trying queue

vLLM Prefix caching, benchmark bug fix, test load script

lib/backend.py

@@ -5,7 +5,7 @@ import base64
 import subprocess
 import dataclasses
 import logging
-from asyncio import wait, sleep, gather, Semaphore, FIRST_COMPLETED, create_task
+from asyncio import wait, sleep, gather, Semaphore, FIRST_COMPLETED, create_task, get_running_loop
 from typing import Tuple, Awaitable, NoReturn, List, Union, Callable, Optional
 from functools import cached_property
 from distutils.util import strtobool
@@ -26,7 +26,8 @@ from lib.data_types import (
     LogAction,
     ApiPayload_T,
     JsonDataException,
-    RequestMetrics
+    RequestMetrics,
+    BenchmarkResult
 )
 
 VERSION = "0.1.0"
@@ -46,7 +47,7 @@ class Backend:
     This class is responsible for:
     1. Tailing logs and updating load time metrics
     2. Taking an EndpointHandler alongside incoming payload, preparing a json to be sent to the model, and
-    sending the request. It also updates metrics as it makes those requests.
+       sending the request. It also updates metrics as it makes those requests.
     3. Running a benchmark from an EndpointHandler
     """
 
@@ -73,6 +74,11 @@ class Backend:
         self._pubkey = self._fetch_pubkey()
         self.__start_healthcheck: bool = False
 
+        # NEW: FIFO queue + worker count
+        self.request_queue: "asyncio.Queue[tuple[EndpointHandler[ApiPayload_T], web.Request, asyncio.Future]]" = asyncio.Queue()
+        # If parallel allowed, let multiple workers drain the queue (order preserved by FIFO per worker; overall start order is FIFO).
+        self._num_workers: int = 1 if not self.allow_parallel_requests else int(os.environ.get("WORKERS", "4"))
+
     @property
     def pubkey(self) -> Optional[RSA.RsaKey]:
         if self._pubkey is None:
@@ -90,6 +96,22 @@ class Backend:
         timeout = ClientTimeout(total=None)
         return ClientSession(self.model_server_url, timeout=timeout, connector=connector)
 
+    async def _worker(self):
+        while True:
+            handler, request, fut = await self.request_queue.get()
+            try:
+                # Skip if already cancelled while waiting in the queue
+                if fut.cancelled():
+                    continue
+                res = await self.__process_enqueued_request(handler, request)
+                if not fut.cancelled():
+                    fut.set_result(res)
+            except Exception as e:
+                if not fut.cancelled():
+                    fut.set_exception(e)
+            finally:
+                self.request_queue.task_done()
+
     def create_handler(
         self,
         handler: EndpointHandler[ApiPayload_T],
@@ -126,7 +148,36 @@ class Backend:
         handler: EndpointHandler[ApiPayload_T],
         request: web.Request,
     ) -> Union[web.Response, web.StreamResponse]:
-        """use this function to forward requests to the model endpoint"""
+        """use this function to enqueue requests for FIFO processing"""
+        loop = get_running_loop()
+        fut: asyncio.Future = loop.create_future()
+
+        # If the client disconnects while waiting in the FIFO, cancel the future so the worker skips it
+        cancel_watch = create_task(request.wait_for_disconnection())
+        def _cancel_if_disconnected(_):
+            if not fut.done():
+                fut.cancel()
+        cancel_watch.add_done_callback(_cancel_if_disconnected)
+
+        try:
+            await self.request_queue.put((handler, request, fut))
+            return await fut
+        except asyncio.CancelledError:
+            # Propagate cancellation to ensure aiohttp doesn't expect a response body
+            raise
+        finally:
+            # Best-effort cleanup of the watcher
+            cancel_watch.cancel()
+
+    async def __process_enqueued_request(
+        self,
+        handler: EndpointHandler[ApiPayload_T],
+        request: web.Request,
+    ) -> Union[web.Response, web.StreamResponse]:
+        """
+        This contains the original __handle_request logic and is invoked by workers,
+        ensuring FIFO execution via asyncio.Queue.
+        """
         try:
             data = await request.json()
             auth_data, payload = handler.get_data_from_request(data)
@@ -134,8 +185,11 @@ class Backend:
             return web.json_response(data=e.message, status=422)
         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")
+        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()
@@ -176,6 +230,8 @@ class Backend:
         acquired = False
         try:
             self.metrics._request_start(request_metrics)
+
+            # Preserve existing semaphore behavior for serializing requests when requested
             if self.allow_parallel_requests is False:
                 log.debug(f"Waiting to aquire Sem for reqnum:{request_metrics.reqnum}")
                 await self.sem.acquire()
@@ -185,6 +241,7 @@ class Backend:
                 )
             else:
                 log.debug(f"Starting request for reqnum:{request_metrics.reqnum}")
+
             done, pending = await wait(
                 [
                     create_task(make_request()),
@@ -252,8 +309,14 @@ class Backend:
                 self.backend_errored(str(e))
 
     async def _start_tracking(self) -> None:
+        # Start the FIFO workers alongside existing loops
+        worker_tasks = tuple(self._worker() for _ in range(self._num_workers))
         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(),
+            self.metrics._send_delete_requests_loop(),
+            *worker_tasks,
         )
 
     def backend_errored(self, msg: str) -> None:
@@ -332,18 +395,26 @@ 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 +428,7 @@ 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,
                         ]
                     )

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
 
@@ -206,6 +206,17 @@ 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 +257,7 @@ 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:

lib/metrics.py

@@ -145,31 +145,56 @@ class Metrics:
     #######################################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
+                "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()
+            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
 
 

lib/test_utils.py

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

utils/endpoint_util.py

@@ -1,5 +1,6 @@
 import logging
-from typing import Any, Dict, Optional
+import time
+from typing import Any, Dict, Optional, Tuple
 
 import requests
 
@@ -16,6 +17,38 @@ class Endpoint:
     Utility class for handling endpoint operations.
     """
 
+    @staticmethod
+    def get_endpoint_info(
+        endpoint_name: str, account_api_key: str, instance: str
+    ) -> Optional[Dict[str, Any]]:
+        headers = {"Authorization": f"Bearer {account_api_key}"}
+        url = f"{Endpoint.get_server_url(instance)}?autoscaler_instance={instance}"
+        # Retry a few times to smooth over transient propagation/network delays
+        for attempt in range(4):
+            try:
+                response = requests.get(url, headers=headers, timeout=8)
+                if response.status_code != 200:
+                    # brief backoff and retry
+                    time.sleep(0.3 * (attempt + 1))
+                    continue
+                try:
+                    data = response.json()
+                except Exception:
+                    # JSON parse failed; backoff and retry
+                    time.sleep(0.3 * (attempt + 1))
+                    continue
+                result = data.get("results", []) if isinstance(data, dict) else []
+                endpoint = next(
+                    (item for item in result if item.get("endpoint_name") == endpoint_name),
+                    None,
+                )
+                if endpoint and endpoint.get("id") and endpoint.get("api_key"):
+                    return {"id": endpoint.get("id"), "api_key": endpoint.get("api_key")}
+            except Exception:
+                # network or other transient error; retry
+                time.sleep(0.3 * (attempt + 1))
+        return None
+
     @staticmethod
     def get_autoscaler_server_url(instance: str) -> str:
         endpoints = {
@@ -23,7 +56,10 @@ class Endpoint:
             "candidate": "run-candidate",
             "prod": "run",
         }
-        return f"https://{endpoints[instance]}.vast.ai/"
+        host = endpoints.get(instance)
+        if host:
+            return f"https://{host}.vast.ai/"
+        return "http://localhost:8080"
 
     @staticmethod
     def get_server_url(instance: str) -> str:
@@ -32,7 +68,8 @@ class Endpoint:
             "candidate": "candidate",
             "prod": "console",
         }
-        return f"https://{endpoints[instance]}.vast.ai/api/v0/endptjobs/"
+        host = endpoints.get(instance, "alpha")
+        return f"https://{host}.vast.ai/api/v0/endptjobs/"
 
     @staticmethod
     def get_endpoint_api_key(
@@ -55,6 +92,7 @@ class Endpoint:
             response = requests.get(
                 f"{Endpoint.get_server_url(instance)}?autoscaler_instance={instance}",
                 headers=headers,
+                timeout=8,
             )
 
             if response.status_code != 200:
@@ -64,14 +102,14 @@ class Endpoint:
 
             try:
                 data = response.json()
-            except requests.exceptions.JSONDecodeError as e:
+            except Exception as e:
                 log.debug(f"Failed to parse JSON response: {e}")
                 return None
 
             result = data.get("results", [])
 
             endpoint: Optional[Dict[str, Any]] = next(
-                (item for item in result if item["endpoint_name"] == endpoint_name),
+                (item for item in result if item.get("endpoint_name") == endpoint_name),
                 None,
             )
             if not endpoint:

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/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/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

@@ -1,8 +1,395 @@
-from lib.test_utils import test_load_cmd, test_args
+from lib.test_utils import test_args
+from utils.endpoint_util import Endpoint
+from utils.ssl import get_cert_file_path
+from lib.data_types import AuthData
 from .data_types.server import CompletionsData
-import os
 
