From f47d89564dc2479be8e2e866a2b17f149318758e Mon Sep 17 00:00:00 2001 From: Mohammad Miadh Angkad <176301910+mmangkad@users.noreply.github.com> Date: Wed, 7 Oct 2026 05:31:41 +0000 Subject: [PATCH] [KDA] Default GLM-5-Next spec verify to the CuTe DSL KDA kernel (#42905) Co-authored-by: Mohammad Angkad --- .../sglang/srt/arg_groups/glm5_next_hook.py | 36 +++++++++++++++++++ python/sglang/srt/arg_groups/model_hook.py | 7 ++++ 2 files changed, 43 insertions(+) create mode 100644 python/sglang/srt/arg_groups/glm5_next_hook.py diff --git a/python/sglang/srt/arg_groups/glm5_next_hook.py b/python/sglang/srt/arg_groups/glm5_next_hook.py new file mode 100644 index 0000000000..347e1308ff --- /dev/null +++ b/python/sglang/srt/arg_groups/glm5_next_hook.py @@ -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." + ) diff --git a/python/sglang/srt/arg_groups/model_hook.py b/python/sglang/srt/arg_groups/model_hook.py index b75a0a51f4..345bb646de 100644 --- a/python/sglang/srt/arg_groups/model_hook.py +++ b/python/sglang/srt/arg_groups/model_hook.py @@ -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", ]: