Issues

Inlining as a first-class, searchable schedule decision

enhancement

#555

· lukstafi opened

on Aug 3, 2026

·

v1.0

Retire the concrete-index tracer: derive virtualization metrics from affine access relations

enhancement

#554

· lukstafi opened

on Aug 3, 2026

·

v1.0

Autotune's baseline is the unscheduled serial form, not the default schedule — the shared cause behind gh-532 and gh-533

#552

· lukstafi opened

on Aug 2, 2026

·

v1.0

bench_gpt has no gate-cost legs: contract item 5's four precision variants are unsatisfiable for gpt2_mini

#551

· lukstafi opened

on Aug 2, 2026

·

v1.0

gpt2_mini cuda/tuned: intermittent OOM during tuning, attributable to gh-541's expanded split-reduce seeding

#550

· lukstafi opened

on Aug 2, 2026

·

v1.0

Metal: a Tensorize wins its arm but never ships — the placement A/B picks the non-tensorized arm

#546

· lukstafi opened

on Aug 2, 2026

·

v1.0

CUDA bf16 mma: 20 candidates time scalar-fallback code under an mma label, so the route is measured but never emitted

#545

· lukstafi opened

on Aug 2, 2026

·

v1.0

Shape inference: 'close down when known' for operations, empty batch for parameters — use-site row resolution is a leaf-tensor rule applied too widely

#544

· lukstafi opened

on Aug 1, 2026

·

v1.0

gpt2_mini inference is 72x off torch cuda and only 2x faster than the 24-thread CPU — tuning and materialization both move nothing

#531

· lukstafi opened

on Jul 30, 2026

·

v1.0

Conv-sketch tuning wins do not port across CPUs: 3.2x/2.5x on Apple silicon vs 1.5x/1.1x on 24-thread x86 (gh-502 follow-up)

#530

· lukstafi opened

on Jul 30, 2026

·

v1.0

Autotune sketch seeding: rank-3 (batched/attention) matmul sites are never proposed, so gpt2_mini cannot reach tensor cores even after #521

#528

· lukstafi opened

on Jul 30, 2026

·

v1.0

16-bit storage with f32 compute on CPU: convert-on-load/store for bandwidth-bound kernels

#517

· lukstafi opened

on Jul 26, 2026

·

v1.0

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

All issues

Inlining as a first-class, searchable schedule decision

Retire the concrete-index tracer: derive virtualization metrics from affine access relations

Autotune's baseline is the unscheduled serial form, not the default schedule — the shared cause behind gh-532 and gh-533

bench_gpt has no gate-cost legs: contract item 5's four precision variants are unsatisfiable for gpt2_mini

gpt2_mini cuda/tuned: intermittent OOM during tuning, attributable to gh-541's expanded split-reduce seeding

Metal: a Tensorize wins its arm but never ships — the placement A/B picks the non-tensorized arm

CUDA bf16 mma: 20 candidates time scalar-fallback code under an mma label, so the route is measured but never emitted

Shape inference: 'close down when known' for operations, empty batch for parameters — use-site row resolution is a leaf-tensor rule applied too widely

gpt2_mini inference is 72x off torch cuda and only 2x faster than the 24-thread CPU — tuning and materialization both move nothing

Conv-sketch tuning wins do not port across CPUs: 3.2x/2.5x on Apple silicon vs 1.5x/1.1x on 24-thread x86 (gh-502 follow-up)

Autotune sketch seeding: rank-3 (batched/attention) matmul sites are never proposed, so gpt2_mini cannot reach tensor cores even after #521

16-bit storage with f32 compute on CPU: convert-on-load/store for bandwidth-bound kernels

Uh oh!

Search results