[Fix] Transformers-fallback (GPT-NeoX) + KV pool config (DeepSeek-VL2) (#35244)

Co-authored-by: YanbingJiang <yanbing.jiang@intel.com>

python/sglang/srt/configs/model_config.py

@@ -907,7 +907,7 @@ class ModelConfig:
             setattr(self.hf_text_config, "head_dim", self.head_dim)
 
         self.v_head_dim = getattr(self.hf_text_config, "v_head_dim", None)
-        if self.v_head_dim is None:
+        if self.v_head_dim is None or self.v_head_dim == 0:
             self.v_head_dim = self.head_dim
             setattr(self.hf_text_config, "v_head_dim", self.v_head_dim)
 

python/sglang/srt/models/transformers.py

@@ -544,6 +544,8 @@ class TransformersBase(nn.Module):
             "model.score.": "classifier.",
             "model.classifier.": "classifier.",
             "transformer.": "model.",
+            "gpt_neox.": "model.",
+            "embed_out.": "lm_head.",
             "model.": "model.",
             "lm_head.": "lm_head.",
             "score.": "classifier.",
@@ -581,7 +583,9 @@ class TransformersBase(nn.Module):
         self.skip_substrs: list[str] = []
         self.ignore_unexpected_prefixes: list[str] = []
         self.ignore_unexpected_suffixes: list[str] = []
-        self.skip_substrs.extend([".attn.bias", ".attn.masked_bias", ".masked_bias"])
+        self.skip_substrs.extend(
+            [".attn.bias", ".attn.masked_bias", ".attention.bias", ".masked_bias"]
+        )
         self.ignore_unexpected_prefixes.extend(["classifier.", "score."])
 
         if self.quant_config is not None: