graphing errors over time

Fix healthcheck with separate session

lib/backend.py

@@ -11,7 +11,8 @@ from functools import cached_property
 from distutils.util import strtobool
 
 from anyio import open_file
-from aiohttp import web, ClientResponse, ClientSession, ClientConnectorError
+from aiohttp import web, ClientResponse, ClientSession, ClientConnectorError, ClientTimeout, TCPConnector
+import asyncio
 
 import requests
 from Crypto.Signature import pkcs1_15
@@ -25,8 +26,11 @@ from lib.data_types import (
     LogAction,
     ApiPayload_T,
     JsonDataException,
+    RequestMetrics
 )
 
+VERSION = "0.1.0"
+
 MSG_HISTORY_LEN = 100
 log = logging.getLogger(__file__)
 
@@ -53,7 +57,9 @@ class Backend:
         EndpointHandler  # this endpoint handler will be used for benchmarking
     )
     log_actions: List[Tuple[LogAction, str]]
+    max_wait_time: float = 10.0
     reqnum = -1
+    version = VERSION
     msg_history = []
     sem: Semaphore = dataclasses.field(default_factory=Semaphore)
     unsecured: bool = dataclasses.field(
@@ -62,6 +68,7 @@ class Backend:
 
     def __post_init__(self):
         self.metrics = Metrics()
+        self.metrics._set_version(self.version)
         self._total_pubkey_fetch_errors = 0
         self._pubkey = self._fetch_pubkey()
         self.__start_healthcheck: bool = False
@@ -75,7 +82,13 @@ class Backend:
     @cached_property
     def session(self):
         log.debug(f"starting session with {self.model_server_url}")
-        return ClientSession(self.model_server_url)
+        connector = TCPConnector(
+            force_close=True, # Required for long running jobs
+            enable_cleanup_closed=True,
+        )
+        
+        timeout = ClientTimeout(total=None)
+        return ClientSession(self.model_server_url, timeout=timeout, connector=connector)
 
     def create_handler(
         self,
@@ -122,58 +135,56 @@ class Backend:
         except json.JSONDecodeError:
             return web.json_response(dict(error="invalid JSON"), status=422)
         workload = payload.count_workload()
+        request_metrics: RequestMetrics = RequestMetrics(request_idx=auth_data.request_idx, reqnum=auth_data.reqnum, workload=workload, status="Created")
 
         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, reqnum=auth_data.reqnum)
-            return web.Response(status=500)
+            log.debug(f"request with reqnum: {request_metrics.reqnum} was canceled")
+            self.metrics._request_canceled(request_metrics)
+            raise asyncio.CancelledError
 
         async def make_request() -> Union[web.Response, web.StreamResponse]:
-            log.debug(f"got request, {auth_data.reqnum}")
-            self.metrics._request_start(workload=workload, reqnum=auth_data.reqnum)
-            if self.allow_parallel_requests is False:
-                log.debug(f"Waiting to aquire Sem for reqnum:{auth_data.reqnum}")
-                await self.sem.acquire()
-                log.debug(
-                    f"Sem acquired for reqnum:{auth_data.reqnum}, starting request..."
-                )
-            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(
                     " ".join(
                         [
-                            f"request with reqnum:{auth_data.reqnum}",
+                            f"request with reqnum:{request_metrics.reqnum}",
                             f"returned status code: {status_code},",
                         ]
                     )
                 )
                 res = await handler.generate_client_response(request, response)
-                self.metrics._request_end(
-                    workload=workload,
-                    req_response_time=time.time() - start_time,
-                    reqnum=auth_data.reqnum,
-                )
+                self.metrics._request_success(request_metrics)
                 return res
             except requests.exceptions.RequestException as e:
                 log.debug(f"[backend] Request error: {e}")
-                self.metrics._request_errored(
-                    workload=workload, reqnum=auth_data.reqnum
-                )
+                self.metrics._request_errored(request_metrics)
                 return web.Response(status=500)
-            finally:
-                self.sem.release()
 
         ###########
 
         if self.__check_signature(auth_data) is False:
+            self.metrics._request_reject(request_metrics)
             return web.Response(status=401)
         
+        if self.metrics.model_metrics.wait_time > self.max_wait_time:
+            self.metrics._request_reject(request_metrics)
+            return web.Response(status=429)
+
+        acquired = False
         try:
+            self.metrics._request_start(request_metrics)
+            if self.allow_parallel_requests is False:
+                log.debug(f"Waiting to aquire Sem for reqnum:{request_metrics.reqnum}")
+                await self.sem.acquire()
+                acquired = True
+                log.debug(
+                    f"Sem acquired for reqnum:{request_metrics.reqnum}, starting request..."
+                )
+            else:
+                log.debug(f"Starting request for reqnum:{request_metrics.reqnum}")
             done, pending = await wait(
                 [
                     create_task(make_request()),
@@ -181,30 +192,52 @@ class Backend:
                 ],
                 return_when=FIRST_COMPLETED,
             )
-            [task.cancel() for task in pending]
-            return done.pop().result()
+            for t in pending:
+                t.cancel()
+            await asyncio.gather(*pending, return_exceptions=True)
+
+            done_task = done.pop()
+            try:
+                return done_task.result()
+            except Exception as e:
+                log.debug(f"Request task raised exception: {e}")
+                return web.Response(status=500)
+        except asyncio.CancelledError:
+            # Client is gone. Do not write a response; just unwind.
+            return web.Response(status=499) 
         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
+            # Always release the semaphore if it was acquired
+            if acquired:
+                self.sem.release()
+            self.metrics._request_end(request_metrics)
+
+    @cached_property  
+    def healthcheck_session(self):
+        """Dedicated session for healthchecks to avoid conflicts with API session"""
+        log.debug("creating dedicated healthcheck session")
+        connector = TCPConnector(
+            force_close=True,  # Keep this for isolation
+            enable_cleanup_closed=True,
         )
+        timeout = ClientTimeout(total=10)  # Reasonable timeout for healthchecks
+        return ClientSession(timeout=timeout, connector=connector)
 
     async def __healthcheck(self):
         health_check_url = self.benchmark_handler.healthcheck_endpoint
         if health_check_url is None:
             log.debug("No healthcheck endpoint defined, skipping healthcheck")
             return
+
         while True:
             await sleep(10)
             if self.__start_healthcheck is False:
                 continue
             try:
                 log.debug(f"Performing healthcheck on {health_check_url}")
-                async with self.session.get(health_check_url) as response:
+                async with self.healthcheck_session.get(health_check_url) as response:
                     if response.status == 200:
                         log.debug("Healthcheck successful")
                     elif response.status == 503:
@@ -213,7 +246,6 @@ class Backend:
                             f"Healthcheck failed with status: {response.status}"
                         )
                     else:
-                        # endpoint not ready yet so bail
                         log.debug(f"Healthcheck Endpoint not ready: {response.status}")
             except Exception as e:
                 log.debug(f"Healthcheck failed with exception: {e}")
@@ -221,7 +253,7 @@ class Backend:
 
     async def _start_tracking(self) -> None:
         await gather(
-            self.__read_logs(), self.metrics._send_metrics_loop(), self.__healthcheck()
+            self.__read_logs(), self.metrics._send_metrics_loop(), self.__healthcheck(), self.metrics._send_delete_requests_loop()
         )
 
     def backend_errored(self, msg: str) -> None:
@@ -289,41 +321,52 @@ class Backend:
                     return float(f.readline())
             except FileNotFoundError:
                 pass
-            max_throughput = 0
-            last_throughput = 0
-            sum_throughput = 0
-            for run in range(self.benchmark_handler.benchmark_runs + 1):
-                start = time.time()
+
+            log.debug("Initial run to trigger model loading...")
             payload = self.benchmark_handler.make_benchmark_payload()
-                res = await self.__call_api(
-                    handler=self.benchmark_handler, payload=payload
+            await self.__call_api(handler=self.benchmark_handler, payload=payload)
+
+            max_throughput = 0
+            sum_throughput = 0
+            concurrent_requests = 10 if self.allow_parallel_requests else 1
+
+            for run in range(1, self.benchmark_handler.benchmark_runs + 1):
+                start = time.time()
+                tasks = []
+                total_workload = 0
+
+                for _ in range(concurrent_requests):
+                    payload = self.benchmark_handler.make_benchmark_payload()
+                    total_workload += payload.count_workload()
+                    tasks.append(
+                        self.__call_api(handler=self.benchmark_handler, payload=payload)
                     )
-                data = await res.json()
+
+                responses = await gather(*tasks)
                 time_elapsed = time.time() - start
-                # first run triggers one-time loading of the model which is very slow, so we skip counting it
-                if run == 0:
-                    continue
-                else:
-                    workload = payload.count_workload()
-                    last_throughput = workload / time_elapsed
-                    sum_throughput += last_throughput
-                    max_throughput = max(max_throughput, last_throughput)
+
+                throughput = total_workload / time_elapsed
+                sum_throughput += throughput
+                max_throughput = max(max_throughput, throughput)
+
+                # Log results for debugging
                 log.debug(
                     "\n".join(
                         [
                             "#" * 60,
-                                f"Run: {run}, workload: {workload} time_elapsed: {time_elapsed}, throughput: {last_throughput}",
-                                "",
-                                f"response: {data}",
+                            f"Run: {run}, concurrent_requests: {concurrent_requests}",
+                            f"Total workload: {total_workload}, time_elapsed: {time_elapsed}s",
+                            f"Throughput: {throughput} workload/s",
+                            f"Successful responses: {len([r for r in responses if r.status == 200])}",
                             "#" * 60,
                         ]
                     )
                 )
+
             average_throughput = sum_throughput / self.benchmark_handler.benchmark_runs
             log.debug(
                 f"benchmark result: avg {average_throughput} workload per second, max {max_throughput}"
             )
-            # save max_throughput so we don't have to run benchmark again on restart of cold instances
             with open(BENCHMARK_INDICATOR_FILE, "w") as f:
                 f.write(str(max_throughput))
             return max_throughput
@@ -368,7 +411,7 @@ class Backend:
                     if line:
                         await handle_log_line(line.rstrip())
                     else:
-                        time.sleep(LOG_POLL_INTERVAL)
+                        await asyncio.sleep(LOG_POLL_INTERVAL)
 
