Suppress matplot debug logs

.gitignore

@@ -3,4 +3,3 @@
 __pycache__
 bin/
 lib64
-.venv

README.md

@@ -39,12 +39,11 @@ reporting these metrics to the autoscaler.
 
 If you are using a Vast.ai template that includes PyWorker integration (marked as autoscaler compatible), it should work out of the box. The template will typically start the appropriate PyWorker server automatically. Here's a few:
 
-*   **vLLM:** [Vast.ai Template](https://cloud.vast.ai?ref_id=62897&template_id=63ae93902bf3978bea033782592b784d)
-*   **TGI (Text Generation Inference):** [Vast.ai Template](https://cloud.vast.ai?ref_id=62897&template_id=6fa6bd5bdf5f0df63db80e40b086037d)
-*   **ComfyUI:** [Vast.ai Template](https://cloud.vast.ai?ref_id=62897&template_id=e6748878ba688e765e3e9fca29541938)
+*   **TGI (Text Generation Inference):** [Vast.ai Template](https://cloud.vast.ai?ref_id=140778&template_id=72d8dcb41ea3a58e06c741e2c725bc00)
+*   **ComfyUI:** [Vast.ai Template](https://cloud.vast.ai?ref_id=140778&template_id=ad72c8bf7cf695c3c9ddf0eaf6da0447)
 
 Currently available workers:
-*   `openai`: A simple example worker for a basic vLLM server.
+*   `hello_world`: A simple example worker for a basic LLM server.
 *   `comfyui`: A worker for the ComfyUI image generation backend.
 *   `tgi`: A worker for the Text Generation Inference backend.
 

lib/backend.py

@@ -12,7 +12,6 @@ from distutils.util import strtobool
 
 from anyio import open_file
 from aiohttp import web, ClientResponse, ClientSession, ClientConnectorError, ClientTimeout, TCPConnector
-import asyncio
 
 import requests
 from Crypto.Signature import pkcs1_15
@@ -26,12 +25,8 @@ from lib.data_types import (
     LogAction,
     ApiPayload_T,
     JsonDataException,
-    RequestMetrics,
-    BenchmarkResult
 )
 
-VERSION = "0.2.0"
-
 MSG_HISTORY_LEN = 100
 log = logging.getLogger(__file__)
 
@@ -58,25 +53,15 @@ 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(
         default_factory=lambda: bool(strtobool(os.environ.get("UNSECURED", "false"))),
     )
-    report_addr: str = dataclasses.field(
-        default_factory=lambda: os.environ.get("REPORT_ADDR", "https://run.vast.ai")
-    )
-    mtoken: str = dataclasses.field(
-        default_factory=lambda: os.environ.get("MASTER_TOKEN", "")
-    )
 
     def __post_init__(self):
         self.metrics = Metrics()
-        self.metrics._set_version(self.version)
-        self.metrics._set_mtoken(self.mtoken)
         self._total_pubkey_fetch_errors = 0
         self._pubkey = self._fetch_pubkey()
         self.__start_healthcheck: bool = False
@@ -111,19 +96,23 @@ class Backend:
 
     #######################################Private#######################################
     def _fetch_pubkey(self):
-        report_addr = self.report_addr.rstrip("/")
-        command = ["curl", "-X", "GET", f"{report_addr}/pubkey/"]
-        try:
-            result = subprocess.check_output(command, universal_newlines=True)
-            log.debug("public key:")
-            log.debug(result)
-            key = RSA.import_key(result)
-            if key is not None:
-                return key
-        except (ValueError , subprocess.CalledProcessError) as e:
-            log.debug(f"Error downloading key: {e}")
-        self.backend_errored("Failed to get autoscaler pubkey")
-       
+        command = ["curl", "-X", "GET", "https://run.vast.ai/pubkey/"]
+        result = subprocess.check_output(command, universal_newlines=True)
+        log.debug("public key:")
+        log.debug(result)
+        key = None
+        for _ in range(5):
+            try:
+                key = RSA.import_key(result)
+                break
+            except ValueError as e:
+                log.debug(f"Error downloading key: {e}")
+                time.sleep(15)
+        if key is None:
+            self._total_pubkey_fetch_errors += 1
+            if self._total_pubkey_fetch_errors >= MAX_PUBKEY_FETCH_ATTEMPTS:
+                self.backend_errored("Failed to get autoscaler pubkey")
+        return key
 
     async def __handle_request(
         self,
@@ -139,56 +128,55 @@ 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: {request_metrics.reqnum} was canceled")
-            self.metrics._request_canceled(request_metrics)
-            raise asyncio.CancelledError
+            log.debug(f"request with reqnum: {auth_data.reqnum} was canceled")
+            self.metrics._request_canceled(workload=workload)
+            return web.Response(status=500)
 
         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:
                 response = await self.__call_api(handler=handler, payload=payload)
                 status_code = response.status
                 log.debug(
                     " ".join(
                         [
-                            f"request with reqnum:{request_metrics.reqnum}",
+                            f"request with reqnum:{auth_data.reqnum}",
                             f"returned status code: {status_code},",
                         ]
                     )
                 )
                 res = await handler.generate_client_response(request, response)
-                self.metrics._request_success(request_metrics)
+                self.metrics._request_success(workload=workload)
                 return res
             except requests.exceptions.RequestException as e:
                 log.debug(f"[backend] Request error: {e}")
-                self.metrics._request_errored(request_metrics)
+                self.metrics._request_errored(workload=workload)
                 return web.Response(status=500)
+            finally:
+                self.metrics._request_end(
+                    workload=workload,
+                    reqnum=auth_data.reqnum,
+                )
+                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()),
@@ -196,27 +184,11 @@ class Backend:
                 ],
                 return_when=FIRST_COMPLETED,
             )
-            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) 
+            [task.cancel() for task in pending]
+            return done.pop().result()
         except Exception as e:
             log.debug(f"Exception in main handler loop {e}")
             return web.Response(status=500)
-        finally:
-            # 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):
@@ -257,7 +229,7 @@ class Backend:
 
     async def _start_tracking(self) -> None:
         await gather(
-            self.__read_logs(), self.metrics._send_metrics_loop(), self.__healthcheck(), self.metrics._send_delete_requests_loop()
+            self.__read_logs(), self.metrics._send_metrics_loop(), self.__healthcheck()
         )
 
     def backend_errored(self, msg: str) -> None:
@@ -289,7 +261,7 @@ class Backend:
         message = {
             key: value
             for (key, value) in (dataclasses.asdict(auth_data).items())
-            if key != "signature" and key != "__request_id"
+            if key != "signature"
         }
         if auth_data.reqnum < (self.reqnum - MSG_HISTORY_LEN):
             log.debug(
@@ -299,7 +271,7 @@ class Backend:
         elif message in self.msg_history:
             log.debug(f"message: {message} already in message history")
             return False
-        elif verify_signature(json.dumps(message, indent=4, sort_keys=True), auth_data.signature):
+        elif verify_signature(json.dumps(message, indent=4), auth_data.signature):
             self.reqnum = max(auth_data.reqnum, self.reqnum)
             self.msg_history.append(message)
             self.msg_history = self.msg_history[-MSG_HISTORY_LEN:]
@@ -318,10 +290,10 @@ class Backend:
                 with open(BENCHMARK_INDICATOR_FILE, "r") as f:
                     log.debug("already ran benchmark")
                     # trigger model load
-                    # payload = self.benchmark_handler.make_benchmark_payload()
-                    # _ = await self.__call_api(
-                    #     handler=self.benchmark_handler, payload=payload
-                    # )
+                    payload = self.benchmark_handler.make_benchmark_payload()
+                    _ = await self.__call_api(
+                        handler=self.benchmark_handler, payload=payload
+                    )
                     return float(f.readline())
             except FileNotFoundError:
                 pass
@@ -336,26 +308,18 @@ class Backend:
 
             for run in range(1, self.benchmark_handler.benchmark_runs + 1):
                 start = time.time()
-                benchmark_requests = []
+                tasks = []
+                total_workload = 0
 
-                for i in range(concurrent_requests):
+                for _ in range(concurrent_requests):
                     payload = self.benchmark_handler.make_benchmark_payload()
-                    workload = payload.count_workload()
-                    task = self.__call_api(handler=self.benchmark_handler, payload=payload)
-                    benchmark_requests.append(
-                        BenchmarkResult(request_idx=i, workload=workload, task=task)
+                    total_workload += payload.count_workload()
+                    tasks.append(
+                        self.__call_api(handler=self.benchmark_handler, payload=payload)
                     )
 
-                responses = await gather(*[br.task for br in benchmark_requests])
-                for br, response in zip(benchmark_requests, responses):
-                    br.response = response
-
-                total_workload = sum(br.workload for br in benchmark_requests if br.is_successful)
+                responses = await gather(*tasks)
                 time_elapsed = time.time() - start
-                successful_responses = sum([1 for br in benchmark_requests if br.is_successful])
-                if successful_responses == 0:
-                    self.backend_errored("No successful responses from benchmark")
-                    log.debug(f"benchmark failed: {successful_responses}/{concurrent_requests} successful responses")
 
                 throughput = total_workload / time_elapsed
                 sum_throughput += throughput
@@ -369,7 +333,7 @@ class Backend:
                             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: {successful_responses}/{concurrent_requests}",
+                            f"Successful responses: {len([r for r in responses if r.status == 200])}",
                             "#" * 60,
                         ]
                     )
@@ -396,7 +360,7 @@ class Backend:
                         )
                         # some backends need a few seconds after logging successful startup before
                         # they can begin accepting requests
-                        # await sleep(5)
+                        await sleep(5)
                         try:
                             max_throughput = await run_benchmark()
                             self.__start_healthcheck = True
@@ -417,13 +381,13 @@ class Backend:
 
         async def tail_log():
             log.debug(f"tailing file: {self.model_log_file}")
-            async with await open_file(self.model_log_file, encoding='utf-8', errors='ignore') as f:
+            async with await open_file(self.model_log_file) as f:
                 while True:
                     line = await f.readline()
                     if line:
                         await handle_log_line(line.rstrip())
                     else:
-                        await asyncio.sleep(LOG_POLL_INTERVAL)
+                        time.sleep(LOG_POLL_INTERVAL)
 
         ###########
 
@@ -431,4 +395,4 @@ class Backend:
             if os.path.isfile(self.model_log_file) is True:
                 return await tail_log()
             else:
-                await asyncio.sleep(1)
+                await sleep(1)

lib/data_types.py

@@ -3,7 +3,7 @@ import logging
 from dataclasses import dataclass, field
 from enum import Enum
 from abc import ABC, abstractmethod
-from typing import Dict, Any, Union, Tuple, Optional, Set, TypeVar, Generic, Type, Awaitable
+from typing import Dict, Any, Union, Tuple, Optional, Set, TypeVar, Generic, Type
 from aiohttp import web, ClientResponse
 import inspect
 
@@ -65,11 +65,10 @@ class ApiPayload(ABC):
 class AuthData:
     """data used to authenticate requester"""
 
+    signature: str
     cost: str
     endpoint: str
     reqnum: int
-    request_idx: int
-    signature: str
     url: str
 
     @classmethod
@@ -190,34 +189,13 @@ class SystemMetrics:
         self.additional_disk_usage = disk_usage - self.last_disk_usage
         self.last_disk_usage = disk_usage
 
-    def reset(self, expected: float | None) -> None:
+    def reset(self):
         # autoscaler excepts model_loading_time to be populated only once, when the instance has
         # finished benchmarking and is ready to receive requests. This applies to restarted instances
         # as well: they should send model_loading_time once when they are done loading
-        if self.model_loading_time == expected:
-            self.model_loading_time = None
+        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 BenchmarkResult:
-    request_idx: int
-    workload: float
-    task: Awaitable[ClientResponse]
-    response: Optional[ClientResponse] = None
-
-    @property
-    def is_successful(self) -> bool:
-        return self.response is not None and self.response.status == 200
-
 @dataclass
 class ModelMetrics:
     """Model specific metrics"""
@@ -227,14 +205,12 @@ class ModelMetrics:
     workload_received: float
     workload_cancelled: float
     workload_errored: float
-    workload_rejected: float
     # these are not
     workload_pending: float
     error_msg: Optional[str]
     max_throughput: float
     requests_recieved: Set[int] = field(default_factory=set)
-    requests_working: dict[int, RequestMetrics] = field(default_factory=dict)
-    requests_deleting: list[RequestMetrics] = field(default_factory=list)
+    requests_working: Set[int] = field(default_factory=set)
     last_update: float = field(default_factory=time.time)
 
     @classmethod
@@ -244,30 +220,19 @@ class ModelMetrics:
             workload_served=0.0,
             workload_cancelled=0.0,
             workload_errored=0.0,
-            workload_rejected=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)
 
-    @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()]) / max(self.max_throughput, 0.00001)
-    
-    @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
@@ -277,21 +242,16 @@ 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 AutoScalerData:
+class AutoScalaerData:
     """Data that is reported to autoscaler"""
 
     id: int
-    mtoken: str
-    version: str
     loadtime: float
     cur_load: float
-    rej_load: float
-    new_load: float
     error_msg: str
     max_perf: float
     cur_perf: float
@@ -300,7 +260,6 @@ class AutoScalerData:
     num_requests_working: int
     num_requests_recieved: int
     additional_disk_usage: float
-    working_request_idxs: list[int]
     url: str
 
 

lib/metrics.py

@@ -5,14 +5,13 @@ 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
 
-from lib.data_types import AutoScalerData, SystemMetrics, ModelMetrics, RequestMetrics
+import requests
+
+from lib.data_types import AutoScalaerData, SystemMetrics, ModelMetrics
 from typing import Awaitable, NoReturn, List
 
 METRICS_UPDATE_INTERVAL = 1
-DELETE_REQUESTS_INTERVAL = 1
 
 log = logging.getLogger(__file__)
 
@@ -27,10 +26,7 @@ def get_url() -> str:
 
 @dataclass
 class Metrics:
-    version: str = "0"
-    mtoken: str = ""
     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(
@@ -39,84 +35,42 @@ 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)
 
-    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:
+    def _request_start(self, workload: float, reqnum: int) -> None:
         """
         this function is called prior to forwarding a request to a model API.
         """
         log.debug("request start")
-        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
+        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, request: RequestMetrics) -> None:
+    def _request_end(self, workload: float, reqnum: int) -> 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()
+        self.model_metrics.workload_pending -= workload
+        self.model_metrics.requests_working.discard(reqnum)
 
-    def _request_success(self, request: RequestMetrics) -> None:
+    def _request_success(self, workload: float) -> 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.model_metrics.workload_served += workload
         self.update_pending = True
 
-    def _request_errored(self, request: RequestMetrics) -> None:
+    def _request_errored(self, workload: float) -> None:
         """
         this function is called if model API returns an error
         """
-        self.model_metrics.workload_errored += request.workload
-        request.status = "Error"
-        request.success = False
-        self.update_pending = True
+        self.model_metrics.workload_errored += workload
 
-    def _request_canceled(self, request: RequestMetrics) -> None:
+    def _request_canceled(self, workload: float) -> None:
         """
         this function is called if client drops connection before model API has responded
         """
-        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()
+        self.model_metrics.workload_cancelled += workload
 
     async def _send_metrics_loop(self) -> Awaitable[NoReturn]:
         while True:
@@ -124,10 +78,10 @@ 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")
-                await self.__send_metrics_and_reset()
+                self.__send_metrics_and_reset(elapsed)
             elif self.update_pending or elapsed > 10:
                 log.debug(f"sending loaded model metrics after {int(elapsed)}s wait")
-                await self.__send_metrics_and_reset()
+                self.__send_metrics_and_reset(elapsed)
 
     def _model_loaded(self, max_throughput: float) -> None:
         self.system_metrics.model_loading_time = (
@@ -140,130 +94,49 @@ 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
-
-    def _set_mtoken(self, mtoken: str) -> None:
-        self.mtoken = mtoken
-
     #######################################Private#######################################
 
-    async def __send_delete_requests_and_reset(self):
-        async def post(report_addr: str, idxs: list[int], success_flag: bool) -> bool:
-            data = {
-                "worker_id": self.id,
-                "mtoken": self.mtoken,
-                "request_idxs": idxs,
-                "success": success_flag,
-            }
-            log.debug(
-                f"Deleting requests that {'succeeded' if success_flag else 'failed'}: {data['request_idxs']}"
-            )
-            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:
-                        log.debug(f"delete_requests response: {res.status}")
-                        res.raise_for_status()
-                    return True
-                except asyncio.TimeoutError:
-                    log.debug("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}")
-            return False
+    def __send_metrics_and_reset(self, elapsed):
 
-        # Take a snapshot of what we plan to send this tick.
-        # New arrivals after this snapshot will remain in the queue for the next tick.
-        snapshot = list(self.model_metrics.requests_deleting)
-        success_idxs = [r.request_idx for r in snapshot if r.success is True]
-        failed_idxs  = [r.request_idx for r in snapshot if r.success is False]
-
-        if not success_idxs and not failed_idxs:
-            return  # nothing to do
-
-        for report_addr in self.report_addr:
-            # TODO: Add a Redis subscriber queue for delete_requests
-            if report_addr == "https://cloud.vast.ai/api/v0":
-                # Patch: ignore the Redis API report_addr
-                continue
-            sent_success = True
-            sent_failed  = True
-
-            if success_idxs:
-                sent_success = await post(report_addr, success_idxs, True)
-            if failed_idxs:
-                sent_failed = await post(report_addr, failed_idxs, False)
-
-            if sent_success and sent_failed:
-                # Remove only the items we actually sent from the live queue.
-                sent_set = set(success_idxs) | set(failed_idxs)
-                self.model_metrics.requests_deleting[:] = [
-                    r for r in self.model_metrics.requests_deleting
-                    if r.request_idx not in sent_set
-                ]
-                break
-
-
-    async def __send_metrics_and_reset(self):
-
-        loadtime_snapshot = self.system_metrics.model_loading_time
-
-        def compute_autoscaler_data() -> AutoScalerData:
-            return AutoScalerData(
+        def compute_autoscaler_data() -> AutoScalaerData:
+            return AutoScalaerData(
                 id=self.id,
-                mtoken=self.mtoken,
-                version=self.version,
-                loadtime=(loadtime_snapshot or 0.0), 
-                new_load=self.model_metrics.workload_processing,
-                cur_load=self.model_metrics.cur_load,
-                rej_load=self.model_metrics.workload_rejected,
+                loadtime=(self.system_metrics.model_loading_time or 0.0),
+                cur_load=(self.model_metrics.workload_processing / elapsed),
                 max_perf=self.model_metrics.max_throughput,
-                cur_perf=self.model_metrics.workload_served,
+                cur_perf=self.model_metrics.cur_perf,
                 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,
             )
 
-        async def send_data(report_addr: str) -> bool:
+        def send_data(report_addr: str) -> bool:
             data = compute_autoscaler_data()
-            log_data = asdict(data)
-            def obfuscate(secret: str) -> str:
-                if secret is None:
-                    return ""
-                return secret[:7] + "..." if len(secret) > 7 else ("*" * len(secret))
-            
-            log_data["mtoken"] = obfuscate(log_data.get("mtoken"))
+            full_path = report_addr.rstrip("/") + "/worker_status/"
             log.debug(
                 "\n".join(
                     [
                         "#" * 60,
                         f"sending data to autoscaler",
-                        f"{json.dumps(log_data, indent=2)}",
+                        f"{json.dumps((asdict(data)), indent=2)}",
                         "#" * 60,
                     ]
                 )
             )
-
-            full_path = report_addr.rstrip("/") + "/worker_status/"
             for attempt in range(1, 4):
                 try:
-                    session = await self.http()
-                    async with session.post(full_path, json=asdict(data)) as res:
-                        res.raise_for_status()
+                    res = requests.post(full_path, json=asdict(data), timeout=1)
+                    res.raise_for_status()
                     return True
-                except asyncio.TimeoutError:
+                except requests.Timeout:
                     log.debug(f"autoscaler status update timed out")
-                except (ClientResponseError, Exception)  as e:
+                except Exception as e:
                     log.debug(f"autoscaler status update failed with error: {e}")
-                await asyncio.sleep(2)
+                time.sleep(2)
                 log.debug(f"retrying autoscaler status update, attempt: {attempt}")
             log.debug(f"failed to send update through {report_addr}")
             return False
@@ -272,15 +145,11 @@ class Metrics:
 
         self.system_metrics.update_disk_usage()
 
-        sent = False
         for report_addr in self.report_addr:
-            if await send_data(report_addr):
-                sent = True
+            success = send_data(report_addr)
+            if success is True:
                 break
-
-        if sent:
-            # clear the one-shot loadtime only if we actually sent *this* value
-            self.system_metrics.reset(expected=loadtime_snapshot)
-            self.update_pending = False
-            self.model_metrics.reset()
-            self.last_metric_update = time.time()
+        self.update_pending = False
+        self.model_metrics.reset()
+        self.system_metrics.reset()
+        self.last_metric_update = time.time()

requirements.txt

@@ -8,4 +8,3 @@ Requests~=2.32
 transformers~=4.52
 utils==1.0.*
 hf_transfer>=0.1.9
-vastai-sdk>=0.2.0

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"
@@ -59,12 +59,12 @@ then
     fi
 
     # Fork testing
-    [[ ! -d $SERVER_DIR ]] && git clone "${PYWORKER_REPO:-https://github.com/vast-ai/pyworker}" "$SERVER_DIR"
+    git clone "${PYWORKER_REPO:-https://github.com/vast-ai/pyworker}" "$SERVER_DIR"
     if [[ -n ${PYWORKER_REF:-} ]]; then
         (cd "$SERVER_DIR" && git checkout "$PYWORKER_REF")
     fi
 
-    uv venv --python-preference only-managed "$ENV_PATH" -p 3.10
+    uv venv --managed-python "$ENV_PATH" -p 3.10
     source "$ENV_PATH/bin/activate"
 
     uv pip install -r "${SERVER_DIR}/requirements.txt"
@@ -124,5 +124,9 @@ cd "$SERVER_DIR"
 
 echo "launching PyWorker server"
 
+# if instance is rebooted, we want to clear out the log file so pyworker doesn't read lines
+# from the run prior to reboot. past logs are saved in $MODEL_LOG.old for debugging only
+[ -e "$MODEL_LOG" ] && cat "$MODEL_LOG" >> "$MODEL_LOG.old" && : > "$MODEL_LOG"
+
 (python3 -m "workers.$BACKEND.server" |& tee -a "$PYWORKER_LOG") &
 echo "launching PyWorker server done"

workers/comfyui-json/README.md

@@ -12,21 +12,9 @@ A docker image is provided but you may use any if the above requirements are met
 
 ## Benchmarking
 
-### Custom Benchmark Workflows
+A simple image generation benchmark runs when each worker initializes to validate GPU performance and identify underperforming machines.
 
-You can provide a custom ComfyUI workflow for benchmarking by creating `workers/comfyui-json/misc/benchmark.json`. This allows you to test performance using your preferred models and workflow complexity.
-
-**Ways to provide the benchmark file:**
-- Fork this repository and add your `benchmark.json` file
-- Write the file during worker provisioning (onstart script or setup phase)
-
-An example file is provided in the repository. To ensure varied generations, use the placeholder `__RANDOM_INT__` in place of static seed values - it will be replaced with a random integer for each benchmark run.
-
-### Default Benchmark (Fallback)
-
-If `benchmark.json` is not available, a simple image generation benchmark runs when each worker initializes. This validates GPU performance and helps identify underperforming machines.
-
-The default benchmark uses Stable Diffusion v1.5 with ComfyUI's standard text-to-image workflow. Configure it using these environment variables:
+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 |
 | -------------------- | ------------- | ----------- |
@@ -36,7 +24,7 @@ The default benchmark uses Stable Diffusion v1.5 with ComfyUI's standard text-to
 
 Each benchmark run uses a random prompt from `misc/test_prompts.txt` and a random seed to ensure consistent GPU load patterns.
 
-#### Calibrating Fallback Benchmark Duration
+### 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.
 

workers/comfyui-json/client.py

@@ -1,35 +1,155 @@
-from .data_types import count_workload
+import logging
 import uuid
 import random
-import asyncio
-import random
+from urllib.parse import urljoin
+import json
 
-from vastai import Serverless
+import requests
 
-async def main():
-    async with Serverless() as client:
-        endpoint = await client.get_endpoint(name="my-comfy-endpoint") # Change this to your endpoint name
+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
 
-        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
-            }
+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
         }
+    }
     
-        response = await endpoint.request("/generate/sync", payload, cost=count_workload())
+    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
 
-        # Get the file from the path on the local machine using SCP or SFTP
-        # or configure S3 to upload to cloud storage.
-        print(response["response"]["output"][0]["local_path"])
 
 if __name__ == "__main__":
-    asyncio.run(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

@@ -5,13 +5,12 @@ import dataclasses
 from typing import Dict, Any
 from functools import cache
 from math import ceil
-from pathlib import Path
-import json
-import logging
 
 from lib.data_types import ApiPayload, JsonDataException
 
-log = logging.getLogger(__file__)
+
+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
@@ -25,32 +24,9 @@ class ComfyWorkflowData(ApiPayload):
     @classmethod
     def for_test(cls):
         """
-        If the user has provided a benchmark workflow we can use it here to properly gauge performance.
-        Otherwise, use the variables available to simulate workflows of the required running time
+        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)
         """
-        # Try to load benchmark.json
-        benchmark_file = Path("workers/comfyui-json/misc/benchmark.json")
-        
-        if benchmark_file.exists():
-            try:
-                with open(benchmark_file, "r") as f:
-                    benchmark_workflow = json.load(f)
-                return cls(
-                    input={
-                        "request_id": f"test-{random.randint(1000, 99999)}",
-                        "workflow_json": benchmark_workflow
-                    }
-                )
-            except (json.JSONDecodeError, IOError):
-                # JSON is malformed or file can't be read, fall through to default
-                log.error(f"Failed to benchmark using {benchmark_file}")
-        
-        # Fallback: read prompts and construct payload
-        log.info("Using fallback method for benchmarking")
-        with open("workers/comfyui-json/misc/test_prompts.txt", "r") as f:
-            test_prompts = f.readlines()
-        
         test_prompt = random.choice(test_prompts).rstrip()
         return cls(
             input={

workers/comfyui-json/misc/benchmark.json.example

@@ -1,107 +0,0 @@
-{
-    "3": {
-        "inputs": {
-            "seed": "__RANDOM_INT__",
-            "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",
-        "_meta": {
-            "title": "KSampler"
-        }
-    },
-    "4": {
-        "inputs": {
-            "ckpt_name": "v1-5-pruned-emaonly-fp16.safetensors"
-        },
-        "class_type": "CheckpointLoaderSimple",
-        "_meta": {
-            "title": "Load Checkpoint"
-        }
-    },
-    "5": {
-        "inputs": {
-            "width": 512,
-            "height": 512,
-            "batch_size": 1
-        },
-        "class_type": "EmptyLatentImage",
-        "_meta": {
-            "title": "Empty Latent Image"
-        }
-    },
-    "6": {
-        "inputs": {
-            "text": "beautiful scenery nature glass bottle landscape, , purple galaxy bottle,",
-            "clip": [
-            "4",
-            1
-            ]
-        },
-        "class_type": "CLIPTextEncode",
-        "_meta": {
-            "title": "CLIP Text Encode (Prompt)"
-        }
-    },
-    "7": {
-        "inputs": {
-            "text": "text, watermark",
-            "clip": [
-            "4",
-            1
-            ]
-        },
-        "class_type": "CLIPTextEncode",
-        "_meta": {
-            "title": "CLIP Text Encode (Prompt)"
-        }
-    },
-    "8": {
-        "inputs": {
-            "samples": [
-            "3",
-            0
-            ],
-            "vae": [
-            "4",
-            2
-            ]
-        },
-        "class_type": "VAEDecode",
-        "_meta": {
-            "title": "VAE Decode"
-        }
-    },
-    "9": {
-        "inputs": {
-            "filename_prefix": "ComfyUI",
-            "images": [
-            "8",
-            0
-            ]
-        },
-        "class_type": "SaveImage",
-        "_meta": {
-            "title": "Save Image"
-        }
-    }
-}

workers/comfyui-json/server.py

@@ -19,7 +19,6 @@ 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
-    "[ERROR] Provisioning Script failed", # Error inserted by provisioning script if models/nodes fail to download
 ]
 
