Merge pull request #4502 from BerriAI/litellm_support_dynamic_rpm_limiting

feat(dynamic_rate_limiter.py): support dynamic rate limiting on rpm

docs/my-website/docs/proxy/team_budgets.md

@@ -152,11 +152,11 @@ litellm_remaining_team_budget_metric{team_alias="QA Prod Bot",team_id="de35b29e-
 ```
 
 
-### Dynamic TPM Allocation 
+### Dynamic TPM/RPM Allocation 
 
-Prevent projects from gobbling too much quota. 
+Prevent projects from gobbling too much tpm/rpm.
 
-Dynamically allocate TPM quota to api keys, based on active keys in that minute. [**See Code**](https://github.com/BerriAI/litellm/blob/9bffa9a48e610cc6886fc2dce5c1815aeae2ad46/litellm/proxy/hooks/dynamic_rate_limiter.py#L125)
+Dynamically allocate TPM/RPM quota to api keys, based on active keys in that minute. [**See Code**](https://github.com/BerriAI/litellm/blob/9bffa9a48e610cc6886fc2dce5c1815aeae2ad46/litellm/proxy/hooks/dynamic_rate_limiter.py#L125)
 
 1. Setup config.yaml 
 
@@ -247,4 +247,90 @@ except RateLimitError as e:
 
 ```
 This was rate limited b/c - Error code: 429 - {'error': {'message': {'error': 'Key=<hashed_token> over available TPM=0. Model TPM=0, Active keys=2'}, 'type': 'None', 'param': 'None', 'code': 429}}
+```
+
+
+#### ✨ [BETA] Set Priority / Reserve Quota
+
+Reserve tpm/rpm capacity for projects in prod.
+
+:::tip
+
+Reserving tpm/rpm on keys based on priority is a premium feature. Please [get an enterprise license](./enterprise.md) for it. 
+:::
+
+
+1. Setup config.yaml
+
+```yaml 
+model_list:
+  - model_name: gpt-3.5-turbo             
+    litellm_params:
+      model: "gpt-3.5-turbo"       
+      api_key: os.environ/OPENAI_API_KEY 
+      rpm: 100   
+
+litellm_settings:
+  callbacks: ["dynamic_rate_limiter"]
+  priority_reservation: {"dev": 0, "prod": 1}
+
+general_settings:
+  master_key: sk-1234 # OR set `LITELLM_MASTER_KEY=".."` in your .env
+  database_url: postgres://.. # OR set `DATABASE_URL=".."` in your .env
+```
+
+
+priority_reservation: 
+- Dict[str, float]
+  - str: can be any string
+  - float: from 0 to 1. Specify the % of tpm/rpm to reserve for keys of this priority.
+
+**Start Proxy**
+
+```
+litellm --config /path/to/config.yaml
+```
+
+2. Create a key with that priority
+
+```bash
+curl -X POST 'http://0.0.0.0:4000/key/generate' \
+-H 'Authorization: Bearer <your-master-key>' \
+-H 'Content-Type: application/json' \
+-D '{
+	"metadata": {"priority": "dev"} # 👈 KEY CHANGE
+}'
+```
+
+**Expected Response**
+
+```
+{
+  ...
+  "key": "sk-.."
+}
+```
+
+
+3. Test it!
+
+```bash
+curl -X POST 'http://0.0.0.0:4000/chat/completions' \
+  -H 'Content-Type: application/json' \
+  -H 'Authorization: sk-...' \ # 👈 key from step 2.
+  -D '{
+  "model": "gpt-3.5-turbo",
+  "messages": [
+      {
+      "role": "user",
+      "content": "what llm are you"
+      }
+  ],
+}'
+```
+
+**Expected Response**
+
+```
+Key=... over available RPM=0. Model RPM=100, Active keys=None
 ```

litellm/__init__.py

@@ -240,6 +240,8 @@ default_user_params: Optional[Dict] = None
 default_team_settings: Optional[List] = None
 max_user_budget: Optional[float] = None
 max_end_user_budget: Optional[float] = None
+#### REQUEST PRIORITIZATION ####
+priority_reservation: Optional[Dict[str, float]] = None
 #### RELIABILITY ####
 request_timeout: float = 6000
 module_level_aclient = AsyncHTTPHandler(timeout=request_timeout)

litellm/proxy/_new_secret_config.yaml

@@ -6,4 +6,4 @@ model_list:
 
 general_settings:
   alerting: ["slack"]
-  alerting_threshold: 10
+  alerting_threshold: 10

litellm/proxy/hooks/dynamic_rate_limiter.py

@@ -3,6 +3,7 @@
 ## Tracks num active projects per minute
 
 import asyncio
+import os
 import sys
 import traceback
 from datetime import datetime
@@ -81,46 +82,109 @@ class _PROXY_DynamicRateLimitHandler(CustomLogger):
     def update_variables(self, llm_router: Router):
         self.llm_router = llm_router
 
-    async def check_available_tpm(
-        self, model: str
-    ) -> Tuple[Optional[int], Optional[int], Optional[int]]:
+    async def check_available_usage(
+        self, model: str, priority: Optional[str] = None
+    ) -> Tuple[
+        Optional[int], Optional[int], Optional[int], Optional[int], Optional[int]
+    ]:
         """
         For a given model, get its available tpm
 
+        Params:
+        - model: str, the name of the model in the router model_list
+        - priority: Optional[str], the priority for the request.
+
         Returns
-        - Tuple[available_tpm, model_tpm, active_projects]
+        - Tuple[available_tpm, available_tpm, model_tpm, model_rpm, active_projects]
+            - available_tpm: int or null - always 0 or positive.
             - available_tpm: int or null - always 0 or positive.
             - remaining_model_tpm: int or null. If available tpm is int, then this will be too.
+            - remaining_model_rpm: int or null. If available rpm is int, then this will be too.
             - active_projects: int or null
         """
-        active_projects = await self.internal_usage_cache.async_get_cache(model=model)
-        current_model_tpm: Optional[int] = await self.llm_router.get_model_group_usage(
-            model_group=model
-        )
-        model_group_info: Optional[ModelGroupInfo] = (
-            self.llm_router.get_model_group_info(model_group=model)
-        )
-        total_model_tpm: Optional[int] = None
-        if model_group_info is not None and model_group_info.tpm is not None:
-            total_model_tpm = model_group_info.tpm
+        try:
+            weight: float = 1
+            if (
+                litellm.priority_reservation is None
+                or priority not in litellm.priority_reservation
+            ):
+                verbose_proxy_logger.error(
+                    "Priority Reservation not set. priority={}, but litellm.priority_reservation is {}.".format(
+                        priority, litellm.priority_reservation
+                    )
+                )
+            elif priority is not None and litellm.priority_reservation is not None:
+                if os.getenv("LITELLM_LICENSE", None) is None:
+                    verbose_proxy_logger.error(
+                        "PREMIUM FEATURE: Reserving tpm/rpm by priority is a premium feature. Please add a 'LITELLM_LICENSE' to your .env to enable this.\nGet a license: https://docs.litellm.ai/docs/proxy/enterprise."
