biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[ROCm]: Enable customop and rope+kvcache fusion for AITER RoPE (#35180)

Browse Source 
Signed-off-by: Rohan138 <rohanpotdar138@gmail.com>
This commit is contained in:
Rohan Potdar
2026-02-24 22:36:40 -06:00
committed by GitHub
parent ec1d30c0f6
commit f38f8c9742
9 changed files with 139 additions and 67 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/envs.py
View File

@@ -105,7 +105,7 @@ if TYPE_CHECKING:
VLLM_ROCM_USE_AITER_MLA: bool = True
VLLM_ROCM_USE_AITER_MHA: bool = True
VLLM_ROCM_USE_AITER_FP4_ASM_GEMM: bool = False
VLLM_ROCM_USE_AITER_TRITON_ROPE: bool = False
VLLM_ROCM_USE_AITER_TRITON_ROPE: bool = True
VLLM_ROCM_USE_AITER_FP8BMM: bool = True
VLLM_ROCM_USE_AITER_FP4BMM: bool = True
VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION: bool = False
@@ -937,9 +937,9 @@ environment_variables: dict[str, Callable[[], Any]] = {
os.getenv("VLLM_ROCM_USE_AITER_FP4_ASM_GEMM", "False").lower() in ("true", "1")
),
# Whether to use aiter rope.
# By default is disabled.
# By default is enabled.
"VLLM_ROCM_USE_AITER_TRITON_ROPE": lambda: (
os.getenv("VLLM_ROCM_USE_AITER_TRITON_ROPE", "False").lower() in ("true", "1")
os.getenv("VLLM_ROCM_USE_AITER_TRITON_ROPE", "True").lower() in ("true", "1")
),
# Whether to use aiter triton fp8 bmm kernel
# By default is enabled.