more concise bpw

llms/mlx_lm/tuner/utils.py

@@ -249,12 +249,13 @@ def remove_lora_layers(model: nn.Module) -> nn.Module:
     return model
 
 
-def print_trainable_parameters(model):
-    def nparams(m):
-        if isinstance(m, (nn.QuantizedLinear, nn.QuantizedEmbedding)):
-            return m.weight.size * (32 // m.bits)
-        return sum(v.size for _, v in tree_flatten(m.parameters()))
+def nparams(module):
+    if isinstance(module, (nn.QuantizedLinear, nn.QuantizedEmbedding)):
+        return module.weight.size * 32 // module.bits
+    return sum(v.size for _, v in tree_flatten(module.parameters()))
 
+
+def print_trainable_parameters(model):
     leaf_modules = tree_flatten(
         model.leaf_modules(), is_leaf=lambda m: isinstance(m, nn.Module)
     )

llms/mlx_lm/utils.py

@@ -15,7 +15,7 @@ from typing import Any, Callable, Dict, Generator, List, Optional, Tuple, Type,
 import mlx.core as mx
 import mlx.nn as nn
 from huggingface_hub import snapshot_download
-from mlx.utils import tree_flatten, tree_reduce
+from mlx.utils import tree_flatten, tree_map, tree_reduce
 from transformers import PreTrainedTokenizer
 
 # Local imports
@@ -23,7 +23,7 @@ from .models import cache
 from .sample_utils import make_logits_processors, make_sampler
 from .tokenizer_utils import TokenizerWrapper, load_tokenizer
 from .tuner.utils import dequantize as dequantize_model
-from .tuner.utils import load_adapters
+from .tuner.utils import load_adapters, nparams
 
 # Constants
 MODEL_REMAPPING = {
@@ -100,6 +100,17 @@ def _get_classes(config: dict):
     return arch.Model, arch.ModelArgs
 
 
+def compute_bits_per_weight(model):
+    model_bytes = tree_reduce(
+        lambda acc, x: acc + x.nbytes if isinstance(x, mx.array) else acc, model, 0
+    )
+    leaf_modules = tree_flatten(
+        model.leaf_modules(), is_leaf=lambda m: isinstance(m, nn.Module)
+    )
+    model_params = sum(nparams(m) for _, m in leaf_modules)
+    return model_bytes * 8 / model_params
+
+
 def get_model_path(path_or_hf_repo: str, revision: Optional[str] = None) -> Path:
     """
     Ensures the model is available locally. If the path does not exist locally,
@@ -404,30 +415,6 @@ def load_config(model_path: Path) -> dict:
     return config
 
 
-def compute_bits_per_weight(config, weights):
-    weight_count = 0
-    bit_count = 0
-    for name, param in weights.items():
-        bit_count += param.size * param.dtype.size * 8
-        if ".scales" in name or ".biases" in name:
-            continue
-        if param.dtype == mx.uint32:
-            base_name = name.replace(".weight", "")
-            quant_config = config["quantization"]
-            quant_layer_config = quant_config.get(base_name)
-            bits = (
-                quant_config["bits"]
-                if not quant_layer_config
-                else quant_layer_config["bits"]
-            )
-            weight_count += param.size * 32 / bits
-        else:
-            weight_count += param.size
-
-    bits_per_weight = bit_count / weight_count
-    return bits_per_weight
-
-
 def load_model(
     model_path: Path,
     lazy: bool = False,
@@ -742,7 +729,7 @@ def quantize_model(
     quantized_config["quantization_config"] = quantized_config["quantization"]
     quantized_weights = dict(tree_flatten(model.parameters()))
 
-    bpw = compute_bits_per_weight(quantized_config, quantized_weights)
+    bpw = compute_bits_per_weight(model)
     print(f"[INFO] Quantized model with {bpw:.3f} bits per weight.")
 
     return quantized_weights, quantized_config