sharpenb commited on
Commit
3ae62f7
1 Parent(s): 32b9203

Upload folder using huggingface_hub (#1)

Browse files

- 2b6e02a7006c8c39f12d2af378d83e4aaeaba931b6d556d2d1b05a4400998b87 (8489ee07f66373e49f008e787f67524d6c6411a0)

README.md ADDED
@@ -0,0 +1,85 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ thumbnail: "https://assets-global.website-files.com/646b351987a8d8ce158d1940/64ec9e96b4334c0e1ac41504_Logo%20with%20white%20text.svg"
3
+ base_model: JingzeShi/Doge-60M-Instruct
4
+ metrics:
5
+ - memory_disk
6
+ - memory_inference
7
+ - inference_latency
8
+ - inference_throughput
9
+ - inference_CO2_emissions
10
+ - inference_energy_consumption
11
+ tags:
12
+ - pruna-ai
13
+ ---
14
+ <!-- header start -->
15
+ <!-- 200823 -->
16
+ <div style="width: auto; margin-left: auto; margin-right: auto">
17
+ <a href="https://docs.pruna.ai/en/latest/setup/pip.html" target="_blank" rel="noopener noreferrer">
18
+ <img src="https://imgur.com/rVAgqMY.png" alt="PrunaAI" style="width: 100%; min-width: 400px; display: block; margin: auto;">
19
+ </a>
20
+ </div>
21
+ <!-- header end -->
22
+
23
+ [![Twitter](https://img.shields.io/twitter/follow/PrunaAI?style=social)](https://twitter.com/PrunaAI)
24
+ [![GitHub](https://img.shields.io/github/followers/PrunaAI?label=Follow%20%40PrunaAI&style=social)](https://github.com/PrunaAI)
25
+ [![LinkedIn](https://img.shields.io/badge/LinkedIn-Connect-blue)](https://www.linkedin.com/company/93832878/admin/feed/posts/?feedType=following)
26
+ [![Discord](https://img.shields.io/badge/Discord-Join%20Us-blue?style=social&logo=discord)](https://discord.gg/rskEr4BZJx)
27
+
28
+ # Simply make AI models cheaper, smaller, faster, and greener!
29
+
30
+ - Give a thumbs up if you like this model!
31
+ - Contact us and tell us which model to compress next [here](https://www.pruna.ai/contact).
32
+ - Request access to easily compress your *own* AI models [here](https://z0halsaff74.typeform.com/pruna-access?typeform-source=www.pruna.ai).
33
+ - Read the documentations to know more [here](https://pruna-ai-pruna.readthedocs-hosted.com/en/latest/)
34
+ - Join Pruna AI community on Discord [here](https://discord.gg/CP4VSgck) to share feedback/suggestions or get help.
35
+
36
+ ## Results
37
+
38
+ ![image info](./plots.png)
39
+
40
+ **Frequently Asked Questions**
41
+ - ***How does the compression work?*** The model is compressed with llm-int8.
42
+ - ***How does the model quality change?*** The quality of the model output might vary compared to the base model.
43
+ - ***How is the model efficiency evaluated?*** These results were obtained with configuration described in `model/smash_config.json` and are obtained after a hardware warmup. The smashed model is directly compared to the original base model. Efficiency results may vary in other settings (e.g. other hardware, image size, batch size, ...). We recommend to directly run them in the use-case conditions to know if the smashed model can benefit you.
44
+ - ***What is the model format?*** We use safetensors.
45
+ - ***What calibration data has been used?*** If needed by the compression method, we used WikiText as the calibration data.
46
+ - ***What is the naming convention for Pruna Huggingface models?*** We take the original model name and append "turbo", "tiny", or "green" if the smashed model has a measured inference speed, inference memory, or inference energy consumption which is less than 90% of the original base model.
47
+ - ***How to compress my own models?*** You can request premium access to more compression methods and tech support for your specific use-cases [here](https://z0halsaff74.typeform.com/pruna-access?typeform-source=www.pruna.ai).
