[perf][rocm]: accelerate strict VIME rollout with direct paged CK and graph-safe execution - #390

Merged

inaniloquentee merged 17 commits into

testfrom

perf/rocm-vime-rollout-overheads

Sep 8, 2026

Commits on Sep 6, 2026

perf(rocm): reduce strict paged decode overheads
Flink-ddd
committed
c9922db
perf(rocm): write deterministic GEMM roots directly
Flink-ddd
committed
b8e1054
perf(rocm): reuse rotary tables across q and k
Flink-ddd
committed
1136880
perf(rocm): write paged attention groups into caller output
Flink-ddd
committed
47e1fd1
perf(rocm): cache prepared LM-head weights
Flink-ddd
committed
af866d2
perf(rocm): bypass RoPE autograd in rollout
Flink-ddd
committed
336a2d7
perf(rocm): cache rotary inverse frequencies
Flink-ddd
committed
c5b122e
perf(rocm): reuse paged bounds validation across layers
Flink-ddd
committed
9df220f
perf(rocm): skip singleton rollout LSE copies
Flink-ddd
committed
b9ed179
perf(rocm): gather paged KV directly in head-major order
Flink-ddd
committed
5c1e4a8
bench(rocm): model the vllm pair-axis KV cache
Flink-ddd
committed
47bc20a
perf(rocm): cache immutable attention provenance inputs
Flink-ddd
committed
c70efd7

Commits on Sep 7, 2026

Commits on Sep 8, 2026

fix(rocm): make piecewise HIP graph replay exact
inaniloquentee
committed
fed8362