+                    )
+                else:
+                    weight = litellm.priority_reservation[priority]
 
-        remaining_model_tpm: Optional[int] = None
-        if total_model_tpm is not None and current_model_tpm is not None:
-            remaining_model_tpm = total_model_tpm - current_model_tpm
-        elif total_model_tpm is not None:
-            remaining_model_tpm = total_model_tpm
+            active_projects = await self.internal_usage_cache.async_get_cache(
+                model=model
+            )
+            current_model_tpm, current_model_rpm = (
+                await self.llm_router.get_model_group_usage(model_group=model)
+            )
+            model_group_info: Optional[ModelGroupInfo] = (
+                self.llm_router.get_model_group_info(model_group=model)
+            )
+            total_model_tpm: Optional[int] = None
+            total_model_rpm: Optional[int] = None
+            if model_group_info is not None:
+                if model_group_info.tpm is not None:
+                    total_model_tpm = model_group_info.tpm
+                if model_group_info.rpm is not None:
+                    total_model_rpm = model_group_info.rpm
 
-        available_tpm: Optional[int] = None
+            remaining_model_tpm: Optional[int] = None
+            if total_model_tpm is not None and current_model_tpm is not None:
+                remaining_model_tpm = total_model_tpm - current_model_tpm
+            elif total_model_tpm is not None:
+                remaining_model_tpm = total_model_tpm
 
-        if remaining_model_tpm is not None:
-            if active_projects is not None:
-                available_tpm = int(remaining_model_tpm / active_projects)
-            else:
-                available_tpm = remaining_model_tpm
+            remaining_model_rpm: Optional[int] = None
+            if total_model_rpm is not None and current_model_rpm is not None:
+                remaining_model_rpm = total_model_rpm - current_model_rpm
+            elif total_model_rpm is not None:
+                remaining_model_rpm = total_model_rpm
 
-        if available_tpm is not None and available_tpm < 0:
-            available_tpm = 0
-        return available_tpm, remaining_model_tpm, active_projects
+            available_tpm: Optional[int] = None
+
+            if remaining_model_tpm is not None:
+                if active_projects is not None:
+                    available_tpm = int(remaining_model_tpm * weight / active_projects)
+                else:
+                    available_tpm = int(remaining_model_tpm * weight)
+
+            if available_tpm is not None and available_tpm < 0:
+                available_tpm = 0
+
+            available_rpm: Optional[int] = None
+
+            if remaining_model_rpm is not None:
+                if active_projects is not None:
+                    available_rpm = int(remaining_model_rpm * weight / active_projects)
+                else:
+                    available_rpm = int(remaining_model_rpm * weight)
+
+            if available_rpm is not None and available_rpm < 0:
+                available_rpm = 0
+            return (
+                available_tpm,
+                available_rpm,
+                remaining_model_tpm,
+                remaining_model_rpm,
+                active_projects,
+            )
+        except Exception as e:
+            verbose_proxy_logger.error(
+                "litellm.proxy.hooks.dynamic_rate_limiter.py::check_available_usage: Exception occurred - {}\n{}".format(
+                    str(e), traceback.format_exc()
+                )
+            )
+            return None, None, None, None, None
 
     async def async_pre_call_hook(
         self,
@@ -140,13 +204,19 @@ class _PROXY_DynamicRateLimitHandler(CustomLogger):
     ]:  # raise exception if invalid, return a str for the user to receive - if rejected, or return a modified dictionary for passing into litellm
         """
         - For a model group
-        - Check if tpm available
-        - Raise RateLimitError if no tpm available
+        - Check if tpm/rpm available
+        - Raise RateLimitError if no tpm/rpm available
         """
         if "model" in data:
-            available_tpm, model_tpm, active_projects = await self.check_available_tpm(
-                model=data["model"]
+            key_priority: Optional[str] = user_api_key_dict.metadata.get(
+                "priority", None
             )
