refactor, handle zombie tasks

Bump pyworker version

lib/backend.py

@@ -9,6 +9,7 @@ from asyncio import wait, sleep, gather, Semaphore, FIRST_COMPLETED, create_task
 from typing import Tuple, Awaitable, NoReturn, List, Union, Callable, Optional
 from functools import cached_property
 from distutils.util import strtobool
+from collections import deque
 
 from anyio import open_file
 from aiohttp import web, ClientResponse, ClientSession, ClientConnectorError, ClientTimeout, TCPConnector
@@ -26,10 +27,11 @@ from lib.data_types import (
     LogAction,
     ApiPayload_T,
     JsonDataException,
-    RequestMetrics
+    RequestMetrics,
+    BenchmarkResult
 )
 
-VERSION = "0.1.0"
+VERSION = "0.2.1"
 
 MSG_HISTORY_LEN = 100
 log = logging.getLogger(__file__)
@@ -62,13 +64,21 @@ class Backend:
     version = VERSION
     msg_history = []
     sem: Semaphore = dataclasses.field(default_factory=Semaphore)
+    queue: deque = dataclasses.field(default_factory=deque, repr=False)
     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
@@ -103,23 +113,19 @@ class Backend:
 
     #######################################Private#######################################
     def _fetch_pubkey(self):
-        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
+        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")
+       
 
     async def __handle_request(
         self,
@@ -137,11 +143,26 @@ class Backend:
         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:
+
+        def advance_queue_after_completion(event: asyncio.Event):
+            """Pop current head and wake next waiter, if any."""
+            # If this event is current head, wake next waiter
+            if self.queue and self.queue[0] is event:
+                self.queue.popleft()
+                if self.queue:
+                    self.queue[0].set()
+            else:
+                # Else, remove it from the queue
+                try:
+                    self.queue.remove(event)
+                except ValueError:
+                    pass
+
+        async def cancel_api_call_if_disconnected() -> None:
             await request.wait_for_disconnection()
-            log.debug(f"request with reqnum: {request_metrics.reqnum} was canceled")
+            log.debug(f"Request with reqnum: {request_metrics.reqnum} was canceled")
             self.metrics._request_canceled(request_metrics)
-            raise asyncio.CancelledError
+            return
 
         async def make_request() -> Union[web.Response, web.StreamResponse]:
             try:
@@ -158,7 +179,9 @@ class Backend:
                 res = await handler.generate_client_response(request, response)
                 self.metrics._request_success(request_metrics)
                 return res
-            except requests.exceptions.RequestException as e:
+            except asyncio.CancelledError:
+                raise
+            except Exception as e:
                 log.debug(f"[backend] Request error: {e}")
                 self.metrics._request_errored(request_metrics)
                 return web.Response(status=500)
@@ -173,46 +196,87 @@ class Backend:
             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()),
-                    create_task(cancel_api_call_if_disconnected()),
-                ],
-                return_when=FIRST_COMPLETED,
-            )
-            for t in pending:
-                t.cancel()
-            await asyncio.gather(*pending, return_exceptions=True)
+        disconnect_task = create_task(cancel_api_call_if_disconnected())
+        next_request_task = None
+        work_task = None
+        event = asyncio.Event() # Used in finally block, so initialize here
 
-            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)
+        self.metrics._request_start(request_metrics)
+
+        try:
+            if self.allow_parallel_requests:
+                log.debug(f"Starting request for reqnum:{request_metrics.reqnum}")
+                work_task = create_task(make_request())
+                done, pending = await wait([work_task, disconnect_task], return_when=FIRST_COMPLETED)
+
+                for t in pending:
+                    t.cancel()
+                await asyncio.gather(*pending, return_exceptions=True)
+
+                if disconnect_task in done:
+                    return web.Response(status=499)
+
+                # otherwise work_task completed
+                return await work_task
+            
+            # FIFO-queue branch
+            else:
+                # Insert a Event into the queue for this request
+                # Event.set() == our request is up next
+                self.queue.append(event)
+                if self.queue and self.queue[0] is event:
+                    event.set()
+
+                # Race between our request being next and request being cancelled
+                next_request_task = create_task(event.wait())
+                first_done, first_pending = await wait(
+                    [next_request_task, disconnect_task], return_when=FIRST_COMPLETED
+                )
+
+                # If the disconnect task wins the race
+                if disconnect_task in first_done:
+                    # Clean up the next_request_task, then exit
+                    for t in first_pending:
+                        t.cancel()
+                    await asyncio.gather(*first_pending, return_exceptions=True)
+                    return web.Response(status=499)
+
+                # We are the next-up request in the queue
+                log.debug(f"Starting work on request {request_metrics.reqnum}...")
+
+                # Race the backend API call with the disconnect task
+                work_task = create_task(make_request())
+
+                done, pending = await wait([work_task, disconnect_task], return_when=FIRST_COMPLETED)
+                for t in pending:
+                    t.cancel()
+                await asyncio.gather(*pending, return_exceptions=True)
+
+                if disconnect_task in done:
+                    return web.Response(status=499)
+
+                # otherwise work_task completed
+                return await work_task
+            
         except asyncio.CancelledError:
-            # Client is gone. Do not write a response; just unwind.
-            return web.Response(status=499) 
+            return web.Response(status=499)
+        
         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()
+            if not self.allow_parallel_requests:
+                advance_queue_after_completion(event)
+
             self.metrics._request_end(request_metrics)
+            cleanup_tasks = [t for t in (next_request_task, work_task, disconnect_task) if t]
+            for t in cleanup_tasks:
+                if not t.done():
+                    t.cancel()
+            if cleanup_tasks:
+                await asyncio.gather(*cleanup_tasks, return_exceptions=True)
+
 
