Commit
•
25a067a
1
Parent(s):
045263c
upload
Browse files- config.json +50 -0
- configuration_qwen.py +65 -0
- generation_config.json +4 -0
- modeling_qwen.py +1421 -0
- pytorch_model-00001-of-00002.bin +3 -0
- pytorch_model-00002-of-00002.bin +3 -0
- pytorch_model.bin.index.json +873 -0
- qwen.tiktoken +0 -0
- special_tokens_map.json +3 -0
- tokenization_qwen.py +598 -0
- tokenizer_config.json +12 -0
- zero_to_fp32.py +587 -0
config.json
ADDED
@@ -0,0 +1,50 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_name_or_path": "OdysseyAgent",
|
3 |
+
"architectures": [
|
4 |
+
"QWenLMHeadModel"
|
5 |
+
],
|
6 |
+
"attn_dropout_prob": 0.0,
|
7 |
+
"auto_map": {
|
8 |
+
"AutoConfig": "configuration_qwen.QWenConfig",
|
9 |
+
"AutoModelForCausalLM": "modeling_qwen.QWenLMHeadModel"
|
10 |
+
},
|
11 |
+
"bf16": true,
|
12 |
+
"emb_dropout_prob": 0.0,
|
13 |
+
"fp16": false,
|
14 |
+
"fp32": false,
|
15 |
+
"hidden_size": 4096,
|
16 |
+
"his_len": 4,
|
17 |
+
"initializer_range": 0.02,
|
18 |
+
"intermediate_size": 22016,
|
19 |
+
"kv_channels": 128,
|
20 |
+
"layer_norm_epsilon": 1e-06,
|
21 |
+
"max_position_embeddings": 8192,
|
22 |
+
"model_type": "qwen",
|
23 |
+
"no_bias": true,
|
24 |
+
"num_attention_heads": 32,
|
25 |
+
"num_hidden_layers": 32,
|
26 |
+
"onnx_safe": null,
|
27 |
+
"rotary_emb_base": 10000,
|
28 |
+
"rotary_pct": 1.0,
|
29 |
+
"scale_attn_weights": true,
|
30 |
+
"seq_length": 2048,
|
31 |
+
"tie_word_embeddings": false,
|
32 |
+
"tokenizer_type": "QWenTokenizer",
|
33 |
+
"torch_dtype": "bfloat16",
|
34 |
+
"transformers_version": "4.32.0",
|
35 |
+
"use_cache": true,
|
36 |
+
"use_dynamic_ntk": true,
|
37 |
+
"use_flash_attn": false,
|
38 |
+
"use_logn_attn": true,
|
39 |
+
"visual": {
|
40 |
+
"heads": 16,
|
41 |
+
"image_size": 448,
|
42 |
+
"image_start_id": 151857,
|
43 |
+
"layers": 48,
|
44 |
+
"mlp_ratio": 4.9231,
|
45 |
+
"output_dim": 4096,
|
46 |
+
"patch_size": 14,
|
47 |
+
"width": 1664
|
48 |
+
},
|
49 |
+
"vocab_size": 151936
|
50 |
+
}
|
configuration_qwen.py
ADDED
@@ -0,0 +1,65 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Alibaba Cloud.
|
2 |
+
#
|
3 |
+
# This source code is licensed under the license found in the
|
4 |
+
# LICENSE file in the root directory of this source tree.
|
5 |
+
|
6 |
+
from transformers import PretrainedConfig
|
7 |
+
|
8 |
+
|
9 |
+
class QWenConfig(PretrainedConfig):
|
10 |
+
model_type = "qwen"
|
11 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
12 |
+
|
13 |
+
def __init__(
|
14 |
+
self,
|
15 |
+
vocab_size=151936,
|
16 |
+
hidden_size=4096,
|
17 |
+
num_hidden_layers=32,
|
18 |
+
num_attention_heads=32,
|
19 |
+
emb_dropout_prob=0.0,
|
20 |
+
attn_dropout_prob=0.0,
|
21 |
+
layer_norm_epsilon=1e-6,
|
22 |
+
initializer_range=0.02,
|
23 |
+
max_position_embeddings=8192,
|
24 |
+
scale_attn_weights=True,
|
25 |
+
use_cache=True,
|
26 |
+
bf16=False,
|
27 |
+
fp16=False,
|
28 |
+
fp32=False,
|
29 |
+
kv_channels=128,
|
30 |
+
rotary_pct=1.0,
|
31 |
+
rotary_emb_base=10000,
|
32 |
+
use_dynamic_ntk=True,
|
33 |
+
use_logn_attn=True,
|
34 |
+
use_flash_attn="auto",
|
35 |
+
intermediate_size=22016,
|
36 |
+
no_bias=True,
|
37 |
+
tie_word_embeddings=False,
|
38 |
+
**kwargs,
|
39 |
+
):
|
40 |
+
self.vocab_size = vocab_size
|
41 |
+
self.hidden_size = hidden_size
|
42 |
+
self.intermediate_size = intermediate_size
|
43 |
+
self.num_hidden_layers = num_hidden_layers
|
44 |
+
self.num_attention_heads = num_attention_heads
|
45 |
+
self.emb_dropout_prob = emb_dropout_prob
|
46 |
+
self.attn_dropout_prob = attn_dropout_prob
|
47 |
+
self.layer_norm_epsilon = layer_norm_epsilon
|
48 |
+
self.initializer_range = initializer_range
|
49 |
+
self.scale_attn_weights = scale_attn_weights
|
50 |
+
self.use_cache = use_cache
|
51 |
+
self.max_position_embeddings = max_position_embeddings
|
52 |
+
self.bf16 = bf16
|
53 |
+
self.fp16 = fp16
|
54 |
+
self.fp32 = fp32
|
55 |
+
self.kv_channels = kv_channels
|
56 |
+
self.rotary_pct = rotary_pct
|
57 |
+
self.rotary_emb_base = rotary_emb_base
|
58 |
+
self.use_dynamic_ntk = use_dynamic_ntk
|
59 |
+
self.use_logn_attn = use_logn_attn
|
60 |
+
self.use_flash_attn = use_flash_attn
|
61 |
+
self.no_bias = no_bias
|
62 |
+
super().__init__(
|
63 |
+
tie_word_embeddings=tie_word_embeddings,
|
64 |
+
**kwargs
|
65 |
+
)
|
generation_config.json
ADDED
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_from_model_config": true,
|
3 |
+
"transformers_version": "4.32.0"
|
4 |
+
}
|
modeling_qwen.py
ADDED
@@ -0,0 +1,1421 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Alibaba Cloud.
|
2 |
+
#
|
3 |
+
# This source code is licensed under the license found in the
|
4 |
+
# LICENSE file in the root directory of this source tree.
|
5 |
+
|
6 |
+
print('OdysseyAgent')
|
7 |
+
|
8 |
+
import importlib
|
9 |
+
import math
|
10 |
+
from typing import TYPE_CHECKING, Optional, Tuple, Union, Callable, List, Any, Generator
|
11 |
+
import os, json
|
12 |
+
import numpy as np
|
13 |
+
import torch
|
14 |
+
import torch.nn.functional as F
|
15 |
+
import torch.utils.checkpoint
|
16 |
+
from torch.cuda.amp import autocast
|
17 |
+
|
18 |
+
from torch.nn import CrossEntropyLoss
|
19 |
+
from transformers import PreTrainedTokenizer, GenerationConfig, StoppingCriteriaList
|
20 |
+
from transformers.generation.logits_process import LogitsProcessorList
|
21 |
+
|
22 |
+
if TYPE_CHECKING:
|
23 |
+
from transformers.generation.streamers import BaseStreamer
|
24 |
+
from transformers.generation.utils import GenerateOutput
|
25 |
+
from transformers.modeling_outputs import (
|
26 |
+
BaseModelOutputWithPast,
|
27 |
+
CausalLMOutputWithPast,
|
28 |
+
)
|
29 |
+
from transformers.modeling_utils import PreTrainedModel
|
30 |
+
from transformers.utils import logging
|
31 |
+
|
32 |
+
try:
|
33 |
+
from einops import rearrange
|
34 |
+
except ImportError:
|
35 |
+
rearrange = None
|
36 |
+
from torch import nn
|
37 |
+
|
38 |
+
SUPPORT_CUDA = torch.cuda.is_available()
|
39 |
+
SUPPORT_BF16 = SUPPORT_CUDA and torch.cuda.is_bf16_supported()
|
40 |
+
SUPPORT_FP16 = SUPPORT_CUDA and torch.cuda.get_device_capability(0)[0] >= 7
|
41 |
+
|
42 |
+
from torch.nn.init import trunc_normal_
|
43 |
+
import sys
|
44 |
+
sys.path.append('../OdysseyAgent')
|
45 |
+
|
46 |
+
from configuration_qwen import QWenConfig
|
47 |
+
from qwen_generation_utils import (
|
48 |
+
HistoryType,
|
49 |
+
make_context,
|
50 |
+
decode_tokens,
|
51 |
+
get_stop_words_ids,
|
52 |
+
StopWordsLogitsProcessor,
|
53 |
+
)
|
54 |
+
from visual import VisionTransformer
|
55 |
+
|
56 |
+
IMAGE_HISTORY = '../data/his_index.json'
|
57 |
+
|
58 |
+
USE_RESAMPLER = True
|
59 |
+
|
60 |
+
print(IMAGE_HISTORY)
|
61 |
+
logger = logging.get_logger(__name__)
|
62 |
+
|
63 |
+
_CHECKPOINT_FOR_DOC = "qwen"
|
64 |
+
_CONFIG_FOR_DOC = "QWenConfig"
|
65 |
+
|
66 |
+
QWen_PRETRAINED_MODEL_ARCHIVE_LIST = ["qwen-7b"]
|
67 |
+
|
68 |
+
_ERROR_BAD_CHAT_FORMAT = """\
|
69 |
+
We detect you are probably using the pretrained model (rather than chat model) for chatting, since the chat_format in generation_config is not "chatml".
|
70 |
+
If you are directly using the model downloaded from Huggingface, please make sure you are using our "Qwen/Qwen-7B-Chat" Huggingface model (rather than "Qwen/Qwen-7B") when you call model.chat().
|
71 |
+
我们检测到您可能在使用预训练模型(而非chat模型)进行多轮chat,因为您当前在generation_config指定的chat_format,并未设置为我们在对话中所支持的"chatml"格式。
|
72 |
+
如果您在直接使用我们从Huggingface提供的模型,请确保您在调用model.chat()时,使用的是"Qwen/Qwen-7B-Chat"模型(而非"Qwen/Qwen-7B"预训练模型)。
|
73 |
+
"""
|
74 |
+
|
75 |
+
_SENTINEL = object()
|
76 |
+
_ERROR_STREAM_IN_CHAT = """\
|
77 |
+
Pass argument `stream` to model.chat() is buggy, deprecated, and marked for removal. Please use model.chat_stream(...) instead of model.chat(..., stream=True).
|
78 |
+
向model.chat()传入参数stream的用法可能存在Bug,该用法已被废弃,将在未来被移除。请使用model.chat_stream(...)代替model.chat(..., stream=True)。
|
79 |
+
"""
|
80 |
+
|
81 |
+
apply_rotary_emb_func = None
|
82 |
+
rms_norm = None
|
83 |
+
|
84 |
+
|
85 |
+
# Copied from transformers.models.bart.modeling_bart._make_causal_mask
|
86 |
+
def _make_causal_mask(
|
87 |
+
input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
|
88 |
+
):
|
89 |
+
"""
|
90 |
+
Make causal mask used for bi-directional self-attention.
|
91 |
+
"""
|
92 |
+
bsz, tgt_len = input_ids_shape
|
93 |
+
mask = torch.full((tgt_len, tgt_len), torch.finfo(dtype).min, device=device)
|
94 |
+
mask_cond = torch.arange(mask.size(-1), device=device)
|
95 |
+
mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
|
96 |
+
mask = mask.to(dtype)
|
97 |
+
|
98 |
+
if past_key_values_length > 0:
|
99 |
+
mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
|
100 |
+
return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)
|
101 |
+
|
102 |
+
|
103 |
+
# Copied from transformers.models.bart.modeling_bart._expand_mask
|
104 |
+
def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
|
105 |
+
"""
|
106 |
+
Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
|
107 |
+
"""
|
108 |
+
bsz, src_len = mask.size()
|
109 |
+
tgt_len = tgt_len if tgt_len is not None else src_len
|
110 |
+
|
111 |
+
expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
|
112 |
+
|
113 |
+
inverted_mask = 1.0 - expanded_mask
|
114 |
+
|
115 |
+
return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
|
116 |
+
|
117 |
+
def get_abs_pos(abs_pos, tgt_size):
|
118 |
+
# abs_pos: L, C
|
119 |
+
# tgt_size: M
|
120 |
+
# return: M, C
|
121 |
+
src_size = int(math.sqrt(abs_pos.size(0)))
|
122 |
+
tgt_size = int(math.sqrt(tgt_size))
|
123 |
+
dtype = abs_pos.dtype
|
124 |
+
|
125 |
+
if src_size != tgt_size:
|
126 |
+
return F.interpolate(
|
127 |
+
abs_pos.float().reshape(1, src_size, src_size, -1).permute(0, 3, 1, 2),
|
128 |
+
size=(tgt_size, tgt_size),
|
129 |
+
mode="bicubic",
|
130 |
+
align_corners=False,
|
131 |
+
).permute(0, 2, 3, 1).flatten(0, 2).to(dtype=dtype)
|
132 |
+
else:
|
133 |
+
return abs_pos
|
134 |
+
|
135 |
+
|
136 |
+
def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False):
|
137 |
+
"""
|
138 |
+
grid_size: int of the grid height and width
|
139 |
+
return:
|
140 |
+
pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
|
141 |
+
"""
|
142 |
+
grid_h = np.arange(grid_size, dtype=np.float32)
|
143 |
+
grid_w = np.arange(grid_size, dtype=np.float32)
|
144 |
+
grid = np.meshgrid(grid_w, grid_h) # here w goes first
|
145 |
+
grid = np.stack(grid, axis=0)
|
146 |
+
|
147 |
+
grid = grid.reshape([2, 1, grid_size, grid_size])
|
148 |
+
pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
|
149 |
+
if cls_token:
|
150 |
+
pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0)
|
151 |
+
return pos_embed
|
152 |
+
|
153 |
+
|
154 |
+
def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
|
155 |
+
assert embed_dim % 2 == 0
|
156 |
+
|
157 |
+
# use half of dimensions to encode grid_h
|
158 |
+
emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2)
|
159 |
+
emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2)
|
160 |
+
|
161 |
+
emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
|
162 |
+
return emb
|
163 |
+
|
164 |
+
|
165 |
+
def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
|
166 |
+
"""
|
167 |
+
embed_dim: output dimension for each position
|
168 |
+
pos: a list of positions to be encoded: size (M,)
|
169 |
+
out: (M, D)
|
170 |
+
"""
|
171 |
+
assert embed_dim % 2 == 0
|
172 |
+
omega = np.arange(embed_dim // 2, dtype=np.float32)
|
173 |
+
omega /= embed_dim / 2.
|
174 |
+
omega = 1. / 10000**omega # (D/2,)
|
175 |
+
|
176 |
+
pos = pos.reshape(-1) # (M,)
|
177 |
+
out = np.einsum('m,d->md', pos, omega) # (M, D/2), outer product
|
178 |
+
|
179 |
+
emb_sin = np.sin(out) # (M, D/2)
|
180 |
+
emb_cos = np.cos(out) # (M, D/2)
|
181 |
+
|
182 |
+
emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D)
|
183 |
+
return emb
|
184 |
+
|
185 |
+
|
186 |
+
|
187 |
+
class HisResampler(nn.Module):
|
188 |
+
"""
|
189 |
+
A 2D perceiver-resampler network with one cross attention layers by
|
190 |
+
(grid_size**2) learnable queries and 2d sincos pos_emb
|
191 |
+
Outputs:
|
192 |
+
A tensor with the shape of (grid_size**2, embed_dim)
|
193 |
+
"""
|
194 |
+
def __init__(
|
195 |
+
self,
|
196 |
+
embed_dim=4096,
|
197 |
+
num_heads=32,
|
198 |
+
grid_size=16,
|
199 |
+
kv_dim=None,
|
200 |
+
norm_layer=nn.LayerNorm
|
201 |
+
):
|
202 |
+
super().__init__()
|
203 |
+
self.num_queries = grid_size ** 2
|
204 |
+
self.embed_dim = embed_dim
|
205 |
+
self.num_heads = num_heads
|
206 |
+
|
207 |
+
self.pos_embed = nn.Parameter(
|
208 |
+
torch.from_numpy(get_2d_sincos_pos_embed(embed_dim, grid_size)).float()
|
209 |
+
).requires_grad_(False)
|
210 |
+
|
211 |
+
self.query = nn.Parameter(torch.zeros(self.num_queries, embed_dim))
|
212 |
+
trunc_normal_(self.query, std=.02)
|
213 |
+
|
214 |
+
if kv_dim is not None and kv_dim != embed_dim:
|
215 |
+
self.kv_proj = nn.Linear(kv_dim, embed_dim, bias=False)
|
216 |
+
else:
|
217 |
+
self.kv_proj = nn.Identity()
|
218 |
+
|
219 |
+
self.attn = nn.MultiheadAttention(embed_dim, num_heads)
|
220 |
+
self.ln_q = norm_layer(embed_dim)
|
221 |
+
self.ln_kv = norm_layer(embed_dim)
|
222 |
+
|
223 |
+
self.ln_post = norm_layer(embed_dim)
|
224 |
+
self.proj = nn.Parameter((embed_dim** -0.5) * torch.randn(embed_dim, embed_dim))
|
225 |
+
|
226 |
+
self.apply(self._init_weights)
|
227 |
+
|
228 |
+
def _init_weights(self, m):
|
229 |
+
if isinstance(m, nn.Linear):
|
230 |
+
trunc_normal_(m.weight, std=.02)
|
231 |
+
if isinstance(m, nn.Linear) and m.bias is not None:
|
232 |
+
nn.init.constant_(m.bias, 0)
|
233 |
+
elif isinstance(m, nn.LayerNorm):
|
234 |
+
nn.init.constant_(m.bias, 0)
|
235 |
+
nn.init.constant_(m.weight, 1.0)
|
236 |
+
|
237 |
+
def forward(self, x, attn_mask=None):
|
238 |
+
|
239 |
+
#pos_embed = get_abs_pos(self.pos_embed, x.size(1))
|
240 |
+
x = self.kv_proj(x)
|
241 |
+
x = self.ln_kv(x).permute(1, 0, 2)
|
242 |
+
|
243 |
+
N = x.shape[1]
|
244 |
+
q = self.ln_q(self.query)
|
245 |
+
out = self.attn(
|
246 |
+
self._repeat(q, N),# + self.pos_embed.unsqueeze(1),
|
247 |
+
x, # + pos_embed.unsqueeze(1),
|
248 |
+
x,
|
249 |
+
attn_mask=attn_mask)[0]
|
250 |
+
out = out.permute(1, 0, 2)
|
251 |
+
out = self.ln_post(out)
|
252 |
+
out = out @ self.proj
|
253 |
+
return out
|
254 |
+
|
255 |
+
def _repeat(self, query, N: int):
|
256 |
+
return query.unsqueeze(1).repeat(1, N, 1)
|
257 |
+
|
258 |
+
|
259 |
+
|
260 |
+
class QWenAttention(nn.Module):
|
261 |
+
def __init__(self, config):
|
262 |
+
super().__init__()
|
263 |
+
|
264 |
+
self.register_buffer("masked_bias", torch.tensor(-1e4), persistent=False)
|
265 |
+
self.seq_length = config.seq_length
|
266 |
+
|
267 |
+
self.hidden_size = config.hidden_size
|
268 |
+
self.split_size = config.hidden_size
|
269 |
+
self.num_heads = config.num_attention_heads
|
270 |
+
self.head_dim = self.hidden_size // self.num_heads
|
271 |
+
|
272 |
+
self.scale_attn_weights = True
|
273 |
+
|
274 |
+
self.projection_size = config.kv_channels * config.num_attention_heads
|
275 |
+
|
276 |
+
assert self.projection_size % config.num_attention_heads == 0
|
277 |
+
self.hidden_size_per_attention_head = (
|
278 |
+
self.projection_size // config.num_attention_heads
|
279 |
+
)
|
280 |
+
|
281 |
+
self.c_attn = nn.Linear(config.hidden_size, 3 * self.projection_size)
|
282 |
+
|
283 |
+
self.c_proj = nn.Linear(
|
284 |
+
config.hidden_size, self.projection_size, bias=not config.no_bias
|
285 |
+
)
|
286 |
+
|
287 |
+
self.is_fp32 = not (config.bf16 or config.fp16)
|
288 |
+
self.bf16 = config.bf16
|
289 |
+
|
290 |
+
self.use_dynamic_ntk = config.use_dynamic_ntk
|
291 |
+
self.use_logn_attn = config.use_logn_attn
|
292 |
+
|
293 |
+
logn_list = [
|
294 |
+
math.log(i, self.seq_length) if i > self.seq_length else 1
|
295 |
+
for i in range(1, 32768)
|
296 |
+
]
|
297 |
+
self.logn_tensor = torch.tensor(logn_list)[None, :, None, None]
|
298 |
+
|
299 |
+
self.attn_dropout = nn.Dropout(config.attn_dropout_prob)
|
300 |
+
|
301 |
+
def _attn(self, query, key, value, registered_causal_mask, attention_mask=None, head_mask=None):
|
302 |
+
attn_weights = torch.matmul(query, key.transpose(-1, -2))
|
303 |
+
|
304 |
+
if self.scale_attn_weights:
|
305 |
+
attn_weights = attn_weights / torch.full(
|
306 |
+
[],
|
307 |
+
value.size(-1) ** 0.5,
|
308 |
+
dtype=attn_weights.dtype,
|
309 |
+
device=attn_weights.device,
|
310 |
+
)
|
311 |
+
|
312 |
+
query_length, key_length = query.size(-2), key.size(-2)
|
313 |
+
# causal_mask = self.bias[
|
314 |
+
# :, :, key_length - query_length : key_length, :key_length
|
315 |
+
# ]
|
316 |
+
# mask_value = torch.finfo(attn_weights.dtype).min
|
317 |
+
# mask_value = torch.full([], mask_value, dtype=attn_weights.dtype).to(
|
318 |
+
# attn_weights.device
|
319 |
+
# )
|
320 |
+
# attn_weights = torch.where(
|
321 |
+
# causal_mask, attn_weights.to(attn_weights.dtype), mask_value
|
322 |
+
# )
|
323 |
+
attn_weights = attn_weights + attention_mask
|
324 |
+
|
325 |
+
attn_weights = nn.functional.softmax(attn_weights, dim=-1)
|
326 |
+
|
327 |
+
attn_weights = attn_weights.type(value.dtype)
|
328 |
+
attn_weights = self.attn_dropout(attn_weights)
|
329 |
+
|
330 |
+
if head_mask is not None:
|
331 |
+
attn_weights = attn_weights * head_mask
|
332 |
+
|
333 |
+
attn_output = torch.matmul(attn_weights, value)
|
334 |
+
attn_output = attn_output.transpose(1, 2)
|
335 |
+
|
336 |
+
return attn_output, attn_weights
|
337 |
+
|
338 |
+
def _upcast_and_reordered_attn(
|
339 |
+
self, query, key, value, registered_causal_mask, attention_mask=None, head_mask=None
|
340 |
+
):
|
341 |
+
bsz, num_heads, q_seq_len, dk = query.size()
|
342 |
+
_, _, k_seq_len, _ = key.size()
|
343 |
+
|
344 |
+
attn_weights = torch.empty(
|
345 |
+
bsz * num_heads,
|
346 |
+
q_seq_len,
|
347 |
+
k_seq_len,
|
348 |
+
dtype=torch.float32,
|
349 |
+
device=query.device,
|
350 |
+
)
|
351 |
+
|
352 |
+
scale_factor = 1.0
|
353 |
+
if self.scale_attn_weights:
|
354 |
+
scale_factor /= float(value.size(-1)) ** 0.5
|
355 |
+
|
356 |
+
with autocast(enabled=False):
|
357 |
+
q, k = query.reshape(-1, q_seq_len, dk), key.transpose(-1, -2).reshape(
|
358 |
+
-1, dk, k_seq_len
|
359 |
+
)
|
360 |
+
attn_weights = torch.baddbmm(
|
361 |
+
attn_weights, q.float(), k.float(), beta=0, alpha=scale_factor
|
362 |
+
)
|
363 |
+
attn_weights = attn_weights.reshape(bsz, num_heads, q_seq_len, k_seq_len)
|
364 |
+
|
365 |
+
query_length, key_length = query.size(-2), key.size(-2)
|
366 |
+
causal_mask = registered_causal_mask[
|
367 |
+
:, :, key_length - query_length : key_length, :key_length
|
368 |
+
]
|
369 |
+
mask_value = torch.finfo(attn_weights.dtype).min
|
370 |
+
mask_value = torch.tensor(mask_value, dtype=attn_weights.dtype).to(
|
371 |
+
attn_weights.device
|
372 |
+
)
|
373 |
+
attn_weights = torch.where(causal_mask, attn_weights, mask_value)
|
374 |
+
|
375 |
+
if attention_mask is not None:
|
376 |
+
attn_weights = attn_weights + attention_mask
|
377 |
+
|
378 |
+
attn_weights = nn.functional.softmax(attn_weights, dim=-1)
|
379 |
+
|
380 |
+
if attn_weights.dtype != torch.float32:
|
381 |
+
raise RuntimeError(
|
382 |
+
"Error with upcasting, attn_weights does not have dtype torch.float32"
|
383 |
+
)
|
384 |
+
attn_weights = attn_weights.type(value.dtype)
|
385 |
+
attn_weights = self.attn_dropout(attn_weights)
|
386 |
+
|
387 |
+
if head_mask is not None:
|
388 |
+
attn_weights = attn_weights * head_mask
|
389 |
+
|
390 |
+
attn_output = torch.matmul(attn_weights, value)
|
391 |
+
|
392 |
+
return attn_output, attn_weights
|
393 |
+
|
394 |
+
def _split_heads(self, tensor, num_heads, attn_head_size):
|
395 |
+
new_shape = tensor.size()[:-1] + (num_heads, attn_head_size)
|
396 |
+
tensor = tensor.view(new_shape)
|
397 |
+
return tensor
|
398 |
+
|
399 |
+
def _merge_heads(self, tensor, num_heads, attn_head_size):
|
400 |
+
tensor = tensor.contiguous()
|
401 |
+
new_shape = tensor.size()[:-2] + (num_heads * attn_head_size,)
|
402 |
+
return tensor.view(new_shape)
|
403 |
+
|
404 |
+
def forward(
|
405 |
+
self,
|
406 |
+
hidden_states: Optional[Tuple[torch.FloatTensor]],
|
407 |
+
rotary_pos_emb: Optional[List[torch.Tensor]] = None,
|
408 |
+
registered_causal_mask: Optional[torch.Tensor] = None,
|
409 |
+
layer_past: Optional[Tuple[torch.Tensor]] = None,
|
410 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
411 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
412 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
413 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
414 |
+
output_attentions: Optional[bool] = False,
