returned to worker-sdk

requirements.txt

@@ -1,6 +1,4 @@
-aiohttp==3.10.1
+aiohttp[speedups]==3.10.1
-aiodns~=3.6.0
-pycares~=4.11.0
 anyio~=4.4
 lib~=4.0
 nltk~=3.9
@@ -10,4 +8,4 @@ Requests~=2.32
 transformers~=4.52
 utils==1.0.*
 hf_transfer>=0.1.9
-vastai-sdk>=0.2.0
+git+https://github.com/vast-ai/vast-sdk.git@worker-sdk

start_server.sh

@@ -22,49 +22,10 @@ function echo_var(){
     echo "$1: ${!1}"
 }
 
-function report_error_and_exit(){
+[ -z "$BACKEND" ] && echo "BACKEND must be set!" && exit 1
-    local error_msg="$1"
+[ -z "$MODEL_LOG" ] && echo "MODEL_LOG must be set!" && exit 1
-    echo "ERROR: $error_msg"
+[ -z "$HF_TOKEN" ] && echo "HF_TOKEN must be set!" && exit 1
-    
+[ "$BACKEND" = "comfyui" ] && [ -z "$COMFY_MODEL" ] && echo "For comfyui backends, COMFY_MODEL must be set!" && exit 1
-    # Report error to autoscaler
-    MTOKEN="${MASTER_TOKEN:-}"
-    VERSION="${PYWORKER_VERSION:-0}"
-    
-    IFS=',' read -r -a REPORT_ADDRS <<< "${REPORT_ADDR}"
-    for addr in "${REPORT_ADDRS[@]}"; do
-        curl -sS -X POST -H 'Content-Type: application/json' \
-            -d "$(cat <<JSON
-{
-  "id": ${CONTAINER_ID:-0},
-  "mtoken": "${MTOKEN}",
-  "version": "${VERSION}",
-  "loadtime": 0,
-  "new_load": 0,
-  "cur_load": 0,
-  "rej_load": 0,
-  "max_perf": 0,
-  "cur_perf": 0,
-  "error_msg": "${error_msg}",
-  "num_requests_working": 0,
-  "num_requests_recieved": 0,
-  "additional_disk_usage": 0,
-  "working_request_idxs": [],
-  "cur_capacity": 0,
-  "max_capacity": 0,
-  "url": "${URL:-}"
-}
-JSON
-)" "${addr%/}/worker_status/" || true
-    done
-    
-    exit 1
-}
-
-[ -z "$BACKEND" ] && report_error_and_exit "BACKEND must be set!"
-[ -z "$MODEL_LOG" ] && report_error_and_exit "MODEL_LOG must be set!"
-[ -z "$HF_TOKEN" ] && report_error_and_exit "HF_TOKEN must be set!"
-[ -z "$CONTAINER_ID" ] && report_error_and_exit "CONTAINER_ID must be set!"
-[ "$BACKEND" = "comfyui" ] && [ -z "$COMFY_MODEL" ] && report_error_and_exit "For comfyui backends, COMFY_MODEL must be set!"
 
 
 echo "start_server.sh"
@@ -84,86 +45,51 @@ echo_var MODEL_LOG
 # from the run prior to reboot. past logs are saved in $MODEL_LOG.old for debugging only
 if [ -e "$MODEL_LOG" ]; then
     echo "Rotating model log at $MODEL_LOG to $MODEL_LOG.old"
-    if ! cat "$MODEL_LOG" >> "$MODEL_LOG.old"; then
+    cat "$MODEL_LOG" >> "$MODEL_LOG.old" 
-        report_error_and_exit "Failed to rotate model log"
+    : > "$MODEL_LOG"
-    fi
-    if ! : > "$MODEL_LOG"; then
-        report_error_and_exit "Failed to truncate model log"
-    fi
 fi
 
 # Populate /etc/environment with quoted values
 if ! grep -q "VAST" /etc/environment; then
-    if ! env -0 | grep -zEv "^(HOME=|SHLVL=)|CONDA" | while IFS= read -r -d '' line; do
+    env -0 | grep -zEv "^(HOME=|SHLVL=)|CONDA" | while IFS= read -r -d '' line; do
             name=${line%%=*}
             value=${line#*=}
             printf '%s="%s"\n' "$name" "$value"
-        done > /etc/environment; then
+        done > /etc/environment
-        echo "WARNING: Failed to populate /etc/environment, continuing anyway"
-    fi
 fi
 
 if [ ! -d "$ENV_PATH" ]
 then
     echo "setting up venv"
     if ! which uv; then
-        if ! curl -LsSf https://astral.sh/uv/install.sh | sh; then
+        curl -LsSf https://astral.sh/uv/install.sh | sh
-            report_error_and_exit "Failed to install uv package manager"
+        source ~/.local/bin/env
-        fi
-        if [[ -f ~/.local/bin/env ]]; then
-            if ! source ~/.local/bin/env; then
-                report_error_and_exit "Failed to source uv environment"
-            fi
-        else
-            echo "WARNING: ~/.local/bin/env not found after uv installation"
-        fi
     fi
 
     # Fork testing
-    if [[ ! -d $SERVER_DIR ]]; then
+    [[ ! -d $SERVER_DIR ]] && git clone "${PYWORKER_REPO:-https://github.com/vast-ai/pyworker}" "$SERVER_DIR"
-        if ! git clone "${PYWORKER_REPO:-https://github.com/vast-ai/pyworker}" "$SERVER_DIR"; then
-            report_error_and_exit "Failed to clone pyworker repository"
-        fi
-    fi
     if [[ -n ${PYWORKER_REF:-} ]]; then
-        if ! (cd "$SERVER_DIR" && git checkout "$PYWORKER_REF"); then
+        (cd "$SERVER_DIR" && git checkout "$PYWORKER_REF")
-            report_error_and_exit "Failed to checkout pyworker reference: $PYWORKER_REF"
-        fi
     fi
 
-    if ! uv venv --python-preference only-managed "$ENV_PATH" -p 3.10; then
+    uv venv --python-preference only-managed "$ENV_PATH" -p 3.10
-        report_error_and_exit "Failed to create virtual environment"
+    source "$ENV_PATH/bin/activate"
-    fi
 
-    if ! source "$ENV_PATH/bin/activate"; then
+    uv pip install -r "${SERVER_DIR}/requirements.txt"
-        report_error_and_exit "Failed to activate virtual environment"
-    fi
 
-    if ! uv pip install -r "${SERVER_DIR}/requirements.txt"; then
+    touch ~/.no_auto_tmux
-        report_error_and_exit "Failed to install Python requirements"
-    fi
-
-    if ! touch ~/.no_auto_tmux; then
-        report_error_and_exit "Failed to create ~/.no_auto_tmux"
-    fi
 else
-    if [[ -f ~/.local/bin/env ]]; then
+    [[ -f ~/.local/bin/env ]] && source ~/.local/bin/env
-        if ! source ~/.local/bin/env; then
+    source "$WORKSPACE_DIR/worker-env/bin/activate"
-            report_error_and_exit "Failed to source uv environment"
-        fi
-    fi
-    if ! source "$WORKSPACE_DIR/worker-env/bin/activate"; then
-        report_error_and_exit "Failed to activate existing virtual environment"
-    fi
     echo "environment activated"
     echo "venv: $VIRTUAL_ENV"
 fi
 
-[ ! -d "$SERVER_DIR/workers/$BACKEND" ] && report_error_and_exit "$BACKEND not supported!"
+[ ! -d "$SERVER_DIR/workers/$BACKEND" ] && echo "$BACKEND not supported!" && exit 1
 
 if [ "$USE_SSL" = true ]; then
 
-    if ! cat << EOF > /etc/openssl-san.cnf
+    cat << EOF > /etc/openssl-san.cnf
     [req]
     default_bits       = 2048
     distinguished_name = req_distinguished_name
@@ -183,25 +109,18 @@ if [ "$USE_SSL" = true ]; then
     [alt_names]
     IP.1   = 0.0.0.0
 EOF
-    then
-        report_error_and_exit "Failed to write OpenSSL config"
-    fi
 
-    if ! openssl req -newkey rsa:2048 -subj "/C=US/ST=CA/CN=pyworker.vast.ai/" \
+    openssl req -newkey rsa:2048 -subj "/C=US/ST=CA/CN=pyworker.vast.ai/" \
         -nodes \
         -sha256 \
         -keyout /etc/instance.key \
         -out /etc/instance.csr \
-        -config /etc/openssl-san.cnf; then
+        -config /etc/openssl-san.cnf
-        report_error_and_exit "Failed to generate SSL certificate request"
-    fi
 
-    if ! curl --header 'Content-Type: application/octet-stream' \
+    curl --header 'Content-Type: application/octet-stream' \
-        --data-binary @/etc/instance.csr \
+        --data-binary @//etc/instance.csr \
         -X \
-        POST "https://console.vast.ai/api/v0/sign_cert/?instance_id=$CONTAINER_ID" > /etc/instance.crt; then
+        POST "https://console.vast.ai/api/v0/sign_cert/?instance_id=$CONTAINER_ID" > /etc/instance.crt;
-        report_error_and_exit "Failed to sign SSL certificate"
-    fi
 fi
 
