[ROCM] [FEAT] Integrate Aiter hipBLASLt GEMM online tuning (#40426)

Signed-off-by: hanlin12 <hanlin12@amd.com>
Signed-off-by: Han Lin <hanlin12@amd.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: TJian <tunjian.tan@embeddedllm.com>
This commit is contained in:
Han Lin
2026-06-04 23:45:36 -07:00
committed by GitHub
co-authored by gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com> TJian
parent c505cd93ef
commit 165b7864d0
6 changed files with 591 additions and 0 deletions
+12
View File
@@ -134,6 +134,7 @@ def _sync_hip_cuda_env_vars():
# Sync at import time - catches misconfigurations from process start.
_sync_hip_cuda_env_vars()
# AMDSMI utils
# Note that NVML is not affected by `{CUDA/HIP}_VISIBLE_DEVICES`,
# all the related functions work on real physical device ids.
@@ -312,6 +313,17 @@ def on_gfx950() -> bool:
return _ON_GFX950
# Enable HIP online tuning early, before hipBLASLt initializes.
# Turn on hipBLASLt online tuning if use AITER hipBLASLt GEMM.
if (
envs.VLLM_ROCM_USE_AITER
and envs.VLLM_ROCM_USE_AITER_LINEAR
and envs.VLLM_ROCM_USE_AITER_LINEAR_HIPBMM
and on_mi3xx()
):
os.environ["HIP_ONLINE_TUNING"] = "1"
@cache
def use_rocm_custom_paged_attention(
qtype: torch.dtype,