[Perf] Integrate FlashInfer KDA kernels (#55364)

Signed-off-by: wzhao18 <wzhao18.sz@gmail.com>
This commit is contained in:
Wei Zhao
2026-09-07 20:29:30 -07:00
committed by GitHub
parent 12579c7f01
commit a69402aaa8
8 changed files with 723 additions and 156 deletions
+14 -2
View File
@@ -772,7 +772,10 @@ class EngineArgs:
fail_on_environ_validation: bool = False
gdn_prefill_backend: Literal["flashinfer", "triton", "cutedsl"] | None = None
kda_prefill_backend: Literal["auto", "triton", "flashkda"] | None = None
kda_prefill_backend: Literal["auto", "triton", "flashkda", "flashinfer"] | None = (
None
)
kda_decode_backend: Literal["auto", "native", "flashinfer", "triton"] | None = None
def __post_init__(self):
# support `EngineArgs(compilation_config={...})`
@@ -1747,10 +1750,17 @@ class EngineArgs:
parser.add_argument(
"--kda-prefill-backend",
dest="kda_prefill_backend",
choices=["auto", "triton", "flashkda"],
choices=["auto", "triton", "flashkda", "flashinfer"],
default=None,
help="Select KDA prefill backend.",
)
parser.add_argument(
"--kda-decode-backend",
dest="kda_decode_backend",
choices=["auto", "native", "flashinfer", "triton"],
default=None,
help="Select KDA decode backend.",
)
return parser
@classmethod
@@ -2567,6 +2577,8 @@ class EngineArgs:
self.additional_config["gdn_prefill_backend"] = self.gdn_prefill_backend
if self.kda_prefill_backend is not None:
self.additional_config["kda_prefill_backend"] = self.kda_prefill_backend
if self.kda_decode_backend is not None:
self.additional_config["kda_decode_backend"] = self.kda_decode_backend
config = VllmConfig(
model_config=model_config,