fix: Gemma2/Gemma3 inference issue caused by ln_1/ln_2 keys (switch to ln1/ln2)

Author: liyang

convert_hf_to_gguf.py

@@ -4873,7 +4873,7 @@ def _xlmroberta_set_vocab(self) -> None:
                 with open(tokenizer_config_path, "r", encoding="utf-8") as fp:
                     tokenizer_config_json = json.load(fp)
 
-            add_prefix = tokenizer.add_prefix_space
+            add_prefix = getattr(tokenizer, "add_prefix_space", False)
             remove_whitespaces = tokenizer.clean_up_tokenization_spaces
             precompiled_charsmap = b64decode(tokenizer_json["normalizer"]["precompiled_charsmap"])
 
@@ -6484,10 +6484,10 @@ def _map_block_tensor(self, layer: int, rest: str, data_torch: Tensor, name: str
         # layer norms
         if rest.startswith('norm1.'):
             suffix = parts[-1]
-            return [(f'v.blk.{layer}.ln_1.{suffix}', data_torch)]
+            return [(f'v.blk.{layer}.ln1.{suffix}', data_torch)]
         if rest.startswith('norm2.'):
             suffix = parts[-1]
-            return [(f'v.blk.{layer}.ln_2.{suffix}', data_torch)]
+            return [(f'v.blk.{layer}.ln2.{suffix}', data_torch)]
         if rest.startswith('attn.inner_attn_ln.'):
             suffix = parts[-1]
             return [(f'v.blk.{layer}.attn_ln.{suffix}', data_torch)]
@@ -6623,8 +6623,8 @@ def _should_be_f32(self, gguf_name: str) -> bool:
         binary-op dtype issues; patch embedding bias is also safer as F32.
         """
         patterns = (
-            ".ln_1.weight", ".ln_1.bias",
-            ".ln_2.weight", ".ln_2.bias",
+            ".ln1.weight", ".ln1.bias",
+            ".ln2.weight", ".ln2.bias",
             ".attn_ln.weight", ".attn_ln.bias",
             ".ffn_norm.weight", ".ffn_norm.bias",
             "v.patch_embd.proj.bias",

tools/mtmd/clip-impl.h

@@ -76,8 +76,8 @@
 #define TN_FFN_GATE        "%s.blk.%d.ffn_gate.%s"
 #define TN_FFN_UP          "%s.blk.%d.ffn_up.%s"
 #define TN_FFN_NORM        "%s.blk.%d.ffn_norm.%s"
-#define TN_LN_1            "%s.blk.%d.ln_1.%s"  // layer norm
-#define TN_LN_2            "%s.blk.%d.ln_2.%s"  // layer norm
+#define TN_LN_1            "%s.blk.%d.ln1.%s"  // layer norm
+#define TN_LN_2            "%s.blk.%d.ln2.%s"  // layer norm
 #define TN_LS_1            "%s.blk.%d.ls1.%s" // layer scale
 #define TN_LS_2            "%s.blk.%d.ls2.%s" // layer scale
 #define TN_LN_PRE          "%s.pre_ln.%s"