crfm-helm 0.4.0__py3-none-any.whl → 0.5.10__py3-none-any.whl

helm/proxy/accounts.py

@@ -23,6 +23,9 @@ DEFAULT_QUOTAS = {
     "jurassic": {"daily": 10000},
     "gooseai": {"daily": 10000},
     "cohere": {"daily": 10000},
+    "dall_e": {"daily": 5},  # In terms of the number of generated images
+    "together_vision": {"daily": 30},
+    "simple": {"daily": 10000},
 }
 
 
@@ -303,7 +306,7 @@ class Accounts:
             model_group: str,
             granularity: str,
             compute_period: Callable[[], str],
-        ):
+        ) -> None:
             """Helper that checks the usage at a certain granularity (e.g., daily, monthly, total)."""
 
             model_group_usages = account.usages.get(model_group)
@@ -321,14 +324,38 @@ class Accounts:
             if not usage.can_use():
                 raise InsufficientQuotaError(f"{granularity} quota ({usage.quota}) for {model_group} already used up")
 
+        def check_non_empty_quota(
+            account: Account,
+            model_group: str,
+        ) -> None:
+            """Helper that checks that the account has quota at some granularity.
+
+            At each granularity, a quota of None means unlimited quota.
+            However, if the quota is None at every granularity, it means that there is no quota.
+            To enforce this rule, this helper raises a InsufficientQuotaError if the quota is None
+            at every granularity."""
+            model_group_usages = account.usages.get(model_group)
+            if model_group_usages is None:
+                raise InsufficientQuotaError(f"No quota for {model_group}")
+            if all(
+                [
+                    granularity_usage.quota is None or granularity_usage.quota <= 0
+                    for granularity_usage in model_group_usages.values()
+                ]
+            ):
+                raise InsufficientQuotaError(f"No quota for {model_group}")
+
         if self.root_mode:
             return
 
         with SqliteDict(self.path) as cache:
             account: Account = from_dict(Account, cache[api_key])
-            granular_check_can_use(account, model_group, "daily", compute_daily_period)
-            granular_check_can_use(account, model_group, "monthly", compute_monthly_period)
-            granular_check_can_use(account, model_group, "total", compute_total_period)
+        if account.is_admin:
+            return
+        granular_check_can_use(account, model_group, "daily", compute_daily_period)
+        granular_check_can_use(account, model_group, "monthly", compute_monthly_period)
+        granular_check_can_use(account, model_group, "total", compute_total_period)
+        check_non_empty_quota(account, model_group)
 
     def use(self, api_key: str, model_group: str, delta: int):
         """

helm/proxy/cli.py

@@ -21,10 +21,10 @@ from typing import List, Dict
 import re
 import sys
 
-from helm.common.hierarchical_logger import hlog
+from helm.common.hierarchical_logger import hlog, setup_default_logging
 from helm.common.authentication import Authentication
-from .accounts import Usage, Account
-from .services.remote_service import RemoteService, add_service_args, create_authentication
+from helm.proxy.accounts import Usage, Account
+from helm.proxy.services.remote_service import RemoteService, add_service_args, create_authentication
 
 GRANULARITIES = ["daily", "monthly", "total"]
 UNLIMITED_QUOTA = "unlimited"
@@ -123,7 +123,7 @@ def do_create_update_command(service: RemoteService, auth: Authentication, args)
 
     # Update quotas
     for quota_str in args.quotas:
-        m = re.match(f"(\w+)\.(\w+)=(\d+|{UNLIMITED_QUOTA})", quota_str)
+        m = re.match(rf"(\w+)\.(\w+)=(\d+|{UNLIMITED_QUOTA})", quota_str)
         if not m:
             raise Exception(
                 f"Invalid format: {quota_str}, expect <model_group>.<granularity>=<quota> "
@@ -198,6 +198,8 @@ def main():
 
     args = parser.parse_args()
 
