Text Generation
Transformers
Safetensors
English
stablelm_alpha
causal-lm
custom_code
5 papers
stable-jon-tow commited on
Commit
ecf8036
1 Parent(s): 7fc523f

Upload folder using huggingface_hub

Browse files
README.md ADDED
@@ -0,0 +1,94 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ datasets:
3
+ - tiiuae/falcon-refinedweb
4
+ - togethercomputer/RedPajama-Data-1T
5
+ - CarperAI/pilev2-dev
6
+ - bigcode/starcoderdata
7
+ - JeanKaddour/minipile
8
+ language:
9
+ - en
10
+ tags:
11
+ - causal-lm
12
+ license: cc-by-sa-4.0
13
+ ---
14
+ # `StableLM-Base-Alpha-3B-v2`
15
+
16
+ ## Model Description
17
+
18
+ `StableLM-Base-Alpha-3B-v2` is a 3 billion parameter decoder-only language model pre-trained on diverse English datasets. This model is the successor to the first [`StableLM-Base-Alpha-3B`](https://huggingface.co/stabilityai/stablelm-base-alpha-3b) model, addressing previous shortcomings through the use of improved data sources and mixture ratios.
19
+
20
+ ## Usage
21
+
22
+ Get started generating text with `StableLM-Base-Alpha-3B-v2` by using the following code snippet:
23
+
24
+ ```python
25
+ from transformers import AutoModelForCausalLM, AutoTokenizer
26
+ tokenizer = AutoTokenizer.from_pretrained("stabilityai/stablelm-base-alpha-3b-v2")
27
+ model = AutoModelForCausalLM.from_pretrained(
28
+ "stabilityai/stablelm-base-alpha-3b-v2",
29
+ trust_remote_code=True,
30
+ torch_dtype="auto",
31
+ )
32
+ model.cuda()
33
+ inputs = tokenizer("The weather is always wonderful", return_tensors="pt").to("cuda")
34
+ tokens = model.generate(
35
+ **inputs,
36
+ max_new_tokens=64,
37
+ temperature=0.75,
38
+ top_p=0.95,
39
+ do_sample=True,
40
+ )
41
+ print(tokenizer.decode(tokens[0], skip_special_tokens=True))
42
+ ```
43
+
44
+ ## Model Details
45
+
46
+ * **Developed by**: [Stability AI](https://stability.ai/)
47
+ * **Model type**: `StableLM-Base-Alpha-v2` models are auto-regressive language models based on the transformer decoder architecture.
48
+ * **Language(s)**: English
49
+ * **Library**: [GPT-NeoX](https://github.com/EleutherAI/gpt-neox)
50
+ * **License**: Model checkpoints are licensed under the Creative Commons license ([CC BY-SA-4.0](https://creativecommons.org/licenses/by-sa/4.0/)). Under this license, you must give [credit](https://creativecommons.org/licenses/by/4.0/#) to Stability AI, provide a link to the license, and [indicate if changes were made](https://creativecommons.org/licenses/by/4.0/#). You may do so in any reasonable manner, but not in any way that suggests the Stability AI endorses you or your use.
51
+ * **Contact**: For questions and comments about the model, please email `lm@stability.ai`
52
+
53
+ ### Model Architecture
54
+
55
+ | Parameters | Hidden Size | Layers | Heads | Sequence Length |
56
+ |----------------|-------------|--------|-------|-----------------|
57
+ | 2,796,431,360 | 2560 | 32 | 32 | 4096 |
58
+
59
+ The model is a decoder-only transformer similar to the `StableLM-Base-Alpha` (v1) with the following configurations:
60
+
61
+ * **Activation**: SwiGLU ([Shazeer, 2020](https://arxiv.org/abs/2002.05202))
62
+ * **Decoder Layer**: Parallel Attention and MLP residuals with a single input LayerNorm ([Wang & Komatsuzaki, 2021](https://github.com/kingoflolz/mesh-transformer-jax/tree/master))
63
+ * **Position Embeddings**: Rotary Position Embeddings ([Su et al., 2021](https://arxiv.org/abs/2104.09864))
64
+ * **Bias**: LayerNorm bias terms only
65
+
66
+ ## Training
67
+
68
+ `StableLM-Base-Alpha-3B-v2` is pre-trained using a multi-stage context length extension schedule following similar work ([Nijkamp et al. 2023](https://blog.salesforceairesearch.com/xgen/)); first pre-training at a context length of 2048 for 1 trillion tokens, then fine-tuning at a context length of 4096 for another 100B tokens.
69
+
70
+ ### Training Dataset
71
+
72
+ The first pre-training stage relies on 1 trillion tokens sourced from a mix of the public Falcon RefinedWeb extract ([Penedo et al., 2023](https://huggingface.co/datasets/tiiuae/falcon-refinedweb)), RedPajama-Data ([Together Computer 2023](https://github.com/togethercomputer/RedPajama-Data), The Pile ([Gao et al., 2020](https://arxiv.org/abs/2101.00027)), and internal datasets with web text sampled at a rate of 71%.
73
+
74
+ In the second stage, we include the StarCoder ([Li et al., 2023](https://arxiv.org/abs/2305.06161)) dataset and down sample web text to 55% while increasing sampling proportions of naturally long text examples in the aforementioned sources.
75
+
76
+ ### Training Procedure
77
+
78
+ The model is pre-trained on the dataset mixes mentioned above in mixed-precision (FP16), optimized with AdamW, and trained using the NeoX tokenizer with a vocabulary size of 50,257. We outline the complete hyperparameters choices in the project's [GitHub repository - config](https://github.com/Stability-AI/StableLM/blob/main/configs/stablelm-base-alpha-3b-v2.yaml).
79
+
80
+ ### Training Infrastructure
81
+
82
+ * **Hardware**: `StableLM-Base-Alpha-3B-v2` was trained on the Stability AI cluster - occupying 256 NVIDIA A100 40GB GPUs across AWS P4d instances. Training took approximately 8.45 days to complete across both stages.
83
+
84
+ * **Software**: We use a fork of gpt-neox ([EleutherAI, 2021](https://github.com/EleutherAI/gpt-neox)) and train under 2D parallelism (Data and Tensor Parallel) with ZeRO-1 ([Rajbhandari et al., 2019](https://arxiv.org/abs/1910.02054v3)) and rely on flash-attention as well as rotary embedding kernels from FlashAttention-2 ([Dao et al., 2023](https://tridao.me/publications/flash2/flash2.pdf))
85
+
86
+ ## Use and Limitations
87
+
88
+ ### Intended Use
89
+
90
+ These models are intended to be used by all individuals as foundational models for application-specific fine-tuning without strict limitations on commercial use.
