Move model log line

Update client scripts

.gitignore

@@ -2,4 +2,5 @@
 .envrc
 __pycache__
 bin/
-lib64
+lib64
+.venv

README.md

@@ -39,11 +39,12 @@ reporting these metrics to the autoscaler.
 
 If you are using a Vast.ai template that includes PyWorker integration (marked as autoscaler compatible), it should work out of the box. The template will typically start the appropriate PyWorker server automatically. Here's a few:
 
-*   **TGI (Text Generation Inference):** [Vast.ai Template](https://cloud.vast.ai?ref_id=140778&template_id=72d8dcb41ea3a58e06c741e2c725bc00)
+*   **vLLM:** [Vast.ai Template](https://cloud.vast.ai?ref_id=62897&template_id=63ae93902bf3978bea033782592b784d)
-*   **ComfyUI:** [Vast.ai Template](https://cloud.vast.ai?ref_id=140778&template_id=ad72c8bf7cf695c3c9ddf0eaf6da0447)
+*   **TGI (Text Generation Inference):** [Vast.ai Template](https://cloud.vast.ai?ref_id=62897&template_id=6fa6bd5bdf5f0df63db80e40b086037d)
+*   **ComfyUI:** [Vast.ai Template](https://cloud.vast.ai?ref_id=62897&template_id=e6748878ba688e765e3e9fca29541938)
 
 Currently available workers:
-*   `hello_world`: A simple example worker for a basic LLM server.
+*   `openai`: A simple example worker for a basic vLLM server.
 *   `comfyui`: A worker for the ComfyUI image generation backend.
 *   `tgi`: A worker for the Text Generation Inference backend.
 

lib/backend.py

@@ -30,7 +30,7 @@ from lib.data_types import (
     BenchmarkResult
 )
 
-VERSION = "0.1.0"
+VERSION = "0.2.0"
 
 MSG_HISTORY_LEN = 100
 log = logging.getLogger(__file__)
@@ -66,10 +66,17 @@ class Backend:
     unsecured: bool = dataclasses.field(
         default_factory=lambda: bool(strtobool(os.environ.get("UNSECURED", "false"))),
     )
+    report_addr: str = dataclasses.field(
+        default_factory=lambda: os.environ.get("REPORT_ADDR", "https://run.vast.ai")
+    )
+    mtoken: str = dataclasses.field(
+        default_factory=lambda: os.environ.get("MASTER_TOKEN", "")
+    )
 
     def __post_init__(self):
         self.metrics = Metrics()
         self.metrics._set_version(self.version)
+        self.metrics._set_mtoken(self.mtoken)
         self._total_pubkey_fetch_errors = 0
         self._pubkey = self._fetch_pubkey()
         self.__start_healthcheck: bool = False
@@ -104,23 +111,19 @@ class Backend:
 
     #######################################Private#######################################
     def _fetch_pubkey(self):
-        command = ["curl", "-X", "GET", "https://run.vast.ai/pubkey/"]
+        report_addr = self.report_addr.rstrip("/")
-        result = subprocess.check_output(command, universal_newlines=True)
+        command = ["curl", "-X", "GET", f"{report_addr}/pubkey/"]
-        log.debug("public key:")
+        try:
-        log.debug(result)
+            result = subprocess.check_output(command, universal_newlines=True)
-        key = None
+            log.debug("public key:")
-        for _ in range(5):
+            log.debug(result)
-            try:
+            key = RSA.import_key(result)
-                key = RSA.import_key(result)
+            if key is not None:
-                break
+                return key
-            except ValueError as e:
+        except (ValueError , subprocess.CalledProcessError) as e:
-                log.debug(f"Error downloading key: {e}")
+            log.debug(f"Error downloading key: {e}")
-                time.sleep(15)
+        self.backend_errored("Failed to get autoscaler pubkey")
-        if key is None:
+       
-            self._total_pubkey_fetch_errors += 1
-            if self._total_pubkey_fetch_errors >= MAX_PUBKEY_FETCH_ATTEMPTS:
-                self.backend_errored("Failed to get autoscaler pubkey")
-        return key
 
     async def __handle_request(
         self,
@@ -286,7 +289,7 @@ 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 +299,7 @@ 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:]
@@ -315,10 +318,10 @@ class Backend:
                 with open(BENCHMARK_INDICATOR_FILE, "r") as f:
                     log.debug("already ran benchmark")
                     # trigger model load
-                    payload = self.benchmark_handler.make_benchmark_payload()
+                    # payload = self.benchmark_handler.make_benchmark_payload()
-                    _ = await self.__call_api(
+                    # _ = await self.__call_api(
-                        handler=self.benchmark_handler, payload=payload
+                    #     handler=self.benchmark_handler, payload=payload
-                    )
+                    # )
                     return float(f.readline())
             except FileNotFoundError:
                 pass
@@ -393,7 +396,7 @@ class Backend:
                         )
                         # some backends need a few seconds after logging successful startup before
                         # they can begin accepting requests
-                        await sleep(5)
+                        # await sleep(5)
                         try:
                             max_throughput = await run_benchmark()
                             self.__start_healthcheck = True
@@ -414,7 +417,7 @@ class Backend:
 
         async def tail_log():
             log.debug(f"tailing file: {self.model_log_file}")
-            async with await open_file(self.model_log_file) as f:
+            async with await open_file(self.model_log_file, encoding='utf-8', errors='ignore') as f:
                 while True:
                     line = await f.readline()
                     if line:

