[ROCm][Perf][DSv4.1] Mono decode on every backbone layer (#61082)

Signed-off-by: Fangzhou Ai <fangzhou.ai@amd.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Fangzhou Ai
2026-10-13 10:16:50 +08:00
committed by GitHub
co-authored by Claude Opus 5.5
parent f47792af7f
commit cc2a6eea1a
13 changed files with 379 additions and 111 deletions
-7
View File
@@ -156,7 +156,6 @@ if TYPE_CHECKING:
VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION: bool = False
VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS: bool = False
VLLM_ROCM_AITER_MEGA_MOE_FUSE_SHARED_EXPERTS: bool = False
VLLM_ROCM_MONO_DECODE: bool = False
VLLM_ROCM_USE_AITER_TRITON_GEMM: bool = True
VLLM_ROCM_USE_SKINNY_GEMM: bool = True
VLLM_ROCM_FP8_PADDING: bool = True
@@ -1398,12 +1397,6 @@ environment_variables: dict[str, Callable[[], Any]] = {
os.getenv("VLLM_ROCM_AITER_MEGA_MOE_FUSE_SHARED_EXPERTS", "False").lower()
in ("true", "1")
),
# Run the eligible layers of decode steps on the model's mono decode
# kernels: persistent launches a layer, the TP all-reduces in-kernel
# (DeepSeek-V4.1 on CDNA4). By default is disabled.
"VLLM_ROCM_MONO_DECODE": lambda: (
os.getenv("VLLM_ROCM_MONO_DECODE", "False").lower() in ("true", "1")
),
# Whether to use aiter triton kernels for gemm ops.
# By default is enabled.
"VLLM_ROCM_USE_AITER_TRITON_GEMM": lambda: (