91
+
92
+ ### Limitations and bias
93
+
94
+ The pre-training dataset may have contained offensive or inappropriate content even after applying data cleansing filters which can be reflected in the model-generated text. We recommend that users exercise caution when using these models in production systems. Do not use the models for any applications that may cause harm or distress to individuals or groups.
config.json ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "StableLMAlphaForCausalLM"
4
+ ],
5
+ "auto_map": {
6
+ "AutoConfig": "configuration_stablelm_alpha.StableLMAlphaConfig",
7
+ "AutoModelForCausalLM": "modeling_stablelm_alpha.StableLMAlphaForCausalLM"
8
+ },
9
+ "bos_token_id": 0,
10
+ "eos_token_id": 0,
11
+ "hidden_act": "silu",
12
+ "hidden_size": 2560,
13
+ "initializer_range": 0.02,
14
+ "max_position_embeddings": 4096,
15
+ "model_type": "stablelm_alpha",
16
+ "norm_eps": 1e-05,
17
+ "num_heads": 32,
18
+ "num_hidden_layers": 32,
19
+ "rotary_emb_base": 10000,
20
+ "rotary_pct": 0.25,
21
+ "rotary_scaling_factor": 1.0,
22
+ "tie_word_embeddings": false,
23
+ "torch_dtype": "float32",
24
+ "transformers_version": "4.30.2",
25
+ "use_cache": true,
26
+ "vocab_size": 50432
27
+ }
configuration_stablelm_alpha.py ADDED
@@ -0,0 +1,113 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2023 Stability and The HuggingFace Inc. team. All rights reserved.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+ """ StableLM β model configuration"""
16
+
17
+ from transformers import PretrainedConfig
18
+ from transformers.utils import logging
19
+
20
+
21
+ logger = logging.get_logger(__name__)
22
+
23
+ STABLE_LM_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
24
+
25
+
26
+ class StableLMAlphaConfig(PretrainedConfig):
27
+ r"""
28
+ Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
29
+ documentation from [`PretrainedConfig`] for more information.
30
+
31
+ Args:
32
+ vocab_size (`int`, *optional*, defaults to 50432):
33
+ Vocabulary size of the StableLM model. Defines the number of different tokens that
34
+ can be represented by the `inputs_ids` passed when calling [`StableLMAlphaModel`].
35
+ hidden_size (`int`, *optional*, defaults to 6144):
36
+ Dimension of the decoder layers and the pooler layer.
37
+ num_hidden_layers (`int`, *optional*, defaults to 44):
38
+ Number of hidden layers in the Transformer decoder.
39
+ num_heads (`int`, *optional*, defaults to 64):
40
+ Number of attention heads for each attention layer in the Transformer decoder.
41
+ hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
42
+ The non-linear activation function (function or string).
43
+ rotary_pct (`float`, *optional*, defaults to 0.25):
44
+ Percentage of hidden dimensions to allocate to rotary embeddings.
45
+ rotary_emb_base (`int`, *optional*, defaults to 10000)
46
+ Base for computing rotary embeddings frequency.
47
+ max_position_embeddings (`int`, *optional*, defaults to 2048):
48
+ The maximum sequence length that this model might ever be used with.
49
+ Typically set this to something large just in case (e.g., 512 or 1024 or 2048).
50
+ initializer_range (`float`, *optional*, defaults to 1e-5):
51
+ The standard deviation of the truncated_normal_initializer for initializing
52
+ all weight matrices.
53
+ norm_eps (`float`, *optional*, defaults to 1e-5):
54
+ The epsilon used by the normalization layers.
55
+ use_cache (`bool`, *optional*, defaults to `True`):
56
+ Whether or not the model should return the last key/values attentions
57
+ (not used by all models). Only relevant if `config.is_decoder=True`.
58
+ tie_word_embeddings(`bool`, *optional*, defaults to `False`):
59
+ Whether to tie weight embeddings
60
+
61
+ Example:
62
+
63
+ ```python
64
+ >>> from transformers import StableLMAlphaConfig, StableLMAlphaModel
65
+
66
+ >>> # Initializing a StableLMAlphaConfig style configuration
67
+ >>> configuration = StableLMAlphaConfig()
68
+
69
+ >>> # Initializing a model (with random weights) from the style configuration
70
+ >>> model = StableLMAlphaModel(configuration) # doctest: +SKIP
71
+
72
+ >>> # Accessing the model configuration
73
+ >>> configuration = model.config # doctest: +SKIP
74
+ ```"""
75
+ model_type = "stablelm_alpha"
76
+ keys_to_ignore_at_inference = ["past_key_values"]
77
+
78
+ def __init__(
79
+ self,
80
+ vocab_size=50_432,
81
+ hidden_size=2_560,
82
+ num_hidden_layers=32,
83
+ num_heads=32,
84
+ hidden_act="silu",
85
+ rotary_pct=0.25,
86
+ rotary_emb_base=10_000,
87
+ max_position_embeddings=2_048,
88
+ initializer_range=0.02,
89
+ norm_eps=1e-5,
90
+ use_cache=True,
91
+ bos_token_id=0,
92
+ eos_token_id=2,
93
+ tie_word_embeddings=False,
94
+ **kwargs,
95
+ ):
96
+ self.vocab_size = vocab_size
97
+ self.max_position_embeddings = max_position_embeddings
98
+ self.hidden_size = hidden_size
99
+ self.num_hidden_layers = num_hidden_layers
100
+ self.num_heads = num_heads
101
+ self.hidden_act = hidden_act
102
+ self.rotary_pct = rotary_pct
103
+ self.rotary_emb_base = rotary_emb_base
104
+ self.initializer_range = initializer_range
105
+ self.norm_eps = norm_eps
106
+ self.use_cache = use_cache
107
+ self.tie_word_embeddings = tie_word_embeddings
108
+ super().__init__(
109
+ bos_token_id=bos_token_id,
110
+ eos_token_id=eos_token_id,
111
+ tie_word_embeddings=tie_word_embeddings,
112
+ **kwargs,
113
+ )
generation_config.json ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "bos_token_id": 0,
4
+ "eos_token_id": 0,
5
+ "transformers_version": "4.30.2"
6
+ }
model-00001-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d8c6314059f1e627e6d04474b0dca9a4972c6aa18d4dd1b7d6457300ab656ec0
3
+ size 9962107344
model-00002-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7f7b27de188b94c4fbb7e753dd58ec0ed556cf6aafdf20fc999f57b4643d7f05
3
+ size 1221654304
model.safetensors.index.json ADDED
@@ -0,0 +1,203 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "metadata": {
3
+ "total_size": 11183738880
4
+ },
5
+ "weight_map": {
6
+ "lm_head.weight": "model-00002-of-00002.safetensors",
7
+ "transformer.embed.weight": "model-00001-of-00002.safetensors",
8
+ "transformer.final_norm.bias": "model-00002-of-00002.safetensors",
9
+ "transformer.final_norm.weight": "model-00002-of-00002.safetensors",
10
+ "transformer.layers.0.attention.out_proj.weight": "model-00001-of-00002.safetensors",
11
+ "transformer.layers.0.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
12
+ "transformer.layers.0.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
13
+ "transformer.layers.0.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
14
+ "transformer.layers.0.norm.bias": "model-00001-of-00002.safetensors",
15
+ "transformer.layers.0.norm.weight": "model-00001-of-00002.safetensors",
16
+ "transformer.layers.1.attention.out_proj.weight": "model-00001-of-00002.safetensors",
17
+ "transformer.layers.1.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
18
+ "transformer.layers.1.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
19
+ "transformer.layers.1.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
20
+ "transformer.layers.1.norm.bias": "model-00001-of-00002.safetensors",
21
+ "transformer.layers.1.norm.weight": "model-00001-of-00002.safetensors",
22
+ "transformer.layers.10.attention.out_proj.weight": "model-00001-of-00002.safetensors",
23
+ "transformer.layers.10.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
24
+ "transformer.layers.10.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
25
+ "transformer.layers.10.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
26
+ "transformer.layers.10.norm.bias": "model-00001-of-00002.safetensors",
27
+ "transformer.layers.10.norm.weight": "model-00001-of-00002.safetensors",
28
+ "transformer.layers.11.attention.out_proj.weight": "model-00001-of-00002.safetensors",