 
@@ -209,19 +128,58 @@ fi
 
 export REPORT_ADDR WORKER_PORT USE_SSL UNSECURED
 
-if ! cd "$SERVER_DIR"; then
+cd "$SERVER_DIR"
-    report_error_and_exit "Failed to cd into SERVER_DIR: $SERVER_DIR"
-fi
 
 echo "launching PyWorker server"
 
 set +e
-python3 -m "workers.$BACKEND.server" |& tee -a "$PYWORKER_LOG"
+
+# Try worker entrypoint first
+echo "trying workers.${BACKEND}.worker"
+python3 -m "workers.${BACKEND}.worker" |& tee -a "$PYWORKER_LOG"
 PY_STATUS=${PIPESTATUS[0]}
+
+# If that fails, fall back to server
+if [ "${PY_STATUS}" -ne 0 ]; then
+  echo "workers.${BACKEND}.worker failed with status ${PY_STATUS}, trying workers.${BACKEND}.server"
+  python3 -m "workers.${BACKEND}.server" |& tee -a "$PYWORKER_LOG"
+  PY_STATUS=${PIPESTATUS[0]}
+fi
+
 set -e
 
 if [ "${PY_STATUS}" -ne 0 ]; then
-  report_error_and_exit "PyWorker exited with status ${PY_STATUS}"
+  echo "PyWorker exited with status ${PY_STATUS}; notifying autoscaler..."
+  ERROR_MSG="PyWorker exited: code ${PY_STATUS}"
+  MTOKEN="${MASTER_TOKEN:-}"
+  VERSION="${PYWORKER_VERSION:-0}"
+
+  IFS=',' read -r -a REPORT_ADDRS <<< "${REPORT_ADDR}"
+  for addr in "${REPORT_ADDRS[@]}"; do
+    curl -sS -X POST -H 'Content-Type: application/json' \
+      -d "$(cat <<JSON
+{
+  "id": ${CONTAINER_ID:-0},
+  "mtoken": "${MTOKEN}",
+  "version": "${VERSION}",
+  "loadtime": 0,
+  "new_load": 0,
+  "cur_load": 0,
+  "rej_load": 0,
+  "max_perf": 0,
+  "cur_perf": 0,
+  "error_msg": "${ERROR_MSG}",
+  "num_requests_working": 0,
+  "num_requests_recieved": 0,
+  "additional_disk_usage": 0,
+  "working_request_idxs": [],
+  "cur_capacity": 0,
+  "max_capacity": 0,
+  "url": "${URL}"
+}
+JSON
+)" "${addr%/}/worker_status/" || true
+  done
 fi
 
 echo "launching PyWorker server done"

workers/ace/worker.py

@@ -0,0 +1,184 @@
+import random
+import sys
+
+from vastai import Worker, WorkerConfig, HandlerConfig, LogActionConfig, BenchmarkConfig
+
+# ComyUI model configuration
+MODEL_SERVER_URL           = 'http://127.0.0.1'
+MODEL_SERVER_PORT          = 18288
+MODEL_LOG_FILE             = '/var/log/portal/comfyui.log'
+MODEL_HEALTHCHECK_ENDPOINT = "/health"
+
+# ComyUI-specific log messages
+MODEL_LOAD_LOG_MSG = [
+    "To see the GUI go to: "
+]
+
+MODEL_ERROR_LOG_MSGS = [
+    "MetadataIncompleteBuffer",
+    "Value not in list: ",
+    "[ERROR] Provisioning Script failed"
+]
+
+MODEL_INFO_LOG_MSGS = [
+    '"message":"Downloading'
+]
+
+benchmark_lyrics = [
+    "[verse]\nGuardian cloaked in twilight hue\nShadows melt where he breaks through\nEchoes swirl in mystic flight\nHooded hero owns the night\n\n[verse]\nThrough the chaos shapes arise\nFeral whispers, glowing eyes\nOrcs and creatures side by side\nMarch within the inky tide\n\n[chorus]\nRise above the fear and gloom\nLet your courage fully bloom\nIn the darkness stand your ground\nHear the night proclaim your sound",
+    "[verse]\nMorning sun on fields of gold\nGentle stories unfold\nEvery breeze a quiet song\nWhere the peaceful hearts belong\n\n[verse]\nLanterns glow at stable doors\nRustling leaves on orchard floors\nSimple joys in every hand\nLife grows soft in fertile land\n\n[chorus]\nLet the day drift slow and free\nRoot your soul where you can be\nIn this haven warm and bright\nFeel the earth breathe pure delight",
+    "[verse]\nLittle feet on dusty ground\nChasing dreams without a sound\nSoccer ball in morning light\nHopes take wing in youthful flight\n\n[verse]\nChrome reflections paint the day\nSwagger in the steps that play\nCopper tones in shining air\nChildhood gleaming everywhere\n\n[chorus]\nKick the world with boundless cheer\nHold the magic close and near\nIn each moment bold and true\nLet the sky belong to you",
+    "[verse]\nSunset bleeds across the street\nGilded calm in summer heat\nLow-rise towers rimmed with fire\nDreams ignite as lights climb higher\n\n[verse]\nFootsteps scatter through the haze\nFutures shimmer in the blaze\nEvery window tells a tale\nFloating through a tangerine veil\n\n[chorus]\nLet the neon softly glow\nLet your restless heartbeat slow\nIn this city forged in light\nCarry hope into the night",
+    "[verse]\nOcean breathes in rolling arcs\nSprays of diamond, glowing sparks\nWaves unfold a perfect line\nNature’s rhythm feels divine\n\n[verse]\nSun above in golden sweep\nPaints the rise of every deep\nShimmer drifting through the blue\nWorld reborn in every view\n\n[chorus]\nLet the tide pull you along\nHear the water’s ancient song\nIn the cresting waves you’ll find\nQuiet peace for heart and mind",
+    "[verse]\nGlass aglow with swirling light\nFruits and mints in colors bright\nIcy whispers clink and chime\nFlowing forms suspend in time\n\n[verse]\nCreamy spirals drift within\nGentle currents slowly spin\nWarm reflections lingering sweet\nMixing flavors at your feet\n\n[chorus]\nSip the glow and let it rise\nTaste the sunset in disguise\nIn this moment clear and true\nLet the warmth flow into you",
+    "[verse]\nEngines rumble down the lane\nCopper clouds of steam and rain\nOilpunk dreams in metal shine\nRider drifting down the line\n\n[verse]\nLeather jacket, steady glare\nStories sparking in the air\nMagazine lights frame his face\nKing of roads in timeless grace\n\n[chorus]\nThrottle up beyond the bend\nFeel the force of steel ascend\nRide the night and hold on tight\nClaim the world in streaks of light",
+    "[verse]\nCut-out shapes in swirling play\nTextures dance in bold array\nCats in denim, grinning wide\nStrut across the patterned tide\n\n[verse]\nPosters hum with neon glow\nSurreal scenes begin to grow\nColors crisp as folded art\nPatchwork beating like a heart\n\n[chorus]\nLet the collage come alive\nWatch the vibrant pieces thrive\nIn this joyful, crafted space\nEvery shape finds its own place",
+    "[verse]\nTiny world in crystal glass\nAncient tales behind the mass\nVillage lights in winter gleam\nFrozen in a mystic dream\n\n[verse]\nLantern beams in swirling air\nSoft enchantment everywhere\nShadows drift with gentle grace\nMagic sealed within the space\n\n[chorus]\nHold the sphere and you will see\nEchoes of a memory\nIn the glow of fragile light\nLives a realm of pure delight",
+    "[verse]\nArmor hums with power bright\nChopping sparks in jungle night\nMecha spirits shift and scream\nThrough the ferns like shattered beams\n\n[verse]\nAxes blaze in glowing arcs\nLighting up the shadowed marks\nNature roars in trembling air\nClash of steel and cosmic flare\n\n[chorus]\nRaise the fire, strike the ground\nLet your legend shake the sound\nIn the wild where echoes roam\nForge the fight and carve your home",
+    "[verse]\nCrowds ignite in vibrant flare\nBeats explode through smoky air\nDJ robes replaced with flame\nPope on decks in holy frame\n\n[verse]\nLeather gleams in blinding light\nTurntables spin with sacred might\nChoirs echo in the bass\nHeaven pulses through the place\n\n[chorus]\nLift the roof and shake the floor\nSacred rhythm evermore\nLet the music take control\nFeel the blessing in your soul",
+]
+
+benchmark_dataset = [
+    {
+        "input": {
+            "request_id": "",
+            "workflow_json": {
+                "14": {
+                "inputs": {
+                    "tags": "funk, pop, soul, rock, melodic, guitar, drums, bass, keyboard, percussion, 105 BPM, energetic, upbeat, groovy, vibrant, dynamic",
+                    "lyrics": lyrics,
+                    "lyrics_strength": 0.99,
+                    "clip": ["40", 1]
+                },
+                "class_type": "TextEncodeAceStepAudio",
+                "_meta": {
+                    "title": "TextEncodeAceStepAudio"
+                }
+                },
+                "17": {
+                "inputs": {
+                    "seconds": 180,
+                    "batch_size": 1
+                },
+                "class_type": "EmptyAceStepLatentAudio",
+                "_meta": {
+                    "title": "EmptyAceStepLatentAudio"
+                }
+                },
+                "18": {
+                "inputs": {
+                    "samples": ["52", 0],
+                    "vae": ["40", 2]
+                },
+                "class_type": "VAEDecodeAudio",
+                "_meta": {
+                    "title": "VAE Decode Audio"
+                }
+                },
+                "40": {
+                "inputs": {
+                    "ckpt_name": "ace_step_v1_3.5b.safetensors"
+                },
+                "class_type": "CheckpointLoaderSimple",
+                "_meta": {
+                    "title": "Load Checkpoint"
+                }
+                },
+                "44": {
+                "inputs": {
+                    "conditioning": ["14", 0]
+                },
+                "class_type": "ConditioningZeroOut",
+                "_meta": {
+                    "title": "ConditioningZeroOut"
+                }
+                },
+                "49": {
+                "inputs": {
+                    "model": ["51", 0],
+                    "operation": ["50", 0]
+                },
+                "class_type": "LatentApplyOperationCFG",
+                "_meta": {
+                    "title": "LatentApplyOperationCFG"
+                }
+                },
+                "50": {
+                "inputs": {
+                    "multiplier": 1.15
+                },
+                "class_type": "LatentOperationTonemapReinhard",
+                "_meta": {
+                    "title": "LatentOperationTonemapReinhard"
+                }
+                },
+                "51": {
+                "inputs": {
+                    "shift": 6,
+                    "model": ["40", 0]
+                },
+                "class_type": "ModelSamplingSD3",
+                "_meta": {
+                    "title": "ModelSamplingSD3"
+                }
+                },
+                "52": {
+                "inputs": {
+                    "seed": "__RANDOM_INT__",
+                    "steps": 65,
+                    "cfg": 4,
+                    "sampler_name": "er_sde",
+                    "scheduler": "linear_quadratic",
+                    "denoise": 1,
+                    "model": ["49", 0],
+                    "positive": ["14", 0],
+                    "negative": ["44", 0],
+                    "latent_image": ["17", 0]
+                },
+                "class_type": "KSampler",
+                "_meta": {
+                    "title": "KSampler"
+                }
+                },
+                "59": {
+                "inputs": {
+                    "filename_prefix": "audio/ComfyUI",
+                    "quality": "V0",
+                    "audioUI": "",
+                    "audio": ["18", 0]
+                },
+                "class_type": "SaveAudioMP3",
+                "_meta": {
+                    "title": "Save Audio (MP3)"
+                }
+                }
+            }
+        }
+    } for lyrics in benchmark_lyrics
+]
+
+worker_config = WorkerConfig(
+    model_server_url=MODEL_SERVER_URL,
+    model_server_port=MODEL_SERVER_PORT,
+    model_log_file=MODEL_LOG_FILE,
+    model_healthcheck_url=MODEL_HEALTHCHECK_ENDPOINT,
+    handlers=[
+        HandlerConfig(
+            route="/generate/sync",
+            allow_parallel_requests=False,
+            max_queue_time=10.0,
+            benchmark_config=BenchmarkConfig(
+                dataset=benchmark_dataset,
+                runs=1
+            ),
+            workload_calculator= lambda _ : 1000.0
+        )
+    ],
+    log_action_config=LogActionConfig(
+        on_load=MODEL_LOAD_LOG_MSG,
+        on_error=MODEL_ERROR_LOG_MSGS,
+        on_info=MODEL_INFO_LOG_MSGS
+    )
+)
+
+Worker(worker_config).run()

workers/comfyui-json/README.md

@@ -1,105 +1,15 @@
 # ComfyUI PyWorker
 
-This is the base PyWorker for ComfyUI. It provides a unified interface for running any ComfyUI workflow through a proxy-based architecture. See the [Serverless documentation](https://docs.vast.ai/serverless) for guides and how-to's.
+This is the base PyWorker for ComfyUI. It provides a unified interface for running any ComfyUI workflow through a proxy-based architecture.
 
 The cost for each request has a static value of `1`.  ComfyUI does not handle concurrent workloads and there is no current provision to load multiple instances of ComfyUI per worker node.
 
-## Instance Setup
-
-1. Pick a template
-
-- [ComfyUI (Serverless)](https://cloud.vast.ai/?ref_id=62897&creator_id=62897&name=ComfyUI%20(Serverless))
-
-2. Follow the [getting started guide](https://docs.vast.ai/documentation/serverless/quickstart) for help with configuring your serverless setup. For testing, we recommend that you use the default options presented by the web interface.
-
 ## Requirements
 
 This worker requires both [ComfyUI](https://github.com/comfyanonymous/ComfyUI) and [ComfyUI API Wrapper](https://github.com/ai-dock/comfyui-api-wrapper).
 
 A docker image is provided but you may use any if the above requirements are met.
 
-## Client
-
-The client demonstrates how to use the Vast Serverless SDK to generate images, save them locally, and optionally upload to S3-compatible storage.
-
-### Setup
-
-1. Clone the PyWorker repository to your local machine and install the necessary requirements for running the test client.
-
-```bash
-git clone https://github.com/vast-ai/pyworker
-cd pyworker
-pip install uv
-uv venv -p 3.12
-source .venv/bin/activate
-uv pip install -r requirements.txt
-```
-
-2. Set your API key:
-
-```bash
-export VAST_API_KEY=<your_api_key>
-```
-
-### Usage
-
-```bash
-# Default prompt
-python -m workers.comfyui-json.client
-
-# Custom prompt
-python -m workers.comfyui-json.client --prompt "a cat sitting on a rainbow"
-
-# With options
-python -m workers.comfyui-json.client --prompt "sunset" --width 1024 --height 1024 --steps 30
-
-# Using a custom workflow file
-python -m workers.comfyui-json.client --workflow my_workflow.json
-
-# With S3 upload
-python -m workers.comfyui-json.client --s3
-```
-
-### CLI Flags
-
-| Flag | Default | Description |
-|------|---------|-------------|
-| `--endpoint` | `my-comfyui-endpoint` | Vast endpoint name |
-| `--prompt` | (default) | Text prompt for image generation |
-| `--workflow` | (none) | Path to custom workflow JSON file |
-| `--width` | 512 | Image width in pixels |
-| `--height` | 512 | Image height in pixels |
-| `--steps` | 20 | Number of denoising steps |
-| `--seed` | (random) | Random seed for reproducibility |
-| `--s3` | (disabled) | Upload generated images to S3 |
-
-### Output
-
-Images are saved to `./generated_images/comfy_{seed}.png`.
-
-### S3 Upload (Optional)
-
-You can optionally upload generated images to an S3-compatible storage service (AWS S3, Cloudflare R2, Backblaze B2, etc.) by using the `--s3` flag.
-
-**1. Set environment variables:**
-
-```bash
-export S3_ENDPOINT_URL="https://your-account.r2.cloudflarestorage.com"
-export S3_BUCKET_NAME="my-bucket"
-export S3_ACCESS_KEY_ID="your-access-key-id"
-export S3_SECRET_ACCESS_KEY="your-secret-access-key"
-```
-
-**2. Run with S3 upload enabled:**
-
-```bash
-python -m workers.comfyui-json.client --prompt "a beautiful landscape" --s3
-```
-
-Images will be saved locally AND uploaded to `s3://{bucket}/comfyui/{filename}`.
-
-**Note:** Requires `boto3` (`pip install boto3`).
-
 ## Benchmarking
 
 ### Custom Benchmark Workflows
@@ -302,3 +212,11 @@ WEBHOOK_TIMEOUT=30                   # Webhook timeout in seconds
   }
 }
 ```
+
+## Client Libraries
+
+See the test client examples for implementation details on how to integrate with the ComfyUI worker.
+
+---
+
+See Vast's serverless documentation for more details on how to use ComfyUI with autoscaler.

workers/comfyui-json/client.py

@@ -1,312 +1,35 @@
-import os
+from .data_types import count_workload
-import sys
-import json
 import uuid
 import random
 import asyncio
-import logging
+import random
-import argparse
-import aiohttp
 
 from vastai import Serverless
 
-# ---------------------- Config ----------------------
+async def main():
-DEFAULT_PROMPT = "a beautiful sunset over mountains, digital art, highly detailed"
+    async with Serverless() as client:
-ENDPOINT_NAME = "my-comfyui-endpoint"
+        endpoint = await client.get_endpoint(name="my-comfy-endpoint") # Change this to your endpoint name
-DEFAULT_WIDTH = 512
-DEFAULT_HEIGHT = 512
-DEFAULT_STEPS = 20
-COST = 100  # Fixed cost for ComfyUI requests
 
-# Optional S3 Configuration (from environment variables)
-S3_ENDPOINT_URL = os.getenv("S3_ENDPOINT_URL")
-S3_BUCKET_NAME = os.getenv("S3_BUCKET_NAME")
-S3_ACCESS_KEY_ID = os.getenv("S3_ACCESS_KEY_ID")
-S3_SECRET_ACCESS_KEY = os.getenv("S3_SECRET_ACCESS_KEY")
-
-logging.basicConfig(level=logging.INFO, format="%(levelname)s - %(message)s")
-log = logging.getLogger(__name__)
-
-
-def get_s3_client():
-    """Create and return an S3 client configured for the S3-compatible endpoint"""
-    try:
-        import boto3
-        from botocore.config import Config
-    except ImportError:
-        log.error("boto3 is required for S3 uploads. Install with: pip install boto3")
-        return None
-
-    if not all([S3_ENDPOINT_URL, S3_BUCKET_NAME, S3_ACCESS_KEY_ID, S3_SECRET_ACCESS_KEY]):
-        log.error("S3 environment variables not fully configured. Required:")
-        log.error("  S3_ENDPOINT_URL, S3_BUCKET_NAME, S3_ACCESS_KEY_ID, S3_SECRET_ACCESS_KEY")
-        return None
-
-    return boto3.client(
-        "s3",
-        endpoint_url=S3_ENDPOINT_URL,
-        aws_access_key_id=S3_ACCESS_KEY_ID,
-        aws_secret_access_key=S3_SECRET_ACCESS_KEY,
-        config=Config(signature_version="s3v4"),
-    )
-
-
-# ---------------------- API Functions ----------------------
-async def call_generate(
-    client: Serverless,
-    *,
-    endpoint_name: str,
-    prompt: str,
-    width: int,
-    height: int,
-    steps: int,
-    seed: int,
-) -> dict:
-    """Generate image using Text2Image modifier"""
-    endpoint = await client.get_endpoint(name=endpoint_name)
         payload = {
             "input": {
                 "request_id": str(uuid.uuid4()),
                 "modifier": "Text2Image",
                 "modifications": {
-                "prompt": prompt,
+                    "prompt": "a beautiful landscape with mountains and lakes",
-                "width": width,
+                    "width": 1024,
-                "height": height,
+                    "height": 1024,
-                "steps": steps,
+                    "steps": 20,
-                "seed": seed,
+                    "seed": random.randint(0, 2**32 - 1)
                 },
+                "workflow_json": {}  # Empty since using modifier approach
             }
         }
-    return await endpoint.request("/generate/sync", payload, cost=COST)
         
+        response = await endpoint.request("/generate/sync", payload, cost=count_workload())
 
-async def call_generate_workflow(
+        # Get the file from the path on the local machine using SCP or SFTP
-    client: Serverless,
+        # or configure S3 to upload to cloud storage.
-    *,
+        print(response["response"]["output"][0]["local_path"])
-    endpoint_name: str,
-    workflow_json: dict,
-) -> dict:
-    """Generate using custom workflow JSON"""
-    endpoint = await client.get_endpoint(name=endpoint_name)
-    payload = {
-        "input": {
-            "request_id": str(uuid.uuid4()),
-            "workflow_json": workflow_json,
-        }
-    }
-    return await endpoint.request("/generate/sync", payload, cost=COST)
-
-
-# ---------------------- Demo Class ----------------------
-class APIDemo:
-    def __init__(self, client: Serverless, endpoint_name: str, upload_s3: bool = False):
-        self.client = client
-        self.endpoint_name = endpoint_name
-        self.upload_s3 = upload_s3
-        self.s3_client = get_s3_client() if upload_s3 else None
-        
-        if upload_s3 and not self.s3_client:
-            log.warning("S3 upload requested but client creation failed. Images will only be saved locally.")
-
-    def extract_filename(self, response: dict) -> str | None:
-        """Extract the generated image filename from ComfyUI response"""
-        if "comfyui_response" in response:
-            for data in response["comfyui_response"].values():
-                if isinstance(data, dict) and "outputs" in data:
-                    for node_output in data["outputs"].values():
-                        if "images" in node_output and node_output["images"]:
-                            return node_output["images"][0].get("filename")
-        return None
-
-    async def save_image(self, worker_url: str, filename: str, local_name: str) -> str | None:
-        """Fetch and save image locally from the worker, optionally upload to S3"""
-        os.makedirs("generated_images", exist_ok=True)
-        return await self._fetch_image(worker_url, filename, local_name)
-
-    def _upload_to_s3(self, local_path: str, s3_key: str) -> str | None:
-        """Upload a local file to S3 and return the S3 URL"""
-        if not self.s3_client:
-            return None
-        
-        try:
-            self.s3_client.upload_file(
-                local_path,
-                S3_BUCKET_NAME,
-                s3_key,
-                ExtraArgs={"ContentType": "image/png"}
-            )
-            s3_url = f"{S3_ENDPOINT_URL}/{S3_BUCKET_NAME}/{s3_key}"
-            print(f"  ☁️  Uploaded to S3: {s3_key}")
-            return s3_url
-        except Exception as e:
-            log.error(f"Failed to upload to S3: {e}")
-            return None
-
-    async def _fetch_image(self, worker_url: str, filename: str, local_name: str) -> str | None:
-        """Fetch image from worker's /view endpoint and save locally"""
-        if not worker_url:
-            return None
-            
-        try:
-            url = f"{worker_url}/view"
-            params = {"filename": filename, "type": "output"}
-            
-            async with aiohttp.ClientSession() as session:
-                async with session.get(url, params=params, ssl=False) as resp:
-                    if resp.status == 200:
-                        path = f"generated_images/{local_name}"
-                        image_data = await resp.read()
-                        with open(path, "wb") as f:
-                            f.write(image_data)
-                        print(f"  💾 Saved: {path}")
-                        
-                        # Upload to S3 if enabled
-                        if self.upload_s3 and self.s3_client:
-                            s3_key = f"comfyui/{local_name}"
-                            self._upload_to_s3(path, s3_key)
-                        
-                        return path
-            return None
-        except Exception:
-            return None
-
-    async def demo_prompt(
-        self,
-        prompt: str,
-        width: int,
-        height: int,
-        steps: int,
-        seed: int | None,
-    ):
-        """Demo: Generate image from text prompt"""
-        print("=" * 60)
-        print("COMFYUI TEXT-TO-IMAGE DEMO")
-        print("=" * 60)
-
-        if seed is None:
-            seed = random.randint(0, 2**32 - 1)
-
-        print(f"Prompt: {prompt[:100]}..." if len(prompt) > 100 else f"Prompt: {prompt}")
-        print(f"Size: {width}x{height}, Steps: {steps}, Seed: {seed}")
-        print("\n🎨 Generating image...")
-
-        response = await call_generate(
-            self.client,
-            endpoint_name=self.endpoint_name,
-            prompt=prompt,
-            width=width,
-            height=height,
-            steps=steps,
-            seed=seed,
-        )
-
-        print("\n✅ Generation complete!")
-
-        # Get worker URL for fetching images
-        worker_url = response.get("url", "")
-        print(f"Worker URL: {worker_url}")
-
-        # Fetch and save image
-        if "response" in response:
-            filename = self.extract_filename(response["response"])
-            if filename:
-                path = await self.save_image(worker_url, filename, f"comfy_{seed}.png")
-                if not path:
-                    print(f"❌ Failed to fetch image")
-            else:
-                print("❌ No image in response")
-        else:
-            print("❌ Unexpected response format")
-
-    async def demo_workflow(self, workflow_file: str):
-        """Demo: Generate using custom workflow file"""
-        print("=" * 60)
-        print("COMFYUI CUSTOM WORKFLOW DEMO")
-        print("=" * 60)
-
-        if not os.path.exists(workflow_file):
-            log.error(f"Workflow file not found: {workflow_file}")
-            return
-
-        with open(workflow_file, "r") as f:
-            workflow_json = json.load(f)
-
-        print(f"Workflow: {workflow_file}")
-        print("\n🎨 Generating...")
-
-        response = await call_generate_workflow(
-            self.client,
-            endpoint_name=self.endpoint_name,
-            workflow_json=workflow_json,
-        )
-
-        print("\n✅ Generation complete!")
-
-        worker_url = response.get("url", "")
-
-        if "response" in response:
-            filename = self.extract_filename(response["response"])
-            if filename:
-                path = await self.save_image(worker_url, filename, "workflow.png")
-                if not path:
-                    print(f"❌ Failed to fetch image")
-            else:
-                print("❌ No image in response")
-        else:
-            print("❌ Unexpected response format")
-
-
-# ---------------------- CLI ----------------------
-def build_arg_parser() -> argparse.ArgumentParser:
-    p = argparse.ArgumentParser(description="Vast ComfyUI-JSON Demo (Serverless SDK)")
-    p.add_argument("--endpoint", default=ENDPOINT_NAME, help=f"Vast endpoint name (default: {ENDPOINT_NAME})")
-    p.add_argument("--prompt", type=str, default=DEFAULT_PROMPT, metavar="TEXT",
-                   help=f"Prompt text (default: '{DEFAULT_PROMPT[:30]}...')")
-    p.add_argument("--workflow", type=str, metavar="FILE", help="Use custom workflow JSON file instead")
-    p.add_argument("--width", type=int, default=DEFAULT_WIDTH, help=f"Image width (default: {DEFAULT_WIDTH})")
-    p.add_argument("--height", type=int, default=DEFAULT_HEIGHT, help=f"Image height (default: {DEFAULT_HEIGHT})")
-    p.add_argument("--steps", type=int, default=DEFAULT_STEPS, help=f"Steps (default: {DEFAULT_STEPS})")
-    p.add_argument("--seed", type=int, default=None, help="Seed (default: random)")
-    p.add_argument("--s3", action="store_true", 
-                   help="Upload generated images to S3 (requires S3_ENDPOINT_URL, S3_BUCKET_NAME, S3_ACCESS_KEY_ID, S3_SECRET_ACCESS_KEY env vars)")
-    return p
-
-
-async def main_async():
-    args = build_arg_parser().parse_args()
-
-    print("=" * 60)
-    print(f"Using endpoint: {args.endpoint}")
-    if args.s3:
-        print(f"S3 upload: enabled (bucket: {S3_BUCKET_NAME})")
-
-    try:
-        async with Serverless() as client:
-            demo = APIDemo(client, args.endpoint, upload_s3=args.s3)
-
-            if args.workflow:
-                await demo.demo_workflow(workflow_file=args.workflow)
-            else:
-                await demo.demo_prompt(
-                    prompt=args.prompt,
-                    width=args.width,
-                    height=args.height,
-                    steps=args.steps,
-                    seed=args.seed,
-                )
-
-    except AttributeError as e:
-        if "API key" in str(e):
-            log.error("API key missing. Set VAST_API_KEY environment variable.")
-        else:
-            log.error(f"Error: {e}")
-        sys.exit(1)
-    except Exception as e:
-        log.error(f"Error: {e}")
-        import traceback
-        traceback.print_exc()
-        sys.exit(1)
-
 
 if __name__ == "__main__":
-    asyncio.run(main_async())
+    asyncio.run(main())

workers/comfyui-json/server.py

@@ -4,7 +4,6 @@ import dataclasses
 import base64
 from typing import Optional, Union, Type
 
-import aiohttp
 from aiohttp import web, ClientResponse
 
 from lib.backend import Backend, LogAction
@@ -14,7 +13,6 @@ from .data_types import ComfyWorkflowData
 
 
 MODEL_SERVER_URL = os.getenv("MODEL_SERVER_URL", "http://127.0.0.1:18288")
-COMFYUI_URL = os.getenv("COMFYUI_URL", "http://127.0.0.1:18188")  # Raw ComfyUI server
 
 # This is the last log line that gets emitted once comfyui+extensions have been fully loaded
 MODEL_SERVER_START_LOG_MSG = "To see the GUI go to: "
@@ -110,39 +108,8 @@ async def handle_ping(_):
     return web.Response(body="pong")
 
 
-async def handle_view(request: web.Request) -> web.Response:
-    """Proxy /view requests to raw ComfyUI server to fetch generated images"""
-    # Forward query params to raw ComfyUI (not the API wrapper)
-    query_string = request.query_string
-    url = f"{COMFYUI_URL}/view?{query_string}"
-    
-    log.debug(f"Proxying /view request to: {url}")
-    
-    try:
-        async with aiohttp.ClientSession() as session:
-            async with session.get(url) as resp:
-                if resp.status == 200:
-                    content = await resp.read()
-                    return web.Response(
-                        body=content,
-                        status=200,
-                        content_type=resp.content_type or "image/png"
-                    )
-                else:
-                    text = await resp.text()
-                    return web.Response(
-                        text=text,
-                        status=resp.status,
-                        content_type="text/plain"
-                    )
-    except Exception as e:
-        log.error(f"Error proxying /view: {e}")
-        return web.Response(text=str(e), status=500)
-
-
 routes = [
     web.post("/generate/sync", backend.create_handler(ComfyWorkflowHandler())),
-    web.get("/view", handle_view),
     web.get("/ping", handle_ping),
 ]
 

workers/comfyui-json/worker.py

@@ -0,0 +1,81 @@
+import random
+import sys
+
+from vastai import Worker, WorkerConfig, HandlerConfig, LogActionConfig, BenchmarkConfig
+
+# ComyUI model configuration
+MODEL_SERVER_URL           = 'http://127.0.0.1'
+MODEL_SERVER_PORT          = 18288
+MODEL_LOG_FILE             = '/var/log/portal/comfyui.log'
+MODEL_HEALTHCHECK_ENDPOINT = "/health"
+
+# ComyUI-specific log messages
+MODEL_LOAD_LOG_MSG = [
+    "To see the GUI go to: "
+]
+
+MODEL_ERROR_LOG_MSGS = [
+    "MetadataIncompleteBuffer",
+    "Value not in list: ",
+    "[ERROR] Provisioning Script failed"
+]
+
+MODEL_INFO_LOG_MSGS = [
+    '"message":"Downloading'
+]
+
+benchmark_prompts = [
+    "Cartoon hoodie hero; orc, anime cat, bunny; black goo; buff; vector on white.",
+    "Cozy farming-game scene with fine details.",
+    "2D vector child with soccer ball; airbrush chrome; swagger; antique copper.",
+    "Realistic futuristic downtown of low buildings at sunset.",
+    "Perfect wave front view; sunny seascape; ultra-detailed water; artful feel.",
+    "Clear cup with ice, fruit, mint; creamy swirls; fluid-sim CGI; warm glow.",
+    "Male biker with backpack on motorcycle; oilpunk; award-worthy magazine cover.",
+    "Collage for textile; surreal cartoon cat in cap/jeans before poster; crisp.",
+    "Medieval village inside glass sphere; volumetric light; macro focus.",
+    "Iron Man with glowing axe; mecha sci-fi; jungle scene; dynamic light.",
+    "Pope Francis DJ in leather jacket, mixing on giant console; dramatic.",
+]
+
+
+
+benchmark_dataset = [
+    {
+        "input": {
+            "request_id": f"test-{random.randint(1000, 99999)}",
+            "modifier": "Text2Image",
+            "modifications": {
+                "prompt": prompt,
+                "width": 512,
+                "height": 512,
+                "steps": 20,
+                "seed": random.randint(0, sys.maxsize)
+            }
+        }
+    } for prompt in benchmark_prompts
+]
+
+worker_config = WorkerConfig(
+    model_server_url=MODEL_SERVER_URL,
+    model_server_port=MODEL_SERVER_PORT,
+    model_log_file=MODEL_LOG_FILE,
+    model_healthcheck_url=MODEL_HEALTHCHECK_ENDPOINT,
+    handlers=[
+        HandlerConfig(
+            route="/generate/sync",
+            allow_parallel_requests=False,
+            max_queue_time=10.0,
+            benchmark_config=BenchmarkConfig(
+                dataset=benchmark_dataset,
+            )
+        )
+    ],
+    log_action_config=LogActionConfig(
+        on_load=MODEL_LOAD_LOG_MSG,
+        on_error=MODEL_ERROR_LOG_MSGS,
+        on_info=MODEL_INFO_LOG_MSGS
+    )
+)
+
+Worker(worker_config).run()

workers/openai/README.md

@@ -8,13 +8,14 @@ This is the base PyWorker for OpenAI compatible inference servers.  See the [Ser
 
 This worker is compatible with any backend API that properly implements the `/v1/completions` and `/v1/chat/completions` endpoints.  We currently have three templates you can choose from but you can also create your own without having to modify the PyWorker.
 
-- [vLLM](https://cloud.vast.ai/?ref_id=62897&creator_id=62897&name=vLLM%20(Serverless)) (recommended)
+- [vLLM](https://cloud.vast.ai/?ref_id=62897&creator_id=62897&name=vLLM%20%2B%20Qwen%2FQwen3-8B%20(Serverless)) (recommended)
 - [Ollama](https://cloud.vast.ai/?ref_id=62897&creator_id=62897&name=Ollama%20%2B%20Qwen3%3A32b%20(Serverless))
+- [HuggingFace TGI](https://cloud.vast.ai/?ref_id=62897&creator_id=62897&name=TGI%20%2B%20Qwen3-8B%20(Serverless))
 
 
 All of these templates can be configured via the template interface.  You may want to change the model or startup arguments, depending on the template you selected.
 
-2. Follow the [getting started guide](https://docs.vast.ai/documentation/serverless/quickstart) for help with configuring your serverless setup.  For testing, we recommend that you use the default options presented by the web interface.
+2. Follow the [getting started guide](https://docs.vast.ai/serverless/getting-started) for help with configuring your serverless setup.  For testing, we recommend that you use the default options presented by the web interface.
 
 ## Client Setup (Demo)
 
@@ -33,30 +34,12 @@ uv pip install -r requirements.txt
 
 Several examples have been provided in the client to help you get started with your own implementation.
 
-First, set your API key as an environment variable:
+### Completions
+
+Call to `/v1/completions` with json response
 
 ```bash
-export VAST_API_KEY=<your_api_key>
+python -m workers.openai.client -k <API_KEY> -e <ENDPOINT_NAME> --completion --model <MODEL_NAME>
-```
-
-The `--model` and `--endpoint` flags are optional. If not provided, they default to `Qwen/Qwen3-8B` and `my-vllm-endpoint` respectively.
-
-### Chat Completion (streaming)
-
-Call to `/v1/chat/completions` with streaming response
-
-```bash
-python -m workers.openai.client --chat-stream --endpoint <ENDPOINT_NAME> --model <MODEL_NAME>
-```
-
-### Interactive Chat (streaming)
-
-Interactive session with calls to `/v1/chat/completions`.
-
-Type `clear` to clear the chat history or `quit` to exit.
-
-```bash
-python -m workers.openai.client --interactive --endpoint <ENDPOINT_NAME> --model <MODEL_NAME>
 ```
 
 ### Chat Completion (json)
@@ -64,7 +47,15 @@ python -m workers.openai.client --interactive --endpoint <ENDPOINT_NAME> --model
 Call to `/v1/chat/completions` with json response
 
 ```bash
-python -m workers.openai.client --chat --endpoint <ENDPOINT_NAME> --model <MODEL_NAME>
+python -m workers.openai.client -k <API_KEY> -e <ENDPOINT_NAME> --chat --model <MODEL_NAME>
+```
+
+### Chat Completion (streaming)
+
+Call to `/v1/chat/completions` with streaming response
+
+```bash
+python -m workers.openai.client -k <API_KEY> -e <ENDPOINT_NAME> --chat-stream --model <MODEL_NAME>
 ```
 
 ### Tool Use (json)
@@ -74,14 +65,16 @@ Call to `/v1/chat/completions` with tool and json response.
 This test defines a simple tool which will list the contents of the local pyworker directory.  The output is then analysed by the model.
 
 ```bash
-python -m workers.openai.client --tools --endpoint <ENDPOINT_NAME> --model <MODEL_NAME>
+python -m workers.openai.client -k <API_KEY> -e <ENDPOINT_NAME> --tools --model <MODEL_NAME>
 ```
 
-### Completions
+### Interactive Chat (streaming)
 
-Call to `/v1/completions` with json response
+Interactive session with calls to `/v1/chat/completions`.
+
+Type `clear` to clear the chat history or `quit` to exit.
 
