perf(BACKEND-ROCM): widen sampling block to 1024 and split-phase random sample - #3010

Open

ghazni101 wants to merge 10 commits into

mudler:mainfrom

ghazni101:row/SAMPLE-SPLIT-PHASE-ROCM

Commits on Sep 12, 2026

perf(GFX1100-TG200): widen sampling block to 1024 and split-phase random sample

ghazni101
authored and
mudler
committed
0b0b6eb
build(BACKEND-ROCM): classify VT_SAMPLE_SPLIT the way its sibling lever is classified
mudler
committed
c000022
fix(BACKEND-ROCM): launch the split-sample reduction at the width its shared arrays are sized for
mudler
committed
59d70f3
spec(BACKEND-ROCM): define split sampler scratch ownership
mudler
committed
a12ea63
test(BACKEND-ROCM): expose shared sampling scratch
mudler
committed
a424c03
test(BACKEND-ROCM): match captured sampler tensor shapes
mudler
committed
c0e7c9b
fix(BACKEND-ROCM): isolate split sampler scratch by queue
mudler
committed
24cd621
spec(BACKEND-ROCM): reject unwarmed sampler capture
mudler
committed
264faa3
test(BACKEND-ROCM): expose captured scratch publication
mudler
committed
c237128
fix(BACKEND-ROCM): refuse graph-owned sampler scratch
mudler
committed
075b435