|
415 |
+
use_cache: Optional[bool] = False,
|
416 |
+
):
|
417 |
+
|
418 |
+
mixed_x_layer = self.c_attn(hidden_states)
|
419 |
+
|
420 |
+
query, key, value = mixed_x_layer.split(self.split_size, dim=2)
|
421 |
+
|
422 |
+
query = self._split_heads(query, self.num_heads, self.head_dim)
|
423 |
+
key = self._split_heads(key, self.num_heads, self.head_dim)
|
424 |
+
value = self._split_heads(value, self.num_heads, self.head_dim)
|
425 |
+
|
426 |
+
if rotary_pos_emb is not None:
|
427 |
+
cur_len = query.shape[1]
|
428 |
+
rotary_pos_emb = [i[:, -cur_len:, :, :] for i in rotary_pos_emb]
|
429 |
+
rotary_pos_emb = (rotary_pos_emb,) * 2
|
430 |
+
q_pos_emb, k_pos_emb = rotary_pos_emb
|
431 |
+
# Slice the pos emb for current inference
|
432 |
+
query = apply_rotary_pos_emb(query, q_pos_emb)
|
433 |
+
key = apply_rotary_pos_emb(key, k_pos_emb)
|
434 |
+
|
435 |
+
if layer_past is not None:
|
436 |
+
past_key, past_value = layer_past[0], layer_past[1]
|
437 |
+
key = torch.cat((past_key, key), dim=1)
|
438 |
+
value = torch.cat((past_value, value), dim=1)
|
439 |
+
|
440 |
+
if use_cache:
|
441 |
+
present = (key, value)
|
442 |
+
else:
|
443 |
+
present = None
|
444 |
+
|
445 |
+
if self.use_logn_attn and not self.training:
|
446 |
+
if self.logn_tensor.device != query.device or self.logn_tensor.dtype != query.dtype:
|
447 |
+
self.logn_tensor = self.logn_tensor.to(query.device).type_as(query)
|
448 |
+
seq_start = key.size(1) - query.size(1)
|
449 |
+
seq_end = key.size(1)
|
450 |
+
logn_tensor = self.logn_tensor[:, seq_start:seq_end, :, :]
|
451 |
+
query = query * logn_tensor.expand_as(query)
|
452 |
+
|
453 |
+
query = query.permute(0, 2, 1, 3)
|
454 |
+
key = key.permute(0, 2, 1, 3)
|
455 |
+
value = value.permute(0, 2, 1, 3)
|
456 |
+
attn_output, attn_weight = self._attn(
|
457 |
+
query, key, value, registered_causal_mask, attention_mask, head_mask
|
458 |
+
)
|
459 |
+
context_layer = self._merge_heads(
|
460 |
+
attn_output, self.num_heads, self.head_dim
|
461 |
+
)
|
462 |
+
|
463 |
+
attn_output = self.c_proj(context_layer)
|
464 |
+
|
465 |
+
outputs = (attn_output, present)
|
466 |
+
if output_attentions:
|
467 |
+
outputs += (attn_weight,)
|
468 |
+
|
469 |
+
return outputs
|
470 |
+
|
471 |
+
|
472 |
+
class QWenMLP(nn.Module):
|
473 |
+
def __init__(self, config):
|
474 |
+
super().__init__()
|
475 |
+
self.w1 = nn.Linear(
|
476 |
+
config.hidden_size, config.intermediate_size // 2, bias=not config.no_bias
|
477 |
+
)
|
478 |
+
self.w2 = nn.Linear(
|
479 |
+
config.hidden_size, config.intermediate_size // 2, bias=not config.no_bias
|
480 |
+
)
|
481 |
+
ff_dim_in = config.intermediate_size // 2
|
482 |
+
self.c_proj = nn.Linear(ff_dim_in, config.hidden_size, bias=not config.no_bias)
|
483 |
+
|
484 |
+
def forward(self, hidden_states):
|
485 |
+
a1 = self.w1(hidden_states)
|
486 |
+
a2 = self.w2(hidden_states)
|
487 |
+
intermediate_parallel = a1 * F.silu(a2)
|
488 |
+
output = self.c_proj(intermediate_parallel)
|
489 |
+
return output
|
490 |
+
|
491 |
+
class QWenBlock(nn.Module):
|
492 |
+
def __init__(self, config):
|
493 |
+
super().__init__()
|
494 |
+
hidden_size = config.hidden_size
|
495 |
+
self.bf16 = config.bf16
|
496 |
+
|
497 |
+
self.ln_1 = RMSNorm(
|
498 |
+
hidden_size,
|
499 |
+
eps=config.layer_norm_epsilon,
|
500 |
+
)
|
501 |
+
self.attn = QWenAttention(config)
|
502 |
+
self.ln_2 = RMSNorm(
|
503 |
+
hidden_size,
|
504 |
+
eps=config.layer_norm_epsilon,
|
505 |
+
)
|
506 |
+
|
507 |
+
self.mlp = QWenMLP(config)
|
508 |
+
|
509 |
+
def forward(
|
510 |
+
self,
|
511 |
+
hidden_states: Optional[Tuple[torch.FloatTensor]],
|
512 |
+
rotary_pos_emb: Optional[List[torch.Tensor]] = None,
|
513 |
+
registered_causal_mask: Optional[torch.Tensor] = None,
|
514 |
+
layer_past: Optional[Tuple[torch.Tensor]] = None,
|
515 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
516 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
517 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
518 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
519 |
+
use_cache: Optional[bool] = False,
|
520 |
+
output_attentions: Optional[bool] = False,
|
521 |
+
):
|
522 |
+
layernorm_output = self.ln_1(hidden_states)
|
523 |
+
|
524 |
+
attn_outputs = self.attn(
|
525 |
+
layernorm_output,
|
526 |
+
rotary_pos_emb,
|
527 |
+
registered_causal_mask=registered_causal_mask,
|
528 |
+
layer_past=layer_past,
|
529 |
+
attention_mask=attention_mask,
|
530 |
+
head_mask=head_mask,
|
531 |
+
use_cache=use_cache,
|
532 |
+
output_attentions=output_attentions,
|
533 |
+
)
|
534 |
+
attn_output = attn_outputs[0]
|
535 |
+
|
536 |
+
outputs = attn_outputs[1:]
|
537 |
+
|
538 |
+
residual = hidden_states
|
539 |
+
layernorm_input = attn_output + residual
|
540 |
+
|
541 |
+
layernorm_output = self.ln_2(layernorm_input)
|
542 |
+
|
543 |
+
residual = layernorm_input
|
544 |
+
mlp_output = self.mlp(layernorm_output)
|
545 |
+
hidden_states = residual + mlp_output
|
546 |
+
|
547 |
+
if use_cache:
|
548 |
+
outputs = (hidden_states,) + outputs
|
549 |
+
else:
|
550 |
+
outputs = (hidden_states,) + outputs[1:]
|
551 |
+
|
552 |
+
return outputs
|
553 |
+
|
554 |
+
|
555 |
+
class QWenPreTrainedModel(PreTrainedModel):
|
556 |
+
config_class = QWenConfig
|
557 |
+
base_model_prefix = "transformer"
|
558 |
+
is_parallelizable = False
|
559 |
+
supports_gradient_checkpointing = True
|
560 |
+
_no_split_modules = ["QWenBlock"]
|
561 |
+
|
562 |
+
def __init__(self, *inputs, **kwargs):
|
563 |
+
super().__init__(*inputs, **kwargs)
|
564 |
+
|
565 |
+
def _init_weights(self, module):
|
566 |
+
"""Initialize the weights."""
|
567 |
+
if isinstance(module, nn.Linear):
|
568 |
+
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
569 |
+
if module.bias is not None:
|
570 |
+
module.bias.data.zero_()
|
571 |
+
elif isinstance(module, nn.Embedding):
|
572 |
+
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
573 |
+
if module.padding_idx is not None:
|
574 |
+
module.weight.data[module.padding_idx].zero_()
|
575 |
+
elif isinstance(module, RMSNorm):
|
576 |
+
module.weight.data.fill_(1.0)
|
577 |
+
|
578 |
+
for name, p in module.named_parameters():
|
579 |
+
if name == "c_proj.weight":
|
580 |
+
p.data.normal_(
|
581 |
+
mean=0.0,
|
582 |
+
std=(
|
583 |
+
self.config.initializer_range
|
584 |
+
/ math.sqrt(2 * self.config.num_hidden_layers)
|
585 |
+
),
|
586 |
+
)
|
587 |
+
|
588 |
+
def _set_gradient_checkpointing(self, module, value=False):
|
589 |
+
if isinstance(module, QWenModel):
|
590 |
+
module.gradient_checkpointing = value
|
591 |
+
|
592 |
+
|
593 |
+
class QWenModel(QWenPreTrainedModel):
|
594 |
+
_keys_to_ignore_on_load_missing = ["attn.masked_bias"]
|
595 |
+
|
596 |
+
def __init__(self, config):
|
597 |
+
super().__init__(config)
|
598 |
+
self.his_len = config.his_len
|
599 |
+
self.vocab_size = config.vocab_size
|
600 |
+
self.num_hidden_layers = config.num_hidden_layers
|
601 |
+
self.embed_dim = config.hidden_size
|
602 |
+
|
603 |
+
self.gradient_checkpointing = False
|
604 |
+
self.use_dynamic_ntk = config.use_dynamic_ntk
|
605 |
+
self.seq_length = config.seq_length
|
606 |
+
|
607 |
+
self.wte = nn.Embedding(self.vocab_size, self.embed_dim)
|
608 |
+
|
609 |
+
self.drop = nn.Dropout(config.emb_dropout_prob)
|
610 |
+
|
611 |
+
if config.rotary_pct == 1.0:
|
612 |
+
self.rotary_ndims = None
|
613 |
+
else:
|
614 |
+
assert config.rotary_pct < 1
|
615 |
+
self.rotary_ndims = int(
|
616 |
+
config.kv_channels * config.rotary_pct
|
617 |
+
)
|
618 |
+
dim = (
|
619 |
+
self.rotary_ndims
|
620 |
+
if self.rotary_ndims is not None
|
621 |
+
else config.kv_channels
|
622 |
+
)
|
623 |
+
self.rotary_emb = RotaryEmbedding(dim, base=config.rotary_emb_base)
|
624 |
+
|
625 |
+
self.use_flash_attn = config.use_flash_attn
|
626 |
+
self.is_fp32 = not (config.bf16 or config.fp16)
|
627 |
+
self.registered_causal_mask = None
|
628 |
+
|
629 |
+
self.h = nn.ModuleList(
|
630 |
+
[
|
631 |
+
QWenBlock(
|
632 |
+
config
|
633 |
+
)
|
634 |
+
for i in range(config.num_hidden_layers)
|
635 |
+
]
|
636 |
+
)
|
637 |
+
self.ln_f = RMSNorm(
|
638 |
+
self.embed_dim,
|
639 |
+
eps=config.layer_norm_epsilon,
|
640 |
+
)
|
641 |
+
|
642 |
+
self.visual = VisionTransformer(**config.visual)
|
643 |
+
|
644 |
+
self.post_init()
|
645 |
+
|
646 |
+
if USE_RESAMPLER:
|
647 |
+
print('init RESAMPLER')
|
648 |
+
self.his_resampler = HisResampler()
|
649 |
+
|
650 |
+
self.imgtoken_dict = {}
|
651 |
+
if os.path.isdir(IMAGE_HISTORY):
|
652 |
+
for subdata in os.listdir(IMAGE_HISTORY):
|
653 |
+
sub_img_dict = json.load(open(os.path.join(IMAGE_HISTORY, subdata)))
|
654 |
+
self.imgtoken_dict.update(sub_img_dict)
|
655 |
+
else:
|
656 |
+
self.imgtoken_dict = json.load(open(IMAGE_HISTORY))
|
657 |
+
|
658 |
+
print('imgtoken_dict cache len:', len(self.imgtoken_dict))
|
659 |
+
|
660 |
+
def set_input_embeddings(self, new_embeddings):
|
661 |
+
self.wte = new_embeddings
|
662 |
+
|
663 |
+
# Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
|
664 |
+
def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
|
665 |
+
# create causal mask
|
666 |
+
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
|
667 |
+
combined_attention_mask = None
|
668 |
+
if input_shape[-1] > 1:
|
669 |
+
combined_attention_mask = _make_causal_mask(
|
670 |
+
input_shape,
|
671 |
+
inputs_embeds.dtype,
|
672 |
+
device=inputs_embeds.device,
|
673 |
+
past_key_values_length=past_key_values_length,
|
674 |
+
)
|
675 |
+
|
676 |
+
if attention_mask is not None:
|
677 |
+
# [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
|
678 |
+
expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
|
679 |
+
inputs_embeds.device
|
680 |
+
)
|
681 |
+
combined_attention_mask = (
|
682 |
+
expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
|
683 |
+
)
|
684 |
+
|
685 |
+
return combined_attention_mask
|
686 |
+
|
687 |
+
|
688 |
+
def forward(
|
689 |
+
self,
|
690 |
+
input_ids: Optional[torch.LongTensor] = None,
|
691 |
+
past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
|
692 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
693 |
+
token_type_ids: Optional[torch.LongTensor] = None,
|
694 |
+
position_ids: Optional[torch.LongTensor] = None,
|
695 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
696 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
697 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
698 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
699 |
+
use_cache: Optional[bool] = None,
|
700 |
+
output_attentions: Optional[bool] = None,
|
701 |
+
output_hidden_states: Optional[bool] = None,
|
702 |
+
return_dict: Optional[bool] = None,
|
703 |
+
):
|
704 |
+
device = input_ids.device if input_ids is not None else inputs_embeds.device
|
705 |
+
|
706 |
+
if past_key_values is None and torch.any(input_ids == self.config.visual['image_start_id']):
|
707 |
+
bos_pos = torch.where(input_ids == self.config.visual['image_start_id'])
|
708 |
+
eos_pos = torch.where(input_ids == self.config.visual['image_start_id'] + 1)
|
709 |
+
assert (bos_pos[0] == eos_pos[0]).all()
|
710 |
+
img_pos = torch.stack((bos_pos[0], bos_pos[1], eos_pos[1]), dim=1)
|
711 |
+
now_images = []
|
712 |
+
his_images = []
|
713 |
+
C_list = []
|
714 |
+
images = []
|
715 |
+
his_idx = []
|
716 |
+
his_image_temp = []
|
717 |
+
for idx, (i, a, b) in enumerate(img_pos):
|
718 |
+
image = input_ids[i][a + 1 : b - 1].tolist()
|
719 |
+
image = image[ : image.index(self.config.visual['image_start_id'] + 2)]
|
720 |
+
image_path = bytes(image).decode('utf-8')
|
721 |
+
|
722 |
+
if image_path.startswith('image-history: '):
|
723 |
+
his_idx.append(idx)
|
724 |
+
image_path = image_path.replace('image-history: ', '')
|
725 |
+
his_list = self.imgtoken_dict[image_path][-self.his_len:] # t0 - tn-1
|
726 |
+
assert len(his_list) > 0, his_list
|
727 |
+
|
728 |
+
his_images.extend(his_list)
|
729 |
+
his_image_temp.append(his_list)
|
730 |
+
|
731 |
+
else:
|
732 |
+
now_images.append(image_path)
|
733 |
+
|
734 |
+
now_images = self.visual.encode(now_images)
|
735 |
+
|
736 |
+
if len(his_images) > 0:
|
737 |
+
his_images = self.visual.encode(his_images)
|
738 |
+
his_tkn = None
|
739 |
+
|
740 |
+
start_pos = 0
|
741 |
+
for his_scr in his_image_temp:
|
742 |
+
his_len = len(his_scr)
|
743 |
+
his_img_feature = his_images[start_pos: start_pos + his_len] # [b, l, d]
|
744 |
+
if USE_RESAMPLER:
|
745 |
+
his_img_feature = his_img_feature.reshape(1, -1, his_img_feature.size(-1))
|
746 |
+
his_vis_tkn = self.his_resampler(his_img_feature) # [l, d]
|
747 |
+
else:
|
748 |
+
raise ValueError("You cannot run without History Redsampler!")
|
749 |
+
his_tkn = his_vis_tkn if his_tkn is None else torch.concat((his_tkn, his_vis_tkn), dim=0)
|
750 |
+
start_pos += his_len
|
751 |
+
assert start_pos == len(his_images)
|
752 |
+
his_images = his_tkn
|
753 |
+
|
754 |
+
now_p, his_p = 0, 0
|
755 |
+
for j in range(len(img_pos)):
|
756 |
+
if j not in his_idx:
|
757 |
+
images.append(now_images[now_p])
|
758 |
+
now_p += 1
|
759 |
+
else:
|
760 |
+
images.append(his_images[his_p])
|
761 |
+
his_p += 1
|
762 |
+
images = torch.stack(images, dim=0)
|
763 |
+
assert len(images) == len(img_pos) == len(now_images) + len(his_images)
|
764 |
+
|
765 |
+
fake_images = None
|
766 |
+
elif self.training:
|
767 |
+
fake_images=torch.zeros(1,3,224,224).to(
|
768 |
+
dtype=self.visual.conv1.weight.dtype, device=self.visual.conv1.weight.device)
|
769 |
+
images = self.visual(fake_images)
|
770 |
+
else:
|
771 |
+
fake_images = None
|
772 |
+
images = None
|
773 |
+
|
774 |
+
output_attentions = (
|
775 |
+
output_attentions
|
776 |
+
if output_attentions is not None
|
777 |
+
else self.config.output_attentions
|
778 |
+
)
|
779 |
+
output_hidden_states = (
|
780 |
+
output_hidden_states
|
781 |
+
if output_hidden_states is not None
|
782 |
+
else self.config.output_hidden_states
|
783 |
+
)
|
784 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
785 |
+
return_dict = (
|
786 |
+
return_dict if return_dict is not None else self.config.use_return_dict
|
787 |
+
)
|
788 |
+
|
789 |
+
if input_ids is not None and inputs_embeds is not None:
|
790 |
+
raise ValueError(
|
791 |
+
"You cannot specify both input_ids and inputs_embeds at the same time"
|
792 |
+
)
|
793 |
+
elif input_ids is not None:
|
794 |
+
input_shape = input_ids.size()
|
795 |
+
input_ids = input_ids.view(-1, input_shape[-1])
|
796 |
+
batch_size = input_ids.shape[0]
|
797 |
+
elif inputs_embeds is not None:
|
798 |
+
input_shape = inputs_embeds.size()[:-1]
|
799 |
+
batch_size = inputs_embeds.shape[0]
|
800 |
+
else:
|
801 |
+
raise ValueError("You have to specify either input_ids or inputs_embeds")
|
802 |
+
|
803 |
+
|
804 |
+
if token_type_ids is not None:
|
805 |
+
token_type_ids = token_type_ids.view(-1, input_shape[-1])
|
806 |
+
if position_ids is not None:
|
807 |
+
position_ids = position_ids.view(-1, input_shape[-1])
|
808 |
+
|
809 |
+
if past_key_values is None:
|
810 |
+
past_length = 0
|
811 |
+
past_key_values = tuple([None] * len(self.h))
|
812 |
+
else:
|
813 |
+
past_length = past_key_values[0][0].size(-2)
|
814 |
+
|
815 |
+
if position_ids is None:
|
816 |
+
position_ids = torch.arange(
|
817 |
+
past_length,
|
818 |
+
input_shape[-1] + past_length,
|
819 |
+
dtype=torch.long,
|
820 |
+
device=device,
|
821 |
+
)
|
822 |
+
position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
|
823 |
+
|
824 |
+
encoder_attention_mask = None
|
825 |
+
head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
|
826 |
+
|
827 |
+
if inputs_embeds is None:
|
828 |
+
inputs_embeds = self.wte(input_ids)
|
829 |
+
|
830 |
+
if batch_size <= 0:
|
831 |
+
raise ValueError("batch_size has to be defined and > 0")
|
832 |
+
attention_mask = self._prepare_decoder_attention_mask(
|
833 |
+
attention_mask, input_shape, inputs_embeds, past_length
|
834 |
+
)
|
835 |
+
|
836 |
+
hidden_states = inputs_embeds
|
837 |
+
|
838 |
+
kv_seq_len = hidden_states.size()[1]
|
839 |
+
if past_key_values[0] is not None:
|
840 |
+
# past key values[0][0] shape: bs * seq_len * head_num * dim
|
841 |
+
kv_seq_len += past_key_values[0][0].shape[1]
|
842 |
+
if (
|
843 |
+
self.use_dynamic_ntk
|
844 |
+
and kv_seq_len == hidden_states.size()[1]
|
845 |
+
and not self.training
|
846 |
+
):
|
847 |
+
context_value = math.log(kv_seq_len / self.seq_length, 2) + 1
|
848 |
+
ntk_alpha = 2 ** math.ceil(context_value) - 1
|
849 |
+
ntk_alpha = max(ntk_alpha, 1)
|
850 |
+
else:
|
851 |
+
ntk_alpha = self.rotary_emb._ntk_alpha_cached
|
852 |
+
|
853 |
+
rotary_pos_emb = self.rotary_emb(kv_seq_len, ntk_alpha=ntk_alpha)
|
854 |
+
for idx in range(len(rotary_pos_emb)):
|
855 |
+
rotary_pos_emb[idx] = rotary_pos_emb[idx].to(hidden_states.device)
|
856 |
+
hidden_states = self.drop(hidden_states).clone()
|
857 |
+
|
858 |
+
if fake_images is not None:
|
859 |
+
hidden_states = hidden_states + images.mean()*0
|
860 |
+
elif images is not None:
|
861 |
+
for idx, (i, a, b) in enumerate(img_pos):
|
862 |
+
hidden_states[i][a + 1 : b] = images[idx]
|
863 |
+
output_shape = input_shape + (hidden_states.size(-1),)
|
864 |
+
|
865 |
+
if self.gradient_checkpointing and self.training:
|
866 |
+
if use_cache:
|
867 |
+
logger.warning_once(
|
868 |
+
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
|
869 |
+
)
|
870 |
+
use_cache = False
|
871 |
+
|
872 |
+
presents = () if use_cache else None
|
873 |
+
all_self_attentions = () if output_attentions else None
|
874 |
+
all_hidden_states = () if output_hidden_states else None
|
875 |
+
for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
|
876 |
+
|
877 |
+
if output_hidden_states:
|
878 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
879 |
+
|
880 |
+
if self.gradient_checkpointing and self.training:
|
881 |
+
|
882 |
+
def create_custom_forward(module):
|
883 |
+
def custom_forward(*inputs):
|
884 |
+
# None for past_key_value
|
885 |
+
return module(*inputs, use_cache, output_attentions)
|
886 |
+
|
887 |
+
return custom_forward
|
888 |
+
|
889 |
+
outputs = torch.utils.checkpoint.checkpoint(
|
890 |
+
create_custom_forward(block),
|
891 |
+
hidden_states,
|
892 |
+
rotary_pos_emb,
|
893 |
+
self.registered_causal_mask,
|
894 |
+
None,
|
895 |
+
attention_mask,
|
896 |
+
head_mask[i],
|
897 |
+
encoder_hidden_states,
|
898 |
+
encoder_attention_mask,
|
899 |
+
)
|
900 |
+
else:
|
901 |
+
outputs = block(
|
902 |
+
hidden_states,
|
903 |
+
layer_past=layer_past,
|
904 |
+
rotary_pos_emb=rotary_pos_emb,
|
905 |
+
registered_causal_mask=self.registered_causal_mask,
|
906 |
+
attention_mask=attention_mask,
|
907 |
+
head_mask=head_mask[i],
|
908 |
+
encoder_hidden_states=encoder_hidden_states,
|
909 |
+
encoder_attention_mask=encoder_attention_mask,
|
910 |
+
use_cache=use_cache,
|
911 |
+
output_attentions=output_attentions,
|
912 |
+
)
|
913 |
+
|
914 |
+
hidden_states = outputs[0]
|
915 |
+
if use_cache is True:
|
916 |
+
presents = presents + (outputs[1],)
|
917 |
+
|
918 |
+
if output_attentions:
|
919 |
+
all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
|
920 |
+
|
921 |
+
hidden_states = self.ln_f(hidden_states)
|
922 |
+
hidden_states = hidden_states.view(output_shape)
|
923 |
+
# Add last hidden state
|
924 |
+
if output_hidden_states:
|
925 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
926 |
+
|
927 |
+
if not return_dict:
|
928 |
+
return tuple(
|
929 |
+
v for v in [hidden_states, presents, all_hidden_states] if v is not None
|
930 |
+
)
|
931 |
+
|
932 |
+
return BaseModelOutputWithPast(
|
933 |
+
last_hidden_state=hidden_states,
|
934 |
+
past_key_values=presents,
|
935 |
+
hidden_states=all_hidden_states,
|
936 |
+
attentions=all_self_attentions,
|
937 |
+
)
|
938 |
+
|
939 |
+
|
940 |
+
class QWenLMHeadModel(QWenPreTrainedModel):
|
941 |
+
_keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.rotary_emb\.inv_freq"]
|
942 |
+
_keys_to_ignore_on_load_unexpected = [r"h\.\d+\.attn\.masked_bias"]
|
943 |
+
|
944 |
+
def __init__(self, config):
|
945 |
+
super().__init__(config)
|
946 |
+
assert (
|
947 |
+
config.bf16 + config.fp16 + config.fp32 <= 1
|
948 |
+
), "Only one of \"bf16\", \"fp16\", \"fp32\" can be true"
|
949 |
+
|
950 |
+
autoset_precision = config.bf16 + config.fp16 + config.fp32 == 0
|
951 |
+
|
952 |
+
if autoset_precision:
|
953 |
+
if SUPPORT_BF16:
|
954 |
+
logger.warn(
|
955 |
+
"The model is automatically converting to bf16 for faster inference. "
|
956 |
+
"If you want to disable the automatic precision, please manually add bf16/fp16/fp32=True to \"AutoModelForCausalLM.from_pretrained\"."
