Upload chatglm.py
Browse files- chatglm.py +608 -0
chatglm.py
ADDED
@@ -0,0 +1,608 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from dataclasses import dataclass
|
2 |
+
from typing import Dict, Optional, Tuple, Union
|
3 |
+
import math
|
4 |
+
|
5 |
+
import mlx.core as mx
|
6 |
+
import mlx.nn as nn
|
7 |
+
|
8 |
+
from .base import BaseModelArgs
|
9 |
+
|
10 |
+
|
11 |
+
@dataclass
|
12 |
+
class ModelArgs(BaseModelArgs):
|
13 |
+
model_type: str
|
14 |
+
add_bias_linear: bool = False
|
15 |
+
add_qkv_bias: bool = True
|
16 |
+
apply_query_key_layer_scaling: bool = True
|
17 |
+
apply_residual_connection_post_layernorm: bool = False
|
18 |
+
attention_dropout: float = 0.0
|
19 |
+
attention_softmax_in_fp32: bool = True
|
20 |
+
bias_dropout_fusion: bool = True
|
21 |
+
ffn_hidden_size: int = 13696
|
22 |
+
fp32_residual_connection: bool = False
|
23 |
+
hidden_dropout: float = 0.0
|
24 |
+
hidden_size: int = 4096
|
25 |
+
kv_channels: int = 128
|
26 |
+
layernorm_epsilon: float = 1.5625e-07
|
27 |
+
multi_query_attention: bool = True
|
28 |
+
multi_query_group_num: int = 2
|
29 |
+
num_attention_heads: int = 32
|
30 |
+
num_hidden_layers: int = 40
|
31 |
+
num_layers: int = 40
|
32 |
+
rope_ratio: int = 500
|
33 |
+
original_rope: bool = True
|
34 |
+
padded_vocab_size: int = 151552
|
35 |
+
post_layer_norm: bool = True
|
36 |
+
rmsnorm: bool = True
|
37 |
+
seq_length: int = 131072
|
38 |
+
use_cache: bool = True
|
39 |
+
torch_dtype: str = "bfloat16"
|
40 |
+
tie_word_embeddings: bool = False
|
41 |
+
|
42 |
+
def __post_init__(self):
|
43 |
+
pass
|
44 |
+
|
45 |
+
class RotaryEmbedding(nn.Module):
|
46 |
+
def __init__(self, dim, rope_ratio=1, original_impl=False, dtype=None):
|
47 |
+
super().__init__()
|
48 |
+
# inv_freq = 1.0 / (10000 ** (mx.arange(0, dim, 2, dtype=dtype) / dim))
|
49 |
+
# self.register_buffer("inv_freq", inv_freq)
|
50 |
+
# self.inv_freq = mx.array(inv_freq, dtype=dtype)
|
51 |
+
self.inv_freq_type = dtype
|
52 |
+
self.dim = dim
|
53 |
+
self.original_impl = original_impl
|
54 |
+
self.rope_ratio = rope_ratio
|
55 |
+
|
56 |
+
def forward_impl(
|
57 |
+
self, seq_len: int, n_elem: int, dtype: mx.Dtype, base: int = 10000
|
58 |
+
):
|
59 |
+
"""Enhanced Transformer with Rotary Position Embedding.
|
60 |
+
Derived from:https://github.com/labmlai/annotated_deep_learning_paper_implementations/blob/master/labml_nn/
|
61 |
+
transformers/rope/__init__.py. MIT License:
|
62 |
+
https://github.com/labmlai/annotated_deep_learning_paper_implementations/blob/master/license.
