Mlx llm package (#301)

* fix converter

* add recursive files

* remove gitignore

* remove gitignore

* add packages properly

* read me update

* remove dup readme

* relative

* fix convert

* fix community name

* fix url

* version

llms/mlx_lm/models/base.py

@@ -0,0 +1,15 @@
+import inspect
+from dataclasses import dataclass
+
+
+@dataclass
+class BaseModelArgs:
+    @classmethod
+    def from_dict(cls, params):
+        return cls(
+            **{
+                k: v
+                for k, v in params.items()
+                if k in inspect.signature(cls).parameters
+            }
+        )

llms/mlx_lm/models/llama.py

@@ -0,0 +1,202 @@
+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):
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    rms_norm_eps: float
+    vocab_size: int
+    num_key_value_heads: int = None
+    rope_theta: float = 10000
+    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
+
+        if self.rope_scaling:
+            required_keys = {"factor", "type"}
+            if not all(key in self.rope_scaling for key in required_keys):
+                raise ValueError(f"rope_scaling must contain keys {required_keys}")
+
+            if self.rope_scaling["type"] != "linear":
+                raise ValueError("rope_scaling 'type' currently only supports 'linear'")
+
+
+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 Attention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        dim = args.hidden_size
+        self.n_heads = n_heads = args.num_attention_heads
+        self.n_kv_heads = n_kv_heads = args.num_key_value_heads
+
+        self.repeats = n_heads // n_kv_heads
+
+        head_dim = args.hidden_size // n_heads
+        self.scale = head_dim**-0.5
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=False)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=False)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=False)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=False)
+
+        rope_scale = (
+            1 / args.rope_scaling["factor"]
+            if args.rope_scaling is not None and args.rope_scaling["type"] == "linear"
+            else 1
+        )
+        self.rope = nn.RoPE(
+            head_dim,
+            traditional=args.rope_traditional,
+            base=args.rope_theta,
+            scale=rope_scale,
+        )
+
+    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.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])
+
+        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 MLP(nn.Module):
+    def __init__(self, dim, hidden_dim):
+        super().__init__()
+        self.gate_proj = nn.Linear(dim, hidden_dim, bias=False)
+        self.down_proj = nn.Linear(hidden_dim, dim, bias=False)
+        self.up_proj = nn.Linear(dim, hidden_dim, bias=False)
+
+    def __call__(self, x) -> mx.array:
+        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.num_attention_heads = args.num_attention_heads
+        self.hidden_size = args.hidden_size
+        self.self_attn = Attention(args)
+        self.mlp = MLP(args.hidden_size, args.intermediate_size)
+        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)
+        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.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r
+        r = self.mlp(self.post_attention_layernorm(h))
+        out = h + r
+        return out, cache
+
+
+class LlamaModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.vocab_size = args.vocab_size
+        self.num_hidden_layers = args.num_hidden_layers
+        assert self.vocab_size > 0
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            TransformerBlock(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
+        if h.shape[1] > 1:
+            mask = nn.MultiHeadAttention.create_additive_causal_mask(h.shape[1])
+            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), cache
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.model = LlamaModel(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/models/phi2.py

@@ -0,0 +1,138 @@
+import math
+from dataclasses import dataclass
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    n_positions: int = 2048
+    vocab_size: int = 51200
+    n_embd: int = 2560
+    n_head: int = 32
+    n_layer: int = 32
+    rotary_dim: int = 32
+
+
+class LayerNorm(nn.LayerNorm):
+    def __call__(self, x: mx.array) -> mx.array:
+        return super().__call__(x.astype(mx.float32)).astype(x.dtype)
+
+
+class RoPEAttention(nn.Module):
+    def __init__(self, dims: int, n_head: int, rotary_dim: int):
+        super().__init__()
+
+        self.n_head = n_head
+
+        self.q_proj = nn.Linear(dims, dims)
+        self.k_proj = nn.Linear(dims, dims)
+        self.v_proj = nn.Linear(dims, dims)
+        self.dense = nn.Linear(dims, dims)
+
+        self.rope = nn.RoPE(rotary_dim, traditional=False)
+
+    def __call__(self, x, mask=None, cache=None):
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        # Extract some shapes
+        n_head = self.n_head
+        B, L, D = queries.shape
+
+        # Prepare the queries, keys and values for the attention computation
+        queries = queries.reshape(B, L, n_head, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, L, n_head, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, n_head, -1).transpose(0, 2, 1, 3)
+
+        # Add RoPE to the queries and keys and combine them with the cache
+        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)
+
+        queries = queries.astype(mx.float32)
+        keys = keys.astype(mx.float32)
+
+        # Finally perform the attention computation
+        scale = math.sqrt(1 / queries.shape[-1])
+        scores = (queries * scale) @ keys.transpose(0, 1, 3, 2)
+        if mask is not None:
+            scores = scores + mask
+
+        scores = mx.softmax(scores, axis=-1).astype(values.dtype)
+        values_hat = (scores @ values).transpose(0, 2, 1, 3).reshape(B, L, -1)
+
+        return self.dense(values_hat), (keys, values)
+
+
+class MLP(nn.Module):
+    def __init__(self, dim, hidden_dim):
+        super().__init__()
+        self.fc1 = nn.Linear(dim, hidden_dim)
+        self.fc2 = nn.Linear(hidden_dim, dim)
+        self.act = nn.GELU(approx="precise")
+
+    def __call__(self, x) -> mx.array:
+        return self.fc2(self.act(self.fc1(x)))
+
+
+class ParallelBlock(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        dims = config.n_embd
+        mlp_dims = dims * 4
+        self.self_attn = RoPEAttention(dims, config.n_head, config.rotary_dim)
+        self.input_layernorm = LayerNorm(dims)
+        self.mlp = MLP(dims, mlp_dims)
+
+    def __call__(self, x, mask, cache):
+        h = self.input_layernorm(x)
+        attn_h, cache = self.self_attn(h, mask, cache)
+        ff_h = self.mlp(h)
+        return attn_h + ff_h + x, cache
+
+
+class Transformer(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.n_embd)
+        self.layers = [ParallelBlock(config) for i in range(config.n_layer)]
+        self.final_layernorm = LayerNorm(config.n_embd)
+
+    def __call__(self, x, mask, cache):
+        x = self.embed_tokens(x)
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        for e, layer in enumerate(self.layers):
+            x, cache[e] = layer(x, mask, cache[e])
+        return self.final_layernorm(x), cache
+
+
+class Model(nn.Module):
+    def __init__(self, config: ModelArgs):
+        super().__init__()
+        self.model = Transformer(config)
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size)
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: mx.array = None,
+        cache: mx.array = None,
+    ) -> tuple[mx.array, mx.array]:
+        mask = None
+        if x.shape[1] > 1:
+            mask = nn.MultiHeadAttention.create_additive_causal_mask(x.shape[1])
+            mask = mask.astype(x.dtype)
+
+        y, cache = self.model(x, mask, cache)
+        return self.lm_head(y), cache