Final cloud stability fixes for PEFT adapter loading

convert_adapters_to_peft.py

@@ -0,0 +1,99 @@
+import os
+import json
+import argparse
+from safetensors.torch import load_file, save_file
+
+def convert_mlx_to_peft(mlx_dir, peft_dir, base_model):
+    os.makedirs(peft_dir, exist_ok=True)
+    
+    # 1. Load MLX weights
+    try:
+        mlx_weights = load_file(os.path.join(mlx_dir, "adapters.safetensors"))
+    except Exception as e:
+        print(f"Error loading {mlx_dir}/adapters.safetensors: {e}")
+        return
+
+    # 2. Convert keys to PEFT format
+    peft_weights = {}
+    target_modules = set()
+    
+    for k, v in mlx_weights.items():
+        is_a = k.endswith(".lora_a")
+        is_b = k.endswith(".lora_b")
+        is_A = k.endswith(".A")
+        is_B = k.endswith(".B")
+        
+        if is_a:
+            module_path = k[:-7] # remove .lora_a
+            new_suffix = ".lora_A.weight"
+        elif is_b:
+            module_path = k[:-7] # remove .lora_b
+            new_suffix = ".lora_B.weight"
+        elif is_A:
+            module_path = k[:-2] # remove .A
+            new_suffix = ".lora_A.weight"
+        elif is_B:
+            module_path = k[:-2] # remove .B
+            new_suffix = ".lora_B.weight"
+        else:
+            print(f"Skipping unknown key: {k}")
+            continue
+            
+        target_module_name = module_path.split(".")[-1]
+        target_modules.add(target_module_name)
+        
+        new_k = f"base_model.model.{module_path}{new_suffix}"
+        peft_weights[new_k] = v
+        
+    save_file(peft_weights, os.path.join(peft_dir, "adapter_model.safetensors"))
+    
+    # 3. Read MLX config to get rank, alpha, dropout
+    with open(os.path.join(mlx_dir, "adapter_config.json"), "r") as f:
+        mlx_config = json.load(f)
+        
+    # 4. Write PEFT config
+    peft_config = {
+        "alpha_pattern": {},
+        "auto_mapping": None,
+        "base_model_name_or_path": base_model,
+        "bias": "none",
+        "fan_in_fan_out": False,
+        "inference_mode": True,
+        "init_lora_weights": True,
+        "layer_replication": None,
+        "layers_pattern": None,
+        "layers_to_transform": None,
+        "loftq_config": {},
+        "lora_alpha": mlx_config.get("alpha", 16),
+        "lora_dropout": mlx_config.get("dropout", 0.05),
+        "megatron_config": None,
+        "megatron_core": "megatron.core",
+        "modules_to_save": None,
+        "peft_type": "LORA",
+        "r": mlx_config.get("rank", 8),
+        "rank_pattern": {},
+        "revision": None,
+        "target_modules": list(target_modules),
+        "task_type": "CAUSAL_LM",
+        "use_dora": False,
+        "use_rslora": False
+    }
+    
+    with open(os.path.join(peft_dir, "adapter_config.json"), "w") as f:
+        json.dump(peft_config, f, indent=2)
+
+    print(f"✅ Converted {mlx_dir} to {peft_dir} successfully!")
+
+if __name__ == "__main__":
+    configs = [
+        ("ministral_adapters_v2", "peft_ministral_adapters_v2", "mistralai/Ministral-3-3B-Instruct-2512"),
+        ("ministral_adapters_v3", "peft_ministral_adapters_v3", "mistralai/Ministral-3-3B-Instruct-2512"),
+        ("ministral_dpo_adapters", "peft_ministral_dpo_adapters", "mistralai/Ministral-3-3B-Instruct-2512"),
+        ("adapters", "peft_adapters", "Qwen/Qwen3.5-4B-Instruct")
+    ]
+    
+    for in_dir, out_dir, base in configs:
+        if os.path.exists(in_dir):
+            convert_mlx_to_peft(in_dir, out_dir, base)
+        else:
+            print(f"Directory not found, skipping: {in_dir}")