+    setup_default_logging()
+
     service = create_remote_service(args)
     auth = create_authentication(args)
 

helm/proxy/critique/mechanical_turk_critique_importer.py

@@ -4,6 +4,7 @@ import os
 from threading import Lock
 from typing import Dict, List, Optional, Tuple, Union
 import re
+import sys
 
 from helm.common.critique_request import (
     CritiqueRequest,
@@ -15,6 +16,8 @@ from helm.common.critique_request import (
 from helm.common.hierarchical_logger import hlog
 from helm.proxy.critique.mechanical_turk_utils import replace_emoji_characters
 
+csv.field_size_limit(sys.maxsize)
+
 # A representation of fields that can be used as a dict key.
 _CritiqueRequestKey = Tuple[Tuple[str, str], ...]
 

helm/proxy/critique/mechanical_turk_utils.py

@@ -38,7 +38,7 @@ def replace_emoji_characters(s: str) -> str:
         highpoints = re.compile("[\U00010000-\U0010ffff]")
     elif sys.maxunicode == 65535:
         # Python was built with '--enable-unicode=ucs2'
-        highpoints = re.compile("[\uD800-\uDBFF][\uDC00-\uDFFF]")
+        highpoints = re.compile("[\ud800-\udbff][\udc00-\udfff]")
     else:
         raise UnicodeError("Unable to determine if Python was built using UCS-2 or UCS-4")
 

helm/proxy/critique/model_critique_client.py

@@ -2,7 +2,7 @@ from typing import Dict, List, Union, Optional
 import string
 import dataclasses
 
-from helm.benchmark.run_specs import get_default_model_deployment_for_model
+from helm.benchmark.run_spec_factory import get_default_model_deployment_for_model
 from helm.common.critique_request import (
     CritiqueRequest,
     CritiqueRequestResult,
@@ -12,9 +12,10 @@ from helm.common.critique_request import (
 )
 from helm.common.hierarchical_logger import hlog
 from helm.common.optional_dependencies import handle_module_not_found_error
-from helm.common.request import Request, RequestResult, Sequence
-from helm.proxy.clients.client import Client
+from helm.common.request import Request, RequestResult, GeneratedOutput
+from helm.clients.client import Client
 from helm.proxy.critique.critique_client import CritiqueClient
+from helm.common.media_object import MultimediaObject, MediaObject
 
 
 class CritiqueParseError(Exception):
@@ -24,6 +25,8 @@ class CritiqueParseError(Exception):
 class ModelCritiqueClient(CritiqueClient):
     """A CritiqueClient that queries a Model to answer CritiqueRequests."""
 
+    VISION_LANGUAGE_MODELS = ["openai/gpt-4-vision", "reka/reka", "huggingface/prometheus-vision"]
+
     def __init__(self, client: Client, model_name):
         self._client = client
         self._model_name = model_name
@@ -31,6 +34,11 @@ class ModelCritiqueClient(CritiqueClient):
             get_default_model_deployment_for_model(model_name, warn_arg_deprecated=False, ignore_deprecated=True)
             or self._model_name
         )
+        self.vision_language = False
+        for vision_language_model_name in self.VISION_LANGUAGE_MODELS:
+            if model_name.startswith(vision_language_model_name):
+                self.vision_language = True
+                break
 
     def _interpolate_fields(self, text: str, fields: Dict[str, str]) -> str:
         for key, value in fields.items():
@@ -58,10 +66,15 @@ class ModelCritiqueClient(CritiqueClient):
 