29
+ "transformer.layers.11.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
30
+ "transformer.layers.11.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
31
+ "transformer.layers.11.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
32
+ "transformer.layers.11.norm.bias": "model-00001-of-00002.safetensors",
33
+ "transformer.layers.11.norm.weight": "model-00001-of-00002.safetensors",
34
+ "transformer.layers.12.attention.out_proj.weight": "model-00001-of-00002.safetensors",
35
+ "transformer.layers.12.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
36
+ "transformer.layers.12.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
37
+ "transformer.layers.12.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
38
+ "transformer.layers.12.norm.bias": "model-00001-of-00002.safetensors",
39
+ "transformer.layers.12.norm.weight": "model-00001-of-00002.safetensors",
40
+ "transformer.layers.13.attention.out_proj.weight": "model-00001-of-00002.safetensors",
41
+ "transformer.layers.13.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
42
+ "transformer.layers.13.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
43
+ "transformer.layers.13.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
44
+ "transformer.layers.13.norm.bias": "model-00001-of-00002.safetensors",
45
+ "transformer.layers.13.norm.weight": "model-00001-of-00002.safetensors",
46
+ "transformer.layers.14.attention.out_proj.weight": "model-00001-of-00002.safetensors",
47
+ "transformer.layers.14.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
48
+ "transformer.layers.14.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
49
+ "transformer.layers.14.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
50
+ "transformer.layers.14.norm.bias": "model-00001-of-00002.safetensors",
51
+ "transformer.layers.14.norm.weight": "model-00001-of-00002.safetensors",
52
+ "transformer.layers.15.attention.out_proj.weight": "model-00001-of-00002.safetensors",
53
+ "transformer.layers.15.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
54
+ "transformer.layers.15.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
55
+ "transformer.layers.15.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
56
+ "transformer.layers.15.norm.bias": "model-00001-of-00002.safetensors",
57
+ "transformer.layers.15.norm.weight": "model-00001-of-00002.safetensors",
58
+ "transformer.layers.16.attention.out_proj.weight": "model-00001-of-00002.safetensors",
59
+ "transformer.layers.16.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
60
+ "transformer.layers.16.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
61
+ "transformer.layers.16.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
62
+ "transformer.layers.16.norm.bias": "model-00001-of-00002.safetensors",
63
+ "transformer.layers.16.norm.weight": "model-00001-of-00002.safetensors",
64
+ "transformer.layers.17.attention.out_proj.weight": "model-00001-of-00002.safetensors",
65
+ "transformer.layers.17.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
66
+ "transformer.layers.17.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
67
+ "transformer.layers.17.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
68
+ "transformer.layers.17.norm.bias": "model-00001-of-00002.safetensors",
69
+ "transformer.layers.17.norm.weight": "model-00001-of-00002.safetensors",
70
+ "transformer.layers.18.attention.out_proj.weight": "model-00001-of-00002.safetensors",
71
+ "transformer.layers.18.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
72
+ "transformer.layers.18.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
73
+ "transformer.layers.18.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
74
+ "transformer.layers.18.norm.bias": "model-00001-of-00002.safetensors",
75
+ "transformer.layers.18.norm.weight": "model-00001-of-00002.safetensors",
76
+ "transformer.layers.19.attention.out_proj.weight": "model-00001-of-00002.safetensors",
77
+ "transformer.layers.19.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
78
+ "transformer.layers.19.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
79
+ "transformer.layers.19.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
80
+ "transformer.layers.19.norm.bias": "model-00001-of-00002.safetensors",
81
+ "transformer.layers.19.norm.weight": "model-00001-of-00002.safetensors",
82
+ "transformer.layers.2.attention.out_proj.weight": "model-00001-of-00002.safetensors",
83
+ "transformer.layers.2.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
84
+ "transformer.layers.2.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
85
+ "transformer.layers.2.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
86
+ "transformer.layers.2.norm.bias": "model-00001-of-00002.safetensors",
87
+ "transformer.layers.2.norm.weight": "model-00001-of-00002.safetensors",
88
+ "transformer.layers.20.attention.out_proj.weight": "model-00001-of-00002.safetensors",
89
+ "transformer.layers.20.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
90
+ "transformer.layers.20.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
91
+ "transformer.layers.20.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
92
+ "transformer.layers.20.norm.bias": "model-00001-of-00002.safetensors",
93
+ "transformer.layers.20.norm.weight": "model-00001-of-00002.safetensors",
94
+ "transformer.layers.21.attention.out_proj.weight": "model-00001-of-00002.safetensors",
95
+ "transformer.layers.21.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
96
+ "transformer.layers.21.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
97
+ "transformer.layers.21.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
98
+ "transformer.layers.21.norm.bias": "model-00001-of-00002.safetensors",
99
+ "transformer.layers.21.norm.weight": "model-00001-of-00002.safetensors",
100
+ "transformer.layers.22.attention.out_proj.weight": "model-00001-of-00002.safetensors",
101
+ "transformer.layers.22.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
102
+ "transformer.layers.22.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
103
+ "transformer.layers.22.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
104
+ "transformer.layers.22.norm.bias": "model-00001-of-00002.safetensors",
105
+ "transformer.layers.22.norm.weight": "model-00001-of-00002.safetensors",
106
+ "transformer.layers.23.attention.out_proj.weight": "model-00001-of-00002.safetensors",
107
+ "transformer.layers.23.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
108
+ "transformer.layers.23.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
109
+ "transformer.layers.23.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
110
+ "transformer.layers.23.norm.bias": "model-00001-of-00002.safetensors",
111
+ "transformer.layers.23.norm.weight": "model-00001-of-00002.safetensors",
112
+ "transformer.layers.24.attention.out_proj.weight": "model-00001-of-00002.safetensors",
113
+ "transformer.layers.24.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
114
+ "transformer.layers.24.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
115
+ "transformer.layers.24.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
116
+ "transformer.layers.24.norm.bias": "model-00001-of-00002.safetensors",
117
+ "transformer.layers.24.norm.weight": "model-00001-of-00002.safetensors",
118
+ "transformer.layers.25.attention.out_proj.weight": "model-00001-of-00002.safetensors",
119
+ "transformer.layers.25.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
120
+ "transformer.layers.25.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
121
+ "transformer.layers.25.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
122
+ "transformer.layers.25.norm.bias": "model-00001-of-00002.safetensors",
123
