diff --git a/python/sglang/srt/models/deepseek_v4.py b/python/sglang/srt/models/deepseek_v4.py index f9508ed95..425261ff9 100644 --- a/python/sglang/srt/models/deepseek_v4.py +++ b/python/sglang/srt/models/deepseek_v4.py @@ -421,6 +421,7 @@ class MQALayer(nn.Module): kv = qkv_a[..., self.q_lora_rank :] else: kv, _ = self.wkv(x) + kv = kv.contiguous() fused_norm_rope_inplace( kv, self.kv_norm.weight.data,