48
+ - ***What are "first" metrics?*** Results mentioning "first" are obtained after the first run of the model. The first run might take more memory or be slower than the subsequent runs due cuda overheads.
49
+ - ***What are "Sync" and "Async" metrics?*** "Sync" metrics are obtained by syncing all GPU processes and stop measurement when all of them are executed. "Async" metrics are obtained without syncing all GPU processes and stop when the model output can be used by the CPU. We provide both metrics since both could be relevant depending on the use-case. We recommend to test the efficiency gains directly in your use-cases.
50
+
51
+ ## Setup
52
+
53
+ You can run the smashed model with these steps:
54
+
55
+ 0. Check requirements from the original repo JingzeShi/Doge-60M-Instruct installed. In particular, check python, cuda, and transformers versions.
56
+ 1. Make sure that you have installed quantization related packages.
57
+ ```bash
58
+ pip install transformers accelerate bitsandbytes>0.37.0
59
+ ```
60
+ 2. Load & run the model.
61
+ ```python
62
+ from transformers import AutoModelForCausalLM, AutoTokenizer
63
+
64
+
65
+ model = AutoModelForCausalLM.from_pretrained("PrunaAI/JingzeShi-Doge-60M-Instruct-bnb-8bit-smashed", trust_remote_code=True, device_map='auto')
66
+ tokenizer = AutoTokenizer.from_pretrained("JingzeShi/Doge-60M-Instruct")
67
+
68
+ input_ids = tokenizer("What is the color of prunes?,", return_tensors='pt').to(model.device)["input_ids"]
69
+
70
+ outputs = model.generate(input_ids, max_new_tokens=216)
71
+ tokenizer.decode(outputs[0])
72
+ ```
73
+
74
+ ## Configurations
75
+
76
+ The configuration info are in `smash_config.json`.
77
+
78
+ ## Credits & License
79
+
80
+ The license of the smashed model follows the license of the original model. Please check the license of the original model JingzeShi/Doge-60M-Instruct before using this model which provided the base model. The license of the `pruna-engine` is [here](https://pypi.org/project/pruna-engine/) on Pypi.
81
+
82
+ ## Want to compress other models?
83
+
84
+ - Contact us and tell us which model to compress next [here](https://www.pruna.ai/contact).
85
+ - Do it by yourself [here](https://docs.pruna.ai/en/latest/setup/pip.html).
config.json ADDED
@@ -0,0 +1,55 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "/covalent/.cache/models/tmp4d4ly8jc9ebsoso6",
3
+ "architectures": [
4
+ "DogeForCausalLM"
5
+ ],
6
+ "attention_dropout": 0.0,
7
+ "auto_map": {
8
+ "AutoConfig": "configuration_doge.DogeConfig",
9
+ "AutoModelForCausalLM": "JingzeShi/Doge-60M-Instruct--modeling_doge.DogeForCausalLM"
10
+ },
11
+ "bos_token_id": 1,
12
+ "eos_token_id": 2,
13
+ "expert_retrieval_size": 256,
14
+ "hidden_act": "silu",
15
+ "hidden_bias": false,
16
+ "hidden_dropout": 0.0,
17
+ "hidden_size": 512,
18
+ "initializer_range": 0.02,
19
+ "intermediate_size": 2048,
20
+ "is_moe": false,
21
+ "max_position_embeddings": 2048,
22
+ "model_type": "doge",
23
+ "num_attention_heads": 4,
24
+ "num_cdmmoe_experts": 4096,
25
+ "num_cdmmoe_experts_per_head": 8,
26
+ "num_cdmmoe_heads": 4,
27
+ "num_hidden_layers": 8,
28
+ "pad_token_id": 0,
29
+ "quantization_config": {
30
+ "_load_in_4bit": false,
31
+ "_load_in_8bit": true,
32
+ "bnb_4bit_compute_dtype": "bfloat16",
33
+ "bnb_4bit_quant_storage": "uint8",
34
+ "bnb_4bit_quant_type": "fp4",
35
+ "bnb_4bit_use_double_quant": false,
36
+ "llm_int8_enable_fp32_cpu_offload": false,
37
+ "llm_int8_has_fp16_weight": false,
38
+ "llm_int8_skip_modules": [
39
+ "lm_head"
40
+ ],
41
+ "llm_int8_threshold": 6.0,
42
+ "load_in_4bit": false,
43
+ "load_in_8bit": true,
44
+ "quant_method": "bitsandbytes"
45
+ },
46
+ "rms_norm_eps": 1e-06,
47
+ "rope_scaling": null,
48
+ "rope_theta": 10000.0,
49
+ "tie_word_embeddings": false,
50
+ "torch_dtype": "float16",
51
+ "transformers_version": "4.46.2",
52
+ "use_cache": true,
53
+ "vocab_size": 32768,
54
+ "api_key": null
55
+ }
configuration_doge.py ADDED
@@ -0,0 +1,189 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2024 Jingze Shi and the HuggingFace Inc. team. All rights reserved.