         ###########
 

lib/data_types.py

@@ -8,7 +8,6 @@ from aiohttp import web, ClientResponse
 import inspect
 
 import psutil
-import requests
 
 
 """
@@ -71,6 +70,7 @@ class AuthData:
     endpoint: str
     reqnum: int
     url: str
+    request_idx: int
 
     @classmethod
     def from_json_msg(cls, json_msg: Dict[str, Any]):
@@ -197,6 +197,15 @@ class SystemMetrics:
         self.model_loading_time = None
 
 
+@dataclass
+class RequestMetrics:
+    """Tracks metrics for an active request."""
+    request_idx: int
+    reqnum: int
+    workload: float
+    status: str
+    success: bool = False
+
 @dataclass
 class ModelMetrics:
     """Model specific metrics"""
@@ -206,13 +215,15 @@ class ModelMetrics:
     workload_received: float
     workload_cancelled: float
     workload_errored: float
-    workload_pending: float
+    workload_rejected: float
     # these are not
-    cur_perf: float
+    workload_pending: float
     error_msg: Optional[str]
     max_throughput: float
     requests_recieved: Set[int] = field(default_factory=set)
-    requests_working: Set[int] = field(default_factory=set)
+    requests_working: dict[int, RequestMetrics] = field(default_factory=dict)
+    requests_deleting: list[RequestMetrics] = field(default_factory=list)
+    last_update: float = field(default_factory=time.time)
 
     @classmethod
     def empty(cls):
@@ -221,7 +232,7 @@ class ModelMetrics:
             workload_served=0.0,
             workload_cancelled=0.0,
             workload_errored=0.0,
-            cur_perf=0.0,
+            workload_rejected=0.0,
             workload_received=0.0,
             error_msg=None,
             max_throughput=0.0,
@@ -231,6 +242,20 @@ class ModelMetrics:
     def workload_processing(self) -> float:
         return max(self.workload_received - self.workload_cancelled, 0.0)
 
+    @property
+    def wait_time(self) -> float:
+        if (len(self.requests_working) == 0):
+            return 0.0
+        return sum([request.workload for request in self.requests_working.values()]) / self.max_throughput
+    
+    @property
+    def cur_load(self) -> float:
+        return sum([request.workload for request in self.requests_working.values()])
+
+    @property
+    def working_request_idxs(self) -> list[int]:
+        return [req.request_idx for req in self.requests_working.values()]
+
     def set_errored(self, error_msg):
         self.reset()
         self.error_msg = error_msg
@@ -240,15 +265,20 @@ class ModelMetrics:
         self.workload_received = 0
         self.workload_cancelled = 0
         self.workload_errored = 0
+        self.workload_rejected = 0
+        self.last_update = time.time()
 
 
 @dataclass
-class AutoScalaerData:
+class AutoScalerData:
     """Data that is reported to autoscaler"""
 
     id: int
+    version: str
     loadtime: float
     cur_load: float
+    rej_load: float
+    new_load: float
     error_msg: str
     max_perf: float
     cur_perf: float
@@ -257,6 +287,7 @@ class AutoScalaerData:
     num_requests_working: int
     num_requests_recieved: int
     additional_disk_usage: float
+    working_request_idxs: list[int]
     url: str
 
 

lib/metrics.py

@@ -5,13 +5,14 @@ import json
 from asyncio import sleep
 from dataclasses import dataclass, asdict, field
 from functools import cache
+import asyncio
+from aiohttp import ClientSession, ClientTimeout, TCPConnector, ClientResponseError
 
-import requests
-
-from lib.data_types import AutoScalaerData, SystemMetrics, ModelMetrics
+from lib.data_types import AutoScalerData, SystemMetrics, ModelMetrics, RequestMetrics
 from typing import Awaitable, NoReturn, List
 
 METRICS_UPDATE_INTERVAL = 1
+DELETE_REQUESTS_INTERVAL = 1
 
 log = logging.getLogger(__file__)
 
@@ -26,7 +27,9 @@ def get_url() -> str:
 
 @dataclass
 class Metrics:
+    version: str = "0"
     last_metric_update: float = 0.0
+    last_request_served: float = 0.0
     update_pending: bool = False
     id: int = field(default_factory=lambda: int(os.environ["CONTAINER_ID"]))
     report_addr: List[str] = field(
@@ -35,44 +38,84 @@ class Metrics:
     url: str = field(default_factory=get_url)
     system_metrics: SystemMetrics = field(default_factory=SystemMetrics.empty)
     model_metrics: ModelMetrics = field(default_factory=ModelMetrics.empty)
+    _session: ClientSession | None = field(default=None, init=False, repr=False)
 
-    def _request_start(self, workload: float, reqnum: int) -> None:
+    async def http(self) -> ClientSession:
+        if self._session is None:
+            self._session = ClientSession(
+                timeout=ClientTimeout(total=10),
+                connector=TCPConnector(limit=8, limit_per_host=4, force_close=True, enable_cleanup_closed=True)
+            )
+        return self._session
+    
+    async def aclose(self) -> None:
+        if self._session is not None:
+            await self._session.close()
+            self._session = None
+
+    def _request_start(self, request: RequestMetrics) -> None:
         """
         this function is called prior to forwarding a request to a model API.
         """
         log.debug("request start")
-        self.model_metrics.workload_pending += workload
-        self.model_metrics.workload_received += workload
-        self.model_metrics.requests_recieved.add(reqnum)
-        self.model_metrics.requests_working.add(reqnum)
-
-    def _request_end(
-        self, workload: float, req_response_time: float, reqnum: int
-    ) -> None:
-        """
-        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
+        request.status = "Started"
+        self.model_metrics.workload_pending += request.workload
+        self.model_metrics.workload_received += request.workload
+        self.model_metrics.requests_recieved.add(request.reqnum)
+        self.model_metrics.requests_working[request.reqnum] = request
         self.update_pending = True
 
-    def _request_errored(self, workload: float, reqnum: int) -> None:
+    def _request_end(self, request: RequestMetrics) -> None:
+        """
+        this function is called after handling of a request ends, regardless of the outcome
+        """
+        self.model_metrics.workload_pending -= request.workload
+        self.model_metrics.requests_working.pop(request.reqnum, None)
+        self.model_metrics.requests_deleting.append(request)
+        self.last_request_served = time.time()
+
+    def _request_success(self, request: RequestMetrics) -> None:
+        """
+        this function is called after a response from model API is received and forwarded.
+        """
+        self.model_metrics.workload_served += request.workload
+        request.status = "Success"
+        request.success = True
+        self.update_pending = True
+
+    def _request_errored(self, request: RequestMetrics) -> 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)
+        self.model_metrics.workload_errored += request.workload
+        request.status = "Error"
+        request.success = False
+        self.update_pending = True
 
-    def _request_canceled(self, workload: float, reqnum: int) -> None:
+    def _request_canceled(self, request: RequestMetrics) -> 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)
+        self.model_metrics.workload_cancelled += request.workload
+        request.success = True
+        request.status = "Cancelled"
+    
+    def _request_reject(self, request: RequestMetrics):
+        """
+        this function is called if the current wait time for the model is above max_wait_time
+        """
+        self.model_metrics.requests_recieved.add(request.reqnum)
+        self.model_metrics.requests_deleting.append(request)
+        self.model_metrics.workload_rejected += request.workload
+        request.success = False
+        request.status = "Rejected"
+        self.update_pending = True
+
+    async def _send_delete_requests_loop(self) -> Awaitable[NoReturn]:
+        while True:
+            await sleep(DELETE_REQUESTS_INTERVAL)
+            if len(self.model_metrics.requests_deleting) > 0:
+                await self.__send_delete_requests_and_reset()
 
     async def _send_metrics_loop(self) -> Awaitable[NoReturn]:
         while True:
@@ -80,10 +123,10 @@ class Metrics:
             elapsed = time.time() - self.last_metric_update
             if self.system_metrics.model_is_loaded is False and elapsed >= 10:
                 log.debug(f"sending loading model metrics after {int(elapsed)}s wait")
-                self.__send_metrics_and_reset(elapsed)
+                await self.__send_metrics_and_reset()
             elif self.update_pending or elapsed > 10:
                 log.debug(f"sending loaded model metrics after {int(elapsed)}s wait")
-                self.__send_metrics_and_reset(elapsed)
+                await self.__send_metrics_and_reset()
 
     def _model_loaded(self, max_throughput: float) -> None:
         self.system_metrics.model_loading_time = (
@@ -96,27 +139,63 @@ class Metrics:
         self.model_metrics.set_errored(error_msg)
         self.system_metrics.model_is_loaded = True
 
+    def _set_version(self, version: str) -> None:
+        self.version = version
+
     #######################################Private#######################################
 
-    def __send_metrics_and_reset(self, elapsed):
+    async def __send_delete_requests_and_reset(self):
 
-        def compute_autoscaler_data() -> AutoScalaerData:
-            return AutoScalaerData(
+        async def send_data(report_addr: str, success: bool) -> bool:
+            data = {
+                "worker_id": self.id,
+                "request_idxs": [r.request_idx for r in self.model_metrics.requests_deleting if r.success == success],
+                "success": success
+            }
+            full_path = report_addr.rstrip("/") + "/delete_requests/"
+            for attempt in range(1, 4):
+                try:
+                    session = await self.http()
+                    async with session.post(full_path, json=data) as res:
+                        res.raise_for_status()
+                    return True
+                except asyncio.TimeoutError:
+                    log.debug(f"delete_requests timed out")
+                except (ClientResponseError, Exception) as e:
+                    log.debug(f"delete_requests failed with error: {e}")
+                await asyncio.sleep(2)
+                log.debug(f"retrying delete_request, attempt: {attempt}")
+
+        for report_addr in self.report_addr:
+            success = await send_data(report_addr, success=True) and await send_data(report_addr, success=False)
+            if success is True:
+                self.model_metrics.requests_deleting.clear()
+                break
+
+
+    async def __send_metrics_and_reset(self):
+
+        def compute_autoscaler_data() -> AutoScalerData:
+            return AutoScalerData(
                 id=self.id,
+                version=self.version,
                 loadtime=(self.system_metrics.model_loading_time or 0.0),
-                cur_load=(self.model_metrics.workload_processing / elapsed),
+                new_load=self.model_metrics.workload_processing,
+                cur_load=self.model_metrics.cur_load,
+                rej_load=self.model_metrics.workload_rejected,
                 max_perf=self.model_metrics.max_throughput,
-                cur_perf=self.model_metrics.cur_perf,
+                cur_perf=self.model_metrics.workload_served,
                 error_msg=self.model_metrics.error_msg or "",
                 num_requests_working=len(self.model_metrics.requests_working),
                 num_requests_recieved=len(self.model_metrics.requests_recieved),
                 additional_disk_usage=self.system_metrics.additional_disk_usage,
+                working_request_idxs=self.model_metrics.working_request_idxs,
                 cur_capacity=0,
                 max_capacity=0,
                 url=self.url,
             )
 
-        def send_data(report_addr: str) -> None:
+        async def send_data(report_addr: str) -> bool:
             data = compute_autoscaler_data()
             full_path = report_addr.rstrip("/") + "/worker_status/"
             log.debug(
@@ -131,21 +210,27 @@ class Metrics:
             )
             for attempt in range(1, 4):
                 try:
-                    requests.post(full_path, json=asdict(data), timeout=1)
-                    break
-                except requests.Timeout:
+                    session = await self.http()
+                    async with session.post(full_path, json=asdict(data)) as res:
+                        res.raise_for_status()
+                    return True
+                except asyncio.TimeoutError:
                     log.debug(f"autoscaler status update timed out")
-                except Exception as e:
+                except (ClientResponseError, Exception)  as e:
                     log.debug(f"autoscaler status update failed with error: {e}")
-                time.sleep(2)
+                await asyncio.sleep(2)
                 log.debug(f"retrying autoscaler status update, attempt: {attempt}")
+            log.debug(f"failed to send update through {report_addr}")
+            return False
 
         ###########
 
         self.system_metrics.update_disk_usage()
 
         for report_addr in self.report_addr:
-            send_data(report_addr)
+            success = await send_data(report_addr)
+            if success is True:
+                break
         self.update_pending = False
         self.model_metrics.reset()
         self.system_metrics.reset()

lib/server.py

@@ -27,7 +27,7 @@ def start_server(backend: Backend, routes: List[web.RouteDef], **kwargs):
         log.debug("starting server...")
         app = web.Application()
         app.add_routes(routes)
-        runner = web.AppRunner(app, handler_cancellation=True)
+        runner = web.AppRunner(app)
         await runner.setup()
         site = web.TCPSite(
             runner,

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(
@@ -289,12 +292,12 @@ def test_load_cmd(
     args = arg_parser.parse_args()
     if hasattr(args, "comfy_model"):
         os.environ["COMFY_MODEL"] = args.comfy_model
-    server_url = dict(
-        prod="https://run.vast.ai",
-        alpha="https://run-alpha.vast.ai",
-        candidate="https://run-candidate.vast.ai",
-        local="http://localhost:8080",
-    )[args.instance]
+    server_url = {
+        "prod": "https://run.vast.ai",
+        "alpha": "https://run-alpha.vast.ai",
+        "candidate": "https://run-candidate.vast.ai",
+        "local": "http://localhost:8080",
+    }.get(args.instance, "http://localhost:8080")
     run_test(
         num_requests=args.num_requests,
         requests_per_second=args.requests_per_second,

requirements.txt

@@ -1,4 +1,4 @@
-aiohttp==3.10.1
+aiohttp[speedups]==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

start_server.sh

@@ -9,7 +9,7 @@ ENV_PATH="$WORKSPACE_DIR/worker-env"
 DEBUG_LOG="$WORKSPACE_DIR/debug.log"
 PYWORKER_LOG="$WORKSPACE_DIR/pyworker.log"
 
-REPORT_ADDR="${REPORT_ADDR:-https://run.vast.ai}"
+REPORT_ADDR="${REPORT_ADDR:-https://cloud.vast.ai/api/v0,https://run.vast.ai}"
 USE_SSL="${USE_SSL:-true}"
 WORKER_PORT="${WORKER_PORT:-3000}"
 mkdir -p "$WORKSPACE_DIR"
@@ -41,24 +41,37 @@ echo_var DEBUG_LOG
 echo_var PYWORKER_LOG
 echo_var MODEL_LOG
 
-env | grep _ >> /etc/environment;
-
+# Populate /etc/environment with quoted values
+if ! grep -q "VAST" /etc/environment; then
+    env -0 | grep -zEv "^(HOME=|SHLVL=)|CONDA" | while IFS= read -r -d '' line; do
+            name=${line%%=*}
+            value=${line#*=}
+            printf '%s="%s"\n' "$name" "$value"
+        done > /etc/environment
+fi
 
 if [ ! -d "$ENV_PATH" ]
 then
     echo "setting up venv"
+    if ! which uv; then
         curl -LsSf https://astral.sh/uv/install.sh | sh
         source ~/.local/bin/env
-    git clone https://github.com/vast-ai/pyworker "$SERVER_DIR"
+    fi
 
-    uv venv --managed-python "$WORKSPACE_DIR/worker-env" -p 3.10
-    source "$WORKSPACE_DIR/worker-env/bin/activate"
+    # Fork testing
+    [[ ! -d $SERVER_DIR ]] && git clone "${PYWORKER_REPO:-https://github.com/vast-ai/pyworker}" "$SERVER_DIR"
+    if [[ -n ${PYWORKER_REF:-} ]]; then
+        (cd "$SERVER_DIR" && git checkout "$PYWORKER_REF")
+    fi
 
-    uv pip install -r vast-pyworker/requirements.txt
+    uv venv --python-preference only-managed "$ENV_PATH" -p 3.10
+    source "$ENV_PATH/bin/activate"
+
+    uv pip install -r "${SERVER_DIR}/requirements.txt"
 
     touch ~/.no_auto_tmux
 else
-    source ~/.local/bin/env
+    [[ -f ~/.local/bin/env ]] && source ~/.local/bin/env
     source "$WORKSPACE_DIR/worker-env/bin/activate"
     echo "environment activated"
     echo "venv: $VIRTUAL_ENV"

utils/endpoint_util.py

@@ -1,5 +1,6 @@
 import logging
-from typing import Any, Dict, Optional
+import time
+from typing import Any, Dict, Optional, Tuple
 
 import requests
 
@@ -16,6 +17,60 @@ class Endpoint:
     Utility class for handling endpoint operations.
     """
 
+    @staticmethod
+    def get_endpoint_info(
+        endpoint_name: str, account_api_key: str, instance: str
+    ) -> Optional[Dict[str, Any]]:
+        headers = {"Authorization": f"Bearer {account_api_key}"}
+        url = f"{Endpoint.get_server_url(instance)}?autoscaler_instance={instance}"
+        # Retry a few times to smooth over transient propagation/network delays
+        for attempt in range(4):
+            try:
+                response = requests.get(url, headers=headers, timeout=8)
+                if response.status_code != 200:
+                    # brief backoff and retry
+                    time.sleep(0.3 * (attempt + 1))
+                    continue
+                try:
+                    data = response.json()
+                except Exception:
+                    # JSON parse failed; backoff and retry
+                    time.sleep(0.3 * (attempt + 1))
+                    continue
+                result = data.get("results", []) if isinstance(data, dict) else []
+                endpoint = next(
+                    (item for item in result if item.get("endpoint_name") == endpoint_name),
+                    None,
+                )
+                if endpoint and endpoint.get("id") and endpoint.get("api_key"):
+                    return {"id": endpoint.get("id"), "api_key": endpoint.get("api_key")}
+            except Exception:
+                # network or other transient error; retry
+                time.sleep(0.3 * (attempt + 1))
+        return None
+
+    @staticmethod
+    def get_autoscaler_server_url(instance: str) -> str:
+        endpoints = {
+            "alpha": "run-alpha",
+            "candidate": "run-candidate",
+            "prod": "run",
+        }
+        host = endpoints.get(instance)
+        if host:
+            return f"https://{host}.vast.ai/"
+        return "http://localhost:8080"
+
+    @staticmethod
+    def get_server_url(instance: str) -> str:
+        endpoints = {
+            "alpha": "alpha",
+            "candidate": "candidate",
+            "prod": "console",
+        }
+        host = endpoints.get(instance, "alpha")
+        return f"https://{host}.vast.ai/api/v0/endptjobs/"
+
     @staticmethod
     def get_endpoint_api_key(
         endpoint_name: str, account_api_key: str, instance: str
@@ -30,13 +85,14 @@ class Endpoint:
         Returns:
             Endpoint API key if successful, None otherwise
         """
-        vast_console_url = "https://console.vast.ai/api/v0/endptjobs/"
         headers = {"Authorization": f"Bearer {account_api_key}"}
 
         try:
             log.debug(f"Fetching endpoint API key for endpoint: {endpoint_name}")
             response = requests.get(
-                f"{vast_console_url}?autoscaler_instance={instance}", headers=headers
+                f"{Endpoint.get_server_url(instance)}?autoscaler_instance={instance}",
+                headers=headers,
+                timeout=8,
             )
 
             if response.status_code != 200:
@@ -46,14 +102,14 @@ class Endpoint:
 
             try:
                 data = response.json()
-            except requests.exceptions.JSONDecodeError as e:
+            except Exception as e:
                 log.debug(f"Failed to parse JSON response: {e}")
                 return None
 
             result = data.get("results", [])
 
             endpoint: Optional[Dict[str, Any]] = next(
-                (item for item in result if item["endpoint_name"] == endpoint_name),
+                (item for item in result if item.get("endpoint_name") == endpoint_name),
                 None,
             )
             if not endpoint:

