feat(mlx_lm): add mixtral support in mlx_lm (#318)

* feat: add mixtral support in mlx_lm

* chore: update doc

llms/README.md

@@ -101,10 +101,12 @@ Here are a few examples of Hugging Face models that work with this example:
 - [deepseek-ai/deepseek-coder-6.7b-instruct](https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-instruct)
 - [01-ai/Yi-6B-Chat](https://huggingface.co/01-ai/Yi-6B-Chat)
 - [microsoft/phi-2](https://huggingface.co/microsoft/phi-2)
+- [mistralai/Mixtral-8x7B-Instruct-v0.1](https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1)
 
 Most
 [Mistral](https://huggingface.co/models?library=transformers,safetensors&other=mistral&sort=trending),
 [Llama](https://huggingface.co/models?library=transformers,safetensors&other=llama&sort=trending),
-and
 [Phi-2](https://huggingface.co/models?library=transformers,safetensors&other=phi&sort=trending)
+and
+[Mixtral](https://huggingface.co/models?library=transformers,safetensors&other=mixtral&sort=trending)
 style models should work out of the box.

llms/mlx_lm/convert.py

@@ -10,7 +10,7 @@ import mlx.nn as nn
 import transformers
 from mlx.utils import tree_flatten
 
-from .utils import get_model_path, load
+from .utils import get_model_path, linear_class_predicate, load
 
 MAX_FILE_SIZE_GB = 15
 
@@ -94,11 +94,10 @@ def quantize_model(
     model, _ = load(hf_path)
     model.load_weights(list(weights.items()))
 
-    nn.QuantizedLinear.quantize_module(model, q_group_size, q_bits)
+    nn.QuantizedLinear.quantize_module(
-    quantized_config["quantization"] = {
+        model, q_group_size, q_bits, linear_class_predicate=linear_class_predicate
-        "group_size": q_group_size,
+    )
-        "bits": q_bits,
+    quantized_config["quantization"] = {"group_size": q_group_size, "bits": q_bits}
-    }
     quantized_weights = dict(tree_flatten(model.parameters()))
 
