[ROCm][MLA] Enable sparse MLA Gluon kernel from Aiter (#53492)

Signed-off-by: Mehmet Cagri Kaymak <mehmet.kaymak@amd.com>
Signed-off-by: Mehmet Cagri <mehmet.kaymak@amd.com>
This commit is contained in:
Mehmet Cagri
2026-09-29 09:08:18 +08:00
committed by GitHub
parent 869278cbb0
commit 70ae0b7435
7 changed files with 989 additions and 22 deletions
+9
View File
@@ -148,6 +148,7 @@ if TYPE_CHECKING:
VLLM_ROCM_AITER_MLA_DCP_VERIFY: Literal["asm", "segmented"] = "segmented"
VLLM_ROCM_USE_AITER_MHA: bool = True
VLLM_ROCM_USE_AITER_FP4_ASM_GEMM: bool = False
VLLM_ROCM_USE_AITER_TRITON_SPARSE_MLA: bool = False
VLLM_ROCM_USE_AITER_TRITON_ROPE: bool = False
VLLM_ROCM_USE_AITER_FP8BMM: bool = True
VLLM_ROCM_USE_AITER_FP4BMM: bool = True
@@ -1343,6 +1344,14 @@ environment_variables: dict[str, Callable[[], Any]] = {
"VLLM_ROCM_USE_AITER_FP4_ASM_GEMM": lambda: (
os.getenv("VLLM_ROCM_USE_AITER_FP4_ASM_GEMM", "False").lower() in ("true", "1")
),
# Whether sparse MLA prefill and decode run on aiter's Triton kernel, which
# reads the KV cache as stored (bf16 or fp8, paged or flat). Used by the
# ROCM_AITER_MLA_SPARSE backend (DeepSeek V3.2, GLM-5.x) and DeepSeek
# V4 / V4.1. gfx950 only. By default is disabled.
"VLLM_ROCM_USE_AITER_TRITON_SPARSE_MLA": lambda: (
os.getenv("VLLM_ROCM_USE_AITER_TRITON_SPARSE_MLA", "False").lower()
in ("true", "1")
),
# Whether to use aiter rope.
# By default is disabled.
"VLLM_ROCM_USE_AITER_TRITON_ROPE": lambda: (