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-json/README.md

@@ -0,0 +1,210 @@
+# ComfyUI PyWorker
+
+This is the base PyWorker for ComfyUI. It provides a unified interface for running any ComfyUI workflow through a proxy-based architecture.
+
+The cost for each request has a static value of `1`.  ComfyUI does not handle concurrent workloads and there is no current provision to load multiple instances of ComfyUI per worker node.
+
+## Requirements
+
+This worker requires both [ComfyUI](https://github.com/comfyanonymous/ComfyUI) and [ComfyUI API Wrapper](https://github.com/ai-dock/comfyui-api-wrapper).
+
+A docker image is provided but you may use any if the above requirements are met.
+
+## Benchmarking
+
+A simple image generation benchmark runs when each worker initializes to validate GPU performance and identify underperforming machines.
+
+The benchmark uses Stable Diffusion v1.5 with ComfyUI's default text-to-image workflow. Configure the benchmark complexity and duration using these variables:
+
+| Environment Variable | Default Value | Description |
+| -------------------- | ------------- | ----------- |
+| BENCHMARK_TEST_WIDTH | 512 | Image width (pixels) |
+| BENCHMARK_TEST_HEIGHT | 512 | Image height (pixels) |
+| BENCHMARK_TEST_STEPS | 20 | Number of denoising steps |
+
+Each benchmark run uses a random prompt from `misc/test_prompts.txt` and a random seed to ensure consistent GPU load patterns.
+
+### Calibrating Benchmark Duration
+
+To screen for underperforming hardware, set `BENCHMARK_TEST_STEPS` to match your expected production workflow duration. This allows you to identify machines that won't meet performance requirements.
+
+**Example:** If your typical workflow should complete in 90 seconds on acceptable hardware:
+
+```bash
+# 1. Measure it/sec on your reference machine
+# RTX 4090 typically achieves ~43 it/sec with SD1.5
+
+# 2. Calculate required steps
+# 90 seconds × 43 it/sec = 3870 steps
+
+# 3. Configure benchmark
+export BENCHMARK_TEST_STEPS=3870
+
+# 4. Machines completing significantly slower than 90s indicate hardware issues
+```
+
+**Performance expectations:**
+- Benchmark duration should remain consistent across identical GPU models
+- Significant variation (>20%) may indicate thermal, power, or configuration issues
+
+## Endpoint
+
+The worker provides a single endpoint:
+
+- `/generate/sync`: Processes ComfyUI workflows using either predefined modifiers or custom workflow JSON
+
+## Request Format
+
+The worker accepts requests in the following format. Choose either modifier mode OR custom workflow mode:
+
+**Modifier Mode:**
+```json
+{
+  "input": {
+    "request_id": "uuid-string",    // optional - UUID generated if not provided
+    "modifier": "RawWorkflow",
+    "modifications": {
+      "prompt": "a beautiful landscape",
+      "width": 1024,
+      "height": 1024,
+      "steps": 20,
+      "seed": 123456789
+    },
+    "s3": { ... },       // optional
+    "webhook": { ... }   // optional
+  }
+}
+```
+
+**Custom Workflow Mode:**
+```json
+{
+  "input": {
+    "request_id": "uuid-string",    // optional - UUID generated if not provided
+    "workflow_json": {
+      // Complete ComfyUI workflow JSON
+    },
+    "s3": { ... },       // optional
+    "webhook": { ... }   // optional
+  }
+}
+```
+
+## Request Fields
+
+### Required Fields
+
+- **`input`**: Contains the main workflow data
+- **`input.request_id`**: Unique identifier for the request
+
+### Workflow Mode (Choose One)
+
+You must provide either `modifier` OR `workflow_json`, but not both:
+
+#### Option 1: Modifier Mode
+- **`input.modifier`**: Name of the predefined workflow modifier (e.g., "Text2Image")
+- **`input.modifications`**: Parameters to pass to the modifier
+
+#### Option 2: Custom Workflow Mode  
+- **`input.workflow_json`**: Complete ComfyUI workflow JSON
+
+### Optional Fields
+
+- **`input.s3`**: S3 configuration for file storage
+- **`input.webhook`**: Webhook configuration for notifications
+
+These configurations can be provided in the request JSON or via environment variables. Request-level configuration takes precedence over environment variables.
+
+#### S3 Configuration
+
+**Via Request JSON:**
+```json
+"s3": {
+  "access_key_id": "your-s3-access-key",
+  "secret_access_key": "your-s3-secret-access-key", 
+  "endpoint_url": "https://my-endpoint.backblaze.com",
+  "bucket_name": "your-bucket",
+  "region": "us-east-1"
+}
+```
+
+**Via Environment Variables:**
+```bash
+S3_ACCESS_KEY_ID=your-key
+S3_SECRET_ACCESS_KEY=your-secret
+S3_BUCKET_NAME=your-bucket
+S3_ENDPOINT_URL=https://s3.amazonaws.com
+S3_REGION=us-east-1
+```
+
+#### Webhook Configuration
+
+**Via Request JSON:**
+```json
+"webhook": {
+  "url": "your-webhook-url",
+  "extra_params": {
+    "custom_field": "value"
+  }
+}
+```
+
+**Via Environment Variables:**
+```bash
+WEBHOOK_URL=https://your-webhook.com  # Default webhook URL
+WEBHOOK_TIMEOUT=30                   # Webhook timeout in seconds
+```
+
+## Examples
+
+### Basic Text-to-Image (Modifier Mode)
+
+```json
+{
+  "input": {
+    "modifier": "Text2Image",
+    "modifications": {
+      "prompt": "a cat sitting on a windowsill",
+      "width": 512,
+      "height": 512,
+      "steps": 20,
+      "seed": 42
+    }
+  }
+}
+```
+
+### Custom Workflow Mode
+
+```json
+{
+  "input": {
+    "request_id": "67890",    // optional - using custom ID for tracking
+    "workflow_json": {
+      "3": {
+        "inputs": {
+          "seed": 42,
+          "steps": 20,
+          "cfg": 8,
+          "sampler_name": "euler",
+          "scheduler": "normal",
+          "denoise": 1,
+          "model": ["4", 0],
+          "positive": ["6", 0],
+          "negative": ["7", 0],
+          "latent_image": ["5", 0]
+        },
+        "class_type": "KSampler"
+      }
+    }
+  }
+}
+```
+
+## Client Libraries
+
+See the test client examples for implementation details on how to integrate with the ComfyUI worker.
+
+---
+
+See Vast's serverless documentation for more details on how to use ComfyUI with autoscaler.