+            available_tpm, available_rpm, model_tpm, model_rpm, active_projects = (
+                await self.check_available_usage(
+                    model=data["model"], priority=key_priority
+                )
+            )
+            ### CHECK TPM ###
             if available_tpm is not None and available_tpm == 0:
                 raise HTTPException(
                     status_code=429,
@@ -159,7 +229,20 @@ class _PROXY_DynamicRateLimitHandler(CustomLogger):
                         )
                     },
                 )
-            elif available_tpm is not None:
+            ### CHECK RPM ###
+            elif available_rpm is not None and available_rpm == 0:
+                raise HTTPException(
+                    status_code=429,
+                    detail={
+                        "error": "Key={} over available RPM={}. Model RPM={}, Active keys={}".format(
+                            user_api_key_dict.api_key,
+                            available_rpm,
+                            model_rpm,
+                            active_projects,
+                        )
+                    },
+                )
+            elif available_rpm is not None or available_tpm is not None:
                 ## UPDATE CACHE WITH ACTIVE PROJECT
                 asyncio.create_task(
                     self.internal_usage_cache.async_set_cache_sadd(  # this is a set
@@ -182,15 +265,24 @@ class _PROXY_DynamicRateLimitHandler(CustomLogger):
                 ), "Model info for model with id={} is None".format(
                     response._hidden_params["model_id"]
                 )
-                available_tpm, remaining_model_tpm, active_projects = (
-                    await self.check_available_tpm(model=model_info["model_name"])
+                key_priority: Optional[str] = user_api_key_dict.metadata.get(
+                    "priority", None
+                )
+                available_tpm, available_rpm, model_tpm, model_rpm, active_projects = (
+                    await self.check_available_usage(
+                        model=model_info["model_name"], priority=key_priority
+                    )
+                )
+                response._hidden_params["additional_headers"] = (
+                    {  # Add additional response headers - easier debugging
+                        "x-litellm-model_group": model_info["model_name"],
+                        "x-ratelimit-remaining-litellm-project-tokens": available_tpm,
+                        "x-ratelimit-remaining-litellm-project-requests": available_rpm,
+                        "x-ratelimit-remaining-model-tokens": model_tpm,
+                        "x-ratelimit-remaining-model-requests": model_rpm,
+                        "x-ratelimit-current-active-projects": active_projects,
+                    }
                 )
-                response._hidden_params["additional_headers"] = {
-                    "x-litellm-model_group": model_info["model_name"],
-                    "x-ratelimit-remaining-litellm-project-tokens": available_tpm,
-                    "x-ratelimit-remaining-model-tokens": remaining_model_tpm,
-                    "x-ratelimit-current-active-projects": active_projects,
-                }
 
                 return response
             return await super().async_post_call_success_hook(

litellm/router.py

@@ -4256,25 +4256,42 @@ class Router:
 
         return model_group_info
 
-    async def get_model_group_usage(self, model_group: str) -> Optional[int]:
+    async def get_model_group_usage(
+        self, model_group: str
+    ) -> Tuple[Optional[int], Optional[int]]:
         """
-        Returns remaining tpm quota for model group
+        Returns remaining tpm/rpm quota for model group
+
+        Returns:
+        - usage: Tuple[tpm, rpm]
         """
         dt = get_utc_datetime()
         current_minute = dt.strftime(
             "%H-%M"
         )  # use the same timezone regardless of system clock
         tpm_keys: List[str] = []
+        rpm_keys: List[str] = []
         for model in self.model_list:
             if "model_name" in model and model["model_name"] == model_group:
                 tpm_keys.append(
                     f"global_router:{model['model_info']['id']}:tpm:{current_minute}"
                 )
+                rpm_keys.append(
+                    f"global_router:{model['model_info']['id']}:rpm:{current_minute}"
+                )
+        combined_tpm_rpm_keys = tpm_keys + rpm_keys
+