         requests: List[Request] = []
         for question in task.questions:
-            prompt: str = base_prompt + "\n\n" + self._question_to_prompt(question, fields)
+            prompt: str
+            if len(question.text) > 0:
+                prompt = base_prompt + "\n\n" + self._question_to_prompt(question, fields)
+            else:
+                # We may don't want to add extra newlines and prompts
+                # if the question text is empty (e.g., the Vibe-Eval evaluator).
+                prompt = base_prompt
             if question.question_type == "free_response":
-                # TODO: Make max_tokens configurable
-                max_tokens = 100
+                max_tokens = 100 if task.max_tokens is None else task.max_tokens
             elif question.question_type == "checkbox":
                 # We multiply by 2 because the model will generate a comma after each option.
                 max_tokens = len(question.options) * 2
@@ -78,12 +91,21 @@ class ModelCritiqueClient(CritiqueClient):
 
                 prompt = anthropic.HUMAN_PROMPT + prompt + anthropic.AI_PROMPT
 
+            multimodal_prompt: Optional[MultimediaObject] = None
+            if self.vision_language:
+                assert question.media_object is not None, "Expect media_object for vision-language models"
+                image_media: MediaObject = question.media_object
+                text_media: MediaObject = MediaObject(text=prompt, content_type="text/plain")
+                multimodal_prompt = MultimediaObject(media_objects=[image_media, text_media])
+                prompt = ""  # set to empty string to avoid conflicts with multimodal_prompt
+
             request = Request(
                 model=self._model_name,
                 model_deployment=self._model_deployment_name,
                 prompt=prompt,
                 max_tokens=max_tokens,
                 echo_prompt=False,
+                multimodal_prompt=multimodal_prompt,
             )
             requests.append(request)
         return requests
@@ -114,7 +136,7 @@ class ModelCritiqueClient(CritiqueClient):
         return answers
 
     def _multiple_choice_completion_to_answer(
-        self, question: CritiqueQuestionTemplate, completion: Sequence
+        self, question: CritiqueQuestionTemplate, completion: GeneratedOutput
     ) -> Optional[str]:
         """Convert a multiple choice completion to an answer."""
         assert question.question_type == "multiple_choice"
@@ -124,14 +146,20 @@ class ModelCritiqueClient(CritiqueClient):
                 raise CritiqueParseError(
                     f"Invalid answer: {completion}. Multiple choice questions should have one answer."
                 )
-            return answers[0]
+            letter_answer = answers[0]
+            choice_rank = string.ascii_uppercase.index(letter_answer)
+            if choice_rank >= len(question.options):
+                raise CritiqueParseError(
+                    f"Invalid answer: {completion}. The answer is out of range of the options: {question.options}"
+                )
+            return letter_answer
         except CritiqueParseError as e:
             # If there was an error parsing the answer, we assume the user did not answer the question.
             hlog(f"Error parsing answer: {e}. Skipping question (and so the respondent entirely)")
             return None
 
     def _checkbox_completion_to_answer(
-        self, question: CritiqueQuestionTemplate, completion: Sequence
+        self, question: CritiqueQuestionTemplate, completion: GeneratedOutput
     ) -> Optional[List[str]]:
         """Convert a checkbox completion to an answer."""
         assert question.question_type == "checkbox"
@@ -147,7 +175,9 @@ class ModelCritiqueClient(CritiqueClient):
             hlog(f"Error parsing answer: {e}. Skipping question (and so the respondent entirely)")
             return None
 
-    def _free_response_completion_to_answer(self, question: CritiqueQuestionTemplate, completion: Sequence) -> str:
+    def _free_response_completion_to_answer(
+        self, question: CritiqueQuestionTemplate, completion: GeneratedOutput
+    ) -> str:
         """Convert a free response completion to an answer."""
         assert question.question_type == "free_response"
         return completion.text

helm/proxy/example_queries.py

@@ -1,6 +1,6 @@
 import textwrap
 
-from .query import Query
+from helm.proxy.query import Query
 
 
 def dedent(text: str) -> str:
@@ -21,6 +21,7 @@ example_queries = [
             """
             temperature: 0.5  # Medium amount of randomness
             stop_sequences: [.]  # Stop when you hit a period
+            model: openai/gpt-4.1-nano-2025-04-14
             """
         ),
         environments="",
@@ -31,28 +32,33 @@ example_queries = [
             """
             temperature: 0.5  # Medium amount of randomness
             stop_sequences: [\\n]  # Stop when you hit a newline
-            num_completions: 10  # Generate many samples
+            num_completions: 5  # Generate many samples
+            model: openai/gpt-4.1-nano-2025-04-14
             """
         ),
         environments="",
     ),
-    Query(
-        prompt="The quick brown fox jumps over the lazy dog.",
-        settings=dedent(
-            """
-            echo_prompt: true  # Analyze the prompt
-            max_tokens: 0  # Don't generate any more
-            top_k_per_token: 10  # Show alternatives for each position
-            """
-        ),
-        environments=dedent(""),
-    ),
+    # Disabled because `max_tokens: 0` no longer works on the OpenAI API
+    # Query(
+    #     prompt="The quick brown fox jumps over the lazy dog.",
+    #     settings=dedent(
+    #         """
+    #         echo_prompt: true  # Analyze the prompt
+    #         max_tokens: 0  # Don't generate any more
+    #         top_k_per_token: 5  # Show alternatives for each position
+    #         model: openai/text-davinci-002
+    #         model_deployment: openai/text-davinci-002
+    #         """
+    #     ),
+    #     environments=dedent(""),
+    # ),
     Query(
         prompt="Odd numbers: 1 -> 3 -> 5",
         settings=dedent(
             """
             temperature: 0  # Deterministic
             max_tokens: 50
+            model: openai/gpt-4.1-nano-2025-04-14
             """
         ),
         environments="",
@@ -63,13 +69,14 @@ example_queries = [
             """
             temperature: 0
             stop_sequences: [.]
-            model_deployment: ${model_deployment}  # Try out multiple models
+            # Try out multiple models
+            model: ${model}
             """
         ),
         environments=dedent(
             """
             occupation: [mathematician, lawyer, doctor]
-            model_deployment: [openai/davinci, ai21/j1-jumbo]
+            model: [openai/gpt-4.1-nano-2025-04-14, openai/gpt-4.1-mini-2025-04-14]
             """
         ),
     ),
@@ -88,12 +95,13 @@ example_queries = [
             temperature: 0.5
             stop_sequences: [\\n]
             num_completions: 5
-            model_deployment: ${model_deployment}  # Try out GPT-3 and Jurassic
+            # Try out multiple models
+            model: ${model}
             """
         ),
         environments=dedent(
             """
-            model_deployment: [openai/davinci, ai21/j1-jumbo]
+            model: [openai/gpt-4.1-nano-2025-04-14, openai/gpt-4.1-mini-2025-04-14]
             """
         ),
     ),
@@ -122,20 +130,21 @@ example_queries = [
             temperature: 0
             max_tokens: 1
             top_k_per_token: 4
-            model_deployment: ${model_deployment}  # Try out GPT-3 and Jurassic
+            # Try out multiple models
+            model: ${model}
             """
         ),
         environments=dedent(
             """
-            model_deployment: [openai/davinci, ai21/j1-jumbo]
+            model: [openai/gpt-4.1-nano-2025-04-14, openai/gpt-4.1-mini-2025-04-14]
             """
         ),
     ),
     Query(
-        prompt="Takes two vectors a and b and returns their Euclidean distance",
+        prompt="Write a Python function that takes two vectors a and b and returns their Euclidean distance.",
         settings=dedent(
             """
-            model_deployment: openai/code-davinci-001  # Codex for code generation
+            model: openai/gpt-4.1-nano-2025-04-14
             """
         ),
         environments="",
@@ -144,19 +153,15 @@ example_queries = [
         prompt="The quick brown fox",
         settings=dedent(
             """
-            model_deployment: ${model_deployment}
             temperature: 0.3
             stop_sequences: [\\n]
+            # Try out multiple models
+            model: ${model}
             """
         ),
         environments=dedent(
             """
-            model_deployment: [
-                "openai/davinci", "openai/text-davinci-002",
-                "openai/text-davinci-003", "ai21/j1-grande-v2-beta",
-                "together/gpt-j-6b", "together/gpt-jt-6b-v1",
-                "together/bloom", "together/opt-175b"
-            ]
+            model: [openai/gpt-4.1-nano-2025-04-14, openai/gpt-4.1-mini-2025-04-14]
             """
         ),
     ),

