support-transformers-inference (#26)

- fix for transformers inference (baded737743fc2c5170986e6e6765b603a086d33)

config.json

@@ -116,10 +116,12 @@
       },
       "format": "pack-quantized",
       "ignore": [
-        "lm_head",
         "re:.*self_attn.*",
         "re:.*shared_experts.*",
-        "re:.*mlp\\.(gate|up|gate_up|down)_proj.*"
+        "re:.*mlp\\.(gate|up|gate_up|down)_proj.*",
+        "re:.*lm_head.*",
+        "re:vision_tower.*",
+        "re:mm_projector.*"
       ],
       "kv_cache_scheme": null,
       "quant_method": "compressed-tensors",

modeling_deepseek.py

@@ -1244,7 +1244,7 @@ class DeepseekV3PreTrainedModel(PreTrainedModel):
 
     def _init_weights(self, module):
         std = self.config.initializer_range
-        if isinstance(module, nn.Linear) and hasattr(module, "weight"):
+        if isinstance(module, nn.Linear):
             module.weight.data.normal_(mean=0.0, std=std)
             if module.bias is not None:
                 module.bias.data.zero_()

modeling_kimi_k25.py

@@ -571,9 +571,8 @@ class MoonViT3dEncoder(nn.Module):
         self.rope_2d = Rope2DPosEmbRepeated(
             block_cfg['hidden_dim'] // block_cfg['num_heads'], 512, 512)
         self.blocks = nn.ModuleList([
-            MoonViTEncoderLayer(
-                **block_cfg,
-                use_deterministic_attn=use_deterministic_attn)
+            MoonViTEncoderLayer(**block_cfg,
+                                use_deterministic_attn=use_deterministic_attn)
             for _ in range(num_layers)
         ])
         self.final_layernorm = nn.LayerNorm(hidden_dim)
@@ -638,20 +637,6 @@ class MoonViT3dPretrainedModel(PreTrainedModel):
     _no_split_modules = ['PackingTransformer']
     _supports_flash_attn_2 = True
     _supports_sdpa = True
-    
-    def _init_weights(self, module):
-        # Default PreTrainedModel._init_weights treats nn.Linear subclasses (e.g.
-        # compressed_tensors.CompressedLinear) as Linear but those layers may have
-        # no `weight` after compression; skip in that case.
-        std = getattr(self.config, "initializer_range", 0.02)
-        if isinstance(module, (nn.Linear, nn.Conv2d)) and hasattr(module, "weight"):
-            module.weight.data.normal_(mean=0.0, std=std)
-            if module.bias is not None:
-                module.bias.data.zero_()
-        elif isinstance(module, nn.Embedding):
-            module.weight.data.normal_(mean=0.0, std=std)
-            if module.padding_idx is not None:
-                module.weight.data[module.padding_idx].zero_()
 
     def __init__(self, config, *inputs, **kwargs):
         super().__init__(config, *inputs, **kwargs)
@@ -800,7 +785,7 @@ class KimiK25PreTrainedModel(PreTrainedModel):
         if hasattr(module, "class_embedding"):
             module.class_embedding.data.normal_(mean=0.0, std=std)
 
-        if isinstance(module, (nn.Linear, nn.Conv2d)) and hasattr(module, "weight"):
+        if isinstance(module, (nn.Linear, nn.Conv2d)):
             module.weight.data.normal_(mean=0.0, std=std)
             if module.bias is not None:
                 module.bias.data.zero_()