feat(mlx-lm): add de-quant for fuse.py (#365)

* feat(mlx-lm): add de-quant for fuse

* chore: disable quant in to linear when de-quant enabled

* chore: add better error handling for adapter file not found

llms/mlx_lm/fuse.py

@@ -8,7 +8,7 @@ from typing import Any, Dict, Union
 from mlx.utils import tree_flatten, tree_unflatten
 
 from .tuner.lora import LoRALinear
-from .tuner.utils import apply_lora_layers
+from .tuner.utils import apply_lora_layers, dequantize
 from .utils import fetch_from_hub, get_model_path, save_weights, upload_to_hub
 
 
@@ -42,6 +42,11 @@ def parse_arguments() -> argparse.Namespace:
         type=str,
         default=None,
     )
+    parser.add_argument(
+        "--de-quantize",
+        help="Generate a de-quantized model.",
+        action="store_true",
+    )
     return parser.parse_args()
 
 
@@ -54,6 +59,7 @@ def main() -> None:
 
     model.freeze()
     model = apply_lora_layers(model, args.adapter_file)
+
     fused_linears = [
         (n, m.to_linear())
         for n, m in model.named_modules()
@@ -61,6 +67,11 @@ def main() -> None:
     ]
 
     model.update_modules(tree_unflatten(fused_linears))
+
+    if args.de_quantize:
+        print("De-quantizing model")
+        model = dequantize(model)
+
     weights = dict(tree_flatten(model.parameters()))
 
     save_path = Path(args.save_path)
@@ -73,6 +84,9 @@ def main() -> None:
 
     tokenizer.save_pretrained(save_path)
 
+    if args.de_quantize:
+        config.pop("quantization", None)
+
     with open(save_path / "config.json", "w") as fid:
         json.dump(config, fid, indent=4)
 

llms/mlx_lm/tuner/lora.py

@@ -29,7 +29,7 @@ class LoRALinear(nn.Module):
         lora_lin.linear = linear
         return lora_lin
 
-    def to_linear(self):
+    def to_linear(self, de_quantize: bool = False):
         linear = self.linear
         bias = "bias" in linear
         weight = linear.weight
@@ -56,7 +56,7 @@ class LoRALinear(nn.Module):
         if bias:
             fused_linear.bias = linear.bias
 
-        if is_quantized:
+        if is_quantized and not de_quantize:
             fused_linear = nn.QuantizedLinear.from_linear(
                 fused_linear,
                 linear.group_size,

llms/mlx_lm/tuner/utils.py

@@ -1,3 +1,5 @@
+import os
+
 import mlx.core as mx
 import mlx.nn as nn
 from mlx.utils import tree_unflatten
@@ -6,18 +8,62 @@ from .lora import LoRALinear
 
 
 def apply_lora_layers(model: nn.Module, adapter_file: str) -> nn.Module:
+    """
+    Apply LoRA layers to the model.
+
+    Args:
+        model (nn.Module): The neural network model.
+        adapter_file (str): Path to the adapter configuration file.
+
+    Returns:
+        nn.Module: The updated model with LoRA layers applied.
+    """
+    if not os.path.exists(adapter_file):
+        raise FileNotFoundError(f"The adapter file does not exist: {adapter_file}")
+
     adapters = list(mx.load(adapter_file).items())
-    linear_replacements = {}
+
+    linear_replacements = []
     lora_layers = set(
         [name.replace(".lora_a", "").replace(".lora_b", "") for name, _ in adapters]
     )
-
     for name, module in model.named_modules():
         if name in lora_layers:
             replacement_module = LoRALinear.from_linear(module)
-            linear_replacements[name] = replacement_module
+            linear_replacements.append((name, replacement_module))
 
-    model.update_modules(tree_unflatten(list(linear_replacements.items())))
-
-    model.update(tree_unflatten(adapters))
+    model.update_modules(tree_unflatten(linear_replacements))
+    return model
+
+
+def dequantize(model: nn.Module) -> nn.Module:
+    """
+    Dequantize the quantized linear layers in the model.
+
+    Args:
+        model (nn.Module): The model with quantized linear layers.
+
+    Returns:
+        nn.Module: The model with dequantized layers.
+    """
+    de_quantize_layers = []
+    for n, m in model.named_modules():
+        if isinstance(m, nn.QuantizedLinear):
+            bias = "bias" in m
+            weight = m.weight
+            weight = mx.dequantize(
+                weight,
+                m.scales,
+                m.biases,
+                m.group_size,
+                m.bits,
+            ).astype(mx.float16)
+            output_dims, input_dims = weight.shape
+            linear = nn.Linear(input_dims, output_dims, bias=bias)
+            linear.weight = weight
+            if bias:
+                linear.bias = m.bias
+            de_quantize_layers.append((n, linear))
+    if len(de_quantize_layers) > 0:
+        model.update_modules(tree_unflatten(de_quantize_layers))
     return model

llms/mlx_lm/utils.py

@@ -24,7 +24,7 @@ MODEL_MAPPING = {
     "qwen": qwen,
     "plamo": plamo,
 }
-MAX_FILE_SIZE_GB = 15
+MAX_FILE_SIZE_GB = 5
 
 linear_class_predicate = (
     lambda m: isinstance(m, nn.Linear)