File size: 14,490 Bytes
50f19fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0680f69
 
 
 
 
 
50f19fa
 
 
 
 
0680f69
 
 
 
 
 
50f19fa
 
 
 
 
 
 
 
 
 
 
0680f69
50f19fa
 
 
0680f69
50f19fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0ad933c
50f19fa
 
 
 
0680f69
50f19fa
 
0680f69
 
 
 
4424c49
0680f69
 
 
50f19fa
 
 
 
4424c49
 
50f19fa
 
 
bbc8453
 
 
 
 
 
50f19fa
4424c49
50f19fa
bbc8453
 
 
 
 
 
50f19fa
 
 
 
0680f69
50f19fa
0680f69
4424c49
 
 
 
50f19fa
0680f69
 
 
50f19fa
0680f69
 
50f19fa
0680f69
4424c49
0680f69
50f19fa
 
 
4424c49
 
bbc8453
50f19fa
bbc8453
0680f69
 
4424c49
50f19fa
 
0680f69
4424c49
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0680f69
 
50f19fa
0680f69
 
50f19fa
0ad933c
 
 
 
 
 
 
 
 
 
 
 
50f19fa
0ad933c
 
 
 
 
 
 
4424c49
 
0ad933c
 
 
 
 
 
 
 
 
 
 
 
 
4424c49
0ad933c
 
 
 
 
 
 
 
 
 
 
50f19fa
0680f69
50f19fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0680f69
50f19fa
 
0680f69
af9cdba
0680f69
50f19fa
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
from gradio.components import Component
import gradio as gr
from abc import ABC, abstractclassmethod
import inspect

class BaseTCOModel(ABC):
    # TO DO: Find way to specify which component should be used for computing cost
    def __setattr__(self, name, value):
        if isinstance(value, Component):
            self._components.append(value)
        self.__dict__[name] = value

    def __init__(self):
        super(BaseTCOModel, self).__setattr__("_components", [])

    def get_components(self) -> list[Component]:
        return self._components
    
    def get_components_for_cost_computing(self):
        return self.components_for_cost_computing
    
    def get_name(self):
        return self.name
    
    def register_components_for_cost_computing(self):
        args = inspect.getfullargspec(self.compute_cost_per_token)[0][1:]
        self.components_for_cost_computing = [self.__getattribute__(arg) for arg in args]
    
    @abstractclassmethod
    def compute_cost_per_token(self):
        pass
    
    @abstractclassmethod
    def render(self):
        pass
    
    def set_name(self, name):
        self.name = name
    
    def set_formula(self, formula):
        self.formula = formula
    
    def get_formula(self):
        return self.formula

class OpenAIModel(BaseTCOModel):

    def __init__(self):
        self.set_name("(SaaS) OpenAI")
        self.set_formula(r"""$CT = \frac{CT\_1K \times 1000}{L}$  <br>
                         with: <br>
                         CT = Cost per output Token <br>
                         CT_1K = Cost per 1000 Tokens (from OpenAI's pricing web page) <br>
                         L = Input Length 
                         """)
        super().__init__()

    def render(self):
        def on_model_change(model):
            
            if model == "GPT-4":
                return gr.Dropdown.update(choices=["8K", "32K"])
            else:
                return gr.Dropdown.update(choices=["4K", "16K"])

        self.model = gr.Dropdown(["GPT-4", "GPT-3.5 Turbo"], value="GPT-4",
                                 label="OpenAI models",
                                 interactive=True, visible=False)
        self.context_length = gr.Dropdown(["8K", "32K"], value="8K", interactive=True,
                                          label="Context size",
                                          visible=False, info="Number of tokens the model considers when processing text")
        self.model.change(on_model_change, inputs=self.model, outputs=self.context_length)
        self.input_length = gr.Number(350, label="Average number of input tokens", 
                                      interactive=True, visible=False)

    def compute_cost_per_token(self, model, context_length, input_length):
        """Cost per token = """
        model = model[0]
        context_length = context_length[0]

        if model == "GPT-4" and context_length == "8K":
            cost_per_1k_input_tokens = 0.03
        elif model == "GPT-4" and context_length == "32K":
            cost_per_1k_input_tokens = 0.06
        elif model == "GPT-3.5" and context_length == "4K":
            cost_per_1k_input_tokens = 0.0015
        else:
            cost_per_1k_input_tokens = 0.003

        cost_per_output_token = cost_per_1k_input_tokens * input_length / 1000

        return cost_per_output_token

class OpenSourceLlama2Model(BaseTCOModel):
    