workers/comfyui-json/client.py

@@ -0,0 +1,155 @@
+import logging
+import uuid
+import random
+from urllib.parse import urljoin
+import json
+
+import requests
+
+from lib.test_utils import print_truncate_res
+from utils.endpoint_util import Endpoint
+from utils.ssl import get_cert_file_path
+from .data_types import count_workload
+
+logging.basicConfig(
+    level=logging.DEBUG,
+    format="%(asctime)s[%(levelname)-5s] %(message)s",
+    datefmt="%Y-%m-%d %H:%M:%S",
+)
+log = logging.getLogger(__file__)
+
+
+def call_text2image_workflow(
+    endpoint_group_name: str, api_key: str, server_url: str
+) -> None:
+    """Simple Text2Image using the new modifier-based approach"""
+    
+    def make_request(url: str, payload: dict, timeout: int = None, verify=True, context: str = "request"):
+        """Helper function for making requests with consistent error handling"""
+        try:
+            response = requests.post(
+                url,
+                json=payload,
+                timeout=timeout,
+                verify=verify
+            )
+            
+            response.raise_for_status()
+            return response.json()
+            
+        except requests.exceptions.HTTPError as http_err:
+            log.error(f"HTTP error occurred during {context}: {http_err}")
+            log.error(f"Status Code: {response.status_code}")
+            log.error("Response content:", response.text)
+            return None
+        except requests.exceptions.Timeout:
+            log.error(f"Timeout occurred during {context}: {url}")
+            return None
+        except requests.exceptions.ConnectionError:
+            log.error(f"Connection error occurred during {context}: {url}")
+            return None
+        except json.JSONDecodeError as json_err:
+            log.error(f"Failed to decode JSON response during {context}: {json_err}")
+            if 'response' in locals():
+                print("Response content:", response.text)
+            return None
+        except Exception as err:
+            log.error(f"An unexpected error occurred during {context}: {err}")
+            if 'response' in locals():
+                log.error("Response content (if available):", response.text)
+            return None
+    
+    WORKER_ENDPOINT = "/generate/sync"
+
+    # This worker has concurrency = 1.  All workloads have cost value 1.0
+    COST = count_workload()
+    
+    # Route to get worker URL
+    route_payload = {
+        "endpoint": endpoint_group_name,
+        "api_key": api_key,
+        "cost": COST,
+    }
+    
+    # First request - get routing information
+    route_response = make_request(
+        url=urljoin(server_url, "/route/"),
+        payload=route_payload,
+        timeout=4,
+        context="route request"
+    )
+    
+    if route_response is None:
+        return None
+    
+    if "url" not in route_response or not route_response["url"]:
+        log.error("Error: No worker in 'Ready' state. Please wait while the serverless engine removes errored workers or finishes loading new workers.")
+        return None
+    
+    if "status" in route_response:
+        print(f"Autoscaler status: {route_response['status']}")
+        return None
+    
+    # Extract data from route response
+    url = route_response["url"]
+    auth_data = dict(
+        signature=route_response["signature"],
+        cost=route_response["cost"],
+        endpoint=route_response["endpoint"],
+        reqnum=route_response["reqnum"],
+        url=route_response["url"],
+    )
+    
+    # Build the payload for the worker request
+    worker_payload = {
+        "input": {
+            "request_id": str(uuid.uuid4()),
+            "modifier": "Text2Image",
+            "modifications": {
+                "prompt": "a beautiful landscape with mountains and lakes",
+                "width": 1024,
+                "height": 1024,
+                "steps": 20,
+                "seed": random.randint(0, 2**32 - 1)
+            },
+            "workflow_json": {}  # Empty since using modifier approach
+        }
+    }
+    
+    req_data = dict(payload=worker_payload, auth_data=auth_data)
+    worker_url = urljoin(url, WORKER_ENDPOINT)
+    print(f"url: {worker_url}")
+    
+    # Second request - call the worker endpoint
+    worker_response = make_request(
+        url=worker_url,
+        payload=req_data,
+        verify=get_cert_file_path(),
+        context="worker request"
+    )
+    
+    return worker_response
+
+
+if __name__ == "__main__":
+    from lib.test_utils import test_args
+
+    args = test_args.parse_args()
+    endpoint_api_key = Endpoint.get_endpoint_api_key(
+        endpoint_name=args.endpoint_group_name,
+        account_api_key=args.api_key,
+        instance=args.instance,
+    )
+
+    if endpoint_api_key:
+        result = call_text2image_workflow(
+            api_key=endpoint_api_key,
+            endpoint_group_name=args.endpoint_group_name,
+            server_url=args.server_url,
+        )
+        if result is None:
+            log.error("Text2Image workflow failed")
+        else:   
+            print(result)
+    else:
+        log.error(f"Failed to get API key for endpoint {args.endpoint_group_name}")

workers/comfyui-json/data_types.py

@@ -0,0 +1,60 @@
+import os
+import sys
+import random
+import dataclasses
+from typing import Dict, Any
+from functools import cache
+from math import ceil
+
+from lib.data_types import ApiPayload, JsonDataException
+
+
+with open("workers/comfyui/misc/test_prompts.txt", "r") as f:
+    test_prompts = f.readlines()
+
+def count_workload() -> float:
+    # Always 100.0 where there is a single instance of ComfyUI handling requests
+    # Results will indicate % or a job completed per second.  Avoids sub 0.1 sec performance indication
+    return 100.0
+
+@dataclasses.dataclass
+class ComfyWorkflowData(ApiPayload):
+    input: dict
+
+    @classmethod
+    def for_test(cls):
+        """
+        Use the variables available to simulate workflows of the required running time
+        Example: SD1.5, simple image gen 10000 steps, 512px x 512px will run for approximately 9 minutes @ ~18 it/s (RTX 4090)
+        """
+        test_prompt = random.choice(test_prompts).rstrip()
+        return cls(
+            input={
+                "request_id": f"test-{random.randint(1000, 99999)}",
+                "modifier": "Text2Image",
+                "modifications": {
+                    "prompt": test_prompt,
+                    "width": os.getenv('BENCHMARK_TEST_WIDTH', 512),
+                    "height": os.getenv('BENCHMARK_TEST_HEIGHT', 512),
+                    "steps": os.getenv('BENCHMARK_TEST_STEPS', 20),
+                    "seed": random.randint(0, sys.maxsize),
+                }
+            }
+        )
+
+    def generate_payload_json(self) -> Dict[str, Any]:
+        # input is already a dict, just return it wrapped in the expected structure
+        return {"input": self.input}
+
+    def count_workload(self) -> float:
+        return count_workload()
+
+    @classmethod
+    def from_json_msg(cls, json_msg: Dict[str, Any]) -> "ComfyWorkflowData":
+        # Extract required fields
+        if "input" not in json_msg:
+            raise JsonDataException({"input": "missing parameter"})
+        
+        return cls(
+            input=json_msg["input"]
+        )

workers/comfyui-json/misc/test_prompts.txt

@@ -0,0 +1,34 @@
+cartoon character of a person with a hoodie , in style of cytus and deemo, ork, gold chains, realistic anime cat, dripping black goo, lineage revolution style, thug life, cute anthropomorphic bunny, balrog, arknights, aliased, very buff, black and red and yellow paint, painting illustration collage style, character composition in vector with white background
+stardew valley, fine details
+2D Vector Illustration of a child with soccer ball Art for Sublimation, Design Art, Chrome Art, Painting and Stunning Artwork, Highly Detailed Digital Painting, Airbrush Art, Highly Detailed Digital Artwork, Dramatic Artwork, stained antique yellow copper paint, digital airbrush art, detailed by Mark Brooks, Chicano airbrush art, Swagger! snake Culture
+realistic futuristic city-downtown with short buildings, sunset
+seascape by Ray Collins and artgerm, front view of a perfect wave, sunny background, ultra detailed water
+inspired by realflow-cinema4d editor features, create image of a transparent luxury cup with ice fruits and mint, connected with white, yellow and pink cream, Slow - High Speed MO Photography, YouTube Video Screenshot, Abstract Clay, Transparent Cup , molecular gastronomy, wheel, 3D fluid,Simulation rendering, still video, 4k polymer clay futras photography, very surreal, Houdini Fluid Simulation, hyperrealistic CGI and FLUIDS & MULTIPHYSICS SIMULATION effect, with Somali Stain Lurex, Metallic Jacquard, Gold Thread, Mulberry Silk, Toub Saree, Warm background, a fantastic image worthy of an award.
+biker with backpack on his back riding a motorcycle, Style by Ade Santora, Oilpunk, Cover photo, craig mullins style, on the cover of a magazine, Outdoor Magazine, inspired by Alex Petruk APe, image of a male biker, Cover of an award-winning magazine, the man has a backpack, photo for magazine, with a backpack, magazine cover
+generate a collage-style illustration inspired by the Procreate raster graphic editor, photographic illustration with the theme, 2D vector, art for textile sublimation, containing surrealistic cartoon cat wearing a baseball cap and jeans standing in front of a poster, inspired by Sadao Watanabe, Doraemon, Japanese cartoon style, Eichiro Oda, Iconic high detail character, Director: Nakahara Nantenbō, Kastuhiro Otomo, image detailed, by Miyamoto, Hidetaka Miyazaki, Katsuhiro illustration, 8k, masterpiece, Minimize noise and grain in photo quality without lose quality and increase brightness and lighting,Symmetry and Alignment, Avoid asymmetrical shapes and out-of-focus points. Focus and Sharpness: Make sure the image is focused and sharp and encourages the viewer to see it as a work of art printed on fabric.
+fantasy medieval village world inside a glass sphere , high detail, fantasy, realistic, light effect, hyper detail, volumetric lighting, cinematic, macro, depth of field, blur, red light and clouds from the back, highly detailed epic cinematic concept art cg render made in maya, blender and photoshop, octane render, excellent composition, dynamic dramatic cinematic lighting, aesthetic, very inspirational, world inside a glass sphere by james gurney by artgerm with james jean, joe fenton and tristan eaton by ross tran, fine details
+Iron Man, (Arnold Tsang, Toru Nakayama), Masterpiece, Studio Quality, 6k , toa, toaair, 1boy, glowing, axe, mecha, science_fiction, solo, weapon, jungle , green_background, nature, outdoors, solo, tree, weapon, mask, dynamic lighting, detailed shading, digital texture painting
+(Pope Francis) wearing leather jacket is a DJ in a nightclub, mixing live on stage, giant mixing table, a masterpiece
+Pope Francis wearing biker (leather jacket), a masterpiece
+Luke Skywalker ordering a burger and fries from the Death Star canteen.
+I want to generate a group avatar for a Feishu group chat. The role of this group is daily software technical communication. Now the subject technology stacks that members of this group discuss daily include: algorithms, data structures, optimization, functional programming, and the programming languages often discussed are: TypeScript, Java, python, etc. I hope this avatar has a simple aesthetic, this avatar is a single person avatar
+portrait Anime black girl cute-fine-face, pretty face, realistic shaded Perfect face, fine details. Anime. realistic shaded lighting by Ilya Kuvshinov Giuseppe Dangelico Pino and Michael Garmash and Rob Rey, IAMAG premiere, WLOP matte print, cute freckles, masterpiece
+young Disney socialite wearing a beige miniskirt, dark brown turtleneck sweater, small neckless, cute-fine-face, anime. illustration, realistic shaded perfect face, brown hair, grey eyes, fine details, realistic shaded lighting by ilya kuvshinov giuseppe dangelico pino and michael garmash and rob rey, iamag premiere, wlop matte print, a masterpiece
+Cute small cat sitting in a movie theater eating chicken wiggs watching a movie ,unreal engine, cozy indoor lighting, artstation, detailed, digital painting,cinematic,character design by mark ryden and pixar and hayao miyazaki, unreal 5, daz, hyperrealistic, octane render
+Cute small dog sitting in a movie theater eating popcorn watching a movie ,unreal engine, cozy indoor lighting, artstation, detailed, digital painting,cinematic,character design by mark ryden and pixar and hayao miyazaki, unreal 5, daz, hyperrealistic, octane render
+fox bracelet made of buckskin with fox features, rich details, fine carvings, studio lighting
+crane buckskin bracelet with crane features, rich details, fine carvings, studio lighting
+london luxurious interior living-room, light walls
+Parisian luxurious interior penthouse bedroom, dark walls, wooden panels
+cute girl, crop-top, blond hair, black glasses, stretching, with background by greg rutkowski makoto shinkai kyoto animation key art feminine mid shot
+houses in front, houses background, straight houses, digital art, smooth, sharp focus, gravity falls style, doraemon style, shinchan style, anime style
+Simplified technical drawing, Leonardo da Vinci, Mechanical Dinosaur Skeleton, Minimalistic annotations, Hand-drawn illustrations, Basic design and engineering, Wonder and curiosity
+High quality 8K painting impressionist style of a Japanese modern city street with a girl on the foreground wearing a traditional wedding dress with a fox mask, staring at the sky, daylight
+a landscape from the Moon with the Earth setting on the horizon, realistic, detailed
+Isometric Atlantis city,great architecture with columns, great details, ornaments,seaweed, blue ambiance, 3D cartoon style, soft light, 45° view
+A hyper realistic avatar of a guy riding on a black honda cbr 650r in leather suit,high detail, high quality,8K,photo realism
+the street of amedieval fantasy town, at dawn, dark, highly detailed
+overwhelmingly beautiful eagle framed with vector flowers, long shiny wavy flowing hair, polished, ultra detailed vector floral illustration mixed with hyper realism, muted pastel colors, vector floral details in background, muted colors, hyper detailed ultra intricate overwhelming realism in detailed complex scene with magical fantasy atmosphere, no signature, no watermark
+a highly detailed matte painting of a man on a hill watching a rocket launch in the distance by studio ghibli, makoto shinkai, by artgerm, by wlop, by greg rutkowski, volumetric lighting, octane render, 4 k resolution, trending on artstation, masterpiece | hyperrealism| highly detailed| insanely detailed| intricate| cinematic lighting| depth of field
+electronik robot and ofice ,unreal engine, cozy indoor lighting, artstation, detailed, digital painting,cinematic,character design by mark ryden and pixar and hayao miyazaki, unreal 5, daz, hyperrealistic, octane render
+exquisitely intricately detailed illustration, of a small world with a lake and a rainbow, inside a closed glass jar.

