Added caller for REPORT_ADDR to backend.py

Redis subscriber queue patch

lib/backend.py

@@ -66,6 +66,9 @@ 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")
+    )
 
     def __post_init__(self):
         self.metrics = Metrics()
@@ -104,23 +107,19 @@ 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,
@@ -286,7 +285,7 @@ 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(
@@ -296,7 +295,7 @@ 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:]

lib/data_types.py

@@ -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

lib/metrics.py

@@ -145,14 +145,15 @@ class Metrics:
     #######################################Private#######################################
 
     async def __send_delete_requests_and_reset(self):
-
+        async def post(report_addr: str, idxs: list[int], success_flag: bool) -> bool:
-        async def send_data(report_addr: str, success: bool) -> bool:
             data = {
                 "worker_id": self.id,
-                "request_idxs": [r.request_idx for r in self.model_metrics.requests_deleting if r.success == success],
+                "request_idxs": idxs,
-                "success": success
+                "success": success_flag,
             }
-            log.debug(f"Deleting requests that {'succeeded' if success else 'failed'}: {data['request_idxs']}")
+            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:
@@ -162,26 +163,54 @@ class Metrics:
                         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)
+            # TODO: Add a Redis subscriber queue for delete_requests
-            if success is True:
+            if report_addr == "https://cloud.vast.ai/api/v0":
-                self.model_metrics.requests_deleting.clear()
+                # 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,
                 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,
@@ -229,11 +258,15 @@ class Metrics:
 
         self.system_metrics.update_disk_usage()
 
+        sent = False
         for report_addr in self.report_addr:
-            success = await send_data(report_addr)
+            if await send_data(report_addr):
-            if success is True:
+                sent = True
                 break
-        self.update_pending = False
+
-        self.model_metrics.reset()
+        if sent:
-        self.system_metrics.reset()
+            # clear the one-shot loadtime only if we actually sent *this* value
-        self.last_metric_update = time.time()
+            self.system_metrics.reset(expected=loadtime_snapshot)
+            self.update_pending = False
+            self.model_metrics.reset()
+            self.last_metric_update = time.time()

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
 
-
+log = logging.getLogger(__file__)
-with open("workers/comfyui/misc/test_prompts.txt", "r") as f:
-    test_prompts = f.readlines()
 
 def count_workload() -> float:
     # Always 100.0 where there is a single instance of ComfyUI handling requests
@@ -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
 ]