 
@@ -71,7 +70,7 @@ class ComfyWorkflowHandler(EndpointHandler[ComfyWorkflowData]):
 
     @property
     def healthcheck_endpoint(self) -> Optional[str]:
-        return f"{MODEL_SERVER_URL}/health"
+        return "/health"
 
     @classmethod
     def payload_cls(cls) -> Type[ComfyWorkflowData]:

workers/comfyui/client.py

@@ -7,13 +7,20 @@ from lib.test_utils import print_truncate_res
 from utils.endpoint_util import Endpoint
 from utils.ssl import get_cert_file_path
 
-from vastai import Serverless
+"""
+NOTE: this client example uses a custom comfy workflow compatible with SD3 only
+"""
+logging.basicConfig(
+    level=logging.DEBUG,
+    format="%(asctime)s[%(levelname)-5s] %(message)s",
+    datefmt="%Y-%m-%d %H:%M:%S",
+)
+log = logging.getLogger(__file__)
 
 
-ENDPOINT_NAME = "my-comfyui-endpoint"
-COST = 100 # Use a constant cost for image generation
-
-def call_default_workflow(client: Serverless) -> None:
+def call_default_workflow(
+    endpoint_group_name: str, api_key: str, server_url: str
+) -> None:
     WORKER_ENDPOINT = "/prompt"
     COST = 100
     route_payload = {
@@ -75,7 +82,6 @@ def call_custom_workflow_for_sd3(
         endpoint=message["endpoint"],
         reqnum=message["reqnum"],
         url=message["url"],
-        request_idx=message["request_idx"],
     )
     workflow = {
         "3": {

workers/openai/client.py

@@ -1,15 +1,14 @@
 import logging
-import json
-import os
 import sys
+import json
 import subprocess
-import argparse
-from typing import Any, Dict, List, Optional
+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
 
-from vastai import Serverless
-import asyncio
-
-# ---------------------- Logging ----------------------
 logging.basicConfig(
     level=logging.DEBUG,
     format="%(asctime)s[%(levelname)-5s] %(message)s",
@@ -17,20 +16,135 @@ logging.basicConfig(
 )
 log = logging.getLogger(__file__)
 
-# ---------------------- Prompts ----------------------
 COMPLETIONS_PROMPT = "the capital of USA is"
 CHAT_PROMPT = "Think step by step: Tell me about the Python programming language."
-TOOLS_PROMPT = (
-    "Can you list the files in the current working directory and tell me what you see? "
-    "What do you think this directory might be for?"
-)
+TOOLS_PROMPT = "Can you list the files in the current working directory and tell me what you see? What do you think this directory might be for?"
+
+
+class APIClient:
+    """Lightweight client focused solely on API communication"""
+
+    # Remove the generic WORKER_ENDPOINT since we're now going direct
+    DEFAULT_COST = 100
+    DEFAULT_TIMEOUT = 4
+
+    def __init__(
+        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.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"""
+        if not self.endpoint_api_key:
+            raise ValueError("No valid endpoint API key available")
+
+        route_payload = {
+            "endpoint": self.endpoint_group_name,
+            "api_key": self.endpoint_api_key,
+            "cost": cost,
+        }
+
+        response = requests.post(
+            urljoin(self.server_url, "/route/"),
+            json=route_payload,
+            timeout=self.DEFAULT_TIMEOUT,
+        )
+        response.raise_for_status()
+        return response.json()
+
+    def _create_auth_data(self, message: Dict[str, Any]) -> Dict[str, Any]:
+        """Create auth data from routing response"""
+        return {
+            "signature": message["signature"],
+            "cost": message["cost"],
+            "endpoint": message["endpoint"],
+            "reqnum": message["reqnum"],
+            "url": message["url"],
+        }
+
+    def _make_request(
+        self,
+        payload: Dict[str, Any],
+        endpoint: str,
+        method: str = "POST",
+        stream: bool = False,
+    ) -> Union[Dict[str, Any], Iterator[str]]:
+        """Make request directly to the specific worker endpoint"""
+        # Get worker URL and auth data
+        cost = payload.get("max_tokens", self.DEFAULT_COST)
+        message = self._get_worker_url(cost=cost)
+        worker_url = message["url"]
+        auth_data = self._create_auth_data(message)
+
+        req_data = {"payload": {"input": payload}, "auth_data": auth_data}
+
+        url = urljoin(worker_url, endpoint)
+        log.debug(f"Making direct request to: {url}")
+        log.debug(f"Payload: {req_data}")
+
+        # Make the request using the specified method
+        if method.upper() == "POST":
+            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, verify=get_cert_file_path()
+            )
+        else:
+            raise ValueError(f"Unsupported HTTP method: {method}")
+
+        response.raise_for_status()
+
+        if stream:
+            return self._handle_streaming_response(response)
+        else:
+            return response.json()
+
+    def _handle_streaming_response(self, response: requests.Response) -> Iterator[str]:
+        """Handle streaming response and yield tokens"""
+        try:
+            for line in response.iter_lines(decode_unicode=True):
+                if line:
+                    if line.startswith("data: "):
+                        data_str = line[6:]
+                        if data_str.strip() == "[DONE]":
+                            break
+                        try:
+                            data = json.loads(data_str)
+                            yield data  # Yield the full chunk
+                        except json.JSONDecodeError:
+                            continue
+        except Exception as e:
+            log.error(f"Error handling streaming response: {e}")
+            raise
+
+    def call_completions(
+        self, config: CompletionConfig
+    ) -> Union[Dict[str, Any], Iterator[str]]:
+        payload = config.to_dict()
+
+        return self._make_request(
+            payload=payload, endpoint="/v1/completions", stream=config.stream
+        )
+
+    def call_chat_completions(
+        self, config: ChatCompletionConfig
+    ) -> Union[Dict[str, Any], Iterator[str]]:
+        payload = config.to_dict()
+
+        return self._make_request(
+            payload=payload, endpoint="/v1/chat/completions", stream=config.stream
+        )
 
-ENDPOINT_NAME = "my-vllm-endpoint"       # change this to your vLLM endpoint name
-DEFAULT_MODEL = "Qwen/Qwen3-8B"          # must support tool calling
-MAX_TOKENS = 1024
-DEFAULT_TEMPERATURE = 0.7
 
-# ---------------------- Tooling ----------------------
 class ToolManager:
     """Handles tool definitions and execution"""
 
@@ -50,7 +164,7 @@ class ToolManager:
 
     @staticmethod
     def get_ls_tool_definition() -> List[Dict[str, Any]]:
-        """OpenAI-compatible tool schema"""
+        """Get the ls tool definition"""
         return [
             {
                 "type": "function",
@@ -64,217 +178,98 @@ class ToolManager:
 
     def execute_tool_call(self, tool_call: Dict[str, Any]) -> str:
         """Execute a tool call and return the result"""
-        function_name = (tool_call.get("function") or {}).get("name")
+        function_name = tool_call["function"]["name"]
+
         if function_name == "list_files":
             return self.list_files()
-        raise ValueError(f"Unknown tool function: {function_name}")
+        else:
+            raise ValueError(f"Unknown tool function: {function_name}")
 
 
-# ----- Helpers to handle streamed tool_calls assembly -----
-def _merge_tool_call_delta(state: Dict[int, Dict[str, Any]], tc_delta: Dict[str, Any]) -> None:
-    """
-    OpenAI-style streaming sends partial tool_calls with an index and partial fields.
-    We merge into a per-index state dict until the assistant message finishes.
-    """
-    idx = tc_delta.get("index")
-    if idx is None:
-        return
-
-    entry = state.setdefault(idx, {"id": None, "function": {"name": None, "arguments": ""}, "type": "function"})
-
-    if tc_delta.get("id"):
-        entry["id"] = tc_delta["id"]
-
-    fn_delta = tc_delta.get("function") or {}
-    if "name" in fn_delta and fn_delta["name"]:
-        entry["function"]["name"] = fn_delta["name"]
-    if "arguments" in fn_delta and fn_delta["arguments"]:
-        entry["function"]["arguments"] += fn_delta["arguments"]
-
-
-def _tool_state_to_message_tool_calls(state: Dict[int, Dict[str, Any]]) -> List[Dict[str, Any]]:
-    return [state[i] for i in sorted(state.keys())]
-
-
-# ---- OpenAI-compatible calls (non-streaming) ----
-async def call_completions(client: Serverless, *, model: str, prompt: str, **kwargs) -> Dict[str, Any]:
-
-    endpoint = await client.get_endpoint(name=ENDPOINT_NAME)
-
-    payload = {
-        "input": {
-            "model": model,
-            "prompt": prompt,
-            "max_tokens": kwargs.get("max_tokens", MAX_TOKENS),
-            "temperature": kwargs.get("temperature", DEFAULT_TEMPERATURE),
-        }
-    }
-    log.debug("POST /v1/completions %s", json.dumps(payload)[:500])
-    resp = await endpoint.request("/v1/completions", payload, cost=payload["input"]["max_tokens"])
-    return resp["response"]
-
-async def call_chat_completions(client: Serverless, *, model: str, messages: List[Dict[str, Any]], **kwargs) -> Dict[str, Any]:
-
-    endpoint = await client.get_endpoint(name=ENDPOINT_NAME)
-
-    payload = {
-        "input": {
-            "model": model,
-            "messages": messages,
-            "max_tokens": kwargs.get("max_tokens", MAX_TOKENS),
-            "temperature": kwargs.get("temperature", DEFAULT_TEMPERATURE),
-            **({"tools": kwargs["tools"]} if "tools" in kwargs else {}),
-            **({"tool_choice": kwargs["tool_choice"]} if "tool_choice" in kwargs else {}),
-        }
-    }
-    log.debug("POST /v1/chat/completions %s", json.dumps(payload)[:500])
-    resp = await endpoint.request("/v1/chat/completions", payload, cost=payload["input"]["max_tokens"])
-    return resp["response"]
-
-# ---- Streaming variants ----
-async def stream_completions(client: Serverless, *, model: str, prompt: str, **kwargs):
-
-    endpoint = await client.get_endpoint(name=ENDPOINT_NAME)
-
-    payload = {
-        "input": {
-            "model": model,
-            "prompt": prompt,
-            "max_tokens": kwargs.get("max_tokens", MAX_TOKENS),
-            "temperature": kwargs.get("temperature", DEFAULT_TEMPERATURE),
-            "stream": True,
-            **({"stop": kwargs["stop"]} if "stop" in kwargs else {}),
-        }
-    }
-    log.debug("STREAM /v1/completions %s", json.dumps(payload)[:500])
-    resp = await endpoint.request("/v1/completions", payload, cost=payload["input"]["max_tokens"], stream=True)
-    return resp["response"]  # async generator
-
-async def stream_chat_completions(client: Serverless, *, model: str, messages: List[Dict[str, Any]], **kwargs):
-
-    endpoint = await client.get_endpoint(name=ENDPOINT_NAME)
-
-    payload = {
-        "input": {
-            "model": model,
-            "messages": messages,
-            "max_tokens": kwargs.get("max_tokens", MAX_TOKENS),
-            "temperature": kwargs.get("temperature", DEFAULT_TEMPERATURE),
-            "stream": True,
-            **({"tools": kwargs["tools"]} if "tools" in kwargs else {}),
-            **({"tool_choice": kwargs["tool_choice"]} if "tool_choice" in kwargs else {}),
-        }
-    }
-    log.debug("STREAM /v1/chat/completions %s", json.dumps(payload)[:500])
-    resp = await endpoint.request("/v1/chat/completions", payload, cost=payload["input"]["max_tokens"], stream=True)
-    return resp["response"]  # async generator
-
-
-# ---------------------- Demo Runner ----------------------
 class APIDemo:
     """Demo and testing functionality for the API client"""
 
-    def __init__(self, client: Serverless, model: str, tool_manager: Optional[ToolManager] = None):
+    def __init__(
+        self, client: APIClient, model: str, tool_manager: Optional[ToolManager] = None
+    ):
         self.client = client
         self.model = model
         self.tool_manager = tool_manager or ToolManager()
 
-    # ----- Streaming handler -----
-    async def handle_streaming_response(self, stream, show_reasoning: bool = True) -> str:
+    def handle_streaming_response(
+        self, response_stream, show_reasoning: bool = True
+    ) -> str:
+        """
+        Handle streaming chat response and display all output.
+        """
+
         full_response = ""
         reasoning_content = ""
-        printed_reasoning = False
-        printed_answer = False
+        reasoning_started = False
+        content_started = False
 
-        async for chunk in stream:
-            choice = (chunk.get("choices") or [{}])[0]
-            delta = choice.get("delta", {})
+        for chunk in response_stream:
+            # Normalize the chunk
+            if isinstance(chunk, str):
+                chunk = chunk.strip()
+                if chunk.startswith("data: "):
+                    chunk = chunk[6:].strip()
+                if chunk in ["[DONE]", ""]:
+                    continue
+                try:
+                    parsed_chunk = json.loads(chunk)
+                except json.JSONDecodeError:
+                    continue
+            elif isinstance(chunk, dict):
+                parsed_chunk = chunk
+            else:
+                continue
 
-            # reasoning tokens
-            rc = delta.get("reasoning_content")
-            if rc and show_reasoning:
-                if not printed_reasoning:
+            # Parse delta from the chunk
+            choices = parsed_chunk.get("choices", [])
+            if not choices:
+                continue
+
+            delta = choices[0].get("delta", {})
+            reasoning_token = delta.get("reasoning_content", "")
+            content_token = delta.get("content", "")
+
+            # Print reasoning token if applicable
+            if show_reasoning and reasoning_token:
+                if not reasoning_started:
                     print("\n🧠 Reasoning: ", end="", flush=True)
-                    printed_reasoning = True
-                print(rc, end="", flush=True)
-                reasoning_content += rc
+                    reasoning_started = True
+                print(f"\033[90m{reasoning_token}\033[0m", end="", flush=True)
+                reasoning_content += reasoning_token
 
-            # content tokens
-            content_part = delta.get("content")
-            if content_part:
-                if not printed_answer:
-                    if show_reasoning and printed_reasoning:
-                        print("\n💬 Response: ", end="", flush=True)
+            # Print content token
+            if content_token:
+                if not content_started:
+                    if show_reasoning and reasoning_started:
+                        print(f"\n💬 Response: ", end="", flush=True)
                     else:
                         print("Assistant: ", end="", flush=True)
-                    printed_answer = True
-                print(content_part, end="", flush=True)
-                full_response += content_part
+                    content_started = True
+                print(content_token, end="", flush=True)
+                full_response += content_token
+
+        print()  # Ensure newline after response
 
-        print()  # newline
         if show_reasoning:
-            if printed_reasoning or printed_answer:
+            if reasoning_started or content_started:
                 print("\nStreaming completed.")
-            if printed_reasoning:
+            if reasoning_started:
                 print(f"Reasoning tokens: {len(reasoning_content.split())}")
-            if printed_answer:
+            if content_started:
                 print(f"Response tokens: {len(full_response.split())}")
 
         return full_response
 
-    async def demo_completions(self) -> None:
-        print("=" * 60)
-        print("COMPLETIONS DEMO")
-        print("=" * 60)
+    def test_tool_support(self) -> bool:
+        """Test if the endpoint supports function calling"""
+        log.debug("Testing endpoint tool calling support...")
 
-        response = await call_completions(
-            client=self.client,
-            model=self.model,
-            prompt=COMPLETIONS_PROMPT,
-            max_tokens=MAX_TOKENS,
-            temperature=DEFAULT_TEMPERATURE,
-        )
-        print("\nResponse:")
-        print(json.dumps(response, indent=2))
-
-    async def demo_chat(self, use_streaming: bool = True) -> None:
-        print("=" * 60)
-        print(f"CHAT COMPLETIONS DEMO {'(STREAMING)' if use_streaming else '(NON-STREAMING)'}")
-        print("=" * 60)
-
-        messages = [{"role": "user", "content": CHAT_PROMPT}]
-
-        if use_streaming:
-            stream = await stream_chat_completions(
-                client=self.client,
-                model=self.model, 
-                messages=messages,
-                max_tokens=MAX_TOKENS,
-                temperature=DEFAULT_TEMPERATURE
-            )
-            try:
-                await self.handle_streaming_response(stream, show_reasoning=True)
-            except Exception as e:
-                log.error("\nError during streaming: %s", e, exc_info=True)
-        else:
-            response = await call_chat_completions(
-                client=self.client,
-                model=self.model, 
-                messages=messages,
-                max_tokens=MAX_TOKENS,
-                temperature=DEFAULT_TEMPERATURE
-            )
-            choice = (response.get("choices") or [{}])[0]
-            message = choice.get("message", {})
-            content = message.get("content", "")
-            reasoning = message.get("reasoning_content", "") or message.get("reasoning", "")
-            if reasoning:
-                print(f"\n🧠 Reasoning: \033[90m{reasoning}\033[0m")
-            print(f"\n💬 Assistant: {content}")
-            print(f"\nFull Response:\n{json.dumps(response, indent=2)}")
-
-    async def test_tool_support(self) -> bool:
-        """Probe that tool schema is accepted (no actual call)"""
+        # Try a simple request with minimal tools to test support
         messages = [{"role": "user", "content": "Hello"}]
         minimal_tool = [
             {
@@ -282,147 +277,170 @@ class APIDemo:
                 "function": {"name": "test_function", "description": "Test function"},
             }
         ]
+
+        config = ChatCompletionConfig(
+            model=self.model,
+            messages=messages,
+            max_tokens=10,
+            tools=minimal_tool,
+            tool_choice="none",  # Don't actually call the tool
+        )
+
         try:
-            _ = await call_chat_completions(
-                client=self.client,
-                model=self.model,
-                messages=messages,
-                tools=minimal_tool,
-                tool_choice="none",
-                max_tokens=10
-            )
+            response = self.client.call_chat_completions(config)
             return True
         except Exception as e:
-            log.error("Endpoint does not support tool calling: %s", e)
+            log.error(f"Error: Endpoint does not support tool calling: {e}")
             return False
 
-    async def demo_ls_tool(self) -> None:
-        """Ask to list files using function calling, then provide final analysis"""
+    def demo_completions(self) -> None:
+        """Demo: test basic completions endpoint"""
+        print("=" * 60)
+        print("COMPLETIONS DEMO")
+        print("=" * 60)
+
+        config = CompletionConfig(
+            model=self.model, prompt=COMPLETIONS_PROMPT, stream=False
+        )
+
+        log.info(
+            f"Testing completions with model '{self.model}' and prompt: '{config.prompt}'"
+        )
+        response = self.client.call_completions(config)
+
+        if isinstance(response, dict):
+            print("\nResponse:")
+            print(json.dumps(response, indent=2))
+        else:
+            log.error("Unexpected response format")
+
+    def demo_chat(self, use_streaming: bool = True) -> None:
+        """
+        Demo: test chat completions endpoint with optional streaming
+        """
+        print("=" * 60)
+        print(
+            f"CHAT COMPLETIONS DEMO {'(STREAMING)' if use_streaming else '(NON-STREAMING)'}"
+        )
+        print("=" * 60)
+
+        config = ChatCompletionConfig(
+            model=self.model,
+            messages=[{"role": "user", "content": CHAT_PROMPT}],
+            stream=use_streaming,
+        )
+
+        log.info(f"Testing chat completions with model '{self.model}'...")
+        response = self.client.call_chat_completions(config)
+
+        if use_streaming:
+            try:
+                self.handle_streaming_response(response, show_reasoning=True)
+            except Exception as e:
+                log.error(f"\nError during streaming: {e}")
+                import traceback
+
+                traceback.print_exc()
+                return
+
+        else:
+            if isinstance(response, dict):
+                choice = response.get("choices", [{}])[0]
+                message = choice.get("message", {})
+                content = message.get("content", "")
+                reasoning = message.get("reasoning_content", "") or message.get(
+                    "reasoning", ""
+                )
+
+                if reasoning:
+                    print(f"\n🧠 Reasoning: \033[90m{reasoning}\033[0m")
+
+                print(f"\n💬 Assistant: {content}")
+                print(f"\nFull Response:")
+                print(json.dumps(response, indent=2))
+            else:
+                log.error("Unexpected response format")
+
+    def demo_ls_tool(self) -> None:
+        """Demo: ask LLM to list files in the current directory and describe what it sees"""
         print("=" * 60)
         print("TOOL USE DEMO: List Directory Contents")
         print("=" * 60)
 
-        if not await self.test_tool_support():
+        # Test if tools are supported first
+        if not self.test_tool_support():
             return
 
-        messages: List[Dict[str, Any]] = [{"role": "user", "content": TOOLS_PROMPT}]
+        # Request with tool available
+        messages = [{"role": "user", "content": TOOLS_PROMPT}]
 
-        # First pass: let the model decide tools, stream tool_calls and partial content
-        stream = await stream_chat_completions(
-            client=self.client,
+        config = ChatCompletionConfig(
             model=self.model,
             messages=messages,
             tools=self.tool_manager.get_ls_tool_definition(),
             tool_choice="auto",
-            max_tokens=MAX_TOKENS,
-            temperature=DEFAULT_TEMPERATURE,
         )
 
-        assistant_content_buf: List[str] = []
-        tool_calls_state: Dict[int, Dict[str, Any]] = {}
-        printed_reasoning = False
-        printed_answer = False
+        log.info(f"Making initial request with tool using model '{self.model}'...")
+        response = self.client.call_chat_completions(config)
 
-        async for chunk in stream:
-            choice = (chunk.get("choices") or [{}])[0]
-            delta = choice.get("delta", {})
+        if not isinstance(response, dict):
+            raise ValueError("Expected dict response for tool use")
 
-            rc = delta.get("reasoning_content")
-            if rc:
-                if not printed_reasoning:
-                    printed_reasoning = True
-                    print("🧠 Reasoning: ", end="", flush=True)
-                print(rc, end="", flush=True)
+        choice = response.get("choices", [{}])[0]
+        message = choice.get("message", {})
 
-            content_part = delta.get("content")
-            if content_part:
-                assistant_content_buf.append(content_part)
-                if not printed_answer:
-                    printed_answer = True
-                    print("\n💬 Response: ", end="", flush=True)
-                print(content_part, end="", flush=True)
+        print(f"Assistant response: {message.get('content', 'No content')}")
 
-            if "tool_calls" in delta and delta["tool_calls"]:
-                for tc_delta in delta["tool_calls"]:
-                    _merge_tool_call_delta(tool_calls_state, tc_delta)
+        # Check for tool calls
+        tool_calls = message.get("tool_calls")
+        if not tool_calls:
+            raise ValueError(
+                "No tool calls made - model may not support function calling"
+            )
 
-        # If no tool calls, we’re done.
-        if not tool_calls_state:
-            print("\n(No tool calls were made.)")
-            return
+        print(f"Tool calls detected: {len(tool_calls)}")
 
-        # Build assistant message with tool_calls
-        assistant_message = {
-            "role": "assistant",
-            "content": "".join(assistant_content_buf) if assistant_content_buf else None,
-            "tool_calls": _tool_state_to_message_tool_calls(tool_calls_state),
-        }
-        messages.append(assistant_message)
+        # Execute the tool call
+        for tool_call in tool_calls:
+            function_name = tool_call["function"]["name"]
+            print(f"Executing tool: {function_name}")
 
-        # Execute tools and feed results back
-        for tc in assistant_message["tool_calls"]:
-            tool_name = (tc.get("function") or {}).get("name")
-            call_id = tc.get("id")
-            raw_args = (tc.get("function") or {}).get("arguments") or "{}"
+            tool_result = self.tool_manager.execute_tool_call(tool_call)
+            print(f"Tool result:\n{tool_result}")
 
-            try:
-                args = json.loads(raw_args) if raw_args.strip() else {}
-            except Exception as e:
-                tool_result = json.dumps({"error": f"Argument parse failed: {str(e)}", "raw_arguments": raw_args})
-                messages.append({"role": "tool", "tool_call_id": call_id, "content": tool_result})
-                continue
+            # Add tool result and continue conversation
+            messages.append(message)  # Add assistant's message with tool call
+            messages.append(
+                {
+                    "role": "tool",
+                    "tool_call_id": tool_call["id"],
+                    "content": tool_result,
+                }
+            )
 
-            try:
-                if tool_name == "list_files":
-                    tool_result = self.tool_manager.list_files()
-                else:
-                    tool_result = json.dumps({"error": f"Unknown tool '{tool_name}'"})
-            except Exception as e:
-                tool_result = json.dumps({"error": f"Tool '{tool_name}' failed: {str(e)}"})
+            # Get final response
+            final_config = ChatCompletionConfig(
+                model=self.model,
+                messages=messages,
+                tools=self.tool_manager.get_ls_tool_definition(),
+            )
 
-            print("\n[Tool executed]", tool_name)
-            print(tool_result[:500] + ("..." if len(tool_result) > 500 else ""))
-            messages.append({"role": "tool", "tool_call_id": call_id, "content": tool_result})
+            print("Getting final response...")
+            final_response = self.client.call_chat_completions(final_config)
 
-        # Second pass: get final streamed answer after tool results
-        stream2 = await stream_chat_completions(
-            client=self.client,
-            model=self.model,
-            messages=messages,
-            max_tokens=MAX_TOKENS,
-            temperature=DEFAULT_TEMPERATURE,
-        )
+            if isinstance(final_response, dict):
+                final_choice = final_response.get("choices", [{}])[0]
+                final_message = final_choice.get("message", {})
+                final_content = final_message.get("content", "")
 
-        final_buf = []
-        printed_reasoning2 = False
-        printed_answer2 = False
+                print("\n" + "=" * 60)
+                print("FINAL LLM ANALYSIS:")
+                print("=" * 60)
+                print(final_content)
+                print("=" * 60)
 
-        async for chunk in stream2:
-            choice = (chunk.get("choices") or [{}])[0]
-            delta = choice.get("delta", {})
-
-            rc2 = delta.get("reasoning_content")
-            if rc2:
-                if not printed_reasoning2:
-                    printed_reasoning2 = True
-                    print("\n🧠 Reasoning (post-tools): ", end="", flush=True)
-                print(rc2, end="", flush=True)
-
-            c2 = delta.get("content")
-            if c2:
-                final_buf.append(c2)
-                if not printed_answer2:
-                    printed_answer2 = True
-                    print("\n💬 Response (final): ", end="", flush=True)
-                print(c2, end="", flush=True)
-
-        print("\n" + "=" * 60)
-        print("FINAL LLM ANALYSIS:")
-        print("=" * 60)
-        print("".join(final_buf))
-        print("=" * 60)
-
-    async def interactive_chat(self) -> None:
+    def interactive_chat(self) -> None:
         """Interactive chat session with streaming"""
         print("=" * 60)
         print("INTERACTIVE STREAMING CHAT")
@@ -431,7 +449,7 @@ class APIDemo:
         print("Type 'quit' to exit, 'clear' to clear history")
         print()
 
-        messages: List[Dict[str, Any]] = []
+        messages = []
 
         while True:
             try:
@@ -449,15 +467,16 @@ class APIDemo:
 
                 messages.append({"role": "user", "content": user_input})
 
-                print("Assistant: ", end="", flush=True)
-                stream = await stream_chat_completions(
-                    client=self.client,
-                    model=self.model, 
-                    messages=messages, 
-                    max_tokens=MAX_TOKENS, 
-                    temperature=0.7
+                config = ChatCompletionConfig(
+                    model=self.model, messages=messages, stream=True, temperature=0.7
+                )
+
+                print("Assistant: ", end="", flush=True)
+
+                response = self.client.call_chat_completions(config)
+                assistant_content = self.handle_streaming_response(
+                    response, show_reasoning=True
                 )
-                assistant_content = await self.handle_streaming_response(stream, show_reasoning=True)
 
                 # Add assistant response to conversation history
                 messages.append({"role": "assistant", "content": assistant_content})
@@ -466,64 +485,115 @@ class APIDemo:
                 print("\n👋 Chat interrupted. Goodbye!")
                 break
             except Exception as e:
-                log.error("\nError: %s", e)
+                log.error(f"\nError: {e}")
                 continue
 
 
-# ---------------------- CLI ----------------------
-def build_arg_parser() -> argparse.ArgumentParser:
-    p = argparse.ArgumentParser(description="Vast vLLM Demo (Serverless SDK)")
-    p.add_argument("--model", required=True, help="Model to use for requests (required)")
-    p.add_argument("--endpoint", default="my-vllm-endpoint", help="Vast endpoint name (default: my-vllm-endpoint)")
+def main():
+    """Main function with CLI switches for different tests"""
+    from lib.test_utils import test_args
 
