From 096acb4f85500f7fca267b519f835633d74d2149 Mon Sep 17 00:00:00 2001 From: wangzaijun Date: Thu, 9 Jul 2026 02:44:41 +0000 Subject: [PATCH] fix qwen3.5 mtp --- lightllm/common/kv_cache_mem_manager/qwen3next_mem_manager.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/lightllm/common/kv_cache_mem_manager/qwen3next_mem_manager.py b/lightllm/common/kv_cache_mem_manager/qwen3next_mem_manager.py index c721ad7e7..f65a849df 100644 --- a/lightllm/common/kv_cache_mem_manager/qwen3next_mem_manager.py +++ b/lightllm/common/kv_cache_mem_manager/qwen3next_mem_manager.py @@ -30,7 +30,8 @@ def __init__( def get_att_input_params(self, layer_index: int) -> Tuple[Any, Any]: if layer_index >= self.linear_config.all_layer_num: - layer_index -= self.linear_layer_num + # MTP draft full-attn layers are packed after the main model layers. + layer_index -= self.linear_config.linear_layer_num else: layer_index = layer_index // self.linear_config.full_attention_interval return super().get_att_input_params(layer_index)