Update models.py

models.py

@@ -14,9 +14,6 @@ class BaseTCOModel(ABC):
     def __init__(self):
         super(BaseTCOModel, self).__setattr__("_components", [])
         self.use_case = None  
-        self.num_users = None
-        self.input_tokens = None
-        self.output_tokens = None
 
     def get_components(self) -> list[Component]:
         return self._components
@@ -61,7 +58,7 @@ class OpenAIModel(BaseTCOModel):
         self.set_formula(r"""$CR = \frac{CIT\_1K \times IT + COT\_1K \times OT}{1000}$  <br>
                          with: <br>
                          CR = Cost per Request <br>
-                         CIT_1K = Cost per 1000 Input Tokens (from OpenAI's pricing web page) <br>
+                         CIT_1K = Cost per 1000 Input Tokens <br>
                          COT_1K = Cost per 1000 Output Tokens <br>
                          IT = Input Tokens <br>
                          OT = Output Tokens
@@ -79,45 +76,59 @@ class OpenAIModel(BaseTCOModel):
                 self.latency = "5s"
                 return gr.Dropdown.update(choices=["4K", "16K"], value="4K")
 
+        def define_cost_per_token(model, context_length):
+            if model == "GPT-4" and context_length == "8K":
+                cost_per_1k_input_tokens = 0.03
+                cost_per_1k_output_tokens = 0.06
+            elif model == "GPT-4" and context_length == "32K":
+                cost_per_1k_input_tokens = 0.06
+                cost_per_1k_output_tokens = 0.12
+            elif model == "GPT-3.5" and context_length == "4K":
+                cost_per_1k_input_tokens = 0.0015
+                cost_per_1k_output_tokens = 0.002
+            else:
+                cost_per_1k_input_tokens = 0.003
+                cost_per_1k_output_tokens = 0.004
+            return cost_per_1k_input_tokens, cost_per_1k_output_tokens
+        
         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)
-
-    def compute_cost_per_token(self, model, context_length):
-        """Cost per token = """
+        self.input_tokens_cost_per_second = gr.Number(0.03, visible=False,
+                                           label="($) Price/1K input prompt tokens",
+                                           interactive=False
+                                           )
+        self.output_tokens_cost_per_second = gr.Number(0.06, visible=False,
+                                           label="($) Price/1K output prompt tokens",
+                                           interactive=False
+                                           )
+        self.info = gr.Markdown("The cost per input and output tokens values are from OpenAI's pricing web page [here](https://openai.com/pricing)", interactive=False, visible=False)
+        self.model.change(on_model_change, inputs=self.model, outputs=self.context_length).then(define_cost_per_token, inputs=[self.model, self.context_length], outputs=[self.input_tokens_cost_per_second, self.output_tokens_cost_per_second])
+        self.context_length.change(define_cost_per_token, inputs=[self.model, self.context_length], outputs=[self.input_tokens_cost_per_second, self.output_tokens_cost_per_second])
+        
+        self.labor = gr.Number(0, visible=False, 
+                                label="($) Labor cost per month", 
+                                interactive=True
+                                )
 
-        if model == "GPT-4" and context_length == "8K":
-            cost_per_1k_input_tokens = 0.03
-            cost_per_1k_output_tokens = 0.06
-        elif model == "GPT-4" and context_length == "32K":
-            cost_per_1k_input_tokens = 0.06
-            cost_per_1k_output_tokens = 0.12
-        elif model == "GPT-3.5" and context_length == "4K":
-            cost_per_1k_input_tokens = 0.0015
-            cost_per_1k_output_tokens = 0.002
-        else:
-            cost_per_1k_input_tokens = 0.003
-            cost_per_1k_output_tokens = 0.004
-        cost_per_input_token = (cost_per_1k_input_tokens / 1000) 
-        cost_per_output_token = (cost_per_1k_output_tokens / 1000)
+    def compute_cost_per_token(self, input_tokens_cost_per_second, output_tokens_cost_per_second, labor):
+        cost_per_input_token = (input_tokens_cost_per_second / 1000) 
+        cost_per_output_token = (output_tokens_cost_per_second / 1000)
 
-        return cost_per_input_token, cost_per_output_token
+        return cost_per_input_token, cost_per_output_token, labor
 
 class OpenSourceLlama2Model(BaseTCOModel):
     
     def __init__(self):
-        self.set_name("(Open source) Llama 2")
-        self.set_formula(r"""$CT = \frac{VM\_CH \times 100}{3600 \times U} \times (\frac{IT}{ITS} + \frac{OT}{OTS})$<br>
+        self.set_name("(Open source) Llama 2 70B")
+        self.set_formula(r"""$CR = \frac{CIT\_1K \times IT + COT\_1K \times OT}{1000}$  <br>
                          with: <br>
-                         CT = Cost per Token <br>
-                         VM_CH = VM Cost per Hour <br>
-                         ITS = Input Tokens per Second <br>
-                         OTS = Output Tokens per Second <br>
-                         U = Used <br>
+                         CR = Cost per Request <br>
+                         CIT_1K = Cost per 1000 Input Tokens <br>
+                         COT_1K = Cost per 1000 Output Tokens <br>
                          IT = Input Tokens <br>
                          OT = Output Tokens
                          """)
@@ -125,118 +136,37 @@ class OpenSourceLlama2Model(BaseTCOModel):
         super().__init__()
     
     def render(self):
-        vm_choices = ["1x Nvidia A100 (Azure NC24ads A100 v4)",
-                      "2x Nvidia A100 (Azure NC24ads A100 v4)", 
-                      "2x Nvidia A100 (Azure ND96amsr 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 benchmark results use for 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(value=694.38),
-                        ]
-            else:
-                not_supported_vm = ["1x Nvidia A100 (Azure NC24ads A100 v4)", "2x Nvidia A100 (Azure NC24ads A100 v4)"]
-                choices = [x for x in vm_choices if x not in not_supported_vm]
-                return [gr.Dropdown.update(choices=choices, value="2x Nvidia A100 (Azure ND96amsr 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=2*37.186),
-                        gr.Number.update(value=2860),
-                        gr.Number.update(value=18.545),
-                        ]
-
-        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=4.777), gr.Number.update(value=694.38), gr.Number.update(value=694.38)]
-            elif model == "Llama 2 7B" and vm == "2x Nvidia A100 (Azure NC24ads A100 v4)":
-                return [gr.Number.update(value=2*4.777), gr.Number.update(value=1388.76), gr.Number.update(value=1388.76)]
-            elif model == "Llama 2 7B" and vm == "2x Nvidia A100 (Azure ND96amsr A100 v4)":
-                return [gr.Number.update(value=2*37.186), gr.Number.update(value=2777.52), gr.Number.update(value=2777.52)]
-            elif model == "Llama 2 70B" and vm == "2x Nvidia A100 (Azure ND96amsr A100 v4)":
-                return [gr.Number.update(value=2*37.186), gr.Number.update(value=2860), gr.Number.update(value=18.449)]
         
-        self.model = gr.Dropdown(["Llama 2 7B", "Llama 2 70B"], value="Llama 2 70B", label="OpenSource models", visible=False)
-        self.vm = gr.Dropdown(choices=["2x Nvidia A100 (Azure ND96amsr A100 v4)"],
-                              value="2x Nvidia A100 (Azure ND96amsr A100 v4)", 
+        self.vm = gr.Textbox(value="2x A100 80GB NVLINK", 
                               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(2*37.186, label="VM instance cost per hour", 
+        self.vm_cost_per_hour = gr.Number(2.21, label="VM instance cost ($) per hour", info="Note that this is the cost for a single VM instance, it is doubled in our case since two GPUs are needed",
                                       interactive=False, visible=False)
-        self.input_tokens_per_second = gr.Number(2860, visible=False,
-                                           label="Number of output tokens per second for this specific model and VM instance",
+        self.input_tokens_cost_per_second = gr.Number(0.00052, visible=False,
+                                           label="($) Price/1K input prompt tokens",
                                            interactive=False
                                            )
-        self.output_tokens_per_second = gr.Number(18.449, visible=False,
-                                           label="Number of output tokens per second for this specific model and VM instance",
+        self.output_tokens_cost_per_second = gr.Number(0.06656, visible=False,
+                                           label="($) Price/1K output prompt tokens",
                                            interactive=False
                                            )
-        self.info = gr.Markdown("To see the script used to benchmark the Llama2-70B model, [click here](https://www.cursor.so/blog/llama-inference#user-content-fn-llama-paper)", interactive=False, visible=False)
+        self.info = gr.Markdown("For the Llama2-70B model, we took the cost per input and output tokens values from the benchmark results [here](https://www.cursor.so/blog/llama-inference#user-content-fn-llama-paper)", interactive=False, visible=False)
         