-    modes = p.add_mutually_exclusive_group(required=False)
-    modes.add_argument("--completion", action="store_true", help="Test completions endpoint")
-    modes.add_argument("--chat", action="store_true", help="Test chat completions endpoint (non-streaming)")
-    modes.add_argument("--chat-stream", action="store_true", help="Test chat completions endpoint with streaming")
-    modes.add_argument("--tools", action="store_true", help="Test function calling with ls tool (non-streaming+streamed phases)")
-    modes.add_argument("--interactive", action="store_true", help="Start interactive streaming chat session")
-    return p
+    # Add mandatory model argument
+    test_args.add_argument(
+        "--model", required=True, help="Model to use for requests (required)"
+    )
 
+    # Add test mode arguments
+    test_args.add_argument(
+        "--completion", action="store_true", help="Test completions endpoint"
+    )
+    test_args.add_argument(
+        "--chat",
+        action="store_true",
+        help="Test chat completions endpoint (non-streaming)",
+    )
+    test_args.add_argument(
+        "--chat-stream",
+        action="store_true",
+        help="Test chat completions endpoint with streaming",
+    )
+    test_args.add_argument(
+        "--tools",
+        action="store_true",
+        help="Test function calling with ls tool (non-streaming)",
+    )
+    test_args.add_argument(
+        "--interactive",
+        action="store_true",
+        help="Start interactive streaming chat session",
+    )
 
-async def main_async():
-    args = build_arg_parser().parse_args()
+    args = test_args.parse_args()
 
-    selected = sum([args.completion, args.chat, args.chat_stream, args.tools, args.interactive])
-    if selected == 0:
+    # Check that only one test mode is selected
+    test_modes = [
+        args.completion,
+        args.chat,
+        args.chat_stream,
+        args.tools,
+        args.interactive,
+    ]
+    selected_count = sum(test_modes)
+
+    if selected_count == 0:
         print("Please specify exactly one test mode:")
         print("  --completion    : Test completions endpoint")
         print("  --chat          : Test chat completions endpoint (non-streaming)")
         print("  --chat-stream   : Test chat completions endpoint with streaming")
-        print("  --tools         : Test function calling with ls tool")
+        print("  --tools         : Test function calling with ls tool (non-streaming)")
         print("  --interactive   : Start interactive streaming chat session")
-        print(f"\nExample: python {os.path.basename(sys.argv[0])} --model Qwen/Qwen3-8B --chat-stream --endpoint my-vllm-endpoint")
+        print(
+            f"\nExample: python {sys.argv[0]} --model Qwen/Qwen3-8B --chat-stream -k YOUR_KEY -e YOUR_ENDPOINT"
+        )
         sys.exit(1)
-    elif selected > 1:
+    elif selected_count > 1:
         print("Please specify exactly one test mode")
         sys.exit(1)
 
-    print(f"Using model: {args.model}")
-    print("=" * 60)
-
     try:
-        async with Serverless() as client:
-            demo = APIDemo(client, args.model, ToolManager())
+        endpoint_api_key = Endpoint.get_endpoint_api_key(
+            endpoint_name=args.endpoint_group_name,
+            account_api_key=args.api_key,
+            instance=args.instance,
+        )
 
-            if args.completion:
-                await demo.demo_completions()
-            elif args.chat:
-                await demo.demo_chat(use_streaming=False)
-            elif args.chat_stream:
-                await demo.demo_chat(use_streaming=True)
-            elif args.tools:
-                await demo.demo_ls_tool()
-            elif args.interactive:
-                await demo.interactive_chat()
+        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=Endpoint.get_autoscaler_server_url(args.instance),
+            endpoint_api_key=endpoint_api_key,
+        )
+
+        # Create tool manager and demo (passing the model parameter)
+        tool_manager = ToolManager()
+        demo = APIDemo(client, args.model, tool_manager)
+
+        print(f"Using model: {args.model}")
+        print("=" * 60)
+
+        # Run the selected test
+        if args.completion:
+            demo.demo_completions()
+        elif args.chat:
+            demo.demo_chat(use_streaming=False)
+        elif args.chat_stream:
+            demo.demo_chat(use_streaming=True)
+        elif args.tools:
+            demo.demo_ls_tool()
+        elif args.interactive:
+            demo.interactive_chat()
 
     except Exception as e:
-        log.error("Error during test: %s", e, exc_info=True)
+        log.error(f"Error during test: {e}", exc_info=True)
         sys.exit(1)
 
 
 if __name__ == "__main__":
-    asyncio.run(main_async())
+    main()

workers/openai/data_types/server.py

@@ -119,25 +119,14 @@ class GenericHandler(EndpointHandler[GenericData], ABC):
 class CompletionsData(GenericData):
     @classmethod
     def for_test(cls) -> "CompletionsData":
-        system_prompt = """You are a helpful AI assistant. You have access to the following knowledge base:
-    
-        Zebras (US: /ˈziːbrəz/, UK: /ˈzɛbrəz, ˈziː-/)[2] (subgenus Hippotigris) are African equines 
-        with distinctive black-and-white striped coats. There are three living species: Grévy's zebra 
-        (Equus grevyi), the plains zebra (E. quagga), and the mountain zebra (E. zebra). Zebras share the 
-        genus Equus with horses and asses, the three groups being the only living members of the family 
-        Equidae. Zebra stripes come in different patterns, unique to each individual. Zebras inhabit eastern 
-        and southern Africa and can be found in a variety of habitats such as savannahs, grasslands, 
-        woodlands, shrublands, and mountainous areas.
-        
-        Please answer the following question based on the above context."""
-        unique_question = " ".join(random.choices(WORD_LIST, k=int(100)))
+        prompt = " ".join(random.choices(WORD_LIST, k=int(250)))
         model = os.environ.get("MODEL_NAME")
         if not model:
             raise ValueError("MODEL_NAME environment variable not set")
 
         test_input = {
             "model": model,
-            "prompt": f"{system_prompt}\n\n{unique_question}",
+            "prompt": prompt,
             "temperature": 0.7,
             "max_tokens": 500,
         }
@@ -164,18 +153,7 @@ class ChatCompletionsData(GenericData):
 
     @classmethod
     def for_test(cls) -> "ChatCompletionsData":
-        system_prompt = """You are a helpful AI assistant. You have access to the following knowledge base:
-    
-        Zebras (US: /ˈziːbrəz/, UK: /ˈzɛbrəz, ˈziː-/)[2] (subgenus Hippotigris) are African equines 
-        with distinctive black-and-white striped coats. There are three living species: Grévy's zebra 
-        (Equus grevyi), the plains zebra (E. quagga), and the mountain zebra (E. zebra). Zebras share the 
-        genus Equus with horses and asses, the three groups being the only living members of the family 
-        Equidae. Zebra stripes come in different patterns, unique to each individual. Zebras inhabit eastern 
-        and southern Africa and can be found in a variety of habitats such as savannahs, grasslands, 
-        woodlands, shrublands, and mountainous areas.
-        
-        Please answer the following question based on the above context."""
-        unique_question = " ".join(random.choices(WORD_LIST, k=int(100)))
+        prompt = " ".join(random.choices(WORD_LIST, k=int(250)))
         model = os.environ.get("MODEL_NAME")
         if not model:
             raise ValueError("MODEL_NAME environment variable not set")
@@ -183,10 +161,7 @@ class ChatCompletionsData(GenericData):
         # Chat completions use messages format instead of prompt
         test_input = {
             "model": model,
-            "messages": [
-                {"role": "system", "content": system_prompt},  # Shared prefix
-                {"role": "user", "content": unique_question}   # Unique per request
-            ],
+            "messages": [{"role": "user", "content": prompt}],
             "temperature": 0.7,
             "max_tokens": 500,
         }

workers/openai/test_load.py

@@ -42,7 +42,6 @@ class ReqResult:
     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
@@ -59,73 +58,31 @@ def do_one(endpoint_name: str,
            route_session,
            worker_session):
     try:
-        workload = payload.count_workload()
-        route_payload = {"endpoint": endpoint_name, "api_key": endpoint_api_key, "cost": workload}
+        u = payload.count_workload()
+        route_payload = {"endpoint": endpoint_name, "api_key": endpoint_api_key, "cost": u}
         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))
+            results_list.append(ReqResult("", (t_after_route - start) * 1000.0, 0.0, (t_after_route - start) * 1000.0, False,
+                                          f"route {r0.status_code} {r0.text}"))
             return
         msg = r0.json()
 
-        # 1) Check if we got a worker back from route
-        worker_url = msg.get("url", "")
-        if not worker_url:
-            status = msg.get("status", "")
+        # 1) "Status" is in the response when no worker is ready
+        worker_sampled = True
+        status = msg.get("status", "")
+        if status:
             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))
+                worker_sampled = False
 
