catch the case where all benchmarks fail (sets error)

lib/backend.py

@@ -26,7 +26,8 @@ from lib.data_types import (
     LogAction,
     ApiPayload_T,
     JsonDataException,
-    RequestMetrics
+    RequestMetrics,
+    BenchmarkResult
 )
 
 VERSION = "0.1.0"
@@ -332,18 +333,26 @@ class Backend:
 
             for run in range(1, self.benchmark_handler.benchmark_runs + 1):
                 start = time.time()
-                tasks = []
+                benchmark_requests = []
-                total_workload = 0
 
-                for _ in range(concurrent_requests):
+                for i in range(concurrent_requests):
                     payload = self.benchmark_handler.make_benchmark_payload()
-                    total_workload += payload.count_workload()
+                    workload = payload.count_workload()
-                    tasks.append(
+                    task = self.__call_api(handler=self.benchmark_handler, payload=payload)
-                        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 +366,7 @@ 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

@@ -152,11 +152,13 @@ class Metrics:
                 "request_idxs": [r.request_idx for r in self.model_metrics.requests_deleting if r.success == success],
                 "success": success
             }
+            log.debug(f"Deleting requests that {'succeeded' if success 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:

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())