+ "transformer.layers.25.norm.weight": "model-00001-of-00002.safetensors",
124
+ "transformer.layers.26.attention.out_proj.weight": "model-00001-of-00002.safetensors",
125
+ "transformer.layers.26.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
126
+ "transformer.layers.26.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
127
+ "transformer.layers.26.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
128
+ "transformer.layers.26.norm.bias": "model-00001-of-00002.safetensors",
129
+ "transformer.layers.26.norm.weight": "model-00001-of-00002.safetensors",
130
+ "transformer.layers.27.attention.out_proj.weight": "model-00001-of-00002.safetensors",
131
+ "transformer.layers.27.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
132
+ "transformer.layers.27.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
133
+ "transformer.layers.27.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
134
+ "transformer.layers.27.norm.bias": "model-00001-of-00002.safetensors",
135
+ "transformer.layers.27.norm.weight": "model-00001-of-00002.safetensors",
136
+ "transformer.layers.28.attention.out_proj.weight": "model-00001-of-00002.safetensors",
137
+ "transformer.layers.28.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
138
+ "transformer.layers.28.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
139
+ "transformer.layers.28.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
140
+ "transformer.layers.28.norm.bias": "model-00001-of-00002.safetensors",
141
+ "transformer.layers.28.norm.weight": "model-00001-of-00002.safetensors",
142
+ "transformer.layers.29.attention.out_proj.weight": "model-00001-of-00002.safetensors",
143
+ "transformer.layers.29.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
144
+ "transformer.layers.29.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
145
+ "transformer.layers.29.mlp.out_proj.weight": "model-00002-of-00002.safetensors",
146
+ "transformer.layers.29.norm.bias": "model-00001-of-00002.safetensors",
147
+ "transformer.layers.29.norm.weight": "model-00001-of-00002.safetensors",
148
+ "transformer.layers.3.attention.out_proj.weight": "model-00001-of-00002.safetensors",
149
+ "transformer.layers.3.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
150
+ "transformer.layers.3.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
151
+ "transformer.layers.3.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
152
+ "transformer.layers.3.norm.bias": "model-00001-of-00002.safetensors",
153
+ "transformer.layers.3.norm.weight": "model-00001-of-00002.safetensors",
154
+ "transformer.layers.30.attention.out_proj.weight": "model-00002-of-00002.safetensors",
155
+ "transformer.layers.30.attention.qkv_proj.weight": "model-00002-of-00002.safetensors",
156
+ "transformer.layers.30.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
157
+ "transformer.layers.30.mlp.out_proj.weight": "model-00002-of-00002.safetensors",
158
+ "transformer.layers.30.norm.bias": "model-00002-of-00002.safetensors",
159
+ "transformer.layers.30.norm.weight": "model-00002-of-00002.safetensors",
160
+ "transformer.layers.31.attention.out_proj.weight": "model-00002-of-00002.safetensors",
161
+ "transformer.layers.31.attention.qkv_proj.weight": "model-00002-of-00002.safetensors",
162
+ "transformer.layers.31.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
163
+ "transformer.layers.31.mlp.out_proj.weight": "model-00002-of-00002.safetensors",
164
+ "transformer.layers.31.norm.bias": "model-00002-of-00002.safetensors",
165
+ "transformer.layers.31.norm.weight": "model-00002-of-00002.safetensors",
166
+ "transformer.layers.4.attention.out_proj.weight": "model-00001-of-00002.safetensors",
167
+ "transformer.layers.4.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
168
+ "transformer.layers.4.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
169
+ "transformer.layers.4.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
170
+ "transformer.layers.4.norm.bias": "model-00001-of-00002.safetensors",
171
+ "transformer.layers.4.norm.weight": "model-00001-of-00002.safetensors",
172
+ "transformer.layers.5.attention.out_proj.weight": "model-00001-of-00002.safetensors",
173
+ "transformer.layers.5.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
174
+ "transformer.layers.5.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
175
+ "transformer.layers.5.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
176
+ "transformer.layers.5.norm.bias": "model-00001-of-00002.safetensors",
177
+ "transformer.layers.5.norm.weight": "model-00001-of-00002.safetensors",
178
+ "transformer.layers.6.attention.out_proj.weight": "model-00001-of-00002.safetensors",
179
+ "transformer.layers.6.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
180
+ "transformer.layers.6.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
181
+ "transformer.layers.6.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
182
+ "transformer.layers.6.norm.bias": "model-00001-of-00002.safetensors",
183
+ "transformer.layers.6.norm.weight": "model-00001-of-00002.safetensors",
184
+ "transformer.layers.7.attention.out_proj.weight": "model-00001-of-00002.safetensors",
185
+ "transformer.layers.7.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
186
+ "transformer.layers.7.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
187
+ "transformer.layers.7.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
188
+ "transformer.layers.7.norm.bias": "model-00001-of-00002.safetensors",
189
+ "transformer.layers.7.norm.weight": "model-00001-of-00002.safetensors",
190
+ "transformer.layers.8.attention.out_proj.weight": "model-00001-of-00002.safetensors",
191
+ "transformer.layers.8.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
192
+ "transformer.layers.8.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
193
+ "transformer.layers.8.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
194
+ "transformer.layers.8.norm.bias": "model-00001-of-00002.safetensors",
195
+ "transformer.layers.8.norm.weight": "model-00001-of-00002.safetensors",
196
+ "transformer.layers.9.attention.out_proj.weight": "model-00001-of-00002.safetensors",
197
+ "transformer.layers.9.attention.qkv_proj.weight": "model-00001-of-00002.safetensors",
198
+ "transformer.layers.9.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
199
+ "transformer.layers.9.mlp.out_proj.weight": "model-00001-of-00002.safetensors",
200
+ "transformer.layers.9.norm.bias": "model-00001-of-00002.safetensors",
201
+ "transformer.layers.9.norm.weight": "model-00001-of-00002.safetensors"
202
+ }
203
+ }
modeling_stablelm_alpha.py ADDED
@@ -0,0 +1,656 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2023 Stability AI, EleutherAI, and The HuggingFace Inc. team. All rights reserved.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+ #
16
+ # This code is based off the following work:
17
+ # https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py
18
+ # https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt_neox/modeling_gpt_neox.py
19
+ """ PyTorch StableLM-Alpha model. """
20
+ from typing import Optional, Tuple, Union
21
+ import math
22
+
23
+ import torch
24
+ import torch.utils.checkpoint
25
+ from torch import nn
26
+ from torch.nn import CrossEntropyLoss
27
+ from transformers.modeling_outputs import (
28
+ BaseModelOutputWithPast,
29
+ CausalLMOutputWithPast,
30
+ )
31
+ from transformers.modeling_utils import PreTrainedModel
32
+ from transformers.utils import logging
33
+
34
+ from .configuration_stablelm_alpha import StableLMAlphaConfig
35
+
36
+
37
+ logger = logging.get_logger(__name__)
38
+
39
+
40
+ def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
41
+ """Expands attention_mask from `[batch_size, seq_len]` to `[batch_size, 1, tgt_seq_len, src_seq_len]`."""
