From 685b9d82bd014067ce62349944ce417554a9e0e0 Mon Sep 17 00:00:00 2001 From: Tianhao Zhou Date: Wed, 26 Nov 2025 14:57:22 -0800 Subject: [PATCH] fix: cuda graph issue while running longcat_flash (#14007) --- python/sglang/srt/models/longcat_flash.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/python/sglang/srt/models/longcat_flash.py b/python/sglang/srt/models/longcat_flash.py index fa7ba01bc..f3860e572 100644 --- a/python/sglang/srt/models/longcat_flash.py +++ b/python/sglang/srt/models/longcat_flash.py @@ -388,6 +388,7 @@ class LongcatFlashDecoderLayer(nn.Module): layer_scatter_modes=self.mlp_layer_scatter_modes[i], input_layernorm=self.input_layernorm[i], post_attention_layernorm=self.post_attention_layernorm[i], + qkv_latent_func=self.self_attn[i].prepare_qkv_latent, ) for i in range(2) ] @@ -402,6 +403,7 @@ class LongcatFlashDecoderLayer(nn.Module): layer_scatter_modes=self.moe_layer_scatter_modes, input_layernorm=self.input_layernorm[0], post_attention_layernorm=self.post_attention_layernorm[0], + qkv_latent_func=self.self_attn[0].prepare_qkv_latent, ) def forward(