     @cached_property  
     def healthcheck_session(self):
@@ -285,7 +349,7 @@ class Backend:
         message = {
             key: value
             for (key, value) in (dataclasses.asdict(auth_data).items())
-            if key != "signature"
+            if key != "signature" and key != "__request_id"
         }
         if auth_data.reqnum < (self.reqnum - MSG_HISTORY_LEN):
             log.debug(
@@ -295,7 +359,7 @@ 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), auth_data.signature):
+        elif verify_signature(json.dumps(message, indent=4, sort_keys=True), 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:]
@@ -314,10 +378,10 @@ 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
@@ -332,18 +396,26 @@ class Backend:
 
             for run in range(1, self.benchmark_handler.benchmark_runs + 1):
                 start = time.time()
-                tasks = []
-                total_workload = 0
+                benchmark_requests = []
 
-                for _ in range(concurrent_requests):
+                for i in range(concurrent_requests):
                     payload = self.benchmark_handler.make_benchmark_payload()
-                    total_workload += payload.count_workload()
-                    tasks.append(
-                        self.__call_api(handler=self.benchmark_handler, payload=payload)
+                    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)
                     )
 
-                responses = await gather(*tasks)
+                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)
                 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
@@ -357,7 +429,7 @@ 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: {len([r for r in responses if r.status == 200])}",
+                            f"Successful responses: {successful_responses}/{concurrent_requests}",
                             "#" * 60,
                         ]
                     )
@@ -384,7 +456,7 @@ 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

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
+from typing import Dict, Any, Union, Tuple, Optional, Set, TypeVar, Generic, Type, Awaitable
 from aiohttp import web, ClientResponse
 import inspect
 
@@ -65,12 +65,12 @@ class ApiPayload(ABC):
 class AuthData:
     """data used to authenticate requester"""
 
-    signature: str
     cost: str
     endpoint: str
     reqnum: int
-    url: str
     request_idx: int
+    signature: str
+    url: str
 
     @classmethod
     def from_json_msg(cls, json_msg: Dict[str, Any]):
@@ -190,11 +190,12 @@ class SystemMetrics:
         self.additional_disk_usage = disk_usage - self.last_disk_usage
         self.last_disk_usage = disk_usage
 
-    def reset(self):
+    def reset(self, expected: float | None) -> None:
         # 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
-        self.model_loading_time = None
+        if self.model_loading_time == expected:
+            self.model_loading_time = None
 
 
 @dataclass
@@ -206,6 +207,17 @@ class RequestMetrics:
     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"""
@@ -246,7 +258,7 @@ class ModelMetrics:
     def wait_time(self) -> float:
         if (len(self.requests_working) == 0):
             return 0.0
-        return sum([request.workload for request in self.requests_working.values()]) / self.max_throughput
+        return sum([request.workload for request in self.requests_working.values()]) / max(self.max_throughput, 0.00001)
     
     @property
     def cur_load(self) -> float:
@@ -274,6 +286,7 @@ class AutoScalerData:
     """Data that is reported to autoscaler"""
 
     id: int
+    mtoken: str
     version: str
     loadtime: float
     cur_load: float

lib/metrics.py

@@ -28,6 +28,7 @@ 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
@@ -142,44 +143,80 @@ class Metrics:
     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 send_data(report_addr: str, success: bool) -> bool:
+        async def post(report_addr: str, idxs: list[int], success_flag: bool) -> bool:
             data = {
                 "worker_id": self.id,
-                "request_idxs": [r.request_idx for r in self.model_metrics.requests_deleting if r.success == success],
-                "success": success
+                "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(f"delete_requests timed out")
+                    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
+
+        # 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:
-            success = await send_data(report_addr, success=True) and await send_data(report_addr, success=False)
-            if success is True:
-                self.model_metrics.requests_deleting.clear()
+            # 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(
                 id=self.id,
+                mtoken=self.mtoken,
                 version=self.version,
-                loadtime=(self.system_metrics.model_loading_time or 0.0),
+                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,
@@ -197,17 +234,25 @@ class Metrics:
 
         async def send_data(report_addr: str) -> bool:
             data = compute_autoscaler_data()
-            full_path = report_addr.rstrip("/") + "/worker_status/"
+            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"))
             log.debug(
                 "\n".join(
                     [
                         "#" * 60,
                         f"sending data to autoscaler",
-                        f"{json.dumps((asdict(data)), indent=2)}",
+                        f"{json.dumps(log_data, indent=2)}",
                         "#" * 60,
                     ]
                 )
             )
+
+            full_path = report_addr.rstrip("/") + "/worker_status/"
             for attempt in range(1, 4):
                 try:
                     session = await self.http()
@@ -227,11 +272,15 @@ class Metrics:
 
         self.system_metrics.update_disk_usage()
 