|
957 |
+
)
|
958 |
+
config.bf16 = True
|
959 |
+
elif SUPPORT_FP16:
|
960 |
+
logger.warn(
|
961 |
+
"The model is automatically converting to fp16 for faster inference. "
|
962 |
+
"If you want to disable the automatic precision, please manually add bf16/fp16/fp32=True to \"AutoModelForCausalLM.from_pretrained\"."
|
963 |
+
)
|
964 |
+
config.fp16 = True
|
965 |
+
else:
|
966 |
+
config.fp32 = True
|
967 |
+
|
968 |
+
if config.bf16 and SUPPORT_CUDA and not SUPPORT_BF16:
|
969 |
+
logger.warn("Your device does NOT seem to support bf16, you can switch to fp16 or fp32 by by passing fp16/fp32=True in \"AutoModelForCausalLM.from_pretrained\".")
|
970 |
+
if config.fp16 and SUPPORT_CUDA and not SUPPORT_FP16:
|
971 |
+
logger.warn("Your device does NOT support faster inference with fp16, please switch to fp32 which is likely to be faster")
|
972 |
+
if config.fp32:
|
973 |
+
if SUPPORT_BF16:
|
974 |
+
logger.warn("Your device support faster inference by passing bf16=True in \"AutoModelForCausalLM.from_pretrained\".")
|
975 |
+
elif SUPPORT_FP16:
|
976 |
+
logger.warn("Your device support faster inference by passing fp16=True in \"AutoModelForCausalLM.from_pretrained\".")
|
977 |
+
|
978 |
+
self.transformer = QWenModel(config)
|
979 |
+
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
980 |
+
|
981 |
+
if config.bf16:
|
982 |
+
self.transformer.bfloat16()
|
983 |
+
self.lm_head.bfloat16()
|
984 |
+
if config.fp16:
|
985 |
+
self.transformer.half()
|
986 |
+
self.lm_head.half()
|
987 |
+
self.post_init()
|
988 |
+
|
989 |
+
def get_output_embeddings(self):
|
990 |
+
return self.lm_head
|
991 |
+
|
992 |
+
def set_output_embeddings(self, new_embeddings):
|
993 |
+
self.lm_head = new_embeddings
|
994 |
+
|
995 |
+
def prepare_inputs_for_generation(
|
996 |
+
self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs
|
997 |
+
):
|
998 |
+
token_type_ids = kwargs.get("token_type_ids", None)
|
999 |
+
if past_key_values:
|
1000 |
+
input_ids = input_ids[:, -1].unsqueeze(-1)
|
1001 |
+
if token_type_ids is not None:
|
1002 |
+
token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
|
1003 |
+
|
1004 |
+
attention_mask = kwargs.get("attention_mask", None)
|
1005 |
+
position_ids = kwargs.get("position_ids", None)
|
1006 |
+
|
1007 |
+
if attention_mask is not None and position_ids is None:
|
1008 |
+
position_ids = attention_mask.long().cumsum(-1) - 1
|
1009 |
+
position_ids.masked_fill_(attention_mask == 0, 1)
|
1010 |
+
if past_key_values:
|
1011 |
+
position_ids = position_ids[:, -1].unsqueeze(-1)
|
1012 |
+
else:
|
1013 |
+
position_ids = None
|
1014 |
+
|
1015 |
+
if inputs_embeds is not None and past_key_values is None:
|
1016 |
+
model_inputs = {"inputs_embeds": inputs_embeds}
|
1017 |
+
else:
|
1018 |
+
model_inputs = {"input_ids": input_ids}
|
1019 |
+
|
1020 |
+
model_inputs.update(
|
1021 |
+
{
|
1022 |
+
"past_key_values": past_key_values,
|
1023 |
+
"use_cache": kwargs.get("use_cache"),
|
1024 |
+
"position_ids": position_ids,
|
1025 |
+
"attention_mask": attention_mask,
|
1026 |
+
"token_type_ids": token_type_ids,
|
1027 |
+
}
|
1028 |
+
)
|
1029 |
+
return model_inputs
|
1030 |
+
|
1031 |
+
def forward(
|
1032 |
+
self,
|
1033 |
+
input_ids: Optional[torch.LongTensor] = None,
|
1034 |
+
past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
|
1035 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
1036 |
+
token_type_ids: Optional[torch.LongTensor] = None,
|
1037 |
+
position_ids: Optional[torch.LongTensor] = None,
|
1038 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
1039 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
1040 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
1041 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
1042 |
+
labels: Optional[torch.LongTensor] = None,
|
1043 |
+
use_cache: Optional[bool] = None,
|
1044 |
+
output_attentions: Optional[bool] = None,
|
1045 |
+
output_hidden_states: Optional[bool] = None,
|
1046 |
+
return_dict: Optional[bool] = None,
|
1047 |
+
) -> Union[Tuple, CausalLMOutputWithPast]:
|
1048 |
+
|
1049 |
+
return_dict = (
|
1050 |
+
return_dict if return_dict is not None else self.config.use_return_dict
|
1051 |
+
)
|
1052 |
+
|
1053 |
+
transformer_outputs = self.transformer(
|
1054 |
+
input_ids,
|
1055 |
+
past_key_values=past_key_values,
|
1056 |
+
attention_mask=attention_mask,
|
1057 |
+
token_type_ids=token_type_ids,
|
1058 |
+
position_ids=position_ids,
|
1059 |
+
head_mask=head_mask,
|
1060 |
+
inputs_embeds=inputs_embeds,
|
1061 |
+
encoder_hidden_states=encoder_hidden_states,
|
1062 |
+
encoder_attention_mask=encoder_attention_mask,
|
1063 |
+
use_cache=use_cache,
|
1064 |
+
output_attentions=output_attentions,
|
1065 |
+
output_hidden_states=output_hidden_states,
|
1066 |
+
return_dict=return_dict,
|
1067 |
+
)
|
1068 |
+
hidden_states = transformer_outputs[0]
|
1069 |
+
|
1070 |
+
lm_logits = self.lm_head(hidden_states)
|
1071 |
+
|
1072 |
+
loss = None
|
1073 |
+
if labels is not None:
|
1074 |
+
labels = labels.to(lm_logits.device)
|
1075 |
+
shift_logits = lm_logits[..., :-1, :].contiguous()
|
1076 |
+
shift_labels = labels[..., 1:].contiguous()
|
1077 |
+
loss_fct = CrossEntropyLoss()
|
1078 |
+
loss = loss_fct(
|
1079 |
+
shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)
|
1080 |
+
)
|
1081 |
+
|
1082 |
+
if not return_dict:
|
1083 |
+
output = (lm_logits,) + transformer_outputs[1:]
|
1084 |
+
return ((loss,) + output) if loss is not None else output
|
1085 |
+
|
1086 |
+
return CausalLMOutputWithPast(
|
1087 |
+
loss=loss,
|
1088 |
+
logits=lm_logits,
|
1089 |
+
past_key_values=transformer_outputs.past_key_values,
|
1090 |
+
hidden_states=transformer_outputs.hidden_states,
|
1091 |
+
attentions=transformer_outputs.attentions,
|
1092 |
+
)
|
1093 |
+
|
1094 |
+
@staticmethod
|
1095 |
+
def _reorder_cache(
|
1096 |
+
past_key_values: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor
|
1097 |
+
) -> Tuple[Tuple[torch.Tensor]]:
|
1098 |
+
|
1099 |
+
return tuple(
|
1100 |
+
tuple(
|
1101 |
+
past_state.index_select(0, beam_idx.to(past_state.device))
|
1102 |
+
for past_state in layer_past
|
1103 |
+
)
|
1104 |
+
for layer_past in past_key_values
|
1105 |
+
)
|
1106 |
+
|
1107 |
+
def chat(
|
1108 |
+
self,
|
1109 |
+
tokenizer: PreTrainedTokenizer,
|
1110 |
+
query: str,
|
1111 |
+
history: Optional[HistoryType],
|
1112 |
+
system: str = "You are a helpful assistant.",
|
1113 |
+
append_history: bool = True,
|
1114 |
+
stream: Optional[bool] = _SENTINEL,
|
1115 |
+
stop_words_ids: Optional[List[List[int]]] = None,
|
1116 |
+
generation_config: Optional[GenerationConfig] = None,
|
1117 |
+
**kwargs,
|
1118 |
+
) -> Tuple[str, HistoryType]:
|
1119 |
+
generation_config = generation_config if generation_config is not None else self.generation_config
|
1120 |
+
|
1121 |
+
assert stream is _SENTINEL, _ERROR_STREAM_IN_CHAT
|
1122 |
+
assert generation_config.chat_format == 'chatml', _ERROR_BAD_CHAT_FORMAT
|
1123 |
+
if history is None:
|
1124 |
+
history = []
|
1125 |
+
if stop_words_ids is None:
|
1126 |
+
stop_words_ids = []
|
1127 |
+
|
1128 |
+
max_window_size = kwargs.get('max_window_size', None)
|
1129 |
+
if max_window_size is None:
|
1130 |
+
max_window_size = generation_config.max_window_size
|
1131 |
+
raw_text, context_tokens = make_context(
|
1132 |
+
tokenizer,
|
1133 |
+
query,
|
1134 |
+
history=history,
|
1135 |
+
system=system,
|
1136 |
+
max_window_size=max_window_size,
|
1137 |
+
chat_format=generation_config.chat_format,
|
1138 |
+
)
|
1139 |
+
|
1140 |
+
stop_words_ids.extend(get_stop_words_ids(
|
1141 |
+
generation_config.chat_format, tokenizer
|
1142 |
+
))
|
1143 |
+
input_ids = torch.tensor([context_tokens]).to(self.device)
|
1144 |
+
outputs = self.generate(
|
1145 |
+
input_ids,
|
1146 |
+
stop_words_ids=stop_words_ids,
|
1147 |
+
return_dict_in_generate=False,
|
1148 |
+
generation_config=generation_config,
|
1149 |
+
**kwargs,
|
1150 |
+
)
|
1151 |
+
|
1152 |
+
response = decode_tokens(
|
1153 |
+
outputs[0],
|
1154 |
+
tokenizer,
|
1155 |
+
raw_text_len=len(raw_text),
|
1156 |
+
context_length=len(context_tokens),
|
1157 |
+
chat_format=generation_config.chat_format,
|
1158 |
+
verbose=False,
|
1159 |
+
errors='replace'
|
1160 |
+
)
|
1161 |
+
|
1162 |
+
if append_history:
|
1163 |
+
history.append((query, response))
|
1164 |
+
|
1165 |
+
return response, history
|
1166 |
+
|
1167 |
+
def chat_stream(
|
1168 |
+
self,
|
1169 |
+
tokenizer: PreTrainedTokenizer,
|
1170 |
+
query: str,
|
1171 |
+
history: Optional[HistoryType],
|
1172 |
+
system: str = "You are a helpful assistant.",
|
1173 |
+
stop_words_ids: Optional[List[List[int]]] = None,
|
1174 |
+
logits_processor: Optional[LogitsProcessorList] = None,
|
1175 |
+
generation_config: Optional[GenerationConfig] = None,
|
1176 |
+
**kwargs,
|
1177 |
+
) -> Generator[str, Any, None]:
|
1178 |
+
generation_config = generation_config if generation_config is not None else self.generation_config
|
1179 |
+
assert generation_config.chat_format == 'chatml', _ERROR_BAD_CHAT_FORMAT
|
1180 |
+
if history is None:
|
1181 |
+
history = []
|
1182 |
+
if stop_words_ids is None:
|
1183 |
+
stop_words_ids = []
|
1184 |
+
|
1185 |
+
max_window_size = kwargs.get('max_window_size', None)
|
1186 |
+
if max_window_size is None:
|
1187 |
+
max_window_size = generation_config.max_window_size
|
1188 |
+
raw_text, context_tokens = make_context(
|
1189 |
+
tokenizer,
|
1190 |
+
query,
|
1191 |
+
history=history,
|
1192 |
+
system=system,
|
1193 |
+
max_window_size=max_window_size,
|
1194 |
+
chat_format=generation_config.chat_format,
|
1195 |
+
)
|
1196 |
+
|
1197 |
+
stop_words_ids.extend(get_stop_words_ids(
|
1198 |
+
generation_config.chat_format, tokenizer
|
1199 |
+
))
|
1200 |
+
if stop_words_ids is not None:
|
1201 |
+
stop_words_logits_processor = StopWordsLogitsProcessor(
|
1202 |
+
stop_words_ids=stop_words_ids,
|
1203 |
+
eos_token_id=generation_config.eos_token_id,
|
1204 |
+
)
|
1205 |
+
if logits_processor is None:
|
1206 |
+
logits_processor = LogitsProcessorList([stop_words_logits_processor])
|
1207 |
+
else:
|
1208 |
+
logits_processor.append(stop_words_logits_processor)
|
1209 |
+
input_ids = torch.tensor([context_tokens]).to(self.device)
|
1210 |
+
|
1211 |
+
from transformers_stream_generator.main import NewGenerationMixin, StreamGenerationConfig
|
1212 |
+
self.__class__.generate_stream = NewGenerationMixin.generate
|
1213 |
+
self.__class__.sample_stream = NewGenerationMixin.sample_stream
|
1214 |
+
stream_config = StreamGenerationConfig(**generation_config.to_dict(), do_stream=True)
|
1215 |
+
|
1216 |
+
def stream_generator():
|
1217 |
+
outputs = []
|
1218 |
+
for token in self.generate_stream(
|
1219 |
+
input_ids,
|
1220 |
+
return_dict_in_generate=False,
|
1221 |
+
generation_config=stream_config,
|
1222 |
+
logits_processor=logits_processor,
|
1223 |
+
seed=-1,
|
1224 |
+
**kwargs):
|
1225 |
+
outputs.append(token.item())
|
1226 |
+
yield tokenizer.decode(outputs, skip_special_tokens=True, errors='ignore', keep_image_special=True)
|
1227 |
+
|
1228 |
+
return stream_generator()
|
1229 |
+
|
1230 |
+
def generate(
|
1231 |
+
self,
|
1232 |
+
inputs: Optional[torch.Tensor] = None,
|
1233 |
+
generation_config: Optional[GenerationConfig] = None,
|
1234 |
+
logits_processor: Optional[LogitsProcessorList] = None,
|
1235 |
+
stopping_criteria: Optional[StoppingCriteriaList] = None,
|
1236 |
+
prefix_allowed_tokens_fn: Optional[
|
1237 |
+
Callable[[int, torch.Tensor], List[int]]
|
1238 |
+
] = None,
|
1239 |
+
synced_gpus: Optional[bool] = None,
|
1240 |
+
assistant_model: Optional["PreTrainedModel"] = None,
|
1241 |
+
streamer: Optional["BaseStreamer"] = None,
|
1242 |
+
**kwargs,
|
1243 |
+
) -> Union[GenerateOutput, torch.LongTensor]:
|
1244 |
+
generation_config = generation_config if generation_config is not None else self.generation_config
|
1245 |
+
|
1246 |
+
# Process stop_words_ids.
|
1247 |
+
stop_words_ids = kwargs.pop("stop_words_ids", None)
|
1248 |
+
if stop_words_ids is None and generation_config is not None:
|
1249 |
+
stop_words_ids = getattr(generation_config, "stop_words_ids", None)
|
1250 |
+
if stop_words_ids is None:
|
1251 |
+
stop_words_ids = getattr(generation_config, "stop_words_ids", None)
|
1252 |
+
|
1253 |
+
if stop_words_ids is not None:
|
1254 |
+
stop_words_logits_processor = StopWordsLogitsProcessor(
|
1255 |
+
stop_words_ids=stop_words_ids,
|
1256 |
+
eos_token_id=generation_config.eos_token_id,
|
1257 |
+
)
|
1258 |
+
if logits_processor is None:
|
1259 |
+
logits_processor = LogitsProcessorList([stop_words_logits_processor])
|
1260 |
+
else:
|
1261 |
+
logits_processor.append(stop_words_logits_processor)
|
1262 |
+
|
1263 |
+
return super().generate(
|
1264 |
+
inputs,
|
1265 |
+
generation_config=generation_config,
|
1266 |
+
logits_processor=logits_processor,
|
1267 |
+
stopping_criteria=stopping_criteria,
|
1268 |
+
prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
|
1269 |
+
synced_gpus=synced_gpus,
|
1270 |
+
assistant_model=assistant_model,
|
1271 |
+
streamer=streamer,
|
1272 |
+
**kwargs,
|
1273 |
+
)
|
1274 |
+
|
1275 |
+
|
1276 |
+
class RotaryEmbedding(torch.nn.Module):
|
1277 |
+
def __init__(self, dim, base=10000):
|
1278 |
+
super().__init__()
|
1279 |
+
self.dim = dim
|
1280 |
+
self.base = base
|
1281 |
+
self.inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
|
1282 |
+
if importlib.util.find_spec("einops") is None:
|
1283 |
+
raise RuntimeError("einops is required for Rotary Embedding")
|
1284 |
+
|
1285 |
+
self._rotary_pos_emb_cache = None
|
1286 |
+
self._seq_len_cached = 0
|
1287 |
+
self._ntk_alpha_cached = 1.0
|
1288 |
+
|
1289 |
+
def update_rotary_pos_emb_cache(self, max_seq_len, offset=0, ntk_alpha=1.0):
|
1290 |
+
seqlen = max_seq_len + offset
|
1291 |
+
if seqlen > self._seq_len_cached or ntk_alpha != self._ntk_alpha_cached:
|
1292 |
+
base = self.base * ntk_alpha ** (self.dim / (self.dim - 2))
|
1293 |
+
self.inv_freq = 1.0 / (
|
1294 |
+
base
|
1295 |
+
** (
|
1296 |
+
torch.arange(0, self.dim, 2, device=self.inv_freq.device).float()
|
1297 |
+
/ self.dim
|
1298 |
+
)
|
1299 |
+
)
|
1300 |
+
self._seq_len_cached = max(2 * seqlen, 16)
|
1301 |
+
self._ntk_alpha_cached = ntk_alpha
|
1302 |
+
seq = torch.arange(self._seq_len_cached, device=self.inv_freq.device)
|
1303 |
+
freqs = torch.outer(seq.type_as(self.inv_freq), self.inv_freq)
|
1304 |
+
|
1305 |
+
emb = torch.cat((freqs, freqs), dim=-1)
|
1306 |
+
from einops import rearrange
|
1307 |
+
|
1308 |
+
emb = rearrange(emb, "n d -> 1 n 1 d")
|
1309 |
+
|
1310 |
+
cos, sin = emb.cos(), emb.sin()
|
1311 |
+
self._rotary_pos_emb_cache = [cos, sin]
|
1312 |
+
|
1313 |
+
def forward(self, max_seq_len, offset=0, ntk_alpha=1.0):
|
1314 |
+
self.update_rotary_pos_emb_cache(max_seq_len, offset, ntk_alpha)
|
1315 |
+
cos, sin = self._rotary_pos_emb_cache
|
1316 |
+
return [cos[:, offset : offset + max_seq_len], sin[:, offset : offset + max_seq_len]]
|
1317 |
+
|
1318 |
+
|
1319 |
+
def _rotate_half(x):
|
1320 |
+
from einops import rearrange
|
1321 |
+
|
1322 |
+
x = rearrange(x, "... (j d) -> ... j d", j=2)
|
1323 |
+
x1, x2 = x.unbind(dim=-2)
|
1324 |
+
return torch.cat((-x2, x1), dim=-1)
|
1325 |
+
|
1326 |
+
|
1327 |
+
def apply_rotary_pos_emb(t, freqs):
|
1328 |
+
cos, sin = freqs
|
1329 |
+
if apply_rotary_emb_func is not None and t.is_cuda:
|
1330 |
+
t_ = t.float()
|
1331 |
+
cos = cos.squeeze(0).squeeze(1)[:, : cos.shape[-1] // 2]
|
1332 |
+
sin = sin.squeeze(0).squeeze(1)[:, : sin.shape[-1] // 2]
|
1333 |
+
output = apply_rotary_emb_func(t_, cos, sin).type_as(t)
|
1334 |
+
return output
|
1335 |
+
else:
|
1336 |
+
rot_dim = freqs[0].shape[-1]
|
1337 |
+
cos, sin = freqs
|
1338 |
+
t_, t_pass_ = t[..., :rot_dim], t[..., rot_dim:]
|
1339 |
+
t_ = t_.float()
|
1340 |
+
t_pass_ = t_pass_.float()
|
1341 |
+
t_ = (t_ * cos) + (_rotate_half(t_) * sin)
|
1342 |
+
return torch.cat((t_, t_pass_), dim=-1).type_as(t)
|
1343 |
+
|
1344 |
+
|
1345 |
+
class RMSNorm(torch.nn.Module):
|
1346 |
+
def __init__(self, dim: int, eps: float = 1e-6):
|
1347 |
+
super().__init__()
|
1348 |
+
self.eps = eps
|
1349 |
+
self.weight = nn.Parameter(torch.ones(dim))
|
1350 |
+
|
1351 |
+
def _norm(self, x):
|
1352 |
+
return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
|
1353 |
+
|
1354 |
+
def forward(self, x):
|
1355 |
+
if rms_norm is not None and x.is_cuda:
|
1356 |
+
return rms_norm(x, self.weight, self.eps)
|
1357 |
+
else:
|
1358 |
+
output = self._norm(x.float()).type_as(x)
|
1359 |
+
return output * self.weight
|
1360 |
+
|
1361 |
+
|
1362 |
+
if __name__ == '__main__':
|
1363 |
+
import json
|
1364 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig
|
1365 |
+
model_path = '/mnt/petrelfs/luquanfeng/Gui-Copilot/Qwenvl-chat'
|
1366 |
+
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
|
1367 |
+
tokenizer.padding_side = 'left'
|
1368 |
+
tokenizer.pad_token_id = tokenizer.eod_id
|
1369 |
+
qs = ["Picture 1: <img>/mnt/petrelfs/luquanfeng/gui_data/aitw_pt/general/img/2971401471600307458_9.png</img>\nPlease show the next action. Goal: What's the price of the new iPhone on eBay?\nPrevious screenshots: <img>image-history: /mnt/petrelfs/luquanfeng/gui_data/aitw_pt/general/img/2971401471600307458_9.png</img>\nPrevious Actions: TYPE: n ebay\nCLICK: (393, 113)\nSCROLL: UP\nSCROLL: UP\nCLICK: (356, 599)\nCLICK: (454, 914)",
|
1370 |
+
"Picture 1: <img>/mnt/petrelfs/luquanfeng/gui_data/aitw_pt/general/img/7681036113908451219_3.png</img>\nPlease show the next action. Goal: What's on the menu at Chipotle?\nPrevious screenshots: <img>image-history: /mnt/petrelfs/luquanfeng/gui_data/aitw_pt/general/img/7681036113908451219_3.png</img>\nPrevious Actions: CLICK: (451, 891)",
|
1371 |
+
"Picture 1: <img>/mnt/petrelfs/luquanfeng/gui_data/aitw_pt/general/img/7681036113908451219_2.png</img>\nPlease show the next action. Goal: What's on the menu at Chipotle?\nPrevious screenshots: <img>image-history: /mnt/petrelfs/luquanfeng/gui_data/aitw_pt/general/img/7681036113908451219_2.png</img>\nPrevious Actions: CLICK: (451, 891)",
|
1372 |
+
"Picture 1: <img>/mnt/petrelfs/luquanfeng/gui_data/aitw_pt/general/img/7681036113908451219_1.png</img>\nPlease show the next action. Goal: What's on the menu at Chipotle?\nPrevious screenshots: <img>image-history: /mnt/petrelfs/luquanfeng/gui_data/aitw_pt/general/img/7681036113908451219_1.png</img>\nPrevious Actions: CLICK: (451, 891)",
|
1373 |
+
"Picture 1: <img>/mnt/petrelfs/luquanfeng/gui_data/aitw_pt/general/img/7681036113908451219_0.png</img>\nPlease show the next action. Goal: What's on the menu at Chipotle?"]