workers/comfyui-json/server.py

@@ -0,0 +1,116 @@
+import os
+import logging
+import dataclasses
+import base64
+from typing import Optional, Union, Type
+
+from aiohttp import web, ClientResponse
+
+from lib.backend import Backend, LogAction
+from lib.data_types import EndpointHandler
+from lib.server import start_server
+from .data_types import ComfyWorkflowData
+
+
+MODEL_SERVER_URL = os.getenv("MODEL_SERVER_URL", "http://127.0.0.1:18288")
+
+# This is the last log line that gets emitted once comfyui+extensions have been fully loaded
+MODEL_SERVER_START_LOG_MSG = "To see the GUI go to: "
+MODEL_SERVER_ERROR_LOG_MSGS = [
+    "MetadataIncompleteBuffer",  # This error is emitted when the downloaded model is corrupted
+    "Value not in list: ",  # This error is emitted when the model file is not there at all
+]
+
+
+logging.basicConfig(
+    level=logging.DEBUG,
+    format="%(asctime)s[%(levelname)-5s] %(message)s",
+    datefmt="%Y-%m-%d %H:%M:%S",
+)
+log = logging.getLogger(__file__)
+
+
+async def generate_client_response(
+        client_request: web.Request, model_response: ClientResponse
+    ) -> Union[web.Response, web.StreamResponse]:
+        # Check if the response is actually streaming based on response headers/content-type
+        is_streaming_response = (
+            model_response.content_type == "text/event-stream"
+            or model_response.content_type == "application/x-ndjson"
+            or model_response.headers.get("Transfer-Encoding") == "chunked"
+            or "stream" in model_response.content_type.lower()
+        )
+
+        if is_streaming_response:
+            log.debug("Detected streaming response...")
+            res = web.StreamResponse()
+            res.content_type = model_response.content_type
+            await res.prepare(client_request)
+            async for chunk in model_response.content:
+                await res.write(chunk)
+            await res.write_eof()
+            log.debug("Done streaming response")
+            return res
+        else:
+            log.debug("Detected non-streaming response...")
+            content = await model_response.read()
+            return web.Response(
+                body=content,
+                status=model_response.status,
+                content_type=model_response.content_type
+            )
+            
+
+@dataclasses.dataclass
+class ComfyWorkflowHandler(EndpointHandler[ComfyWorkflowData]):
+
+    @property
+    def endpoint(self) -> str:
+        return "/generate/sync"
+
+    @property
+    def healthcheck_endpoint(self) -> Optional[str]:
+        return f"{MODEL_SERVER_URL}/health"
+
+    @classmethod
+    def payload_cls(cls) -> Type[ComfyWorkflowData]:
+        return ComfyWorkflowData
+
+    def make_benchmark_payload(self) -> ComfyWorkflowData:
+        return ComfyWorkflowData.for_test()
+
+    async def generate_client_response(
+        self, client_request: web.Request, model_response: ClientResponse
+    ) -> Union[web.Response, web.StreamResponse]:
+        return await generate_client_response(client_request, model_response)
+
+
+backend = Backend(
+    model_server_url=MODEL_SERVER_URL,
+    model_log_file=os.environ["MODEL_LOG"],
+    allow_parallel_requests=False,
+    benchmark_handler=ComfyWorkflowHandler(
+        benchmark_runs=3, benchmark_words=100
+    ),
+    log_actions=[
+        (LogAction.ModelLoaded, MODEL_SERVER_START_LOG_MSG),
+        (LogAction.Info, "Downloading:"),
+        *[
+            (LogAction.ModelError, error_msg)
+            for error_msg in MODEL_SERVER_ERROR_LOG_MSGS
+        ],
+    ],
+)
+
+
+async def handle_ping(_):
+    return web.Response(body="pong")
+
+
+routes = [
+    web.post("/generate/sync", backend.create_handler(ComfyWorkflowHandler())),
+    web.get("/ping", handle_ping),
+]
+
+if __name__ == "__main__":
+    start_server(backend, routes)

workers/comfyui-json/test_load.py

@@ -0,0 +1,8 @@
+from lib.test_utils import test_load_cmd, test_args
+from .data_types import ComfyWorkflowData
+
+WORKER_ENDPOINT = "/generate/sync"
+
+
+if __name__ == "__main__":
+    test_load_cmd(ComfyWorkflowData, WORKER_ENDPOINT, arg_parser=test_args)

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/comfyui/server.py

@@ -13,7 +13,7 @@ from lib.server import start_server
 from .data_types import DefaultComfyWorkflowData, CustomComfyWorkflowData
 
 
-MODEL_SERVER_URL = "http://0.0.0.0:38188"
+MODEL_SERVER_URL = "http://127.0.0.1:18288" # API Wrapper Service
 
 # This is the last log line that gets emitted once comfyui+extensions have been fully loaded
 MODEL_SERVER_START_LOG_MSG = "To see the GUI go to: http://127.0.0.1:18188"

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(
@@ -28,24 +29,16 @@ class APIClient:
     DEFAULT_TIMEOUT = 4
 
     def __init__(
-        self, endpoint_group_name: str, api_key: str, server_url: str, instance: str
+        self,
+        endpoint_group_name: str,
+        api_key: str,
+        server_url: str,
+        endpoint_api_key: str,
     ):
         self.endpoint_group_name = endpoint_group_name
         self.api_key = api_key
         self.server_url = server_url
-        self.instance = instance
-        self.endpoint_api_key = self._get_endpoint_api_key()
-
-    def _get_endpoint_api_key(self) -> Optional[str]:
-        """Get the endpoint API key"""
-        endpoint_api_key = Endpoint.get_endpoint_api_key(
-            endpoint_name=self.endpoint_group_name,
-            account_api_key=self.api_key,
-            instance=self.instance,
-        )
-        if not endpoint_api_key:
-            log.error(f"Failed to get API key for endpoint {self.endpoint_group_name}")
-        return endpoint_api_key
+        self.endpoint_api_key = endpoint_api_key
 
     def _get_worker_url(self, cost: int = DEFAULT_COST) -> Dict[str, Any]:
         """Get worker URL and auth data from routing service"""
@@ -98,9 +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}")
 
@@ -554,12 +551,24 @@ def main():
         sys.exit(1)
 
     try:
+        endpoint_api_key = Endpoint.get_endpoint_api_key(
+            endpoint_name=args.endpoint_group_name,
+            account_api_key=args.api_key,
+            instance=args.instance,
+        )
+
+        if not endpoint_api_key:
+            log.error(
+                f"Could not retrieve API key for endpoint '{args.endpoint_group_name}'. Exiting."
+            )
+            sys.exit(1)
+
         # Create the core API client
         client = APIClient(
             endpoint_group_name=args.endpoint_group_name,
             api_key=args.api_key,
-            server_url=args.server_url,
-            instance=args.instance,
+            server_url=Endpoint.get_autoscaler_server_url(args.instance),
+            endpoint_api_key=endpoint_api_key,
         )
 