+        sent = False
         for report_addr in self.report_addr:
-            success = await send_data(report_addr)
-            if success is True:
+            if await send_data(report_addr):
+                sent = True
                 break
-        self.update_pending = False
-        self.model_metrics.reset()
-        self.system_metrics.reset()
-        self.last_metric_update = time.time()
+
+        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()

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="${REPORT_ADDR:-https://cloud.vast.ai/api/v0,https://run.vast.ai}"
+REPORT_ADDR="${REPORT_ADDR:-https://run.vast.ai}"
 USE_SSL="${USE_SSL:-true}"
 WORKER_PORT="${WORKER_PORT:-3000}"
 mkdir -p "$WORKSPACE_DIR"

workers/comfyui-json/README.md

@@ -12,9 +12,21 @@ A docker image is provided but you may use any if the above requirements are met
 
 ## Benchmarking
 
-A simple image generation benchmark runs when each worker initializes to validate GPU performance and identify underperforming machines.
+### Custom Benchmark Workflows
 
-The benchmark uses Stable Diffusion v1.5 with ComfyUI's default text-to-image workflow. Configure the benchmark complexity and duration using these variables:
+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:
 
 | Environment Variable | Default Value | Description |
 | -------------------- | ------------- | ----------- |
@@ -24,7 +36,7 @@ The benchmark uses Stable Diffusion v1.5 with ComfyUI's default text-to-image wo
 
 Each benchmark run uses a random prompt from `misc/test_prompts.txt` and a random seed to ensure consistent GPU load patterns.
 
-### Calibrating Benchmark Duration
+#### Calibrating Fallback 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

@@ -98,6 +98,7 @@ def call_text2image_workflow(
         endpoint=route_response["endpoint"],
         reqnum=route_response["reqnum"],
         url=route_response["url"],
+        request_idx=route_response["request_idx"],
     )
     
     # Build the payload for the worker request

workers/comfyui-json/data_types.py

@@ -5,12 +5,13 @@ 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
 
-
-with open("workers/comfyui/misc/test_prompts.txt", "r") as f:
-    test_prompts = f.readlines()
+log = logging.getLogger(__file__)
 
 def count_workload() -> float:
     # Always 100.0 where there is a single instance of ComfyUI handling requests
@@ -24,9 +25,32 @@ class ComfyWorkflowData(ApiPayload):
     @classmethod
     def for_test(cls):
         """
-        Use the variables available to simulate workflows of the required running time
+        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
         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

@@ -0,0 +1,107 @@
+{
+    "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

@@ -19,6 +19,7 @@ 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
 ]
 
 

workers/comfyui/client.py

@@ -82,6 +82,7 @@ def call_custom_workflow_for_sd3(
         endpoint=message["endpoint"],
         reqnum=message["reqnum"],
         url=message["url"],
+        request_idx=message["request_idx"],
     )
     workflow = {
         "3": {

workers/openai/data_types/server.py

@@ -119,14 +119,25 @@ class GenericHandler(EndpointHandler[GenericData], ABC):
 class CompletionsData(GenericData):
     @classmethod
     def for_test(cls) -> "CompletionsData":
-        prompt = " ".join(random.choices(WORD_LIST, k=int(250)))
+        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)))
         model = os.environ.get("MODEL_NAME")
         if not model:
             raise ValueError("MODEL_NAME environment variable not set")
 
         test_input = {
             "model": model,
-            "prompt": prompt,
+            "prompt": f"{system_prompt}\n\n{unique_question}",
             "temperature": 0.7,
             "max_tokens": 500,
         }
@@ -153,7 +164,18 @@ class ChatCompletionsData(GenericData):
 
     @classmethod
     def for_test(cls) -> "ChatCompletionsData":
-        prompt = " ".join(random.choices(WORD_LIST, k=int(250)))
+        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)))
         model = os.environ.get("MODEL_NAME")
         if not model:
             raise ValueError("MODEL_NAME environment variable not set")
@@ -161,7 +183,10 @@ class ChatCompletionsData(GenericData):
         # Chat completions use messages format instead of prompt
         test_input = {
             "model": model,
-            "messages": [{"role": "user", "content": prompt}],
+            "messages": [
+                {"role": "system", "content": system_prompt},  # Shared prefix
+                {"role": "user", "content": unique_question}   # Unique per request
+            ],
             "temperature": 0.7,
             "max_tokens": 500,
         }

workers/openai/test_load.py

@@ -82,6 +82,7 @@ def do_one(endpoint_name: str,
         # 1) Check if we got a worker back from route
         worker_url = msg.get("url", "")
         if not worker_url:
+            status = msg.get("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())