 ```bash
-python -m workers.openai.client --completion --endpoint <ENDPOINT_NAME> --model <MODEL_NAME>
+python -m workers.openai.client -k <API_KEY> -e <ENDPOINT_NAME> --interactive --model <MODEL_NAME>
 ```
 

workers/openai/client.py

@@ -18,7 +18,7 @@ logging.basicConfig(
 log = logging.getLogger(__file__)
 
 # ---------------------- Prompts ----------------------
-COMPLETIONS_PROMPT = "Zebras are primarily grazers and can subsist on lower-quality vegetation. They are preyed on mainly by"
+COMPLETIONS_PROMPT = "the capital of USA is"
 CHAT_PROMPT = "Think step by step: Tell me about the Python programming language."
 TOOLS_PROMPT = (
     "Can you list the files in the current working directory and tell me what you see? "
@@ -97,9 +97,9 @@ def _tool_state_to_message_tool_calls(state: Dict[int, Dict[str, Any]]) -> List[
 
 
 # ---- OpenAI-compatible calls (non-streaming) ----
-async def call_completions(client: Serverless, *, model: str, prompt: str, endpoint_name: str, **kwargs) -> Dict[str, Any]:
+async def call_completions(client: Serverless, *, model: str, prompt: str, **kwargs) -> Dict[str, Any]:
 
-    endpoint = await client.get_endpoint(name=endpoint_name)
+    endpoint = await client.get_endpoint(name=ENDPOINT_NAME)
 
     payload = {
         "input": {
@@ -113,9 +113,9 @@ async def call_completions(client: Serverless, *, model: str, prompt: str, endpo
     resp = await endpoint.request("/v1/completions", payload, cost=payload["input"]["max_tokens"])
     return resp["response"]
 
-async def call_chat_completions(client: Serverless, *, model: str, messages: List[Dict[str, Any]], endpoint_name: str, **kwargs) -> Dict[str, Any]:
+async def call_chat_completions(client: Serverless, *, model: str, messages: List[Dict[str, Any]], **kwargs) -> Dict[str, Any]:
 
-    endpoint = await client.get_endpoint(name=endpoint_name)
+    endpoint = await client.get_endpoint(name=ENDPOINT_NAME)
 
     payload = {
         "input": {
@@ -132,9 +132,9 @@ async def call_chat_completions(client: Serverless, *, model: str, messages: Lis
     return resp["response"]
 
 # ---- Streaming variants ----
-async def stream_completions(client: Serverless, *, model: str, prompt: str, endpoint_name: str, **kwargs):
+async def stream_completions(client: Serverless, *, model: str, prompt: str, **kwargs):
 
-    endpoint = await client.get_endpoint(name=endpoint_name)
+    endpoint = await client.get_endpoint(name=ENDPOINT_NAME)
 
     payload = {
         "input": {
@@ -150,9 +150,9 @@ async def stream_completions(client: Serverless, *, model: str, prompt: str, end
     resp = await endpoint.request("/v1/completions", payload, cost=payload["input"]["max_tokens"], stream=True)
     return resp["response"]  # async generator
 
-async def stream_chat_completions(client: Serverless, *, model: str, messages: List[Dict[str, Any]], endpoint_name: str, **kwargs):
+async def stream_chat_completions(client: Serverless, *, model: str, messages: List[Dict[str, Any]], **kwargs):
 
-    endpoint = await client.get_endpoint(name=endpoint_name)
+    endpoint = await client.get_endpoint(name=ENDPOINT_NAME)
 
     payload = {
         "input": {
@@ -174,10 +174,9 @@ async def stream_chat_completions(client: Serverless, *, model: str, messages: L
 class APIDemo:
     """Demo and testing functionality for the API client"""
 
-    def __init__(self, client: Serverless, model: str, endpoint_name: str, tool_manager: Optional[ToolManager] = None):
+    def __init__(self, client: Serverless, model: str, tool_manager: Optional[ToolManager] = None):
         self.client = client
         self.model = model
-        self.endpoint_name = endpoint_name
         self.tool_manager = tool_manager or ToolManager()
 
     # ----- Streaming handler -----
@@ -186,16 +185,11 @@ class APIDemo:
         reasoning_content = ""
         printed_reasoning = False
         printed_answer = False
-        finish_reason = None
 
         async for chunk in stream:
             choice = (chunk.get("choices") or [{}])[0]
             delta = choice.get("delta", {})
 
-            # Track finish reason
-            if choice.get("finish_reason"):
-                finish_reason = choice.get("finish_reason")
-
             # reasoning tokens
             rc = delta.get("reasoning_content")
             if rc and show_reasoning:
@@ -225,8 +219,6 @@ class APIDemo:
                 print(f"Reasoning tokens: {len(reasoning_content.split())}")
             if printed_answer:
                 print(f"Response tokens: {len(full_response.split())}")
-            if finish_reason:
-                print(f"Finish reason: {finish_reason}")
 
         return full_response
     
@@ -239,7 +231,6 @@ class APIDemo:
             client=self.client,
             model=self.model,
             prompt=COMPLETIONS_PROMPT,
-            endpoint_name=self.endpoint_name,
             max_tokens=MAX_TOKENS,
             temperature=DEFAULT_TEMPERATURE,
         )
@@ -258,7 +249,6 @@ class APIDemo:
                 client=self.client,
                 model=self.model, 
                 messages=messages,
-                endpoint_name=self.endpoint_name,
                 max_tokens=MAX_TOKENS,
                 temperature=DEFAULT_TEMPERATURE
             )
@@ -271,7 +261,6 @@ class APIDemo:
                 client=self.client,
                 model=self.model, 
                 messages=messages,
-                endpoint_name=self.endpoint_name,
                 max_tokens=MAX_TOKENS,
                 temperature=DEFAULT_TEMPERATURE
             )
@@ -298,7 +287,6 @@ class APIDemo:
                 client=self.client,
                 model=self.model,
                 messages=messages,
-                endpoint_name=self.endpoint_name,
                 tools=minimal_tool,
                 tool_choice="none",
                 max_tokens=10
@@ -324,7 +312,6 @@ class APIDemo:
             client=self.client,
             model=self.model,
             messages=messages,
-            endpoint_name=self.endpoint_name,
             tools=self.tool_manager.get_ls_tool_definition(),
             tool_choice="auto",
             max_tokens=MAX_TOKENS,
@@ -402,7 +389,6 @@ class APIDemo:
             client=self.client,
             model=self.model,
             messages=messages,
-            endpoint_name=self.endpoint_name,
             max_tokens=MAX_TOKENS,
             temperature=DEFAULT_TEMPERATURE,
         )
@@ -441,6 +427,7 @@ class APIDemo:
         print("=" * 60)
         print("INTERACTIVE STREAMING CHAT")
         print("=" * 60)
+        print(f"Using model: {self.model}")
         print("Type 'quit' to exit, 'clear' to clear history")
         print()
 
@@ -467,7 +454,6 @@ class APIDemo:
                     client=self.client,
                     model=self.model, 
                     messages=messages, 
-                    endpoint_name=self.endpoint_name,
                     max_tokens=MAX_TOKENS, 
                     temperature=0.7
                 )
@@ -487,8 +473,8 @@ class APIDemo:
 # ---------------------- CLI ----------------------
 def build_arg_parser() -> argparse.ArgumentParser:
     p = argparse.ArgumentParser(description="Vast vLLM Demo (Serverless SDK)")
-    p.add_argument("--model", default=DEFAULT_MODEL, help=f"Model to use for requests (default: {DEFAULT_MODEL})")
+    p.add_argument("--model", required=True, help="Model to use for requests (required)")
-    p.add_argument("--endpoint", default=ENDPOINT_NAME, help=f"Vast endpoint name (default: {ENDPOINT_NAME})")
+    p.add_argument("--endpoint", default="my-vllm-endpoint", help="Vast endpoint name (default: my-vllm-endpoint)")
 
     modes = p.add_mutually_exclusive_group(required=False)
     modes.add_argument("--completion", action="store_true", help="Test completions endpoint")
@@ -516,14 +502,12 @@ async def main_async():
         print("Please specify exactly one test mode")
         sys.exit(1)
 
-    print("=" * 60)
     print(f"Using model: {args.model}")
-    print(f"Using endpoint: {args.endpoint}")
+    print("=" * 60)
-
 
     try:
         async with Serverless() as client:
-            demo = APIDemo(client, args.model, args.endpoint, ToolManager())
+            demo = APIDemo(client, args.model, ToolManager())
 
             if args.completion:
                 await demo.demo_completions()

workers/openai/server.py

@@ -35,7 +35,6 @@ backend = Backend(
     model_server_url=os.environ["MODEL_SERVER_URL"],
     model_log_file=os.environ["MODEL_LOG"],
     allow_parallel_requests=True,
-    max_wait_time=600.0,
     benchmark_handler=CompletionsHandler(benchmark_runs=3, benchmark_words=256),
     log_actions=[
         *[(LogAction.ModelLoaded, info_msg) for info_msg in MODEL_SERVER_START_LOG_MSG],

workers/openai/worker.py

@@ -0,0 +1,78 @@
+import nltk
+import random
+import os
+
+from vastai import Worker, WorkerConfig, HandlerConfig, LogActionConfig, BenchmarkConfig
+
+# vLLM model configuration
+MODEL_SERVER_URL           = 'http://127.0.0.1'
+MODEL_SERVER_PORT          = 18000
+MODEL_LOG_FILE             = '/var/log/portal/vllm.log'
+MODEL_HEALTHCHECK_ENDPOINT = "/health"
+
+# vLLM-specific log messages
+MODEL_LOAD_LOG_MSG = [
+    "Application startup complete.",
+]
+
+MODEL_ERROR_LOG_MSGS = [
+    "INFO exited: vllm",
+    "RuntimeError: Engine",
+    "Traceback (most recent call last):"
+]
+
+MODEL_INFO_LOG_MSGS = [
+    '"message":"Download'
+]
+
+nltk.download("words")
+WORD_LIST = nltk.corpus.words.words()
+
+
+def completions_benchmark_generator() -> dict:
+    prompt = " ".join(random.choices(WORD_LIST, k=int(250)))
+    model = os.environ.get("MODEL_NAME")
+    if not model:
+        raise ValueError("MODEL_NAME environment variable not set")
+
+    benchmark_data = {
+        "model": model,
+        "prompt": prompt,
+        "temperature": 0.7,
+        "max_tokens": 500,
+    }
+
+    return benchmark_data
+
+worker_config = WorkerConfig(
+    model_server_url=MODEL_SERVER_URL,
+    model_server_port=MODEL_SERVER_PORT,
+    model_log_file=MODEL_LOG_FILE,
+    model_healthcheck_url=MODEL_HEALTHCHECK_ENDPOINT,
+    handlers=[
+        HandlerConfig(
+            route="/v1/completions",
+            workload_calculator= lambda data: data.get("max_tokens", 0),
+            allow_parallel_requests=True,
+            max_queue_time=60.0,
+            benchmark_config=BenchmarkConfig(
+                generator=completions_benchmark_generator,
+                concurrency=100,
+                runs=2
+            )
+        ),
+        HandlerConfig(
+            route="/v1/chat/completions",
+            workload_calculator= lambda data: data.get("max_tokens", 0),
+            allow_parallel_requests=True,
+            max_queue_time=60.0,
+        )
+    ],
+    log_action_config=LogActionConfig(
+        on_load=MODEL_LOAD_LOG_MSG,
+        on_error=MODEL_ERROR_LOG_MSGS,
+        on_info=MODEL_INFO_LOG_MSGS
+    )
+)
+
+Worker(worker_config).run()

workers/tgi/README.md

@@ -1,103 +1,19 @@
-# HuggingFace TGI PyWorker
+This is the base PyWorker for TGI, designed to create PyWorkers that can utilize various LLMs. It offers two primary endpoints:
 
-This is the base PyWorker for HuggingFace Text Generation Inference (TGI) servers. See the [Serverless documentation](https://docs.vast.ai/serverless) for guides and how-to's.
+1. `generate`: Generates the LLM's response to a given prompt in a single request.
+2. `generate_stream`: Streams the LLM's response token by token.
 
-## Instance Setup
+Both endpoints use the following API payload format:
-
-1. Pick a template
-
-This worker is compatible with any TGI backend. We have a template you can use or you can create your own.
-
-- [HuggingFace TGI](https://cloud.vast.ai/?ref_id=62897&creator_id=62897&name=TGI%20(Serverless))
-
-The template can be configured via the template interface. You may want to change the model or startup arguments.
-
-2. Follow the [getting started guide](https://docs.vast.ai/documentation/serverless/quickstart) for help with configuring your serverless setup. For testing, we recommend that you use the default options presented by the web interface.
-
-## Client Setup (Demo)
-
-1. Clone the PyWorker repository to your local machine and install the necessary requirements for running the test client.
-
-```bash
-git clone https://github.com/vast-ai/pyworker
-cd pyworker
-pip install uv
-uv venv -p 3.12
-source .venv/bin/activate
-uv pip install -r requirements.txt
-```
-
-## Using the Test Client
-
-The test client demonstrates both streaming and non-streaming generation using TGI's native API.
-
-First, set your API key as an environment variable:
-
-```bash
-export VAST_API_KEY=<your_api_key>
-```
-
-The `--endpoint` flag is optional. If not provided, it defaults to `my-tgi-endpoint`.
-
-### Generate (Streaming)
-
-Call to `/generate_stream` with streaming response:
-
-```bash
-python -m workers.tgi.client --generate-stream --endpoint <ENDPOINT_NAME>
-```
-
-### Generate (Non-Streaming)
-
-Call to `/generate` with json response:
-
-```bash
-python -m workers.tgi.client --generate --endpoint <ENDPOINT_NAME>
-```
-
-### Interactive Session (Streaming)
-
-Interactive session with streaming responses. Type `quit` to exit.
-
-```bash
-python -m workers.tgi.client --interactive --endpoint <ENDPOINT_NAME>
-```
-
-## API Endpoints
-
-TGI provides two primary endpoints:
-
-### Generate (Non-Streaming)
-
-`/generate` - Returns the complete response in a single request.
 
