File size: 6,448 Bytes
818282c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 | #!/usr/bin/env python3
"""Inference-only Wisp architecture for the packaged MTP reference runtime.
This file intentionally contains no checkpoint, training, loss, data, or
release-repository dependencies. Ship it beside ``wisp_mtp_reference.py`` in
the Hugging Face package. Its parameter tree is identical to the training
model's inference tree.
"""
from __future__ import annotations
from dataclasses import dataclass
import mlx.core as mx
import mlx.nn as nn
@dataclass(frozen=True)
class ModelArgs:
vocab_size: int
dim: int
n_layers: int
n_heads: int
n_kv_heads: int
ffn_hidden: int
max_seq_len: int
rope_theta: float
norm_eps: float
tie_embeddings: bool
mtp_layers: int
mtp_depth: int
ce_chunk: int = 0
@property
def head_dim(self) -> int:
return self.dim // self.n_heads
def causal_mask(length: int, dtype=mx.float32) -> mx.array:
"""Return Wisp's additive square causal mask."""
upper = mx.triu(mx.ones((length, length), dtype=mx.bool_), k=1)
return mx.where(
upper,
mx.array(-1e9, dtype=dtype),
mx.array(0.0, dtype=dtype),
)
class Attention(nn.Module):
def __init__(self, args: ModelArgs):
super().__init__()
self.n_heads = args.n_heads
self.n_kv_heads = args.n_kv_heads
self.head_dim = args.head_dim
self.scale = self.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 = nn.RoPE(
args.head_dim,
traditional=False,
base=args.rope_theta,
)
def __call__(self, x, mask=None, cache=None):
batch, length, _ = x.shape
query = self.wq(x).reshape(
batch,
length,
self.n_heads,
self.head_dim,
)
key = self.wk(x).reshape(
batch,
length,
self.n_kv_heads,
self.head_dim,
)
value = self.wv(x).reshape(
batch,
length,
self.n_kv_heads,
self.head_dim,
)
query = query.transpose(0, 2, 1, 3)
key = key.transpose(0, 2, 1, 3)
value = value.transpose(0, 2, 1, 3)
offset = 0 if cache is None else cache[0].shape[2]
query = self.rope(query, offset=offset)
key = self.rope(key, offset=offset)
if cache is not None:
key = mx.concatenate([cache[0], key], axis=2)
value = mx.concatenate([cache[1], value], axis=2)
new_cache = (key, value)
output = mx.fast.scaled_dot_product_attention(
query,
key,
value,
scale=self.scale,
mask=mask,
)
output = output.transpose(0, 2, 1, 3).reshape(
batch,
length,
-1,
)
return self.wo(output), new_cache
class FeedForward(nn.Module):
def __init__(self, args: ModelArgs):
super().__init__()
self.w1 = nn.Linear(args.dim, args.ffn_hidden, bias=False)
self.w3 = nn.Linear(args.dim, args.ffn_hidden, bias=False)
self.w2 = nn.Linear(args.ffn_hidden, args.dim, bias=False)
def __call__(self, value):
return self.w2(nn.silu(self.w1(value)) * self.w3(value))
class Block(nn.Module):
def __init__(self, args: ModelArgs):
super().__init__()
self.attn_norm = nn.RMSNorm(args.dim, eps=args.norm_eps)
self.attn = Attention(args)
self.ffn_norm = nn.RMSNorm(args.dim, eps=args.norm_eps)
self.ffn = FeedForward(args)
def __call__(self, value, mask=None, cache=None):
attention, new_cache = self.attn(
self.attn_norm(value),
mask,
cache,
)
value = value + attention
value = value + self.ffn(self.ffn_norm(value))
return value, new_cache
class MTPModule(nn.Module):
def __init__(self, args: ModelArgs):
super().__init__()
self.h_norm = nn.RMSNorm(args.dim, eps=args.norm_eps)
self.e_norm = nn.RMSNorm(args.dim, eps=args.norm_eps)
self.proj = nn.Linear(2 * args.dim, args.dim, bias=False)
self.blocks = [Block(args) for _ in range(args.mtp_layers)]
def __call__(self, hidden, token_embeddings, mask=None, caches=None):
value = mx.concatenate(
[
self.h_norm(hidden),
self.e_norm(token_embeddings),
],
axis=-1,
)
value = self.proj(value)
new_caches = []
for index, block in enumerate(self.blocks):
cache = None if caches is None else caches[index]
value, new_cache = block(value, mask, cache)
new_caches.append(new_cache)
return value, new_caches
class Wisp(nn.Module):
def __init__(self, args: ModelArgs):
super().__init__()
self.args = args
self.tok_emb = nn.Embedding(args.vocab_size, args.dim)
self.blocks = [Block(args) for _ in range(args.n_layers)]
self.norm = nn.RMSNorm(args.dim, eps=args.norm_eps)
if not args.tie_embeddings:
self.lm_head = nn.Linear(args.dim, args.vocab_size, bias=False)
self.mtp = MTPModule(args)
def head(self, hidden):
normalized = self.norm(hidden)
if self.args.tie_embeddings:
return self.tok_emb.as_linear(normalized)
return self.lm_head(normalized)
def trunk(self, tokens, mask=None, caches=None):
hidden = self.tok_emb(tokens)
new_caches = []
for index, block in enumerate(self.blocks):
cache = None if caches is None else caches[index]
hidden, new_cache = block(hidden, mask, cache)
new_caches.append(new_cache)
return hidden, new_caches
def __call__(self, tokens, mask=None, caches=None):
hidden, new_caches = self.trunk(tokens, mask, caches)
return self.head(hidden), hidden, new_caches
|