Add llms subdir + update README (#145)

* add llms subdir + update README

* nits

* use same pre-commit as mlx

* update readmes a bit

* format

llms/mixtral/README.md

@@ -0,0 +1,73 @@
+## Mixtral 8x7B
+
+Run the Mixtral[^mixtral] 8x7B mixture-of-experts (MoE) model in MLX on Apple silicon.
+
+This example also supports the instruction fine-tuned Mixtral model.[^instruct]
+
+Note, for 16-bit precision this model needs a machine with substantial RAM (~100GB) to run.
+
+### Setup
+
+Install [Git Large File
+Storage](https://docs.github.com/en/repositories/working-with-files/managing-large-files/installing-git-large-file-storage).
+For example with Homebrew:
+
+```
+brew install git-lfs
+```
+
+Download the models from Hugging Face:
+
+For the base model use:
+
+```
+export MIXTRAL_MODEL=Mixtral-8x7B-v0.1
+```
+
+For the instruction fine-tuned model use:
+
+```
+export MIXTRAL_MODEL=Mixtral-8x7B-Instruct-v0.1
+```
+
+Then run:
+
+```
+GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/mistralai/${MIXTRAL_MODEL}/
+cd $MIXTRAL_MODEL/ && \
+  git lfs pull --include "consolidated.*.pt" && \
+  git lfs pull --include "tokenizer.model"
+```
+
+Now from `mlx-exmaples/mixtral` convert and save the weights as NumPy arrays so
+MLX can read them:
+
+```
+python convert.py --model_path $MIXTRAL_MODEL/
+```
+
+The conversion script will save the converted weights in the same location.
+
+### Generate
+
+As easy as:
+
+```
+python mixtral.py --model_path $MIXTRAL_MODEL/
+```
+
+For more options including how to prompt the model, run:
+
+```
+python mixtral.py --help
+```
+
+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

llms/mixtral/convert.py

@@ -0,0 +1,59 @@
+# Copyright © 2023 Apple Inc.
+
+import argparse
+import glob
+import json
+from pathlib import Path
+
+import numpy as np
+import torch
+
+
+def convert(k, v, config):
+    v = v.to(torch.float16).numpy()
+    if "block_sparse_moe" not in k:
+        return [(k, v)]
+    if "gate" in k:
+        return [(k.replace("block_sparse_moe", "feed_forward"), v)]
+
+    # From: layers.N.block_sparse_moe.w
+    # To: layers.N.experts.M.w
+    num_experts = args["moe"]["num_experts"]
+    key_path = k.split(".")
+    v = np.split(v, num_experts, axis=0)
+    if key_path[-1] == "w2":
+        v = [u.T for u in v]
+
+    w_name = key_path.pop()
+    key_path[-1] = "feed_forward.experts"
+    return [
+        (".".join(key_path + [str(e), w_name, "weight"]), u) for e, u in enumerate(v)
+    ]
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Convert Mixtral weights to MLX.")
+    parser.add_argument(
+        "--model_path",
+        type=str,
+        default="Mixtral-8x7B-v0.1/",
+        help="The path to the Mixtral model. The MLX model weights will also be saved there.",
+    )
+    args = parser.parse_args()
+    model_path = Path(args.model_path)
+
+    with open("params.json") as fid:
+        args = json.load(fid)
+        args["model_type"] = "mixtral"
+    with open(model_path / "config.json", "w") as f:
+        json.dump(args, f, indent=4)
+
+    torch_files = glob.glob(str(model_path / "consolidated.*.pt"))
+    torch_files = sorted(torch_files, key=lambda tf: int(tf.split(".")[-2]))
+    for e, tf in enumerate(torch_files):
+        print(f"[INFO] Converting file {e + 1}/{len(torch_files)}")
+        state = torch.load(tf)
+        new_state = {}
+        for k, v in state.items():
+            new_state.update(convert(k, v, args))
+        np.savez(str(model_path / f"weights.{e}.npz"), **new_state)

