[ROCm][Perf] Use AITER tgemm for DeepSeek V4 compressors - #48727

Closed

Fangzhou-Ai wants to merge 10 commits into

vllm-project:mainfrom

Fangzhou-Ai:afz/dsv4-aiter-tgemm-compressors

Commits on Aug 7, 2026

[ROCm] Use AITER tgemm for DeepSeek V4 compressors
Fangzhou-Ai
committed
d25536b
cache the init time predicate values

tjtanaa
authored and
Fangzhou-Ai
committed
1a81b4b
fix(rocm): import shared fused QKV RMSNorm op
Fangzhou-Ai
committed
a09d2e4
refactor(rocm): simplify DeepSeek V4 tgemm capture
Fangzhou-Ai
committed
0690d0a
refactor(rocm): inline DeepSeek V4 tgemm selection
Fangzhou-Ai
committed
692f3b8
refactor(rocm): use DeepSeek V4 tgemm across modes
Fangzhou-Ai
committed
bf677d8
fix(rocm): preserve tgemm breakable graph boundary
Fangzhou-Ai
committed
9412bd2
refactor(rocm): reuse DeepSeek V4 attention pipeline
Fangzhou-Ai
committed
77c826a
fix(rocm): limit compressor tgemm to full graphs
Fangzhou-Ai
committed
4c13bdd
fix(rocm): limit compressor tgemm to tuned shapes
Fangzhou-Ai
committed
aaa3fee