-        # 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:
+        # 2) Otherwise (successful request), sample via /get_endpoint_workers/ for eligible (idle) worker tracking
+        if worker_sampled:
             try:
                 r_status = route_session.post(
                     urljoin(server_url, "/get_endpoint_workers/"),
@@ -143,18 +100,29 @@ def do_one(endpoint_name: str,
                     status_samples.append((time.time() - t0, idle))
             except Exception:
                 pass
+
+        # 3) Send the request
+        worker_address = msg["url"]
+        req = dict(payload=payload.__dict__, auth_data=AuthData.from_json_msg(msg).__dict__)
+        t1 = time.time()
+        # Use explicit connect/read timeouts to avoid long hangs
+        r1 = worker_session.post(
+            urljoin(worker_address, worker_endpoint),
+            json=req,
+            verify=get_cert_file_path(),
+            timeout=(4, 120),
+        )
+        t2 = time.time()
+        if r1.status_code != 200:
+            results_list.append(ReqResult(worker_address, (t_after_route - start) * 1000.0, (t2 - t1) * 1000.0,
+                                          (t2 - start) * 1000.0, False,
+                                          f"infer {r1.status_code} {r1.text}"))
+            return
+        results_list.append(ReqResult(worker_address, (t_after_route - start) * 1000.0, (t2 - t1) * 1000.0, (t2 - start) * 1000.0,
+                                        True, "", t_start=start - t0, t_end=t2 - t0, workload=u))
     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))
+        results_list.append(ReqResult("", (t - start) * 1000.0, 0.0, (t - start) * 1000.0, False, str(e)))
 
 def run_load_with_metrics(num_requests: int,
                           requests_per_second: float,
@@ -164,7 +132,7 @@ def run_load_with_metrics(num_requests: int,
                           worker_endpoint: str,
                           instance: str,
                           out_path: str):
-
+    # Resolve endpoint id + endpoint-scoped API key
     ep_info = Endpoint.get_endpoint_info(endpoint_name=endpoint_group_name,
                                          account_api_key=account_api_key,
                                          instance=instance)
@@ -177,7 +145,8 @@ def run_load_with_metrics(num_requests: int,
     t0 = time.time()
     results = []
     status_samples = []
-    max_concurrency = int(os.environ.get("MAX_CONCURRENCY", "8192"))
+    # Concurrency control
+    max_concurrency = int(os.environ.get("MAX_CONCURRENCY", "1024"))
     submit_queue_factor = 2  # cap queued tasks to reduce memory
 
     # Shared HTTP sessions with connection pooling (persistent connections)
@@ -189,9 +158,9 @@ def run_load_with_metrics(num_requests: int,
         return sess
 
     # Router: mostly single host, small connection pool is sufficient
-    route_session = make_session(pool_connections=1, pool_maxsize=max_concurrency)
+    route_session = make_session(pool_connections=8, 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)
+    worker_session = make_session(pool_connections=max(256, max_concurrency), pool_maxsize=max_concurrency)
 
     # Fire requests using a thread pool, scheduling at requested RPS
     inflight = set()
@@ -240,12 +209,11 @@ def run_load_with_metrics(num_requests: int,
 
     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([])
+    #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)
+    total_compute_time_ms = float(np.sum(worker_ms)) if succ else 0.0
 
     # Distribution over workers (by host:port)
     hosts = [urlparse(r.worker_url).netloc for r in oks if r.worker_url]
@@ -272,11 +240,11 @@ def run_load_with_metrics(num_requests: int,
 
     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}")
+    print(f"Total compute time (sum worker latency, s): {total_compute_time_ms/1000.0:.2f}")
     if errs:
         print("Sample errors:")
         for e in errs[:5]:
-            print(f"  {e.status_code} {e.error}")
+            print(f"  {e.error}")
 
     # Plot: 2x3 grid
     fig, axes = plt.subplots(2, 3, figsize=(15, 8))
@@ -296,7 +264,7 @@ def run_load_with_metrics(num_requests: int,
     # Latency histogram (total)
     ax1 = axes[0, 1]
     if succ:
-        ax1.hist(total_ms, bins=30)
+        ax1.hist(total_ms, bins=30, color="#4e79a7")
     ax1.set_title("Total latency (ms)")
     ax1.set_xlabel("ms")
     ax1.set_ylabel("freq")
@@ -322,7 +290,7 @@ def run_load_with_metrics(num_requests: int,
         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")
+    ax_idle.set_ylabel("req/s")
 
     # Summary text
     ax3 = axes[1, 1]
@@ -330,36 +298,22 @@ def run_load_with_metrics(num_requests: int,
     text = (
         f"Total requests: {total_reqs}\n"
         f"Success: {succ}  Errors: {len(errs)}\n"
-        f"Avg total latency: {avg_total:.1f} ms\n"
+        f"Avg 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"
+        f"Total compute time: {total_compute_time_ms/1000.0:.2f} s"
     )
     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")
+    # Latency CDF (total_ms)
+    ax_cdf = axes[1, 2]
+    if succ:
+        x = np.sort(total_ms)
+        y = np.linspace(0, 1, len(x), endpoint=True)
+        ax_cdf.plot(x, y)
+    ax_cdf.set_title("Latency CDF")
+    ax_cdf.set_xlabel("ms")
+    ax_cdf.set_ylabel("fraction ≤ x")
 
     # Ensure unique output path and create directory if needed
     final_out_path = get_incremented_path(out_path)

workers/tgi/client.py

@@ -1,61 +1,125 @@
-from vastai import Serverless
-import asyncio
+import logging
+import sys
+import json
+from urllib.parse import urljoin
+import requests
+from utils.endpoint_util import Endpoint
+from utils.ssl import get_cert_file_path
 
-ENDPOINT_NAME = "my-tgi-endpoint" # Change this to match your endpoint name
-MAX_TOKENS = 1024
-PROMPT = "Think step by step: Tell me about the Python programming language."
+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 call_generate(client: Serverless) -> None:
-    endpoint = await client.get_endpoint(name=ENDPOINT_NAME)
 
-    payload = {
-        "inputs": PROMPT,
-        "parameters": {
-            "max_new_tokens": MAX_TOKENS,
-            "temperature": 0.7,
-            "return_full_text": False
-        }
+def call_generate(endpoint_group_name: str, api_key: str, server_url: str) -> None:
+    WORKER_ENDPOINT = "/generate"
+    COST = 100
+    route_payload = {
+        "endpoint": endpoint_group_name,
+        "api_key": api_key,
+        "cost": COST,
     }
-
-    resp = await endpoint.request("/generate", payload, cost=MAX_TOKENS)
-
-    print(resp["response"]["generated_text"])
-
-
-async def call_generate_stream(client: Serverless) -> None:
-    endpoint = await client.get_endpoint(name=ENDPOINT_NAME)
-
-    payload = {
-        "inputs": PROMPT,
-        "parameters": {
-            "max_new_tokens": MAX_TOKENS,
-            "temperature": 0.7,
-            "do_sample": True,
-            "return_full_text": False,
-        }
-    }
-
-    resp = await endpoint.request(
-        "/generate_stream",
-        payload,
-        cost=MAX_TOKENS,
-        stream=True,
+    response = requests.post(
+        urljoin(server_url, "/route/"),
+        json=route_payload,
+        timeout=4,
     )
-    stream = resp["response"]
+    response.raise_for_status()  # Raise an exception for bad status codes
+    message = response.json()
+    url = message["url"]
 
-    printed_answer = False
-    async for event in stream:
-        tok = (event.get("token") or {}).get("text")
-        if tok:
-            if not printed_answer:
-                printed_answer = True
-                print("Answer:\n", end="", flush=True)
-            print(tok, end="", flush=True)
+    auth_data = dict(
+        signature=message["signature"],
+        cost=message["cost"],
+        endpoint=message["endpoint"],
+        reqnum=message["reqnum"],
+        url=url,
+    )
+
+    payload = dict(inputs="tell me about cats", parameters=dict(max_new_tokens=500))
+    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,
+        verify=get_cert_file_path(),
+    )
+    response.raise_for_status()
+    res = response.json()
+    print(res)
+
+
+def call_generate_stream(
+    endpoint_group_name: str, api_key: str, server_url: str
+) -> None:
+    WORKER_ENDPOINT = "/generate_stream"
+    COST = 100
+    route_payload = {
+        "endpoint": endpoint_group_name,
+        "api_key": api_key,
+        "cost": COST,
+    }
+    response = requests.post(
+        urljoin(server_url, "/route/"),
+        json=route_payload,
+        timeout=4,
+    )
+    response.raise_for_status()  # Raise an exception for bad status codes
+    message = response.json()
+    url = message["url"]
+    print(f"url: {url}")
+    auth_data = dict(
+        signature=message["signature"],
+        cost=message["cost"],
+        endpoint=message["endpoint"],
+        reqnum=message["reqnum"],
+        url=message["url"],
+    )
+    payload = dict(inputs="tell me about dogs", parameters=dict(max_new_tokens=500))
+    req_data = dict(payload=payload, auth_data=auth_data)
+    url = urljoin(url, WORKER_ENDPOINT)
+    response = requests.post(url, json=req_data, stream=True)
+    response.raise_for_status()  # Raise an exception for bad status codes
+    for line in response.iter_lines():
+        payload = line.decode().lstrip("data:").rstrip()
+        if payload:
+            try:
+                data = json.loads(payload)
+                print(data["token"]["text"], end="")
+                sys.stdout.flush()
+            except (json.JSONDecodeError, KeyError) as e:
+                log.warning(f"Failed to parse streaming response: {e}")
+                continue
+    print()
 
-async def main():
-    async with Serverless() as client:
-        await call_generate(client)
-        await call_generate_stream(client)
 
 if __name__ == "__main__":
-    asyncio.run(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:
+        try:
+            call_generate(
+                api_key=endpoint_api_key,
+                endpoint_group_name=args.endpoint_group_name,
+                server_url=args.server_url,
+            )
+            call_generate_stream(
+                api_key=endpoint_api_key,
+                endpoint_group_name=args.endpoint_group_name,
+                server_url=args.server_url,
+            )
+        except Exception as e:
+            log.error(f"Error during API call: {e}")
+    else:
+        log.error(f"Failed to get API key for endpoint {args.endpoint_group_name} ")