|
63 |
+
"""
|
64 |
+
# $\Theta = {\theta_i = 10000^{\frac{2(i-1)}{d}}, i \in [1, 2, ..., \frac{d}{2}]}$
|
65 |
+
base = base * self.rope_ratio
|
66 |
+
theta = 1.0 / (base ** (mx.arange(0, n_elem, 2, dtype=mx.float16) / n_elem))
|
67 |
+
|
68 |
+
# Create position indexes `[0, 1, ..., seq_len - 1]`
|
69 |
+
seq_idx = mx.arange(seq_len, dtype=mx.float16)
|
70 |
+
|
71 |
+
# Calculate the product of position index and $\theta_i$
|
72 |
+
idx_theta = mx.outer(seq_idx, theta).astype(mx.float16)
|
73 |
+
|
74 |
+
cache = mx.stack([mx.cos(idx_theta), mx.sin(idx_theta)], axis=-1)
|
75 |
+
|
76 |
+
# this is to mimic the behaviour of complex32, else we will get different results
|
77 |
+
if dtype in (mx.float16, mx.bfloat16, mx.int8):
|
78 |
+
cache = cache.astype(mx.bfloat16) if dtype == mx.bfloat16 else cache.astype(mx.float16)
|
79 |
+
return cache
|
80 |
+
|
81 |
+
def __call__(self, max_seq_len, offset=0):
|
82 |
+
return self.forward_impl(
|
83 |
+
max_seq_len, self.dim, dtype=self.inv_freq_type,
|
84 |
+
)
|
85 |
+
|
86 |
+
def apply_rotary_pos_emb(x: mx.array, rope_cache: mx.array) -> mx.array:
|
87 |
+
# x: [b, np, sq, hn]
|
88 |
+
b, np, sq, hn = x.shape[0], x.shape[1], x.shape[2], x.shape[3]
|
89 |
+
rot_dim = rope_cache.shape[-2] * 2
|
90 |
+
x, x_pass = x[..., :rot_dim], x[..., rot_dim:]
|
91 |
+
# truncate to support variable sizes
|
92 |
+
rope_cache = rope_cache[:, :sq]
|
93 |
+
xshaped = x.reshape(b, np, sq, rot_dim // 2, 2)
|
94 |
+
rope_cache = rope_cache.reshape(-1, 1, sq, xshaped.shape[3], 2)
|
95 |
+
x_out2 = mx.stack(
|
96 |
+
[
|
97 |
+
xshaped[..., 0] * rope_cache[..., 0] - xshaped[..., 1] * rope_cache[..., 1],
|
98 |
+
xshaped[..., 1] * rope_cache[..., 0] + xshaped[..., 0] * rope_cache[..., 1],
|
99 |
+
],
|
100 |
+
-1,
|
101 |
+
)
|
102 |
+
x_out2 = x_out2.flatten(3)
|
103 |
+
return mx.concatenate((x_out2, x_pass), axis=-1)
|
104 |
+
|
105 |
+
# class RMSNorm(nn.Module):
|
106 |
+
# def __init__(self, normalized_shape, eps=1e-5, device=None, dtype=None, **kwargs):
|
107 |
+
# super().__init__()
|
108 |
+
# self.weight = nn.empty(normalized_shape, device=device, dtype=dtype)
|
109 |
+
# self.eps = eps
|
110 |
+
|
111 |
+
# def __call__(self, hidden_states: mx.array):
|
112 |
+
# input_dtype = hidden_states.dtype
|
113 |
+
# variance = hidden_states.astype("float32").power(2).mean(-1, keepdims=True)
|
114 |
+
# hidden_states = hidden_states * variance.rsqrt()
|
115 |
+
|
116 |
+
# return (self.weight * hidden_states).astype(input_dtype)
|
117 |
+
|
118 |
+
|
119 |
+
class CoreAttention(nn.Module):
|
120 |
+
def __init__(self, args: ModelArgs, layer_number):
|
121 |
+
super().__init__()
|
122 |
+
|
123 |
+
self.apply_query_key_layer_scaling = args.apply_query_key_layer_scaling
|
124 |
+
self.attention_softmax_in_fp32 = args.attention_softmax_in_fp32
|
125 |
+
if self.apply_query_key_layer_scaling:
|
126 |
+
self.attention_softmax_in_fp32 = True
|
127 |
+
self.layer_number = max(1, layer_number)
