[Attention] Add MLA prefill backend: trtllm_ragged_attention_deepseek (#26397)

Signed-off-by: Ming Yang <minos.future@gmail.com>
This commit is contained in:
Ming Yang
2025-10-24 10:24:08 -07:00
committed by GitHub
parent 7e1d697b56
commit 0f67d4d962
2 changed files with 113 additions and 2 deletions
+6
View File
@@ -183,6 +183,7 @@ if TYPE_CHECKING:
VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB: int | None = None
VLLM_NIXL_ABORT_REQUEST_TIMEOUT: int = 480
VLLM_USE_CUDNN_PREFILL: bool = False
VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL: bool = False
VLLM_ENABLE_CUDAGRAPH_GC: bool = False
VLLM_LOOPBACK_IP: str = ""
VLLM_ALLOW_CHUNKED_LOCAL_ATTN_WITH_HYBRID_KV_CACHE: bool = False
@@ -1250,6 +1251,10 @@ environment_variables: dict[str, Callable[[], Any]] = {
"VLLM_USE_CUDNN_PREFILL": lambda: bool(
int(os.getenv("VLLM_USE_CUDNN_PREFILL", "0"))
),
# Controls whether to use TRT-LLM ragged DeepSeek prefill
"VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL": lambda: bool(
int(os.getenv("VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL", "0"))
),
# If set to 1/True, use the TRTLLM attention backend in flashinfer.
# If set to 0/False, use the default attention backend in flashinfer.
# If not set, auto-detect the attention backend in flashinfer.
@@ -1481,6 +1486,7 @@ def compute_hash() -> str:
"VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS",
"VLLM_USE_FLASHINFER_MOE_MXFP4_BF16",
"VLLM_USE_CUDNN_PREFILL",
"VLLM_USE_TRTLLM_RAGGED_DEEPSEEK_PREFILL",
"VLLM_USE_TRTLLM_ATTENTION",
"VLLM_FLASHINFER_DISABLE_Q_QUANTIZATION",
"VLLM_ROCM_USE_AITER",