try allow parallel requests

Use mtoken

lib/backend.py

@@ -30,7 +30,7 @@ from lib.data_types import (
     BenchmarkResult
 )
 
-VERSION = "0.1.0"
+VERSION = "0.2.0"
 
 MSG_HISTORY_LEN = 100
 log = logging.getLogger(__file__)
@@ -66,10 +66,17 @@ class Backend:
     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
@@ -104,23 +111,19 @@ class Backend:
 
     #######################################Private#######################################
     def _fetch_pubkey(self):
-        command = ["curl", "-X", "GET", "https://run.vast.ai/pubkey/"]
+        report_addr = self.report_addr.rstrip("/")
-        result = subprocess.check_output(command, universal_newlines=True)
+        command = ["curl", "-X", "GET", f"{report_addr}/pubkey/"]
-        log.debug("public key:")
+        try:
-        log.debug(result)
+            result = subprocess.check_output(command, universal_newlines=True)
-        key = None
+            log.debug("public key:")
-        for _ in range(5):
+            log.debug(result)
-            try:
+            key = RSA.import_key(result)
-                key = RSA.import_key(result)
+            if key is not None:
-                break
+                return key
-            except ValueError as e:
+        except (ValueError , subprocess.CalledProcessError) as e:
-                log.debug(f"Error downloading key: {e}")
+            log.debug(f"Error downloading key: {e}")
-                time.sleep(15)
+        self.backend_errored("Failed to get autoscaler pubkey")
-        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,
@@ -315,10 +318,10 @@ 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()
+                    # payload = self.benchmark_handler.make_benchmark_payload()
-                    _ = await self.__call_api(
+                    # _ = await self.__call_api(
-                        handler=self.benchmark_handler, payload=payload
+                    #     handler=self.benchmark_handler, payload=payload
-                    )
+                    # )
                     return float(f.readline())
             except FileNotFoundError:
                 pass
@@ -393,7 +396,7 @@ 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

@@ -286,6 +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
@@ -142,12 +143,16 @@ 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 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,
             }
@@ -209,6 +214,7 @@ class Metrics:
         def compute_autoscaler_data() -> AutoScalerData:
             return AutoScalerData(
                 id=self.id,
+                mtoken=self.mtoken,
                 version=self.version,
                 loadtime=(loadtime_snapshot or 0.0), 
                 new_load=self.model_metrics.workload_processing,
@@ -228,17 +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()

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/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/server.py

@@ -89,7 +89,7 @@ class ComfyWorkflowHandler(EndpointHandler[ComfyWorkflowData]):
 backend = Backend(
     model_server_url=MODEL_SERVER_URL,
     model_log_file=os.environ["MODEL_LOG"],
-    allow_parallel_requests=False,
+    allow_parallel_requests=True,
     benchmark_handler=ComfyWorkflowHandler(
         benchmark_runs=3, benchmark_words=100
     ),

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": {