(feat) cost tracking for azure llms

litellm/llms/azure.py

@@ -157,6 +157,7 @@ class AzureChatCompletion(BaseLLM):
                     azure_client_params["azure_ad_token"] = azure_ad_token
                 azure_client = AzureOpenAI(**azure_client_params)
                 response = azure_client.chat.completions.create(**data) # type: ignore
+                response.model = "azure/" + str(response.model)
                 return convert_to_model_response_object(response_object=json.loads(response.model_dump_json()), model_response_object=model_response)
         except AzureOpenAIError as e: 
             exception_mapping_worked = True
@@ -193,6 +194,7 @@ class AzureChatCompletion(BaseLLM):
                 azure_client_params["azure_ad_token"] = azure_ad_token
             azure_client = AsyncAzureOpenAI(**azure_client_params)
             response = await azure_client.chat.completions.create(**data) 
+            response.model = "azure" + str(response.model)
             return convert_to_model_response_object(response_object=json.loads(response.model_dump_json()), model_response_object=model_response)
        except Exception as e: 
            if isinstance(e,httpx.TimeoutException):
@@ -335,7 +337,7 @@ class AzureChatCompletion(BaseLLM):
                 )
             model_response["object"] = "list"
             model_response["data"] = output_data
-            model_response["model"] = model
+            model_response["model"] = "azure/" + model
             model_response["usage"] = embedding_response["usage"]
             return model_response
         except AzureOpenAIError as e: 

litellm/tests/test_get_model_cost_map.py

@@ -68,7 +68,7 @@ def test_cost_azure_gpt_35():
             id='chatcmpl-e41836bb-bb8b-4df2-8e70-8f3e160155ac', 
             choices=[Choices(finish_reason=None, index=0, 
                             message=Message(content=' Sure! Here is a short poem about the sky:\n\nA canvas of blue, a', role='assistant'))], 
-                            model='chatGPT-deployment-LiteLLM-isAMAZING', 
+                            model='azure/gpt-35-turbo',  # azure always has model written like this
                             usage=Usage(prompt_tokens=21, completion_tokens=17, total_tokens=38)
         )
 

litellm/utils.py

@@ -1602,6 +1602,15 @@ def cost_per_token(model="", prompt_tokens=0, completion_tokens=0):
     prompt_tokens_cost_usd_dollar = 0
     completion_tokens_cost_usd_dollar = 0
     model_cost_ref = litellm.model_cost
+
+    # see this https://learn.microsoft.com/en-us/azure/ai-services/openai/concepts/models
+    azure_llms = {
+        "gpt-35-turbo": "azure/gpt-3.5-turbo",
+        "gpt-35-turbo-16k": "azure/gpt-3.5-turbo-16k",
+        "gpt-35-turbo-instruct": "azure/gpt-3.5-turbo-instruct"
+    }
+    if "azure/" in  model:
+        model = model.replace("azure/", "")
     if model in model_cost_ref:
         prompt_tokens_cost_usd_dollar = (
             model_cost_ref[model]["input_cost_per_token"] * prompt_tokens
@@ -1619,6 +1628,14 @@ def cost_per_token(model="", prompt_tokens=0, completion_tokens=0):
             model_cost_ref["ft:gpt-3.5-turbo"]["output_cost_per_token"] * completion_tokens
         )
         return prompt_tokens_cost_usd_dollar, completion_tokens_cost_usd_dollar
+    elif model in azure_llms:
+        model = azure_llms[model]
+        prompt_tokens_cost_usd_dollar = (
+            model_cost_ref[model]["input_cost_per_token"] * prompt_tokens
+        )
+        completion_tokens_cost_usd_dollar = (
+            model_cost_ref[model]["output_cost_per_token"] * completion_tokens
+        )
     else:
         # calculate average input cost, azure/gpt-deployments can potentially go here if users don't specify, gpt-4, gpt-3.5-turbo. LLMs litellm knows
         input_cost_sum = 0