3
+ #
4
+ # This code is based on the Wonderful Matrices paper implementation.
5
+ #
6
+ # https://arxiv.org/abs/2412.11834
7
+ #
8
+ # Licensed under the Apache License, Version 2.0 (the "License");
9
+ # you may not use this file except in compliance with the License.
10
+ # You may obtain a copy of the License at
11
+ #
12
+ # http://www.apache.org/licenses/LICENSE-2.0
13
+ #
14
+ # Unless required by applicable law or agreed to in writing, software
15
+ # distributed under the License is distributed on an "AS IS" BASIS,
16
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
17
+ # See the License for the specific language governing permissions and
18
+ # limitations under the License.
19
+ """PyTorch Doge model configuration"""
20
+
21
+ from transformers.configuration_utils import PretrainedConfig
22
+ from transformers.modeling_rope_utils import rope_config_validation
23
+
24
+
25
+ class DogeConfig(PretrainedConfig):
26
+ r"""
27
+ This is the configuration class to store the configuration of a [`DogeModel`]. It is used to instantiate an Doge
28
+ model according to the specified arguments, defining the model architecture like [LoserCheems/doge-tiny-test](https://huggingface.co/LoserCheems/doge-tiny-test)
29
+
30
+ Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
31
+ documentation from [`PretrainedConfig`] for more information.
32
+
33
+ Args:
34
+ vocab_size (`int`, *optional*, defaults to 32768):
35
+ Vocabulary size of the Doge model. Defines the number of different tokens that can be represented by the
36
+ `inputs_ids` passed when calling [`DogeModel`]
37
+ hidden_size (`int`, *optional*, defaults to 1024):
38
+ Dimension of the hidden representations.
39
+ intermediate_size (`int`, *optional*, defaults to 4096):
40
+ Dimension of the CDMoE representations.
41
+ num_hidden_layers (`int`, *optional*, defaults to 16):
42
+ Number of hidden layers in the Transformer decoder.
43
+ hidden_bias (`bool`, *optional*, defaults to `False`):
44
+ Whether to use bias in the hidden layers.
45
+ hidden_dropout (`float`, *optional*, defaults to 0.0):
46
+ Dropout probability for each sequence transformation and state transformation module.
47
+ hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
48
+ The non-linear activation function (function or string) in the decoder.
49
+ max_position_embeddings (`int`, *optional*, defaults to 2048):
50
+ The maximum sequence length that this model might ever be used with.
51
+ rope_theta (`float`, *optional*, defaults to 10000.0):
52
+ The base period of the RoPE embeddings.
53
+ rope_scaling (`Dict`, *optional*):
54
+ Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
55
+ and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
56
+ accordingly.
57
+ Expected contents:
58
+ `rope_type` (`str`):
59
+ The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
60
+ 'llama3'], with 'default' being the original RoPE implementation.
61
+ `factor` (`float`, *optional*):
62
+ Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
63
+ most scaling types, a `factor` of x will enable the model to handle sequences of length x *
64
+ original maximum pre-trained length.
65
+ `original_max_position_embeddings` (`int`, *optional*):
66
+ Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
67
+ pretraining.