|
1374 |
+
device = 'cuda'
|
1375 |
+
|
1376 |
+
batch_raw_text = []
|
1377 |
+
for q in qs:
|
1378 |
+
raw_text, _ = make_context(tokenizer, q, system="You are a helpful assistant.", max_window_size=6144, chat_format='chatml')
|
1379 |
+
batch_raw_text.append(raw_text)
|
1380 |
+
|
1381 |
+
batch_input = tokenizer(batch_raw_text, return_tensors='pt', padding='longest')
|
1382 |
+
batch_input_ids = batch_input['input_ids'].to(device)
|
1383 |
+
batch_input_attention_mask = batch_input['attention_mask'].to(device)
|
1384 |
+
print(batch_input_ids.shape)
|
1385 |
+
|
1386 |
+
model = AutoModelForCausalLM.from_pretrained(model_path, device_map=device, trust_remote_code=True).eval()
|
1387 |
+
import pdb; pdb.set_trace()
|
1388 |
+
batch_out_ids = model.generate(
|
1389 |
+
input_ids=batch_input_ids,
|
1390 |
+
attention_mask=batch_input_attention_mask,
|
1391 |
+
do_sample=False,
|
1392 |
+
#top_k=0,
|
1393 |
+
#top_p=0.5,
|
1394 |
+
num_beams=1,
|
1395 |
+
length_penalty=1,
|
1396 |
+
num_return_sequences=1,
|
1397 |
+
use_cache=True,
|
1398 |
+
pad_token_id=tokenizer.eod_id,
|
1399 |
+
eos_token_id=tokenizer.eod_id,
|
1400 |
+
min_new_tokens=1,
|
1401 |
+
max_new_tokens=10,
|
1402 |
+
)
|
1403 |
+
print(batch_out_ids.shape)
|
1404 |
+
print(batch_out_ids)
|
1405 |
+
print('-----------------------------------------------')
|
1406 |
+
|
1407 |
+
padding_lens = [batch_input_ids[i].eq(tokenizer.pad_token_id).sum().item() for i in range(batch_input_ids.size(0))]
|
1408 |
+
|
1409 |
+
batch_response = [
|
1410 |
+
decode_tokens(
|
1411 |
+
batch_out_ids[i][padding_lens[i]:],
|
1412 |
+
tokenizer,
|
1413 |
+
raw_text_len=len(batch_raw_text[i]),
|
1414 |
+
context_length=(batch_input_ids[i].size(0)-padding_lens[i]),
|
1415 |
+
chat_format="chatml",
|
1416 |
+
verbose=False,
|
1417 |
+
errors='replace'
|
1418 |
+
) for i in range(len(qs))
|
1419 |
+
]
|
1420 |
+
print(batch_input_ids.shape, batch_out_ids.shape)
|
1421 |
+
print(batch_response)
|
pytorch_model-00001-of-00002.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:e611779e7dab747371bece50a1e5c51f031314d65bcca3b1ed2037c375009306
|
3 |
+
size 9969772092
|
pytorch_model-00002-of-00002.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:3adeac8626039d5bd2e307758ef95e4f46ddf764ad58e8ee029df1556e301f1d
|
3 |
+
size 9516452100
|
pytorch_model.bin.index.json
ADDED
@@ -0,0 +1,873 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"metadata": {
|
3 |
+
"total_size": 19485918720
|
4 |
+
},
|
5 |
+
"weight_map": {
|
6 |
+
"lm_head.weight": "pytorch_model-00002-of-00002.bin",
|
7 |
+
"transformer.h.0.attn.c_attn.bias": "pytorch_model-00001-of-00002.bin",
|
8 |
+
"transformer.h.0.attn.c_attn.weight": "pytorch_model-00001-of-00002.bin",
|
9 |
+
"transformer.h.0.attn.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
10 |
+
"transformer.h.0.ln_1.weight": "pytorch_model-00001-of-00002.bin",
|
11 |
+
"transformer.h.0.ln_2.weight": "pytorch_model-00001-of-00002.bin",
|
12 |
+
"transformer.h.0.mlp.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
13 |
+
"transformer.h.0.mlp.w1.weight": "pytorch_model-00001-of-00002.bin",
|
14 |
+
"transformer.h.0.mlp.w2.weight": "pytorch_model-00001-of-00002.bin",
|
15 |
+
"transformer.h.1.attn.c_attn.bias": "pytorch_model-00001-of-00002.bin",
|
16 |
+
"transformer.h.1.attn.c_attn.weight": "pytorch_model-00001-of-00002.bin",
|
17 |
+
"transformer.h.1.attn.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
18 |
+
"transformer.h.1.ln_1.weight": "pytorch_model-00001-of-00002.bin",
|
19 |
+
"transformer.h.1.ln_2.weight": "pytorch_model-00001-of-00002.bin",
|
20 |
+
"transformer.h.1.mlp.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
21 |
+
"transformer.h.1.mlp.w1.weight": "pytorch_model-00001-of-00002.bin",
|
22 |
+
"transformer.h.1.mlp.w2.weight": "pytorch_model-00001-of-00002.bin",
|
23 |
+
"transformer.h.10.attn.c_attn.bias": "pytorch_model-00001-of-00002.bin",
|
24 |
+
"transformer.h.10.attn.c_attn.weight": "pytorch_model-00001-of-00002.bin",
|
25 |
+
"transformer.h.10.attn.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
26 |
+
"transformer.h.10.ln_1.weight": "pytorch_model-00001-of-00002.bin",
|
27 |
+
"transformer.h.10.ln_2.weight": "pytorch_model-00001-of-00002.bin",
|
28 |
+
"transformer.h.10.mlp.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
29 |
+
"transformer.h.10.mlp.w1.weight": "pytorch_model-00001-of-00002.bin",
|
30 |
+
"transformer.h.10.mlp.w2.weight": "pytorch_model-00001-of-00002.bin",
|
31 |
+
"transformer.h.11.attn.c_attn.bias": "pytorch_model-00001-of-00002.bin",
|
32 |
+
"transformer.h.11.attn.c_attn.weight": "pytorch_model-00001-of-00002.bin",
|
33 |
+
"transformer.h.11.attn.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
34 |
+
"transformer.h.11.ln_1.weight": "pytorch_model-00001-of-00002.bin",
|
35 |
+
"transformer.h.11.ln_2.weight": "pytorch_model-00001-of-00002.bin",
|
36 |
+
"transformer.h.11.mlp.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
37 |
+
"transformer.h.11.mlp.w1.weight": "pytorch_model-00001-of-00002.bin",
|
38 |
+
"transformer.h.11.mlp.w2.weight": "pytorch_model-00001-of-00002.bin",
|
39 |
+
"transformer.h.12.attn.c_attn.bias": "pytorch_model-00001-of-00002.bin",
|
40 |
+
"transformer.h.12.attn.c_attn.weight": "pytorch_model-00001-of-00002.bin",
|
41 |
+
"transformer.h.12.attn.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
42 |
+
"transformer.h.12.ln_1.weight": "pytorch_model-00001-of-00002.bin",
|
43 |
+
"transformer.h.12.ln_2.weight": "pytorch_model-00001-of-00002.bin",
|
44 |
+
"transformer.h.12.mlp.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
45 |
+
"transformer.h.12.mlp.w1.weight": "pytorch_model-00001-of-00002.bin",
|
46 |
+
"transformer.h.12.mlp.w2.weight": "pytorch_model-00001-of-00002.bin",
|
47 |
+
"transformer.h.13.attn.c_attn.bias": "pytorch_model-00001-of-00002.bin",
|
48 |
+
"transformer.h.13.attn.c_attn.weight": "pytorch_model-00001-of-00002.bin",
|
49 |
+
"transformer.h.13.attn.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
50 |
+
"transformer.h.13.ln_1.weight": "pytorch_model-00001-of-00002.bin",
|
51 |
+
"transformer.h.13.ln_2.weight": "pytorch_model-00001-of-00002.bin",
|
52 |
+
"transformer.h.13.mlp.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
53 |
+
"transformer.h.13.mlp.w1.weight": "pytorch_model-00001-of-00002.bin",
|
54 |
+
"transformer.h.13.mlp.w2.weight": "pytorch_model-00001-of-00002.bin",
|
55 |
+
"transformer.h.14.attn.c_attn.bias": "pytorch_model-00001-of-00002.bin",
|
56 |
+
"transformer.h.14.attn.c_attn.weight": "pytorch_model-00001-of-00002.bin",
|
57 |
+
"transformer.h.14.attn.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
58 |
+
"transformer.h.14.ln_1.weight": "pytorch_model-00001-of-00002.bin",
|
59 |
+
"transformer.h.14.ln_2.weight": "pytorch_model-00001-of-00002.bin",
|
60 |
+
"transformer.h.14.mlp.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
61 |
+
"transformer.h.14.mlp.w1.weight": "pytorch_model-00001-of-00002.bin",
|
62 |
+
"transformer.h.14.mlp.w2.weight": "pytorch_model-00001-of-00002.bin",
|
63 |
+
"transformer.h.15.attn.c_attn.bias": "pytorch_model-00001-of-00002.bin",
|
64 |
+
"transformer.h.15.attn.c_attn.weight": "pytorch_model-00001-of-00002.bin",
|
65 |
+
"transformer.h.15.attn.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
66 |
+
"transformer.h.15.ln_1.weight": "pytorch_model-00001-of-00002.bin",
|
67 |
+
"transformer.h.15.ln_2.weight": "pytorch_model-00001-of-00002.bin",
|
68 |
+
"transformer.h.15.mlp.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
69 |
+
"transformer.h.15.mlp.w1.weight": "pytorch_model-00001-of-00002.bin",
|
70 |
+
"transformer.h.15.mlp.w2.weight": "pytorch_model-00001-of-00002.bin",
|
71 |
+
"transformer.h.16.attn.c_attn.bias": "pytorch_model-00001-of-00002.bin",
|
72 |
+
"transformer.h.16.attn.c_attn.weight": "pytorch_model-00001-of-00002.bin",
|
73 |
+
"transformer.h.16.attn.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
74 |
+
"transformer.h.16.ln_1.weight": "pytorch_model-00001-of-00002.bin",
|
75 |
+
"transformer.h.16.ln_2.weight": "pytorch_model-00001-of-00002.bin",
|
76 |
+
"transformer.h.16.mlp.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
77 |
+
"transformer.h.16.mlp.w1.weight": "pytorch_model-00001-of-00002.bin",
|
78 |
+
"transformer.h.16.mlp.w2.weight": "pytorch_model-00001-of-00002.bin",
|
79 |
+
"transformer.h.17.attn.c_attn.bias": "pytorch_model-00001-of-00002.bin",
|
80 |
+
"transformer.h.17.attn.c_attn.weight": "pytorch_model-00001-of-00002.bin",
|
81 |
+
"transformer.h.17.attn.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
82 |
+
"transformer.h.17.ln_1.weight": "pytorch_model-00001-of-00002.bin",
|
83 |
+
"transformer.h.17.ln_2.weight": "pytorch_model-00001-of-00002.bin",
|
84 |
+
"transformer.h.17.mlp.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
85 |
+
"transformer.h.17.mlp.w1.weight": "pytorch_model-00001-of-00002.bin",
|
86 |
+
"transformer.h.17.mlp.w2.weight": "pytorch_model-00001-of-00002.bin",
|
87 |
+
"transformer.h.18.attn.c_attn.bias": "pytorch_model-00001-of-00002.bin",
|
88 |
+
"transformer.h.18.attn.c_attn.weight": "pytorch_model-00001-of-00002.bin",
|
89 |
+
"transformer.h.18.attn.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
90 |
+
"transformer.h.18.ln_1.weight": "pytorch_model-00001-of-00002.bin",
|
91 |
+
"transformer.h.18.ln_2.weight": "pytorch_model-00001-of-00002.bin",
|
92 |
+
"transformer.h.18.mlp.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
93 |
+
"transformer.h.18.mlp.w1.weight": "pytorch_model-00001-of-00002.bin",
|
94 |
+
"transformer.h.18.mlp.w2.weight": "pytorch_model-00001-of-00002.bin",
|
95 |
+
"transformer.h.19.attn.c_attn.bias": "pytorch_model-00001-of-00002.bin",
|
96 |
+
"transformer.h.19.attn.c_attn.weight": "pytorch_model-00001-of-00002.bin",
|
97 |
+
"transformer.h.19.attn.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
98 |
+
"transformer.h.19.ln_1.weight": "pytorch_model-00001-of-00002.bin",
|
99 |
+
"transformer.h.19.ln_2.weight": "pytorch_model-00001-of-00002.bin",
|
100 |
+
"transformer.h.19.mlp.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
101 |
+
"transformer.h.19.mlp.w1.weight": "pytorch_model-00001-of-00002.bin",
|
102 |
+
"transformer.h.19.mlp.w2.weight": "pytorch_model-00001-of-00002.bin",
|
103 |
+
"transformer.h.2.attn.c_attn.bias": "pytorch_model-00001-of-00002.bin",
|
104 |
+
"transformer.h.2.attn.c_attn.weight": "pytorch_model-00001-of-00002.bin",
|
105 |
+
"transformer.h.2.attn.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
106 |
+
"transformer.h.2.ln_1.weight": "pytorch_model-00001-of-00002.bin",
|
107 |
+
"transformer.h.2.ln_2.weight": "pytorch_model-00001-of-00002.bin",
|
108 |
+
"transformer.h.2.mlp.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
109 |
+
"transformer.h.2.mlp.w1.weight": "pytorch_model-00001-of-00002.bin",
|
110 |
+
"transformer.h.2.mlp.w2.weight": "pytorch_model-00001-of-00002.bin",
|
111 |
+
"transformer.h.20.attn.c_attn.bias": "pytorch_model-00001-of-00002.bin",
|
112 |
+
"transformer.h.20.attn.c_attn.weight": "pytorch_model-00001-of-00002.bin",
|
113 |
+
"transformer.h.20.attn.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
114 |
+
"transformer.h.20.ln_1.weight": "pytorch_model-00001-of-00002.bin",
|
115 |
+
"transformer.h.20.ln_2.weight": "pytorch_model-00001-of-00002.bin",
|
116 |
+
"transformer.h.20.mlp.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
117 |
+
"transformer.h.20.mlp.w1.weight": "pytorch_model-00001-of-00002.bin",
|
118 |
+
"transformer.h.20.mlp.w2.weight": "pytorch_model-00001-of-00002.bin",
|
119 |
+
"transformer.h.21.attn.c_attn.bias": "pytorch_model-00001-of-00002.bin",
|
120 |
+
"transformer.h.21.attn.c_attn.weight": "pytorch_model-00001-of-00002.bin",
|
121 |
+
"transformer.h.21.attn.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
122 |
+
"transformer.h.21.ln_1.weight": "pytorch_model-00001-of-00002.bin",
|
123 |
+
"transformer.h.21.ln_2.weight": "pytorch_model-00001-of-00002.bin",
|
124 |
+
"transformer.h.21.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
125 |
+
"transformer.h.21.mlp.w1.weight": "pytorch_model-00001-of-00002.bin",
|
126 |
+
"transformer.h.21.mlp.w2.weight": "pytorch_model-00002-of-00002.bin",
|
127 |
+
"transformer.h.22.attn.c_attn.bias": "pytorch_model-00002-of-00002.bin",
|
128 |
+
"transformer.h.22.attn.c_attn.weight": "pytorch_model-00002-of-00002.bin",
|
129 |
+
"transformer.h.22.attn.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
130 |
+
"transformer.h.22.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
131 |
+
"transformer.h.22.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
132 |
+
"transformer.h.22.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
133 |
+
"transformer.h.22.mlp.w1.weight": "pytorch_model-00002-of-00002.bin",
|
134 |
+
"transformer.h.22.mlp.w2.weight": "pytorch_model-00002-of-00002.bin",
|
135 |
+
"transformer.h.23.attn.c_attn.bias": "pytorch_model-00002-of-00002.bin",
|
136 |
+
"transformer.h.23.attn.c_attn.weight": "pytorch_model-00002-of-00002.bin",
|
137 |
+
"transformer.h.23.attn.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
138 |
+
"transformer.h.23.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
139 |
+
"transformer.h.23.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
140 |
+
"transformer.h.23.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
141 |
+
"transformer.h.23.mlp.w1.weight": "pytorch_model-00002-of-00002.bin",
|
142 |
+
"transformer.h.23.mlp.w2.weight": "pytorch_model-00002-of-00002.bin",
|
143 |
+
"transformer.h.24.attn.c_attn.bias": "pytorch_model-00002-of-00002.bin",
|
144 |
+
"transformer.h.24.attn.c_attn.weight": "pytorch_model-00002-of-00002.bin",
|
145 |
+
"transformer.h.24.attn.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
146 |
+
"transformer.h.24.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
147 |
+
"transformer.h.24.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
148 |
+
"transformer.h.24.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
149 |
+
"transformer.h.24.mlp.w1.weight": "pytorch_model-00002-of-00002.bin",
|
150 |
+
"transformer.h.24.mlp.w2.weight": "pytorch_model-00002-of-00002.bin",
|
151 |
+
"transformer.h.25.attn.c_attn.bias": "pytorch_model-00002-of-00002.bin",
|
152 |
+
"transformer.h.25.attn.c_attn.weight": "pytorch_model-00002-of-00002.bin",
|
153 |
+
"transformer.h.25.attn.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
154 |
+
"transformer.h.25.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
155 |
+
"transformer.h.25.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
156 |
+
"transformer.h.25.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
157 |
+
"transformer.h.25.mlp.w1.weight": "pytorch_model-00002-of-00002.bin",
|
158 |
+
"transformer.h.25.mlp.w2.weight": "pytorch_model-00002-of-00002.bin",
|
159 |
+
"transformer.h.26.attn.c_attn.bias": "pytorch_model-00002-of-00002.bin",
|
160 |
+
"transformer.h.26.attn.c_attn.weight": "pytorch_model-00002-of-00002.bin",
|
161 |
+
"transformer.h.26.attn.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
162 |
+
"transformer.h.26.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
163 |
+
"transformer.h.26.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
164 |
+
"transformer.h.26.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
165 |
+
"transformer.h.26.mlp.w1.weight": "pytorch_model-00002-of-00002.bin",
|
166 |
+
"transformer.h.26.mlp.w2.weight": "pytorch_model-00002-of-00002.bin",
|
167 |
+
"transformer.h.27.attn.c_attn.bias": "pytorch_model-00002-of-00002.bin",
|
168 |
+
"transformer.h.27.attn.c_attn.weight": "pytorch_model-00002-of-00002.bin",
|
169 |
+
"transformer.h.27.attn.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
170 |
+
"transformer.h.27.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
171 |
+
"transformer.h.27.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
172 |
+
"transformer.h.27.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
173 |
+
"transformer.h.27.mlp.w1.weight": "pytorch_model-00002-of-00002.bin",
|
174 |
+
"transformer.h.27.mlp.w2.weight": "pytorch_model-00002-of-00002.bin",
|
175 |
+
"transformer.h.28.attn.c_attn.bias": "pytorch_model-00002-of-00002.bin",
|
176 |
+
"transformer.h.28.attn.c_attn.weight": "pytorch_model-00002-of-00002.bin",
|
177 |
+
"transformer.h.28.attn.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
178 |
+
"transformer.h.28.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
179 |
+
"transformer.h.28.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
180 |
+
"transformer.h.28.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
181 |
+
"transformer.h.28.mlp.w1.weight": "pytorch_model-00002-of-00002.bin",
|
182 |
+
"transformer.h.28.mlp.w2.weight": "pytorch_model-00002-of-00002.bin",
|
183 |
+
"transformer.h.29.attn.c_attn.bias": "pytorch_model-00002-of-00002.bin",
|
184 |
+
"transformer.h.29.attn.c_attn.weight": "pytorch_model-00002-of-00002.bin",
|
185 |
+
"transformer.h.29.attn.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
186 |
+
"transformer.h.29.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
187 |
+
"transformer.h.29.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
188 |
+
"transformer.h.29.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
189 |
+
"transformer.h.29.mlp.w1.weight": "pytorch_model-00002-of-00002.bin",
|
190 |
+
"transformer.h.29.mlp.w2.weight": "pytorch_model-00002-of-00002.bin",
|
191 |
+
"transformer.h.3.attn.c_attn.bias": "pytorch_model-00001-of-00002.bin",
|
192 |
+
"transformer.h.3.attn.c_attn.weight": "pytorch_model-00001-of-00002.bin",
|
193 |
+
"transformer.h.3.attn.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
194 |
+
"transformer.h.3.ln_1.weight": "pytorch_model-00001-of-00002.bin",
|
195 |
+
"transformer.h.3.ln_2.weight": "pytorch_model-00001-of-00002.bin",
|
196 |
+
"transformer.h.3.mlp.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
197 |
+
"transformer.h.3.mlp.w1.weight": "pytorch_model-00001-of-00002.bin",
|
198 |
+
"transformer.h.3.mlp.w2.weight": "pytorch_model-00001-of-00002.bin",
|
199 |
+
"transformer.h.30.attn.c_attn.bias": "pytorch_model-00002-of-00002.bin",
|
200 |
+
"transformer.h.30.attn.c_attn.weight": "pytorch_model-00002-of-00002.bin",
|
201 |
+
"transformer.h.30.attn.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
202 |
+
"transformer.h.30.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
203 |
+
"transformer.h.30.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
204 |
+
"transformer.h.30.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
205 |
+
"transformer.h.30.mlp.w1.weight": "pytorch_model-00002-of-00002.bin",
|
206 |
+
"transformer.h.30.mlp.w2.weight": "pytorch_model-00002-of-00002.bin",
|
207 |
+
"transformer.h.31.attn.c_attn.bias": "pytorch_model-00002-of-00002.bin",
|
208 |
+
"transformer.h.31.attn.c_attn.weight": "pytorch_model-00002-of-00002.bin",
|
209 |
+
"transformer.h.31.attn.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
210 |
+
"transformer.h.31.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
211 |
+
"transformer.h.31.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
212 |
+
"transformer.h.31.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
213 |
+
"transformer.h.31.mlp.w1.weight": "pytorch_model-00002-of-00002.bin",
|
214 |
+
"transformer.h.31.mlp.w2.weight": "pytorch_model-00002-of-00002.bin",
|
215 |
+
"transformer.h.4.attn.c_attn.bias": "pytorch_model-00001-of-00002.bin",
|
216 |
+
"transformer.h.4.attn.c_attn.weight": "pytorch_model-00001-of-00002.bin",
|
217 |
+
"transformer.h.4.attn.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
218 |
+
"transformer.h.4.ln_1.weight": "pytorch_model-00001-of-00002.bin",
|
219 |
+
"transformer.h.4.ln_2.weight": "pytorch_model-00001-of-00002.bin",
|
220 |
+
"transformer.h.4.mlp.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
221 |
+
"transformer.h.4.mlp.w1.weight": "pytorch_model-00001-of-00002.bin",
|
222 |
+
"transformer.h.4.mlp.w2.weight": "pytorch_model-00001-of-00002.bin",
|
223 |
+
"transformer.h.5.attn.c_attn.bias": "pytorch_model-00001-of-00002.bin",
|
224 |
+
"transformer.h.5.attn.c_attn.weight": "pytorch_model-00001-of-00002.bin",
|
225 |
+
"transformer.h.5.attn.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
226 |
+
"transformer.h.5.ln_1.weight": "pytorch_model-00001-of-00002.bin",
|
227 |
+
"transformer.h.5.ln_2.weight": "pytorch_model-00001-of-00002.bin",
|
228 |
+
"transformer.h.5.mlp.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
229 |
+
"transformer.h.5.mlp.w1.weight": "pytorch_model-00001-of-00002.bin",
|
230 |
+
"transformer.h.5.mlp.w2.weight": "pytorch_model-00001-of-00002.bin",
|
231 |
+
"transformer.h.6.attn.c_attn.bias": "pytorch_model-00001-of-00002.bin",
|
232 |
+
"transformer.h.6.attn.c_attn.weight": "pytorch_model-00001-of-00002.bin",
|
233 |
+
"transformer.h.6.attn.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
234 |
+
"transformer.h.6.ln_1.weight": "pytorch_model-00001-of-00002.bin",
|
235 |
+
"transformer.h.6.ln_2.weight": "pytorch_model-00001-of-00002.bin",
|
236 |
+
"transformer.h.6.mlp.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
237 |
+
"transformer.h.6.mlp.w1.weight": "pytorch_model-00001-of-00002.bin",
|
238 |
+
"transformer.h.6.mlp.w2.weight": "pytorch_model-00001-of-00002.bin",
|
239 |
+
"transformer.h.7.attn.c_attn.bias": "pytorch_model-00001-of-00002.bin",
|
240 |
+
"transformer.h.7.attn.c_attn.weight": "pytorch_model-00001-of-00002.bin",
|
241 |
+
"transformer.h.7.attn.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
242 |
+
"transformer.h.7.ln_1.weight": "pytorch_model-00001-of-00002.bin",
|
243 |
+
"transformer.h.7.ln_2.weight": "pytorch_model-00001-of-00002.bin",
|
244 |
+
"transformer.h.7.mlp.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
245 |
+
"transformer.h.7.mlp.w1.weight": "pytorch_model-00001-of-00002.bin",
|
246 |
+
"transformer.h.7.mlp.w2.weight": "pytorch_model-00001-of-00002.bin",
|
247 |
+
"transformer.h.8.attn.c_attn.bias": "pytorch_model-00001-of-00002.bin",
|
248 |
+
"transformer.h.8.attn.c_attn.weight": "pytorch_model-00001-of-00002.bin",
|
249 |
+
"transformer.h.8.attn.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
250 |
+
"transformer.h.8.ln_1.weight": "pytorch_model-00001-of-00002.bin",
|
251 |
+
"transformer.h.8.ln_2.weight": "pytorch_model-00001-of-00002.bin",
|
252 |
+
"transformer.h.8.mlp.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
253 |
+
"transformer.h.8.mlp.w1.weight": "pytorch_model-00001-of-00002.bin",
|
254 |
+
"transformer.h.8.mlp.w2.weight": "pytorch_model-00001-of-00002.bin",
|
255 |
+
"transformer.h.9.attn.c_attn.bias": "pytorch_model-00001-of-00002.bin",
|
256 |
+
"transformer.h.9.attn.c_attn.weight": "pytorch_model-00001-of-00002.bin",
|
257 |
+
"transformer.h.9.attn.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
258 |
+
"transformer.h.9.ln_1.weight": "pytorch_model-00001-of-00002.bin",
|
259 |
+
"transformer.h.9.ln_2.weight": "pytorch_model-00001-of-00002.bin",
|
260 |
+
"transformer.h.9.mlp.c_proj.weight": "pytorch_model-00001-of-00002.bin",
|
261 |
+
"transformer.h.9.mlp.w1.weight": "pytorch_model-00001-of-00002.bin",
|
262 |
+
"transformer.h.9.mlp.w2.weight": "pytorch_model-00001-of-00002.bin",
|
263 |
+
"transformer.his_resampler.attn.in_proj_bias": "pytorch_model-00002-of-00002.bin",
|
264 |
+
"transformer.his_resampler.attn.in_proj_weight": "pytorch_model-00002-of-00002.bin",
|
265 |
+
"transformer.his_resampler.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
266 |
+
"transformer.his_resampler.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
267 |
+
"transformer.his_resampler.ln_kv.bias": "pytorch_model-00002-of-00002.bin",
|
268 |
+
"transformer.his_resampler.ln_kv.weight": "pytorch_model-00002-of-00002.bin",
|
269 |
+
"transformer.his_resampler.ln_post.bias": "pytorch_model-00002-of-00002.bin",
|
270 |
+
"transformer.his_resampler.ln_post.weight": "pytorch_model-00002-of-00002.bin",
|
271 |
+
"transformer.his_resampler.ln_q.bias": "pytorch_model-00002-of-00002.bin",
|
272 |
+
"transformer.his_resampler.ln_q.weight": "pytorch_model-00002-of-00002.bin",
|
273 |
+
"transformer.his_resampler.pos_embed": "pytorch_model-00002-of-00002.bin",
|
274 |
+
"transformer.his_resampler.proj": "pytorch_model-00002-of-00002.bin",
|
275 |
+
"transformer.his_resampler.query": "pytorch_model-00002-of-00002.bin",
|
276 |
+
"transformer.ln_f.weight": "pytorch_model-00002-of-00002.bin",
|
277 |
+
"transformer.visual.attn_pool.attn.in_proj_bias": "pytorch_model-00002-of-00002.bin",
|
278 |
+
"transformer.visual.attn_pool.attn.in_proj_weight": "pytorch_model-00002-of-00002.bin",
|
279 |
+
"transformer.visual.attn_pool.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
280 |
+
"transformer.visual.attn_pool.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
281 |
+
"transformer.visual.attn_pool.kv_proj.weight": "pytorch_model-00002-of-00002.bin",
|
282 |
+
"transformer.visual.attn_pool.ln_kv.bias": "pytorch_model-00002-of-00002.bin",
|
283 |
+
"transformer.visual.attn_pool.ln_kv.weight": "pytorch_model-00002-of-00002.bin",
|
284 |
+
"transformer.visual.attn_pool.ln_q.bias": "pytorch_model-00002-of-00002.bin",
|
285 |
+
"transformer.visual.attn_pool.ln_q.weight": "pytorch_model-00002-of-00002.bin",
|
286 |
+
"transformer.visual.attn_pool.pos_embed": "pytorch_model-00002-of-00002.bin",
|
287 |
+
"transformer.visual.attn_pool.query": "pytorch_model-00002-of-00002.bin",
|
288 |
+
"transformer.visual.conv1.weight": "pytorch_model-00002-of-00002.bin",
|
289 |
+
"transformer.visual.ln_post.bias": "pytorch_model-00002-of-00002.bin",
|
290 |
+
"transformer.visual.ln_post.weight": "pytorch_model-00002-of-00002.bin",
|
291 |
+
"transformer.visual.ln_pre.bias": "pytorch_model-00002-of-00002.bin",
|
292 |
+
"transformer.visual.ln_pre.weight": "pytorch_model-00002-of-00002.bin",
|
293 |
+
"transformer.visual.positional_embedding": "pytorch_model-00002-of-00002.bin",
|
294 |
+
"transformer.visual.proj": "pytorch_model-00002-of-00002.bin",
|
295 |
+
"transformer.visual.transformer.resblocks.0.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
296 |
+
"transformer.visual.transformer.resblocks.0.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
297 |
+
"transformer.visual.transformer.resblocks.0.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
298 |
+
"transformer.visual.transformer.resblocks.0.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
299 |
+
"transformer.visual.transformer.resblocks.0.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
300 |
+
"transformer.visual.transformer.resblocks.0.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
301 |
+
"transformer.visual.transformer.resblocks.0.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
302 |
+
"transformer.visual.transformer.resblocks.0.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
303 |
+
"transformer.visual.transformer.resblocks.0.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
304 |
+
"transformer.visual.transformer.resblocks.0.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
305 |
+
"transformer.visual.transformer.resblocks.0.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
306 |
+
"transformer.visual.transformer.resblocks.0.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
307 |
+
"transformer.visual.transformer.resblocks.1.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
308 |
+
"transformer.visual.transformer.resblocks.1.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
309 |
+
"transformer.visual.transformer.resblocks.1.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
310 |
+
"transformer.visual.transformer.resblocks.1.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
311 |
+
"transformer.visual.transformer.resblocks.1.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
312 |
+
"transformer.visual.transformer.resblocks.1.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
313 |
+
"transformer.visual.transformer.resblocks.1.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
314 |
+
"transformer.visual.transformer.resblocks.1.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
315 |
+
"transformer.visual.transformer.resblocks.1.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
316 |
+
"transformer.visual.transformer.resblocks.1.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
317 |
+
"transformer.visual.transformer.resblocks.1.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
318 |
+
"transformer.visual.transformer.resblocks.1.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
319 |
+
"transformer.visual.transformer.resblocks.10.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
320 |
+
"transformer.visual.transformer.resblocks.10.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
321 |
+
"transformer.visual.transformer.resblocks.10.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
322 |
+
"transformer.visual.transformer.resblocks.10.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
323 |
+
"transformer.visual.transformer.resblocks.10.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
324 |
+
"transformer.visual.transformer.resblocks.10.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
325 |
+
"transformer.visual.transformer.resblocks.10.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
326 |
+
"transformer.visual.transformer.resblocks.10.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
327 |
+
"transformer.visual.transformer.resblocks.10.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
328 |
+
"transformer.visual.transformer.resblocks.10.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
329 |
+
"transformer.visual.transformer.resblocks.10.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
330 |
+
"transformer.visual.transformer.resblocks.10.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
331 |
+
"transformer.visual.transformer.resblocks.11.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
332 |
+
"transformer.visual.transformer.resblocks.11.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
333 |
+
"transformer.visual.transformer.resblocks.11.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
334 |
+
"transformer.visual.transformer.resblocks.11.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
335 |
+
"transformer.visual.transformer.resblocks.11.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
336 |
+
"transformer.visual.transformer.resblocks.11.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
337 |
+