llms/mixtral/mixtral.py

@@ -0,0 +1,332 @@
+# Copyright © 2023 Apple Inc.
+
+import argparse
+import glob
+import json
+from dataclasses import dataclass
+from pathlib import Path
+from typing import List, Optional, Tuple
+
+import mlx.core as mx
+import mlx.nn as nn
+import numpy as np
+from mlx.utils import tree_map, tree_unflatten
+from sentencepiece import SentencePieceProcessor
+
+
+@dataclass
+class ModelArgs:
+    dim: int
+    n_layers: int
+    head_dim: int
+    hidden_dim: int
+    n_heads: int
+    n_kv_heads: int
+    norm_eps: float
+    vocab_size: int
+    moe: dict = None
+
+
+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 RoPE(nn.RoPE):
+    def __init__(self, dims: int, traditional: bool = False):
+        super().__init__(dims, traditional)
+
+    def __call__(self, x, offset: int = 0):
+        shape = x.shape
+        x = mx.reshape(x, (-1, shape[-2], shape[-1]))
+        N = x.shape[1] + offset
+        costheta, sintheta = RoPE.create_cos_sin_theta(
+            N, self.dims, offset=offset, base=1000000, dtype=x.dtype
+        )
+
+        rope = (
+            self._compute_traditional_rope if self.traditional else self._compute_rope
+        )
+        rx = rope(costheta, sintheta, x)
+
+        return mx.reshape(rx, shape)
+
+
+class Attention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+
+        self.n_heads: int = args.n_heads
+        self.n_kv_heads: int = args.n_kv_heads
+
+        self.repeats = self.n_heads // self.n_kv_heads
+
+        self.scale = self.args.head_dim**-0.5
+
+        self.wq = nn.Linear(args.dim, args.n_heads * args.head_dim, bias=False)
+        self.wk = nn.Linear(args.dim, args.n_kv_heads * args.head_dim, bias=False)
+        self.wv = nn.Linear(args.dim, args.n_kv_heads * args.head_dim, bias=False)
+        self.wo = nn.Linear(args.n_heads * args.head_dim, args.dim, bias=False)
+        self.rope = RoPE(args.head_dim, traditional=True)
+
+    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.wq(x), self.wk(x), self.wv(x)
+
+        # Prepare the queries, keys and values for the attention computation
+        queries = queries.reshape(B, L, self.n_heads, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.n_kv_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.n_heads, L, -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.wo(output), (keys, values)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        self.w1 = nn.Linear(args.dim, args.hidden_dim, bias=False)
+        self.w2 = nn.Linear(args.hidden_dim, args.dim, bias=False)
+        self.w3 = nn.Linear(args.dim, args.hidden_dim, bias=False)
+
+    def __call__(self, x) -> mx.array:
+        return self.w2(nn.silu(self.w1(x)) * self.w3(x))
+
+
+class MOEFeedForward(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        self.num_experts = args.moe["num_experts"]
+        self.num_experts_per_tok = args.moe["num_experts_per_tok"]
+        self.experts = [FeedForward(args) for _ in range(self.num_experts)]
+        self.gate = nn.Linear(args.dim, self.num_experts, bias=False)
+
+    def __call__(self, x) -> 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 MOETransformerBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.n_heads = args.n_heads
+        self.dim = args.dim
+        self.attention = Attention(args)
+        self.feed_forward = MOEFeedForward(args=args)
+        self.attention_norm = RMSNorm(args.dim, eps=args.norm_eps)
+        self.ffn_norm = RMSNorm(args.dim, eps=args.norm_eps)
+        self.args = args
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Tuple[mx.array, mx.array]] = None,
+    ) -> mx.array:
+        r, cache = self.attention(self.attention_norm(x), mask, cache)
+        h = x + r
+        r = self.feed_forward(self.ffn_norm(h))
+        out = h + r
+        return out, cache
+
+
+class Mixtral(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.vocab_size = args.vocab_size
+        self.n_layers = args.n_layers
+        assert self.vocab_size > 0
+        self.tok_embeddings = nn.Embedding(args.vocab_size, args.dim)
+        self.layers = [MOETransformerBlock(args=args) for _ in range(args.n_layers)]