+        combined_tpm_rpm_values = await self.cache.async_batch_get_cache(
+            keys=combined_tpm_rpm_keys
+        )
+
+        if combined_tpm_rpm_values is None:
+            return None, None
+
+        tpm_usage_list: Optional[List] = combined_tpm_rpm_values[: len(tpm_keys)]
+        rpm_usage_list: Optional[List] = combined_tpm_rpm_values[len(tpm_keys) :]
 
         ## TPM
-        tpm_usage_list: Optional[List] = await self.cache.async_batch_get_cache(
-            keys=tpm_keys
-        )
         tpm_usage: Optional[int] = None
         if tpm_usage_list is not None:
             for t in tpm_usage_list:
@@ -4282,8 +4299,15 @@ class Router:
                     if tpm_usage is None:
                         tpm_usage = 0
                     tpm_usage += t
-
-        return tpm_usage
+        ## RPM
+        rpm_usage: Optional[int] = None
+        if rpm_usage_list is not None:
+            for t in rpm_usage_list:
+                if isinstance(t, int):
+                    if rpm_usage is None:
+                        rpm_usage = 0
+                    rpm_usage += t
+        return tpm_usage, rpm_usage
 
     def get_model_ids(self) -> List[str]:
         """

litellm/tests/test_dynamic_rate_limit_handler.py

@@ -109,17 +109,56 @@ async def test_available_tpm(num_projects, dynamic_rate_limit_handler):
 
     ## CHECK AVAILABLE TPM PER PROJECT
 
-    availability, _, _ = await dynamic_rate_limit_handler.check_available_tpm(
-        model=model
-    )
+    resp = await dynamic_rate_limit_handler.check_available_usage(model=model)
+
+    availability = resp[0]
 
     expected_availability = int(model_tpm / num_projects)
 
     assert availability == expected_availability
 
 
+@pytest.mark.parametrize("num_projects", [1, 2, 100])
 @pytest.mark.asyncio
-async def test_rate_limit_raised(dynamic_rate_limit_handler, user_api_key_auth):
+async def test_available_rpm(num_projects, dynamic_rate_limit_handler):
+    model = "my-fake-model"
+    ## SET CACHE W/ ACTIVE PROJECTS
+    projects = [str(uuid.uuid4()) for _ in range(num_projects)]
+
+    await dynamic_rate_limit_handler.internal_usage_cache.async_set_cache_sadd(
+        model=model, value=projects
+    )
+
+    model_rpm = 100
+    llm_router = Router(
+        model_list=[
+            {
+                "model_name": model,
+                "litellm_params": {
+                    "model": "gpt-3.5-turbo",
+                    "api_key": "my-key",
+                    "api_base": "my-base",
+                    "rpm": model_rpm,
+                },
+            }
+        ]
+    )
+    dynamic_rate_limit_handler.update_variables(llm_router=llm_router)
+
+    ## CHECK AVAILABLE rpm PER PROJECT
+
+    resp = await dynamic_rate_limit_handler.check_available_usage(model=model)
+
+    availability = resp[1]
+
+    expected_availability = int(model_rpm / num_projects)
+
+    assert availability == expected_availability
+
+
+@pytest.mark.parametrize("usage", ["rpm", "tpm"])
+@pytest.mark.asyncio
+async def test_rate_limit_raised(dynamic_rate_limit_handler, user_api_key_auth, usage):
     """
     Unit test. Tests if rate limit error raised when quota exhausted.
     """
@@ -133,7 +172,7 @@ async def test_rate_limit_raised(dynamic_rate_limit_handler, user_api_key_auth):
         model=model, value=projects
     )
 
-    model_tpm = 0
+    model_usage = 0
     llm_router = Router(
         model_list=[
             {
@@ -142,7 +181,7 @@ async def test_rate_limit_raised(dynamic_rate_limit_handler, user_api_key_auth):
                     "model": "gpt-3.5-turbo",
                     "api_key": "my-key",
                     "api_base": "my-base",
-                    "tpm": model_tpm,
+                    usage: model_usage,
                 },
             }
         ]
@@ -151,11 +190,14 @@ async def test_rate_limit_raised(dynamic_rate_limit_handler, user_api_key_auth):
 
     ## CHECK AVAILABLE TPM PER PROJECT
 
-    availability, _, _ = await dynamic_rate_limit_handler.check_available_tpm(
-        model=model
-    )
+    resp = await dynamic_rate_limit_handler.check_available_usage(model=model)
 
-    expected_availability = int(model_tpm / 1)
+    if usage == "tpm":
+        availability = resp[0]
+    else:
+        availability = resp[1]
+
+    expected_availability = 0
 
     assert availability == expected_availability
 
@@ -217,9 +259,9 @@ async def test_base_case(dynamic_rate_limit_handler, mock_response):
     for _ in range(2):