42
+ batch_size, src_len = mask.size()
43
+ tgt_len = tgt_len if tgt_len is not None else src_len
44
+
45
+ expanded_mask = mask[:, None, None, :].expand(batch_size, 1, tgt_len, src_len).to(dtype)
46
+ inverted_mask = 1.0 - expanded_mask
47
+
48
+ return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
49
+
50
+
51
+ class LayerNorm(nn.LayerNorm):
52
+ def __init__(self, normalized_shape: torch.Size, bias: bool = True, **kwargs):
53
+ r"""
54
+ use_cache (`bool`, default = True): whether to use the bias term.
55
+ """
56
+ super().__init__(normalized_shape, **kwargs)
57
+ if not bias:
58
+ self.bias = None
59
+
60
+
61
+ class DecoderLayer(nn.Module):
62
+ def __init__(self, config: StableLMAlphaConfig):
63
+ super().__init__()
64
+
65
+ self.norm = LayerNorm(config.hidden_size, eps=config.norm_eps)
66
+ self.attention = Attention(config)
67
+ self.mlp = MLP(config)
68
+
69
+ def forward(
70
+ self,
71
+ hidden_states: Optional[torch.FloatTensor],
72
+ attention_mask: Optional[torch.FloatTensor] = None,
73
+ position_ids: Optional[torch.LongTensor] = None,
74
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
75
+ output_attentions: Optional[bool] = False,
76
+ use_cache: Optional[bool] = False,
77
+ ) -> Union[Tuple[torch.Tensor], Optional[Tuple[torch.Tensor, Tuple[torch.FloatTensor, ...]]]]:
78
+ residual = hidden_states
79
+
80
+ # Pre-Norm
81
+ hidden_states = self.norm(hidden_states)
82
+
83
+ # Self-Attention
84
+ attn_output, attn_weights, present_key_value = self.attention(
85
+ hidden_states=hidden_states,
86
+ attention_mask=attention_mask,
87
+ position_ids=position_ids,
88
+ past_key_value=past_key_value,
89
+ use_cache=use_cache,
90
+ output_attentions=output_attentions,
91
+ )
92
+
93
+ # Feed-forward
94
+ mlp_output = self.mlp(hidden_states)
95
+
96
+ hidden_states = residual + attn_output + mlp_output
97
+
98
+ outputs = (hidden_states,)
99
+ if output_attentions:
100
+ outputs += (attn_weights,)
101
+ if use_cache:
102
+ outputs += (present_key_value,)
103
+ return outputs # hidden_states, (optional: attn_weights), (optional: present_key_value)
104
+
105
+
106
+ class MLP(nn.Module):
107
+ def __init__(self, config: StableLMAlphaConfig):
108
+ super().__init__()
109
+
110
+ hidden_size = config.hidden_size
111
+ multiple_of = 256
112
+ ff_dim = int(8 * hidden_size / 3)
113
+ intermediate_size = multiple_of * ((ff_dim + multiple_of - 1) // multiple_of)
114
+
115
+ self.gate_proj = torch.nn.Linear(hidden_size, 2 * intermediate_size, bias=False)
116
+ self.out_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
117
+ self.act = nn.SiLU()
118
+
119
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
120
+ ff, ff_gate = self.gate_proj(x).chunk(2, dim=-1)
121
+ return self.out_proj(ff * self.act(ff_gate))
122
+
123
+
124
+ class RotaryEmbedding(torch.nn.Module):
125
+ def __init__(
126
+ self,
127
+ dim: int,
128
+ max_position_embeddings: int,
129
+ base: int = 10_000,
130
+ device: Optional[torch.device] = None,
131
+ ):
132
+ super().__init__()
133
+
134
+ self.dim = dim
135
+ self.max_position_embeddings = max_position_embeddings
136
+ self.base = base
137
+ inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, device=device, dtype=torch.float32) / self.dim))
138
+ self.register_buffer("inv_freq", inv_freq, persistent=False)
139
+
140
+ # Build here to make `torch.jit.trace` work.
141
+ self._set_cos_sin_cache(
142
+ seq_len=max_position_embeddings, device=self.inv_freq.device, dtype=torch.get_default_dtype()
143
+ )
144
+
145
+ def _set_cos_sin_cache(self, seq_len: int, device: torch.device, dtype: torch.dtype):
146
+ self.max_seq_len_cached = seq_len
147
+ t = torch.arange(self.max_seq_len_cached, device=device, dtype=self.inv_freq.dtype)
148
+ freqs = torch.outer(t, self.inv_freq)
149
+ # Different from paper, but it uses a different permutation in order to obtain the same calculation
150
+ emb = torch.cat((freqs, freqs), dim=-1)
151
+ self.register_buffer("cos_cached", emb.cos()[None, None, :, :].to(dtype), persistent=False)
152
+ self.register_buffer("sin_cached", emb.sin()[None, None, :, :].to(dtype), persistent=False)
153
+
154
+ def forward(self, x: torch.Tensor, seq_len: Optional[int] = None):
155
+ # x: [batch_size, num_heads, seq_len, head_size]
156
+ if seq_len > self.max_seq_len_cached:
157
+ self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=torch.get_default_dtype())
158
+ return (
159
+ self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
160
+ self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
161
+ )
162
+
163
+
164
+ def rotate_half(x: torch.Tensor):
165
+ """Rotates half the hidden dims of the input."""
