download vast.ai's root certificate in order to make pyworker requests

lib/backend.py

@@ -126,7 +126,7 @@ class Backend:
         async def cancel_api_call_if_disconnected() -> web.Response:
             await request.wait_for_disconnection()
             log.debug(f"request with reqnum: {auth_data.reqnum} was canceled")
-            self.metrics._request_canceled(workload=workload)
+            self.metrics._request_canceled(workload=workload, reqnum=auth_data.reqnum)
             return web.Response(status=500)
 
         async def make_request() -> Union[web.Response, web.StreamResponse]:
@@ -141,6 +141,7 @@ class Backend:
             else:
                 log.debug(f"Starting request for reqnum:{auth_data.reqnum}")
             try:
+                start_time = time.time()
                 response = await self.__call_api(handler=handler, payload=payload)
                 status_code = response.status
                 log.debug(
@@ -152,17 +153,19 @@ class Backend:
                     )
                 )
                 res = await handler.generate_client_response(request, response)
-                self.metrics._request_success(workload=workload)
+                self.metrics._request_end(
+                    workload=workload,
+                    req_response_time=time.time() - start_time,
+                    reqnum=auth_data.reqnum,
+                )
                 return res
             except requests.exceptions.RequestException as e:
                 log.debug(f"[backend] Request error: {e}")
-                self.metrics._request_errored(workload=workload)
+                self.metrics._request_errored(
+                    workload=workload, reqnum=auth_data.reqnum
+                )
                 return web.Response(status=500)
             finally:
-                self.metrics._request_end(
-                    workload=workload,
-                    reqnum=auth_data.reqnum,
-                )
                 self.sem.release()
 
         ###########
@@ -183,6 +186,12 @@ class Backend:
         except Exception as e:
             log.debug(f"Exception in main handler loop {e}")
             return web.Response(status=500)
+        finally:
+            if request.task.cancelled():
+                log.debug(f"request with reqnum: {auth_data.reqnum} was canceled")
+                self.metrics._request_canceled(
+                    workload=workload, reqnum=auth_data.reqnum
+                )
 
     async def __healthcheck(self):
         health_check_url = self.benchmark_handler.healthcheck_endpoint

lib/data_types.py

@@ -8,6 +8,7 @@ from aiohttp import web, ClientResponse
 import inspect
 
 import psutil
+import requests
 
 
 """
@@ -205,13 +206,13 @@ class ModelMetrics:
     workload_received: float
     workload_cancelled: float
     workload_errored: float
-    # these are not
     workload_pending: float
+    # these are not
+    cur_perf: float
     error_msg: Optional[str]
     max_throughput: float
     requests_recieved: Set[int] = field(default_factory=set)
     requests_working: Set[int] = field(default_factory=set)
-    last_update: float = field(default_factory=time.time)
 
     @classmethod
     def empty(cls):
@@ -220,15 +221,12 @@ class ModelMetrics:
             workload_served=0.0,
             workload_cancelled=0.0,
             workload_errored=0.0,
+            cur_perf=0.0,
             workload_received=0.0,
             error_msg=None,
             max_throughput=0.0,
         )
 
-    @property
-    def cur_perf(self) -> float:
-        return max(self.workload_served / (time.time() - self.last_update), 0.0)
-
     @property
     def workload_processing(self) -> float:
         return max(self.workload_received - self.workload_cancelled, 0.0)
@@ -242,7 +240,6 @@ class ModelMetrics:
         self.workload_received = 0
         self.workload_cancelled = 0
         self.workload_errored = 0
-        self.last_update = time.time()
 
 
 @dataclass

lib/metrics.py

@@ -46,31 +46,33 @@ class Metrics:
         self.model_metrics.requests_recieved.add(reqnum)
         self.model_metrics.requests_working.add(reqnum)
 
-    def _request_end(self, workload: float, reqnum: int) -> None:
+    def _request_end(
+        self, workload: float, req_response_time: float, reqnum: int
+    ) -> None:
         """
-        this function is called after handling of a request ends, regardless of the outcome
-        """
-        self.model_metrics.workload_pending -= workload
-        self.model_metrics.requests_working.discard(reqnum)
-
-    def _request_success(self, workload: float) -> None:
-        """
-        this function is called after a response from model API is received and forwarded.
+        this function is called after a response from model API is received.
         """
         self.model_metrics.workload_served += workload
+        self.model_metrics.workload_pending -= workload
+        self.model_metrics.requests_working.discard(reqnum)
+        self.model_metrics.cur_perf = workload / req_response_time
         self.update_pending = True
 
-    def _request_errored(self, workload: float) -> None:
+    def _request_errored(self, workload: float, reqnum: int) -> None:
         """
         this function is called if model API returns an error
         """
+        self.model_metrics.workload_pending -= workload
         self.model_metrics.workload_errored += workload
+        self.model_metrics.requests_working.discard(reqnum)
 