68
+ `attention_factor` (`float`, *optional*):
69
+ Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
70
+ computation. If unspecified, it defaults to value recommended by the implementation, using the
71
+ `factor` field to infer the suggested value.
72
+ `beta_fast` (`float`, *optional*):
73
+ Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
74
+ ramp function. If unspecified, it defaults to 32.
75
+ `beta_slow` (`float`, *optional*):
76
+ Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
77
+ ramp function. If unspecified, it defaults to 1.
78
+ `short_factor` (`List[float]`, *optional*):
79
+ Only used with 'longrope'. The scaling factor to be applied to short contexts (<
80
+ `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
81
+ size divided by the number of attention heads divided by 2
82
+ `long_factor` (`List[float]`, *optional*):
83
+ Only used with 'longrope'. The scaling factor to be applied to long contexts (<
84
+ `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
85
+ size divided by the number of attention heads divided by 2
86
+ `low_freq_factor` (`float`, *optional*):
87
+ Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
88
+ `high_freq_factor` (`float`, *optional*):
89
+ Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
90
+ initializer_range (`float`, *optional*, defaults to 0.02):
91
+ The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
92
+ rms_norm_eps (`float`, *optional*, defaults to 1e-06):
93
+ The epsilon used by the rms normalization layers.
94
+ use_cache (`bool`, *optional*, defaults to `True`):
95
+ Whether or not the model should return the last key/values attentions (not used by all models). Only
96
+ relevant if `config.is_decoder=True`.
97
+ pad_token_id (`int`, *optional*, defaults to 0):
98
+ Padding token id.
99
+ bos_token_id (`int`, *optional*, defaults to 1):
100
+ Beginning of stream token id.
101
+ eos_token_id (`int`, *optional*, defaults to 2):
102
+ End of stream token id.
103
+ tie_word_embeddings (`bool`, *optional*, defaults to `False`):
104
+ Whether to tie weight embeddings
105
+ num_attention_heads (`int`, *optional*, defaults to 8):
106
+ Number of attention heads for each attention layer in the Transformer decoder.
107
+ attention_dropout (`float`, *optional*, defaults to 0.0):
108
+ The dropout ratio for the attention probabilities.
109
+ is_moe (`bool`, *optional*, defaults to `False`):
110
+ Whether to use the Cross Domain Mixture of Experts, if `True`, the MoE will inherit the MLP to initialize
111
+ num_cdmmoe_experts (`int`, *optional*, defaults to 4096):
112
+ Number of Private Experts for the Cross Domain Mixture of Experts.
113
+ num_cdmmoe_heads (`int`, *optional*, defaults to 4):
114
+ Number of heads of Private Experts for the Cross Domain Mixture of Experts.
115
+ num_cdmmoe_experts_per_head (`int`, *optional*, defaults to 8):
116
+ Number of Private Experts per head for the Cross Domain Mixture of Experts.
117
+ expert_retrieval_size (`int`, *optional*, defaults to 256):
118
+ Dimension of the Expert retrieval states for the Cross Domain Mixture of Experts.