166
+ x1, x2 = torch.chunk(x, 2, dim=-1)
167
+ return torch.cat((-x2, x1), dim=-1)
168
+
169
+
170
+ def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
171
+ # The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
172
+ cos = cos.squeeze(1).squeeze(0) # [seq_len, dim]
173
+ sin = sin.squeeze(1).squeeze(0) # [seq_len, dim]
174
+ cos = cos[position_ids].unsqueeze(1) # [batch_size, 1, seq_len, dim]
175
+ sin = sin[position_ids].unsqueeze(1) # [batch_size, 1, seq_len, dim]
176
+ q_embed = (q * cos) + (rotate_half(q) * sin)
177
+ k_embed = (k * cos) + (rotate_half(k) * sin)
178
+ return q_embed, k_embed
179
+
180
+
181
+ class Attention(nn.Module):
182
+ def __init__(self, config: StableLMAlphaConfig):
183
+ super().__init__()
184
+
185
+ self.config = config
186
+ self.hidden_size = config.hidden_size
187
+ self.num_heads = config.num_heads
188
+ self.head_dim = self.hidden_size // self.num_heads
189
+ self.max_position_embeddings = config.max_position_embeddings
190
+ if self.hidden_size % self.num_heads != 0:
191
+ raise ValueError(
192
+ "`hidden_size` is not divisble by the number of attention heads! Make sure to update them"
193
+ )
194
+
195
+ self.qkv_proj = nn.Linear(self.hidden_size, 3 * self.hidden_size, bias=False)
196
+ self.out_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
197
+ self._init_rope()
198
+
199
+ def _init_rope(self):
200
+ self.rotary_ndims = int(self.head_dim * self.config.rotary_pct)
201
+ self.rotary_emb = RotaryEmbedding(
202
+ self.rotary_ndims,
203
+ max_position_embeddings=self.config.max_position_embeddings,
204
+ base=self.config.rotary_emb_base,
205
+ )
206
+
207
+ def forward(
208
+ self,
209
+ hidden_states: torch.FloatTensor,
210
+ attention_mask: torch.FloatTensor,
211
+ position_ids: torch.LongTensor,
212
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
213
+ output_attentions: Optional[bool] = False,
214
+ use_cache: Optional[bool] = False,
215
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
216
+ has_past_key_value = past_key_value is not None
217
+
218
+ # Compute QKV
219
+ # [batch_size, seq_len, (num_heads * 3 * head_dim)]
220
+ qkv = self.qkv_proj(hidden_states)
221
+
222
+ # [batch_size, seq_len, num_heads, 3 * head_dim]
223
+ new_qkv_shape = qkv.size()[:-1] + (self.num_heads, 3 * self.head_dim)
224
+ qkv = qkv.view(*new_qkv_shape)
225
+
226
+ # 3 * [batch_size, num_heads, seq_len, head_dim]
227
+ query = qkv[..., : self.head_dim].permute(0, 2, 1, 3)
228
+ key = qkv[..., self.head_dim:(2 * self.head_dim)].permute(0, 2, 1, 3)
229
+ value = qkv[..., (2 * self.head_dim):].permute(0, 2, 1, 3)
230
+
231
+ # Compute rotary embeddings on rotary_ndims
232
+ # [batch_size, num_heads, seq_len, rotary_ndims]
233
+ query_rot = query[..., :self.rotary_ndims]
234
+ query_pass = query[..., self.rotary_ndims:]
235
+ key_rot = key[..., :self.rotary_ndims]
236
+ key_pass = key[..., self.rotary_ndims:]
237
+
238
+ # Compute token offset for rotary embeddings (when decoding)
239
+ kv_seq_len = key.shape[-2]
240
+ if has_past_key_value:
241
+ kv_seq_len += past_key_value[0].shape[-2]
242
+
243
+ # Add rotary embeddings to query and key
244
+ cos, sin = self.rotary_emb(value, seq_len=kv_seq_len)
245
+ query, key = apply_rotary_pos_emb(query_rot, key_rot, cos, sin, position_ids)
246
+
247
+ # Concatenate rotary embeddings with pass-through query and key
248
+ # [batch_size, num_heads, seq_len, head_dim]
249
+ query = torch.cat((query, query_pass), dim=-1)
250
+ key = torch.cat((key, key_pass), dim=-1)
251
+
252
+ # Reuse past key-value states
253
+ if has_past_key_value:
254
+ key = torch.cat((past_key_value[0], key), dim=2)
255
+ value = torch.cat((past_key_value[1], value), dim=2)
256
+ present_key_value = (key, value) if use_cache else None
257
+
258
+ # [batch_size, num_heads, seq_len, head_dim]
259
+ query = query.transpose(1, 2).contiguous()
260
+ key = key.transpose(1, 2).contiguous()
261
+ value = value.transpose(1, 2).contiguous()
262
+
263
+ # Compute attention
264
+ softmax_scale = 1 / math.sqrt(self.head_dim)
265
+ attn_scores = torch.einsum('bthd,bshd->bhts', query, key * softmax_scale)
266
+ # Apply the attention mask
267
+ if attention_mask is not None:
268
+ attn_scores = attn_scores + attention_mask
269
+ attn_weights = nn.functional.softmax(attn_scores, dim=-1, dtype=torch.float32).to(query.dtype)
270
+ attn_output = torch.einsum('bhts,bshd->bthd', attn_weights, value)
271
+
272
+ # Merge heads
273
+ attn_output = attn_output.reshape(attn_output.shape[0], attn_output.shape[1], -1)
274
+
275
+ # Final linear projection
276
+ attn_output = self.out_proj(attn_output)
277
+
278
+ if not output_attentions:
279
+ attn_weights = None
280
+
281
+ return attn_output, attn_weights, present_key_value
282
+
283
+
284
+ def attention_mask_func(attention_scores: torch.Tensor, ltor_mask: torch.Tensor):
285
+ attention_scores.masked_fill_(~ltor_mask, torch.finfo(attention_scores.dtype).min)
286
+ return attention_scores
287
+
288
+
289
+ class StableLMAlphaPreTrainedModel(PreTrainedModel):
290
+ """An abstract class to handle weights initialization and a simple interface
291
+ for downloading and loading pretrained models.
292
+ """
293
+
294
+ config_class = StableLMAlphaConfig
295
+ base_model_prefix = "transformer"
296
+ supports_gradient_checkpointing = True
297
+ _no_split_modules = ["DecoderLayer"]
298
+ _skip_keys_device_placement = "past_key_values"
299
+
300
+ def _init_weights(self, module: nn.Module):
301
+ """Initialize the weights"""
302
+ if isinstance(module, nn.Linear):
303
+ module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
304
+ if module.bias is not None:
305
+ module.bias.data.zero_()
306
+ elif isinstance(module, nn.Embedding):
307
+ module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
308
+ if module.padding_idx is not None:
309
+ module.weight.data[module.padding_idx].zero_()
310
+ elif isinstance(module, nn.LayerNorm):
311
+ module.bias.data.zero_()
312
+ module.weight.data.fill_(1.0)
313
+
314
+ def _set_gradient_checkpointing(self, module: nn.Module, value=False):
315
+ if isinstance(module, StableLMAlphaModel):
316
+ module.gradient_checkpointing = value
317
+
318
+
319
+ def _make_causal_mask(
320
+ input_ids_shape: torch.Size,
321
+ dtype: torch.dtype,
322
+ device: torch.device,
323
+ past_key_values_length: int = 0
324
+ ):
325
+ """Make causal mask used for bi-directional self-attention."""