helm/proxy/retry.py

@@ -5,6 +5,7 @@ from retrying import Retrying
 from helm.common.request import RequestResult
 from helm.common.tokenization_request import TokenizationRequestResult
 from helm.common.hierarchical_logger import hlog
+import os
 import traceback
 import threading
 
@@ -19,6 +20,10 @@ Example usage:
         ...
 """
 
+# TODO: make these configurable at a config / cli level
+HELM_RETRIES = int(os.environ.get("HELM_RETRIES", "5"))
+HELM_TOKENIZER_RETRIES = int(os.environ.get("HELM_TOKENIZER_RETRIES", HELM_RETRIES))
+
 # The lock is used to prevent multiple threads from printing at the same time.
 # This can cause issues when printing the stack trace.
 # (The stack traces can get mixed up and become unreadable.)

helm/proxy/server.py

@@ -1,5 +1,3 @@
-# mypy: check_untyped_defs = False
-
 """
 Starts a REST server for the frontend to interact with.
 Look at `index.js` to see how the functionality is invoked.
@@ -8,6 +6,7 @@ Look at `index.js` to see how the functionality is invoked.
 from urllib.parse import unquote_plus
 import argparse
 import dataclasses
+import importlib_resources as resources
 import json
 import os
 import sys
@@ -20,15 +19,20 @@ from helm.benchmark.config_registry import (
     register_configs_from_directory,
     register_builtin_configs_from_helm_package,
 )
+from helm.benchmark.model_deployment_registry import get_default_model_deployment_for_model
 from helm.common.authentication import Authentication
-from helm.common.hierarchical_logger import hlog
+from helm.common.cache_backend_config import CacheBackendConfig, MongoCacheBackendConfig, SqliteCacheBackendConfig
+from helm.common.general import ensure_directory_exists
+from helm.common.hierarchical_logger import hlog, setup_default_logging
 from helm.common.optional_dependencies import handle_module_not_found_error
 from helm.common.request import Request
 from helm.common.perspective_api_request import PerspectiveAPIRequest
+from helm.common.moderations_api_request import ModerationAPIRequest
 from helm.common.tokenization_request import TokenizationRequest, DecodeRequest
-from .accounts import Account
-from .services.server_service import ServerService
-from .query import Query
+from helm.proxy.services.service import CACHE_DIR
+from helm.proxy.accounts import Account
+from helm.proxy.services.server_service import ServerService
+from helm.proxy.query import Query
 
 try:
     import gunicorn  # noqa
@@ -39,6 +43,7 @@ except ModuleNotFoundError as e:
 bottle.BaseRequest.MEMFILE_MAX = 1024 * 1024
 
 app = bottle.default_app()
+service: ServerService
 
 
 def safe_call(func, to_json=True):
@@ -82,30 +87,30 @@ def handle_root():
 
 @app.get("/static/<filename:path>")
 def handle_static_filename(filename):
-    resp = bottle.static_file(filename, root=os.path.join(os.path.dirname(__file__), "static"))
+    resp = bottle.static_file(filename, root=app.config["helm.staticpath"])
     resp.add_header("Cache-Control", "no-store, must-revalidate ")
     return resp
 
 
+@app.get("/output/<filename:path>")
+def handle_output_filename(filename):
+    resp = bottle.static_file(filename, root=app.config["crfm.proxy.outputpath"])
+    return resp
+
+
 @app.get("/api/general_info")
 def handle_get_general_info():
     def perform(args):
+        global service
         return dataclasses.asdict(service.get_general_info())
 
     return safe_call(perform)
 