-    def _request_canceled(self, workload: float) -> None:
+    def _request_canceled(self, workload: float, reqnum: int) -> None:
         """
         this function is called if client drops connection before model API has responded
         """
+        self.model_metrics.workload_pending -= workload
         self.model_metrics.workload_cancelled += workload
+        self.model_metrics.requests_working.discard(reqnum)
 
     async def _send_metrics_loop(self) -> Awaitable[NoReturn]:
         while True:

lib/test_utils.py

@@ -10,6 +10,7 @@ from collections import Counter
 from dataclasses import dataclass, field, asdict
 from urllib.parse import urljoin
 from utils.endpoint_util import Endpoint
+from utils.ssl import get_cert_file_path
 import requests
 
 from lib.data_types import AuthData, ApiPayload
@@ -120,9 +121,11 @@ class ClientState:
         self.url = worker_address
         url = urljoin(worker_address, self.worker_endpoint)
         self.status = ClientStatus.Generating
+
         response = requests.post(
             url,
             json=req_data,
+            verify=get_cert_file_path(),
         )
         if response.status_code != 200:
             self.infer_error.append(

requirements.txt

@@ -1,4 +1,4 @@
-aiohttp[speedups]==3.10.1
+aiohttp==3.10.1
 anyio~=4.4
 lib~=4.0
 nltk~=3.9
@@ -6,5 +6,5 @@ psutil~=6.0
 pycryptodome~=3.20
 Requests~=2.32
 transformers~=4.52
-utils==1.0.*
+utils~=1.0
 hf_transfer>=0.1.9

utils/ssl.py

@@ -0,0 +1,15 @@
+import tempfile
+from functools import cache
+
+import requests
+
+
+@cache
+def get_cert_file_path():
+    cert_url = "https://console.vast.ai/static/jvastai_root.cer"
+    response = requests.get(cert_url)
+    response.raise_for_status()
+    # Use a temporary file that is not deleted on close
+    with tempfile.NamedTemporaryFile(delete=False, suffix=".cer", mode="wb") as f:
+        f.write(response.content)
+        return f.name

workers/comfyui/client.py

@@ -5,6 +5,7 @@ import requests
 
 from lib.test_utils import print_truncate_res
 from utils.endpoint_util import Endpoint
+from utils.ssl import get_cert_file_path
 
 """
 NOTE: this client example uses a custom comfy workflow compatible with SD3 only
@@ -51,6 +52,7 @@ def call_default_workflow(
     response = requests.post(
         url,
         json=req_data,
+        verify=get_cert_file_path(),
     )
     response.raise_for_status()
     print_truncate_res(str(response.json()))
@@ -141,6 +143,7 @@ def call_custom_workflow_for_sd3(
     response = requests.post(
         url,
         json=req_data,
+        verify=get_cert_file_path(),
     )
     response.raise_for_status()
     print_truncate_res(str(response.json()))

workers/openai/client.py

@@ -6,6 +6,7 @@ from urllib.parse import urljoin
 from typing import Dict, Any, Optional, Iterator, Union, List
 import requests
 from utils.endpoint_util import Endpoint
+from utils.ssl import get_cert_file_path
 from .data_types.client import CompletionConfig, ChatCompletionConfig
 
 logging.basicConfig(
@@ -90,9 +91,13 @@ class APIClient:
 
         # Make the request using the specified method
         if method.upper() == "POST":
-            response = requests.post(url, json=req_data, stream=stream)
+            response = requests.post(
+                url, json=req_data, stream=stream, verify=get_cert_file_path()
+            )
         elif method.upper() == "GET":
-            response = requests.get(url, params=req_data, stream=stream)
+            response = requests.get(
+                url, params=req_data, stream=stream, verify=get_cert_file_path()
+            )
         else:
             raise ValueError(f"Unsupported HTTP method: {method}")
 

workers/tgi/client.py

@@ -4,6 +4,7 @@ import json
 from urllib.parse import urljoin
 import requests
 from utils.endpoint_util import Endpoint
+from utils.ssl import get_cert_file_path
 
 logging.basicConfig(
     level=logging.DEBUG,
@@ -42,7 +43,11 @@ def call_generate(endpoint_group_name: str, api_key: str, server_url: str) -> No
     req_data = dict(payload=payload, auth_data=auth_data)
     url = urljoin(url, WORKER_ENDPOINT)
     print(f"url: {url}")
-    response = requests.post(url, json=req_data)
+    response = requests.post(
+        url,
+        json=req_data,
+        verify=get_cert_file_path(),
+    )
     response.raise_for_status()
     res = response.json()
     print(res)