 ```json
 {
-  "inputs": "Your prompt here",
+  "inputs": "PROMPT",
   "parameters": {
-    "max_new_tokens": 1024,
+    "max_new_tokens": 250
-    "temperature": 0.7,
-    "return_full_text": false
   }
 }
 ```
 
-### Generate Stream (Streaming)
+Note that the max_new_tokens parameter, rather than the prompt size, impacts performance. For example, if an
-
+instance is benchmarked to process 100 tokens per second, a request with max_new_tokens = 200 will take
-`/generate_stream` - Streams the response token by token.
+approximately 2 seconds to complete.
-
-```json
-{
-  "inputs": "Your prompt here",
-  "parameters": {
-    "max_new_tokens": 1024,
-    "temperature": 0.7,
-    "do_sample": true,
-    "return_full_text": false
-  }
-}
-```
-
-## Performance Notes
-
-The `max_new_tokens` parameter (not the prompt size) primarily impacts performance. For example, if an instance is benchmarked to process 100 tokens per second, a request with `max_new_tokens = 200` will take approximately 2 seconds to complete.

workers/tgi/client.py

@@ -1,222 +1,61 @@
-import logging
-import json
-import os
-import sys
-import argparse
-
 from vastai import Serverless
 import asyncio
 
-# ---------------------- Logging ----------------------
+ENDPOINT_NAME = "my-tgi-endpoint" # Change this to match your endpoint name
-logging.basicConfig(
-    level=logging.DEBUG,
-    format="%(asctime)s[%(levelname)-5s] %(message)s",
-    datefmt="%Y-%m-%d %H:%M:%S",
-)
-log = logging.getLogger(__file__)
-
-# ---------------------- Defaults ----------------------
-DEFAULT_PROMPT = "Think step by step: Tell me about the Python programming language."
-
-ENDPOINT_NAME = "TGI-Prod2"       # change this to your TGI endpoint name
 MAX_TOKENS = 1024
-DEFAULT_TEMPERATURE = 0.7
+PROMPT = "Think step by step: Tell me about the Python programming language."
 
-
+async def call_generate(client: Serverless) -> None:
-# ---------------------- API Calls ----------------------
+    endpoint = await client.get_endpoint(name=ENDPOINT_NAME)
-async def call_generate(client: Serverless, *, endpoint_name: str, prompt: str, **kwargs) -> dict:
-    """Non-streaming generation via /generate endpoint"""
-    endpoint = await client.get_endpoint(name=endpoint_name)
 
     payload = {
-        "inputs": prompt,
+        "inputs": PROMPT,
         "parameters": {
-            "max_new_tokens": kwargs.get("max_tokens", MAX_TOKENS),
+            "max_new_tokens": MAX_TOKENS,
-            "temperature": kwargs.get("temperature", DEFAULT_TEMPERATURE),
+            "temperature": 0.7,
-            "return_full_text": False,
+            "return_full_text": False
         }
     }
-    log.debug("POST /generate %s", json.dumps(payload)[:500])
+
-    resp = await endpoint.request("/generate", payload, cost=payload["parameters"]["max_new_tokens"])
+    resp = await endpoint.request("/generate", payload, cost=MAX_TOKENS)
-    return resp["response"]
+
+    print(resp["response"]["generated_text"])
 
 
-async def call_generate_stream(client: Serverless, *, endpoint_name: str, prompt: str, **kwargs):
+async def call_generate_stream(client: Serverless) -> None:
-    """Streaming generation via /generate_stream endpoint"""
+    endpoint = await client.get_endpoint(name=ENDPOINT_NAME)
-    endpoint = await client.get_endpoint(name=endpoint_name)
 
     payload = {
-        "inputs": prompt,
+        "inputs": PROMPT,
         "parameters": {
-            "max_new_tokens": kwargs.get("max_tokens", MAX_TOKENS),
+            "max_new_tokens": MAX_TOKENS,
-            "temperature": kwargs.get("temperature", DEFAULT_TEMPERATURE),
+            "temperature": 0.7,
             "do_sample": True,
             "return_full_text": False,
         }
     }
-    log.debug("STREAM /generate_stream %s", json.dumps(payload)[:500])
+
     resp = await endpoint.request(
         "/generate_stream",
         payload,
-        cost=payload["parameters"]["max_new_tokens"],
+        cost=MAX_TOKENS,
         stream=True,
     )
-    return resp["response"]  # async generator
+    stream = resp["response"]
 
-
-# ---------------------- Demo Runner ----------------------
-class APIDemo:
-    """Demo and testing functionality for the TGI API client"""
-
-    def __init__(self, client: Serverless, endpoint_name: str):
-        self.client = client
-        self.endpoint_name = endpoint_name
-
-    async def handle_streaming_response(self, stream) -> str:
-        """Process streaming response and print tokens"""
-        full_response = ""
     printed_answer = False
-
     async for event in stream:
         tok = (event.get("token") or {}).get("text")
         if tok:
             if not printed_answer:
                 printed_answer = True
-                    print("\n💬 Response: ", end="", flush=True)
+                print("Answer:\n", end="", flush=True)
             print(tok, end="", flush=True)
-                full_response += tok
 
-        print()  # newline
+async def main():
-        if printed_answer:
-            print(f"\nStreaming completed. Response tokens: {len(full_response.split())}")
-
-        return full_response
-
-    async def demo_generate(self) -> None:
-        """Demo non-streaming generation"""
-        print("=" * 60)
-        print("GENERATE DEMO (NON-STREAMING)")
-        print("=" * 60)
-
-        response = await call_generate(
-            client=self.client,
-            endpoint_name=self.endpoint_name,
-            prompt=DEFAULT_PROMPT,
-            max_tokens=MAX_TOKENS,
-            temperature=DEFAULT_TEMPERATURE,
-        )
-        
-        print(f"\n💬 Response: {response.get('generated_text', '')}")
-        print(f"\nFull Response:\n{json.dumps(response, indent=2)}")
-
-    async def demo_generate_stream(self) -> None:
-        """Demo streaming generation"""
-        print("=" * 60)
-        print("GENERATE DEMO (STREAMING)")
-        print("=" * 60)
-
-        stream = await call_generate_stream(
-            client=self.client,
-            endpoint_name=self.endpoint_name,
-            prompt=DEFAULT_PROMPT,
-            max_tokens=MAX_TOKENS,
-            temperature=DEFAULT_TEMPERATURE,
-        )
-
-        try:
-            await self.handle_streaming_response(stream)
-        except Exception as e:
-            log.error("\nError during streaming: %s", e, exc_info=True)
-
-    async def interactive_chat(self) -> None:
-        """Interactive session with streaming generation"""
-        print("=" * 60)
-        print("INTERACTIVE STREAMING SESSION")
-        print("=" * 60)
-        print(f"Using endpoint: {self.endpoint_name}")
-        print("Type 'quit' to exit")
-        print()
-
-        while True:
-            try:
-                user_input = input("You: ").strip()
-
-                if user_input.lower() == "quit":
-                    print("👋 Goodbye!")
-                    break
-                elif not user_input:
-                    continue
-
-                print("Assistant: ", end="", flush=True)
-                stream = await call_generate_stream(
-                    client=self.client,
-                    endpoint_name=self.endpoint_name,
-                    prompt=user_input,
-                    max_tokens=MAX_TOKENS,
-                    temperature=DEFAULT_TEMPERATURE,
-                )
-
-                full_response = ""
-                async for event in stream:
-                    tok = (event.get("token") or {}).get("text")
-                    if tok:
-                        print(tok, end="", flush=True)
-                        full_response += tok
-                print()  # newline
-
-            except KeyboardInterrupt:
-                print("\n👋 Session interrupted. Goodbye!")
-                break
-            except Exception as e:
-                log.error("\nError: %s", e)
-                continue
-
-
-# ---------------------- CLI ----------------------
-def build_arg_parser() -> argparse.ArgumentParser:
-    p = argparse.ArgumentParser(description="Vast TGI Demo (Serverless SDK)")
-    p.add_argument("--endpoint", default=ENDPOINT_NAME, help=f"Vast endpoint name (default: {ENDPOINT_NAME})")
-
-    modes = p.add_mutually_exclusive_group(required=False)
-    modes.add_argument("--generate", action="store_true", help="Test generate endpoint (non-streaming)")
-    modes.add_argument("--generate-stream", action="store_true", help="Test generate endpoint with streaming")
-    modes.add_argument("--interactive", action="store_true", help="Start interactive streaming session")
-    return p
-
-
-async def main_async():
-    args = build_arg_parser().parse_args()
-
-    selected = sum([args.generate, args.generate_stream, args.interactive])
-    if selected == 0:
-        print("Please specify exactly one test mode:")
-        print("  --generate        : Test generate endpoint (non-streaming)")
-        print("  --generate-stream : Test generate endpoint with streaming")
-        print("  --interactive     : Start interactive streaming session")
-        print(f"\nExample: python {os.path.basename(sys.argv[0])} --generate-stream --endpoint my-tgi-endpoint")
-        sys.exit(1)
-    elif selected > 1:
-        print("Please specify exactly one test mode")
-        sys.exit(1)
-
-    print("=" * 60)
-    print(f"Using endpoint: {args.endpoint}")
-
-    try:
     async with Serverless() as client:
-            demo = APIDemo(client, args.endpoint)
+        await call_generate(client)
-
+        await call_generate_stream(client)
-            if args.generate:
-                await demo.demo_generate()
-            elif args.generate_stream:
-                await demo.demo_generate_stream()
-            elif args.interactive:
-                await demo.interactive_chat()
-
-    except Exception as e:
-        log.error("Error during test: %s", e, exc_info=True)
-        sys.exit(1)
-
 
 if __name__ == "__main__":
-    asyncio.run(main_async())
+    asyncio.run(main())

workers/tgi/worker.py

@@ -0,0 +1,76 @@
+import nltk
+import random
+
+from vastai import Worker, WorkerConfig, HandlerConfig, LogActionConfig, BenchmarkConfig
+
+# TGI model configuration
+MODEL_SERVER_URL           = 'http://0.0.0.0'
+MODEL_SERVER_PORT          = 5001
+MODEL_LOG_FILE             = "/workspace/infer.log"
+MODEL_HEALTHCHECK_ENDPOINT = "/health"
+
+# TGI-specific log messages
+MODEL_LOAD_LOG_MSG = [
+    '"message":"Connected","target":"text_generation_router"',
+    '"message":"Connected","target":"text_generation_router::server"',
+]
+
+MODEL_ERROR_LOG_MSGS = [
+    "Error: WebserverFailed",
+    "Error: DownloadError",
+    "Error: ShardCannotStart",
+]
+
+MODEL_INFO_LOG_MSGS = [
+    '"message":"Download'
+]
+
+nltk.download("words")
+WORD_LIST = nltk.corpus.words.words()
+
+
+def benchmark_generator() -> dict:
+    prompt = " ".join(random.choices(WORD_LIST, k=int(250)))
+
+    benchmark_data = {
+        "inputs": prompt,