|
128 |
+
|
129 |
+
projection_size = args.kv_channels * args.num_attention_heads
|
130 |
+
|
131 |
+
# Per attention head and per partition values.
|
132 |
+
self.hidden_size_per_partition = projection_size
|
133 |
+
self.hidden_size_per_attention_head = projection_size // args.num_attention_heads
|
134 |
+
self.num_attention_heads_per_partition = args.num_attention_heads
|
135 |
+
|
136 |
+
coeff = None
|
137 |
+
self.norm_factor = math.sqrt(self.hidden_size_per_attention_head)
|
138 |
+
if self.apply_query_key_layer_scaling:
|
139 |
+
coeff = self.layer_number
|
140 |
+
self.norm_factor *= coeff
|
141 |
+
self.coeff = coeff
|
142 |
+
|
143 |
+
self.attention_dropout = nn.Dropout(args.attention_dropout)
|
144 |
+
|
145 |
+
def __call__(self, query_layer, key_layer, value_layer, attention_mask):
|
146 |
+
# scale_factor = 1 / math.sqrt(query_layer.shape[-1])
|
147 |
+
scale_factor = query_layer.shape[-1] ** -0.5
|
148 |
+
# if self.layer_number == 1:
|
149 |
+
# print(f"== |{self.layer_number}| query_layer:{query_layer.shape} key_layer:{key_layer.shape} value_layer:{value_layer.shape}")
|
150 |
+
if attention_mask is None and query_layer.shape[2] == key_layer.shape[2]:
|
151 |
+
attention_mask = nn.MultiHeadAttention.create_additive_causal_mask(query_layer.shape[2]).astype(query_layer.dtype)
|
152 |
+
context_layer = mx.fast.scaled_dot_product_attention(query_layer, key_layer, value_layer, scale=scale_factor,mask=attention_mask)
|
153 |
+
else:
|
154 |
+
if attention_mask is not None:
|
155 |
+
attention_mask = ~attention_mask
|
156 |
+
context_layer = mx.fast.scaled_dot_product_attention(query_layer, key_layer, value_layer, scale=scale_factor, mask=attention_mask)
|
157 |
+
context_layer = context_layer.transpose((0,2,1,3))
|
158 |
+
new_context_layer_shape = context_layer.shape[:-2] + (self.hidden_size_per_partition,)
|
159 |
+
context_layer = context_layer.reshape(*new_context_layer_shape)
|
160 |
+
|
161 |
+
return context_layer
|
162 |
+
|
163 |
+
class SelfAttention(nn.Module):
|
164 |
+
def __init__(self, args: ModelArgs, layer_number):
|
165 |
+
super(SelfAttention, self).__init__()
|
166 |
+
self.layer_number = max(1, layer_number)
|
167 |
+
|
168 |
+
self.projection_size = args.kv_channels * args.num_attention_heads
|
169 |
+
|
170 |
+
# Per attention head and per partition values.
|
171 |
+
self.hidden_size_per_attention_head = self.projection_size // args.num_attention_heads
|
172 |
+
self.num_attention_heads_per_partition = args.num_attention_heads
|
173 |
+
self.multi_query_attention = args.multi_query_attention
|
174 |
+
self.qkv_hidden_size = 3 * self.projection_size
|
175 |
+
if self.multi_query_attention:
|
176 |
+
self.num_multi_query_groups_per_partition = args.multi_query_group_num
|
177 |
+
self.qkv_hidden_size = (
|
178 |
+
self.projection_size + 2 * self.hidden_size_per_attention_head * args.multi_query_group_num
|
179 |
+
)
|
180 |
+
self.query_key_value = nn.Linear(args.hidden_size, self.qkv_hidden_size,
|
181 |
+
bias=args.add_bias_linear or args.add_qkv_bias)
|
182 |
+
|
183 |
+
self.core_attention = CoreAttention(args, self.layer_number)
|
184 |
+
|
185 |
+
# Output.
|
186 |
+
self.dense = nn.Linear(self.projection_size, args.hidden_size, bias=args.add_bias_linear)
|
187 |
+
|
188 |
+
def __call__(self, hidden_states, attention_mask, rotary_pos_emb, kv_cache=None, use_cache=True):
|
189 |
+
# hidden_states: [b, sq, h]
|
190 |
+
|
191 |
+
# =================================================
|
192 |
+
# Pre-allocate memory for key-values for inference.
|
193 |
+
# =================================================
|
194 |
+
# =====================
|
195 |
+
# Query, Key, and Value
|
196 |
+
# =====================
|
197 |
+
|
198 |
+
# Attention heads [b, sq, h] --> [b, sq, (np * 3 * hn)]
|
199 |
+
mixed_x_layer = self.query_key_value(hidden_states)
|
200 |
+
|
201 |
+
if self.multi_query_attention:
|
202 |
+
q_k_v_len = [
|
203 |
+
self.num_attention_heads_per_partition * self.hidden_size_per_attention_head,
|
204 |
+
self.num_multi_query_groups_per_partition * self.hidden_size_per_attention_head,
|
205 |
+
self.num_multi_query_groups_per_partition * self.hidden_size_per_attention_head,
|
206 |
+
]
|
207 |
+
mixs = mixed_x_layer.split([
|
208 |
+
q_k_v_len[0],
|
209 |
+
q_k_v_len[0]+q_k_v_len[1],
|
210 |
+
q_k_v_len[0]+q_k_v_len[1]+q_k_v_len[2],
|
211 |
+
],
|
212 |
+
axis=-1,
|
213 |
+
)
|
214 |
+
|
215 |
+
query_layer, key_layer, value_layer = mixs[0], mixs[1], mixs[2]
|
216 |
+
query_layer = query_layer.reshape(
|
217 |
+
query_layer.shape[:-1] + (self.num_attention_heads_per_partition, self.hidden_size_per_attention_head)
|
218 |
+
)
|
219 |
+
key_layer = key_layer.reshape( key_layer.shape[:-1] + (self.num_multi_query_groups_per_partition, self.hidden_size_per_attention_head))
|
220 |
+
value_layer = value_layer.reshape(
|
221 |
+
value_layer.shape[:-1]
|
222 |
+
+ (self.num_multi_query_groups_per_partition, self.hidden_size_per_attention_head)
|
223 |
+
)
|
224 |
+
else:
|
225 |
+
new_tensor_shape = mixed_x_layer.shape[:-1] + \
|
226 |
+
(self.num_attention_heads_per_partition,
|
227 |
+
3 * self.hidden_size_per_attention_head)
|
228 |
+
mixed_x_layer = mixed_x_layer.reshape(*new_tensor_shape)
|
229 |
+
|
230 |
+
# [b, sq, np, 3 * hn] --> 3 [b, sq, np, hn]
|
231 |
+
(query_layer, key_layer, value_layer) = mx.split_along_last_dim(mixed_x_layer, 3)
|
232 |
+
|
233 |
+
# [b, sq, np, hn] -> [b, np, sq, hn]
|
234 |
+
query_layer, key_layer, value_layer = [k.transpose((0,2,1,3)) for k in [query_layer, key_layer, value_layer]]
|
235 |
+
|
236 |
+
# apply relative positional encoding (rotary embedding)
|
237 |
+
if rotary_pos_emb is not None:
|
238 |
+
query_layer = apply_rotary_pos_emb(query_layer, rotary_pos_emb)
|
239 |
+
key_layer = apply_rotary_pos_emb(key_layer, rotary_pos_emb)
|
240 |
+
|
241 |
+
|
242 |
+
# adjust key and value for inference
|
243 |
+
if use_cache:
|
244 |
+
key_layer, value_layer = kv_cache.update_and_fetch(key_layer, value_layer)
|
245 |
+
else:
|
246 |
+
kv_cache = None
|
247 |
+
|
248 |
+
# if self.multi_query_attention:
|
249 |
+
# # key_layer = key_layer.unsqueeze(2)
|
250 |
+
# key_layer = mx.expand_dims(key_layer,2)
|
251 |
+
# key_layer_shape = key_layer.shape
|
252 |
+
# key_layer = mx.broadcast_to(key_layer,[
|
253 |
+
# key_layer_shape[0], key_layer_shape[1], self.num_attention_heads_per_partition // self.num_multi_query_groups_per_partition, key_layer_shape[3], key_layer_shape[4]]
|
254 |
+
# )
|
255 |
+
# key_layer = key_layer.reshape(
|
256 |
+
# key_layer.shape[:1] + (self.num_attention_heads_per_partition,) + key_layer.shape[3:]
|
257 |
+
# )
|
258 |
+
|
259 |
+
# # value_layer = value_layer.unsqueeze(2)
|
260 |
+
# value_layer = mx.expand_dims(value_layer,2)
|
261 |
+
# value_layer_shape = value_layer.shape
|
262 |
+
# value_layer = mx.broadcast_to(value_layer,[
|
263 |
+
# value_layer_shape[0], value_layer_shape[1], self.num_attention_heads_per_partition // self.num_multi_query_groups_per_partition, value_layer_shape[3], value_layer_shape[4]]
|
264 |
+
# )
|
265 |
+
# value_layer = value_layer.reshape(
|
266 |
+
# value_layer.shape[:1] + (self.num_attention_heads_per_partition,) + value_layer.shape[3:]
|
267 |
+
# )
|
268 |
+
|
269 |
+
# ==================================
|
270 |
+
# core attention computation
|
271 |
+
# ==================================
|
272 |
+
context_layer = self.core_attention(query_layer, key_layer, value_layer, attention_mask)
|
273 |
+
|
274 |
+
# =================
|
275 |
+
# Output. [sq, b, h]
|
276 |
+
# =================
|
277 |
+
|
278 |
+
output = self.dense(context_layer)
|
279 |
+
|
280 |
+
return output
|
281 |
+
|
282 |
+
class MLP(nn.Module):
|
283 |
+
def __init__(self, args: ModelArgs):
|
284 |
+
super().__init__()
|
285 |
+
|
286 |
+
self.add_bias = args.add_bias_linear
|
287 |
+
|
288 |
+
# Project to 4h. If using swiglu double the output width, see https://arxiv.org/pdf/2002.05202.pdf
|
289 |
+
self.dense_h_to_4h = nn.Linear(
|
290 |
+
args.hidden_size,
|
291 |
+
args.ffn_hidden_size * 2,
|
292 |
+
bias=self.add_bias,
|
293 |
+
)
|
294 |
+
|
295 |
+
def swiglu(x):
|
296 |
+
x = mx.split(x, 2, axis=-1)
|
297 |
+
return nn.silu(x[0]) * x[1]
|
298 |
+
|
299 |
+
self.activation_func = swiglu
|
300 |
+
|
301 |
+
# Project back to h.
|
302 |
+
self.dense_4h_to_h = nn.Linear(
|
303 |
+
args.ffn_hidden_size,
|
304 |
+
args.hidden_size,
|
305 |
+
bias=self.add_bias,
|
306 |
+
)
|
307 |
+
|
308 |
+
def __call__(self, hidden_states):
|
309 |
+
# [s, b, 4hp]
|
310 |
+
intermediate_parallel = self.dense_h_to_4h(hidden_states)
|
311 |
+
intermediate_parallel = self.activation_func(intermediate_parallel)
|
312 |
+
# [s, b, h]
|
313 |
+
output = self.dense_4h_to_h(intermediate_parallel)
|
314 |
+
return output
|
315 |
+
|
316 |
+
|
317 |
+
class GLMBlock(nn.Module):
|
318 |
+
def __init__(self, args: ModelArgs, layer_number):
|
319 |
+
super(GLMBlock, self).__init__()
|
320 |
+
self.layer_number = layer_number
|
321 |
+
|
322 |
+
self.apply_residual_connection_post_layernorm = args.apply_residual_connection_post_layernorm
|
323 |
+
|
324 |
+
self.fp32_residual_connection = args.fp32_residual_connection
|
325 |
+
|
326 |
+
LayerNormFunc = nn.RMSNorm if args.rmsnorm else nn.LayerNorm
|
327 |
+
# Layernorm on the input data.
|
328 |
+
self.input_layernorm = LayerNormFunc(args.hidden_size, eps=args.layernorm_epsilon)
|
329 |
+
|
330 |
+
# Self attention.
|
331 |
+
self.self_attention = SelfAttention(args, layer_number)
|
332 |
+
self.hidden_dropout = args.hidden_dropout
|
333 |
+
|
334 |
+
self.dropout = nn.Dropout(self.hidden_dropout)
|
335 |
+
|
336 |
+
# Layernorm on the attention output
|
337 |
+
self.post_attention_layernorm = LayerNormFunc(args.hidden_size, eps=args.layernorm_epsilon)
|
338 |
+
|
339 |
+
# MLP
|
340 |
+
self.mlp = MLP(args)
|
341 |
+
|
342 |
+
def __call__(
|
343 |
+
self, hidden_states, attention_mask, rotary_pos_emb, kv_cache=None, use_cache=True,
|
344 |
+
):
|
345 |
+
# hidden_states: [s, b, h]
|
346 |
+
|
347 |
+
# Layer norm at the beginning of the transformer layer.
|
348 |
+
layernorm_output = self.input_layernorm(hidden_states)
|
349 |
+
# Self attention.
|
350 |
+
attention_output = self.self_attention(
|
351 |
+
layernorm_output,
|
352 |
+
attention_mask,
|
353 |
+
rotary_pos_emb,
|
354 |
+
kv_cache=kv_cache,
|
355 |
+
use_cache=use_cache
|
356 |
+
)
|
357 |
+
|
358 |
+
# Residual connection.
|
359 |
+
if self.apply_residual_connection_post_layernorm:
|
360 |
+
residual = layernorm_output
|
361 |
+
else:
|
362 |
+
residual = hidden_states
|
363 |
+
|
364 |
+
layernorm_input = self.dropout(attention_output)
|
365 |
+
layernorm_input = residual + layernorm_input
|
366 |
+
|
367 |
+
# Layer norm post the self attention.
|
368 |
+
layernorm_output = self.post_attention_layernorm(layernorm_input)
|
369 |
+
|
370 |
+
# MLP.
|
371 |
+
mlp_output = self.mlp(layernorm_output)
|
372 |
+
|
373 |
+
# Second residual connection.
|
374 |
+
if self.apply_residual_connection_post_layernorm:
|
375 |
+
residual = layernorm_output
|
376 |
+
else:
|
377 |
+
residual = layernorm_input
|
378 |
+
|
379 |
+
output = self.dropout(mlp_output)
|
380 |
+
output = residual + output
|
381 |
+
|
382 |
+
return output
|
383 |
+
|
384 |
+
class GLMTransformer(nn.Module):
|
385 |
+
def __init__(self, args: ModelArgs):
|
386 |
+
super().__init__()
|
387 |
+
|
388 |
+
self.fp32_residual_connection = args.fp32_residual_connection
|
389 |
+
self.post_layer_norm = args.post_layer_norm
|
390 |
+
|
391 |
+
# Number of layers.
|
392 |
+
self.num_layers = args.num_layers
|
393 |
+
|
394 |
+
# Transformer layers.
|
395 |
+
def build_layer(layer_number):
|
396 |
+
return GLMBlock(args, layer_number)
|
397 |
+
|
398 |
+
self.layers = [build_layer(i + 1) for i in range(self.num_layers)]
|
399 |
+
|
400 |
+
if self.post_layer_norm:
|
401 |
+
LayerNormFunc = nn.RMSNorm if args.rmsnorm else nn.LayerNorm
|
402 |
+
# Final layer norm before output.
|
403 |
+
self.final_layernorm = LayerNormFunc(args.hidden_size, eps=args.layernorm_epsilon)
|
404 |
+
|
405 |
+
self.gradient_checkpointing = False
|
406 |
+
|
407 |
+
def _get_layer(self, layer_number):
|
408 |
+
return self.layers[layer_number]
|
409 |
+
|
410 |
+
def __call__(
|
411 |
+
self, hidden_states, attention_mask, rotary_pos_emb, kv_caches=None,
|
412 |
+
use_cache: Optional[bool] = True,
|
413 |
+
):
|
414 |
+
if not kv_caches:
|
415 |
+
kv_caches = [None for _ in range(self.num_layers)]
|
416 |
+
|
417 |
+
for index in range(self.num_layers):
|
418 |
+
layer = self._get_layer(index)
|
419 |
+
layer_ret = layer(
|
420 |
+
hidden_states,
|
421 |
+
attention_mask,
|
422 |
+
rotary_pos_emb,
|
423 |
+
kv_cache=kv_caches[index],
|
424 |
+
use_cache=use_cache
|
425 |
+
)
|
426 |
+
hidden_states = layer_ret
|
427 |
+
|
428 |
+
# Final layer norm.
|
429 |
+
if self.post_layer_norm:
|
430 |
+
hidden_states = self.final_layernorm(hidden_states)
|
431 |
+
|
432 |
+
return hidden_states
|
433 |
+
|
434 |
+
class Embedding(nn.Module):
|
435 |
+
def __init__(self, args: ModelArgs):
|
436 |
+
super().__init__()
|
437 |
+
|
438 |
+
self.hidden_size = args.hidden_size
|
439 |
+
# Word embeddings (parallel).
|
440 |
+
self.word_embeddings = nn.Embedding(
|
441 |
+
args.padded_vocab_size,
|
442 |
+
self.hidden_size,
|
443 |
+
)
|
444 |
+
self.fp32_residual_connection = args.fp32_residual_connection
|
445 |
+
|
446 |
+
def __call__(self, input_ids):
|
447 |
+
# Embeddings.
|
448 |
+
words_embeddings = self.word_embeddings(input_ids)
|
449 |
+
embeddings = words_embeddings
|
450 |
+
# If the input flag for fp32 residual connection is set, convert for float.
|
451 |
+
if self.fp32_residual_connection:
|
452 |
+
embeddings = embeddings.float()
|
453 |
+
return embeddings
|
454 |
+
|
455 |
+
|
456 |
+
class ChatGLMModel(nn.Module):
|
457 |
+
def __init__(self, args: ModelArgs):
|
458 |
+
super().__init__()
|
459 |
+
|
460 |
+
self.embedding = Embedding(args)
|
461 |
+
self.num_layers = args.num_layers
|
462 |
+
self.multi_query_group_num = args.multi_query_group_num
|
463 |
+
|
464 |
+
self.kv_channels = args.kv_channels
|
465 |
+
self.use_cache = args.use_cache
|
466 |
+
self.use_return_dict = False
|
467 |
+
self.output_hidden_states = False
|
468 |
+
|
469 |
+
# Rotary positional embeddings
|
470 |
+
self.seq_length = args.seq_length
|
471 |
+
rotary_dim = (
|
472 |
+
args.hidden_size // args.num_attention_heads if args.kv_channels is None else args.kv_channels
|
473 |
+
)
|
474 |
+
|
475 |
+
self.rotary_pos_emb = RotaryEmbedding(rotary_dim // 2, rope_ratio=args.rope_ratio, original_impl=args.original_rope,dtype=args.torch_dtype)
|
476 |
+
self.encoder = GLMTransformer(args)
|
477 |
+
self.output_layer = nn.Linear(args.hidden_size, args.padded_vocab_size, bias=False)
|
478 |
+
|
479 |
+
self.new_position_id = None
|
480 |
+
self.is_first_forward = True
|
481 |
+
|
482 |
+
def get_input_embeddings(self):
|
483 |
+
return self.embedding.word_embeddings
|
484 |
+
|
485 |
+
def set_input_embeddings(self, value):
|
486 |
+
self.embedding.word_embeddings = value
|
487 |
+
|
488 |
+
def get_masks(self, input_ids, past_key_values, padding_mask=None):
|
489 |
+
batch_size, seq_length = input_ids.shape
|
490 |
+
full_attention_mask = mx.ones((batch_size, seq_length, seq_length), dtype=input_ids.dtype)
|
491 |
+
full_attention_mask = mx.tril(full_attention_mask)
|
492 |
+
past_length = 0
|
493 |
+
if past_key_values and past_key_values[0].keys is not None:
|
494 |
+
past_length = past_key_values[0].offset
|
495 |
+
if past_length:
|
496 |
+
full_attention_mask = mx.concatenate((mx.ones((batch_size, seq_length, past_length), dtype=input_ids.dtype),
|
497 |
+
full_attention_mask), axis=-1)
|
498 |
+
if padding_mask is not None:
|
499 |
+
full_attention_mask = full_attention_mask * mx.expand_dims(padding_mask,1)
|
500 |
+
if not past_length and padding_mask is not None:
|
501 |
+
full_attention_mask -= mx.expand_dims(padding_mask,-1) - 1
|
502 |
+
full_attention_mask = (full_attention_mask < 0.5)
|
503 |
+
full_attention_mask = mx.expand_dims(full_attention_mask,1)
|
504 |
+
return full_attention_mask
|
505 |
+
|
506 |
+
def get_position_ids(self, input_ids):
|
507 |
+
batch_size, seq_length = input_ids.shape
|
508 |
+
position_ids = mx.arange(seq_length, dtype=mx.int32)
|
509 |
+
position_ids = mx.broadcast_to(position_ids, (batch_size, seq_length))
|
510 |
+
return position_ids
|
511 |
+
|
512 |
+
def __call__(
|
513 |
+
self,
|
514 |
+
input_ids,
|
515 |
+
position_ids: Optional[mx.array] = None,
|
516 |
+
attention_mask: Optional[mx.array] = None,
|
517 |
+
full_attention_mask: Optional[mx.array] = None,