+        self.norm = RMSNorm(args.dim, eps=args.norm_eps)
+        self.output = nn.Linear(args.dim, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+    ):
+        h = self.tok_embeddings(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.output(self.norm(h[:, T - 1 : T, :])), cache
+
+
+class Tokenizer:
+    def __init__(self, model_path: str):
+        assert Path(model_path).exists(), model_path
+        self._model = SentencePieceProcessor(model_file=model_path)
+        self._sep = "▁"
+        assert self._model.vocab_size() == self._model.get_piece_size()
+
+    @property
+    def eos_id(self) -> int:
+        return self._model.eos_id()
+
+    @property
+    def pad_id(self) -> int:
+        return self._model.pad_id()
+
+    def encode(self, s: str) -> List[int]:
+        return [self._model.bos_id(), *self._model.encode(s)]
+
+    def decode(self, t: List[int]) -> str:
+        out = self._model.decode(t)
+        if t and self._model.id_to_piece(t[0])[0] == self._sep:
+            return " " + out
+        return out
+
+
+def load_model(folder: str, dtype=mx.float16):
+    model_path = Path(folder)
+    tokenizer = Tokenizer(str(model_path / "tokenizer.model"))
+    with open(model_path / "config.json", "r") as f:
+        config = json.loads(f.read())
+        config.pop("model_type", None)
+        model_args = ModelArgs(**config)
+    weight_files = glob.glob(str(model_path / "weights.*.npz"))
+    weights = {}
+    for wf in weight_files:
+        weights.update(mx.load(wf).items())
+    weights = tree_unflatten(list(weights.items()))
+    weights = tree_map(lambda p: p.astype(dtype), weights)
+    model = Mixtral(model_args)
+    model.update(weights)
+    return model, tokenizer
+
+
+def generate(prompt: mx.array, model: Mixtral, temp: Optional[float] = 0.0):
+    def sample(logits):
+        if temp == 0:
+            return mx.argmax(logits, axis=-1)
+        else:
+            return mx.random.categorical(logits * (1 / temp))
+
+    logits, cache = model(prompt[None])
+    y = sample(logits[:, -1, :])
+    yield y
+
+    while True:
+        logits, cache = model(y[:, None], cache)
+        y = sample(logits.squeeze(1))
+        yield y
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Mixtral inference script")
+    parser.add_argument(
+        "--model_path",
+        type=str,
+        default="Mixtral-8x7B-v0.1",
+        help="The path to the model weights, tokenizer, and config",
+    )
+    parser.add_argument(
+        "--prompt",
+        help="The message to be processed by the model",
+        default="In the beginning the Universe was created.",
+    )
+    parser.add_argument(
+        "--max_tokens",
+        "-m",
+        type=int,
+        default=100,
+        help="Maximum number of tokens to generate",
+    )
+    parser.add_argument(
+        "--temp",
+        help="The sampling temperature.",
+        type=float,
+        default=0.0,
+    )
+    parser.add_argument("--seed", type=int, default=0, help="The PRNG seed")
+
+    args = parser.parse_args()
+
+    mx.random.seed(args.seed)
+    print("[INFO] Loading model from disk.")
+    model, tokenizer = load_model(args.model_path)
+
+    print("[INFO] Starting generation...")
+
+    print(args.prompt, end="", flush=True)
+    prompt = mx.array(tokenizer.encode(args.prompt))
+    tokens = []
+    for token, _ in zip(generate(prompt, model, args.temp), range(args.max_tokens)):
+        tokens.append(token)
+
+        if (len(tokens) % 10) == 0:
+            mx.eval(tokens)
+            s = tokenizer.decode([t.item() for t in tokens])
+            print(s, end="", flush=True)
+            tokens = []
+
+    mx.eval(tokens)
+    s = tokenizer.decode([t.item() for t in tokens])
+    print(s, flush=True)

llms/mixtral/params.json

@@ -0,0 +1 @@
+{"dim": 4096, "n_layers": 32, "head_dim": 128, "hidden_dim": 14336, "n_heads": 32, "n_kv_heads": 8, "norm_eps": 1e-05, "vocab_size": 32000, "moe": {"num_experts_per_tok": 2, "num_experts": 8}}

llms/mixtral/requirements.txt

@@ -0,0 +1,4 @@
+mlx
+sentencepiece
+torch
+numpy