     return quantized_weights, quantized_config

llms/mlx_lm/models/mixtral.py

@@ -0,0 +1,247 @@
+from dataclasses import dataclass
+from typing import Dict, Optional, Tuple, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    vocab_size: int = 32000
+    max_position_embeddings: int = 4096 * 32
+    hidden_size: int = 4096
+    intermediate_size: int = 14336
+    num_hidden_layers: int = 32
+    num_attention_heads: int = 32
+    num_experts_per_tok: int = 2
+    num_key_value_heads: int = 8
+    num_local_experts: int = 8
+    rms_norm_eps: float = 1e-5
+    vocab_size: int
+    rope_theta: float = 1e6
+    rope_traditional: bool = False
+    model_type: str = None
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+
+    def __post_init__(self):
+        if self.num_key_value_heads is None:
+            self.num_key_value_heads = self.num_attention_heads
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, dims: int, eps: float = 1e-5):
+        super().__init__()
+        self.weight = mx.ones((dims,))
+        self.eps = eps
+
+    def _norm(self, x):
+        return x * mx.rsqrt(x.square().mean(-1, keepdims=True) + self.eps)
+
+    def __call__(self, x):
+        output = self._norm(x.astype(mx.float32)).astype(x.dtype)
+        return self.weight * output
+
+
+class MixtralAttention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.hidden_size = args.hidden_size
+        self.num_heads = args.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.num_key_value_heads = args.num_key_value_heads
+        self.max_position_embeddings = args.max_position_embeddings
+        self.rope_theta = args.rope_theta
+
+        self.repeats = self.num_heads // self.num_key_value_heads
+
+        self.scale = self.head_dim**-0.5
+
+        self.q_proj = nn.Linear(
+            self.hidden_size, self.num_heads * self.head_dim, bias=False
+        )
+        self.k_proj = nn.Linear(
+            self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False
+        )
+        self.v_proj = nn.Linear(
+            self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False
+        )
+        self.o_proj = nn.Linear(
+            self.num_heads * self.head_dim, self.hidden_size, bias=False
+        )
+
+        self.rope = nn.RoPE(
+            self.head_dim,
+            traditional=args.rope_traditional,
+            base=args.rope_theta,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Tuple[mx.array, mx.array]] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        # Prepare the queries, keys and values for the attention computation
+        queries = queries.reshape(B, L, self.num_heads, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, L, self.num_key_value_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.num_key_value_heads, -1).transpose(
+            0, 2, 1, 3
+        )
+
+        def repeat(a):
+            a = mx.concatenate([mx.expand_dims(a, 2)] * self.repeats, axis=2)
+            return a.reshape([B, self.num_heads, L, -1])
+
+        if self.repeats > 1:
+            keys, values = map(repeat, (keys, values))
+
+        if cache is not None:
+            key_cache, value_cache = cache
+            queries = self.rope(queries, offset=key_cache.shape[2])
+            keys = self.rope(keys, offset=key_cache.shape[2])
+            keys = mx.concatenate([key_cache, keys], axis=2)
+            values = mx.concatenate([value_cache, values], axis=2)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        scores = (queries * self.scale) @ keys.transpose(0, 1, 3, 2)
+        if mask is not None:
+            scores += mask
+        scores = mx.softmax(scores.astype(mx.float32), axis=-1).astype(scores.dtype)
+        output = (scores @ values).transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output), (keys, values)
+
+
+class MixtralBLockSparseTop2MLP(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.ffn_dim = args.intermediate_size
+        self.hidden_dim = args.hidden_size
+
+        self.w1 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)
+        self.w2 = nn.Linear(self.ffn_dim, self.hidden_dim, bias=False)
+        self.w3 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)
+
+        self.act_fn = nn.silu
+
+    def __call__(self, x: mx.array) -> mx.array:
+        current_hidden_states = self.act_fn(self.w1(x)) * self.w3(x)
+        current_hidden_states = self.w2(current_hidden_states)
+        return current_hidden_states
+
+
+class MixtralSparseMoeBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.hidden_dim = args.hidden_size
+        self.ffn_dim = args.intermediate_size
+        self.num_experts = args.num_local_experts
+        self.num_experts_per_tok = args.num_experts_per_tok
+
+        # gating
+        self.gate = nn.Linear(self.hidden_dim, self.num_experts, bias=False)
+
+        self.experts = [
+            MixtralBLockSparseTop2MLP(args=args) for _ in range(self.num_experts)
+        ]
+
+    def __call__(self, x: mx.array) -> mx.array:
+        ne = self.num_experts_per_tok
+        orig_shape = x.shape
+        x = x.reshape(-1, x.shape[-1])
+
+        gates = self.gate(x)
+        inds = mx.argpartition(-gates, kth=ne, axis=-1)[:, :ne]
+        scores = mx.softmax(
+            mx.take_along_axis(gates, inds, axis=-1).astype(mx.float32),
+            axis=-1,
+        ).astype(gates.dtype)
+
+        y = []
+        for xt, st, it in zip(x, scores, inds.tolist()):
+            yt = mx.concatenate([self.experts[e](xt)[:, None] for e in it], axis=-1)
+            yt = (yt * st).sum(axis=-1)
+            y.append(yt[None, :])
+        y = mx.concatenate(y)
+
+        return y.reshape(orig_shape)
+
+
+class MixtralDecoderLayer(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.hidden_size = args.hidden_size
+
+        self.self_attn = MixtralAttention(args)
+
+        self.block_sparse_moe = MixtralSparseMoeBlock(args)
+        self.input_layernorm = RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Tuple[mx.array, mx.array]] = None,
+    ) -> mx.array:
+        r, cache = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r
+        r = self.block_sparse_moe(self.post_attention_layernorm(h))
+        out = h + r
+        return out, cache
+
+
+class MixtralModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.vocab_size = args.vocab_size
+        self.num_hidden_layers = args.num_hidden_layers
+
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            MixtralDecoderLayer(args=args) for _ in range(args.num_hidden_layers)
+        ]
+        self.norm = RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+    ):
+        h = self.embed_tokens(inputs)
+
+        mask = None
+        T = h.shape[1]
+        if T > 1:
+            mask = nn.MultiHeadAttention.create_additive_causal_mask(T)
+            mask = mask.astype(h.dtype)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        for e, layer in enumerate(self.layers):
+            h, cache[e] = layer(h, mask, cache[e])
+
+        return self.norm(h[:, T - 1 : T, :]), cache
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.model = MixtralModel(args)
+        self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+    ):
+        out, cache = self.model(inputs, cache)
+        return self.lm_head(out), cache

llms/mlx_lm/utils.py

@@ -10,16 +10,21 @@ from huggingface_hub import snapshot_download
 from transformers import AutoTokenizer, PreTrainedTokenizer
 
 # Local imports
-from .models import llama, phi2
+from .models import llama, mixtral, phi2
 from .models.base import BaseModelArgs
 
 # Constants
 MODEL_MAPPING = {
     "llama": llama,
     "mistral": llama,  # mistral is compatible with llama
+    "mixtral": mixtral,
     "phi": phi2,
 }
 
+linear_class_predicate = (
+    lambda m: isinstance(m, nn.Linear) and m.weight.shape[0] % 32 == 0
+)  # TODO remove this once we support quantization for non-multiples of 32
+
 
 def _get_classes(config: dict):
     """
@@ -171,7 +176,11 @@ def load(path_or_hf_repo: str) -> Tuple[nn.Module, PreTrainedTokenizer]:
     model = model_class(model_args)
 
     if quantization is not None:
-        nn.QuantizedLinear.quantize_module(model, **quantization)
+        nn.QuantizedLinear.quantize_module(
+            model,
+            **quantization,
+            linear_class_predicate=linear_class_predicate,
+        )
 
     model.load_weights(list(weights.items()))