         try:
             # check availability
-            availability, _, _ = await dynamic_rate_limit_handler.check_available_tpm(
-                model=model
-            )
+            resp = await dynamic_rate_limit_handler.check_available_usage(model=model)
+
+            availability = resp[0]
 
             print(
                 "prev_availability={}, availability={}".format(
@@ -273,9 +315,9 @@ async def test_update_cache(
     dynamic_rate_limit_handler.update_variables(llm_router=llm_router)
 
     ## INITIAL ACTIVE PROJECTS - ASSERT NONE
-    _, _, active_projects = await dynamic_rate_limit_handler.check_available_tpm(
-        model=model
-    )
+    resp = await dynamic_rate_limit_handler.check_available_usage(model=model)
+
+    active_projects = resp[-1]
 
     assert active_projects is None
 
@@ -289,9 +331,9 @@ async def test_update_cache(
 
     await asyncio.sleep(2)
     ## INITIAL ACTIVE PROJECTS - ASSERT 1
-    _, _, active_projects = await dynamic_rate_limit_handler.check_available_tpm(
-        model=model
-    )
+    resp = await dynamic_rate_limit_handler.check_available_usage(model=model)
+
+    active_projects = resp[-1]
 
     assert active_projects == 1
 
@@ -357,9 +399,9 @@ async def test_multiple_projects(
     for i in range(expected_runs + 1):
         # check availability
 
-        availability, _, _ = await dynamic_rate_limit_handler.check_available_tpm(
-            model=model
-        )
+        resp = await dynamic_rate_limit_handler.check_available_usage(model=model)
+
+        availability = resp[0]
 
         ## assert availability updated
         if prev_availability is not None and availability is not None:
@@ -389,12 +431,63 @@ async def test_multiple_projects(
         await asyncio.sleep(3)
 
     # check availability
-    availability, _, _ = await dynamic_rate_limit_handler.check_available_tpm(
-        model=model
-    )
+    resp = await dynamic_rate_limit_handler.check_available_usage(model=model)
+
+    availability = resp[0]
+
     assert availability == 0
 
 
+@pytest.mark.parametrize("num_projects", [1, 2, 100])
+@pytest.mark.asyncio
+async def test_priority_reservation(num_projects, dynamic_rate_limit_handler):
+    """
+    If reservation is set + `mock_testing_reservation` passed in
+
+    assert correct rpm is reserved
+    """
+    model = "my-fake-model"
+    ## SET CACHE W/ ACTIVE PROJECTS
+    projects = [str(uuid.uuid4()) for _ in range(num_projects)]
+
+    await dynamic_rate_limit_handler.internal_usage_cache.async_set_cache_sadd(
+        model=model, value=projects
+    )
+
+    litellm.priority_reservation = {"dev": 0.1, "prod": 0.9}
+
+    model_usage = 100
+
+    llm_router = Router(
+        model_list=[
+            {
+                "model_name": model,
+                "litellm_params": {
+                    "model": "gpt-3.5-turbo",
+                    "api_key": "my-key",
+                    "api_base": "my-base",
+                    "rpm": model_usage,
+                },
+            }
+        ]
+    )
+    dynamic_rate_limit_handler.update_variables(llm_router=llm_router)
+
+    ## CHECK AVAILABLE TPM PER PROJECT
+
+    resp = await dynamic_rate_limit_handler.check_available_usage(
+        model=model, priority="prod"
+    )
+
+    availability = resp[1]
+
+    expected_availability = int(
+        model_usage * litellm.priority_reservation["prod"] / num_projects
+    )
+
+    assert availability == expected_availability
+
+
 @pytest.mark.skip(
     reason="Unstable on ci/cd due to curr minute changes. Refactor to handle minute changing"
 )
@@ -456,9 +549,9 @@ async def test_multiple_projects_e2e(
     print("expected_runs: {}".format(expected_runs))
     for i in range(expected_runs + 1):
         # check availability
-        availability, _, _ = await dynamic_rate_limit_handler.check_available_tpm(
-            model=model
-        )
+        resp = await dynamic_rate_limit_handler.check_available_usage(model=model)
+
+        availability = resp[0]
 
         ## assert availability updated
         if prev_availability is not None and availability is not None:
@@ -488,7 +581,7 @@ async def test_multiple_projects_e2e(
         await asyncio.sleep(3)
 
     # check availability
-    availability, _, _ = await dynamic_rate_limit_handler.check_available_tpm(
-        model=model
-    )
+    resp = await dynamic_rate_limit_handler.check_available_usage(model=model)
+
+    availability = resp[0]
     assert availability == 0