326
+ batch_size, tgt_len = input_ids_shape
327
+ mask = torch.full((tgt_len, tgt_len), torch.finfo(torch.float16).min, device=device)
328
+ mask_cond = torch.arange(mask.size(-1), device=device)
329
+ mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
330
+ mask = mask.to(dtype)
331
+ if past_key_values_length > 0:
332
+ mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
333
+ return mask[None, None, :, :].expand(batch_size, 1, tgt_len, tgt_len + past_key_values_length)
334
+
335
+
336
+ class StableLMAlphaModel(StableLMAlphaPreTrainedModel):
337
+ def __init__(self, config: StableLMAlphaConfig):
338
+ super().__init__(config)
339
+ self.config = config
340
+
341
+ self.embed = nn.Embedding(config.vocab_size, config.hidden_size)
342
+ self.layers = nn.ModuleList([DecoderLayer(config) for _ in range(config.num_hidden_layers)])
343
+ self.final_norm = LayerNorm(config.hidden_size, eps=config.norm_eps)
344
+
345
+ self.gradient_checkpointing = False
346
+ self.post_init()
347
+
348
+ def get_input_embeddings(self):
349
+ return self.embed
350
+
351
+ def set_input_embeddings(self, value: nn.Module):
352
+ self.embed = value
353
+
354
+ # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
355
+ def _prepare_decoder_attention_mask(
356
+ self,
357
+ attention_mask: torch.Tensor,
358
+ input_shape: torch.Size,
359
+ inputs_embeds: torch.Tensor,
360
+ past_key_values_length: int,
361
+ ):
362
+ # Create causal mask
363
+ # [batch_size, seq_len] -> [batch_size, 1, tgt_seq_len, src_seq_len]
364
+ combined_attention_mask = None
365
+ if input_shape[-1] > 1:
366
+ combined_attention_mask = _make_causal_mask(
367
+ input_shape,
368
+ inputs_embeds.dtype,
369
+ device=inputs_embeds.device,
370
+ past_key_values_length=past_key_values_length,
371
+ )
372
+
373
+ if attention_mask is not None:
374
+ # [batch_size, seq_len] -> [batch_size, 1, tgt_seq_len, src_seq_len]
375
+ expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
376
+ inputs_embeds.device
377
+ )
378
+ combined_attention_mask = (
379
+ expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
380
+ )
381
+
382
+ return combined_attention_mask
383
+
384
+ def forward(
385
+ self,
386
+ input_ids: Optional[torch.LongTensor] = None,
387
+ attention_mask: Optional[torch.FloatTensor] = None,
388
+ position_ids: Optional[torch.LongTensor] = None,
389
+ inputs_embeds: Optional[torch.FloatTensor] = None,
390
+ past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
391
+ use_cache: Optional[bool] = None,
392
+ output_attentions: Optional[bool] = None,
393
+ output_hidden_states: Optional[bool] = None,
394
+ return_dict: Optional[bool] = None,
395
+ ) -> Union[Tuple, BaseModelOutputWithPast]:
396
+ r"""
397
+ past_key_values (`tuple(tuple(torch.FloatTensor))` of length `config.n_layers`
398
+ with each tuple having 4 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
399
+ Contains precomputed key and value hidden states of the attention blocks.
400
+ Can be used to speed up decoding. If `past_key_values` are used, the user
401
+ can optionally input only the last `decoder_input_ids` (those that don't
402
+ have their past key value states given to this model) of shape `(batch_size, 1)`
403
+ instead of all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
404
+ use_cache (`bool`, *optional*):
405
+ If set to `True`, `past_key_values` key value states are returned and
406
+ can be used to speed up decoding (see `past_key_values`).
407
+ """
408
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
409
+ output_hidden_states = (
410
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
411
+ )
412
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
413
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
414
+
415
+ if input_ids is not None and inputs_embeds is not None:
416
+ raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
417
+ elif input_ids is not None:
418
+ input_shape = input_ids.size()
419
+ elif inputs_embeds is not None:
420
+ input_shape = inputs_embeds.size()[:-1]
421
+ else:
422
+ raise ValueError("You have to specify either input_ids or inputs_embeds")
423
+
424
+ batch_size, seq_length = input_shape
425
+
426
+ if past_key_values is None:
427
+ past_key_values_length = 0
428
+ past_key_values = tuple([None] * self.config.num_hidden_layers)
429
+ seq_length_with_past = seq_length
430
+ else:
431
+ past_key_values_length = past_key_values[0][0].shape[2]
432
+ seq_length_with_past = seq_length + past_key_values_length
433
+
434
+ if position_ids is None:
435
+ device = input_ids.device if input_ids is not None else inputs_embeds.device
436
+ position_ids = torch.arange(past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device)
437
+ position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
438
+ else:
439
+ position_ids = position_ids.view(-1, seq_length).long()
440
+
441
+ if inputs_embeds is None:
442
+ inputs_embeds = self.embed(input_ids)
443
+
444
+ # Attention mask.