"transformer.visual.transformer.resblocks.11.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
338 |
+
"transformer.visual.transformer.resblocks.11.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
339 |
+
"transformer.visual.transformer.resblocks.11.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
340 |
+
"transformer.visual.transformer.resblocks.11.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
341 |
+
"transformer.visual.transformer.resblocks.11.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
342 |
+
"transformer.visual.transformer.resblocks.11.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
343 |
+
"transformer.visual.transformer.resblocks.12.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
344 |
+
"transformer.visual.transformer.resblocks.12.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
345 |
+
"transformer.visual.transformer.resblocks.12.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
346 |
+
"transformer.visual.transformer.resblocks.12.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
347 |
+
"transformer.visual.transformer.resblocks.12.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
348 |
+
"transformer.visual.transformer.resblocks.12.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
349 |
+
"transformer.visual.transformer.resblocks.12.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
350 |
+
"transformer.visual.transformer.resblocks.12.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
351 |
+
"transformer.visual.transformer.resblocks.12.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
352 |
+
"transformer.visual.transformer.resblocks.12.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
353 |
+
"transformer.visual.transformer.resblocks.12.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
354 |
+
"transformer.visual.transformer.resblocks.12.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
355 |
+
"transformer.visual.transformer.resblocks.13.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
356 |
+
"transformer.visual.transformer.resblocks.13.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
357 |
+
"transformer.visual.transformer.resblocks.13.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
358 |
+
"transformer.visual.transformer.resblocks.13.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
359 |
+
"transformer.visual.transformer.resblocks.13.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
360 |
+
"transformer.visual.transformer.resblocks.13.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
361 |
+
"transformer.visual.transformer.resblocks.13.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
362 |
+
"transformer.visual.transformer.resblocks.13.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
363 |
+
"transformer.visual.transformer.resblocks.13.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
364 |
+
"transformer.visual.transformer.resblocks.13.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
365 |
+
"transformer.visual.transformer.resblocks.13.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
366 |
+
"transformer.visual.transformer.resblocks.13.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
367 |
+
"transformer.visual.transformer.resblocks.14.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
368 |
+
"transformer.visual.transformer.resblocks.14.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
369 |
+
"transformer.visual.transformer.resblocks.14.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
370 |
+
"transformer.visual.transformer.resblocks.14.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
371 |
+
"transformer.visual.transformer.resblocks.14.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
372 |
+
"transformer.visual.transformer.resblocks.14.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
373 |
+
"transformer.visual.transformer.resblocks.14.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
374 |
+
"transformer.visual.transformer.resblocks.14.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
375 |
+
"transformer.visual.transformer.resblocks.14.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
376 |
+
"transformer.visual.transformer.resblocks.14.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
377 |
+
"transformer.visual.transformer.resblocks.14.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
378 |
+
"transformer.visual.transformer.resblocks.14.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
379 |
+
"transformer.visual.transformer.resblocks.15.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
380 |
+
"transformer.visual.transformer.resblocks.15.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
381 |
+
"transformer.visual.transformer.resblocks.15.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
382 |
+
"transformer.visual.transformer.resblocks.15.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
383 |
+
"transformer.visual.transformer.resblocks.15.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
384 |
+
"transformer.visual.transformer.resblocks.15.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
385 |
+
"transformer.visual.transformer.resblocks.15.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
386 |
+
"transformer.visual.transformer.resblocks.15.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
387 |
+
"transformer.visual.transformer.resblocks.15.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
388 |
+
"transformer.visual.transformer.resblocks.15.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
389 |
+
"transformer.visual.transformer.resblocks.15.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
390 |
+
"transformer.visual.transformer.resblocks.15.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
391 |
+
"transformer.visual.transformer.resblocks.16.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
392 |
+
"transformer.visual.transformer.resblocks.16.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
393 |
+
"transformer.visual.transformer.resblocks.16.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
394 |
+
"transformer.visual.transformer.resblocks.16.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
395 |
+
"transformer.visual.transformer.resblocks.16.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
396 |
+
"transformer.visual.transformer.resblocks.16.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
397 |
+
"transformer.visual.transformer.resblocks.16.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
398 |
+
"transformer.visual.transformer.resblocks.16.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
399 |
+
"transformer.visual.transformer.resblocks.16.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
400 |
+
"transformer.visual.transformer.resblocks.16.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
401 |
+
"transformer.visual.transformer.resblocks.16.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
402 |
+
"transformer.visual.transformer.resblocks.16.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
403 |
+
"transformer.visual.transformer.resblocks.17.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
404 |
+
"transformer.visual.transformer.resblocks.17.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
405 |
+
"transformer.visual.transformer.resblocks.17.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
406 |
+
"transformer.visual.transformer.resblocks.17.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
407 |
+
"transformer.visual.transformer.resblocks.17.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
408 |
+
"transformer.visual.transformer.resblocks.17.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
409 |
+
"transformer.visual.transformer.resblocks.17.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
410 |
+
"transformer.visual.transformer.resblocks.17.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
411 |
+
"transformer.visual.transformer.resblocks.17.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
412 |
+
"transformer.visual.transformer.resblocks.17.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
413 |
+
"transformer.visual.transformer.resblocks.17.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
414 |
+
"transformer.visual.transformer.resblocks.17.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
415 |
+
"transformer.visual.transformer.resblocks.18.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
416 |
+
"transformer.visual.transformer.resblocks.18.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
417 |
+
"transformer.visual.transformer.resblocks.18.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
418 |
+
"transformer.visual.transformer.resblocks.18.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
419 |
+
"transformer.visual.transformer.resblocks.18.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
420 |
+
"transformer.visual.transformer.resblocks.18.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
421 |
+
"transformer.visual.transformer.resblocks.18.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
422 |
+
"transformer.visual.transformer.resblocks.18.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
423 |
+
"transformer.visual.transformer.resblocks.18.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
424 |
+
"transformer.visual.transformer.resblocks.18.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
425 |
+
"transformer.visual.transformer.resblocks.18.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
426 |
+
"transformer.visual.transformer.resblocks.18.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
427 |
+
"transformer.visual.transformer.resblocks.19.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
428 |
+
"transformer.visual.transformer.resblocks.19.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
429 |
+
"transformer.visual.transformer.resblocks.19.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
430 |
+
"transformer.visual.transformer.resblocks.19.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
431 |
+
"transformer.visual.transformer.resblocks.19.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
432 |
+
"transformer.visual.transformer.resblocks.19.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
433 |
+
"transformer.visual.transformer.resblocks.19.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
434 |
+
"transformer.visual.transformer.resblocks.19.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
435 |
+
"transformer.visual.transformer.resblocks.19.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
436 |
+
"transformer.visual.transformer.resblocks.19.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
437 |
+
"transformer.visual.transformer.resblocks.19.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
438 |
+
"transformer.visual.transformer.resblocks.19.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
439 |
+
"transformer.visual.transformer.resblocks.2.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
440 |
+
"transformer.visual.transformer.resblocks.2.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
441 |
+
"transformer.visual.transformer.resblocks.2.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
442 |
+
"transformer.visual.transformer.resblocks.2.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
443 |
+
"transformer.visual.transformer.resblocks.2.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
444 |
+
"transformer.visual.transformer.resblocks.2.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
445 |
+
"transformer.visual.transformer.resblocks.2.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
446 |
+
"transformer.visual.transformer.resblocks.2.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
447 |
+
"transformer.visual.transformer.resblocks.2.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
448 |
+
"transformer.visual.transformer.resblocks.2.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
449 |
+
"transformer.visual.transformer.resblocks.2.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
450 |
+
"transformer.visual.transformer.resblocks.2.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
451 |
+
"transformer.visual.transformer.resblocks.20.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
452 |
+
"transformer.visual.transformer.resblocks.20.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
453 |
+
"transformer.visual.transformer.resblocks.20.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
454 |
+
"transformer.visual.transformer.resblocks.20.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
455 |
+
"transformer.visual.transformer.resblocks.20.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
456 |
+
"transformer.visual.transformer.resblocks.20.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
457 |
+
"transformer.visual.transformer.resblocks.20.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
458 |
+
"transformer.visual.transformer.resblocks.20.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
459 |
+
"transformer.visual.transformer.resblocks.20.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
460 |
+
"transformer.visual.transformer.resblocks.20.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
461 |
+
"transformer.visual.transformer.resblocks.20.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
462 |
+
"transformer.visual.transformer.resblocks.20.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
463 |
+
"transformer.visual.transformer.resblocks.21.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
464 |
+
"transformer.visual.transformer.resblocks.21.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
465 |
+
"transformer.visual.transformer.resblocks.21.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
466 |
+
"transformer.visual.transformer.resblocks.21.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
467 |
+
"transformer.visual.transformer.resblocks.21.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
468 |
+
"transformer.visual.transformer.resblocks.21.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
469 |
+
"transformer.visual.transformer.resblocks.21.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
470 |
+
"transformer.visual.transformer.resblocks.21.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
471 |
+
"transformer.visual.transformer.resblocks.21.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
472 |
+
"transformer.visual.transformer.resblocks.21.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
473 |
+
"transformer.visual.transformer.resblocks.21.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
474 |
+
"transformer.visual.transformer.resblocks.21.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
475 |
+
"transformer.visual.transformer.resblocks.22.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
476 |
+
"transformer.visual.transformer.resblocks.22.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
477 |
+
"transformer.visual.transformer.resblocks.22.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
478 |
+
"transformer.visual.transformer.resblocks.22.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
479 |
+
"transformer.visual.transformer.resblocks.22.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
480 |
+
"transformer.visual.transformer.resblocks.22.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
481 |
+
"transformer.visual.transformer.resblocks.22.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
482 |
+
"transformer.visual.transformer.resblocks.22.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
483 |
+
"transformer.visual.transformer.resblocks.22.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
484 |
+
"transformer.visual.transformer.resblocks.22.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
485 |
+
"transformer.visual.transformer.resblocks.22.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
486 |
+
"transformer.visual.transformer.resblocks.22.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
487 |
+
"transformer.visual.transformer.resblocks.23.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
488 |
+
"transformer.visual.transformer.resblocks.23.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
489 |
+
"transformer.visual.transformer.resblocks.23.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
490 |
+
"transformer.visual.transformer.resblocks.23.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
491 |
+
"transformer.visual.transformer.resblocks.23.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
492 |
+
"transformer.visual.transformer.resblocks.23.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
493 |
+
"transformer.visual.transformer.resblocks.23.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
494 |
+
"transformer.visual.transformer.resblocks.23.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
495 |
+
"transformer.visual.transformer.resblocks.23.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
496 |
+
"transformer.visual.transformer.resblocks.23.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
497 |
+
"transformer.visual.transformer.resblocks.23.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
498 |
+
"transformer.visual.transformer.resblocks.23.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
499 |
+
"transformer.visual.transformer.resblocks.24.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
500 |
+
"transformer.visual.transformer.resblocks.24.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
501 |
+
"transformer.visual.transformer.resblocks.24.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
502 |
+
"transformer.visual.transformer.resblocks.24.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
503 |
+
"transformer.visual.transformer.resblocks.24.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
504 |
+
"transformer.visual.transformer.resblocks.24.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
505 |
+
"transformer.visual.transformer.resblocks.24.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
506 |
+
"transformer.visual.transformer.resblocks.24.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
507 |
+
"transformer.visual.transformer.resblocks.24.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
508 |
+
"transformer.visual.transformer.resblocks.24.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
509 |
+
"transformer.visual.transformer.resblocks.24.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
510 |
+
"transformer.visual.transformer.resblocks.24.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
511 |
+
"transformer.visual.transformer.resblocks.25.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
512 |
+
"transformer.visual.transformer.resblocks.25.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
513 |
+
"transformer.visual.transformer.resblocks.25.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
514 |
+
"transformer.visual.transformer.resblocks.25.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
515 |
+
"transformer.visual.transformer.resblocks.25.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
516 |
+
"transformer.visual.transformer.resblocks.25.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
517 |
+
"transformer.visual.transformer.resblocks.25.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
518 |
+
"transformer.visual.transformer.resblocks.25.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
519 |
+
"transformer.visual.transformer.resblocks.25.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
520 |
+
"transformer.visual.transformer.resblocks.25.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
521 |
+
"transformer.visual.transformer.resblocks.25.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
522 |
+
"transformer.visual.transformer.resblocks.25.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
523 |
+
"transformer.visual.transformer.resblocks.26.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
524 |
+
"transformer.visual.transformer.resblocks.26.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
525 |
+
"transformer.visual.transformer.resblocks.26.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
526 |
+
"transformer.visual.transformer.resblocks.26.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
527 |
+
"transformer.visual.transformer.resblocks.26.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
528 |
+
"transformer.visual.transformer.resblocks.26.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
529 |
+
"transformer.visual.transformer.resblocks.26.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
530 |
+
"transformer.visual.transformer.resblocks.26.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
531 |
+
"transformer.visual.transformer.resblocks.26.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
532 |
+
"transformer.visual.transformer.resblocks.26.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
533 |
+
"transformer.visual.transformer.resblocks.26.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
534 |
+
"transformer.visual.transformer.resblocks.26.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
535 |
+
"transformer.visual.transformer.resblocks.27.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
536 |
+
"transformer.visual.transformer.resblocks.27.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
537 |
+
"transformer.visual.transformer.resblocks.27.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
538 |
+
"transformer.visual.transformer.resblocks.27.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
539 |
+
"transformer.visual.transformer.resblocks.27.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
540 |
+
"transformer.visual.transformer.resblocks.27.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
541 |
+
"transformer.visual.transformer.resblocks.27.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
542 |
+
"transformer.visual.transformer.resblocks.27.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
543 |
+
"transformer.visual.transformer.resblocks.27.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
544 |
+
"transformer.visual.transformer.resblocks.27.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
545 |
+
"transformer.visual.transformer.resblocks.27.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
546 |
+
"transformer.visual.transformer.resblocks.27.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
547 |
+
"transformer.visual.transformer.resblocks.28.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
548 |
+
"transformer.visual.transformer.resblocks.28.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
549 |
+
"transformer.visual.transformer.resblocks.28.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
550 |
+
"transformer.visual.transformer.resblocks.28.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
551 |
+
"transformer.visual.transformer.resblocks.28.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
552 |
+
"transformer.visual.transformer.resblocks.28.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
553 |
+
"transformer.visual.transformer.resblocks.28.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
554 |
+
"transformer.visual.transformer.resblocks.28.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
555 |
+
"transformer.visual.transformer.resblocks.28.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
556 |
+
"transformer.visual.transformer.resblocks.28.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
557 |
+
"transformer.visual.transformer.resblocks.28.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
558 |
+
"transformer.visual.transformer.resblocks.28.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
559 |
+
"transformer.visual.transformer.resblocks.29.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
560 |
+
"transformer.visual.transformer.resblocks.29.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
561 |
+
"transformer.visual.transformer.resblocks.29.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
562 |
+
"transformer.visual.transformer.resblocks.29.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
563 |
+
"transformer.visual.transformer.resblocks.29.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
564 |
+
"transformer.visual.transformer.resblocks.29.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
565 |
+
"transformer.visual.transformer.resblocks.29.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
566 |
+
"transformer.visual.transformer.resblocks.29.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
567 |
+
"transformer.visual.transformer.resblocks.29.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
568 |
+
"transformer.visual.transformer.resblocks.29.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
569 |
+
"transformer.visual.transformer.resblocks.29.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
570 |
+
"transformer.visual.transformer.resblocks.29.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
571 |
+
"transformer.visual.transformer.resblocks.3.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
572 |
+
"transformer.visual.transformer.resblocks.3.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
573 |
+
"transformer.visual.transformer.resblocks.3.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
574 |
+
"transformer.visual.transformer.resblocks.3.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
575 |
+
"transformer.visual.transformer.resblocks.3.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
576 |
+
"transformer.visual.transformer.resblocks.3.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
577 |
+
"transformer.visual.transformer.resblocks.3.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
578 |
+
"transformer.visual.transformer.resblocks.3.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
579 |
+
"transformer.visual.transformer.resblocks.3.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
580 |
+
"transformer.visual.transformer.resblocks.3.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
581 |
+
"transformer.visual.transformer.resblocks.3.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
582 |
+
"transformer.visual.transformer.resblocks.3.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
583 |
+
"transformer.visual.transformer.resblocks.30.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
584 |
+
"transformer.visual.transformer.resblocks.30.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
585 |
+
"transformer.visual.transformer.resblocks.30.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
586 |
+
"transformer.visual.transformer.resblocks.30.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
587 |
+
"transformer.visual.transformer.resblocks.30.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
588 |
+
"transformer.visual.transformer.resblocks.30.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
589 |
+
"transformer.visual.transformer.resblocks.30.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
590 |
+
"transformer.visual.transformer.resblocks.30.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
591 |
+
"transformer.visual.transformer.resblocks.30.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
592 |
+
"transformer.visual.transformer.resblocks.30.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
593 |
+
"transformer.visual.transformer.resblocks.30.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
594 |
+
"transformer.visual.transformer.resblocks.30.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
595 |
+
"transformer.visual.transformer.resblocks.31.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
596 |
+
"transformer.visual.transformer.resblocks.31.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
597 |
+
"transformer.visual.transformer.resblocks.31.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
598 |
+
"transformer.visual.transformer.resblocks.31.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
599 |
+
"transformer.visual.transformer.resblocks.31.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
600 |
+
"transformer.visual.transformer.resblocks.31.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
601 |
+
"transformer.visual.transformer.resblocks.31.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
602 |
+
"transformer.visual.transformer.resblocks.31.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
603 |
+
"transformer.visual.transformer.resblocks.31.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
604 |
+
"transformer.visual.transformer.resblocks.31.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
605 |
+
"transformer.visual.transformer.resblocks.31.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
606 |
+
"transformer.visual.transformer.resblocks.31.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
607 |
+
"transformer.visual.transformer.resblocks.32.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
608 |
+
"transformer.visual.transformer.resblocks.32.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
609 |
+
"transformer.visual.transformer.resblocks.32.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
610 |
+
"transformer.visual.transformer.resblocks.32.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
611 |
+
"transformer.visual.transformer.resblocks.32.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
612 |
+
"transformer.visual.transformer.resblocks.32.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
613 |
+
"transformer.visual.transformer.resblocks.32.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
614 |
+
"transformer.visual.transformer.resblocks.32.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
615 |
+
"transformer.visual.transformer.resblocks.32.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
616 |
+
"transformer.visual.transformer.resblocks.32.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
617 |
+
"transformer.visual.transformer.resblocks.32.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
618 |
+
"transformer.visual.transformer.resblocks.32.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
619 |
+
"transformer.visual.transformer.resblocks.33.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
620 |
+
"transformer.visual.transformer.resblocks.33.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
621 |
+
"transformer.visual.transformer.resblocks.33.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
622 |
+
"transformer.visual.transformer.resblocks.33.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
623 |
+
"transformer.visual.transformer.resblocks.33.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
624 |
+
"transformer.visual.transformer.resblocks.33.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
625 |
+
"transformer.visual.transformer.resblocks.33.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
626 |
+
"transformer.visual.transformer.resblocks.33.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
627 |
+
"transformer.visual.transformer.resblocks.33.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
628 |
+
"transformer.visual.transformer.resblocks.33.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
629 |
+
"transformer.visual.transformer.resblocks.33.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
630 |
+
"transformer.visual.transformer.resblocks.33.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
631 |
+
"transformer.visual.transformer.resblocks.34.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
632 |
+
"transformer.visual.transformer.resblocks.34.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
633 |
+
"transformer.visual.transformer.resblocks.34.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
634 |
+
"transformer.visual.transformer.resblocks.34.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
635 |
+
"transformer.visual.transformer.resblocks.34.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
636 |
+
"transformer.visual.transformer.resblocks.34.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
637 |
+
"transformer.visual.transformer.resblocks.34.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
638 |
+
"transformer.visual.transformer.resblocks.34.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
639 |
+
"transformer.visual.transformer.resblocks.34.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
640 |
+
"transformer.visual.transformer.resblocks.34.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
641 |
+
"transformer.visual.transformer.resblocks.34.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
642 |
+
"transformer.visual.transformer.resblocks.34.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
643 |
+
"transformer.visual.transformer.resblocks.35.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
644 |
+
"transformer.visual.transformer.resblocks.35.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
645 |
+
"transformer.visual.transformer.resblocks.35.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
646 |
+
"transformer.visual.transformer.resblocks.35.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
647 |
+
"transformer.visual.transformer.resblocks.35.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
648 |
+
"transformer.visual.transformer.resblocks.35.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
649 |
+
"transformer.visual.transformer.resblocks.35.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
650 |
+
"transformer.visual.transformer.resblocks.35.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
651 |
+
"transformer.visual.transformer.resblocks.35.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
652 |
+
"transformer.visual.transformer.resblocks.35.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
653 |
+
"transformer.visual.transformer.resblocks.35.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
654 |
+
"transformer.visual.transformer.resblocks.35.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
655 |
+
"transformer.visual.transformer.resblocks.36.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
656 |
+
"transformer.visual.transformer.resblocks.36.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
657 |
+
"transformer.visual.transformer.resblocks.36.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
658 |
+
"transformer.visual.transformer.resblocks.36.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
659 |
+
"transformer.visual.transformer.resblocks.36.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
660 |
+
"transformer.visual.transformer.resblocks.36.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
661 |
+
"transformer.visual.transformer.resblocks.36.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
662 |
+
"transformer.visual.transformer.resblocks.36.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
663 |
+
"transformer.visual.transformer.resblocks.36.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
664 |
+
"transformer.visual.transformer.resblocks.36.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
665 |
+
"transformer.visual.transformer.resblocks.36.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
666 |
+
"transformer.visual.transformer.resblocks.36.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
667 |
+
"transformer.visual.transformer.resblocks.37.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
668 |
+
"transformer.visual.transformer.resblocks.37.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
669 |
+
"transformer.visual.transformer.resblocks.37.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
670 |
+
"transformer.visual.transformer.resblocks.37.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
671 |
+
"transformer.visual.transformer.resblocks.37.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
672 |
+
"transformer.visual.transformer.resblocks.37.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
673 |
+
"transformer.visual.transformer.resblocks.37.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
674 |
+
"transformer.visual.transformer.resblocks.37.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
675 |
+
"transformer.visual.transformer.resblocks.37.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
676 |
+
"transformer.visual.transformer.resblocks.37.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
677 |
+
"transformer.visual.transformer.resblocks.37.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
678 |
+
"transformer.visual.transformer.resblocks.37.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
679 |
+
"transformer.visual.transformer.resblocks.38.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
680 |
+
"transformer.visual.transformer.resblocks.38.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
681 |
+
"transformer.visual.transformer.resblocks.38.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
682 |
+
"transformer.visual.transformer.resblocks.38.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
683 |
+
"transformer.visual.transformer.resblocks.38.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
684 |
+
"transformer.visual.transformer.resblocks.38.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
685 |
+
"transformer.visual.transformer.resblocks.38.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
686 |
+
"transformer.visual.transformer.resblocks.38.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
687 |
+
"transformer.visual.transformer.resblocks.38.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
688 |
+
"transformer.visual.transformer.resblocks.38.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
689 |
+
"transformer.visual.transformer.resblocks.38.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
690 |
+
"transformer.visual.transformer.resblocks.38.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
691 |
+
"transformer.visual.transformer.resblocks.39.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
692 |
+
"transformer.visual.transformer.resblocks.39.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
693 |
+
"transformer.visual.transformer.resblocks.39.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
694 |
+
"transformer.visual.transformer.resblocks.39.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
695 |
+
"transformer.visual.transformer.resblocks.39.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
696 |
+
"transformer.visual.transformer.resblocks.39.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
697 |
+
"transformer.visual.transformer.resblocks.39.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
698 |
+
"transformer.visual.transformer.resblocks.39.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
699 |
+
"transformer.visual.transformer.resblocks.39.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
700 |
+
"transformer.visual.transformer.resblocks.39.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
701 |
+
"transformer.visual.transformer.resblocks.39.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
702 |
+
"transformer.visual.transformer.resblocks.39.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
703 |
+
"transformer.visual.transformer.resblocks.4.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
704 |
+
"transformer.visual.transformer.resblocks.4.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
705 |
+
"transformer.visual.transformer.resblocks.4.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
706 |
+
"transformer.visual.transformer.resblocks.4.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
707 |
+
"transformer.visual.transformer.resblocks.4.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
708 |
+
"transformer.visual.transformer.resblocks.4.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
709 |
+
"transformer.visual.transformer.resblocks.4.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
710 |
+
"transformer.visual.transformer.resblocks.4.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
711 |
+
"transformer.visual.transformer.resblocks.4.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
712 |
+
"transformer.visual.transformer.resblocks.4.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
713 |
+
"transformer.visual.transformer.resblocks.4.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
714 |
+
"transformer.visual.transformer.resblocks.4.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
715 |
+
"transformer.visual.transformer.resblocks.40.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
716 |
+
"transformer.visual.transformer.resblocks.40.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
717 |
+
"transformer.visual.transformer.resblocks.40.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
718 |
+
"transformer.visual.transformer.resblocks.40.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
719 |
+
"transformer.visual.transformer.resblocks.40.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
720 |
+
"transformer.visual.transformer.resblocks.40.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
721 |
+
"transformer.visual.transformer.resblocks.40.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
722 |
+
"transformer.visual.transformer.resblocks.40.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
723 |
+
"transformer.visual.transformer.resblocks.40.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
724 |
+
"transformer.visual.transformer.resblocks.40.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
725 |
+
"transformer.visual.transformer.resblocks.40.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
726 |
+
"transformer.visual.transformer.resblocks.40.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
727 |
+
"transformer.visual.transformer.resblocks.41.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
728 |
+
"transformer.visual.transformer.resblocks.41.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
729 |
+
"transformer.visual.transformer.resblocks.41.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
730 |
+
"transformer.visual.transformer.resblocks.41.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
731 |
+
"transformer.visual.transformer.resblocks.41.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
732 |
+
"transformer.visual.transformer.resblocks.41.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
733 |
+
"transformer.visual.transformer.resblocks.41.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
734 |
+
"transformer.visual.transformer.resblocks.41.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
735 |
+
"transformer.visual.transformer.resblocks.41.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
736 |
+
"transformer.visual.transformer.resblocks.41.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
737 |
+
"transformer.visual.transformer.resblocks.41.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
738 |
+
"transformer.visual.transformer.resblocks.41.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
739 |
+
"transformer.visual.transformer.resblocks.42.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
740 |
+
"transformer.visual.transformer.resblocks.42.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
741 |
+
"transformer.visual.transformer.resblocks.42.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
742 |
+
"transformer.visual.transformer.resblocks.42.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
743 |
+
"transformer.visual.transformer.resblocks.42.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
744 |
+
"transformer.visual.transformer.resblocks.42.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
745 |
+
"transformer.visual.transformer.resblocks.42.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
746 |
+
"transformer.visual.transformer.resblocks.42.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
747 |
+
"transformer.visual.transformer.resblocks.42.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
748 |
+
"transformer.visual.transformer.resblocks.42.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
749 |
+
"transformer.visual.transformer.resblocks.42.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
750 |
+
"transformer.visual.transformer.resblocks.42.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
751 |
+
"transformer.visual.transformer.resblocks.43.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
752 |
+
"transformer.visual.transformer.resblocks.43.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
753 |
+
"transformer.visual.transformer.resblocks.43.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
754 |
+
"transformer.visual.transformer.resblocks.43.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
755 |
+
"transformer.visual.transformer.resblocks.43.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
756 |
+
"transformer.visual.transformer.resblocks.43.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
757 |
+
"transformer.visual.transformer.resblocks.43.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
758 |
+
"transformer.visual.transformer.resblocks.43.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
759 |
+
"transformer.visual.transformer.resblocks.43.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
760 |
+
"transformer.visual.transformer.resblocks.43.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
761 |
+
"transformer.visual.transformer.resblocks.43.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
762 |
+
"transformer.visual.transformer.resblocks.43.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
763 |
+
"transformer.visual.transformer.resblocks.44.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
764 |
+
"transformer.visual.transformer.resblocks.44.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
765 |
+
"transformer.visual.transformer.resblocks.44.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
766 |
+
"transformer.visual.transformer.resblocks.44.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
767 |
+
"transformer.visual.transformer.resblocks.44.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
768 |
+
"transformer.visual.transformer.resblocks.44.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
769 |
+
"transformer.visual.transformer.resblocks.44.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
770 |
+
"transformer.visual.transformer.resblocks.44.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
771 |
+
"transformer.visual.transformer.resblocks.44.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
772 |
+
"transformer.visual.transformer.resblocks.44.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
773 |
+
"transformer.visual.transformer.resblocks.44.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
774 |
+
"transformer.visual.transformer.resblocks.44.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
775 |
+
"transformer.visual.transformer.resblocks.45.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
776 |
+
"transformer.visual.transformer.resblocks.45.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
777 |
+
"transformer.visual.transformer.resblocks.45.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
778 |
+
"transformer.visual.transformer.resblocks.45.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
779 |
+
"transformer.visual.transformer.resblocks.45.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
780 |
+
"transformer.visual.transformer.resblocks.45.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
781 |
+
"transformer.visual.transformer.resblocks.45.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
782 |
+
"transformer.visual.transformer.resblocks.45.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
783 |
+
"transformer.visual.transformer.resblocks.45.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
784 |
+
"transformer.visual.transformer.resblocks.45.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
785 |
+
"transformer.visual.transformer.resblocks.45.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
786 |
+
"transformer.visual.transformer.resblocks.45.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
787 |
+
"transformer.visual.transformer.resblocks.46.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
788 |
+
"transformer.visual.transformer.resblocks.46.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
789 |
+
"transformer.visual.transformer.resblocks.46.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
790 |
+
"transformer.visual.transformer.resblocks.46.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
791 |
+
"transformer.visual.transformer.resblocks.46.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
792 |
+
"transformer.visual.transformer.resblocks.46.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
793 |
+
"transformer.visual.transformer.resblocks.46.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
794 |
+
"transformer.visual.transformer.resblocks.46.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
795 |
+
"transformer.visual.transformer.resblocks.46.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
796 |
+
"transformer.visual.transformer.resblocks.46.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
797 |
+
"transformer.visual.transformer.resblocks.46.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
798 |
+
"transformer.visual.transformer.resblocks.46.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
799 |
+
"transformer.visual.transformer.resblocks.47.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
800 |
+
"transformer.visual.transformer.resblocks.47.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
801 |
+
"transformer.visual.transformer.resblocks.47.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
802 |
+
"transformer.visual.transformer.resblocks.47.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
803 |
+
"transformer.visual.transformer.resblocks.47.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
804 |
+
"transformer.visual.transformer.resblocks.47.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
805 |
+
"transformer.visual.transformer.resblocks.47.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
806 |
+
"transformer.visual.transformer.resblocks.47.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
807 |
+
"transformer.visual.transformer.resblocks.47.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
808 |
+
"transformer.visual.transformer.resblocks.47.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
809 |
+
"transformer.visual.transformer.resblocks.47.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
810 |
+
"transformer.visual.transformer.resblocks.47.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
811 |
+
"transformer.visual.transformer.resblocks.5.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
812 |
+
"transformer.visual.transformer.resblocks.5.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
813 |
+
"transformer.visual.transformer.resblocks.5.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
814 |
+
"transformer.visual.transformer.resblocks.5.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
815 |
+
"transformer.visual.transformer.resblocks.5.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
816 |
+
"transformer.visual.transformer.resblocks.5.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
817 |
+
"transformer.visual.transformer.resblocks.5.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
818 |
+
"transformer.visual.transformer.resblocks.5.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
819 |
+
"transformer.visual.transformer.resblocks.5.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
820 |
+
"transformer.visual.transformer.resblocks.5.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
821 |
+
"transformer.visual.transformer.resblocks.5.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
822 |
+
"transformer.visual.transformer.resblocks.5.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
823 |
+
"transformer.visual.transformer.resblocks.6.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
824 |
+
"transformer.visual.transformer.resblocks.6.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
825 |
+
"transformer.visual.transformer.resblocks.6.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
826 |
+
"transformer.visual.transformer.resblocks.6.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
827 |
+
"transformer.visual.transformer.resblocks.6.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
828 |
+
"transformer.visual.transformer.resblocks.6.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
829 |
+
"transformer.visual.transformer.resblocks.6.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
830 |
+
"transformer.visual.transformer.resblocks.6.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
831 |
+
"transformer.visual.transformer.resblocks.6.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
832 |
+
"transformer.visual.transformer.resblocks.6.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
833 |
+
"transformer.visual.transformer.resblocks.6.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
834 |
+
"transformer.visual.transformer.resblocks.6.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
835 |
+
"transformer.visual.transformer.resblocks.7.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
836 |
+
"transformer.visual.transformer.resblocks.7.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
837 |
+
"transformer.visual.transformer.resblocks.7.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
838 |
+
"transformer.visual.transformer.resblocks.7.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
839 |
+
"transformer.visual.transformer.resblocks.7.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
840 |
+
"transformer.visual.transformer.resblocks.7.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
841 |
+
"transformer.visual.transformer.resblocks.7.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
842 |
+
"transformer.visual.transformer.resblocks.7.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
843 |
+
"transformer.visual.transformer.resblocks.7.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
844 |
+
"transformer.visual.transformer.resblocks.7.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
845 |
+
"transformer.visual.transformer.resblocks.7.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
846 |
+
"transformer.visual.transformer.resblocks.7.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
847 |
+
"transformer.visual.transformer.resblocks.8.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
848 |
+
"transformer.visual.transformer.resblocks.8.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
849 |
+
"transformer.visual.transformer.resblocks.8.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
850 |
+
"transformer.visual.transformer.resblocks.8.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
851 |
+
"transformer.visual.transformer.resblocks.8.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
852 |
+
"transformer.visual.transformer.resblocks.8.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
853 |
+
"transformer.visual.transformer.resblocks.8.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
854 |
+
"transformer.visual.transformer.resblocks.8.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
855 |
+
"transformer.visual.transformer.resblocks.8.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
856 |
+
"transformer.visual.transformer.resblocks.8.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
857 |
+
"transformer.visual.transformer.resblocks.8.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
858 |
+
"transformer.visual.transformer.resblocks.8.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
859 |
+
"transformer.visual.transformer.resblocks.9.attn.in_proj.bias": "pytorch_model-00002-of-00002.bin",
|
860 |
+
"transformer.visual.transformer.resblocks.9.attn.in_proj.weight": "pytorch_model-00002-of-00002.bin",
|
861 |
+
"transformer.visual.transformer.resblocks.9.attn.out_proj.bias": "pytorch_model-00002-of-00002.bin",
|
862 |
+
"transformer.visual.transformer.resblocks.9.attn.out_proj.weight": "pytorch_model-00002-of-00002.bin",
|
863 |
+
"transformer.visual.transformer.resblocks.9.ln_1.bias": "pytorch_model-00002-of-00002.bin",
|
864 |
+
"transformer.visual.transformer.resblocks.9.ln_1.weight": "pytorch_model-00002-of-00002.bin",
|
865 |
+
"transformer.visual.transformer.resblocks.9.ln_2.bias": "pytorch_model-00002-of-00002.bin",
|
866 |
+
"transformer.visual.transformer.resblocks.9.ln_2.weight": "pytorch_model-00002-of-00002.bin",
|
867 |
+
"transformer.visual.transformer.resblocks.9.mlp.c_fc.bias": "pytorch_model-00002-of-00002.bin",
|
868 |
+
"transformer.visual.transformer.resblocks.9.mlp.c_fc.weight": "pytorch_model-00002-of-00002.bin",
|
869 |
+
"transformer.visual.transformer.resblocks.9.mlp.c_proj.bias": "pytorch_model-00002-of-00002.bin",
|
870 |
+
"transformer.visual.transformer.resblocks.9.mlp.c_proj.weight": "pytorch_model-00002-of-00002.bin",
|
871 |
+
"transformer.wte.weight": "pytorch_model-00001-of-00002.bin"
|
872 |
+
}
|
873 |
+
}
|
qwen.tiktoken
ADDED
The diff for this file is too large to render.
See raw diff
|
|
special_tokens_map.json
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"pad_token": "<|endoftext|>"
|
3 |
+
}
|
tokenization_qwen.py
ADDED
@@ -0,0 +1,598 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Alibaba Cloud.
|
2 |
+
#
|
3 |
+
# This source code is licensed under the license found in the
|
4 |
+
# LICENSE file in the root directory of this source tree.
|
5 |
+
|
6 |
+
"""Tokenization classes for QWen."""
|
7 |
+
|
8 |
+
import base64
|
9 |
+
import logging
|
10 |
+
import os
|
11 |
+
import requests
|
12 |
+
import unicodedata
|
13 |
+
from typing import Collection, Dict, List, Set, Tuple, Union, Any, Callable, Optional
|
14 |
+
|
15 |
+
import tiktoken
|
16 |
+
import numpy as np
|
17 |
+
from PIL import Image
|
18 |
+
from PIL import ImageFont
|
19 |
+
from PIL import ImageDraw
|
20 |
+
from transformers import PreTrainedTokenizer, AddedToken
|
21 |
+
from transformers.utils import try_to_load_from_cache
|
22 |
+
|
23 |
+
import matplotlib.colors as mcolors
|
24 |
+
from matplotlib.font_manager import FontProperties
|
25 |
+
|
26 |
+
logger = logging.getLogger(__name__)
|
27 |
+
|
28 |
+
|
29 |
+
VOCAB_FILES_NAMES = {"vocab_file": "qwen.tiktoken", "ttf": "SimSun.ttf"}
|
30 |
+
FONT_PATH = try_to_load_from_cache("Qwen/Qwen-VL-Chat", "SimSun.ttf")
|
31 |
+
if FONT_PATH is None:
|
32 |
+
if not os.path.exists("SimSun.ttf"):
|
33 |
+
ttf = requests.get("https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/SimSun.ttf")
|
34 |
+
open("SimSun.ttf", "wb").write(ttf.content)
|
35 |
+
FONT_PATH = "SimSun.ttf"
|
36 |
+
|
37 |
+
PAT_STR = r"""(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+"""
|
38 |
+
ENDOFTEXT = "<|endoftext|>"
|
39 |
+
IMSTART = "<|im_start|>"
|
40 |
+
IMEND = "<|im_end|>"
|
41 |
+
# as the default behavior is changed to allow special tokens in
|
42 |
+
# regular texts, the surface forms of special tokens need to be
|
43 |
+
# as different as possible to minimize the impact
|
44 |
+
EXTRAS = tuple((f"<|extra_{i}|>" for i in range(205)))
|
45 |
+
SPECIAL_TOKENS = (
|
46 |
+
ENDOFTEXT,
|
47 |
+
IMSTART,
|
48 |
+
IMEND,
|
49 |
+
) + EXTRAS
|
50 |
+
IMG_TOKEN_SPAN = 256
|
51 |
+
|
52 |
+
|
53 |
+
def _load_tiktoken_bpe(tiktoken_bpe_file: str) -> Dict[bytes, int]:
|
54 |
+
with open(tiktoken_bpe_file, "rb") as f:
|
55 |
+
contents = f.read()
|
56 |
+
return {
|
57 |
+
base64.b64decode(token): int(rank)
|
58 |
+
for token, rank in (line.split() for line in contents.splitlines() if line)
|
59 |
+
}
|
60 |
+
|
61 |
+
def _list_find(
|
62 |
+
input_list: List[Any],
|
63 |
+
candidates: Tuple[Any],
|
64 |
+
start: int = 0,
|
65 |
+
):
|
66 |
+
for i in range(start, len(input_list)):
|
67 |
+
if input_list[i] in candidates:
|
68 |
+
return i
|
69 |
+
return -1
|
70 |
+
|
71 |
+
def _replace_closed_tag(
|
72 |
+
input_tokens: List[Any],
|
73 |
+
start_tags: Union[Any, Tuple[Any]],
|
74 |
+
end_tags: Union[Any, Tuple[Any]],
|
75 |
+
inclusive_replace_func: Callable,
|
76 |
+
exclusive_replace_func: Callable = lambda x: x,
|
77 |
+
):
|
78 |
+
if isinstance(start_tags, (str, int)):
|
79 |
+
start_tags = (start_tags,)
|
80 |
+
if isinstance(end_tags, (str, int)):
|
81 |
+
end_tags = (end_tags,)
|
82 |
+
assert len(start_tags) == len(end_tags)
|
83 |
+
|
84 |
+
output_tokens = []
|
85 |
+
end = 0
|
86 |
+
while True:
|
87 |
+
start = _list_find(input_tokens, start_tags, end)
|
88 |
+
if start == -1:
|
89 |
+
break
|
90 |
+
output_tokens.extend(exclusive_replace_func(input_tokens[end : start]))
|
91 |
+
tag_idx = start_tags.index(input_tokens[start])
|
92 |
+
end = _list_find(input_tokens, (end_tags[tag_idx],), start)
|
93 |
+
if end == -1:
|
94 |
+
raise ValueError("Unclosed image token")
|
95 |
+
output_tokens.extend(inclusive_replace_func(input_tokens[start : end + 1]))
|
96 |
+
end += 1
|
97 |
+
output_tokens.extend(exclusive_replace_func(input_tokens[end : ]))
|
98 |
+
return output_tokens
|
99 |
+
|
100 |
+
class QWenTokenizer(PreTrainedTokenizer):
|
101 |
+
"""QWen tokenizer."""
|
102 |
+
|
103 |
+
vocab_files_names = VOCAB_FILES_NAMES
|
104 |
+
|
105 |
+
def __init__(
|
106 |
+
self,
|
107 |
+
vocab_file,
|
108 |
+
errors="replace",
|
109 |
+
image_start_tag='<img>',
|
110 |
+
image_end_tag='</img>',
|
111 |
+
image_pad_tag='<imgpad>',
|
112 |
+
ref_start_tag='<ref>',
|
113 |
+
ref_end_tag='</ref>',
|
114 |
+
box_start_tag='<box>',
|
115 |
+
box_end_tag='</box>',
|
116 |
+
quad_start_tag='<quad>',
|
117 |
+
quad_end_tag='</quad>',
|
118 |
+
**kwargs,
|
119 |
+
):
|
120 |
+
super().__init__(**kwargs)
|
121 |
+
self.image_start_tag = image_start_tag
|
122 |
+
self.image_end_tag = image_end_tag
|
123 |
+
self.image_pad_tag = image_pad_tag
|
124 |
+
self.ref_start_tag = ref_start_tag
|
125 |
+
self.ref_end_tag = ref_end_tag
|
126 |
+
self.box_start_tag = box_start_tag
|
127 |
+
self.box_end_tag = box_end_tag
|
128 |
+
self.quad_start_tag = quad_start_tag
|
129 |
+
self.quad_end_tag = quad_end_tag
|
130 |
+
self.IMAGE_ST = (
|
131 |
+
ref_start_tag, ref_end_tag,
|
132 |
+
box_start_tag, box_end_tag,
|
133 |
+
quad_start_tag, quad_end_tag,
|
134 |
+
image_start_tag, image_end_tag,
|
135 |
+
image_pad_tag
|
136 |
+
)
|
137 |
+
|
138 |
+
self.errors = errors # how to handle errors in decoding
|
139 |
+
|
140 |
+
self.mergeable_ranks = _load_tiktoken_bpe(vocab_file) # type: dict[bytes, int]
|
141 |
+
self.special_tokens = {
|
142 |
+
token: index
|
143 |
+
for index, token in enumerate(
|
144 |
+
SPECIAL_TOKENS + self.IMAGE_ST, start=len(self.mergeable_ranks)
|
145 |
+
)
|
146 |
+
}
|
147 |
+
self.img_start_id = self.special_tokens[self.image_start_tag]
|
148 |
+
self.img_end_id = self.special_tokens[self.image_end_tag]
|
149 |
+
self.img_pad_id = self.special_tokens[self.image_pad_tag]
|
150 |
+
self.ref_start_id = self.special_tokens[self.ref_start_tag]
|
151 |
+
self.ref_end_id = self.special_tokens[self.ref_end_tag]
|
152 |
+
self.box_start_id = self.special_tokens[self.box_start_tag]
|
153 |
+
self.box_end_id = self.special_tokens[self.box_end_tag]
|
154 |
+
self.quad_start_id = self.special_tokens[self.quad_start_tag]
|
155 |
+
self.quad_end_id = self.special_tokens[self.quad_end_tag]
|
156 |
+
self.image_special_tokens = set([
|
157 |
+
self.ref_start_id, self.ref_end_id, self.box_start_id, self.box_end_id,
|
158 |
+
self.quad_start_id, self.quad_end_id,
|
159 |
+
])
|
160 |
+
|
161 |
+
enc = tiktoken.Encoding(
|
162 |
+
"Qwen",
|
163 |
+
pat_str=PAT_STR,
|
164 |
+
mergeable_ranks=self.mergeable_ranks,
|
165 |
+
special_tokens=self.special_tokens,
|
166 |
+
)
|
167 |
+
assert (
|
168 |
+
len(self.mergeable_ranks) + len(self.special_tokens) == enc.n_vocab
|
169 |
+
), f"{len(self.mergeable_ranks) + len(self.special_tokens)} != {enc.n_vocab} in encoding"
|
170 |
+
|
171 |
+
self.decoder = {
|
172 |
+
v: k for k, v in self.mergeable_ranks.items()
|
173 |
+
} # type: dict[int, bytes|str]
|
174 |
+
self.decoder.update({v: k for k, v in self.special_tokens.items()})
|
175 |
+
|
176 |
+
self.tokenizer = enc # type: tiktoken.Encoding
|
177 |
+
|
178 |
+
self.eod_id = self.tokenizer.eot_token
|
179 |
+
self.im_start_id = self.special_tokens[IMSTART]
|
180 |
+
self.im_end_id = self.special_tokens[IMEND]
|
181 |
+
|
182 |
+
def __getstate__(self):
|
183 |
+
# for pickle lovers
|
184 |
+
state = self.__dict__.copy()
|
185 |
+
del state['tokenizer']
|
186 |
+
return state
|
187 |
+
|
188 |
+
def __setstate__(self, state):
|
189 |
+
# tokenizer is not python native; don't pass it; rebuild it
|
190 |
+
self.__dict__.update(state)
|
191 |
+
enc = tiktoken.Encoding(
|
192 |
+
"Qwen",
|
193 |
+
pat_str=PAT_STR,
|
194 |
+
mergeable_ranks=self.mergeable_ranks,
|
195 |
+
special_tokens=self.special_tokens,
|
196 |
+
)
|
197 |
+
self.tokenizer = enc
|
198 |
+
|
199 |
+
|
200 |
+
def __len__(self) -> int:
|
201 |
+
return self.tokenizer.n_vocab
|
202 |
+
|
203 |
+
def get_vocab(self) -> Dict[bytes, int]:
|
204 |
+
return self.mergeable_ranks
|
205 |
+
|
206 |
+
def convert_tokens_to_ids(
|
207 |
+
self, tokens: Union[bytes, str, List[Union[bytes, str]]]
|
208 |
+
) -> List[int]:
|
209 |
+
ids = []
|
210 |
+
if isinstance(tokens, (str, bytes)):
|
211 |
+
if tokens in self.special_tokens:
|
212 |
+
return self.special_tokens[tokens]
|
213 |
+
else:
|
214 |
+
return self.mergeable_ranks.get(tokens)
|
215 |
+
for token in tokens:
|
216 |
+
if token in self.special_tokens:
|
217 |
+
ids.append(self.special_tokens[token])
|
218 |
+
else:
|
219 |
+
ids.append(self.mergeable_ranks.get(token))
|
220 |
+
return ids
|
221 |
+
|
222 |
+
def _add_tokens(self, new_tokens: Union[List[str], List[AddedToken]], special_tokens: bool = False) -> int:
|
223 |
+
if not special_tokens and new_tokens:
|
224 |
+
raise ValueError('Adding regular tokens is not supported')
|
225 |
+
for token in new_tokens:
|
226 |
+
surface_form = token.content if isinstance(token, AddedToken) else token
|
227 |
+
if surface_form not in SPECIAL_TOKENS + self.IMAGE_ST:
|
228 |
+
raise ValueError('Adding unknown special tokens is not supported')
|
229 |
+
return 0
|
230 |
+
|
231 |
+
def save_vocabulary(self, save_directory: str, **kwargs) -> Tuple[str]:
|
232 |
+
"""
|
233 |
+
Save only the vocabulary of the tokenizer (vocabulary).
|
234 |
+
|
235 |
+
Returns:
|
236 |
+
`Tuple(str)`: Paths to the files saved.
|
237 |
+
"""
|
238 |
+
file_path = os.path.join(save_directory, "qwen.tiktoken")
|
239 |
+
with open(file_path, "w", encoding="utf8") as w:
|
240 |
+
for k, v in self.mergeable_ranks.items():
|
241 |
+
line = base64.b64encode(k).decode("utf8") + " " + str(v) + "\n"
|
242 |
+
w.write(line)
|
243 |
+
return (file_path,)
|
244 |
+
|
245 |
+
def tokenize(
|
246 |
+
self,
|
247 |
+
text: str,
|
248 |
+
allowed_special: Union[Set, str] = "all",
|
249 |
+
disallowed_special: Union[Collection, str] = (),
|
250 |
+
**kwargs,
|
251 |
+
) -> List[Union[bytes, str]]:
|
252 |
+
"""
|
253 |
+
Converts a string in a sequence of tokens.
|
254 |
+
|
255 |
+
Args:
|
256 |
+
text (`str`):
|
257 |
+
The sequence to be encoded.
|
258 |
+
allowed_special (`Literal["all"]` or `set`):
|
259 |
+
The surface forms of the tokens to be encoded as special tokens in regular texts.
|
260 |
+
Default to "all".
|
261 |
+
disallowed_special (`Literal["all"]` or `Collection`):
|
262 |
+
The surface forms of the tokens that should not be in regular texts and trigger errors.
|
263 |
+
Default to an empty tuple.
|
264 |
+
|
265 |
+
kwargs (additional keyword arguments, *optional*):
|
266 |
+
Will be passed to the underlying model specific encode method.
|
267 |
+
|
268 |
+
Returns:
|
269 |
+
`List[bytes|str]`: The list of tokens.
|
270 |
+
"""
|
271 |
+
tokens = []
|
272 |
+
text = unicodedata.normalize("NFC", text)
|
273 |
+
|
274 |
+
# this implementation takes a detour: text -> token id -> token surface forms
|
275 |
+
for t in self.tokenizer.encode(
|
276 |
+
text, allowed_special=allowed_special, disallowed_special=disallowed_special
|
277 |
+
):
|
278 |
+
tokens.append(self.decoder[t])
|
279 |
+
|
280 |
+
def _encode_imgurl(img_tokens):
|
281 |
+
assert img_tokens[0] == self.image_start_tag and img_tokens[-1] == self.image_end_tag
|
282 |
+
img_tokens = img_tokens[1:-1]
|
283 |
+
img_url = b''.join(img_tokens)
|
284 |
+
out_img_tokens = list(map(self.decoder.get, img_url))
|
285 |
+
if len(out_img_tokens) > IMG_TOKEN_SPAN:
|
286 |
+
raise ValueError("The content in {}..{} is too long".format(
|
287 |
+
self.image_start_tag, self.image_end_tag))
|
288 |
+
out_img_tokens.extend([self.image_pad_tag] * (IMG_TOKEN_SPAN - len(out_img_tokens)))
|
289 |
+
out_img_tokens = [self.image_start_tag] + out_img_tokens + [self.image_end_tag]
|
290 |
+
return out_img_tokens
|
291 |
+
|
292 |
+
return _replace_closed_tag(tokens, self.image_start_tag, self.image_end_tag, _encode_imgurl)
|
293 |
+
|
294 |
+
def convert_tokens_to_string(self, tokens: List[Union[bytes, str]]) -> str:
|
295 |
+
"""
|
296 |
+
Converts a sequence of tokens in a single string.
|
297 |
+
"""
|
298 |
+
text = ""
|
299 |
+
temp = b""
|
300 |
+
for t in tokens:
|
301 |
+
if isinstance(t, str):
|
302 |
+
if temp:
|
303 |
+
text += temp.decode("utf-8", errors=self.errors)
|
304 |
+
temp = b""
|
305 |
+
text += t
|
306 |
+
elif isinstance(t, bytes):
|
307 |
+
temp += t
|
308 |
+
else:
|
309 |
+
raise TypeError("token should only be of type types or str")
|
310 |
+
if temp:
|
311 |
+
text += temp.decode("utf-8", errors=self.errors)
|
312 |
+
return text
|
313 |
+
|
314 |
+
@property
|
315 |
+
def vocab_size(self):
|
316 |
+
return self.tokenizer.n_vocab
|
317 |
+
|
318 |
+
def _convert_id_to_token(self, index: int) -> Union[bytes, str]:
|
319 |
+
"""Converts an id to a token, special tokens included"""
|
320 |
+
if index in self.decoder:
|
321 |
+
return self.decoder[index]
|
322 |
+
raise ValueError("unknown ids")
|
323 |
+
|
324 |
+
def _convert_token_to_id(self, token: Union[bytes, str]) -> int:
|
325 |
+
"""Converts a token to an id using the vocab, special tokens included"""
|
326 |
+
if token in self.special_tokens:
|
327 |
+
return self.special_tokens[token]
|
328 |
+
if token in self.mergeable_ranks:
|
329 |
+
return self.mergeable_ranks[token]
|
330 |
+
raise ValueError("unknown token")
|
331 |
+
|
332 |
+
def _tokenize(self, text: str, **kwargs):
|
333 |
+
"""
|
334 |
+
Converts a string in a sequence of tokens (string), using the tokenizer. Split in words for word-based
|
335 |
+
vocabulary or sub-words for sub-word-based vocabularies (BPE/SentencePieces/WordPieces).
|
336 |
+
|
337 |
+
Do NOT take care of added tokens.
|
338 |
+
"""
|
339 |
+
raise NotImplementedError
|
340 |
+
|
341 |
+
def _decode(
|
342 |
+
self,
|
343 |
+
token_ids: Union[int, List[int]],
|
344 |
+
skip_special_tokens: bool = False,
|
345 |
+
errors: str = None,
|
346 |
+
**kwargs,
|
347 |
+
) -> str:
|
348 |
+
if isinstance(token_ids, int):
|
349 |
+
token_ids = [token_ids]
|
350 |
+
|
351 |
+
def _decode_imgurl(img_token_ids):
|
352 |
+
assert img_token_ids[0] == self.img_start_id and img_token_ids[-1] == self.img_end_id
|
353 |
+
img_token_ids = img_token_ids[1:-1]
|
354 |
+
img_token_ids = img_token_ids[ : img_token_ids.index(self.img_pad_id)]
|
355 |
+
img_url = bytes(img_token_ids).decode('utf-8')
|
356 |
+
return [self.img_start_id] + self.tokenizer.encode(img_url) + [self.img_end_id]
|
357 |
+
|
358 |
+
token_ids = _replace_closed_tag(token_ids, self.img_start_id, self.img_end_id, _decode_imgurl)
|
359 |
+
|
360 |
+
if skip_special_tokens:
|
361 |
+
if kwargs.get('keep_image_special', False):
|
362 |
+
token_ids = [i for i in token_ids if i < self.eod_id
|
363 |
+
or i in self.image_special_tokens]
|
364 |
+
else:
|
365 |
+
token_ids = [i for i in token_ids if i < self.eod_id]
|
366 |
+
return self.tokenizer.decode(token_ids, errors=errors or self.errors)
|
367 |
+
|
368 |
+
def to_list_format(self, text: str):
|
369 |
+
text = unicodedata.normalize("NFC", text)
|
370 |
+
token_ids = self.tokenizer.encode(
|
371 |
+
text, allowed_special=set(self.IMAGE_ST + (ENDOFTEXT,)))
|
372 |
+
|
373 |
+
def _encode_vl_info(tokens):
|
374 |
+
if len(tokens) == 0:
|
375 |
+
return []
|
376 |
+
if tokens[0] == self.img_start_id and tokens[-1] == self.img_end_id:
|
377 |
+
key = 'image'
|
378 |
+
elif tokens[0] == self.ref_start_id and tokens[-1] == self.ref_end_id:
|
379 |
+
key = 'ref'
|
380 |
+
elif tokens[0] == self.box_start_id and tokens[-1] == self.box_end_id:
|
381 |
+
key = 'box'
|
382 |
+
elif tokens[0] == self.quad_start_id and tokens[-1] == self.quad_end_id:
|
383 |
+
key = 'quad'
|
384 |
+
else:
|
385 |
+
_tobytes = lambda x: x.encode('utf-8') if isinstance(x, str) else x
|
386 |
+
return [{'text': b''.join(map(_tobytes, map(self.decoder.get, tokens))).decode('utf-8')}]
|
387 |
+
_tobytes = lambda x: x.encode('utf-8') if isinstance(x, str) else x
|
388 |
+
val = b''.join(map(_tobytes, map(self.decoder.get, tokens[1:-1]))).decode('utf-8')
|
389 |
+
return [{key: val}]
|
390 |
+
|
391 |
+
return _replace_closed_tag(
|
392 |
+
token_ids,
|
393 |
+
(self.img_start_id, self.ref_start_id, self.box_start_id, self.quad_start_id),
|
394 |
+
(self.img_end_id, self.ref_end_id, self.box_end_id, self.quad_end_id),
|
395 |
+
_encode_vl_info,
|
396 |
+
_encode_vl_info,
|
397 |
+
)
|
398 |
+
|
399 |
+
def from_list_format(self, list_format: List[Dict]):
|
400 |
+
text = ''
|
401 |
+
num_images = 0
|
402 |
+
for ele in list_format:
|
403 |
+
if 'image' in ele:
|
404 |
+
num_images += 1
|
405 |
+
text += f'Picture {num_images}: '
|
406 |
+
text += self.image_start_tag + ele['image'] + self.image_end_tag
|
407 |
+
text += '\n'
|
408 |
+
elif 'text' in ele:
|
409 |
+
text += ele['text']
|
410 |
+
elif 'box' in ele:
|
411 |
+
if 'ref' in ele:
|
412 |
+
text += self.ref_start_tag + ele['ref'] + self.ref_end_tag
|
413 |
+
for box in ele['box']:
|
414 |
+
text += self.box_start_tag + '(%d,%d),(%d,%d)' % (box[0], box[1], box[2], box[3]) + self.box_end_tag
|
415 |
+
else:
|
416 |
+
raise ValueError("Unsupport element: " + str(ele))
|
417 |
+
return text
|
418 |
+
|
419 |
+
def _fetch_latest_picture(self, response, history):
|
420 |
+
if history is None:
|
421 |
+
history = []
|
422 |
+
_history = history + [(response, None)]
|
423 |
+
for q, r in _history[::-1]:
|
424 |
+
for ele in self.to_list_format(q)[::-1]:
|
425 |
+
if 'image' in ele:
|
426 |
+
return ele['image']
|
427 |
+
return None
|
428 |
+
|
429 |
+
def _fetch_all_box_with_ref(self, text):
|
430 |
+
list_format = self.to_list_format(text)
|
431 |
+
output = []
|
432 |
+
for i, ele in enumerate(list_format):
|
433 |
+
if 'box' in ele:
|
434 |
+
bbox = tuple(map(int, ele['box'].replace('(', '').replace(')', '').split(',')))
|
435 |
+
assert len(bbox) == 4
|
436 |
+
output.append({'box': bbox})
|
437 |
+
if i > 0 and 'ref' in list_format[i-1]:
|
438 |
+
output[-1]['ref'] = list_format[i-1]['ref'].strip()
|
439 |
+
return output
|
440 |
+
|
441 |
+
def draw_bbox_on_latest_picture(
|
442 |
+
self,
|
443 |
+
response,
|
444 |
+
history=None,
|
445 |
+
) -> Optional[Image.Image]:
|
446 |
+
image = self._fetch_latest_picture(response, history)
|
447 |
+
if image is None:
|
448 |
+
return None
|
449 |
+
if image.startswith("http://") or image.startswith("https://"):
|
450 |
+
image = Image.open(requests.get(image, stream=True).raw).convert("RGB")
|
451 |
+
h, w = image.height, image.width
|
452 |
+
else:
|
453 |
+
image = np.asarray(Image.open(image).convert("RGB"))
|
454 |
+
h, w = image.shape[0], image.shape[1]
|
455 |
+
visualizer = Visualizer(image)
|
456 |
+
|
457 |
+
boxes = self._fetch_all_box_with_ref(response)
|
458 |
+
if not boxes:
|
459 |
+
return None
|
460 |
+
color = random.choice([_ for _ in mcolors.TABLEAU_COLORS.keys()]) # init color
|
461 |
+
for box in boxes:
|
462 |
+
if 'ref' in box: # random new color for new refexps
|
463 |
+
color = random.choice([_ for _ in mcolors.TABLEAU_COLORS.keys()])
|
464 |
+
x1, y1, x2, y2 = box['box']
|
465 |
+
x1, y1, x2, y2 = (int(x1 / 1000 * w), int(y1 / 1000 * h), int(x2 / 1000 * w), int(y2 / 1000 * h))
|
466 |
+
visualizer.draw_box((x1, y1, x2, y2), alpha=1, edge_color=color)
|
467 |
+
if 'ref' in box:
|
468 |
+
visualizer.draw_text(box['ref'], (x1, y1), color=color, horizontal_alignment="left")
|
469 |
+
return visualizer.output
|
470 |
+
|
471 |
+
|
472 |
+
import colorsys
|
473 |
+
import logging
|
474 |
+
import math
|
475 |
+
import numpy as np
|
476 |
+
import matplotlib as mpl
|
477 |
+
import matplotlib.colors as mplc
|
478 |
+
import matplotlib.figure as mplfigure
|
479 |
+
import torch
|
480 |
+
from matplotlib.backends.backend_agg import FigureCanvasAgg
|
481 |
+
from PIL import Image
|
482 |
+
import random
|
483 |
+
|
484 |
+
logger = logging.getLogger(__name__)
|
485 |
+
|
486 |
+
|
487 |
+
class VisImage:
|
488 |
+
def __init__(self, img, scale=1.0):
|
489 |
+
self.img = img
|
490 |
+
self.scale = scale
|
491 |
+
self.width, self.height = img.shape[1], img.shape[0]
|
492 |
+
self._setup_figure(img)
|
493 |
+
|
494 |
+
def _setup_figure(self, img):
|
495 |
+
fig = mplfigure.Figure(frameon=False)
|
496 |
+
self.dpi = fig.get_dpi()
|
497 |
+
# add a small 1e-2 to avoid precision lost due to matplotlib's truncation
|
498 |
+
# (https://github.com/matplotlib/matplotlib/issues/15363)
|
499 |
+
fig.set_size_inches(
|
500 |
+
(self.width * self.scale + 1e-2) / self.dpi,
|
501 |
+
(self.height * self.scale + 1e-2) / self.dpi,
|
502 |
+
)
|
503 |
+
self.canvas = FigureCanvasAgg(fig)
|
504 |
+
# self.canvas = mpl.backends.backend_cairo.FigureCanvasCairo(fig)
|
505 |
+
ax = fig.add_axes([0.0, 0.0, 1.0, 1.0])
|
506 |
+
ax.axis("off")
|
507 |
+
self.fig = fig
|
508 |
+
self.ax = ax
|
509 |
+
self.reset_image(img)
|
510 |
+
|
511 |
+
def reset_image(self, img):
|
512 |
+
img = img.astype("uint8")
|
513 |
+
self.ax.imshow(img, extent=(0, self.width, self.height, 0), interpolation="nearest")
|
514 |
+
|
515 |
+
def save(self, filepath):
|
516 |
+
self.fig.savefig(filepath)
|
517 |
+
|
518 |
+
def get_image(self):
|
519 |
+
canvas = self.canvas
|
520 |
+
s, (width, height) = canvas.print_to_buffer()
|
521 |
+
|
522 |
+
buffer = np.frombuffer(s, dtype="uint8")
|
523 |
+
|
524 |
+
img_rgba = buffer.reshape(height, width, 4)
|
525 |
+
rgb, alpha = np.split(img_rgba, [3], axis=2)
|
526 |
+
return rgb.astype("uint8")
|
527 |
+
|
528 |
+
|
529 |
+
class Visualizer:
|
530 |
+
def __init__(self, img_rgb, metadata=None, scale=1.0):
|
531 |
+
self.img = np.asarray(img_rgb).clip(0, 255).astype(np.uint8)
|
532 |
+
self.font_path = FONT_PATH
|
533 |
+
self.output = VisImage(self.img, scale=scale)
|
534 |
+
self.cpu_device = torch.device("cpu")
|
535 |
+
|
536 |
+
# too small texts are useless, therefore clamp to 14
|
537 |
+
self._default_font_size = max(
|
538 |
+
np.sqrt(self.output.height * self.output.width) // 30, 15 // scale
|
539 |
+
)
|
540 |
+
|
541 |
+
def draw_text(
|
542 |
+
self,
|
543 |
+
text,
|
544 |
+
position,
|
545 |
+
*,
|
546 |
+
font_size=None,
|
547 |
+
color="g",
|
548 |
+
horizontal_alignment="center",
|
549 |
+
rotation=0,
|
550 |
+
):
|
551 |
+
if not font_size:
|
552 |
+
font_size = self._default_font_size
|
553 |
+
|
554 |
+
# since the text background is dark, we don't want the text to be dark
|
555 |
+
color = np.maximum(list(mplc.to_rgb(color)), 0.2)
|
556 |
+
color[np.argmax(color)] = max(0.8, np.max(color))
|
557 |
+
|
558 |
+
x, y = position
|
559 |
+
self.output.ax.text(
|
560 |
+
x,
|
561 |
+
y,
|
562 |
+
text,
|
563 |
+
size=font_size * self.output.scale,
|
564 |
+
fontproperties=FontProperties(fname=self.font_path),
|
565 |
+
bbox={"facecolor": "black", "alpha": 0.8, "pad": 0.7, "edgecolor": "none"},
|
566 |
+
verticalalignment="top",
|
567 |
+
horizontalalignment=horizontal_alignment,
|
568 |
+
color=color,
|
569 |
+
zorder=10,
|
570 |
+
rotation=rotation,
|
571 |
+
)
|
572 |
+
return self.output
|
573 |
+
|
574 |
+
def draw_box(self, box_coord, alpha=0.5, edge_color="g", line_style="-"):
|
575 |
+
|
576 |
+
x0, y0, x1, y1 = box_coord
|
577 |
+
width = x1 - x0
|
578 |
+
height = y1 - y0
|
579 |
+
|
580 |
+
linewidth = max(self._default_font_size / 4, 1)
|
581 |
+
|
582 |
+
self.output.ax.add_patch(
|
583 |
+
mpl.patches.Rectangle(
|
584 |
+
(x0, y0),
|
585 |
+
width,
|
586 |
+
height,
|
587 |
+
fill=False,
|
588 |
+
edgecolor=edge_color,
|
589 |
+
linewidth=linewidth * self.output.scale,
|
590 |
+
alpha=alpha,
|
591 |
+
linestyle=line_style,
|
592 |
+
)
|
593 |
+
)
|
594 |
+
return self.output
|
595 |
+
|
596 |
+
def get_output(self):
|
597 |
+
|
598 |
+
return self.output
|
tokenizer_config.json
ADDED
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"auto_map": {
|
3 |
+
"AutoTokenizer": [
|
4 |
+
"tokenization_qwen.QWenTokenizer",
|
5 |
+
null
|
6 |
+
]
|
7 |
+
},
|
8 |
+
"clean_up_tokenization_spaces": true,
|
9 |
+
"model_max_length": 800,
|
10 |
+
"padding_side": "right",
|
11 |
+
"tokenizer_class": "QWenTokenizer"