119
+ """
120
+
121
+ model_type = "doge"
122
+ keys_to_ignore_at_inference = ["past_key_values"]
123
+
124
+ def __init__(
125
+ self,
126
+ vocab_size=32768,
127
+ hidden_size=1024,
128
+ intermediate_size=4096,
129
+ num_hidden_layers=16,
130
+ hidden_bias=False,
131
+ hidden_dropout=0.0,
132
+ hidden_act="silu",
133
+ max_position_embeddings=2048,
134
+ rope_theta=10000.0,
135
+ rope_scaling=None,
136
+ initializer_range=0.02,
137
+ rms_norm_eps=1e-06,
138
+ use_cache=True,
139
+ pad_token_id=0,
140
+ bos_token_id=1,
141
+ eos_token_id=2,
142
+ tie_word_embeddings=False,
143
+ num_attention_heads=8,
144
+ attention_dropout=0.0,
145
+ is_moe=False,
146
+ num_cdmmoe_experts=4096,
147
+ num_cdmmoe_heads=4,
148
+ num_cdmmoe_experts_per_head=8,
149
+ expert_retrieval_size=256,
150
+ **kwargs,
151
+ ):
152
+ self.vocab_size = vocab_size
153
+ self.hidden_size = hidden_size
154
+ self.intermediate_size = intermediate_size
155
+ self.num_hidden_layers = num_hidden_layers
156
+ self.hidden_bias = hidden_bias
157
+ self.hidden_dropout = hidden_dropout
158
+ self.hidden_act = hidden_act
159
+ self.max_position_embeddings = max_position_embeddings
160
+ self.rope_theta = rope_theta
161
+ self.rope_scaling = rope_scaling
162
+ self.initializer_range = initializer_range
163
+ self.rms_norm_eps = rms_norm_eps
164
+ self.use_cache = use_cache
165
+ self.pad_token_id = pad_token_id
166
+ self.bos_token_id = bos_token_id
167
+ self.eos_token_id = eos_token_id
168
+ self.tie_word_embeddings = tie_word_embeddings
169
+ self.num_attention_heads = num_attention_heads
170
+ self.attention_dropout = attention_dropout
171
+ self.is_moe = is_moe
172
+ self.num_cdmmoe_experts = num_cdmmoe_experts
173
+ self.num_cdmmoe_heads = num_cdmmoe_heads
174
+ self.num_cdmmoe_experts_per_head = num_cdmmoe_experts_per_head
175
+ self.expert_retrieval_size = expert_retrieval_size
176
+
177
+ # Validate the correctness of rotary position embeddings parameters
178
+ # BC: if there is a 'type' field, copy it it to 'rope_type'.
179
+ if self.rope_scaling is not None and "type" in self.rope_scaling:
180
+ self.rope_scaling["rope_type"] = self.rope_scaling["type"]
181
+ rope_config_validation(self)
182
+
183
+ super().__init__(
184
+ pad_token_id=pad_token_id,
185
+ bos_token_id=bos_token_id,
186
+ eos_token_id=eos_token_id,
187
+ tie_word_embeddings=tie_word_embeddings,
188
+ **kwargs,
189
+ )
generation_config.json ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "bos_token_id": 1,
4
+ "eos_token_id": 2,
5
+ "pad_token_id": 0,
6
+ "transformers_version": "4.46.2"
7
+ }
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5624fafb4a46256215f44812717687fb53d273ca656c16513a24a32952300f7c
3
+ size 100951424
smash_config.json ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "comp_cgenerate_active": false,
3
+ "comp_ctranslate_active": false,
4
+ "comp_cwhisper_active": false,
5
+ "comp_diffusers2_active": false,
6
+ "comp_ifw_active": false,
7
+ "comp_onediff_active": false,
8
+ "comp_step_caching_active": false,
9
+ "comp_torch_compile_active": false,
10
+ "comp_ws2t_active": false,
11
+ "comp_x-fast_active": false,
12
+ "prune_torch-structured_active": false,
13
+ "quant_aqlm_active": false,
14
+ "quant_awq_active": false,
15
+ "quant_gptq_active": false,
16
+ "quant_half_active": false,
17
+ "quant_hqq_active": false,
18
+ "quant_llm-int8_active": true,
19
+ "quant_quanto_active": false,
20
+ "quant_torch_dynamic_active": false,
21
+ "quant_torch_static_active": false,
22
+ "quant_llm-int8_compute_dtype": "bfloat16",
23
+ "quant_llm-int8_double_quant": false,
24
+ "quant_llm-int8_enable_fp32_cpu_offload": false,
25
+ "quant_llm-int8_has_fp16_weight": false,
26
+ "quant_llm-int8_quant_type": "fp4",
27
+ "quant_llm-int8_threshold": 6.0,
28
+ "quant_llm-int8_weight_bits": 8,
29
+ "max_batch_size": 1,
30
+ "device": "cuda",
31
+ "cache_dir": "/covalent/.cache/models/tmp4d4ly8jc",
32
+ "task": "",
33
+ "save_load_fn": "bitsandbytes",
34
+ "save_load_fn_args": {}
35
+ }