[KDA] Default GLM-5-Next spec verify to the CuTe DSL KDA kernel (#42905)

Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai>
This commit is contained in:
Mohammad Miadh Angkad
2026-10-06 22:31:41 -07:00
committed by GitHub
co-authored by Mohammad Angkad
parent 9ddbba50dd
commit f47d89564d
2 changed files with 43 additions and 0 deletions
@@ -0,0 +1,36 @@
from __future__ import annotations
import logging
from typing import TYPE_CHECKING
from sglang.srt.arg_groups.overrides import (
declare_resolution,
resolving_view,
)
from sglang.srt.runtime_context import get_platform
if TYPE_CHECKING:
from sglang.srt.server_args import ServerArgs
logger = logging.getLogger(__name__)
def apply_glm5_next_spec_backend_defaults(server_args: ServerArgs) -> None:
cfg = resolving_view(server_args)
if (
cfg.speculative_algorithm is None
or cfg.linear_attn_verify_backend is not None
or not get_platform().is_sm100
):
return
declare_resolution(
server_args,
"apply_glm5_next_spec_backend_defaults",
linear_attn_verify_backend="nv_cutedsl",
)
logger.info(
"GLM-5-Next with speculative decoding: defaulting "
"--linear-attn-verify-backend to nv_cutedsl."
)
@@ -233,6 +233,13 @@ def handle_model_specific_adjustments(server_args: Any):
apply_kimi_k3_linear_attn_defaults(server_args)
apply_kimi_k3_spec_backend_defaults(server_args)
if model_arch == "Glm5NextForConditionalGeneration":
from sglang.srt.arg_groups.glm5_next_hook import (
apply_glm5_next_spec_backend_defaults,
)
apply_glm5_next_spec_backend_defaults(server_args)
if model_arch in [
"DeepseekV4ForCausalLM",
]: