proteinglm
commited on
Commit
•
487d5a5
1
Parent(s):
d679c98
Upload 9 files
Browse files- config.json +52 -0
- configuration_proteinglm.py +86 -0
- generation_config.json +4 -0
- modeling_proteinglm.py +1572 -0
- quantization.py +188 -0
- special_tokens_map.json +17 -0
- tokenization_proteinglm.py +140 -0
- tokenizer.model +3 -0
- tokenizer_config.json +100 -0
config.json
ADDED
@@ -0,0 +1,52 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_name_or_path": "proteinglm-100b-int4",
|
3 |
+
"add_bias_linear": true,
|
4 |
+
"add_qkv_bias": true,
|
5 |
+
"apply_query_key_layer_scaling": true,
|
6 |
+
"apply_residual_connection_post_layernorm": true,
|
7 |
+
"architectures": [
|
8 |
+
"ProteinGLMModel"
|
9 |
+
],
|
10 |
+
"attention_dropout": 0.0,
|
11 |
+
"attention_softmax_in_fp32": true,
|
12 |
+
"auto_map": {
|
13 |
+
"AutoConfig": "configuration_proteinglm.ProteinGLMConfig",
|
14 |
+
"AutoModel": "modeling_proteinglm.ProteinGLMForMaskedLM",
|
15 |
+
"AutoModelForCausalLM": "modeling_proteinglm.ProteinGLMForCasualLM",
|
16 |
+
"AutoModelForMaskedLM": "modeling_proteinglm.ProteinGLMForMaskedLM",
|
17 |
+
"AutoModelForSequenceClassification": "modeling_proteinglm.ProteinGLMForSequenceClassification",
|
18 |
+
"AutoModelForTokenClassification": "modeling_proteinglm.ProteinGLMForTokenClassification"
|
19 |
+
},
|
20 |
+
"bias_dropout_fusion": true,
|
21 |
+
"deepnorm": true,
|
22 |
+
"experts_per_token": 0,
|
23 |
+
"ffn_hidden_size": 31744,
|
24 |
+
"fp32_residual_connection": false,
|
25 |
+
"glu_activation": "geglu",
|
26 |
+
"initializer_range": 0.02,
|
27 |
+
"head_num": 1,
|
28 |
+
"hidden_dropout": 0.0,
|
29 |
+
"hidden_size": 10240,
|
30 |
+
"is_causal": false,
|
31 |
+
"use_cache":true,
|
32 |
+
"kv_channels": 128,
|
33 |
+
"layernorm_epsilon": 1e-05,
|
34 |
+
"model_type": "ProteinGLM",
|
35 |
+
"moe": false,
|
36 |
+
"multi_query_attention": false,
|
37 |
+
"multi_query_group_num": 1,
|
38 |
+
"num_attention_heads": 80,
|
39 |
+
"num_experts": 0,
|
40 |
+
"num_layers": 72,
|
41 |
+
"padded_vocab_size": 128,
|
42 |
+
"post_layer_norm": true,
|
43 |
+
"quantization_bit": 4,
|
44 |
+
"rmsnorm": false,
|
45 |
+
"rotary_embedding_2d": true,
|
46 |
+
"seq_length": 2048,
|
47 |
+
"torch_dtype": "float32",
|
48 |
+
"transformers_version": "4.41.2",
|
49 |
+
"untie_head": false,
|
50 |
+
"use_pytorch_sdpa": true,
|
51 |
+
"vocab_size": 128
|
52 |
+
}
|
configuration_proteinglm.py
ADDED
@@ -0,0 +1,86 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import PretrainedConfig
|
2 |
+
|
3 |
+
|
4 |
+
class ProteinGLMConfig(PretrainedConfig):
|
5 |
+
model_type = "ProteinGLM"
|
6 |
+
def __init__(
|
7 |
+
self,
|
8 |
+
num_layers=72,
|
9 |
+
padded_vocab_size=128,
|
10 |
+
hidden_size=10240,
|
11 |
+
ffn_hidden_size=31744,
|
12 |
+
kv_channels=128,
|
13 |
+
num_attention_heads=80,
|
14 |
+
seq_length=2048,
|
15 |
+
hidden_dropout=0.0,
|
16 |
+
attention_dropout=0.0,
|
17 |
+
layernorm_epsilon=1e-5,
|
18 |
+
initializer_range=0.02,
|
19 |
+
glu_activation='geglu',
|
20 |
+
rmsnorm=False,
|
21 |
+
deepnorm=True,
|
22 |
+
apply_residual_connection_post_layernorm=True,
|
23 |
+
post_layer_norm=True,
|
24 |
+
add_bias_linear=True,
|
25 |
+
add_qkv_bias=True,
|
26 |
+
bias_dropout_fusion=True,
|
27 |
+
multi_query_attention=False,
|
28 |
+
multi_query_group_num=1,
|
29 |
+
apply_query_key_layer_scaling=True,
|
30 |
+
attention_softmax_in_fp32=True,
|
31 |
+
fp32_residual_connection=False,
|
32 |
+
quantization_bit=0,
|
33 |
+
rotary_embedding_2d=True,
|
34 |
+
use_pytorch_sdpa=True,
|
35 |
+
is_causal=False,
|
36 |
+
use_cache=True,
|
37 |
+
moe=False,
|
38 |
+
num_experts=0,
|
39 |
+
experts_per_token=0,
|
40 |
+
untie_head=False,
|
41 |
+
head_num=1,
|
42 |
+
**kwargs
|
43 |
+
):
|
44 |
+
|
45 |
+
if not deepnorm and apply_residual_connection_post_layernorm:
|
46 |
+
print(f"Warning: deepnorm is False and apply_residual_connection_post_layernorm is True")
|
47 |
+
|
48 |
+
if deepnorm:
|
49 |
+
apply_residual_connection_post_layernorm = True
|
50 |
+
|
51 |
+
self.num_layers = num_layers
|
52 |
+
self.vocab_size = padded_vocab_size
|
53 |
+
self.padded_vocab_size = padded_vocab_size
|
54 |
+
self.hidden_size = hidden_size
|
55 |
+
self.ffn_hidden_size = ffn_hidden_size
|
56 |
+
self.kv_channels = kv_channels
|
57 |
+
self.num_attention_heads = num_attention_heads
|
58 |
+
self.seq_length = seq_length
|
59 |
+
self.hidden_dropout = hidden_dropout
|
60 |
+
self.attention_dropout = attention_dropout
|
61 |
+
self.layernorm_epsilon = layernorm_epsilon
|
62 |
+
self.glu_activation = glu_activation
|
63 |
+
self.initializer_range = initializer_range
|
64 |
+
self.rmsnorm = rmsnorm
|
65 |
+
self.deepnorm = deepnorm
|
66 |
+
self.apply_residual_connection_post_layernorm = apply_residual_connection_post_layernorm
|
67 |
+
self.post_layer_norm = post_layer_norm
|
68 |
+
self.add_bias_linear = add_bias_linear
|
69 |
+
self.add_qkv_bias = add_qkv_bias
|
70 |
+
self.bias_dropout_fusion = bias_dropout_fusion
|
71 |
+
self.multi_query_attention = multi_query_attention
|
72 |
+
self.multi_query_group_num = multi_query_group_num
|
73 |
+
self.apply_query_key_layer_scaling = apply_query_key_layer_scaling
|
74 |
+
self.attention_softmax_in_fp32 = attention_softmax_in_fp32
|
75 |
+
self.fp32_residual_connection = fp32_residual_connection
|
76 |
+
self.quantization_bit = quantization_bit
|
77 |
+
self.rotary_embedding_2d = rotary_embedding_2d
|
78 |
+
self.is_causal = is_causal
|
79 |
+
self.use_cache=use_cache
|
80 |
+
self.use_pytorch_sdpa = use_pytorch_sdpa
|
81 |
+
self.moe = moe
|
82 |
+
self.num_experts = num_experts
|
83 |
+
self.experts_per_token = experts_per_token
|
84 |
+
self.untie_head = untie_head
|
85 |
+
self.head_num=head_num
|
86 |
+
super().__init__(**kwargs)
|
generation_config.json
ADDED
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_from_model_config": true,
|
3 |
+
"transformers_version": "4.41.2"
|
4 |
+
}
|
modeling_proteinglm.py
ADDED
@@ -0,0 +1,1572 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
""" PyTorch ProteinGLM model. """
|
2 |
+
|
3 |
+
import math
|
4 |
+
import copy
|
5 |
+
import warnings
|
6 |
+
import re
|
7 |
+
import sys
|
8 |
+
import os
|
9 |
+
import pathlib
|
10 |
+
import time
|
11 |
+
import random
|
12 |
+
import numpy as np
|
13 |
+
from tqdm.auto import tqdm
|
14 |
+
|
15 |
+
import torch, deepspeed
|
16 |
+
import torch.utils.checkpoint
|
17 |
+
import torch.nn.functional as F
|
18 |
+
from torch import nn
|
19 |
+
from torch.nn import CrossEntropyLoss, LayerNorm, MSELoss, BCEWithLogitsLoss
|
20 |
+
from torch.nn.utils import skip_init
|
21 |
+
from typing import Optional, Tuple, Union, List, Callable, Dict, Any
|
22 |
+
from copy import deepcopy
|
23 |
+
from collections import namedtuple
|
24 |
+
|
25 |
+
from transformers.modeling_outputs import (
|
26 |
+
BaseModelOutputWithPast,
|
27 |
+
MaskedLMOutput,
|
28 |
+
CausalLMOutputWithPast,
|
29 |
+
SequenceClassifierOutput,
|
30 |
+
TokenClassifierOutput
|
31 |
+
)
|
32 |
+
from transformers import PreTrainedModel
|
33 |
+
from transformers.utils import logging
|
34 |
+
from transformers.generation.logits_process import LogitsProcessor
|
35 |
+
from transformers.generation.utils import LogitsProcessorList, StoppingCriteriaList, GenerationConfig, ModelOutput
|
36 |
+
|
37 |
+
from .configuration_proteinglm import ProteinGLMConfig
|
38 |
+
from .quantization import quantize
|
39 |
+
|
40 |
+
def get_checkpoint_fn():
|
41 |
+
if deepspeed.checkpointing.is_configured():
|
42 |
+
checkpoint = deepspeed.checkpointing.checkpoint
|
43 |
+
else:
|
44 |
+
checkpoint = torch.utils.checkpoint.checkpoint
|
45 |
+
return checkpoint
|
46 |
+
|
47 |
+
# flags required to enable jit fusion kernels
|
48 |
+
|
49 |
+
if sys.platform != 'darwin':
|
50 |
+
torch._C._jit_set_profiling_mode(False)
|
51 |
+
torch._C._jit_set_profiling_executor(False)
|
52 |
+
torch._C._jit_override_can_fuse_on_cpu(True)
|
53 |
+
torch._C._jit_override_can_fuse_on_gpu(True)
|
54 |
+
|
55 |
+
logger = logging.get_logger(__name__)
|
56 |
+
|
57 |
+
_CHECKPOINT_FOR_DOC = "proteinglm-100b-int4"
|
58 |
+
_CONFIG_FOR_DOC = "ProteinGLMConfig"
|
59 |
+
DeepNormCoefficients = namedtuple("DeepNormCoefficients", ["alpha", "beta"])
|
60 |
+
|
61 |
+
def default_init(cls, *args, **kwargs):
|
62 |
+
return cls(*args, **kwargs)
|
63 |
+
|
64 |
+
|
65 |
+
def get_deepnorm_coefficients(config: ProteinGLMConfig):
|
66 |
+
"""
|
67 |
+
DeepNorm coefficients from : https://kexue.fm/archives/8978
|
68 |
+
"""
|
69 |
+
num_layers = config.num_layers
|
70 |
+
return DeepNormCoefficients(alpha=(2 * num_layers) ** 0.5, beta=(2 * num_layers) ** -0.5)
|
71 |
+
|
72 |
+
|
73 |
+
class InvalidScoreLogitsProcessor(LogitsProcessor):
|
74 |
+
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
|
75 |
+
if torch.isnan(scores).any() or torch.isinf(scores).any():
|
76 |
+
scores.zero_()
|
77 |
+
scores[..., 5] = 5e4
|
78 |
+
return scores
|
79 |
+
|
80 |
+
|
81 |
+
def split_tensor_along_last_dim(
|
82 |
+
tensor: torch.Tensor,
|
83 |
+
num_partitions: int,
|
84 |
+
contiguous_split_chunks: bool = False,
|
85 |
+
) -> List[torch.Tensor]:
|
86 |
+
"""Split a tensor along its last dimension.
|
87 |
+
|
88 |
+
Arguments:
|
89 |
+
tensor: input tensor.
|
90 |
+
num_partitions: number of partitions to split the tensor
|
91 |
+
contiguous_split_chunks: If True, make each chunk contiguous
|
92 |
+
in memory.
|
93 |
+
|
94 |
+
Returns:
|
95 |
+
A list of Tensors
|
96 |
+
"""
|
97 |
+
# Get the size and dimension.
|
98 |
+
last_dim = tensor.dim() - 1
|
99 |
+
last_dim_size = tensor.size()[last_dim] // num_partitions
|
100 |
+
# Split.
|
101 |
+
tensor_list = torch.split(tensor, last_dim_size, dim=last_dim)
|
102 |
+
# Note: torch.split does not create contiguous tensors by default.
|
103 |
+
if contiguous_split_chunks:
|
104 |
+
return tuple(chunk.contiguous() for chunk in tensor_list)
|
105 |
+
|
106 |
+
return tensor_list
|
107 |
+
|
108 |
+
class RotaryEmbedding(torch.nn.Module):
|
109 |
+
|
110 |
+
def __init__(self, dim, base=10000, precision=torch.half, learnable=False):
|
111 |
+
super().__init__()
|
112 |
+
inv_freq = 1. / (base ** (torch.arange(0, dim, 2).float() / dim)).to(precision)
|
113 |
+
self.dim = dim
|
114 |
+
self.base = base
|
115 |
+
self.learnable = learnable
|
116 |
+
if learnable:
|
117 |
+
self.inv_freq = torch.nn.Parameter(inv_freq)
|
118 |
+
self.max_seq_len_cached = None
|
119 |
+
else:
|
120 |
+
self.register_buffer('inv_freq', inv_freq)
|
121 |
+
self.max_seq_len_cached = None
|
122 |
+
self.cos_cached = None
|
123 |
+
self.sin_cached = None
|
124 |
+
self.precision = precision
|
125 |
+
|
126 |
+
def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs):
|
127 |
+
if f'{prefix}inv_freq' in state_dict:
|
128 |
+
super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs)
|
129 |
+
else:
|
130 |
+
self.inv_freq.copy_(1. / (self.base ** (torch.arange(0, self.dim, 2).float() / self.dim)).to(self.precision))
|
131 |
+
|
132 |
+
def forward(self, x, seq_dim=1, seq_len=None):
|
133 |
+
if seq_len is None:
|
134 |
+
seq_len = x.shape[seq_dim]
|
135 |
+
if self.max_seq_len_cached is None or (seq_len > self.max_seq_len_cached):
|
136 |
+
self.max_seq_len_cached = None if self.learnable else seq_len
|
137 |
+
t = torch.arange(seq_len, device=x.device, dtype=torch.float32)
|
138 |
+
freqs = torch.einsum('i,j->ij', t, self.inv_freq.to(x.device))
|
139 |
+
# Different from paper, but it uses a different permutation in order to obtain the same calculation
|
140 |
+
emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
|
141 |
+
if self.precision == torch.bfloat16 or self.precision == torch.half:
|
142 |
+
emb = emb.float()
|
143 |
+
# [sx, 1 (b * np), hn]
|
144 |
+
cos_cached = emb.cos()[:, None, :]
|
145 |
+
sin_cached = emb.sin()[:, None, :]
|
146 |
+
if self.precision == torch.bfloat16:
|
147 |
+
cos_cached = cos_cached.bfloat16()
|
148 |
+
sin_cached = sin_cached.bfloat16()
|
149 |
+
elif self.precision == torch.half:
|
150 |
+
cos_cached = cos_cached.half()
|
151 |
+
sin_cached = sin_cached.half()
|
152 |
+
if self.learnable:
|
153 |
+
return cos_cached, sin_cached
|
154 |
+
self.cos_cached, self.sin_cached = cos_cached, sin_cached
|
155 |
+
return self.cos_cached[:seq_len, ...], self.sin_cached[:seq_len, ...]
|
156 |
+
|
157 |
+
def rotate_half(x):
|
158 |
+
x1, x2 = x[..., :x.shape[-1] // 2], x[..., x.shape[-1] // 2:]
|
159 |
+
return torch.cat((-x2, x1), dim=x1.ndim - 1) # dim=-1 triggers a bug in earlier torch versions
|
160 |
+
|
161 |
+
def assert_dim_check(tensor, ndim=None, shape=None):
|
162 |
+
if ndim is not None:
|
163 |
+
assert tensor.ndim == ndim, f"Exepct tensor.ndim={ndim}. gut got tensor.shape={tensor.shape}"
|
164 |
+
if shape is not None:
|
165 |
+
assert list(tensor.shape) == list(shape), f"Exepct tensor.shape={shape}. gut got tensor.shape={tensor.shape}"
|
166 |
+
|
167 |
+
def apply_rotary_pos_emb_index_torch(q, k, cos, sin, position_id): # jitting fails with bf16
|
168 |
+
# position_id: [sq, b], q, k: [sq, b, np, hn], cos: [sq, 1, hn] -> [sq, b, 1, hn]
|
169 |
+
cos, sin = F.embedding(position_id, cos.squeeze(1)).unsqueeze(2), \
|
170 |
+
F.embedding(position_id, sin.squeeze(1)).unsqueeze(2)
|
171 |
+
q, k = (q * cos) + (rotate_half(q) * sin), (k * cos) + (rotate_half(k) * sin)
|
172 |
+
return q, k
|
173 |
+
|
174 |
+
class RMSNorm(torch.nn.Module):
|
175 |
+
def __init__(self, normalized_shape, eps=1e-5, device=None, dtype=None, **kwargs):
|
176 |
+
super().__init__()
|
177 |
+
self.weight = torch.nn.Parameter(torch.empty(normalized_shape, device=device, dtype=dtype))
|
178 |
+
self.eps = eps
|
179 |
+
|
180 |
+
def forward(self, hidden_states: torch.Tensor):
|
181 |
+
input_dtype = hidden_states.dtype
|
182 |
+
variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
|
183 |
+
hidden_states = hidden_states * torch.rsqrt(variance + self.eps)
|
184 |
+
|
185 |
+
return (self.weight * hidden_states).to(input_dtype)
|
186 |
+
|
187 |
+
class CoreAttention(torch.nn.Module):
|
188 |
+
def __init__(self, config: ProteinGLMConfig, layer_number):
|
189 |
+
super(CoreAttention, self).__init__()
|
190 |
+
|
191 |
+
self.apply_query_key_layer_scaling = config.apply_query_key_layer_scaling
|
192 |
+
self.attention_softmax_in_fp32 = config.attention_softmax_in_fp32
|
193 |
+
if self.apply_query_key_layer_scaling:
|
194 |
+
self.attention_softmax_in_fp32 = True
|
195 |
+
self.layer_number = max(1, layer_number)
|
196 |
+
|
197 |
+
projection_size = config.kv_channels * config.num_attention_heads
|
198 |
+
|
199 |
+
# Per attention head and per partition values.
|
200 |
+
self.hidden_size_per_partition = projection_size
|
201 |
+
self.hidden_size_per_attention_head = projection_size // config.num_attention_heads
|
202 |
+
self.num_attention_heads_per_partition = config.num_attention_heads
|
203 |
+
|
204 |
+
coeff = None
|
205 |
+
self.norm_factor = math.sqrt(self.hidden_size_per_attention_head)
|
206 |
+
if self.apply_query_key_layer_scaling:
|
207 |
+
coeff = self.layer_number
|
208 |
+
self.norm_factor *= coeff
|
209 |
+
self.coeff = coeff
|
210 |
+
|
211 |
+
self.attention_dropout = torch.nn.Dropout(config.attention_dropout)
|
212 |
+
|
213 |
+
self.is_causal = config.is_causal
|
214 |
+
self.use_pytorch_sdpa = config.use_pytorch_sdpa
|
215 |
+
|
216 |
+
def forward(self, query_layer, key_layer, value_layer, attention_mask):
|
217 |
+
# query_layer, key_layer, value_layer: [seq_len, batch_size, num_heads, head_dim]
|
218 |
+
# import pdb; pdb.set_trace();
|
219 |
+
pytorch_major_version = int(torch.__version__.split('.')[0])
|
220 |
+
# assert pytorch_major_version >= 2, f"Expect PyTorch version > 2.0"
|
221 |
+
if pytorch_major_version >= 2 and self.use_pytorch_sdpa:
|
222 |
+
dropout_p = self.attention_dropout.p if self.training else 0
|
223 |
+
# [seq_len, batch_size, num_heads, head_dim] -> [batch_size, num_heads, seq_len, head_dim]
|
224 |
+
query_layer, key_layer, value_layer = [k.permute(1, 2, 0, 3) for k in [query_layer, key_layer, value_layer]]
|
225 |
+
# import pdb; pdb.set_trace();
|
226 |
+
if attention_mask is None and query_layer.shape[2] == key_layer.shape[2]:
|
227 |
+
# context_layer: [batch_size, num_heads, seq_len, head_dim]
|
228 |
+
context_layer = torch.nn.functional.scaled_dot_product_attention(query_layer, key_layer, value_layer, is_causal=self.is_causal, dropout_p=dropout_p)
|
229 |
+
else:
|
230 |
+
if (attention_mask is not None) and (attention_mask.dtype == torch.bool):
|
231 |
+
attention_mask = attention_mask.logical_not() ## DO NOT inplace operation!!!!
|
232 |
+
context_layer = torch.nn.functional.scaled_dot_product_attention(query_layer, key_layer, value_layer, attention_mask, dropout_p=dropout_p)
|
233 |
+
# [batch_size, num_heads, seq_len, head_dim] -> [seq_len, batch_size, num_heads, head_dim]
|
234 |
+
context_layer = context_layer.permute(2, 0, 1, 3)
|
235 |
+
# [seq_len, batch_size, 2560]
|
236 |
+
new_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size_per_partition,)
|
237 |
+
context_layer = context_layer.reshape(*new_context_layer_shape)
|
238 |
+
else:
|
239 |
+
# Raw attention scores
|
240 |
+
|
241 |
+
# [b, np, sq, sk]
|
242 |
+
output_size = (query_layer.size(1), query_layer.size(2), query_layer.size(0), key_layer.size(0))
|
243 |
+
|
244 |
+
# [sq, b, np, hn] -> [sq, b * np, hn]
|
245 |
+
query_layer = query_layer.view(output_size[2], output_size[0] * output_size[1], -1)
|
246 |
+
# [sk, b, np, hn] -> [sk, b * np, hn]
|
247 |
+
key_layer = key_layer.view(output_size[3], output_size[0] * output_size[1], -1)
|
248 |
+
|
249 |
+
# preallocting input tensor: [b * np, sq, sk]
|
250 |
+
matmul_input_buffer = torch.empty(
|
251 |
+
output_size[0] * output_size[1], output_size[2], output_size[3], dtype=query_layer.dtype,
|
252 |
+
device=query_layer.device
|
253 |
+
)
|
254 |
+
|
255 |
+
# Raw attention scores. [b * np, sq, sk]
|
256 |
+
matmul_result = torch.baddbmm(
|
257 |
+
matmul_input_buffer,
|
258 |
+
query_layer.transpose(0, 1), # [b * np, sq, hn]
|
259 |
+
key_layer.transpose(0, 1).transpose(1, 2), # [b * np, hn, sk]
|
260 |
+
beta=0.0,
|
261 |
+
alpha=(1.0 / self.norm_factor),
|
262 |
+
)
|
263 |
+
|
264 |
+
# change view to [b, np, sq, sk]
|
265 |
+
attention_scores = matmul_result.view(*output_size)
|
266 |
+
|
267 |
+
# ===========================
|
268 |
+
# Attention probs and dropout
|
269 |
+
# ===========================
|
270 |
+
|
271 |
+
# attention scores and attention mask [b, np, sq, sk]
|
272 |
+
if self.attention_softmax_in_fp32:
|
273 |
+
attention_scores = attention_scores.float()
|
274 |
+
if self.coeff is not None:
|
275 |
+
attention_scores = attention_scores * self.coeff
|
276 |
+
if self.is_causal and attention_mask is None and attention_scores.shape[2] == attention_scores.shape[3]:
|
277 |
+
attention_mask = torch.ones(output_size[0], 1, output_size[2], output_size[3],
|
278 |
+
device=attention_scores.device, dtype=torch.bool)
|
279 |
+
attention_mask.tril_()
|
280 |
+
attention_mask = ~attention_mask
|
281 |
+
if attention_mask is not None:
|
282 |
+
attention_scores = attention_scores.masked_fill(attention_mask, float("-inf"))
|
283 |
+
attention_probs = F.softmax(attention_scores, dim=-1)
|
284 |
+
attention_probs = attention_probs.type_as(value_layer)
|
285 |
+
|
286 |
+
# This is actually dropping out entire tokens to attend to, which might
|
287 |
+
# seem a bit unusual, but is taken from the original Transformer paper.
|
288 |
+
attention_probs = self.attention_dropout(attention_probs)
|
289 |
+
# =========================
|
290 |
+
# Context layer. [sq, b, hp]
|
291 |
+
# =========================
|
292 |
+
|
293 |
+
# value_layer -> context layer.
|
294 |
+
# [sk, b, np, hn] --> [b, np, sq, hn]
|
295 |
+
|
296 |
+
# context layer shape: [b, np, sq, hn]
|
297 |
+
output_size = (value_layer.size(1), value_layer.size(2), query_layer.size(0), value_layer.size(3))
|
298 |
+
# change view [sk, b * np, hn]
|
299 |
+
value_layer = value_layer.view(value_layer.size(0), output_size[0] * output_size[1], -1)
|
300 |
+
# change view [b * np, sq, sk]
|
301 |
+
attention_probs = attention_probs.view(output_size[0] * output_size[1], output_size[2], -1)
|
302 |
+
# matmul: [b * np, sq, hn]
|
303 |
+
context_layer = torch.bmm(attention_probs, value_layer.transpose(0, 1))
|
304 |
+
# change view [b, np, sq, hn]
|
305 |
+
context_layer = context_layer.view(*output_size)
|
306 |
+
# [b, np, sq, hn] --> [sq, b, np, hn]
|
307 |
+
context_layer = context_layer.permute(2, 0, 1, 3).contiguous()
|
308 |
+
# [sq, b, np, hn] --> [sq, b, hp]
|
309 |
+
new_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size_per_partition,)
|
310 |
+
context_layer = context_layer.view(*new_context_layer_shape)
|
311 |
+
|
312 |
+
return context_layer
|
313 |
+
|
314 |
+
|
315 |
+
class SelfAttention(torch.nn.Module):
|
316 |
+
"""Parallel self-attention layer abstract class.
|
317 |
+
|
318 |
+
Self-attention layer takes input with size [s, b, h]
|
319 |
+
and returns output of the same size.
|
320 |
+
"""
|
321 |
+
|
322 |
+
def __init__(self, config: ProteinGLMConfig, layer_number, device=None):
|
323 |
+
super(SelfAttention, self).__init__()
|
324 |
+
self.layer_number = max(1, layer_number)
|
325 |
+
|
326 |
+
self.projection_size = config.kv_channels * config.num_attention_heads
|
327 |
+
|
328 |
+
# Per attention head and per partition values.
|
329 |
+
self.hidden_size_per_attention_head = self.projection_size // config.num_attention_heads
|
330 |
+
self.num_attention_heads_per_partition = config.num_attention_heads
|
331 |
+
|
332 |
+
self.multi_query_attention = config.multi_query_attention
|
333 |
+
self.qkv_hidden_size = 3 * self.projection_size
|
334 |
+
if self.multi_query_attention:
|
335 |
+
self.num_multi_query_groups_per_partition = config.multi_query_group_num
|
336 |
+
self.qkv_hidden_size = (
|
337 |
+
self.projection_size + 2 * self.hidden_size_per_attention_head * config.multi_query_group_num
|
338 |
+
)
|
339 |
+
self.query_key_value = nn.Linear(config.hidden_size, self.qkv_hidden_size,
|
340 |
+
bias=config.add_bias_linear or config.add_qkv_bias,
|
341 |
+
device=device, **_config_to_kwargs(config)
|
342 |
+
)
|
343 |
+
|
344 |
+
self.core_attention = CoreAttention(config, self.layer_number)
|
345 |
+
|
346 |
+
# Output.
|
347 |
+
self.dense = nn.Linear(self.projection_size, config.hidden_size, bias=config.add_bias_linear, device=device, **_config_to_kwargs(config))
|
348 |
+
|
349 |
+
self.rotary_embedding_2d = config.rotary_embedding_2d
|
350 |
+
# dim, base=10000, precision=torch.half, learnable=False
|
351 |
+
self.rotary_emb = RotaryEmbedding(self.hidden_size_per_attention_head // 2 if self.rotary_embedding_2d else self.hidden_size_per_attention_head,
|
352 |
+
base=10000, precision=config.torch_dtype, learnable=False)
|
353 |
+
|
354 |
+
|
355 |
+
def forward(
|
356 |
+
self, hidden_states, attention_mask, position_ids, kv_cache=None, use_cache=True
|
357 |
+
):
|
358 |
+
# hidden_states: [sq, b, h]
|
359 |
+
|
360 |
+
# =================================================
|
361 |
+
# Pre-allocate memory for key-values for inference.
|
362 |
+
# =================================================
|
363 |
+
# =====================
|
364 |
+
# Query, Key, and Value
|
365 |
+
# =====================
|
366 |
+
|
367 |
+
# Attention heads [sq, b, h] --> [sq, b, (np * 3 * hn)]
|
368 |
+
mixed_x_layer = self.query_key_value(hidden_states)
|
369 |
+
|
370 |
+
if self.multi_query_attention:
|
371 |
+
(query_layer, key_layer, value_layer) = mixed_x_layer.split(
|
372 |
+
[
|
373 |
+
self.num_attention_heads_per_partition * self.hidden_size_per_attention_head,
|
374 |
+
self.num_multi_query_groups_per_partition * self.hidden_size_per_attention_head,
|
375 |
+
self.num_multi_query_groups_per_partition * self.hidden_size_per_attention_head,
|
376 |
+
],
|
377 |
+
dim=-1,
|
378 |
+
)
|
379 |
+
query_layer = query_layer.view(
|
380 |
+
query_layer.size()[:-1] + (self.num_attention_heads_per_partition, self.hidden_size_per_attention_head)
|
381 |
+
)
|
382 |
+
key_layer = key_layer.view(
|
383 |
+
key_layer.size()[:-1] + (self.num_multi_query_groups_per_partition, self.hidden_size_per_attention_head)
|
384 |
+
)
|
385 |
+
value_layer = value_layer.view(
|
386 |
+
value_layer.size()[:-1]
|
387 |
+
+ (self.num_multi_query_groups_per_partition, self.hidden_size_per_attention_head)
|
388 |
+
)
|
389 |
+
else:
|
390 |
+
new_tensor_shape = mixed_x_layer.size()[:-1] + (self.num_attention_heads_per_partition, 3 * self.hidden_size_per_attention_head)
|
391 |
+
mixed_x_layer = mixed_x_layer.view(*new_tensor_shape)
|
392 |
+
# [sq, b, np, 3 * hn] --> 3 [sq, b, np, hn]
|
393 |
+
(query_layer, key_layer, value_layer) = split_tensor_along_last_dim(mixed_x_layer, 3)
|
394 |
+
|
395 |
+
# apply relative positional encoding (rotary embedding)
|
396 |
+
if position_ids is not None: # [seq_len, 2, batch_size, 32, 2]
|
397 |
+
|
398 |
+
if self.rotary_embedding_2d:
|
399 |
+
q1, q2 = query_layer.chunk(2, dim=(query_layer.ndim - 1)) # 32
|
400 |
+
k1, k2 = key_layer.chunk(2, dim=(key_layer.ndim - 1))
|
401 |
+
# import pdb; pdb.set_trace();
|
402 |
+
cos, sin = self.rotary_emb(q1, seq_len=position_ids.max() + 1) # 32
|
403 |
+
position_ids, block_position_ids = \
|
404 |
+
position_ids[:, 0, :].transpose(0, 1).contiguous(), \
|
405 |
+
position_ids[:, 1, :].transpose(0, 1).contiguous()
|
406 |
+
q1, k1 = apply_rotary_pos_emb_index_torch(q1, k1, cos, sin, position_ids)
|
407 |
+
q2, k2 = apply_rotary_pos_emb_index_torch(q2, k2, cos, sin, block_position_ids)
|
408 |
+
query_layer = torch.concat([q1, q2], dim=(q1.ndim - 1))
|
409 |
+
key_layer = torch.concat([k1, k2], dim=(k1.ndim - 1))
|
410 |
+
else:
|
411 |
+
# [b, sq] -> [sq, b]
|
412 |
+
position_ids = position_ids.transpose(0, 1)
|
413 |
+
cos, sin = self.rotary_emb(value_layer, seq_len=position_ids.max() + 1)
|
414 |
+
query_layer, key_layer = apply_rotary_pos_emb_index_torch(query_layer, key_layer, cos, sin, position_ids)
|
415 |
+
|
416 |
+
# adjust key and value for inference
|
417 |
+
if kv_cache is not None:
|
418 |
+
cache_k, cache_v = kv_cache
|
419 |
+
key_layer = torch.cat((cache_k, key_layer), dim=0)
|
420 |
+
value_layer = torch.cat((cache_v, value_layer), dim=0)
|
421 |
+
if use_cache:
|
422 |
+
kv_cache = (key_layer, value_layer)
|
423 |
+
else:
|
424 |
+
kv_cache = None
|
425 |
+
|
426 |
+
if self.multi_query_attention:
|
427 |
+
key_layer = key_layer.unsqueeze(-2)
|
428 |
+
key_layer = key_layer.expand(-1, -1, -1, self.num_attention_heads_per_partition // self.num_multi_query_groups_per_partition, -1)
|
429 |
+
key_layer = key_layer.contiguous().view(key_layer.size()[:2] + (self.num_attention_heads_per_partition, self.hidden_size_per_attention_head))
|
430 |
+
value_layer = value_layer.unsqueeze(-2)
|
431 |
+
value_layer = value_layer.expand(-1, -1, -1, self.num_attention_heads_per_partition // self.num_multi_query_groups_per_partition, -1)
|
432 |
+
value_layer = value_layer.contiguous().view(value_layer.size()[:2] + (self.num_attention_heads_per_partition, self.hidden_size_per_attention_head))
|
433 |
+
|
434 |
+
# ==================================
|
435 |
+
# core attention computation
|
436 |
+
# ==================================
|
437 |
+
|
438 |
+
context_layer = self.core_attention(query_layer, key_layer, value_layer, attention_mask) # context_layer: [seq_len, batch_size, num_heads*head_dim]
|
439 |
+
output = self.dense(context_layer)
|
440 |
+
# =================
|
441 |
+
# Output. [sq, b, h]
|
442 |
+
# =================
|
443 |
+
|
444 |
+
# output = context_layer @ self.dense.weight.T + self.dense.bias
|
445 |
+
return output, kv_cache
|
446 |
+
|
447 |
+
|
448 |
+
def _config_to_kwargs(args):
|
449 |
+
common_kwargs = {
|
450 |
+
"dtype": args.torch_dtype,
|
451 |
+
}
|
452 |
+
return common_kwargs
|
453 |
+
|
454 |
+
|
455 |
+
class MLP(torch.nn.Module):
|
456 |
+
"""MLP.
|
457 |
+
|
458 |
+
MLP will take the input with h hidden state, project it to 4*h
|
459 |
+
hidden dimension, perform nonlinear transformation, and project the
|
460 |
+
state back into h hidden dimension.
|
461 |
+
"""
|
462 |
+
|
463 |
+
def __init__(self, config: ProteinGLMConfig, device=None):
|
464 |
+
super(MLP, self).__init__()
|
465 |
+
|
466 |
+
self.add_bias = config.add_bias_linear
|
467 |
+
self.moe = config.moe
|
468 |
+
self.num_experts = config.num_experts
|
469 |
+
self.experts_per_token = config.experts_per_token # 2
|
470 |
+
|
471 |
+
# Project to 4h. If using swiglu double the output width, see https://arxiv.org/pdf/2002.05202.pdf
|
472 |
+
self.dense_h_to_4h = nn.Linear(
|
473 |
+
config.hidden_size,
|
474 |
+
config.ffn_hidden_size * 2,
|
475 |
+
bias=self.add_bias,
|
476 |
+
device=device,
|
477 |
+
**_config_to_kwargs(config)
|
478 |
+
)
|
479 |
+
|
480 |
+
def swiglu(x):
|
481 |
+
x = torch.chunk(x, 2, dim=-1)
|
482 |
+
return x[0] * F.silu(x[1])
|
483 |
+
|
484 |
+
def geglu(x):
|
485 |
+
x = torch.chunk(x, 2, dim=-1)
|
486 |
+
return x[0] * F.gelu(x[1])
|
487 |
+
|
488 |
+
if config.glu_activation == 'geglu':
|
489 |
+
self.activation_func = geglu
|
490 |
+
elif config.glu_activation == 'swiglu':
|
491 |
+
self.activation_func = swiglu
|
492 |
+
else:
|
493 |
+
assert RuntimeError(f"Unsupported glu_activation: {config.glu_activation}")
|
494 |
+
|
495 |
+
# Project back to h.
|
496 |
+
self.dense_4h_to_h = nn.Linear(
|
497 |
+
config.ffn_hidden_size,
|
498 |
+
config.hidden_size,
|
499 |
+
bias=self.add_bias,
|
500 |
+
device=device,
|
501 |
+
**_config_to_kwargs(config)
|
502 |
+
)
|
503 |
+
|
504 |
+
if self.moe:
|
505 |
+
assert self.num_experts > 1
|
506 |
+
del self.dense_h_to_4h
|
507 |
+
del self.dense_4h_to_h
|
508 |
+
self.router = nn.Linear(
|
509 |
+
config.hidden_size,
|
510 |
+
config.num_experts,
|
511 |
+
bias=False,
|
512 |
+
device=device,
|
513 |
+
dtype=torch.float32
|
514 |
+
)
|
515 |
+
for i in range(0, self.num_experts):
|
516 |
+
self.register_module(f"dense_h_to_4h_{i}", nn.Linear(
|
517 |
+
config.hidden_size,
|
518 |
+
config.ffn_hidden_size * 2,
|
519 |
+
bias=self.add_bias,
|
520 |
+
device=device,
|
521 |
+
**_config_to_kwargs(config)
|
522 |
+
))
|
523 |
+
self.register_module(f"dense_4h_to_h_{i}", nn.Linear(
|
524 |
+
config.ffn_hidden_size,
|
525 |
+
config.hidden_size,
|
526 |
+
bias=self.add_bias,
|
527 |
+
device=device,
|
528 |
+
**_config_to_kwargs(config)
|
529 |
+
))
|
530 |
+
|
531 |
+
def moe_forward(self, hidden_states, expert_idx):
|
532 |
+
intermediate_parallel = getattr(self, f"dense_h_to_4h_{expert_idx}")(hidden_states)
|
533 |
+
intermediate_parallel = self.activation_func(intermediate_parallel)
|
534 |
+
output = getattr(self, f"dense_4h_to_h_{expert_idx}")(intermediate_parallel)
|
535 |
+
return output
|
536 |
+
|
537 |
+
def forward(self, hidden_states):
|
538 |
+
if self.moe:
|
539 |
+
# import pdb; pdb.set_trace();
|
540 |
+
s, b, n = hidden_states.shape
|
541 |
+
dtype = hidden_states.dtype
|
542 |
+
hidden_states = hidden_states.view(-1, hidden_states.size(2)) # [s*b h]
|
543 |
+
route = self.router(hidden_states).to(dtype)
|
544 |
+
|
545 |
+
weights, selected_experts = torch.topk(route, self.experts_per_token)
|
546 |
+
weights = F.softmax(weights, dim=1, dtype=torch.float).to(hidden_states.dtype)
|
547 |
+
output = torch.zeros_like(hidden_states, dtype=hidden_states.dtype, device=hidden_states.device)
|
548 |
+
for expert_idx in range(self.num_experts):
|
549 |
+
batch_idx, nth_expert = torch.where(selected_experts == expert_idx)
|
550 |
+
if nth_expert.shape[0] == 0:
|
551 |
+
continue
|
552 |
+
cur_out = self.moe_forward(hidden_states[batch_idx], expert_idx)
|
553 |
+
output[batch_idx] += weights[batch_idx, nth_expert, None] * cur_out
|
554 |
+
output = output.reshape(s, b, n)
|
555 |
+
else:
|
556 |
+
# [s, b, 4hp]
|
557 |
+
intermediate_parallel = self.dense_h_to_4h(hidden_states)
|
558 |
+
intermediate_parallel = self.activation_func(intermediate_parallel)
|
559 |
+
# [s, b, h]
|
560 |
+
output = self.dense_4h_to_h(intermediate_parallel)
|
561 |
+
return output
|
562 |
+
|
563 |
+
class ProteinGLMBlock(torch.nn.Module):
|
564 |
+
"""A single transformer layer.
|
565 |
+
|
566 |
+
Transformer layer takes input with size [s, b, h] and returns an
|
567 |
+
output of the same size.
|
568 |
+
"""
|
569 |
+
|
570 |
+
def __init__(self, config: ProteinGLMConfig, layer_number, device=None):
|
571 |
+
super(ProteinGLMBlock, self).__init__()
|
572 |
+
self.layer_number = layer_number
|
573 |
+
|
574 |
+
self.apply_residual_connection_post_layernorm = config.apply_residual_connection_post_layernorm
|
575 |
+
|
576 |
+
self.fp32_residual_connection = config.fp32_residual_connection
|
577 |
+
|
578 |
+
LayerNormFunc = RMSNorm if config.rmsnorm else LayerNorm
|
579 |
+
# Layernorm on the input data.
|
580 |
+
self.input_layernorm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon)
|
581 |
+
|
582 |
+
# Self attention.
|
583 |
+
self.self_attention = SelfAttention(config, layer_number, device=device)
|
584 |
+
self.hidden_dropout = config.hidden_dropout
|
585 |
+
|
586 |
+
# Layernorm on the attention output
|
587 |
+
self.post_attention_layernorm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon)
|
588 |
+
|
589 |
+
# MLP
|
590 |
+
self.mlp = MLP(config, device=device)
|
591 |
+
|
592 |
+
self.deepnorm_coeff = get_deepnorm_coefficients(config) if config.deepnorm else None
|
593 |
+
|
594 |
+
def forward(
|
595 |
+
self, hidden_states, attention_mask, position_ids, kv_cache=None, use_cache=True,
|
596 |
+
):
|
597 |
+
# hidden_states: [s, b, h]
|
598 |
+
# Layer norm at the beginning of the transformer layer.
|
599 |
+
layernorm_output = self.input_layernorm(hidden_states)
|
600 |
+
# Self attention.
|
601 |
+
attention_output, kv_cache = self.self_attention(
|
602 |
+
layernorm_output,
|
603 |
+
attention_mask,
|
604 |
+
position_ids, # [batch_size, 2, seq_len, 32, 2]
|
605 |
+
kv_cache=kv_cache,
|
606 |
+
use_cache=use_cache
|
607 |
+
)
|
608 |
+
|
609 |
+
# Residual connection.
|
610 |
+
if self.apply_residual_connection_post_layernorm:
|
611 |
+
residual = layernorm_output
|
612 |
+
else:
|
613 |
+
residual = hidden_states
|
614 |
+
|
615 |
+
layernorm_input = torch.nn.functional.dropout(attention_output, p=self.hidden_dropout, training=self.training)
|
616 |
+
if self.deepnorm_coeff is not None:
|
617 |
+
layernorm_input = residual*self.deepnorm_coeff.alpha + layernorm_input
|
618 |
+
else:
|
619 |
+
layernorm_input = residual + layernorm_input
|
620 |
+
|
621 |
+
# Layer norm post the self attention.
|
622 |
+
layernorm_output = self.post_attention_layernorm(layernorm_input)
|
623 |
+
|
624 |
+
# MLP.
|
625 |
+
mlp_output = self.mlp(layernorm_output)
|
626 |
+
|
627 |
+
# Second residual connection.
|
628 |
+
if self.apply_residual_connection_post_layernorm:
|
629 |
+
residual = layernorm_output
|
630 |
+
else:
|
631 |
+
residual = layernorm_input
|
632 |
+
|
633 |
+
output = torch.nn.functional.dropout(mlp_output, p=self.hidden_dropout, training=self.training)
|
634 |
+
if self.deepnorm_coeff is not None:
|
635 |
+
output = residual*self.deepnorm_coeff.alpha + output
|
636 |
+
else:
|
637 |
+
#print(f"2 self.deepnorm_coeff is None")
|
638 |
+
output = residual + output
|
639 |
+
|
640 |
+
return output, kv_cache
|
641 |
+
|
642 |
+
|
643 |
+
class ProteinGLMTransformer(torch.nn.Module):
|
644 |
+
"""Transformer class."""
|
645 |
+
|
646 |
+
def __init__(self, config: ProteinGLMConfig, device=None):
|
647 |
+
super(ProteinGLMTransformer, self).__init__()
|
648 |
+
|
649 |
+
self.fp32_residual_connection = config.fp32_residual_connection
|
650 |
+
self.post_layer_norm = config.post_layer_norm
|
651 |
+
|
652 |
+
# Number of layers.
|
653 |
+
self.num_layers = config.num_layers
|
654 |
+
|
655 |
+
# Transformer layers.
|
656 |
+
def build_layer(layer_number):
|
657 |
+
return ProteinGLMBlock(config, layer_number, device=device)
|
658 |
+
|
659 |
+
self.layers = torch.nn.ModuleList([build_layer(i + 1) for i in range(self.num_layers)])
|
660 |
+
|
661 |
+
if self.post_layer_norm:
|
662 |
+
LayerNormFunc = RMSNorm if config.rmsnorm else LayerNorm
|
663 |
+
# Final layer norm before output.
|
664 |
+
self.final_layernorm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon)
|
665 |
+
|
666 |
+
self.gradient_checkpointing = False
|
667 |
+
|
668 |
+
def _get_layer(self, layer_number):
|
669 |
+
return self.layers[layer_number]
|
670 |
+
|
671 |
+
def forward(
|
672 |
+
self, hidden_states, attention_mask, position_ids, kv_caches=None,
|
673 |
+
use_cache: Optional[bool] = True,
|
674 |
+
output_hidden_states: Optional[bool] = False,
|
675 |
+
):
|
676 |
+
if not kv_caches:
|
677 |
+
kv_caches = [None for _ in range(self.num_layers)]
|
678 |
+
presents = () if use_cache else None
|
679 |
+
if self.gradient_checkpointing and self.training:
|
680 |
+
if use_cache:
|
681 |
+
logger.warning_once(
|
682 |
+
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
|
683 |
+
)
|
684 |
+
use_cache = False
|
685 |
+
|
686 |
+
all_self_attentions = None
|
687 |
+
all_hidden_states = () if output_hidden_states else None
|
688 |
+
for index in range(self.num_layers):
|
689 |
+
if output_hidden_states:
|
690 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
691 |
+
|
692 |
+
layer = self._get_layer(index)
|
693 |
+
if self.gradient_checkpointing and self.training and torch.is_grad_enabled():
|
694 |
+
layer_ret = get_checkpoint_fn()(
|
695 |
+
layer,
|
696 |
+
hidden_states,
|
697 |
+
attention_mask,
|
698 |
+
position_ids,
|
699 |
+
kv_caches[index],
|
700 |
+
use_cache
|
701 |
+
)
|
702 |
+
else:
|
703 |
+
layer_ret = layer(
|
704 |
+
hidden_states,
|
705 |
+
attention_mask,
|
706 |
+
position_ids,
|
707 |
+
kv_cache=kv_caches[index],
|
708 |
+
use_cache=use_cache
|
709 |
+
)
|
710 |
+
hidden_states, kv_cache = layer_ret
|
711 |
+
if use_cache:
|
712 |
+
presents = presents + (kv_cache,)
|
713 |
+
|
714 |
+
|
715 |
+
# Final layer norm.
|
716 |
+
if self.post_layer_norm:
|
717 |
+
hidden_states = self.final_layernorm(hidden_states)
|
718 |
+
|
719 |
+
if output_hidden_states:
|
720 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
721 |
+
|
722 |
+
return hidden_states, presents, all_hidden_states, all_self_attentions
|
723 |
+
|
724 |
+
|
725 |
+
class ProteinGLMPreTrainedModel(PreTrainedModel):
|
726 |
+
"""
|
727 |
+
An abstract class to handle weights initialization and
|
728 |
+
a simple interface for downloading and loading pretrained models.
|
729 |
+
"""
|
730 |
+
|
731 |
+
is_parallelizable = False
|
732 |
+
supports_gradient_checkpointing = True
|
733 |
+
config_class = ProteinGLMConfig
|
734 |
+
base_model_prefix = "transformer"
|
735 |
+
_no_split_modules = ["ProteinGLMBlock"]
|
736 |
+
|
737 |
+
_quantized = False
|
738 |
+
|
739 |
+
|
740 |
+
def get_masks(self, input_ids, past_key_values, padding_mask=None, is_causal=True):
|
741 |
+
batch_size, seq_length = input_ids.shape
|
742 |
+
full_attention_mask = torch.ones(batch_size, seq_length, seq_length, device=input_ids.device)
|
743 |
+
if is_causal:
|
744 |
+
full_attention_mask.tril_()
|
745 |
+
past_length = 0
|
746 |
+
if past_key_values:
|
747 |
+
past_length = past_key_values[0][0].shape[0]
|
748 |
+
if past_length:
|
749 |
+
full_attention_mask = torch.cat((torch.ones(batch_size, seq_length, past_length,
|
750 |
+
device=input_ids.device), full_attention_mask), dim=-1)
|
751 |
+
if padding_mask is not None:
|
752 |
+
full_attention_mask = full_attention_mask * padding_mask.unsqueeze(1)
|
753 |
+
if not past_length and padding_mask is not None:
|
754 |
+
full_attention_mask -= padding_mask.unsqueeze(-1) - 1
|
755 |
+
full_attention_mask = (full_attention_mask < 0.5).bool()
|
756 |
+
full_attention_mask.unsqueeze_(1)
|
757 |
+
return full_attention_mask
|
758 |
+
|
759 |
+
def get_position_ids(self, input_ids, device, context_length=0):
|
760 |
+
batch_size, seq_length = input_ids.shape
|
761 |
+
if self.config.rotary_embedding_2d:
|
762 |
+
if self.config.is_causal: # 100b model
|
763 |
+
position_ids_1 = torch.zeros(seq_length, dtype=torch.long, device=device).unsqueeze(0).repeat(batch_size, 1) # [batch_size, seq_len]
|
764 |
+
position_ids_2 = torch.arange(seq_length, dtype=torch.long, device=device).unsqueeze(0).repeat(batch_size, 1) # [batch_size, seq_len]
|
765 |
+
position_ids = torch.stack([position_ids_1, position_ids_2], axis=1) # [batch_size, 2, seq_len]
|
766 |
+
else:
|
767 |
+
position_ids_1 = torch.arange(seq_length, dtype=torch.long, device=device).unsqueeze(0).repeat(batch_size, 1) # [batch_size, seq_len]
|
768 |
+
position_ids_2 = torch.zeros(seq_length, dtype=torch.long, device=device).unsqueeze(0).repeat(batch_size, 1) # [batch_size, seq_len]
|
769 |
+
position_ids = torch.stack([position_ids_1, position_ids_2], axis=1) # [batch_size, 2, seq_len]
|
770 |
+
else:
|
771 |
+
position_ids = torch.arange(seq_length, dtype=torch.long, device=device).unsqueeze(0).repeat(batch_size, 1) # [batch_size, 1, seq_len]
|
772 |
+
return position_ids
|
773 |
+
|
774 |
+
def _set_gradient_checkpointing(self, module, value=False):
|
775 |
+
if isinstance(module, ProteinGLMTransformer):
|
776 |
+
module.gradient_checkpointing = value
|
777 |
+
|
778 |
+
|
779 |
+
# Copied from transformers.models.bert.modeling_bert.BertPreTrainedModel._init_weights
|
780 |
+
def _init_weights(self, module):
|
781 |
+
std = self.config.initializer_range
|
782 |
+
"""Initialize the weights"""
|
783 |
+
if isinstance(module, nn.Linear):
|
784 |
+
# Slightly different from the TF version which uses truncated_normal for initialization
|
785 |
+
# cf https://github.com/pytorch/pytorch/pull/5617
|
786 |
+
module.weight.data.normal_(mean=0.0, std=std)
|
787 |
+
if module.bias is not None:
|
788 |
+
module.bias.data.zero_()
|
789 |
+
elif isinstance(module, nn.Embedding):
|
790 |
+
module.weight.data.normal_(mean=0.0, std=std)
|
791 |
+
if module.padding_idx is not None:
|
792 |
+
module.weight.data[module.padding_idx].zero_()
|
793 |
+
elif isinstance(module, nn.LayerNorm):
|
794 |
+
module.bias.data.zero_()
|
795 |
+
module.weight.data.fill_(1.0)
|
796 |
+
|
797 |
+
def quantize(self, weight_bit_width: int, empty_init=True, device=None):
|
798 |
+
if self._quantized:
|
799 |
+
print(f"Model has been quantized...")
|
800 |
+
return
|
801 |
+
self.transformer.encoder = quantize(self.transformer.encoder, weight_bit_width, empty_init, device)
|
802 |
+
self._quantized = True
|
803 |
+
return self
|
804 |
+
|
805 |
+
class Embedding(torch.nn.Module):
|
806 |
+
"""Language model embeddings."""
|
807 |
+
|
808 |
+
def __init__(self, config: ProteinGLMConfig, device=None):
|
809 |
+
super(Embedding, self).__init__()
|
810 |
+
|
811 |
+
self.hidden_size = config.hidden_size
|
812 |
+
# Word embeddings (parallel).
|
813 |
+
self.word_embeddings = nn.Embedding(
|
814 |
+
config.padded_vocab_size,
|
815 |
+
self.hidden_size,
|
816 |
+
dtype=config.torch_dtype,
|
817 |
+
device=device
|
818 |
+
)
|
819 |
+
self.fp32_residual_connection = config.fp32_residual_connection
|
820 |
+
|
821 |
+
|
822 |
+
def forward(self, input_ids):
|
823 |
+
# Embeddings.
|
824 |
+
words_embeddings = self.word_embeddings(input_ids)
|
825 |
+
embeddings = words_embeddings
|
826 |
+
# Data format change to avoid explicit tranposes : [b s h] --> [s b h].
|
827 |
+
embeddings = embeddings.transpose(0, 1).contiguous()
|
828 |
+
# If the input flag for fp32 residual connection is set, convert for float.
|
829 |
+
if self.fp32_residual_connection:
|
830 |
+
embeddings = embeddings.float()
|
831 |
+
return embeddings
|
832 |
+
|
833 |
+
class ProteinGLMModel(ProteinGLMPreTrainedModel):
|
834 |
+
def __init__(self, config: ProteinGLMConfig, device=None, empty_init=True):
|
835 |
+
super().__init__(config)
|
836 |
+
if empty_init:
|
837 |
+
init_method = skip_init
|
838 |
+
else:
|
839 |
+
init_method = default_init
|
840 |
+
init_kwargs = {}
|
841 |
+
if device is not None:
|
842 |
+
init_kwargs["device"] = device
|
843 |
+
self.embedding = init_method(Embedding, config, **init_kwargs)
|
844 |
+
self.num_layers = config.num_layers
|
845 |
+
self.multi_query_group_num = config.multi_query_group_num
|
846 |
+
self.kv_channels = config.kv_channels
|
847 |
+
|
848 |
+
# Rotary positional embeddings
|
849 |
+
self.seq_length = config.seq_length
|
850 |
+
rotary_dim = (
|
851 |
+
config.hidden_size // config.num_attention_heads if config.kv_channels is None else config.kv_channels
|
852 |
+
)
|
853 |
+
|
854 |
+
# self.rotary_pos_emb = RotaryEmbedding(rotary_dim // 2, base=10000, precision=config.torch_dtype, learnable=False)
|
855 |
+
self.encoder = init_method(ProteinGLMTransformer, config, **init_kwargs)
|
856 |
+
|
857 |
+
self.output_layer = init_method(nn.Linear, config.hidden_size, config.padded_vocab_size, bias=False,
|
858 |
+
dtype=config.torch_dtype, **init_kwargs)
|
859 |
+
|
860 |
+
def get_input_embeddings(self):
|
861 |
+
return self.embedding.word_embeddings
|
862 |
+
|
863 |
+
def set_input_embeddings(self, value):
|
864 |
+
self.embedding.word_embeddings = value
|
865 |
+
|
866 |
+
def forward(
|
867 |
+
self,
|
868 |
+
input_ids,
|
869 |
+
position_ids: Optional[torch.Tensor] = None, # position_ids: [batch_size, 2, seq_len]
|
870 |
+
attention_mask: Optional[torch.BoolTensor] = None,
|
871 |
+
full_attention_mask: Optional[torch.BoolTensor] = None,
|
872 |
+
past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
|
873 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
874 |
+
use_cache: Optional[bool] = None,
|
875 |
+
output_hidden_states: Optional[bool] = None,
|
876 |
+
return_dict: Optional[bool] = None,
|
877 |
+
):
|
878 |
+
output_hidden_states = (
|
879 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
880 |
+
)
|
881 |
+
if self.config.is_causal:
|
882 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
883 |
+
else:
|
884 |
+
use_cache = False
|
885 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
886 |
+
|
887 |
+
batch_size, seq_length = input_ids.shape
|
888 |
+
|
889 |
+
if inputs_embeds is None:
|
890 |
+
inputs_embeds = self.embedding(input_ids)
|
891 |
+
|
892 |
+
if full_attention_mask is None:
|
893 |
+
if (attention_mask is not None and not attention_mask.all()) or (past_key_values and seq_length != 1):
|
894 |
+
full_attention_mask = self.get_masks(input_ids, past_key_values, padding_mask=attention_mask)
|
895 |
+
# Run encoder.
|
896 |
+
hidden_states, presents, all_hidden_states, all_self_attentions = self.encoder(
|
897 |
+
inputs_embeds, full_attention_mask, position_ids=position_ids,
|
898 |
+
kv_caches=past_key_values, use_cache=use_cache, output_hidden_states=output_hidden_states
|
899 |
+
)
|
900 |
+
|
901 |
+
if not return_dict:
|
902 |
+
return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None)
|
903 |
+
|
904 |
+
return BaseModelOutputWithPast(
|
905 |
+
last_hidden_state=hidden_states,
|
906 |
+
past_key_values=presents,
|
907 |
+
hidden_states=all_hidden_states,
|
908 |
+
attentions=all_self_attentions,
|
909 |
+
)
|
910 |
+
|
911 |
+
|
912 |
+
class ProteinGLMForMaskedLM(ProteinGLMPreTrainedModel):
|
913 |
+
def __init__(self, config: ProteinGLMConfig, empty_init=True, device=None):
|
914 |
+
super().__init__(config)
|
915 |
+
|
916 |
+
self.max_sequence_length = config.max_length
|
917 |
+
self.transformer = ProteinGLMModel(config, empty_init=empty_init, device=device)
|
918 |
+
self.config = config
|
919 |
+
if self.config.quantization_bit:
|
920 |
+
print(f"Begin Quantization to {self.config.quantization_bit} bit")
|
921 |
+
self.quantize(self.config.quantization_bit, empty_init=True, device=device)
|
922 |
+
|
923 |
+
def forward(
|
924 |
+
self,
|
925 |
+
input_ids: Optional[torch.Tensor] = None,
|
926 |
+
position_ids: Optional[torch.Tensor] = None,
|
927 |
+
attention_mask: Optional[torch.Tensor] = None,
|
928 |
+
past_key_values: Optional[Tuple[torch.FloatTensor]] = None,
|
929 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
930 |
+
labels: Optional[torch.Tensor] = None,
|
931 |
+
use_cache: Optional[bool] = None,
|
932 |
+
output_attentions: Optional[bool] = None,
|
933 |
+
output_hidden_states: Optional[bool] = None,
|
934 |
+
return_dict: Optional[bool] = None,
|
935 |
+
return_last_logit: Optional[bool] = None,
|
936 |
+
return_last_hidden_state: Optional[bool] = None
|
937 |
+
):
|
938 |
+
if self.config.is_causal:
|
939 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
940 |
+
else:
|
941 |
+
use_cache = False
|
942 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
943 |
+
|
944 |
+
if position_ids is None:
|
945 |
+
position_ids = self.get_position_ids(input_ids, device=input_ids.device)
|
946 |
+
|
947 |
+
full_attention_mask = self.get_masks(input_ids, past_key_values, padding_mask=attention_mask, is_causal=self.config.is_causal)
|
948 |
+
|
949 |
+
transformer_outputs = self.transformer(
|
950 |
+
input_ids=input_ids,
|
951 |
+
position_ids=position_ids, # position_ids: [batch_size, 2, seq_len]
|
952 |
+
full_attention_mask=full_attention_mask,
|
953 |
+
past_key_values=past_key_values,
|
954 |
+
inputs_embeds=inputs_embeds,
|
955 |
+
use_cache=use_cache,
|
956 |
+
output_hidden_states=output_hidden_states,
|
957 |
+
return_dict=return_dict,
|
958 |
+
)
|
959 |
+
|
960 |
+
hidden_states = transformer_outputs[0]
|
961 |
+
if return_last_logit:
|
962 |
+
hidden_states = hidden_states[-1:]
|
963 |
+
lm_logits = self.transformer.output_layer(hidden_states)
|
964 |
+
lm_logits = lm_logits.transpose(0, 1).contiguous()
|
965 |
+
|
966 |
+
masked_lm_loss = None
|
967 |
+
if labels is not None:
|
968 |
+
lm_logits = lm_logits.to(torch.float32)
|
969 |
+
|
970 |
+
# Flatten the tokens
|
971 |
+
loss_fct = CrossEntropyLoss(ignore_index=-100) # -100 for padding token.
|
972 |
+
masked_lm_loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.view(-1))
|
973 |
+
|
974 |
+
lm_logits = lm_logits.to(hidden_states.dtype)
|
975 |
+
loss = loss.to(hidden_states.dtype)
|
976 |
+
|
977 |
+
if not return_dict:
|
978 |
+
output = (lm_logits,) + transformer_outputs[1:]
|
979 |
+
return ((loss,) + output) if loss is not None else output
|
980 |
+
return MaskedLMOutput(
|
981 |
+
loss = masked_lm_loss,
|
982 |
+
logits=lm_logits,
|
983 |
+
hidden_states=transformer_outputs.last_hidden_state if return_last_hidden_state else transformer_outputs.hidden_states,
|
984 |
+
attentions=transformer_outputs.attentions,
|
985 |
+
)
|
986 |
+
|
987 |
+
|
988 |
+
|
989 |
+
|
990 |
+
class ProteinGLMForSequenceClassification(ProteinGLMPreTrainedModel):
|
991 |
+
def __init__(self, config: ProteinGLMConfig, empty_init=True, device=None):
|
992 |
+
super().__init__(config)
|
993 |
+
self.config = config
|
994 |
+
self.num_labels = config.num_labels
|
995 |
+
|
996 |
+
self.transformer = ProteinGLMModel(config, empty_init=empty_init, device=device)
|
997 |
+
self.classifier = ProteinGLMClassificationHead(config)
|
998 |
+
if self.config.quantization_bit:
|
999 |
+
print(f"Begin Quantization to {self.config.quantization_bit} bit")
|
1000 |
+
self.quantize(self.config.quantization_bit, empty_init=True, device=device)
|
1001 |
+
|
1002 |
+
def forward(
|
1003 |
+
self,
|
1004 |
+
input_ids: Optional[torch.Tensor] = None,
|
1005 |
+
position_ids: Optional[torch.Tensor] = None,
|
1006 |
+
attention_mask: Optional[torch.Tensor] = None,
|
1007 |
+
past_key_values: Optional[Tuple[torch.FloatTensor]] = None,
|
1008 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
1009 |
+
labels: Optional[torch.Tensor] = None,
|
1010 |
+
use_cache: Optional[bool] = None,
|
1011 |
+
output_attentions: Optional[bool] = None,
|
1012 |
+
output_hidden_states: Optional[bool] = None,
|
1013 |
+
return_dict: Optional[bool] = None,
|
1014 |
+
return_last_logit: Optional[bool] = None,
|
1015 |
+
return_last_hidden_state: Optional[bool] = None,
|
1016 |
+
**kwargs
|
1017 |
+
) -> Union[Tuple, SequenceClassifierOutput]:
|
1018 |
+
r"""
|
1019 |
+
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
1020 |
+
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
|
1021 |
+
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
|
1022 |
+
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
1023 |
+
"""
|
1024 |
+
if self.config.is_causal:
|
1025 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
1026 |
+
else:
|
1027 |
+
use_cache = False
|
1028 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1029 |
+
|
1030 |
+
if position_ids is None:
|
1031 |
+
position_ids = self.get_position_ids(input_ids, device=input_ids.device)
|
1032 |
+
|
1033 |
+
full_attention_mask = self.get_masks(input_ids, past_key_values, padding_mask=attention_mask, is_causal=self.config.is_causal)
|
1034 |
+
|
1035 |
+
transformer_outputs = self.transformer(
|
1036 |
+
input_ids=input_ids,
|
1037 |
+
position_ids=position_ids, # position_ids: [batch_size, 2, seq_len]
|
1038 |
+
full_attention_mask=full_attention_mask,
|
1039 |
+
past_key_values=past_key_values,
|
1040 |
+
inputs_embeds=inputs_embeds,
|
1041 |
+
use_cache=use_cache,
|
1042 |
+
output_hidden_states=output_hidden_states,
|
1043 |
+
return_dict=return_dict,
|
1044 |
+
)
|
1045 |
+
if self.config.add_special_tokens:
|
1046 |
+
hidden_states = transformer_outputs[0][:-1] # get rid of <eos> token
|
1047 |
+
else:
|
1048 |
+
hidden_states = transformer_outputs[0]
|
1049 |
+
logits = self.classifier(hidden_states, add_pooling=True)
|
1050 |
+
loss = None
|
1051 |
+
if labels is not None:
|
1052 |
+
labels = labels.to(logits.device)
|
1053 |
+
|
1054 |
+
if self.config.problem_type is None:
|
1055 |
+
if self.num_labels == 1:
|
1056 |
+
self.config.problem_type = "regression"
|
1057 |
+
elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
|
1058 |
+
self.config.problem_type = "single_label_classification"
|
1059 |
+
else:
|
1060 |
+
self.config.problem_type = "multi_label_classification"
|
1061 |
+
|
1062 |
+
if self.config.problem_type == "regression":
|
1063 |
+
loss_fct = MSELoss()
|
1064 |
+
if self.num_labels == 1:
|
1065 |
+
loss = loss_fct(logits.squeeze(), labels.squeeze())
|
1066 |
+
else:
|
1067 |
+
loss = loss_fct(logits, labels)
|
1068 |
+
elif self.config.problem_type == "single_label_classification":
|
1069 |
+
loss_fct = CrossEntropyLoss()
|
1070 |
+
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
|
1071 |
+
elif self.config.problem_type == "multi_label_classification":
|
1072 |
+
loss_fct = BCEWithLogitsLoss()
|
1073 |
+
loss = loss_fct(logits, labels)
|
1074 |
+
|
1075 |
+
if not return_dict:
|
1076 |
+
output = (logits,) + transformer_outputs[2:]
|
1077 |
+
return ((loss,) + output) if loss is not None else output
|
1078 |
+
|
1079 |
+
return SequenceClassifierOutput(
|
1080 |
+
loss=loss,
|
1081 |
+
logits=logits,
|
1082 |
+
hidden_states=transformer_outputs.hidden_states,
|
1083 |
+
attentions=transformer_outputs.attentions,
|
1084 |
+
)
|
1085 |
+
|
1086 |
+
class ProteinGLMForTokenClassification(ProteinGLMPreTrainedModel):
|
1087 |
+
def __init__(self, config: ProteinGLMConfig, empty_init=True, device=None):
|
1088 |
+
super().__init__(config)
|
1089 |
+
self.config = config
|
1090 |
+
self.num_labels = config.num_labels
|
1091 |
+
|
1092 |
+
self.transformer = ProteinGLMModel(config, empty_init=empty_init, device=device)
|
1093 |
+
if config.task_modality == "token":
|
1094 |
+
self.classifier = ProteinGLMClassificationHead(config)
|
1095 |
+
elif config.task_modality == 'pair':
|
1096 |
+
self.classifier = ProteinGLMContactHead(config)
|
1097 |
+
|
1098 |
+
self.quantized = False
|
1099 |
+
|
1100 |
+
if self.config.quantization_bit:
|
1101 |
+
print(f"Begin Quantization to {self.config.quantization_bit} bit")
|
1102 |
+
self.quantize(self.config.quantization_bit, empty_init=True, device=device)
|
1103 |
+
|
1104 |
+
|
1105 |
+
def forward(
|
1106 |
+
self,
|
1107 |
+
input_ids: Optional[torch.Tensor] = None,
|
1108 |
+
position_ids: Optional[torch.Tensor] = None,
|
1109 |
+
attention_mask: Optional[torch.Tensor] = None,
|
1110 |
+
past_key_values: Optional[Tuple[torch.FloatTensor]] = None,
|
1111 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
1112 |
+
labels: Optional[torch.Tensor] = None,
|
1113 |
+
use_cache: Optional[bool] = None,
|
1114 |
+
output_attentions: Optional[bool] = None,
|
1115 |
+
output_hidden_states: Optional[bool] = None,
|
1116 |
+
return_dict: Optional[bool] = None,
|
1117 |
+
return_last_logit: Optional[bool] = None,
|
1118 |
+
return_last_hidden_state: Optional[bool] = None,
|
1119 |
+
**kwargs
|
1120 |
+
) -> Union[Tuple, SequenceClassifierOutput]:
|
1121 |
+
r"""
|
1122 |
+
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
1123 |
+
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
|
1124 |
+
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
|
1125 |
+
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
1126 |
+
"""
|
1127 |
+
if self.config.is_causal:
|
1128 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
1129 |
+
else:
|
1130 |
+
use_cache = False
|
1131 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1132 |
+
|
1133 |
+
if position_ids is None:
|
1134 |
+
position_ids = self.get_position_ids(input_ids, device=input_ids.device)
|
1135 |
+
|
1136 |
+
full_attention_mask = self.get_masks(input_ids, past_key_values, padding_mask=attention_mask, is_causal = self.config.is_causal)
|
1137 |
+
|
1138 |
+
transformer_outputs = self.transformer(
|
1139 |
+
input_ids=input_ids,
|
1140 |
+
position_ids=position_ids, # position_ids: [batch_size, 2, seq_len]
|
1141 |
+
full_attention_mask=full_attention_mask,
|
1142 |
+
past_key_values=past_key_values,
|
1143 |
+
inputs_embeds=inputs_embeds,
|
1144 |
+
use_cache=use_cache,
|
1145 |
+
output_hidden_states=output_hidden_states,
|
1146 |
+
return_dict=return_dict,
|
1147 |
+
)
|
1148 |
+
if self.config.add_special_tokens:
|
1149 |
+
hidden_states = transformer_outputs[0][:-1] # get rid of <eos> token
|
1150 |
+
else:
|
1151 |
+
hidden_states = transformer_outputs[0]
|
1152 |
+
|
1153 |
+
logits = self.classifier(hidden_states, add_pooling=False)
|
1154 |
+
loss = None
|
1155 |
+
if labels is not None:
|
1156 |
+
labels = labels.to(logits.device)
|
1157 |
+
loss_fct = CrossEntropyLoss()
|
1158 |
+
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
|
1159 |
+
|
1160 |
+
if not return_dict:
|
1161 |
+
output = (logits,) + transformer_outputs[2:]
|
1162 |
+
return ((loss,) + output) if loss is not None else output
|
1163 |
+
|
1164 |
+
|
1165 |
+
return TokenClassifierOutput(
|
1166 |
+
loss=loss,
|
1167 |
+
logits=logits,
|
1168 |
+
hidden_states=transformer_outputs.hidden_states,
|
1169 |
+
attentions=transformer_outputs.attentions,
|
1170 |
+
)
|
1171 |
+
|
1172 |
+
|
1173 |
+
|
1174 |
+
class ProteinGLMClassificationHead(nn.Module):
|
1175 |
+
"""Head for classification tasks."""
|
1176 |
+
def __init__(self, config):
|
1177 |
+
super().__init__()
|
1178 |
+
self.activation_func = config.activation_func
|
1179 |
+
self.layers = torch.nn.ModuleList()
|
1180 |
+
last_size = config.hidden_size
|
1181 |
+
for sz in config.inter_hidden_size:
|
1182 |
+
this_layer = torch.nn.Linear(last_size, sz, bias=config.bias)
|
1183 |
+
last_size = sz
|
1184 |
+
self.layers.append(this_layer)
|
1185 |
+
|
1186 |
+
def forward(self,
|
1187 |
+
input_features,
|
1188 |
+
add_pooling: Optional[bool] = True
|
1189 |
+
):
|
1190 |
+
# [s, b, h] -> [b, s ,h]
|
1191 |
+
input_features = input_features.transpose(0,1).contiguous()
|
1192 |
+
if add_pooling:
|
1193 |
+
# [b, h]
|
1194 |
+
input_features = torch.mean(input_features, dim = 1)
|
1195 |
+
for i, layer in enumerate(self.layers):
|
1196 |
+
if i > 0:
|
1197 |
+
input_features = self.activation_func(input_features)
|
1198 |
+
input_features = layer(input_features)
|
1199 |
+
return input_features
|
1200 |
+
|
1201 |
+
class ProteinGLMContactHead(nn.Module):
|
1202 |
+
"""Head for sentence-level classification tasks."""
|
1203 |
+
def __init__(self, config):
|
1204 |
+
super().__init__()
|
1205 |
+
self.activation_func = config.activation_func
|
1206 |
+
self.layers = torch.nn.ModuleList()
|
1207 |
+
last_size = config.hidden_size * 2
|
1208 |
+
for sz in config.inter_hidden_size:
|
1209 |
+
this_layer = torch.nn.Linear(last_size, sz, bias=config.bias)
|
1210 |
+
last_size = sz
|
1211 |
+
self.layers.append(this_layer)
|
1212 |
+
|
1213 |
+
def outer_concat(self, x):
|
1214 |
+
batch_size, seq_len, features = x.shape
|
1215 |
+
|
1216 |
+
# Permute to [batch_size, features, seq_len]
|
1217 |
+
x = x.permute(0, 2, 1)
|
1218 |
+
|
1219 |
+
# Introduce new dimensions for broadcasting
|
1220 |
+
x_1 = x[:, None, :, :, None] # [batch_size, 1, features, seq_len, 1]
|
1221 |
+
x_2 = x[:, None, :, None, :] # [batch_size, 1, features, 1, seq_len]
|
1222 |
+
|
1223 |
+
# Repeat along new dimensions
|
1224 |
+
x_1 = x_1.repeat(1, 1, 1, 1, seq_len) # [batch_size, 1, features, seq_len, seq_len]
|
1225 |
+
x_2 = x_2.repeat(1, 1, 1, seq_len, 1) # [batch_size, 1, features, seq_len, seq_len]
|
1226 |
+
|
1227 |
+
# Concatenate along the second dimension
|
1228 |
+
x = torch.cat((x_1, x_2), dim=1) # [batch_size, 2, features, seq_len, seq_len]
|
1229 |
+
|
1230 |
+
# Get lower triangular indices
|
1231 |
+
I, J = torch.tril_indices(seq_len, seq_len, -1)
|
1232 |
+
|
1233 |
+
# Symmetrize
|
1234 |
+
x[:, :, :, I, J] = x[:, :, :, J, I]
|
1235 |
+
|
1236 |
+
# Permute to desired shape and make contiguous
|
1237 |
+
x = x.permute(0, 3, 4, 2, 1).contiguous() # [batch_size, seq_len, seq_len, features, 2]
|
1238 |
+
|
1239 |
+
# Reshape to combine the last two dimensions
|
1240 |
+
x = x.view(batch_size, seq_len, seq_len, features * 2) # [batch_size, seq_len, seq_len, features * 2]
|
1241 |
+
|
1242 |
+
return x
|
1243 |
+
|
1244 |
+
def forward(self,
|
1245 |
+
input_features,
|
1246 |
+
add_pooling: Optional[bool] = True
|
1247 |
+
):
|
1248 |
+
# [s, b, h] -> [b, s ,h]
|
1249 |
+
input_features = input_features.transpose(0,1).contiguous()
|
1250 |
+
input_features = self.outer_concat(input_features)
|
1251 |
+
for i, layer in enumerate(self.layers):
|
1252 |
+
if i > 0:
|
1253 |
+
input_features = self.activation_func(input_features)
|
1254 |
+
input_features = layer(input_features)
|
1255 |
+
return input_features
|
1256 |
+
|
1257 |
+
|
1258 |
+
|
1259 |
+
|
1260 |
+
|
1261 |
+
class ProteinGLMForCasualLM(ProteinGLMPreTrainedModel):
|
1262 |
+
def __init__(self, config: ProteinGLMConfig, empty_init=True, device=None):
|
1263 |
+
super().__init__(config)
|
1264 |
+
|
1265 |
+
self.max_sequence_length = config.max_length
|
1266 |
+
self.transformer = ProteinGLMModel(config, empty_init=empty_init, device=device)
|
1267 |
+
self.config = config
|
1268 |
+
if self.config.quantization_bit:
|
1269 |
+
print(f"Begin Quantization to {self.config.quantization_bit} bit")
|
1270 |
+
self.quantize(self.config.quantization_bit, empty_init=True, device=device)
|
1271 |
+
|
1272 |
+
def _update_model_kwargs_for_generation(
|
1273 |
+
self,
|
1274 |
+
outputs: ModelOutput,
|
1275 |
+
model_kwargs: Dict[str, Any],
|
1276 |
+
is_encoder_decoder: bool = False,
|
1277 |
+
standardize_cache_format: bool = False,
|
1278 |
+
) -> Dict[str, Any]:
|
1279 |
+
# update past_key_values
|
1280 |
+
cache_name, cache = self._extract_past_from_model_output(outputs)
|
1281 |
+
model_kwargs[cache_name] = cache
|
1282 |
+
|
1283 |
+
# update attention mask
|
1284 |
+
if "attention_mask" in model_kwargs:
|
1285 |
+
attention_mask = model_kwargs["attention_mask"]
|
1286 |
+
model_kwargs["attention_mask"] = torch.cat(
|
1287 |
+
[attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1
|
1288 |
+
)
|
1289 |
+
|
1290 |
+
# update position ids
|
1291 |
+
if "position_ids" in model_kwargs:
|
1292 |
+
position_ids = model_kwargs["position_ids"]
|
1293 |
+
new_position_id = position_ids[..., -1:].clone() # [batch_size, 2, 1]
|
1294 |
+
if self.config.rotary_embedding_2d:
|
1295 |
+
new_position_id[:, 1] += 1 # Only update the 2nd dimension
|
1296 |
+
else:
|
1297 |
+
new_position_id[:] += 1
|
1298 |
+
model_kwargs["position_ids"] = torch.cat(
|
1299 |
+
[position_ids, new_position_id], dim=-1
|
1300 |
+
) # [batch_size, 2, seq_len+1]
|
1301 |
+
|
1302 |
+
model_kwargs["is_first_forward"] = False
|
1303 |
+
return model_kwargs
|
1304 |
+
|
1305 |
+
def prepare_inputs_for_generation(
|
1306 |
+
self,
|
1307 |
+
input_ids: torch.LongTensor,
|
1308 |
+
past_key_values: Optional[torch.Tensor] = None,
|
1309 |
+
attention_mask: Optional[torch.Tensor] = None,
|
1310 |
+
position_ids: Optional[torch.Tensor] = None,
|
1311 |
+
use_cache: Optional[bool] = None,
|
1312 |
+
is_first_forward: bool = True,
|
1313 |
+
**kwargs
|
1314 |
+
) -> dict:
|
1315 |
+
# only last token for input_ids if past is not None
|
1316 |
+
if position_ids is None:
|
1317 |
+
position_ids = self.get_position_ids(input_ids, device=input_ids.device) # position_ids: [batch_size, 2, seq_len]
|
1318 |
+
if not is_first_forward:
|
1319 |
+
if past_key_values is not None:
|
1320 |
+
position_ids = position_ids[..., -1:]
|
1321 |
+
input_ids = input_ids[:, -1:]
|
1322 |
+
return {
|
1323 |
+
"input_ids": input_ids,
|
1324 |
+
"past_key_values": past_key_values,
|
1325 |
+
"position_ids": position_ids,
|
1326 |
+
"attention_mask": attention_mask,
|
1327 |
+
"return_last_logit": True,
|
1328 |
+
"use_cache": use_cache
|
1329 |
+
}
|
1330 |
+
|
1331 |
+
def forward(
|
1332 |
+
self,
|
1333 |
+
input_ids: Optional[torch.Tensor] = None,
|
1334 |
+
position_ids: Optional[torch.Tensor] = None,
|
1335 |
+
attention_mask: Optional[torch.Tensor] = None,
|
1336 |
+
past_key_values: Optional[Tuple[torch.FloatTensor]] = None,
|
1337 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
1338 |
+
labels: Optional[torch.Tensor] = None,
|
1339 |
+
use_cache: Optional[bool] = None,
|
1340 |
+
output_attentions: Optional[bool] = None,
|
1341 |
+
output_hidden_states: Optional[bool] = None,
|
1342 |
+
return_dict: Optional[bool] = None,
|
1343 |
+
return_last_logit: Optional[bool] = False
|
1344 |
+
):
|
1345 |
+
if self.config.is_causal:
|
1346 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
1347 |
+
else:
|
1348 |
+
use_cache = False
|
1349 |
+
|
1350 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1351 |
+
|
1352 |
+
if position_ids is None:
|
1353 |
+
position_ids = self.get_position_ids(input_ids, device=input_ids.device)
|
1354 |
+
|
1355 |
+
transformer_outputs = self.transformer(
|
1356 |
+
input_ids=input_ids,
|
1357 |
+
position_ids=position_ids, # position_ids: [batch_size, 2, seq_len]
|
1358 |
+
attention_mask=attention_mask,
|
1359 |
+
past_key_values=past_key_values,
|
1360 |
+
inputs_embeds=inputs_embeds,
|
1361 |
+
use_cache=use_cache,
|
1362 |
+
output_hidden_states=output_hidden_states,
|
1363 |
+
return_dict=return_dict
|
1364 |
+
)
|
1365 |
+
hidden_states = transformer_outputs[0]
|
1366 |
+
if return_last_logit:
|
1367 |
+
hidden_states = hidden_states[-1:]
|
1368 |
+
lm_logits = self.transformer.output_layer(hidden_states)
|
1369 |
+
lm_logits = lm_logits.transpose(0, 1).contiguous()
|
1370 |
+
|
1371 |
+
loss = None
|
1372 |
+
if labels is not None:
|
1373 |
+
lm_logits = lm_logits.to(torch.float32)
|
1374 |
+
|
1375 |
+
# Shift so that tokens < n predict n
|
1376 |
+
shift_logits = lm_logits[..., :-1, :].contiguous()
|
1377 |
+
shift_labels = labels[..., 1:].contiguous()
|
1378 |
+
# Flatten the tokens
|
1379 |
+
loss_fct = CrossEntropyLoss(ignore_index=-100)
|
1380 |
+
loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
|
1381 |
+
|
1382 |
+
lm_logits = lm_logits.to(hidden_states.dtype)
|
1383 |
+
loss = loss.to(hidden_states.dtype)
|
1384 |
+
|
1385 |
+
if not return_dict:
|
1386 |
+
output = (lm_logits,) + transformer_outputs[1:]
|
1387 |
+
return ((loss,) + output) if loss is not None else output
|
1388 |
+
|
1389 |
+
return CausalLMOutputWithPast(
|
1390 |
+
loss=loss,
|
1391 |
+
logits=lm_logits,
|
1392 |
+
past_key_values=transformer_outputs.past_key_values,
|
1393 |
+
hidden_states=transformer_outputs.hidden_states,
|
1394 |
+
attentions=transformer_outputs.attentions,
|
1395 |
+
)
|
1396 |
+
|
1397 |
+
@staticmethod
|
1398 |
+
def _reorder_cache(
|
1399 |
+
past: Tuple[Tuple[torch.Tensor, torch.Tensor], ...], beam_idx: torch.LongTensor
|
1400 |
+
) -> Tuple[Tuple[torch.Tensor, torch.Tensor], ...]:
|
1401 |
+
"""
|
1402 |
+
This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
|
1403 |
+
[`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
|
1404 |
+
beam_idx at every generation step.
|
1405 |
+
|
1406 |
+
Output shares the same memory storage as `past`.
|
1407 |
+
"""
|
1408 |
+
return tuple(
|
1409 |
+
(
|
1410 |
+
layer_past[0].index_select(1, beam_idx.to(layer_past[0].device)),
|
1411 |
+
layer_past[1].index_select(1, beam_idx.to(layer_past[1].device)),
|
1412 |
+
)
|
1413 |
+
for layer_past in past
|
1414 |
+
)
|
1415 |
+
|
1416 |
+
@torch.inference_mode()
|
1417 |
+
def chat(self, tokenizer, query: str, max_length: int = 256, num_beams=1, do_sample=True,
|
1418 |
+
top_p=1.0, temperature=1.0, logits_processor=None, **kwargs):
|
1419 |
+
if logits_processor is None:
|
1420 |
+
logits_processor = LogitsProcessorList()
|
1421 |
+
logits_processor.append(InvalidScoreLogitsProcessor())
|
1422 |
+
gen_kwargs = {"max_length": max_length, "num_beams": num_beams, "do_sample": do_sample, "top_p": top_p,
|
1423 |
+
"temperature": temperature, "logits_processor": logits_processor, **kwargs}
|
1424 |
+
inputs = tokenizer.apply_chat_template(query, add_generation_prompt=True, tokenize=True,
|
1425 |
+
return_tensors="pt", return_dict=True)
|
1426 |
+
position_ids = self.get_position_ids(inputs['input_ids'], device=self.device) # TODO: ADD BATCH
|
1427 |
+
eos_token_id = [tokenizer.eos_token_id, tokenizer.convert_tokens_to_ids("<eop>")]
|
1428 |
+
inputs["position_ids"] = position_ids
|
1429 |
+
inputs = inputs.to(self.device)
|
1430 |
+
outputs = self.generate(**inputs, **gen_kwargs, eos_token_id=eos_token_id)
|
1431 |
+
outputs = outputs.tolist()[0][3:] # 3 for generation prompt "<gmask><sop><eos>"
|
1432 |
+
if outputs[-1] in eos_token_id:
|
1433 |
+
outputs = outputs[:-1]
|
1434 |
+
response = tokenizer.decode(outputs)
|
1435 |
+
return response
|
1436 |
+
|
1437 |
+
# TODO: fix bug in streaming chat
|
1438 |
+
@torch.inference_mode()
|
1439 |
+
def stream_chat(self, tokenizer, query: str, max_length: int = 56, num_beams=1, do_sample=True,
|
1440 |
+
top_p=0.8, temperature=0.8, logits_processor=None, past_key_values = None, **kwargs):
|
1441 |
+
if logits_processor is None:
|
1442 |
+
logits_processor = LogitsProcessorList()
|
1443 |
+
logits_processor.append(InvalidScoreLogitsProcessor())
|
1444 |
+
eos_token_id = [tokenizer.eos_token_id, tokenizer.convert_tokens_to_ids("<eop>")]
|
1445 |
+
gen_kwargs = {"max_length": max_length, "do_sample": do_sample, "top_p": top_p,
|
1446 |
+
"temperature": temperature, "logits_processor": logits_processor, **kwargs}
|
1447 |
+
inputs = tokenizer.apply_chat_template(query, add_generation_prompt=True, tokenize=True,
|
1448 |
+
return_tensors="pt", return_dict=True)
|
1449 |
+
position_ids = self.get_position_ids(inputs['input_ids'], device=self.device) # TODO: ADD BATCH
|
1450 |
+
eos_token_id = [tokenizer.eos_token_id, tokenizer.convert_tokens_to_ids("<eop>")]
|
1451 |
+
inputs["position_ids"] = position_ids
|
1452 |
+
inputs = inputs.to(self.device)
|
1453 |
+
offset = 3 # 3 for generation prompt
|
1454 |
+
for outputs in self.stream_generate(**inputs, past_key_values=past_key_values,
|
1455 |
+
eos_token_id=eos_token_id, return_past_key_values=False,
|
1456 |
+
**gen_kwargs):
|
1457 |
+
outputs = outputs.tolist()[0][3:]
|
1458 |
+
if outputs[-1] in eos_token_id:
|
1459 |
+
outputs = outputs[:-1]
|
1460 |
+
# offset = 3 + len(outputs)
|
1461 |
+
response = tokenizer.decode(outputs)
|
1462 |
+
if response:
|
1463 |
+
yield response
|
1464 |
+
|
1465 |
+
@torch.inference_mode()
|
1466 |
+
def stream_generate(
|
1467 |
+
self,
|
1468 |
+
input_ids,
|
1469 |
+
generation_config: Optional[GenerationConfig] = None,
|
1470 |
+
logits_processor: Optional[LogitsProcessorList] = None,
|
1471 |
+
stopping_criteria: Optional[StoppingCriteriaList] = None,
|
1472 |
+
prefix_allowed_tokens_fn: Optional[Callable[[int, torch.Tensor], List[int]]] = None,
|
1473 |
+
return_past_key_values=False,
|
1474 |
+
**kwargs,
|
1475 |
+
):
|
1476 |
+
breakpoint()
|
1477 |
+
batch_size, input_ids_seq_length = input_ids.shape[0], input_ids.shape[-1]
|
1478 |
+
|
1479 |
+
if generation_config is None:
|
1480 |
+
generation_config = self.generation_config
|
1481 |
+
generation_config = copy.deepcopy(generation_config)
|
1482 |
+
model_kwargs = generation_config.update(**kwargs)
|
1483 |
+
model_kwargs["use_cache"] = generation_config.use_cache
|
1484 |
+
bos_token_id, eos_token_id = generation_config.bos_token_id, generation_config.eos_token_id
|
1485 |
+
|
1486 |
+
if isinstance(eos_token_id, int):
|
1487 |
+
eos_token_id = [eos_token_id]
|
1488 |
+
eos_token_id_tensor = torch.tensor(eos_token_id).to(input_ids.device) if eos_token_id is not None else None
|
1489 |
+
|
1490 |
+
has_default_max_length = kwargs.get("max_length") is None and generation_config.max_length is not None
|
1491 |
+
if has_default_max_length and generation_config.max_new_tokens is None:
|
1492 |
+
warnings.warn(
|
1493 |
+
f"Using `max_length`'s default ({generation_config.max_length}) to control the generation length. "
|
1494 |
+
"This behaviour is deprecated and will be removed from the config in v5 of Transformers -- we"
|
1495 |
+
" recommend using `max_new_tokens` to control the maximum length of the generation.",
|
1496 |
+
UserWarning,
|
1497 |
+
)
|
1498 |
+
elif generation_config.max_new_tokens is not None:
|
1499 |
+
generation_config.max_length = generation_config.max_new_tokens + input_ids_seq_length
|
1500 |
+
if not has_default_max_length:
|
1501 |
+
logger.warn(
|
1502 |
+
f"Both `max_new_tokens` (={generation_config.max_new_tokens}) and `max_length`(="
|
1503 |
+
f"{generation_config.max_length}) seem to have been set. `max_new_tokens` will take precedence. "
|
1504 |
+
"Please refer to the documentation for more information. "
|
1505 |
+
"(https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)",
|
1506 |
+
UserWarning,
|
1507 |
+
)
|
1508 |
+
|
1509 |
+
if input_ids_seq_length >= generation_config.max_length:
|
1510 |
+
input_ids_string = "decoder_input_ids" if self.config.is_encoder_decoder else "input_ids"
|
1511 |
+
logger.warning(
|
1512 |
+
f"Input length of {input_ids_string} is {input_ids_seq_length}, but `max_length` is set to"
|
1513 |
+
f" {generation_config.max_length}. This can lead to unexpected behavior. You should consider"
|
1514 |
+
" increasing `max_new_tokens`."
|
1515 |
+
)
|
1516 |
+
|
1517 |
+
# 2. Set generation parameters if not already defined
|
1518 |
+
logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
|
1519 |
+
stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
|
1520 |
+
|
1521 |
+
logits_processor = self._get_logits_processor(
|
1522 |
+
generation_config=generation_config,
|
1523 |
+
input_ids_seq_length=input_ids_seq_length,
|
1524 |
+
encoder_input_ids=input_ids,
|
1525 |
+
prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
|
1526 |
+
logits_processor=logits_processor,
|
1527 |
+
)
|
1528 |
+
|
1529 |
+
stopping_criteria = self._get_stopping_criteria(
|
1530 |
+
generation_config=generation_config, stopping_criteria=stopping_criteria
|
1531 |
+
)
|
1532 |
+
logits_warper = self._get_logits_warper(generation_config)
|
1533 |
+
|
1534 |
+
unfinished_sequences = input_ids.new(input_ids.shape[0]).fill_(1)
|
1535 |
+
scores = None
|
1536 |
+
while True:
|
1537 |
+
model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)
|
1538 |
+
# forward pass to get next token
|
1539 |
+
outputs = self(
|
1540 |
+
**model_inputs,
|
1541 |
+
return_dict=True,
|
1542 |
+
output_attentions=False,
|
1543 |
+
output_hidden_states=False,
|
1544 |
+
)
|
1545 |
+
|
1546 |
+
next_token_logits = outputs.logits[:, -1, :]
|
1547 |
+
|
1548 |
+
# pre-process distribution
|
1549 |
+
next_token_scores = logits_processor(input_ids, next_token_logits)
|
1550 |
+
next_token_scores = logits_warper(input_ids, next_token_scores)
|
1551 |
+
|
1552 |
+
# sample
|
1553 |
+
probs = nn.functional.softmax(next_token_scores, dim=-1)
|
1554 |
+
if generation_config.do_sample:
|
1555 |
+
next_tokens = torch.multinomial(probs, num_samples=1).squeeze(1)
|
1556 |
+
else:
|
1557 |
+
next_tokens = torch.argmax(probs, dim=-1)
|
1558 |
+
# update generated ids, model inputs, and length for next step
|
1559 |
+
input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1)
|
1560 |
+
model_kwargs = self._update_model_kwargs_for_generation(
|
1561 |
+
outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
|
1562 |
+
)
|
1563 |
+
unfinished_sequences = unfinished_sequences.mul(
|
1564 |
+
next_tokens.tile(eos_token_id_tensor.shape[0], 1).ne(eos_token_id_tensor.unsqueeze(1)).prod(dim=0)
|
1565 |
+
)
|
1566 |
+
if return_past_key_values:
|
1567 |
+
yield input_ids, outputs.past_key_values
|
1568 |
+
else:
|
1569 |
+
yield input_ids
|
1570 |
+
# stop when each sentence is finished, or if we exceed the maximum length
|
1571 |
+
if unfinished_sequences.max() == 0 or stopping_criteria(input_ids, scores):
|
1572 |
+
break
|
quantization.py
ADDED
@@ -0,0 +1,188 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from torch.nn import Linear
|
2 |
+
from torch.nn.parameter import Parameter
|
3 |
+
|
4 |
+
import bz2
|
5 |
+
import torch
|
6 |
+
import base64
|
7 |
+
import ctypes
|
8 |
+
from transformers.utils import logging
|
9 |
+
|
10 |
+
from typing import List
|
11 |
+
from functools import partial
|
12 |
+
|
13 |
+
logger = logging.get_logger(__name__)
|
14 |
+
|
15 |
+
try:
|
16 |
+
from cpm_kernels.kernels.base import LazyKernelCModule, KernelFunction, round_up
|
17 |
+
|
18 |
+
class Kernel:
|
19 |
+
def __init__(self, code: bytes, function_names: List[str]):
|
20 |
+
self.code = code
|
21 |
+
self._function_names = function_names
|
22 |
+
self._cmodule = LazyKernelCModule(self.code)
|
23 |
+
|
24 |
+
for name in self._function_names:
|
25 |
+
setattr(self, name, KernelFunction(self._cmodule, name))
|
26 |
+
|
27 |
+
quantization_code = "$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"
|
28 |
+
|
29 |
+
kernels = Kernel(
|
30 |
+
bz2.decompress(base64.b64decode(quantization_code)),
|
31 |
+
[
|
32 |
+
"int4WeightCompression",
|
33 |
+
"int4WeightExtractionFloat",
|
34 |
+
"int4WeightExtractionHalf",
|
35 |
+
"int8WeightExtractionFloat",
|
36 |
+
"int8WeightExtractionHalf"
|
37 |
+
],
|
38 |
+
)
|
39 |
+
except Exception as exception:
|
40 |
+
kernels = None
|
41 |
+
logger.warning("Failed to load cpm_kernels:" + str(exception))
|
42 |
+
|
43 |
+
|
44 |
+
class W8A16Linear(torch.autograd.Function):
|
45 |
+
@staticmethod
|
46 |
+
def forward(ctx, inp: torch.Tensor, quant_w: torch.Tensor, scale_w: torch.Tensor, weight_bit_width):
|
47 |
+
ctx.inp_shape = inp.size()
|
48 |
+
ctx.weight_bit_width = weight_bit_width
|
49 |
+
out_features = quant_w.size(0)
|
50 |
+
inp = inp.contiguous().view(-1, inp.size(-1))
|
51 |
+
weight = extract_weight_to_half(quant_w, scale_w, weight_bit_width)
|
52 |
+
ctx.weight_shape = weight.size()
|
53 |
+
output = inp.mm(weight.t())
|
54 |
+
ctx.save_for_backward(inp, quant_w, scale_w)
|
55 |
+
return output.view(*(ctx.inp_shape[:-1] + (out_features,)))
|
56 |
+
|
57 |
+
@staticmethod
|
58 |
+
def backward(ctx, grad_output: torch.Tensor):
|
59 |
+
inp, quant_w, scale_w = ctx.saved_tensors
|
60 |
+
weight = extract_weight_to_half(quant_w, scale_w, ctx.weight_bit_width)
|
61 |
+
grad_output = grad_output.contiguous().view(-1, weight.size(0))
|
62 |
+
grad_input = grad_output.mm(weight)
|
63 |
+
grad_weight = grad_output.t().mm(inp)
|
64 |
+
return grad_input.view(ctx.inp_shape), grad_weight.view(ctx.weight_shape), None, None
|
65 |
+
|
66 |
+
|
67 |
+
def compress_int4_weight(weight: torch.Tensor): # (n, m)
|
68 |
+
with torch.cuda.device(weight.device):
|
69 |
+
n, m = weight.size(0), weight.size(1)
|
70 |
+
assert m % 2 == 0
|
71 |
+
m = m // 2
|
72 |
+
out = torch.empty(n, m, dtype=torch.int8, device="cuda")
|
73 |
+
stream = torch.cuda.current_stream()
|
74 |
+
|
75 |
+
gridDim = (n, 1, 1)
|
76 |
+
blockDim = (min(round_up(m, 32), 1024), 1, 1)
|
77 |
+
|
78 |
+
kernels.int4WeightCompression(
|
79 |
+
gridDim,
|
80 |
+
blockDim,
|
81 |
+
0,
|
82 |
+
stream,
|
83 |
+
[ctypes.c_void_p(weight.data_ptr()), ctypes.c_void_p(out.data_ptr()), ctypes.c_int32(n), ctypes.c_int32(m)],
|
84 |
+
)
|
85 |
+
return out
|
86 |
+
|
87 |
+
|
88 |
+
def extract_weight_to_half(weight: torch.Tensor, scale_list: torch.Tensor, source_bit_width: int):
|
89 |
+
assert scale_list.dtype in [torch.half, torch.bfloat16]
|
90 |
+
assert weight.dtype in [torch.int8]
|
91 |
+
if source_bit_width == 8:
|
92 |
+
return weight.to(scale_list.dtype) * scale_list[:, None]
|
93 |
+
elif source_bit_width == 4:
|
94 |
+
func = (
|
95 |
+
kernels.int4WeightExtractionHalf if scale_list.dtype == torch.half else kernels.int4WeightExtractionBFloat16
|
96 |
+
)
|
97 |
+
else:
|
98 |
+
assert False, "Unsupported bit-width"
|
99 |
+
|
100 |
+
with torch.cuda.device(weight.device):
|
101 |
+
n, m = weight.size(0), weight.size(1)
|
102 |
+
out = torch.empty(n, m * (8 // source_bit_width), dtype=scale_list.dtype, device="cuda")
|
103 |
+
stream = torch.cuda.current_stream()
|
104 |
+
|
105 |
+
gridDim = (n, 1, 1)
|
106 |
+
blockDim = (min(round_up(m, 32), 1024), 1, 1)
|
107 |
+
|
108 |
+
func(
|
109 |
+
gridDim,
|
110 |
+
blockDim,
|
111 |
+
0,
|
112 |
+
stream,
|
113 |
+
[
|
114 |
+
ctypes.c_void_p(weight.data_ptr()),
|
115 |
+
ctypes.c_void_p(scale_list.data_ptr()),
|
116 |
+
ctypes.c_void_p(out.data_ptr()),
|
117 |
+
ctypes.c_int32(n),
|
118 |
+
ctypes.c_int32(m),
|
119 |
+
],
|
120 |
+
)
|
121 |
+
return out
|
122 |
+
|
123 |
+
|
124 |
+
class QuantizedLinear(torch.nn.Module):
|
125 |
+
def __init__(self, weight_bit_width: int, weight, bias=None, device="cpu", dtype=None, empty_init=False, *args,
|
126 |
+
**kwargs):
|
127 |
+
super().__init__()
|
128 |
+
self.weight_bit_width = weight_bit_width
|
129 |
+
|
130 |
+
shape = weight.shape
|
131 |
+
|
132 |
+
if weight is None or empty_init:
|
133 |
+
self.weight = torch.empty(shape[0], shape[1] * weight_bit_width // 8, dtype=torch.int8, device=device)
|
134 |
+
self.weight_scale = torch.empty(shape[0], dtype=dtype, device=device)
|
135 |
+
else:
|
136 |
+
self.weight_scale = weight.abs().max(dim=-1).values / ((2 ** (weight_bit_width - 1)) - 1)
|
137 |
+
self.weight = torch.round(weight / self.weight_scale[:, None]).to(torch.int8)
|
138 |
+
if weight_bit_width == 4:
|
139 |
+
self.weight = compress_int4_weight(self.weight)
|
140 |
+
|
141 |
+
self.weight = Parameter(self.weight.to(device), requires_grad=False)
|
142 |
+
self.weight_scale = Parameter(self.weight_scale.to(device), requires_grad=False)
|
143 |
+
self.bias = Parameter(bias.to(device), requires_grad=False) if bias is not None else None
|
144 |
+
|
145 |
+
def forward(self, input):
|
146 |
+
output = W8A16Linear.apply(input, self.weight, self.weight_scale, self.weight_bit_width)
|
147 |
+
if self.bias is not None:
|
148 |
+
output = output + self.bias
|
149 |
+
return output
|
150 |
+
|
151 |
+
|
152 |
+
def quantize(model, weight_bit_width, empty_init=False, device=None):
|
153 |
+
"""Replace fp16 linear with quantized linear"""
|
154 |
+
for layer in model.layers:
|
155 |
+
layer.self_attention.query_key_value = QuantizedLinear(
|
156 |
+
weight_bit_width=weight_bit_width,
|
157 |
+
weight=layer.self_attention.query_key_value.weight.to(torch.cuda.current_device()),
|
158 |
+
bias=layer.self_attention.query_key_value.bias,
|
159 |
+
dtype=layer.self_attention.query_key_value.weight.dtype,
|
160 |
+
device=layer.self_attention.query_key_value.weight.device if device is None else device,
|
161 |
+
empty_init=empty_init
|
162 |
+
)
|
163 |
+
layer.self_attention.dense = QuantizedLinear(
|
164 |
+
weight_bit_width=weight_bit_width,
|
165 |
+
weight=layer.self_attention.dense.weight.to(torch.cuda.current_device()),
|
166 |
+
bias=layer.self_attention.dense.bias,
|
167 |
+
dtype=layer.self_attention.dense.weight.dtype,
|
168 |
+
device=layer.self_attention.dense.weight.device if device is None else device,
|
169 |
+
empty_init=empty_init
|
170 |
+
)
|
171 |
+
layer.mlp.dense_h_to_4h = QuantizedLinear(
|
172 |
+
weight_bit_width=weight_bit_width,
|
173 |
+
weight=layer.mlp.dense_h_to_4h.weight.to(torch.cuda.current_device()),
|
174 |
+
bias=layer.mlp.dense_h_to_4h.bias,
|
175 |
+
dtype=layer.mlp.dense_h_to_4h.weight.dtype,
|
176 |
+
device=layer.mlp.dense_h_to_4h.weight.device if device is None else device,
|
177 |
+
empty_init=empty_init
|
178 |
+
)
|
179 |
+
layer.mlp.dense_4h_to_h = QuantizedLinear(
|
180 |
+
weight_bit_width=weight_bit_width,
|
181 |
+
weight=layer.mlp.dense_4h_to_h.weight.to(torch.cuda.current_device()),
|
182 |
+
bias=layer.mlp.dense_4h_to_h.bias,
|
183 |
+
dtype=layer.mlp.dense_4h_to_h.weight.dtype,
|
184 |
+
device=layer.mlp.dense_4h_to_h.weight.device if device is None else device,
|
185 |
+
empty_init=empty_init
|
186 |
+
)
|
187 |
+
|
188 |
+
return model
|
special_tokens_map.json
ADDED
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"additional_special_tokens": [
|
3 |
+
"<pad>",
|
4 |
+
"<mask>",
|
5 |
+
"<gmask>",
|
6 |
+
"<smask>",
|
7 |
+
"<eod>",
|
8 |
+
"<sop>",
|
9 |
+
"<eop>",
|
10 |
+
"<eos>",
|
11 |
+
"<unk>"
|
12 |
+
],
|
13 |
+
"eos_token": "<eos>",
|
14 |
+
"mask_token": "<mask>",
|
15 |
+
"pad_token": "<pad>",
|
16 |
+
"unk_token": "<unk>"
|
17 |
+
}
|
tokenization_proteinglm.py
ADDED
@@ -0,0 +1,140 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""Tokenization classes for ProteinGLM."""
|
2 |
+
|
3 |
+
import os
|
4 |
+
from typing import List, Optional, Union, Dict, Any
|
5 |
+
from torch import TensorType
|
6 |
+
from transformers import PreTrainedTokenizer, PreTrainedTokenizerFast
|
7 |
+
from transformers.tokenization_utils_base import EncodedInput, BatchEncoding
|
8 |
+
|
9 |
+
VOCAB_FILES_NAMES = {"vocab_file": "tokenizer.model"}
|
10 |
+
|
11 |
+
|
12 |
+
def load_vocab_file(vocab_file: str) -> List[str]:
|
13 |
+
with open(vocab_file, "r") as f:
|
14 |
+
lines = f.read().splitlines()
|
15 |
+
return [line.strip() for line in lines]
|
16 |
+
|
17 |
+
|
18 |
+
class ProteinGLMTokenizer(PreTrainedTokenizer):
|
19 |
+
"""
|
20 |
+
Constructs a ProteinGLM tokenizer.
|
21 |
+
"""
|
22 |
+
|
23 |
+
vocab_files_names = VOCAB_FILES_NAMES
|
24 |
+
model_input_names = ["input_ids", "attention_mask", "position_ids"]
|
25 |
+
def __init__(
|
26 |
+
self,
|
27 |
+
vocab_file: str,
|
28 |
+
unk_token: str = "<unk>",
|
29 |
+
pad_token: str = "<pad>",
|
30 |
+
mask_token: str = "<mask>",
|
31 |
+
eos_token: str = "<eos>",
|
32 |
+
model_max_length: int = 2048,
|
33 |
+
additional_special_tokens: Optional[List[str]] = None,
|
34 |
+
**kwargs,
|
35 |
+
):
|
36 |
+
self.all_tokens = load_vocab_file(vocab_file)
|
37 |
+
self._id_to_token = dict(enumerate(self.all_tokens))
|
38 |
+
self._token_to_id = {tok: ind for ind, tok in enumerate(self.all_tokens)}
|
39 |
+
|
40 |
+
if additional_special_tokens is None:
|
41 |
+
additional_special_tokens = ['<pad>', '<mask>', '<gmask>', '<smask>', '<eod>', '<sop>', '<eop>', '<eos>', '<unk>']
|
42 |
+
|
43 |
+
super().__init__(
|
44 |
+
unk_token=unk_token,
|
45 |
+
pad_token=pad_token,
|
46 |
+
mask_token=mask_token,
|
47 |
+
eos_token=eos_token,
|
48 |
+
model_max_length=model_max_length,
|
49 |
+
additional_special_tokens=additional_special_tokens,
|
50 |
+
**kwargs,
|
51 |
+
)
|
52 |
+
|
53 |
+
self.unique_no_split_tokens = self.all_tokens
|
54 |
+
self._update_trie(self.unique_no_split_tokens)
|
55 |
+
|
56 |
+
def _convert_id_to_token(self, index: int) -> str:
|
57 |
+
return self._id_to_token.get(index, self.unk_token)
|
58 |
+
|
59 |
+
def _convert_token_to_id(self, token: str) -> int:
|
60 |
+
return self._token_to_id.get(token, self._token_to_id.get(self.unk_token))
|
61 |
+
|
62 |
+
def _tokenize(self, text: str, **kwargs) -> List[str]:
|
63 |
+
return text.split()
|
64 |
+
|
65 |
+
def get_vocab(self) -> dict:
|
66 |
+
base_vocab = self._token_to_id.copy()
|
67 |
+
base_vocab.update(self.added_tokens_encoder)
|
68 |
+
return base_vocab
|
69 |
+
|
70 |
+
def token_to_id(self, token: str) -> int:
|
71 |
+
return self._token_to_id.get(token, self._token_to_id.get(self.unk_token))
|
72 |
+
|
73 |
+
def id_to_token(self, index: int) -> str:
|
74 |
+
return self._id_to_token.get(index, self.unk_token)
|
75 |
+
|
76 |
+
def build_inputs_with_special_tokens(
|
77 |
+
self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
|
78 |
+
) -> List[int]:
|
79 |
+
sep = [self.eos_token_id]
|
80 |
+
if token_ids_1 is None:
|
81 |
+
if self.eos_token_id is None:
|
82 |
+
return token_ids_0
|
83 |
+
else:
|
84 |
+
return token_ids_0 + sep
|
85 |
+
elif self.eos_token_id is None:
|
86 |
+
raise ValueError("Cannot tokenize multiple sequences when EOS token is not set!")
|
87 |
+
return token_ids_0 + sep + token_ids_1 + sep # Multiple inputs always have an EOS token
|
88 |
+
|
89 |
+
|
90 |
+
def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> tuple:
|
91 |
+
vocab_file = os.path.join(save_directory, (filename_prefix + "-" if filename_prefix else "") + "tokenizer.model")
|
92 |
+
with open(vocab_file, "w") as f:
|
93 |
+
f.write("\n".join(self.all_tokens))
|
94 |
+
return (vocab_file,)
|
95 |
+
|
96 |
+
@property
|
97 |
+
def vocab_size(self) -> int:
|
98 |
+
return len(self.all_tokens)
|
99 |
+
|
100 |
+
def apply_chat_template(
|
101 |
+
self,
|
102 |
+
query,
|
103 |
+
add_generation_prompt: bool = True,
|
104 |
+
tokenize: bool = True,
|
105 |
+
padding: bool = False,
|
106 |
+
truncation: bool = False,
|
107 |
+
max_length: Optional[int] = None,
|
108 |
+
return_tensors: Optional[Union[str, TensorType]] = None,
|
109 |
+
return_dict: bool = False,
|
110 |
+
tokenizer_kwargs: Optional[Dict[str, Any]] = None,
|
111 |
+
add_special_tokens: bool = True,
|
112 |
+
**kwargs,
|
113 |
+
) -> Union[str, List[int], List[str], List[List[int]], BatchEncoding]:
|
114 |
+
|
115 |
+
generation_prompt = "<gmask><sop><eos>"
|
116 |
+
if isinstance(query, str):
|
117 |
+
query = [query]
|
118 |
+
prompt_query = []
|
119 |
+
if add_generation_prompt:
|
120 |
+
for each in query:
|
121 |
+
assert isinstance(each, str)
|
122 |
+
prompt_query.append(generation_prompt+each)
|
123 |
+
else:
|
124 |
+
prompt_query = query
|
125 |
+
if tokenize:
|
126 |
+
output = self.batch_encode_plus(
|
127 |
+
prompt_query,
|
128 |
+
padding=padding,
|
129 |
+
truncation=truncation,
|
130 |
+
max_length=max_length,
|
131 |
+
return_tensors=return_tensors,
|
132 |
+
is_split_into_words=True,
|
133 |
+
add_special_tokens=False
|
134 |
+
)
|
135 |
+
if return_dict:
|
136 |
+
return output
|
137 |
+
else:
|
138 |
+
return output["input_ids"]
|
139 |
+
else:
|
140 |
+
return prompt_query
|
tokenizer.model
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:9bd2746ab3ae26b1ae1c4246e0d35f895c86c9685fcd85f8aa89a9a08e534da0
|
3 |
+
size 112
|
tokenizer_config.json
ADDED
@@ -0,0 +1,100 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"added_tokens_decoder": {
|
3 |
+
"0": {
|
4 |
+
"content": "<pad>",
|
5 |
+
"lstrip": false,
|
6 |
+
"normalized": false,
|
7 |
+
"rstrip": false,
|
8 |
+
"single_word": false,
|
9 |
+
"special": true
|
10 |
+
},
|
11 |
+
"28": {
|
12 |
+
"content": "<mask>",
|
13 |
+
"lstrip": false,
|
14 |
+
"normalized": false,
|
15 |
+
"rstrip": false,
|
16 |
+
"single_word": false,
|
17 |
+
"special": true
|
18 |
+
},
|
19 |
+
"29": {
|
20 |
+
"content": "<gmask>",
|
21 |
+
"lstrip": false,
|
22 |
+
"normalized": false,
|
23 |
+
"rstrip": false,
|
24 |
+
"single_word": false,
|
25 |
+
"special": true
|
26 |
+
},
|
27 |
+
"30": {
|
28 |
+
"content": "<smask>",
|
29 |
+
"lstrip": false,
|
30 |
+
"normalized": false,
|
31 |
+
"rstrip": false,
|
32 |
+
"single_word": false,
|
33 |
+
"special": true
|
34 |
+
},
|
35 |
+
"31": {
|
36 |
+
"content": "<eod>",
|
37 |
+
"lstrip": false,
|
38 |
+
"normalized": false,
|
39 |
+
"rstrip": false,
|
40 |
+
"single_word": false,
|
41 |
+
"special": true
|
42 |
+
},
|
43 |
+
"32": {
|
44 |
+
"content": "<sop>",
|
45 |
+
"lstrip": false,
|
46 |
+
"normalized": false,
|
47 |
+
"rstrip": false,
|
48 |
+
"single_word": false,
|
49 |
+
"special": true
|
50 |
+
},
|
51 |
+
"33": {
|
52 |
+
"content": "<eop>",
|
53 |
+
"lstrip": false,
|
54 |
+
"normalized": false,
|
55 |
+
"rstrip": false,
|
56 |
+
"single_word": false,
|
57 |
+
"special": true
|
58 |
+
},
|
59 |
+
"34": {
|
60 |
+
"content": "<eos>",
|
61 |
+
"lstrip": false,
|
62 |
+
"normalized": false,
|
63 |
+
"rstrip": false,
|
64 |
+
"single_word": false,
|
65 |
+
"special": true
|
66 |
+
},
|
67 |
+
"35": {
|
68 |
+
"content": "<unk>",
|
69 |
+
"lstrip": false,
|
70 |
+
"normalized": false,
|
71 |
+
"rstrip": false,
|
72 |
+
"single_word": false,
|
73 |
+
"special": true
|
74 |
+
}
|
75 |
+
},
|
76 |
+
"additional_special_tokens": [
|
77 |
+
"<pad>",
|
78 |
+
"<mask>",
|
79 |
+
"<gmask>",
|
80 |
+
"<smask>",
|
81 |
+
"<eod>",
|
82 |
+
"<sop>",
|
83 |
+
"<eop>",
|
84 |
+
"<eos>",
|
85 |
+
"<unk>"
|
86 |
+
],
|
87 |
+
"auto_map": {
|
88 |
+
"AutoTokenizer": [
|
89 |
+
"tokenization_proteinglm.ProteinGLMTokenizer",
|
90 |
+
null
|
91 |
+
]
|
92 |
+
},
|
93 |
+
"clean_up_tokenization_spaces": true,
|
94 |
+
"eos_token": "<eos>",
|
95 |
+
"mask_token": "<mask>",
|
96 |
+
"model_max_length": 2048,
|
97 |
+
"pad_token": "<pad>",
|
98 |
+
"tokenizer_class": "ProteinGLMTokenizer",
|
99 |
+
"unk_token": "<unk>"
|
100 |
+
}
|