+        "parameters": {
+            "max_new_tokens": 128,
+            "temperature": 0.7,
+            "return_full_text": False
+        }
+    }
+
+    return benchmark_data
+
+worker_config = WorkerConfig(
+    model_server_url=MODEL_SERVER_URL,
+    model_server_port=MODEL_SERVER_PORT,
+    model_log_file=MODEL_LOG_FILE,
+    model_healthcheck_url=MODEL_HEALTHCHECK_ENDPOINT,
+    handlers=[
+        HandlerConfig(
+            route="/generate",
+            allow_parallel_requests=True,
+            max_queue_time=60.0,
+            benchmark_config=BenchmarkConfig(
+                generator=benchmark_generator,
+                concurrency=50
+            ),
+            workload_calculator= lambda x: x["parameters"]["max_new_tokens"]
+        ),
+        HandlerConfig(
+            route="/generate_stream",
+            allow_parallel_requests=True,
+            max_queue_time=60.0,
+            workload_calculator= lambda x: x["parameters"]["max_new_tokens"]
+        )
+    ],
+    log_action_config=LogActionConfig(
+        on_load=MODEL_LOAD_LOG_MSG,
+        on_error=MODEL_ERROR_LOG_MSGS,
+        on_info=MODEL_INFO_LOG_MSGS
+    )
+)
+
+Worker(worker_config).run()

workers/wan/worker.py

@@ -0,0 +1,288 @@
+import random
+import sys
+
+from vastai import Worker, WorkerConfig, HandlerConfig, LogActionConfig, BenchmarkConfig
+
+# ComyUI model configuration
+MODEL_SERVER_URL           = 'http://127.0.0.1'
+MODEL_SERVER_PORT          = 18288
+MODEL_LOG_FILE             = '/var/log/portal/comfyui.log'
+MODEL_HEALTHCHECK_ENDPOINT = "/health"
+
+# ComyUI-specific log messages
+MODEL_LOAD_LOG_MSG = [
+    "To see the GUI go to: "
+]
+
+MODEL_ERROR_LOG_MSGS = [
+    "MetadataIncompleteBuffer",
+    "Value not in list: ",
+    "[ERROR] Provisioning Script failed"
+]
+
+MODEL_INFO_LOG_MSGS = [
+    '"message":"Downloading'
+]
+
+benchmark_prompts = [
+    "Cartoon hoodie hero; orc, anime cat, bunny; black goo; buff; vector on white.",
+    "Cozy farming-game scene with fine details.",
+    "2D vector child with soccer ball; airbrush chrome; swagger; antique copper.",
+    "Realistic futuristic downtown of low buildings at sunset.",
+    "Perfect wave front view; sunny seascape; ultra-detailed water; artful feel.",
+    "Clear cup with ice, fruit, mint; creamy swirls; fluid-sim CGI; warm glow.",
+    "Male biker with backpack on motorcycle; oilpunk; award-worthy magazine cover.",
+    "Collage for textile; surreal cartoon cat in cap/jeans before poster; crisp.",
+    "Medieval village inside glass sphere; volumetric light; macro focus.",
+    "Iron Man with glowing axe; mecha sci-fi; jungle scene; dynamic light.",
+    "Pope Francis DJ in leather jacket, mixing on giant console; dramatic.",
+]
+
+benchmark_dataset = [
+    {
+        "input": {
+            "workflow_json": {
+                "90": {
+                    "inputs": {
+                    "clip_name": "umt5_xxl_fp8_e4m3fn_scaled.safetensors",
+                    "type": "wan",
+                    "device": "default"
+                    },
+                    "class_type": "CLIPLoader",
+                    "_meta": {
+                    "title": "Load CLIP"
+                    }
+                },
+                "91": {
+                    "inputs": {
+                    "text": "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走，裸露，NSFW",
+                    "clip": [
+                        "90",
+                        0
+                    ]
+                    },
+                    "class_type": "CLIPTextEncode",
+                    "_meta": {
+                    "title": "CLIP Text Encode (Negative Prompt)"
+                    }
+                },
+                "92": {
+                    "inputs": {
+                    "vae_name": "wan_2.1_vae.safetensors"
+                    },
+                    "class_type": "VAELoader",
+                    "_meta": {
+                    "title": "Load VAE"
+                    }
+                },
+                "93": {
+                    "inputs": {
+                    "shift": 8.000000000000002,
+                    "model": [
+                        "101",
+                        0
+                    ]
+                    },
+                    "class_type": "ModelSamplingSD3",
+                    "_meta": {
+                    "title": "ModelSamplingSD3"
+                    }
+                },
+                "94": {
+                    "inputs": {
+                    "shift": 8,
+                    "model": [
+                        "102",
+                        0
+                    ]
+                    },
+                    "class_type": "ModelSamplingSD3",
+                    "_meta": {
+                    "title": "ModelSamplingSD3"
+                    }
+                },
+                "95": {
+                    "inputs": {
+                    "add_noise": "disable",
+                    "noise_seed": 0,
+                    "steps": 20,
+                    "cfg": 3.5,
+                    "sampler_name": "euler",
+                    "scheduler": "simple",
+                    "start_at_step": 10,
+                    "end_at_step": 10000,
+                    "return_with_leftover_noise": "disable",
+                    "model": [
+                        "94",
+                        0
+                    ],
+                    "positive": [
+                        "99",
+                        0
+                    ],
+                    "negative": [
+                        "91",
+                        0
+                    ],
+                    "latent_image": [
+                        "96",
+                        0
+                    ]
+                    },
+                    "class_type": "KSamplerAdvanced",
+                    "_meta": {
+                    "title": "KSampler (Advanced)"
+                    }
+                },
+                "96": {
+                    "inputs": {
+                    "add_noise": "enable",
+                    "noise_seed": "__RANDOM_INT__",
+                    "steps": 20,
+                    "cfg": 3.5,
+                    "sampler_name": "euler",
+                    "scheduler": "simple",
+                    "start_at_step": 0,
+                    "end_at_step": 10,
+                    "return_with_leftover_noise": "enable",
+                    "model": [
+                        "93",
+                        0
+                    ],
+                    "positive": [
+                        "99",
+                        0
+                    ],
+                    "negative": [
+                        "91",
+                        0
+                    ],
+                    "latent_image": [
+                        "104",
+                        0
+                    ]
+                    },
+                    "class_type": "KSamplerAdvanced",
+                    "_meta": {
+                    "title": "KSampler (Advanced)"
+                    }
+                },
+                "97": {
+                    "inputs": {
+                    "samples": [
+                        "95",
+                        0
+                    ],
+                    "vae": [
+                        "92",
+                        0
+                    ]
+                    },
+                    "class_type": "VAEDecode",
+                    "_meta": {
+                    "title": "VAE Decode"
+                    }
+                },
+                "98": {
+                    "inputs": {
+                    "filename_prefix": "video/ComfyUI",
+                    "format": "auto",
+                    "codec": "auto",
+                    "video": [
+                        "100",
+                        0
+                    ]
+                    },
+                    "class_type": "SaveVideo",
+                    "_meta": {
+                    "title": "Save Video"
+                    }
+                },
+                "99": {
+                    "inputs": {
+                    "text":prompt,
+                    "clip": [
+                        "90",
+                        0
+                    ]
+                    },
+                    "class_type": "CLIPTextEncode",
+                    "_meta": {
+                    "title": "CLIP Text Encode (Positive Prompt)"
+                    }
+                },
+                "100": {
+                    "inputs": {
+                    "fps": 16,
+                    "images": [
+                        "97",
+                        0
+                    ]
+                    },
+                    "class_type": "CreateVideo",
+                    "_meta": {
+                    "title": "Create Video"
+                    }
+                },
+                "101": {
+                    "inputs": {
+                    "unet_name": "wan2.2_t2v_high_noise_14B_fp8_scaled.safetensors",
+                    "weight_dtype": "default"
+                    },
+                    "class_type": "UNETLoader",
+                    "_meta": {
+                    "title": "Load Diffusion Model"
+                    }
+                },
+                "102": {
+                    "inputs": {
+                    "unet_name": "wan2.2_t2v_low_noise_14B_fp8_scaled.safetensors",
+                    "weight_dtype": "default"
+                    },
+                    "class_type": "UNETLoader",
+                    "_meta": {
+                    "title": "Load Diffusion Model"
+                    }
+                },
+                "104": {
+                    "inputs": {
+                    "width": 640,
+                    "height": 640,
+                    "length": 81,
+                    "batch_size": 1
+                    },
+                    "class_type": "EmptyHunyuanLatentVideo",
+                    "_meta": {
+                    "title": "EmptyHunyuanLatentVideo"
+                    }
+                }
+            }
+        }
+    } for prompt in benchmark_prompts
+]
+
+worker_config = WorkerConfig(
+    model_server_url=MODEL_SERVER_URL,
+    model_server_port=MODEL_SERVER_PORT,
+    model_log_file=MODEL_LOG_FILE,
+    model_healthcheck_url=MODEL_HEALTHCHECK_ENDPOINT,
+    handlers=[
+        HandlerConfig(
+            route="/generate/sync",
+            allow_parallel_requests=False,
+            max_queue_time=10.0,
+            benchmark_config=BenchmarkConfig(
+                dataset=benchmark_dataset,
+                runs=1
+            ),
+            workload_calculator= lambda _ : 10000.0
+        )
+    ],
+    log_action_config=LogActionConfig(
+        on_load=MODEL_LOAD_LOG_MSG,
+        on_error=MODEL_ERROR_LOG_MSGS,
+        on_info=MODEL_INFO_LOG_MSGS
+    )
+)
+
+Worker(worker_config).run()