    def __init__(self):
        self.set_name("(Open source) Llama 2")
        self.set_formula(r"""$CT = \frac{VM\_CH}{TS \times 3600 \times MO \times U}$<br>
                         with: <br>
                         CT = Cost per Token <br>
                         VM_CH = VM Cost per Hour <br>
                         TS = Tokens per Second <br>
                         MO = Maxed Out <br>
                         U = Used
                         """)
        super().__init__()
    
    def render(self):
        vm_choices = ["1x Nvidia A100 (Azure NC24ads A100 v4)",
                      "2x Nvidia A100 (Azure NC48ads A100 v4)", 
                      "4x Nvidia A100 (Azure NC48ads A100 v4)"]
        
        def on_model_change(model):
            if model == "Llama 2 7B":
                return [gr.Dropdown.update(choices=vm_choices), 
                        gr.Markdown.update(value="To see the script used to benchmark the Llama2-7B model, [click here](https://example.com/script)"), 
                        gr.Number.update(value=3.6730),
                        gr.Number.update(value=694.38),
                        gr.Number.update(visible=True)
                        ]
            else:
                not_supported_vm = ["1x Nvidia A100 (Azure NC24ads A100 v4)", "2x Nvidia A100 (Azure NC48ads A100 v4)"]
                choices = [x for x in vm_choices if x not in not_supported_vm]
                return [gr.Dropdown.update(choices=choices, value="4x Nvidia A100 (Azure NC48ads A100 v4)"), 
                        gr.Markdown.update(value="To see the benchmark results used for the Llama2-70B model, [click here](https://www.cursor.so/blog/llama-inference#user-content-fn-llama-paper)"),
                        gr.Number.update(value=14.692),
                        gr.Number.update(value=18.6),
                        gr.Number.update(visible=False)
                        ]

        def on_vm_change(model, vm):
            # TO DO: load info from CSV
            if model == "Llama 2 7B" and vm == "1x Nvidia A100 (Azure NC24ads A100 v4)":
                return [gr.Number.update(value=3.6730), gr.Number.update(value=694.38)]
            elif model == "Llama 2 7B" and vm == "2x Nvidia A100 (Azure NC48ads A100 v4)":
                return [gr.Number.update(value=7.346), gr.Number.update(value=1388.76)]
            elif model == "Llama 2 7B" and vm == "4x Nvidia A100 (Azure NC48ads A100 v4)":
                return [gr.Number.update(value=14.692), gr.Number.update(value=2777.52)]
            elif model == "Llama 2 70B" and vm == "4x Nvidia A100 (Azure NC48ads A100 v4)":
                return [gr.Number.update(value=14.692), gr.Number.update(value=18.6)]
        
        self.model = gr.Dropdown(["Llama 2 7B", "Llama 2 70B"], value="Llama 2 7B", label="OpenSource models", visible=False)
        self.vm = gr.Dropdown(vm_choices,
                              value="1x Nvidia A100 (Azure NC24ads A100 v4)", 
                              visible=False,
                              label="Instance of VM with GPU",
                              info="Your options for this choice depend on the model you previously chose"
                              )
        self.vm_cost_per_hour = gr.Number(3.6730, label="VM instance cost per hour", 
                                      interactive=False, visible=False)
        self.tokens_per_second = gr.Number(694.38, visible=False,
                                           label="Number of tokens per second for this specific model and VM instance",
                                           interactive=False
                                           )
        self.input_length = gr.Number(233, label="Average number of input tokens", info="This is the number of input tokens used when the model was benchmarked to get the number of tokens/second it processes",
                                      interactive=False, visible=False)
        self.info = gr.Markdown("To see the script used to benchmark the Llama2-7B model, [click here](https://example.com/script)", interactive=False, visible=False)
        
        self.model.change(on_model_change, inputs=self.model, outputs=[self.vm, self.info, self.vm_cost_per_hour, self.tokens_per_second, self.input_length])
        self.vm.change(on_vm_change, inputs=[self.model, self.vm], outputs=[self.vm_cost_per_hour, self.tokens_per_second])
        self.maxed_out = gr.Slider(minimum=0.01, value=50., step=0.01, label="% maxed out", 
                                   info="How much the GPU is fully used",
                                   interactive=True,
                                   visible=False)
        self.used = gr.Slider(minimum=0.01, value=50., step=0.01, label="% used", 
                                   info="Percentage of time the GPU is used",
                                   interactive=True,
                                   visible=False)

    def compute_cost_per_token(self, vm_cost_per_hour, tokens_per_second, maxed_out, used):
        cost_per_token = vm_cost_per_hour / (tokens_per_second * 3600 * maxed_out * used)
        return cost_per_token 

class OpenSourceDIY(BaseTCOModel):
    