 
-@app.get("/api/window_service_info")
-def handle_get_window_service_info():
-    def perform(args):
-        return dataclasses.asdict(service.get_window_service_info(args["model_name"]))
-
-    return safe_call(perform)
-
-
 @app.post("/api/account")
 def handle_create_account():
     def perform(args):
+        global service
         auth = Authentication(**json.loads(args["auth"]))
         return dataclasses.asdict(service.create_account(auth))
 
@@ -115,6 +120,7 @@ def handle_create_account():
 @app.delete("/api/account")
 def handle_delete_account():
     def perform(args):
+        global service
         auth = Authentication(**json.loads(args["auth"]))
         api_key = args["api_key"]
         return dataclasses.asdict(service.delete_account(auth, api_key))
@@ -125,6 +131,7 @@ def handle_delete_account():
 @app.get("/api/account")
 def handle_get_account():
     def perform(args):
+        global service
         auth = Authentication(**json.loads(args["auth"]))
         if "all" in args and args["all"].lower() == "true":
             return [dataclasses.asdict(account) for account in service.get_accounts(auth)]
@@ -137,6 +144,7 @@ def handle_get_account():
 @app.put("/api/account")
 def handle_update_account():
     def perform(args):
+        global service
         auth = Authentication(**json.loads(args["auth"]))
         account = from_dict(Account, json.loads(args["account"]))
         return dataclasses.asdict(service.update_account(auth, account))
@@ -147,6 +155,7 @@ def handle_update_account():
 @app.put("/api/account/api_key")
 def handle_update_api_key():
     def perform(args):
+        global service
         auth = Authentication(**json.loads(args["auth"]))
         account = from_dict(Account, json.loads(args["account"]))
         return dataclasses.asdict(service.rotate_api_key(auth, account))
@@ -157,6 +166,7 @@ def handle_update_api_key():
 @app.get("/api/query")
 def handle_query():
     def perform(args):
+        global service
         query = Query(**args)
         return dataclasses.asdict(service.expand_query(query))
 
@@ -166,9 +176,28 @@ def handle_query():
 @app.get("/api/request")
 def handle_request():
     def perform(args):
+        global service
         auth = Authentication(**json.loads(args["auth"]))
         request = Request(**json.loads(args["request"]))
-        return dataclasses.asdict(service.make_request(auth, request))
+        # Hack to maintain reverse compatibility with clients with version <= 0.3.0.
+        # Clients with version <= 0.3.0 do not set model_deployment, but this is now
+        # required by Request.
+        if not request.model_deployment:
+            model_deployment = get_default_model_deployment_for_model(request.model)
+            if model_deployment is None:
+                raise ValueError(f"Unknown model '{request.model}'")
+            request = dataclasses.replace(request, model_deployment=model_deployment)
+
+        raw_response = dataclasses.asdict(service.make_request(auth, request))
+
+        # Hack to maintain reverse compatibility with clients with version <= 1.0.0.
+        # Clients with version <= 1.0.0 expect each token to contain a `top_logprobs`
+        # field of type dict.
+        for completion in raw_response["completions"]:
+            for token in completion["tokens"]:
+                token["top_logprobs"] = {}
+
+        return raw_response
 
     return safe_call(perform)
 
