[feature] fix dependency, add legacy weights convertor

misc/legacy_weights_convertor.py

@@ -0,0 +1,94 @@
+import json
+import os
+from typing import Dict, List, Optional
+
+import fire
+import torch
+from huggingface_hub import snapshot_download
+from transformers import AutoModelForCausalLM
+
+import mixlora
+
+legacy_proj_names = {
+    "w1_proj": "gate_proj",
+    "w2_proj": "down_proj",
+    "w3_proj": "up_proj",
+}
+
+modern_proj_names = {
+    "gate_proj": "w1_proj",
+    "down_proj": "w2_proj",
+    "up_proj": "w3_proj",
+}
+
+
+def from_legacy(name_or_path: str, output_dir: Optional[str] = None):
+    if not os.path.exists(name_or_path):
+        assert output_dir is not None
+        name_or_path = snapshot_download(repo_id=name_or_path, repo_type="model")
+
+    if output_dir is None:
+        output_dir = name_or_path
+
+    if not os.path.exists(output_dir):
+        os.makedirs(output_dir)
+
+    with open(
+        name_or_path + os.sep + "adapter_config.json", "r", encoding="utf8"
+    ) as fp:
+        config = json.load(fp)
+        assert "routing_strategy" in config and config["routing_strategy"] == "mixtral"
+        config["routing_strategy"] = "mixlora"
+        target_modules: List[str] = []
+        assert isinstance(config["target_modules"], List)
+        for target in config["target_modules"]:
+            if target in legacy_proj_names:
+                target = legacy_proj_names[target]
+            if target in mixlora.config.lora_target_modules:
+                target_modules.append(target)
+        config["target_modules"] = target_modules
+        config = mixlora.MixLoraConfig.from_config(config)
+
+    config.check()
+
+    weights: Dict[str, torch.Tensor] = torch.load(
+        name_or_path + os.sep + "adapter_model.bin", map_location="cpu"
+    )
+
+    model = AutoModelForCausalLM.from_pretrained(
+        config.base_model_,
+        torch_dtype=torch.float16,
+        device_map="cpu",
+    )
+
+    for layer_idx, layer in enumerate(model.model.layers):
+        weights[f"mixlora.layers.{layer_idx}.mlp.moe_gate.weight"] = weights.pop(
+            f"mixlora.layers.{layer_idx}.gate.weight"
+        )
+        for proj_name, inject in config.target_modules_.items():
+            if not inject or not hasattr(layer.mlp, proj_name):
+                continue
+            for expert_idx in range(config.num_experts_):
+                new_layer_prefix_name = (
+                    f"mixlora.layers.{layer_idx}.mlp.{proj_name}.experts.{expert_idx}"
+                )
+                old_layer_prefix_name = (
+                    f"mixlora.layers.{layer_idx}.experts.{expert_idx}.{proj_name}"
+                )
+                if f"{old_layer_prefix_name}.lora_A.weight" not in weights:
+                    old_layer_prefix_name = f"mixlora.layers.{layer_idx}.experts.{expert_idx}.{modern_proj_names[proj_name]}"
+                weights[f"{new_layer_prefix_name}.lora_A.weight"] = weights.pop(
+                    f"{old_layer_prefix_name}.lora_A.weight"
+                )
+                weights[f"{new_layer_prefix_name}.lora_B.weight"] = weights.pop(
+                    f"{old_layer_prefix_name}.lora_B.weight"
+                )
+
+    torch.save(weights, output_dir + os.sep + "adapter_model.bin")
+
+    with open(output_dir + os.sep + "adapter_config.json", "w") as f:
+        json.dump(config.export(), f, indent=4)
+
+
+if __name__ == "__main__":
+    fire.Fire(from_legacy)

mixlora/model.py

@@ -351,7 +351,9 @@ def load_adapter_weights(
         config.adapter_name_ = adapter_name
         config.dtype_ = dtype
 
-    weights = torch.load(
+    config.check()
+
+    weights: Dict[str, torch.Tensor] = torch.load(
         name_or_path + os.sep + "adapter_model.bin", map_location=device
     )
 

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "mixlora"
-version = "0.2.1"
+version = "0.2.2"
 description = "State-of-the-art Parameter-Efficient MoE Fine-tuning Method"
 readme = "README.md"
 requires-python = ">=3.8"
@@ -14,8 +14,8 @@ classifiers = [
     "Operating System :: OS Independent",
 ]
 dependencies = [
-    "torch>=2.3.0,<2.4.0",
-    "transformers>=4.43.0,<4.44.0",
+    "torch>=2.2.0",
+    "transformers>=4.43.0",
     "huggingface_hub",
 ]