Add autotuning - #247

Draft

AntonOresten wants to merge 18 commits into

mainfrom

Commits on Aug 23, 2026

Add autotuning to Experimental submodule.
AntonOresten
committed
efaeae7
Lock cufunction's codegen pipeline against concurrent compiles.
AntonOresten
committed
b053320
Restore shared inference cache via ScopedValue.
AntonOresten
committed
3fac480
Add hygienic @cutile macro.
AntonOresten
committed
b923736
Add `@autotune` macro and static num_ctas/occupancy hints.
AntonOresten
committed
b5bc4e4
Preserve `cache_hit=true` when a race fills the cache mid-tune.
AntonOresten
committed
d8dbf68
Pipeline compile→measure during autotune.
AntonOresten
committed
e8f8020
Move to src/experimental/
AntonOresten
committed
c7097dd
Trim autotune API: drop key_fn, accept literal grid/args, untype kwargs.
AntonOresten
committed
a5459fc
Drop single-value `nothing` axes from autotune test fixtures.
AntonOresten
committed
5e8b437
Refactor experimental autotuning internals
AntonOresten
committed
4c434e1
Keep autotune candidate compiles temporary
AntonOresten
committed
e5e4ec2
Fix TileCacheKey call
AntonOresten
committed
9032854
Retain context
AntonOresten
committed
df4fcea
cleanup
AntonOresten
committed
a75d15e
Support `num_worker_warps` entry hint in autotuning API
AntonOresten
committed
142a6f4
Remove `@cutile` macro
AntonOresten
committed
00a7b57

Commits on Aug 27, 2026

Drop the inference lock: interned cache keys, GPUCompiler.inference_batch
AntonOresten
committed
77b82da