@@ -176,6 +205,7 @@ def handle_request():
 @app.get("/api/tokenize")
 def handle_tokenization():
     def perform(args):
+        global service
         auth = Authentication(**json.loads(args["auth"]))
         request = TokenizationRequest(**json.loads(args["request"]))
         return dataclasses.asdict(service.tokenize(auth, request))
@@ -186,6 +216,7 @@ def handle_tokenization():
 @app.get("/api/decode")
 def handle_decode():
     def perform(args):
+        global service
         auth = Authentication(**json.loads(args["auth"]))
         request = DecodeRequest(**json.loads(args["request"]))
         return dataclasses.asdict(service.decode(auth, request))
@@ -196,6 +227,7 @@ def handle_decode():
 @app.get("/api/toxicity")
 def handle_toxicity_request():
     def perform(args):
+        global service
         auth = Authentication(**json.loads(args["auth"]))
         request = PerspectiveAPIRequest(**json.loads(args["request"]))
         return dataclasses.asdict(service.get_toxicity_scores(auth, request))
@@ -203,9 +235,21 @@ def handle_toxicity_request():
     return safe_call(perform)
 
 
+@app.get("/api/moderation")
+def handle_moderation_request():
+    def perform(args):
+        global service
+        auth = Authentication(**json.loads(args["auth"]))
+        request = ModerationAPIRequest(**json.loads(args["request"]))
+        return dataclasses.asdict(service.get_moderation_results(auth, request))
+
+    return safe_call(perform)
+
+
 @app.get("/api/shutdown")
 def handle_shutdown():
     def perform(args):
+        global service
         auth = Authentication(**json.loads(args["auth"]))
         service.shutdown(auth)
 
@@ -218,6 +262,7 @@ def main():
     parser.add_argument("-p", "--port", type=int, help="What port to listen on", default=1959)
     parser.add_argument("--ssl-key-file", type=str, help="Path to SSL key file")
     parser.add_argument("--ssl-cert-file", type=str, help="Path to SSL cert file")
+    parser.add_argument("--ssl-ca-certs", type=str, help="Path to SSL CA certs")
     parser.add_argument("-b", "--base-path", help="What directory has credentials, etc.", default="prod_env")
     parser.add_argument("-w", "--workers", type=int, help="Number of worker processes to handle requests", default=8)
     parser.add_argument("-t", "--timeout", type=int, help="Request timeout in seconds", default=5 * 60)
@@ -228,21 +273,40 @@ def main():
         default="",
     )
     args = parser.parse_args()
+    setup_default_logging()
 
     register_builtin_configs_from_helm_package()
     register_configs_from_directory(args.base_path)
 
-    service = ServerService(base_path=args.base_path, mongo_uri=args.mongo_uri)
+    cache_backend_config: CacheBackendConfig
+    if args.mongo_uri:
+        cache_backend_config = MongoCacheBackendConfig(args.mongo_uri)
+    else:
+        sqlite_cache_path = os.path.join(args.base_path, CACHE_DIR)
+        ensure_directory_exists(sqlite_cache_path)
+        cache_backend_config = SqliteCacheBackendConfig(sqlite_cache_path)
+
+    static_package_name = "helm.proxy.static"
+    resource_path = resources.files(static_package_name).joinpath("index.html")
+    with resources.as_file(resource_path) as resource_filename:
+        static_path = str(resource_filename.parent)
+    app.config["helm.staticpath"] = static_path
+
+    service = ServerService(base_path=args.base_path, cache_backend_config=cache_backend_config)
 
     gunicorn_args = {
         "workers": args.workers,
         "timeout": args.timeout,
         "limit_request_line": 0,  # Controls the maximum size of HTTP request line in bytes. 0 = unlimited.
     }
-    if args.ssl_key_file and args.ssl_cert_file:
+    if args.ssl_key_file:
         gunicorn_args["keyfile"] = args.ssl_key_file
+    if args.ssl_cert_file:
         gunicorn_args["certfile"] = args.ssl_cert_file
+    if args.ssl_ca_certs:
+        gunicorn_args["ca_certs"] = args.ssl_ca_certs
 
     # Clear arguments before running gunicorn as it also uses argparse
     sys.argv = [sys.argv[0]]
+    app.config["crfm.proxy.outputpath"] = os.path.join(os.path.realpath(args.base_path), "cache", "output")
     app.run(host="0.0.0.0", port=args.port, server="gunicorn", **gunicorn_args)