format

llama/llama.py

@@ -315,7 +315,7 @@ def load_model(model_path):
             config["hidden_dim"] = weights["layers.0.feed_forward.w1.weight"].shape[0]
         if config.get("vocab_size", -1) < 0:
             config["vocab_size"] = weights["output.weight"].shape[-1]
-        unused = ["multiple_of", "ffn_dim_multiplier", 'rope_theta']
+        unused = ["multiple_of", "ffn_dim_multiplier", "rope_theta"]
         for k in unused:
             if k in config:
                 config.pop(k)

mixtral/README.md

@@ -62,6 +62,12 @@ For more options including how to prompt the model, run:
 python mixtral.py --help
 ```
 
-[^mixtral]: Refer to Mistral's [blog post](https://mistral.ai/news/mixtral-of-experts/) for more details.
+For the Instruction model, make sure to follow the prompt format:
+
+```
+[INST] Instruction prompt [/INST]
+```
+
+[^mixtral]: Refer to Mistral's [blog post](https://mistral.ai/news/mixtral-of-experts/) and the [Hugging Face blog post](https://huggingface.co/blog/mixtral) for more details.
 [^instruc]: Refer to the [Hugging Face repo](https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1) for more
 details

phi2/convert.py

@@ -1,6 +1,7 @@
 from transformers import AutoModelForCausalLM
 import numpy as np
 
+
 def replace_key(key: str) -> str:
     if "wte.weight" in key:
         key = "wte.weight"

whisper/test.py

@@ -65,7 +65,6 @@ class TestWhisper(unittest.TestCase):
         logits = mlx_model(mels, tokens)
         self.assertEqual(logits.dtype, mx.float16)
 
-
     def test_decode_lang(self):
         options = decoding.DecodingOptions(task="lang_id", fp16=False)
         result = decoding.decode(self.model, self.mels, options)

whisper/whisper/load_models.py

@@ -44,7 +44,7 @@ _ALIGNMENT_HEADS = {
     "large-v1": b"ABzY8r9j$a0{>%R7#4sLmoOs{s)o3~84-RPdcFk!JR<kSfC2yj",
     "large-v2": b"ABzY8zd+h!0{>%R7=D0pU<_bnWW*tkYAhobTNnu$jnkEkXqp)j;w1Tzk)UH3X%SZd&fFZ2fC2yj",
     "large-v3": b"ABzY8gWO1E0{>%R7(9S+Kn!D~%ngiGaR?*L!iJG9p-nab0JQ=-{D1-g00",
-    "large": b"ABzY8gWO1E0{>%R7(9S+Kn!D~%ngiGaR?*L!iJG9p-nab0JQ=-{D1-g00"
+    "large": b"ABzY8gWO1E0{>%R7(9S+Kn!D~%ngiGaR?*L!iJG9p-nab0JQ=-{D1-g00",
 }
 
 
@@ -166,7 +166,8 @@ def convert(model, rules=None):
 
 
 def torch_to_mlx(
-    torch_model: torch_whisper.Whisper, dtype: mx.Dtype = mx.float16,
+    torch_model: torch_whisper.Whisper,
+    dtype: mx.Dtype = mx.float16,
 ) -> whisper.Whisper:
     def convert_rblock(model, rules):
         children = dict(model.named_children())

whisper/whisper/whisper.py

@@ -37,6 +37,7 @@ def sinusoids(length, channels, max_timescale=10000):
     scaled_time = mx.arange(length)[:, None] * inv_timescales[None, :]
     return mx.concatenate([mx.sin(scaled_time), mx.cos(scaled_time)], axis=1)
 
+
 class LayerNorm(nn.LayerNorm):
     def __call__(self, x: mx.array) -> mx.array:
         return super().__call__(x.astype(mx.float32)).astype(x.dtype)
@@ -123,7 +124,13 @@ class ResidualAttentionBlock(nn.Module):
 
 class AudioEncoder(nn.Module):
     def __init__(
-        self, n_mels: int, n_ctx: int, n_state: int, n_head: int, n_layer: int, dtype: mx.Dtype = mx.float16,
+        self,
+        n_mels: int,
+        n_ctx: int,
+        n_state: int,
+        n_head: int,
+        n_layer: int,
+        dtype: mx.Dtype = mx.float16,
     ):
         super().__init__()
         self.conv1 = nn.Conv1d(n_mels, n_state, kernel_size=3, padding=1)
@@ -148,7 +155,13 @@ class AudioEncoder(nn.Module):
 
 class TextDecoder(nn.Module):
     def __init__(
-        self, n_vocab: int, n_ctx: int, n_state: int, n_head: int, n_layer: int, dtype: mx.Dtype = mx.float16,
+        self,
+        n_vocab: int,
+        n_ctx: int,
+        n_state: int,
+        n_head: int,
+        n_layer: int,
+        dtype: mx.Dtype = mx.float16,
     ):
         super().__init__()
 
@@ -160,7 +173,9 @@ class TextDecoder(nn.Module):
             for _ in range(n_layer)
         ]
         self.ln = LayerNorm(n_state)
-        self._mask = nn.MultiHeadAttention.create_additive_causal_mask(n_ctx).astype(dtype)
+        self._mask = nn.MultiHeadAttention.create_additive_causal_mask(n_ctx).astype(
+            dtype
+        )
 
     def __call__(self, x, xa, kv_cache=None):
         """