445
+ if attention_mask is None:
446
+ attention_mask = torch.ones(
447
+ (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
448
+ )
449
+ attention_mask = self._prepare_decoder_attention_mask(
450
+ attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
451
+ )
452
+
453
+ hidden_states = inputs_embeds
454
+
455
+ if self.gradient_checkpointing and self.training:
456
+ if use_cache:
457
+ logger.warning(
458
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
459
+ )
460
+ use_cache = False
461
+
462
+ all_hidden_states = () if output_hidden_states else None
463
+ all_attentions = () if output_attentions else None
464
+ present_key_values = () if use_cache else None
465
+
466
+ for _, (decoder_layer, past_key_value) in enumerate(zip(self.layers, past_key_values)):
467
+ if output_hidden_states:
468
+ all_hidden_states = all_hidden_states + (hidden_states,)
469
+
470
+ if self.gradient_checkpointing and self.training:
471
+
472
+ def create_custom_forward(module):
473
+ def custom_forward(*inputs):
474
+ # `None` for `use_cache`
475
+ return module(*inputs, output_attentions, None)
476
+
477
+ return custom_forward
478
+
479
+ outputs = torch.utils.checkpoint.checkpoint(
480
+ create_custom_forward(decoder_layer),
481
+ hidden_states,
482
+ attention_mask,
483
+ position_ids,
484
+ # `None` for `past_key_value`
485
+ None,
486
+ )
487
+ else:
488
+ outputs = decoder_layer(
489
+ hidden_states,
490
+ attention_mask=attention_mask,
491
+ position_ids=position_ids,
492
+ past_key_value=past_key_value,
493
+ output_attentions=output_attentions,
494
+ use_cache=use_cache,
495
+ )
496
+
497
+ hidden_states = outputs[0]
498
+
499
+ if output_attentions:
500
+ all_attentions = all_attentions + (outputs[1],)
501
+
502
+ if use_cache:
503
+ present_key_values += (outputs[2 if output_attentions else 1],)
504
+
505
+ hidden_states = self.final_norm(hidden_states)
506
+
507
+ # Add last hidden state
508
+ if output_hidden_states:
509
+ all_hidden_states += (hidden_states,)
510
+
511
+ present_key_values = present_key_values if use_cache else None
512
+ if not return_dict:
513
+ return tuple(v for v in [hidden_states, present_key_values, all_hidden_states, all_attentions] if v is not None)
514
+
515
+ return BaseModelOutputWithPast(
516
+ last_hidden_state=hidden_states,
517
+ past_key_values=present_key_values,
518
+ hidden_states=all_hidden_states,
519
+ attentions=all_attentions,
520
+ )
521
+
522
+
523
+ class StableLMAlphaForCausalLM(StableLMAlphaPreTrainedModel):
524
+ _tied_weights_keys = ["lm_head.weight"]
525
+
526
+ def __init__(self, config: StableLMAlphaConfig):
527
+ super().__init__(config)
528
+
529
+ self.transformer = StableLMAlphaModel(config)
530
+ self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
531
+
532
+ self.post_init()
533
+
534
+ def get_output_embeddings(self):
535
+ return self.lm_head
536
+
537
+ def set_output_embeddings(self, new_embeddings: nn.Module):
538
+ self.lm_head = new_embeddings
539
+
540
+ def forward(
541
+ self,
542
+ input_ids: Optional[torch.LongTensor] = None,
543
+ attention_mask: Optional[torch.FloatTensor] = None,
544
+ position_ids: Optional[torch.LongTensor] = None,
545
+ inputs_embeds: Optional[torch.FloatTensor] = None,
546
+ past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
547
+ labels: Optional[torch.LongTensor] = None,
548
+ use_cache: Optional[bool] = None,
549
+ output_attentions: Optional[bool] = None,
550
+ output_hidden_states: Optional[bool] = None,
551
+ return_dict: Optional[bool] = None,
552
+ ) -> Union[Tuple, CausalLMOutputWithPast]:
553
+ r"""
554
+ Example:
555
+
556
+ ```python
557
+ >>> from transformers import AutoTokenizer, StableLMAlphaForCausalLM, StableLMAlphaConfig
558
+ >>> import torch
559
+
560
+ >>> tokenizer = AutoTokenizer.from_pretrained("stabilityai/stablelm-base-alpha-3b-v2", trust_remote_code=True)
561
+ >>> config = StableLMAlphaConfig.from_pretrained("stabilityai/stablelm-base-alpha-3b-v2")
562
+ >>> config.is_decoder = True
563
+ >>> model = StableLMAlphaForCausalLM.from_pretrained("stabilityai/stablelm-base-alpha-3b-v2", config=config)
564
+
565
+ >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
566
+ >>> outputs = model(**inputs)
567
+
568
+ >>> logits = outputs.logits
569
+ ```
570
+ """
571
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
572
+
573
+ outputs = self.transformer(
574
+ input_ids,
575
+ attention_mask=attention_mask,
576
+ position_ids=position_ids,
577
+ inputs_embeds=inputs_embeds,
578
+ past_key_values=past_key_values,
579
+ use_cache=use_cache,
580
+ output_attentions=output_attentions,
581
+ output_hidden_states=output_hidden_states,
582
+ return_dict=return_dict,
583
+ )
584
+
585
+ hidden_states = outputs[0]
586
+ logits = self.lm_head(hidden_states)
587
+
588
+ lm_loss = None
589
+ if labels is not None:
590
+ # move labels to correct device to enable model parallelism
591
+ labels = labels.to(logits.device)
592
+ # we are doing next-token prediction; shift prediction scores and input ids by one
593
+ shift_logits = logits[:, :-1, :].contiguous()
594
+ labels = labels[:, 1:].contiguous()
595
+ loss_fct = CrossEntropyLoss()
596
+ lm_loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), labels.view(-1))
597
+
598
+ if not return_dict:
599
+ output = (logits,) + outputs[1:]
600
+ return ((lm_loss,) + output) if lm_loss is not None else output
601
+
602
+ return CausalLMOutputWithPast(
603
+ loss=lm_loss,
604
+ logits=logits,
605
+ past_key_values=outputs.past_key_values,
606
+ hidden_states=outputs.hidden_states,
607
+ attentions=outputs.attentions,
608
+ )
609
+
610
+ def prepare_inputs_for_generation(
611
+ self,
612
+ input_ids,
613
+ past_key_values: Optional[torch.Tensor] = None,
614
+ attention_mask: Optional[torch.Tensor] = None,
615
+ inputs_embeds: Optional[torch.Tensor] = None,
616
+ **kwargs
617
+ ):
618
+ # Cut decoder_input_ids if past is used
619
+ if past_key_values and past_key_values[0] is not None:
620
+ input_ids = input_ids[:, -1:]
621
+
622
+ position_ids = kwargs.get("position_ids", None)
623
+ if attention_mask is not None and position_ids is None:
624
+ # Create position_ids on the fly for batch generation
625
+ position_ids = attention_mask.long().cumsum(-1) - 1
626
+ position_ids.masked_fill_(attention_mask == 0, 1)
627
+ if past_key_values:
628
+ position_ids = position_ids[:, -1].unsqueeze(-1)
629
+
630
+ # If `inputs_embeds` are passed, we only want to use them in the 1st generation step
631
+ if inputs_embeds is not None and past_key_values is None:
632
+ model_inputs = {"inputs_embeds": inputs_embeds}
633
+ else:
634
+ model_inputs = {"input_ids": input_ids}
635
+
636
+ model_inputs.update(
637
+ {
638
+ "attention_mask": attention_mask,
639
+ "past_key_values": past_key_values,
640
+ "position_ids": position_ids,
641
+ }
642
+ )
643
+
644
+ return model_inputs
645
+
646
+ def _reorder_cache(self, past_key_values: torch.Tensor, beam_idx: int):
647
+ reordered_past = ()
648
+ for past_key_value in past_key_values:
649
+ reordered_past += (
650
+ tuple(past_state.index_select(0, beam_idx) for past_state in past_key_value[:2]) + past_key_value[2:],
651
+ )
652
+ return reordered_past
653
+
654
+
655
+ StableLMAlphaConfig.register_for_auto_class()
656
+ StableLMAlphaForCausalLM.register_for_auto_class("AutoModelForCausalLM")
special_tokens_map.json ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ {
2
+ "bos_token": "<|endoftext|>",
3
+ "eos_token": "<|endoftext|>",
4
+ "unk_token": "<|endoftext|>"
5
+ }
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "tokenizer_class": "GPTNeoXTokenizer",
3
+ "bos_token": "<|endoftext|>",
4
+ "eos_token": "<|endoftext|>",
5
+ "unk_token": "<|endoftext|>",
6
+ "clean_up_tokenization_spaces": true,
7
+ "model_max_length": 1000000000000000019884624838656
8
+ }