    def __init__(self):
        self.set_name("(Open source) DIY")
        self.set_formula(r"""$CT = \frac{VM\_CH}{TS \times 3600 \times MO \times U}$<br>
                         with: <br>
                         CT = Cost per Token <br>
                         VM_CH = VM Cost per Hour <br>
                         TS = Tokens per Second <br>
                         MO = Maxed Out <br>
                         U = Used
                         """)
        super().__init__()
    
    def render(self):   
        self.info = gr.Markdown("Compute the cost/token based on our formula below, using your own parameters", visible=False)
        self.display_formula = gr.Markdown(r"""$CT = \frac{VM\_CH}{TS \times 3600 \times MO \times U}$<br>
                         with: <br>
                         CT = Cost per Token <br>
                         VM_CH = VM Cost per Hour <br>
                         TS = Tokens per Second <br>
                         MO = Maxed Out <br>
                         U = Used
                         """, visible=False)
        self.vm_cost_per_hour = gr.Number(3.5, label="VM instance cost per hour", 
                                      interactive=True, visible=False)
        self.tokens_per_second = gr.Number(700, visible=False,
                                           label="Number of tokens per second for this specific model and VM instance",
                                           interactive=True
                                           )
        self.maxed_out = gr.Slider(minimum=0.01, value=50., step=0.01, label="% maxed out", 
                                   info="How much the GPU is fully used",
                                   interactive=True,
                                   visible=False)
        self.used = gr.Slider(minimum=0.01, value=50., step=0.01, label="% used", 
                                   info="Percentage of time the GPU is used",
                                   interactive=True,
                                   visible=False)

    def compute_cost_per_token(self, vm_cost_per_hour, tokens_per_second, maxed_out, used):
        cost_per_token = vm_cost_per_hour / (tokens_per_second * 3600 * maxed_out * used)
        return cost_per_token 

class CohereModel(BaseTCOModel):
    
    def __init__(self):
        self.set_name("(SaaS) Cohere")
        self.set_formula(r"""$CT = \frac{CT\_1K \times 1000}{L}$  <br>
                         with: <br>
                         CT = Cost per output Token <br>
                         CT_1M = Cost per one million Tokens (from Cohere's pricing web page) <br>
                         L = Input Length 
                         """)
        super().__init__()
    
    def render(self):
        def on_use_case_change(use_case):
            if use_case == "Summarize":
                return gr.Dropdown.update(choices=["Default"])
            else:
                return gr.Dropdown.update(choices=["Default", "Custom"])
            
        self.use_case = gr.Dropdown(["Generate", "Summarize"], value="Generate",
                                 label="API",
                                 interactive=True, visible=False)
        self.model = gr.Dropdown(["Default", "Custom"], value="Default",
                                 label="Model",
                                 interactive=True, visible=False)
        self.use_case.change(on_use_case_change, inputs=self.use_case, outputs=self.model)        
        self.input_length = gr.Number(350, label="Average number of input tokens", 
                                      interactive=True, visible=False)

    def compute_cost_per_token(self, use_case, model, input_length):
        """Cost per token = """
        use_case = use_case[0]        
        model = model[0]

        if use_case == "Generate":
            if model == "Default":
                cost_per_1M_input_tokens = 15
            else: 
                cost_per_1M_input_tokens = 30
        else:
            cost_per_1M_input_tokens = 15

        cost_per_output_token = cost_per_1M_input_tokens * input_length / 1000000

        return cost_per_output_token
   
class ModelPage:
    
    def __init__(self, Models: BaseTCOModel):
        self.models: list[BaseTCOModel] = []
        for Model in Models:
            model = Model()
            self.models.append(model)

    def render(self):
        for model in self.models:
            model.render()
            model.register_components_for_cost_computing() 

    def get_all_components(self) -> list[Component]:
        output = []
        for model in self.models:
            output += model.get_components()
        return output
    
    def get_all_components_for_cost_computing(self) -> list[Component]:
        output = []
        for model in self.models:
            output += model.get_components_for_cost_computing()
        return output

    def make_model_visible(self, name:str):
        # First decide which indexes
        output = []
        for model in self.models:
            if model.get_name() == name:
                output+= [gr.update(visible=True)] * len(model.get_components())
            else:
                output+= [gr.update(visible=False)] * len(model.get_components())
        return output
    
    def compute_cost_per_token(self, *args):
        begin=0
        current_model = args[-1]
        for model in self.models:
            model_n_args = len(model.get_components_for_cost_computing())
            if current_model == model.get_name():
                
                model_args = args[begin:begin+model_n_args]
                model_tco = model.compute_cost_per_token(*model_args)
                formula = model.get_formula()
                return f"Model {current_model} has a TCO of: ${model_tco}", model_tco, formula
            
            begin = begin+model_n_args