|
518 |
+
past_key_values: Optional[Tuple[Tuple[mx.array, mx.array], ...]] = None,
|
519 |
+
inputs_embeds: Optional[mx.array] = None,
|
520 |
+
use_cache: Optional[bool] = None,
|
521 |
+
):
|
522 |
+
|
523 |
+
# prepare_inputs_for_generation
|
524 |
+
if self.new_position_id is None:
|
525 |
+
position_ids = self.get_position_ids(input_ids)
|
526 |
+
else:
|
527 |
+
position_ids = self.new_position_id
|
528 |
+
|
529 |
+
new_position_id = position_ids[..., -1:]
|
530 |
+
# print(f"== new_position_id:{new_position_id}")
|
531 |
+
new_position_id += 1
|
532 |
+
# print(f"== new_position_id:{new_position_id}")
|
533 |
+
new_position_id = mx.concatenate(
|
534 |
+
[position_ids, new_position_id], axis=-1
|
535 |
+
)
|
536 |
+
# print(f"== new_position_id:{new_position_id}")
|
537 |
+
self.new_position_id = new_position_id
|
538 |
+
|
539 |
+
if past_key_values and past_key_values[0].offset > 0: # TODO: check pre_seq
|
540 |
+
position_ids = position_ids[..., -1:]
|
541 |
+
input_ids = input_ids[:, -1:]
|
542 |
+
|
543 |
+
# print(f"== position_ids:{position_ids} input_ids:{input_ids}")
|
544 |
+
batch_size, seq_length = input_ids.shape
|
545 |
+
|
546 |
+
if inputs_embeds is None:
|
547 |
+
inputs_embeds = self.embedding(input_ids)
|
548 |
+
|
549 |
+
# Rotary positional embeddings
|
550 |
+
rotary_pos_emb = self.rotary_pos_emb(self.seq_length)
|
551 |
+
if position_ids is not None:
|
552 |
+
rotary_pos_emb = rotary_pos_emb[position_ids]
|
553 |
+
else:
|
554 |
+
rotary_pos_emb = rotary_pos_emb[None, :seq_length]
|
555 |
+
# print(f"== rotary_pos_emb:{rotary_pos_emb.shape}")
|
556 |
+
|
557 |
+
# Run encoder.
|
558 |
+
hidden_states = self.encoder(
|
559 |
+
inputs_embeds, full_attention_mask, rotary_pos_emb=rotary_pos_emb,
|
560 |
+
kv_caches=past_key_values, use_cache=use_cache
|
561 |
+
)
|
562 |
+
|
563 |
+
return hidden_states
|
564 |
+
|
565 |
+
|
566 |
+
class Model(nn.Module):
|
567 |
+
def __init__(self, args: ModelArgs):
|
568 |
+
super().__init__()
|
569 |
+
self.args = args
|
570 |
+
self.model_type = args.model_type
|
571 |
+
self.transformer = ChatGLMModel(args)
|
572 |
+
|
573 |
+
def __call__(
|
574 |
+
self,
|
575 |
+
inputs: mx.array,
|
576 |
+
cache=None,
|
577 |
+
):
|
578 |
+
out = self.transformer(inputs, None, None, None, cache, None, True)
|
579 |
+
if self.args.tie_word_embeddings:
|
580 |
+
out = self.model.embedding.as_linear(out)
|
581 |
+
else:
|
582 |
+
out = self.model.output_layer(out)
|
583 |
+
return out
|
584 |
+
|
585 |
+
def sanitize(self, weights):
|
586 |
+
# Remove unused precomputed rotary freqs
|
587 |
+
return {
|
588 |
+
k: v for k, v in weights.items() if "transformer.rotary_pos_emb.inv_freq" not in k
|
589 |
+
}
|
590 |
+
# return weights
|
591 |
+
|
592 |
+
@property
|
593 |
+
def layers(self):
|
594 |
+
return self.model.encoder.layers
|
595 |
+
|
596 |
+
@property
|
597 |
+
def head_dim(self):
|
598 |
+
return self.args.hidden_size // self.args.num_attention_heads
|
599 |
+
|
600 |
+
@property
|
601 |
+
def n_kv_heads(self):
|
602 |
+
return self.args.multi_query_group_num
|
603 |
+
|
604 |
+
@property
|
605 |
+
def model(self):
|
606 |
+
return self.transformer
|
607 |
+
|
608 |
+
|