-WORKER_ENDPOINT = "/v1/completions"
+import os
+import time
+import threading
+import requests
+from dataclasses import dataclass
+from collections import Counter
+from urllib.parse import urljoin, urlparse
+import re
+
+# Headless plotting
+import matplotlib
+matplotlib.use("Agg")
+import logging
+logging.getLogger("matplotlib.font_manager").setLevel(logging.WARNING)
+import matplotlib.pyplot as plt
+import numpy as np
+from concurrent.futures import ThreadPoolExecutor, wait, FIRST_COMPLETED
+from requests.adapters import HTTPAdapter
+
+def get_incremented_path(path: str) -> str:
+    base, ext = os.path.splitext(path)
+    if not os.path.exists(path):
+        return path
+    i = 1
+    while os.path.exists(f"{base}-{i}{ext}"):
+        i += 1
+    return f"{base}-{i}{ext}"
+
+WORKER_ENDPOINT = "/v1/completions" # This will return the full text output at once. Latency metrics reflect that (ie not measuring TTFT)
+
+@dataclass
+class ReqResult:
+    worker_url: str
+    route_ms: float
+    worker_ms: float
+    total_ms: float
+    ok: bool
+    error: str = ""
+    status_code: int = 0
+    t_start: float = 0.0
+    t_end: float = 0.0
+    workload: float = 0.0
+
+def do_one(endpoint_name: str,
+           endpoint_id: int,
+           endpoint_api_key: str,
+           server_url: str,
+           worker_endpoint: str,
+           payload,
+           results_list,
+           t0,
+           status_samples,
+           route_session,
+           worker_session):
+    try:
+        workload = payload.count_workload()
+        route_payload = {"endpoint": endpoint_name, "api_key": endpoint_api_key, "cost": workload}
+        headers = {"Authorization": f"Bearer {endpoint_api_key}"}
+        start = time.time()
+        r0 = route_session.post(urljoin(server_url, "/route/"), json=route_payload, headers=headers, timeout=4)
+        t_after_route = time.time()
+        if r0.status_code != 200:
+            results_list.append(ReqResult(worker_url="", 
+                                            route_ms=(t_after_route - start) * 1000.0, 
+                                            worker_ms=0.0, 
+                                            total_ms=(t_after_route - start) * 1000.0, 
+                                            ok=False, 
+                                            error=f"route error {r0.reason} {r0.text}",
+                                            status_code=r0.status_code,
+                                            t_start=start - t0, 
+                                            t_end=t_after_route - t0, 
+                                            workload=workload))
+            return
+        msg = r0.json()
+
+        # 1) Check if we got a worker back from route
+        worker_url = msg.get("url", "")
+        if not worker_url:
+            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())
+                idle = max(tot - loading - standby - err, 0)
+                status_samples.append((time.time() - t0, idle))
+
+        # 2) If we got a worker, send the request
+        if worker_url:
+            req = dict(payload=payload.__dict__, auth_data=AuthData.from_json_msg(msg).__dict__)
+            t_before_worker = time.time()
+            r1 = worker_session.post(
+                urljoin(worker_url, worker_endpoint),
+                json=req,
+                verify=get_cert_file_path(),
+                timeout=(4, 120),
+            )
+            t_after_worker = time.time()
+            if r1.status_code != 200:
+                results_list.append(ReqResult(worker_url=worker_url, 
+                                                route_ms=(t_after_route - start) * 1000.0, 
+                                                worker_ms=(t_after_worker - t_before_worker) * 1000.0,
+                                                total_ms=(t_after_worker - start) * 1000.0, 
+                                                ok=False,
+                                                error=f"worker inference error {r1.reason} {r1.text}",
+                                                status_code=r1.status_code,
+                                                t_start=start - t0, 
+                                                t_end=t_after_worker - t0, 
+                                                workload=workload))
+                return
+            # Success case
+            results_list.append(ReqResult(worker_url=worker_url, 
+                                            route_ms=(t_after_route - start) * 1000.0, 
+                                            worker_ms=(t_after_worker - t_before_worker) * 1000.0, 
+                                            total_ms=(t_after_worker - start) * 1000.0,
+                                            ok=True,
+                                            error="",
+                                            status_code=200,
+                                            t_start=start - t0, 
+                                            t_end=t_after_worker - t0, 
+                                            workload=workload))
+
+        # 3) If so, sample via /get_endpoint_workers/ for eligible (idle) worker tracking
+        if worker_url:
+            try:
+                r_status = route_session.post(
+                    urljoin(server_url, "/get_endpoint_workers/"),
+                    json={"id": endpoint_id},