|
12 |
+
}
|
zero_to_fp32.py
ADDED
@@ -0,0 +1,587 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python
|
2 |
+
|
3 |
+
# Copyright (c) Microsoft Corporation.
|
4 |
+
# SPDX-License-Identifier: Apache-2.0
|
5 |
+
|
6 |
+
# DeepSpeed Team
|
7 |
+
|
8 |
+
# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
|
9 |
+
# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
|
10 |
+
# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
|
11 |
+
# application.
|
12 |
+
#
|
13 |
+
# example: python zero_to_fp32.py . pytorch_model.bin
|
14 |
+
|
15 |
+
import argparse
|
16 |
+
import torch
|
17 |
+
import glob
|
18 |
+
import math
|
19 |
+
import os
|
20 |
+
import re
|
21 |
+
from collections import OrderedDict
|
22 |
+
from dataclasses import dataclass
|
23 |
+
|
24 |
+
# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
|
25 |
+
# DeepSpeed data structures it has to be available in the current python environment.
|
26 |
+
from deepspeed.utils import logger
|
27 |
+
from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
|
28 |
+
FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
|
29 |
+
FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
|
30 |
+
|
31 |
+
|
32 |
+
@dataclass
|
33 |
+
class zero_model_state:
|
34 |
+
buffers: dict()
|
35 |
+
param_shapes: dict()
|
36 |
+
shared_params: list
|
37 |
+
ds_version: int
|
38 |
+
frozen_param_shapes: dict()
|
39 |
+
frozen_param_fragments: dict()
|
40 |
+
|
41 |
+
|
42 |
+
debug = 0
|
43 |
+
|
44 |
+
# load to cpu
|
45 |
+
device = torch.device('cpu')
|
46 |
+
|
47 |
+
|
48 |
+
def atoi(text):
|
49 |
+
return int(text) if text.isdigit() else text
|
50 |
+
|
51 |
+
|
52 |
+
def natural_keys(text):
|
53 |
+
'''
|
54 |
+
alist.sort(key=natural_keys) sorts in human order
|
55 |
+
http://nedbatchelder.com/blog/200712/human_sorting.html
|
56 |
+
(See Toothy's implementation in the comments)
|
57 |
+
'''
|
58 |
+
return [atoi(c) for c in re.split(r'(\d+)', text)]
|
59 |
+
|
60 |
+
|
61 |
+
def get_model_state_file(checkpoint_dir, zero_stage):
|
62 |
+
if not os.path.isdir(checkpoint_dir):
|
63 |
+
raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
|
64 |
+
|
65 |
+
# there should be only one file
|
66 |
+
if zero_stage <= 2:
|
67 |
+
file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
|
68 |
+
elif zero_stage == 3:
|
69 |
+
file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
|
70 |
+
|
71 |
+
if not os.path.exists(file):
|
72 |
+
raise FileNotFoundError(f"can't find model states file at '{file}'")
|
73 |
+
|
74 |
+
return file
|
75 |
+
|
76 |
+
|
77 |
+
def get_checkpoint_files(checkpoint_dir, glob_pattern):
|
78 |
+
# XXX: need to test that this simple glob rule works for multi-node setup too
|
79 |
+
ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
|
80 |
+
|
81 |
+
if len(ckpt_files) == 0:
|
82 |
+
raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
|
83 |
+
|
84 |
+
return ckpt_files
|
85 |
+
|
86 |
+
|
87 |
+
def get_optim_files(checkpoint_dir):
|
88 |
+
return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
|
89 |
+
|
90 |
+
|
91 |
+
def get_model_state_files(checkpoint_dir):
|
92 |
+
return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
|
93 |
+
|
94 |
+
|
95 |
+
def parse_model_states(files):
|
96 |
+
zero_model_states = []
|
97 |
+
for file in files:
|
98 |
+
state_dict = torch.load(file, map_location=device)
|
99 |
+
|
100 |
+
if BUFFER_NAMES not in state_dict:
|
101 |
+
raise ValueError(f"{file} is not a model state checkpoint")
|
102 |
+
buffer_names = state_dict[BUFFER_NAMES]
|
103 |
+
if debug:
|
104 |
+
print("Found buffers:", buffer_names)
|
105 |
+
|
106 |
+
# recover just the buffers while restoring them to fp32 if they were saved in fp16
|
107 |
+
buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
|
108 |
+
param_shapes = state_dict[PARAM_SHAPES]
|
109 |
+
|
110 |
+
# collect parameters that are included in param_shapes
|
111 |
+
param_names = []
|
112 |
+
for s in param_shapes:
|
113 |
+
for name in s.keys():
|
114 |
+
param_names.append(name)
|
115 |
+
|
116 |
+
# update with frozen parameters
|
117 |
+
frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
|
118 |
+
if frozen_param_shapes is not None:
|
119 |
+
if debug:
|
120 |
+
print(f"Found frozen_param_shapes: {frozen_param_shapes}")
|
121 |
+
param_names += list(frozen_param_shapes.keys())
|
122 |
+
|
123 |
+
# handle shared params
|
124 |
+
shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
|
125 |
+
|
126 |
+
ds_version = state_dict.get(DS_VERSION, None)
|
127 |
+
|
128 |
+
frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
|
129 |
+
|
130 |
+
z_model_state = zero_model_state(buffers=buffers,
|
131 |
+
param_shapes=param_shapes,
|
132 |
+
shared_params=shared_params,
|
133 |
+
ds_version=ds_version,
|
134 |
+
frozen_param_shapes=frozen_param_shapes,
|
135 |
+
frozen_param_fragments=frozen_param_fragments)
|
136 |
+
zero_model_states.append(z_model_state)
|
137 |
+
|
138 |
+
return zero_model_states
|
139 |
+
|
140 |
+
|
141 |
+
def parse_optim_states(files, ds_checkpoint_dir):
|
142 |
+
|
143 |
+
total_files = len(files)
|
144 |
+
state_dicts = []
|
145 |
+
for f in files:
|
146 |
+
state_dict = torch.load(f, map_location=device)
|
147 |
+
# immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
|
148 |
+
# and also handle the case where it was already removed by another helper script
|
149 |
+
state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
|
150 |
+
state_dicts.append(state_dict)
|
151 |
+
|
152 |
+
if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
|
153 |
+
raise ValueError(f"{files[0]} is not a zero checkpoint")
|
154 |
+
zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
|
155 |
+
world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
|
156 |
+
|
157 |
+
# For ZeRO-2 each param group can have different partition_count as data parallelism for expert
|
158 |
+
# parameters can be different from data parallelism for non-expert parameters. So we can just
|
159 |
+
# use the max of the partition_count to get the dp world_size.
|
160 |
+
|
161 |
+
if type(world_size) is list:
|
162 |
+
world_size = max(world_size)
|
163 |
+
|
164 |
+
if world_size != total_files:
|
165 |
+
raise ValueError(
|
166 |
+
f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
|
167 |
+
"Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
|
168 |
+
)
|
169 |
+
|
170 |
+
# the groups are named differently in each stage
|
171 |
+
if zero_stage <= 2:
|
172 |
+
fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
|
173 |
+
elif zero_stage == 3:
|
174 |
+
fp32_groups_key = FP32_FLAT_GROUPS
|
175 |
+
else:
|
176 |
+
raise ValueError(f"unknown zero stage {zero_stage}")
|
177 |
+
|
178 |
+
if zero_stage <= 2:
|
179 |
+
fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
|
180 |
+
elif zero_stage == 3:
|
181 |
+
# if there is more than one param group, there will be multiple flattened tensors - one
|
182 |
+
# flattened tensor per group - for simplicity merge them into a single tensor
|
183 |
+
#
|
184 |
+
# XXX: could make the script more memory efficient for when there are multiple groups - it
|
185 |
+
# will require matching the sub-lists of param_shapes for each param group flattened tensor
|
186 |
+
|
187 |
+
fp32_flat_groups = [
|
188 |
+
torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
|
189 |
+
]
|
190 |
+
|
191 |
+
return zero_stage, world_size, fp32_flat_groups
|
192 |
+
|
193 |
+
|
194 |
+
def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
|
195 |
+
"""
|
196 |
+
Returns fp32 state_dict reconstructed from ds checkpoint
|
197 |
+
|
198 |
+
Args:
|
199 |
+
- ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
|
200 |
+
|
201 |
+
"""
|
202 |
+
print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
|
203 |
+
|
204 |
+
optim_files = get_optim_files(ds_checkpoint_dir)
|
205 |
+
zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
|
206 |
+
print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
|
207 |
+
|
208 |
+
model_files = get_model_state_files(ds_checkpoint_dir)
|
209 |
+
|
210 |
+
zero_model_states = parse_model_states(model_files)
|
211 |
+
print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
|
212 |
+
|
213 |
+
if zero_stage <= 2:
|
214 |
+
return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
|
215 |
+
elif zero_stage == 3:
|
216 |
+
return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
|
217 |
+
|
218 |
+
|
219 |
+
def _zero2_merge_frozen_params(state_dict, zero_model_states):
|
220 |
+
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
221 |
+
return
|
222 |
+
|
223 |
+
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
224 |
+
frozen_param_fragments = zero_model_states[0].frozen_param_fragments
|
225 |
+
|
226 |
+
if debug:
|
227 |
+
num_elem = sum(s.numel() for s in frozen_param_shapes.values())
|
228 |
+
print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
229 |
+
|
230 |
+
wanted_params = len(frozen_param_shapes)
|
231 |
+
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
232 |
+
avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
|
233 |
+
print(f'Frozen params: Have {avail_numel} numels to process.')
|
234 |
+
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
235 |
+
|
236 |
+
total_params = 0
|
237 |
+
total_numel = 0
|
238 |
+
for name, shape in frozen_param_shapes.items():
|
239 |
+
total_params += 1
|
240 |
+
unpartitioned_numel = shape.numel()
|
241 |
+
total_numel += unpartitioned_numel
|
242 |
+
|
243 |
+
state_dict[name] = frozen_param_fragments[name]
|
244 |
+
|
245 |
+
if debug:
|
246 |
+
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
247 |
+
|
248 |
+
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
249 |
+
|
250 |
+
|
251 |
+
def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
252 |
+
param_shapes = zero_model_states[0].param_shapes
|
253 |
+
|
254 |
+
# Reconstruction protocol:
|
255 |
+
#
|
256 |
+
# XXX: document this
|
257 |
+
|
258 |
+
if debug:
|
259 |
+
for i in range(world_size):
|
260 |
+
for j in range(len(fp32_flat_groups[0])):
|
261 |
+
print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
|
262 |
+
|
263 |
+
# XXX: memory usage doubles here (zero2)
|
264 |
+
num_param_groups = len(fp32_flat_groups[0])
|
265 |
+
merged_single_partition_of_fp32_groups = []
|
266 |
+
for i in range(num_param_groups):
|
267 |
+
merged_partitions = [sd[i] for sd in fp32_flat_groups]
|
268 |
+
full_single_fp32_vector = torch.cat(merged_partitions, 0)
|
269 |
+
merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
|
270 |
+
avail_numel = sum(
|
271 |
+
[full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
|
272 |
+
|
273 |
+
if debug:
|
274 |
+
wanted_params = sum([len(shapes) for shapes in param_shapes])
|
275 |
+
wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
|
276 |
+
# not asserting if there is a mismatch due to possible padding
|
277 |
+
print(f"Have {avail_numel} numels to process.")
|
278 |
+
print(f"Need {wanted_numel} numels in {wanted_params} params.")
|
279 |
+
|
280 |
+
# params
|
281 |
+
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
282 |
+
# out-of-core computing solution
|
283 |
+
total_numel = 0
|
284 |
+
total_params = 0
|
285 |
+
for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
|
286 |
+
offset = 0
|
287 |
+
avail_numel = full_single_fp32_vector.numel()
|
288 |
+
for name, shape in shapes.items():
|
289 |
+
|
290 |
+
unpartitioned_numel = shape.numel()
|
291 |
+
total_numel += unpartitioned_numel
|
292 |
+
total_params += 1
|
293 |
+
|
294 |
+
if debug:
|
295 |
+
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
|
296 |
+
state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
|
297 |
+
offset += unpartitioned_numel
|
298 |
+
|
299 |
+
# Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
|
300 |
+
# avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
|
301 |
+
# paddings performed in the code it's almost impossible to predict the exact numbers w/o the
|
302 |
+
# live optimizer object, so we are checking that the numbers are within the right range
|
303 |
+
align_to = 2 * world_size
|
304 |
+
|
305 |
+
def zero2_align(x):
|
306 |
+
return align_to * math.ceil(x / align_to)
|
307 |
+
|
308 |
+
if debug:
|
309 |
+
print(f"original offset={offset}, avail_numel={avail_numel}")
|
310 |
+
|
311 |
+
offset = zero2_align(offset)
|
312 |
+
avail_numel = zero2_align(avail_numel)
|
313 |
+
|
314 |
+
if debug:
|
315 |
+
print(f"aligned offset={offset}, avail_numel={avail_numel}")
|
316 |
+
|
317 |
+
# Sanity check
|
318 |
+
if offset != avail_numel:
|
319 |
+
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
320 |
+
|
321 |
+
print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
|
322 |
+
|
323 |
+
|
324 |
+
def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
|
325 |
+
state_dict = OrderedDict()
|
326 |
+
|
327 |
+
# buffers
|
328 |
+
buffers = zero_model_states[0].buffers
|
329 |
+
state_dict.update(buffers)
|
330 |
+
if debug:
|
331 |
+
print(f"added {len(buffers)} buffers")
|
332 |
+
|
333 |
+
_zero2_merge_frozen_params(state_dict, zero_model_states)
|
334 |
+
|
335 |
+
_zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
336 |
+
|
337 |
+
# recover shared parameters
|
338 |
+
for pair in zero_model_states[0].shared_params:
|
339 |
+
if pair[1] in state_dict:
|
340 |
+
state_dict[pair[0]] = state_dict[pair[1]]
|
341 |
+
|
342 |
+
return state_dict
|
343 |
+
|
344 |
+
|
345 |
+
def zero3_partitioned_param_info(unpartitioned_numel, world_size):
|
346 |
+
remainder = unpartitioned_numel % world_size
|
347 |
+
padding_numel = (world_size - remainder) if remainder else 0
|
348 |
+
partitioned_numel = math.ceil(unpartitioned_numel / world_size)
|
349 |
+
return partitioned_numel, padding_numel
|
350 |
+
|
351 |
+
|
352 |
+
def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
|
353 |
+
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
|
354 |
+
return
|
355 |
+
|
356 |
+
if debug:
|
357 |
+
for i in range(world_size):
|
358 |
+
num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
|
359 |
+
print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
|
360 |
+
|
361 |
+
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
362 |
+
wanted_params = len(frozen_param_shapes)
|
363 |
+
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
364 |
+
avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
|
365 |
+
print(f'Frozen params: Have {avail_numel} numels to process.')
|
366 |
+
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
|
367 |
+
|
368 |
+
total_params = 0
|
369 |
+
total_numel = 0
|
370 |
+
for name, shape in zero_model_states[0].frozen_param_shapes.items():
|
371 |
+
total_params += 1
|
372 |
+
unpartitioned_numel = shape.numel()
|
373 |
+
total_numel += unpartitioned_numel
|
374 |
+
|
375 |
+
param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
|
376 |
+
state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
|
377 |
+
|
378 |
+
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
379 |
+
|
380 |
+
if debug:
|
381 |
+
print(
|
382 |
+
f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
383 |
+
)
|
384 |
+
|
385 |
+
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
|
386 |
+
|
387 |
+
|
388 |
+
def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
|
389 |
+
param_shapes = zero_model_states[0].param_shapes
|
390 |
+
avail_numel = fp32_flat_groups[0].numel() * world_size
|
391 |
+
# Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
|
392 |
+
# param, re-consolidating each param, while dealing with padding if any
|
393 |
+
|
394 |
+
# merge list of dicts, preserving order
|
395 |
+
param_shapes = {k: v for d in param_shapes for k, v in d.items()}
|
396 |
+
|
397 |
+
if debug:
|
398 |
+
for i in range(world_size):
|
399 |
+
print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
|
400 |
+
|
401 |
+
wanted_params = len(param_shapes)
|
402 |
+
wanted_numel = sum(shape.numel() for shape in param_shapes.values())
|
403 |
+
# not asserting if there is a mismatch due to possible padding
|
404 |
+
avail_numel = fp32_flat_groups[0].numel() * world_size
|
405 |
+
print(f"Trainable params: Have {avail_numel} numels to process.")
|
406 |
+
print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
|
407 |
+
|
408 |
+
# params
|
409 |
+
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
|
410 |
+
# out-of-core computing solution
|
411 |
+
offset = 0
|
412 |
+
total_numel = 0
|
413 |
+
total_params = 0
|
414 |
+
for name, shape in param_shapes.items():
|
415 |
+
|
416 |
+
unpartitioned_numel = shape.numel()
|
417 |
+
total_numel += unpartitioned_numel
|
418 |
+
total_params += 1
|
419 |
+
|
420 |
+
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
|
421 |
+
|
422 |
+
if debug:
|
423 |
+
print(
|
424 |
+
f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
|
425 |
+
)
|
426 |
+
|
427 |
+
# XXX: memory usage doubles here
|
428 |
+
state_dict[name] = torch.cat(
|
429 |
+
tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
|
430 |
+
0).narrow(0, 0, unpartitioned_numel).view(shape)
|
431 |
+
offset += partitioned_numel
|
432 |
+
|
433 |
+
offset *= world_size
|
434 |
+
|
435 |
+
# Sanity check
|
436 |
+
if offset != avail_numel:
|
437 |
+
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
|
438 |
+
|
439 |
+
print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
|
440 |
+
|
441 |
+
|
442 |
+
def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
|
443 |
+
state_dict = OrderedDict()
|
444 |
+
|
445 |
+
# buffers
|
446 |
+
buffers = zero_model_states[0].buffers
|
447 |
+
state_dict.update(buffers)
|
448 |
+
if debug:
|
449 |
+
print(f"added {len(buffers)} buffers")
|
450 |
+
|
451 |
+
_zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
|
452 |
+
|
453 |
+
_zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
|
454 |
+
|
455 |
+
# recover shared parameters
|
456 |
+
for pair in zero_model_states[0].shared_params:
|
457 |
+
if pair[1] in state_dict:
|
458 |
+
state_dict[pair[0]] = state_dict[pair[1]]
|
459 |
+
|
460 |
+
return state_dict
|
461 |
+
|
462 |
+
|
463 |
+
def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
|
464 |
+
"""
|
465 |
+
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
|
466 |
+
``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
|
467 |
+
via a model hub.
|
468 |
+
|
469 |
+
Args:
|
470 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder
|
471 |
+
- ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
|
472 |
+
|
473 |
+
Returns:
|
474 |
+
- pytorch ``state_dict``
|
475 |
+
|
476 |
+
Note: this approach may not work if your application doesn't have sufficient free CPU memory and
|
477 |
+
you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
|
478 |
+
the checkpoint.
|
479 |
+
|
480 |
+
A typical usage might be ::
|
481 |
+
|
482 |
+
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
|
483 |
+
# do the training and checkpoint saving
|
484 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
|
485 |
+
model = model.cpu() # move to cpu
|
486 |
+
model.load_state_dict(state_dict)
|
487 |
+
# submit to model hub or save the model to share with others
|
488 |
+
|
489 |
+
In this example the ``model`` will no longer be usable in the deepspeed context of the same
|
490 |
+
application. i.e. you will need to re-initialize the deepspeed engine, since
|
491 |
+
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
492 |
+
|
493 |
+
If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
|
494 |
+
|
495 |
+
"""
|
496 |
+
if tag is None:
|
497 |
+
latest_path = os.path.join(checkpoint_dir, 'latest')
|
498 |
+
if os.path.isfile(latest_path):
|
499 |
+
with open(latest_path, 'r') as fd:
|
500 |
+
tag = fd.read().strip()
|
501 |
+
else:
|
502 |
+
raise ValueError(f"Unable to find 'latest' file at {latest_path}")
|
503 |
+
|
504 |
+
ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
|
505 |
+
|
506 |
+
if not os.path.isdir(ds_checkpoint_dir):
|
507 |
+
raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
|
508 |
+
|
509 |
+
return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
|
510 |
+
|
511 |
+
|
512 |
+
def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
|
513 |
+
"""
|
514 |
+
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
|
515 |
+
loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
|
516 |
+
|
517 |
+
Args:
|
518 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
519 |
+
- ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
|
520 |
+
- ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
|
521 |
+
"""
|
522 |
+
|
523 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
524 |
+
print(f"Saving fp32 state dict to {output_file}")
|
525 |
+
torch.save(state_dict, output_file)
|
526 |
+
|
527 |
+
|
528 |
+
def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
|
529 |
+
"""
|
530 |
+
1. Put the provided model to cpu
|
531 |
+
2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
|
532 |
+
3. Load it into the provided model
|
533 |
+
|
534 |
+
Args:
|
535 |
+
- ``model``: the model object to update
|
536 |
+
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
|
537 |
+
- ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
|
538 |
+
|
539 |
+
Returns:
|
540 |
+
- ``model`: modified model
|
541 |
+
|
542 |
+
Make sure you have plenty of CPU memory available before you call this function. If you don't
|
543 |
+
have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
|
544 |
+
conveniently placed for you in the checkpoint folder.
|
545 |
+
|
546 |
+
A typical usage might be ::
|
547 |
+
|
548 |
+
from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
|
549 |
+
model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
|
550 |
+
# submit to model hub or save the model to share with others
|
551 |
+
|
552 |
+
Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
|
553 |
+
of the same application. i.e. you will need to re-initialize the deepspeed engine, since
|
554 |
+
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
|
555 |
+
|
556 |
+
"""
|
557 |
+
logger.info(f"Extracting fp32 weights")
|
558 |
+
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
|
559 |
+
|
560 |
+
logger.info(f"Overwriting model with fp32 weights")
|
561 |
+
model = model.cpu()
|
562 |
+
model.load_state_dict(state_dict, strict=False)
|
563 |
+
|
564 |
+
return model
|
565 |
+
|
566 |
+
|
567 |
+
if __name__ == "__main__":
|
568 |
+
|
569 |
+
parser = argparse.ArgumentParser()
|
570 |
+
parser.add_argument("checkpoint_dir",
|
571 |
+
type=str,
|
572 |
+
help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
|
573 |
+
parser.add_argument(
|
574 |
+
"output_file",
|
575 |
+
type=str,
|
576 |
+
help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
|
577 |
+
parser.add_argument("-t",
|
578 |
+
"--tag",
|
579 |
+
type=str,
|
580 |
+
default=None,
|
581 |
+
help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
|
582 |
+
parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
|
583 |
+
args = parser.parse_args()
|
584 |
+
|
585 |
+
debug = args.debug
|
586 |
+
|
587 |
+
convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file, tag=args.tag)
|