array:
    # BM / BN are concrete ints at compile time (static_argnames)
    return pallas_matmul(a, b, BM=BM, BN=BN)

Calling it

# First call: sweeps all configs, compiles in parallel, times sequentially,
# writes the winner to .tonno-cache/matmul.json
c = matmul(a, b)

# Subsequent calls with the same input shapes on the same device: cache hit,
# no sweep, runs immediately with the best config
c = matmul(a, b)

# Explicit override — bypass autotune entirely
c = matmul(a, b, BM=16, BN=32)

The cache key is derived from the input shapes and dtypes — exactly the same information jax.jit uses to decide whether to recompile. No explicit key= parameter needed.

Non-tunable static args

If your kernel has static args beyond the tunable ones, declare them in jax.jit as usual. They are automatically part of the cache key:

@autotune(configs=[{"BM": 32}, {"BM": 64}])
@jax.jit(static_argnames=["BM", "transpose"])
def matmul(a: jax.Array, b: jax.Array, transpose: bool = False, BM: int = 32) -> jax.Array:
    ...

How it works

1. On first call (cache miss): dummy inputs are built from the args' shapes/dtypes via jax.ensure_compile_time_eval. All configs are compiled in parallel via ThreadPoolExecutor (XLA compilation is CPU-bound). Each compiled artifact is then timed sequentially for accurate device timing. The winner is written to .tonno-cache/<fn>.json.

2. On subsequent calls (cache hit): the best config is loaded from disk and injected as static kwargs. JAX's own compilation cache takes over from there.

3. Inside jax.jit / jax.grad / jax.vmap: the sweep runs as a side channel during the first trace, then the winning config is baked into the jaxpr as a compile-time constant.

API reference

autotune(
    configs: list[dict[str, Any]],  # configs to sweep; all dicts must share keys
    *,
    num_warmup: int = 1,            # warmup calls per config after compilation
    num_timing: int = 3,            # timed calls per config (median used)
)

Contract:

• @autotune must wrap a @jax.jit-decorated function.
• The tunable param keys must appear in static_argnames of the jax.jit call.
• Tunable params must have defaults in the function signature.
• Config values must be JSON-serialisable (int, float, str, bool).

Cache

Results are stored in .tonno-cache/<qualname>.json (or $TONNO_CACHE_DIR). The file is human-readable JSON; you can inspect or delete entries manually.

{
  "NVIDIA H100 80GB": {
    "{\"__arg0\":{\"dtype\":\"float32\",\"shape\":[4096,4096]}, ...}": {
      "config": {"BM": 64, "BN": 128},
      "time_ms": 0.312,
      "key_values": {...}
    }
  }
}

Example

See examples/matmul.py for a complete autotuned tiled GEMM.