From 8b395cb8c44763326ade23b3e4aea46bdeddee72 Mon Sep 17 00:00:00 2001
From: &LED-M& <105789115+xledx@users.noreply.github.com>
Date: Tue, 8 Sep 2026 09:44:27 +0900
Subject: [PATCH] fix: preserve BF16 embedding weights for get_rows

---
 src/model/common/ggml_block.hpp | 1 +
 src/model/te/clip.hpp           | 4 ++++
 2 files changed, 5 insertions(+)

diff --git a/src/model/common/ggml_block.hpp b/src/model/common/ggml_block.hpp
index cc0f69419..0b3cb8650 100644
--- a/src/model/common/ggml_block.hpp
+++ b/src/model/common/ggml_block.hpp
@@ -309,6 +309,7 @@ class Linear : public UnaryBlock {
 __STATIC_INLINE__ bool support_get_rows(ggml_type wtype) {
     switch (wtype) {
         case GGML_TYPE_F16:
+        case GGML_TYPE_BF16:
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_Q5_1:
         case GGML_TYPE_Q5_0:
diff --git a/src/model/te/clip.hpp b/src/model/te/clip.hpp
index 441a26c2a..cac613eda 100644
--- a/src/model/te/clip.hpp
+++ b/src/model/te/clip.hpp
@@ -145,6 +145,10 @@ class CLIPEmbeddings : public GGMLBlock {
         params["position_embedding.weight"] = ggml_new_tensor_2d(ctx, position_wtype, embed_dim, num_positions);
     }
 
+    enum ggml_op param_usage_op(const std::string& name) const override {
+        return name == "token_embedding.weight" ? GGML_OP_GET_ROWS : GGML_OP_NONE;
+    }
+
 public:
     CLIPEmbeddings(int64_t embed_dim,
                    int64_t vocab_size    = 49408,