+                    headers={"Authorization": f"Bearer {endpoint_api_key}"},
+                    timeout=3,
+                )
+                if r_status.status_code == 200:
+                    workers = r_status.json()
+                    idle = 0
+                    for w in workers:
+                        st = str(w.get("status", "")).lower()
+                        if (st in ("idle")):
+                            idle += 1
+                    status_samples.append((time.time() - t0, idle))
+            except Exception:
+                pass
+    except Exception as e:
+        t = time.time()
+        results_list.append(ReqResult(worker_url="", 
+                                        route_ms=0.0, 
+                                        worker_ms=0.0, 
+                                        total_ms=0.0, 
+                                        ok=False, 
+                                        error=f"unknown error {e}",
+                                        status_code=0,
+                                        t_start=t - t0, 
+                                        t_end=t - t0, 
+                                        workload=0.0))
+
+def run_load_with_metrics(num_requests: int,
+                          requests_per_second: float,
+                          endpoint_group_name: str,
+                          account_api_key: str,
+                          server_url: str,
+                          worker_endpoint: str,
+                          instance: str,
+                          out_path: str):
+
+    ep_info = Endpoint.get_endpoint_info(endpoint_name=endpoint_group_name,
+                                         account_api_key=account_api_key,
+                                         instance=instance)
+    if not ep_info or not ep_info.get("api_key") or not ep_info.get("id"):
+        print(f"Endpoint {endpoint_group_name} not found for API key")
+        return
+    endpoint_id = int(ep_info["id"])
+    endpoint_api_key = ep_info["api_key"]
+
+    t0 = time.time()
+    results = []
+    status_samples = []
+    max_concurrency = int(os.environ.get("MAX_CONCURRENCY", "8192"))
+    submit_queue_factor = 2  # cap queued tasks to reduce memory
+
+    # Shared HTTP sessions with connection pooling (persistent connections)
+    def make_session(pool_connections: int, pool_maxsize: int) -> requests.Session:
+        sess = requests.Session()
+        adapter = HTTPAdapter(pool_connections=pool_connections, pool_maxsize=pool_maxsize, max_retries=0)
+        sess.mount("https://", adapter)
+        sess.mount("http://", adapter)
+        return sess
+
+    # Router: mostly single host, small connection pool is sufficient
+    route_session = make_session(pool_connections=1, pool_maxsize=max_concurrency)
+    # Workers: many hosts; allow many pools and per-host concurrency up to max_concurrency
+    worker_session = make_session(pool_connections=64, pool_maxsize=max_concurrency // 8)
+
+    # Fire requests using a thread pool, scheduling at requested RPS
+    inflight = set()
+    with ThreadPoolExecutor(max_workers=max_concurrency) as executor:
+        for i in range(num_requests):
+            # Pace submissions to RPS
+            target_time = t0 + i / max(requests_per_second, 1e-9)
+            sleep_s = target_time - time.time()
+            if sleep_s > 0:
+                time.sleep(min(sleep_s, 0.5))  # sleep in chunks to stay responsive
+
+            payload = CompletionsData.for_test()
+            fut = executor.submit(
+                do_one,
+                endpoint_group_name,
+                endpoint_id,
+                endpoint_api_key,
+                server_url,
+                worker_endpoint,
+                payload,
+                results,
+                t0,
+                status_samples,
+                route_session,
+                worker_session,
+            )
+            inflight.add(fut)
+            # Prevent unbounded queue growth
+            if len(inflight) >= max_concurrency * submit_queue_factor:
+                done, not_done = wait(inflight, return_when=FIRST_COMPLETED)
+                inflight = not_done
+        # Wait for all outstanding tasks
+        if inflight:
+            wait(inflight)
+    # Close sessions
+    try:
+        route_session.close()
+    finally:
+        worker_session.close()
+
+    # Aggregate results
+    oks = [r for r in results if r.ok]
+    errs = [r for r in results if not r.ok]
+    total_reqs = len(results)
+    succ = len(oks)
+
+    total_ms = np.array([r.total_ms for r in oks]) if succ else np.array([])
+    worker_ms = np.array([r.worker_ms for r in oks]) if succ else np.array([])
+    route_ms = np.array([r.route_ms for r in oks]) if succ else np.array([])
+
+    avg_total = float(np.mean(total_ms)) if succ else 0.0
+    avg_worker = float(np.mean(worker_ms)) if succ else 0.0
+    avg_route = float(np.mean(route_ms)) if succ else 0.0
+    p50_total, p95_total = (float(np.percentile(total_ms, 50)), float(np.percentile(total_ms, 95))) if succ else (0.0, 0.0)
+
+    # Distribution over workers (by host:port)
+    hosts = [urlparse(r.worker_url).netloc for r in oks if r.worker_url]
+    dist = Counter(hosts)
+
+    # Idle over time (mode per second)
+    idle_ts, idle_vals = [], []
+    if status_samples:
+        buckets = {}
+        for ts, idle in status_samples:
