build(llama.cu): 更新 InfiniNN

YdrMaster · YdrMaster · commit 525978f6377c · 2025-07-09T11:20:48.000+08:00
Signed-off-by: YdrMaster &lt;ydrml@hotmail.com&gt;
diff --git a/Cargo.lock b/Cargo.lock
diff --git a/llama.cu/Cargo.toml b/llama.cu/Cargo.toml
@@ -7,7 +7,7 @@ edition.workspace = true
 operators = { git = "https://github.com/YdrMaster/operators-rs", rev = "01b39e8", default-features = false, features = [
     "nvidia-gpu",
 ] }
-nn = { git = "https://github.com/YdrMaster/InfiniNN", rev = "7b01a0e" }
+nn = { git = "https://github.com/YdrMaster/InfiniNN", rev = "e3061d6" }
 ggus = { git = "https://github.com/InfiniTensor/gguf", rev = "23c362f" }
 tokeneer = { git = "https://github.com/InfiniTensor/tokeneer", rev = "c48f39f" }
 
diff --git a/llama.cu/src/model/llama.rs b/llama.cu/src/model/llama.rs
@@ -68,6 +68,8 @@ impl GGufModel<'_> {
                                 get(&format!("blk.{iblk}.attn_qkv.weight")),
                                 dt_bias.map(|dt| (dt, get(&format!("blk.{iblk}.attn_qkv.bias")))),
                             ),
+                            q_norm: None,
+                            k_norm: None,
                             rope: Some(RoPE {
                                 multimodal: false,
                                 nctx,