         # Create tool manager and demo (passing the model parameter)

workers/openai/data_types/server.py

@@ -124,7 +124,12 @@ class CompletionsData(GenericData):
         if not model:
             raise ValueError("MODEL_NAME environment variable not set")
 
-        test_input = {"model": model, "prompt": prompt, "temperature": 0.7}
+        test_input = {
+            "model": model,
+            "prompt": prompt,
+            "temperature": 0.7,
+            "max_tokens": 500,
+        }
         return cls(input=test_input)
 
 
@@ -158,6 +163,7 @@ class ChatCompletionsData(GenericData):
             "model": model,
             "messages": [{"role": "user", "content": prompt}],
             "temperature": 0.7,
+            "max_tokens": 500,
         }
         return cls(input=test_input)
 

workers/openai/test_load.py

@@ -1,8 +1,394 @@
-from lib.test_utils import test_load_cmd, test_args
+from lib.test_utils import test_args
+from utils.endpoint_util import Endpoint
+from utils.ssl import get_cert_file_path
+from lib.data_types import AuthData
 from .data_types.server import CompletionsData
-import os
 
-WORKER_ENDPOINT = "/v1/completions"
+import os
+import time
+import threading
+import requests
+from dataclasses import dataclass
+from collections import Counter
+from urllib.parse import urljoin, urlparse
+import re
+
+# Headless plotting
+import matplotlib
+matplotlib.use("Agg")
+import logging
+logging.getLogger("matplotlib.font_manager").setLevel(logging.WARNING)
+import matplotlib.pyplot as plt
+import numpy as np
+from concurrent.futures import ThreadPoolExecutor, wait, FIRST_COMPLETED
+from requests.adapters import HTTPAdapter
+
+def get_incremented_path(path: str) -> str:
+    base, ext = os.path.splitext(path)
+    if not os.path.exists(path):
+        return path
+    i = 1
+    while os.path.exists(f"{base}-{i}{ext}"):
+        i += 1
+    return f"{base}-{i}{ext}"
+
+WORKER_ENDPOINT = "/v1/completions" # This will return the full text output at once. Latency metrics reflect that (ie not measuring TTFT)
+
+@dataclass
+class ReqResult:
+    worker_url: str
+    route_ms: float
+    worker_ms: float
+    total_ms: float
+    ok: bool
+    error: str = ""
+    status_code: int = 0
+    t_start: float = 0.0
+    t_end: float = 0.0
+    workload: float = 0.0
+
+def do_one(endpoint_name: str,
+           endpoint_id: int,
+           endpoint_api_key: str,
+           server_url: str,
+           worker_endpoint: str,
+           payload,
+           results_list,
+           t0,
+           status_samples,
+           route_session,
+           worker_session):
+    try:
+        workload = payload.count_workload()
+        route_payload = {"endpoint": endpoint_name, "api_key": endpoint_api_key, "cost": workload}
+        headers = {"Authorization": f"Bearer {endpoint_api_key}"}
+        start = time.time()
+        r0 = route_session.post(urljoin(server_url, "/route/"), json=route_payload, headers=headers, timeout=4)
+        t_after_route = time.time()
+        if r0.status_code != 200:
+            results_list.append(ReqResult(worker_url="", 
+                                            route_ms=(t_after_route - start) * 1000.0, 
+                                            worker_ms=0.0, 
+                                            total_ms=(t_after_route - start) * 1000.0, 
+                                            ok=False, 
+                                            error=f"route error {r0.reason} {r0.text}",
+                                            status_code=r0.status_code,
+                                            t_start=start - t0, 
+                                            t_end=t_after_route - t0, 
+                                            workload=workload))
+            return
+        msg = r0.json()
+
+        # 1) Check if we got a worker back from route
+        worker_url = msg.get("url", "")
+        if not worker_url:
+            m = re.search(r"total workers:\s*(\d+).*loading workers:\s*(\d+).*standby workers:\s*(\d+).*error workers:\s*(\d+)", status, re.I | re.S)
+            if m:
+                tot, loading, standby, err = map(int, m.groups())
+                idle = max(tot - loading - standby - err, 0)
+                status_samples.append((time.time() - t0, idle))
+
+        # 2) If we got a worker, send the request
+        if worker_url:
+            req = dict(payload=payload.__dict__, auth_data=AuthData.from_json_msg(msg).__dict__)
+            t_before_worker = time.time()
+            r1 = worker_session.post(
+                urljoin(worker_url, worker_endpoint),
+                json=req,
+                verify=get_cert_file_path(),
+                timeout=(4, 120),
+            )
+            t_after_worker = time.time()
+            if r1.status_code != 200:
+                results_list.append(ReqResult(worker_url=worker_url, 
+                                                route_ms=(t_after_route - start) * 1000.0, 
+                                                worker_ms=(t_after_worker - t_before_worker) * 1000.0,
+                                                total_ms=(t_after_worker - start) * 1000.0, 
+                                                ok=False,
+                                                error=f"worker inference error {r1.reason} {r1.text}",
+                                                status_code=r1.status_code,
+                                                t_start=start - t0, 
+                                                t_end=t_after_worker - t0, 
+                                                workload=workload))
+                return
+            # Success case
+            results_list.append(ReqResult(worker_url=worker_url, 
+                                            route_ms=(t_after_route - start) * 1000.0, 
+                                            worker_ms=(t_after_worker - t_before_worker) * 1000.0, 
+                                            total_ms=(t_after_worker - start) * 1000.0,
+                                            ok=True,
+                                            error="",
+                                            status_code=200,
+                                            t_start=start - t0, 
+                                            t_end=t_after_worker - t0, 
+                                            workload=workload))
+
+        # 3) If so, sample via /get_endpoint_workers/ for eligible (idle) worker tracking
+        if worker_url:
+            try:
+                r_status = route_session.post(
+                    urljoin(server_url, "/get_endpoint_workers/"),
+                    json={"id": endpoint_id},
+                    headers={"Authorization": f"Bearer {endpoint_api_key}"},
+                    timeout=3,
+                )
+                if r_status.status_code == 200:
+                    workers = r_status.json()
+                    idle = 0
+                    for w in workers:
+                        st = str(w.get("status", "")).lower()
+                        if (st in ("idle")):
+                            idle += 1
+                    status_samples.append((time.time() - t0, idle))
+            except Exception:
+                pass
+    except Exception as e:
+        t = time.time()
+        results_list.append(ReqResult(worker_url="", 
+                                        route_ms=0.0, 
+                                        worker_ms=0.0, 
+                                        total_ms=0.0, 
+                                        ok=False, 
+                                        error=f"unknown error {e}",
+                                        status_code=0,
+                                        t_start=t - t0, 
+                                        t_end=t - t0, 
+                                        workload=0.0))
+
+def run_load_with_metrics(num_requests: int,
+                          requests_per_second: float,
+                          endpoint_group_name: str,
+                          account_api_key: str,
+                          server_url: str,
+                          worker_endpoint: str,
+                          instance: str,
+                          out_path: str):
+
+    ep_info = Endpoint.get_endpoint_info(endpoint_name=endpoint_group_name,
+                                         account_api_key=account_api_key,
+                                         instance=instance)
+    if not ep_info or not ep_info.get("api_key") or not ep_info.get("id"):
+        print(f"Endpoint {endpoint_group_name} not found for API key")
+        return
+    endpoint_id = int(ep_info["id"])
+    endpoint_api_key = ep_info["api_key"]
+
+    t0 = time.time()
+    results = []
+    status_samples = []
+    max_concurrency = int(os.environ.get("MAX_CONCURRENCY", "8192"))
+    submit_queue_factor = 2  # cap queued tasks to reduce memory
+
+    # Shared HTTP sessions with connection pooling (persistent connections)
+    def make_session(pool_connections: int, pool_maxsize: int) -> requests.Session:
+        sess = requests.Session()
+        adapter = HTTPAdapter(pool_connections=pool_connections, pool_maxsize=pool_maxsize, max_retries=0)
+        sess.mount("https://", adapter)
+        sess.mount("http://", adapter)
+        return sess
+
+    # Router: mostly single host, small connection pool is sufficient
+    route_session = make_session(pool_connections=1, pool_maxsize=max_concurrency)
+    # Workers: many hosts; allow many pools and per-host concurrency up to max_concurrency
+    worker_session = make_session(pool_connections=64, pool_maxsize=max_concurrency // 8)
+
+    # Fire requests using a thread pool, scheduling at requested RPS
+    inflight = set()
+    with ThreadPoolExecutor(max_workers=max_concurrency) as executor:
+        for i in range(num_requests):
+            # Pace submissions to RPS
+            target_time = t0 + i / max(requests_per_second, 1e-9)
+            sleep_s = target_time - time.time()
+            if sleep_s > 0:
+                time.sleep(min(sleep_s, 0.5))  # sleep in chunks to stay responsive
+
+            payload = CompletionsData.for_test()
+            fut = executor.submit(
+                do_one,
+                endpoint_group_name,
+                endpoint_id,
+                endpoint_api_key,
+                server_url,
+                worker_endpoint,
+                payload,
+                results,
+                t0,
+                status_samples,
+                route_session,
+                worker_session,
+            )
+            inflight.add(fut)
+            # Prevent unbounded queue growth
+            if len(inflight) >= max_concurrency * submit_queue_factor:
+                done, not_done = wait(inflight, return_when=FIRST_COMPLETED)
+                inflight = not_done
+        # Wait for all outstanding tasks
+        if inflight:
+            wait(inflight)
+    # Close sessions
+    try:
+        route_session.close()
+    finally:
+        worker_session.close()
+
+    # Aggregate results
+    oks = [r for r in results if r.ok]
+    errs = [r for r in results if not r.ok]
+    total_reqs = len(results)
+    succ = len(oks)
+
+    total_ms = np.array([r.total_ms for r in oks]) if succ else np.array([])
+    worker_ms = np.array([r.worker_ms for r in oks]) if succ else np.array([])
+    route_ms = np.array([r.route_ms for r in oks]) if succ else np.array([])
+
+    avg_total = float(np.mean(total_ms)) if succ else 0.0
+    avg_worker = float(np.mean(worker_ms)) if succ else 0.0
+    avg_route = float(np.mean(route_ms)) if succ else 0.0
+    p50_total, p95_total = (float(np.percentile(total_ms, 50)), float(np.percentile(total_ms, 95))) if succ else (0.0, 0.0)
+
+    # Distribution over workers (by host:port)
+    hosts = [urlparse(r.worker_url).netloc for r in oks if r.worker_url]
+    dist = Counter(hosts)
+
+    # Idle over time (mode per second)
+    idle_ts, idle_vals = [], []
+    if status_samples:
+        buckets = {}
+        for ts, idle in status_samples:
+            k = int(ts)
+            buckets.setdefault(k, []).append(idle)
+        keys = sorted(buckets.keys())
+        idle_ts = keys
+        # Use the most frequent sampled value per second (mode) to keep integer counts
+        idle_vals = []
+        for k in keys:
+            vals_k = [int(v) for v in buckets[k]]
+            if vals_k:
+                cnt = Counter(vals_k)
+                idle_vals.append(cnt.most_common(1)[0][0])
+            else:
+                idle_vals.append(0)
+
+    print(f"\nResults: total={total_reqs} success={succ} errors={len(errs)}")
+    print(f"Avg latency (ms): {avg_total:.1f}  p50: {p50_total:.1f}  p95: {p95_total:.1f}")
+    print(f"Avg route latency (ms): {avg_route:.1f}  Avg worker latency (ms): {avg_worker:.1f}")
+    if errs:
+        print("Sample errors:")
+        for e in errs[:5]:
+            print(f"  {e.status_code} {e.error}")
+
+    # Plot: 2x3 grid
+    fig, axes = plt.subplots(2, 3, figsize=(15, 8))
+    fig.suptitle(f"Load test: {endpoint_group_name}  n={total_reqs}, rps={requests_per_second}, success={succ}")
+
+    # Dist per worker
+    ax0 = axes[0, 0]
+    if dist:
+        items = sorted(dist.items(), key=lambda kv: kv[1], reverse=True)
+        labels, counts = zip(*items)
+        ax0.bar(range(len(labels)), counts)
+        ax0.set_xticks(range(len(labels)))
+        ax0.set_xticklabels(labels, rotation=45, ha="right", fontsize=8)
+    ax0.set_title("Request distribution over workers")
+    ax0.set_ylabel("count")
+
+    # Latency histogram (total)
+    ax1 = axes[0, 1]
+    if succ:
+        ax1.hist(total_ms, bins=30)
+    ax1.set_title("Total latency (ms)")
+    ax1.set_xlabel("ms")
+    ax1.set_ylabel("freq")
+
+    # Eligible workers over time
+    ax_idle = axes[0, 2]
+    if idle_ts:
+        ax_idle.plot(idle_ts, idle_vals, "-o", ms=3)
+    ax_idle.set_title("Eligible workers over time")
+    ax_idle.set_xlabel("time (s)")
+    ax_idle.set_ylabel("eligible count")
+
+    # Throughput over time (completions/sec)
+    ax_idle = axes[1, 0]
+    ax_idle.clear()
+    if succ:
+        per_sec = {}
+        for r in oks:
+            s = int(r.t_end)
+            per_sec[s] = per_sec.get(s, 0) + 1
+        ts = sorted(per_sec.keys())
+        vals = [per_sec[t] for t in ts]
+        ax_idle.plot(ts, vals, "-o", ms=3)
+    ax_idle.set_title("Completions per second")
+    ax_idle.set_xlabel("time (s)")
+    ax_idle.set_ylabel("completions / sec")
+
+    # Summary text
+    ax3 = axes[1, 1]
+    ax3.axis("off")
+    text = (
+        f"Total requests: {total_reqs}\n"
+        f"Success: {succ}  Errors: {len(errs)}\n"
+        f"Avg total latency: {avg_total:.1f} ms\n"
+        f"p50: {p50_total:.1f} ms  p95: {p95_total:.1f} ms\n"
+        f"Avg route latency: {avg_route:.1f} ms\n"  
+        f"Avg worker latency: {avg_worker:.1f} ms\n"
+        f"300 errors: {len([r for r in errs if r.status_code >= 300 and r.status_code < 400])}\n"
+        f"429 errors: {len([r for r in errs if r.status_code == 429])}\n"
+        f"500 errors: {len([r for r in errs if r.status_code >= 500])}\n"
+        f"Other errors: {len([r for r in errs if r.status_code not in [300, 429, 500]])}\n"
+    )
+    ax3.set_title("Summary")
+    ax3.text(0.02, 0.98, text, va="top", ha="left", fontsize=11, transform=ax3.transAxes)
+
+    # Error count over time
+    ax_errors = axes[1, 2]
+    all_end_times = [int(r.t_end) for r in results if r.t_end > 0]
+    if all_end_times:
+        min_second = min(all_end_times)
+        max_second = max(all_end_times)
+        # Count errors per second
+        errors_per_second = {}
+        for result in errs:
+            second = int(result.t_end)
+            errors_per_second[second] = errors_per_second.get(second, 0) + 1
+        # Create complete timeline including zeros
+        time_seconds = list(range(min_second, max_second + 1))
+        error_counts = [errors_per_second.get(sec, 0) for sec in time_seconds]
+        ax_errors.plot(time_seconds, error_counts, "-o", ms=3)
+    ax_errors.set_title("Errors per second")
+    ax_errors.set_xlabel("time (s)")
+    ax_errors.set_ylabel("errors / sec")
+
+    # Ensure unique output path and create directory if needed
+    final_out_path = get_incremented_path(out_path)
+    out_dir = os.path.dirname(final_out_path)
+    if out_dir:
+        os.makedirs(out_dir, exist_ok=True)
+
+    plt.tight_layout(rect=[0, 0, 1, 0.96])
+    plt.savefig(final_out_path, dpi=120)
+    print(f"Saved report to: {final_out_path}")
+
+    # Per-worker latency boxplot (top 12 by volume)
+    groups = {}
+    for r in oks:
+        host = urlparse(r.worker_url).netloc
+        groups.setdefault(host, []).append(r.total_ms)
+    items = sorted(groups.items(), key=lambda kv: len(kv[1]), reverse=True)[:12]
+    if items:
+        labels, data = zip(*items)
+        fig2, axb = plt.subplots(1, 1, figsize=(12, 5))
+        axb.boxplot(data, showfliers=False)
+        axb.set_xticklabels(labels, rotation=45, ha="right", fontsize=8)
+        axb.set_title("Per-worker latency (ms)")
+        axb.set_ylabel("ms")
+        plt.tight_layout()
+        extra_out = get_incremented_path(os.path.splitext(out_path)[0] + "-workers.png")
+        plt.savefig(extra_out, dpi=120)
+        fig2.tight_layout()
+        fig2.savefig(extra_out, dpi=120)
+        print(f"Saved worker latency plot to: {extra_out}")
 