-        self.model.change(on_model_change, inputs=self.model, outputs=[self.vm, self.info, self.vm_cost_per_hour, self.input_tokens_per_second, self.output_tokens_per_second])
-        self.vm.change(on_vm_change, inputs=[self.model, self.vm], outputs=[self.vm_cost_per_hour, self.input_tokens_per_second, self.output_tokens_per_second])
-        self.used = gr.Slider(minimum=0.01, value=30., 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, input_tokens_per_second, output_tokens_per_second, used):
-        cost_per_input_token =  vm_cost_per_hour * 100 / (3600 * used * input_tokens_per_second) 
-        cost_per_output_token =  vm_cost_per_hour * 100 / (3600 * used * output_tokens_per_second) 
-        return cost_per_input_token,  cost_per_output_token
-
-class OpenSourceDIY(BaseTCOModel):
-    
-    def __init__(self):
-        self.set_name("(Open source) DIY")
-        self.set_formula(r"""$CT = \frac{VM\_CH \times 100}{3600 \times U} \times (\frac{IT}{ITS} + \frac{OT}{OTS})$<br>
-                         with: <br>
-                         CT = Cost per Token <br>
-                         VM_CH = VM Cost per Hour <br>
-                         ITS = Input Tokens per Second <br>
-                         OTS = Output Tokens per Second <br>
-                         U = Used <br>
-                         IT = Input Tokens <br>
-                         OT = Output Tokens
-                         """)
-        self.set_latency("The latency can't be estimated in the DIY scenario for the model isn't defined")
-        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"""$CT = \frac{VM\_CH \times 100}{3600 \times U} \times (\frac{IT}{ITS} + \frac{OT}{OTS})$<br>
-                         with: <br>
-                         CT = Cost per Token <br>
-                         VM_CH = VM Cost per Hour <br>
-                         ITS = Input Tokens per Second <br>
-                         OTS = Output Tokens per Second <br>
-                         U = Used <br>
-                         IT = Input Tokens <br>
-                         OT = Output Tokens
-                         """, visible=False)
-        self.vm_cost_per_hour = gr.Number(3.5, label="VM instance cost per hour", 
-                                      interactive=True, visible=False)
-        self.input_tokens_per_second = gr.Number(300, visible=False,
-                                           label="Number of input tokens per second processed for this specific model and VM instance",
-                                           interactive=True
-                                           )
-        self.output_tokens_per_second = gr.Number(300, visible=False,
-                                           label="Number of output tokens per second processed for this specific model and VM instance",
-                                           interactive=True
-                                           )
-        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)
+        self.labor = gr.Number(1000, visible=False, 
+                                label="($) Labor cost per month",
+                                interactive=True
+                                )
+        
+        # self.used = gr.Slider(minimum=0.01, value=30., 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, input_tokens_per_second, output_tokens_per_second, used):
-        cost_per_input_token =  vm_cost_per_hour * 100 / (3600 * used * input_tokens_per_second) 
-        cost_per_output_token =  vm_cost_per_hour * 100 / (3600 * used * output_tokens_per_second) 
-        return cost_per_input_token, cost_per_output_token  
+    def compute_cost_per_token(self, input_tokens_cost_per_second, output_tokens_cost_per_second, labor):
+        cost_per_input_token = (input_tokens_cost_per_second / 1000) 
+        cost_per_output_token = (output_tokens_cost_per_second / 1000)
+        return cost_per_input_token,  cost_per_output_token, labor
 
 class CohereModel(BaseTCOModel):
     
@@ -262,8 +192,13 @@ class CohereModel(BaseTCOModel):
             self.model: gr.Dropdown.update(choices=["Default", "Custom"])
         else: 
             self.model: gr.Dropdown.update(choices=["Default", "Custom"])
+            
+        self.labor = gr.Number(0, visible=False, 
+                                label="($) Labor cost per month", 
+                                interactive=True
+                                )
 
-    def compute_cost_per_token(self, model):
+    def compute_cost_per_token(self, model, labor):
         """Cost per token = """
         use_case = self.use_case  
 
@@ -279,7 +214,7 @@ class CohereModel(BaseTCOModel):
         cost_per_input_token = cost_per_1M_tokens / 1000000
         cost_per_output_token = cost_per_1M_tokens / 1000000
 
-        return cost_per_input_token, cost_per_output_token
+        return cost_per_input_token, cost_per_output_token, labor
 
 class ModelPage:
     
@@ -306,7 +241,7 @@ class ModelPage:
             output += model.get_components_for_cost_computing()
         return output
 
-    def make_model_visible(self, name:str, use_case: gr.Dropdown, num_users: gr.Number, input_tokens: gr.Slider, output_tokens: gr.Slider):
+    def make_model_visible(self, name:str, use_case: gr.Dropdown):
         # First decide which indexes
         output = []
         for model in self.models:
@@ -314,26 +249,25 @@ class ModelPage:
                 output+= [gr.update(visible=True)] * len(model.get_components()) 
                 # Set use_case and num_users values in the model
                 model.use_case = use_case
-                model.num_users = num_users
-                model.input_tokens = input_tokens
-                model.output_tokens = output_tokens
             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]
+        current_model = args[-3]
+        current_input_tokens = args[-2]
+        current_output_tokens = 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]
-                cost_per_input_token, cost_per_output_token = model.compute_cost_per_token(*model_args)
-                model_tco = cost_per_input_token * model.input_tokens + cost_per_output_token * model.output_tokens
+                cost_per_input_token, cost_per_output_token, labor_cost = model.compute_cost_per_token(*model_args)
+                model_tco = cost_per_input_token * current_input_tokens + cost_per_output_token * current_output_tokens 
                 formula = model.get_formula()
                 latency = model.get_latency()
                 
-                return f"Model {current_model} has a cost/request of: ${model_tco}", model_tco, formula, f"The average latency of this model is {latency}"
+                return f"Model {current_model} has a cost/request of: ${model_tco}", model_tco, formula, f"The average latency of this model is {latency}", labor_cost
             
             begin = begin+model_n_args