+            k = int(ts)
+            buckets.setdefault(k, []).append(idle)
+        keys = sorted(buckets.keys())
+        idle_ts = keys
+        # Use the most frequent sampled value per second (mode) to keep integer counts
+        idle_vals = []
+        for k in keys:
+            vals_k = [int(v) for v in buckets[k]]
+            if vals_k:
+                cnt = Counter(vals_k)
+                idle_vals.append(cnt.most_common(1)[0][0])
+            else:
+                idle_vals.append(0)
+
+    print(f"\nResults: total={total_reqs} success={succ} errors={len(errs)}")
+    print(f"Avg latency (ms): {avg_total:.1f}  p50: {p50_total:.1f}  p95: {p95_total:.1f}")
+    print(f"Avg route latency (ms): {avg_route:.1f}  Avg worker latency (ms): {avg_worker:.1f}")
+    if errs:
+        print("Sample errors:")
+        for e in errs[:5]:
+            print(f"  {e.status_code} {e.error}")
+
+    # Plot: 2x3 grid
+    fig, axes = plt.subplots(2, 3, figsize=(15, 8))
+    fig.suptitle(f"Load test: {endpoint_group_name}  n={total_reqs}, rps={requests_per_second}, success={succ}")
+
+    # Dist per worker
+    ax0 = axes[0, 0]
+    if dist:
+        items = sorted(dist.items(), key=lambda kv: kv[1], reverse=True)
+        labels, counts = zip(*items)
+        ax0.bar(range(len(labels)), counts)
+        ax0.set_xticks(range(len(labels)))
+        ax0.set_xticklabels(labels, rotation=45, ha="right", fontsize=8)
+    ax0.set_title("Request distribution over workers")
+    ax0.set_ylabel("count")
+
+    # Latency histogram (total)
+    ax1 = axes[0, 1]
+    if succ:
+        ax1.hist(total_ms, bins=30)
+    ax1.set_title("Total latency (ms)")
+    ax1.set_xlabel("ms")
+    ax1.set_ylabel("freq")
+
+    # Eligible workers over time
+    ax_idle = axes[0, 2]
+    if idle_ts:
+        ax_idle.plot(idle_ts, idle_vals, "-o", ms=3)
+    ax_idle.set_title("Eligible workers over time")
+    ax_idle.set_xlabel("time (s)")
+    ax_idle.set_ylabel("eligible count")
+
+    # Throughput over time (completions/sec)
+    ax_idle = axes[1, 0]
+    ax_idle.clear()
+    if succ:
+        per_sec = {}
+        for r in oks:
+            s = int(r.t_end)
+            per_sec[s] = per_sec.get(s, 0) + 1
+        ts = sorted(per_sec.keys())
+        vals = [per_sec[t] for t in ts]
+        ax_idle.plot(ts, vals, "-o", ms=3)
+    ax_idle.set_title("Completions per second")
+    ax_idle.set_xlabel("time (s)")
+    ax_idle.set_ylabel("completions / sec")
+
+    # Summary text
+    ax3 = axes[1, 1]
+    ax3.axis("off")
+    text = (
+        f"Total requests: {total_reqs}\n"
+        f"Success: {succ}  Errors: {len(errs)}\n"
+        f"Avg total latency: {avg_total:.1f} ms\n"
+        f"p50: {p50_total:.1f} ms  p95: {p95_total:.1f} ms\n"
+        f"Avg route latency: {avg_route:.1f} ms\n"  
+        f"Avg worker latency: {avg_worker:.1f} ms\n"
+        f"300 errors: {len([r for r in errs if r.status_code >= 300 and r.status_code < 400])}\n"
+        f"429 errors: {len([r for r in errs if r.status_code == 429])}\n"
+        f"500 errors: {len([r for r in errs if r.status_code >= 500])}\n"
+        f"Other errors: {len([r for r in errs if r.status_code not in [300, 429, 500]])}\n"
+    )
+    ax3.set_title("Summary")
+    ax3.text(0.02, 0.98, text, va="top", ha="left", fontsize=11, transform=ax3.transAxes)
+
+    # Error count over time
+    ax_errors = axes[1, 2]
+    all_end_times = [int(r.t_end) for r in results if r.t_end > 0]
+    if all_end_times:
+        min_second = min(all_end_times)
+        max_second = max(all_end_times)
+        # Count errors per second
+        errors_per_second = {}
+        for result in errs:
+            second = int(result.t_end)
+            errors_per_second[second] = errors_per_second.get(second, 0) + 1
+        # Create complete timeline including zeros
+        time_seconds = list(range(min_second, max_second + 1))
+        error_counts = [errors_per_second.get(sec, 0) for sec in time_seconds]
+        ax_errors.plot(time_seconds, error_counts, "-o", ms=3)
+    ax_errors.set_title("Errors per second")
+    ax_errors.set_xlabel("time (s)")
+    ax_errors.set_ylabel("errors / sec")
+
+    # Ensure unique output path and create directory if needed
+    final_out_path = get_incremented_path(out_path)
+    out_dir = os.path.dirname(final_out_path)
+    if out_dir:
+        os.makedirs(out_dir, exist_ok=True)
+
+    plt.tight_layout(rect=[0, 0, 1, 0.96])
+    plt.savefig(final_out_path, dpi=120)
+    print(f"Saved report to: {final_out_path}")
+
+    # Per-worker latency boxplot (top 12 by volume)
+    groups = {}
+    for r in oks:
+        host = urlparse(r.worker_url).netloc
+        groups.setdefault(host, []).append(r.total_ms)
+    items = sorted(groups.items(), key=lambda kv: len(kv[1]), reverse=True)[:12]
+    if items:
+        labels, data = zip(*items)
+        fig2, axb = plt.subplots(1, 1, figsize=(12, 5))
+        axb.boxplot(data, showfliers=False)
+        axb.set_xticklabels(labels, rotation=45, ha="right", fontsize=8)
+        axb.set_title("Per-worker latency (ms)")
+        axb.set_ylabel("ms")
+        plt.tight_layout()
+        extra_out = get_incremented_path(os.path.splitext(out_path)[0] + "-workers.png")
+        plt.savefig(extra_out, dpi=120)
+        fig2.tight_layout()
+        fig2.savefig(extra_out, dpi=120)
+        print(f"Saved worker latency plot to: {extra_out}")
 