 if __name__ == "__main__":
     # Check if MODEL_NAME environment variable is set
@@ -16,13 +402,32 @@ if __name__ == "__main__":
         help="Model to use for completions request (required if MODEL_NAME env var not set)",
     )
 
-    # Parse known args to get model early, before test_load_cmd adds its args
+    # Parse known args to get model early, before adding load args
     known_args, _ = test_args.parse_known_args()
-
-    # Set environment variable if model was provided
     if hasattr(known_args, "model") and known_args.model:
         os.environ["MODEL_NAME"] = known_args.model
         print(f"Set MODEL_NAME environment variable to: {known_args.model}")
 
-    # Now call test_load_cmd normally - it will add its own args and re-parse
-    test_load_cmd(CompletionsData, WORKER_ENDPOINT, arg_parser=test_args)
+    # Load test args
+    test_args.add_argument("-n", dest="num_requests", type=int, required=True, help="total number of requests")
+    test_args.add_argument("-rps", dest="requests_per_second", type=float, required=True, help="requests per second")
+    test_args.add_argument("--out", dest="out_path", type=str, default="load_test_report.png", help="path to save the report image")
+    args = test_args.parse_args()
+
+    server_url = {
+        "prod": "https://run.vast.ai",
+        "alpha": "https://run-alpha.vast.ai",
+        "candidate": "https://run-candidate.vast.ai",
+        "local": "http://localhost:8080"
+    }.get(args.instance, "http://localhost:8080")
+
+    run_load_with_metrics(
+        num_requests=args.num_requests,
+        requests_per_second=args.requests_per_second,
+        endpoint_group_name=args.endpoint_group_name,
+        account_api_key=args.api_key,
+        server_url=server_url,
+        worker_endpoint=WORKER_ENDPOINT,
+        instance=args.instance,
+        out_path=args.out_path,
+    )

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)