Quantize embedding / Update quantize API (#680)

* more async eval

* quantize embedding / update quantize api

* more updates for quantize

* update for quantize embeddings

* update sd quant API

* update sdxl quants

* error for datasets < batch_size

* async

* fix config loading

* fix quant

* fix tests

* fix req

* remove lm head if tie weights is true

* fix test

llms/llama/convert.py

@@ -134,7 +134,7 @@ def quantize(weights, config, args):
     model.update(tree_unflatten(list(weights.items())))
 
     # Quantize the model:
-    nn.QuantizedLinear.quantize_module(model, args.q_group_size, args.q_bits)
+    nn.quantize(model, args.q_group_size, args.q_bits)
 
     # Update the config:
     quantized_config["quantization"] = {

llms/llama/llama.py

@@ -339,7 +339,7 @@ def load_model(model_path):
         quantization = config.pop("quantization", None)
     model = Llama(ModelArgs(**config))
     if quantization is not None:
-        nn.QuantizedLinear.quantize_module(model, **quantization)
+        nn.quantize(model, **quantization)
     model.update(tree_unflatten(list(weights.items())))
     tokenizer = SentencePieceProcessor(model_file=str(model_path / "tokenizer.model"))
     return model, tokenizer

llms/llama/requirements.txt

@@ -1,4 +1,4 @@
-mlx>=0.8.0
+mlx>=0.11.0
 sentencepiece
 torch
 numpy