 if __name__ == "__main__":
     # Check if MODEL_NAME environment variable is set
@@ -16,13 +403,32 @@ if __name__ == "__main__":
         help="Model to use for completions request (required if MODEL_NAME env var not set)",
     )
 
-    # Parse known args to get model early, before test_load_cmd adds its args
+    # Parse known args to get model early, before adding load args
     known_args, _ = test_args.parse_known_args()
-
-    # Set environment variable if model was provided
     if hasattr(known_args, "model") and known_args.model:
         os.environ["MODEL_NAME"] = known_args.model
         print(f"Set MODEL_NAME environment variable to: {known_args.model}")
 
-    # Now call test_load_cmd normally - it will add its own args and re-parse
-    test_load_cmd(CompletionsData, WORKER_ENDPOINT, arg_parser=test_args)
+    # Load test args
+    test_args.add_argument("-n", dest="num_requests", type=int, required=True, help="total number of requests")
+    test_args.add_argument("-rps", dest="requests_per_second", type=float, required=True, help="requests per second")
+    test_args.add_argument("--out", dest="out_path", type=str, default="load_test_report.png", help="path to save the report image")
+    args = test_args.parse_args()
+
+    server_url = {
+        "prod": "https://run.vast.ai",
+        "alpha": "https://run-alpha.vast.ai",
+        "candidate": "https://run-candidate.vast.ai",
+        "local": "http://localhost:8080"
+    }.get(args.instance, "http://localhost:8080")
+
+    run_load_with_metrics(
+        num_requests=args.num_requests,
+        requests_per_second=args.requests_per_second,
+        endpoint_group_name=args.endpoint_group_name,
+        account_api_key=args.api_key,
+        server_url=server_url,
+        worker_endpoint=WORKER_ENDPOINT,
+        instance=args.instance,
+        out_path=args.out_path,
+    )