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]):
@@ -286,6 +286,7 @@ class AutoScalerData:
     """Data that is reported to autoscaler"""
 
     id: int
+    mtoken: str
     version: str
     loadtime: float
     cur_load: float

lib/metrics.py

@@ -28,6 +28,7 @@ def get_url() -> str:
 @dataclass
 class Metrics:
     version: str = "0"
+    mtoken: str = ""
     last_metric_update: float = 0.0
     last_request_served: float = 0.0
     update_pending: bool = False
@@ -142,12 +143,16 @@ class Metrics:
     def _set_version(self, version: str) -> None:
         self.version = version
 
+    def _set_mtoken(self, mtoken: str) -> None:
+        self.mtoken = mtoken
+
     #######################################Private#######################################
 
     async def __send_delete_requests_and_reset(self):
         async def post(report_addr: str, idxs: list[int], success_flag: bool) -> bool:
             data = {
                 "worker_id": self.id,
+                "mtoken": self.mtoken,
                 "request_idxs": idxs,
                 "success": success_flag,
             }
@@ -180,6 +185,10 @@ class Metrics:
             return  # nothing to do
 
         for report_addr in self.report_addr:
+            # TODO: Add a Redis subscriber queue for delete_requests
+            if report_addr == "https://cloud.vast.ai/api/v0":
+                # Patch: ignore the Redis API report_addr
+                continue
             sent_success = True
             sent_failed  = True
 
@@ -205,6 +214,7 @@ class Metrics:
         def compute_autoscaler_data() -> AutoScalerData:
             return AutoScalerData(
                 id=self.id,
+                mtoken=self.mtoken,
                 version=self.version,
                 loadtime=(loadtime_snapshot or 0.0), 
                 new_load=self.model_metrics.workload_processing,
@@ -224,17 +234,25 @@ class Metrics:
 
         async def send_data(report_addr: str) -> bool:
             data = compute_autoscaler_data()
-            full_path = report_addr.rstrip("/") + "/worker_status/"
+            log_data = asdict(data)
+            def obfuscate(secret: str) -> str:
+                if secret is None:
+                    return ""
+                return secret[:7] + "..." if len(secret) > 7 else ("*" * len(secret))
+            
+            log_data["mtoken"] = obfuscate(log_data.get("mtoken"))
             log.debug(
                 "\n".join(
                     [
                         "#" * 60,
                         f"sending data to autoscaler",
-                        f"{json.dumps((asdict(data)), indent=2)}",
+                        f"{json.dumps(log_data, indent=2)}",
                         "#" * 60,
                     ]
                 )
             )
+
+            full_path = report_addr.rstrip("/") + "/worker_status/"
             for attempt in range(1, 4):
                 try:
                     session = await self.http()

requirements.txt

@@ -8,3 +8,4 @@ Requests~=2.32
 transformers~=4.52
 utils==1.0.*
 hf_transfer>=0.1.9
+vastai-sdk>=0.2.0

start_server.sh

@@ -9,7 +9,7 @@ ENV_PATH="$WORKSPACE_DIR/worker-env"
 DEBUG_LOG="$WORKSPACE_DIR/debug.log"
 PYWORKER_LOG="$WORKSPACE_DIR/pyworker.log"
 
-REPORT_ADDR="${REPORT_ADDR:-https://cloud.vast.ai/api/v0,https://run.vast.ai}"
+REPORT_ADDR="${REPORT_ADDR:-https://run.vast.ai}"
 USE_SSL="${USE_SSL:-true}"
 WORKER_PORT="${WORKER_PORT:-3000}"
 mkdir -p "$WORKSPACE_DIR"
@@ -41,6 +41,14 @@ echo_var DEBUG_LOG
 echo_var PYWORKER_LOG
 echo_var MODEL_LOG
 
+# if instance is rebooted, we want to clear out the log file so pyworker doesn't read lines
+# 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 $MODEL_LOG.old"
+    cat "$MODEL_LOG" >> "$MODEL_LOG.old" 
+    : > "$MODEL_LOG"
+fi
+
 # Populate /etc/environment with quoted values
 if ! grep -q "VAST" /etc/environment; then
     env -0 | grep -zEv "^(HOME=|SHLVL=)|CONDA" | while IFS= read -r -d '' line; do
@@ -124,9 +132,5 @@ cd "$SERVER_DIR"
 
 echo "launching PyWorker server"
 
-# if instance is rebooted, we want to clear out the log file so pyworker doesn't read lines
-# from the run prior to reboot. past logs are saved in $MODEL_LOG.old for debugging only
-[ -e "$MODEL_LOG" ] && cat "$MODEL_LOG" >> "$MODEL_LOG.old" && : > "$MODEL_LOG"
-
 (python3 -m "workers.$BACKEND.server" |& tee -a "$PYWORKER_LOG") &
 echo "launching PyWorker server done"

workers/comfyui-json/client.py

@@ -1,155 +1,35 @@
-import logging
+from .data_types import count_workload
 import uuid
 import random
-from urllib.parse import urljoin
+import asyncio
-import json
+import random
 
-import requests
+from vastai import Serverless
 
-from lib.test_utils import print_truncate_res
+async def main():
-from utils.endpoint_util import Endpoint
+    async with Serverless() as client:
-from utils.ssl import get_cert_file_path
+        endpoint = await client.get_endpoint(name="my-comfy-endpoint") # Change this to your endpoint name
-from .data_types import count_workload
 
-logging.basicConfig(
+        payload = {
-    level=logging.DEBUG,
+            "input": {
-    format="%(asctime)s[%(levelname)-5s] %(message)s",
+                "request_id": str(uuid.uuid4()),
-    datefmt="%Y-%m-%d %H:%M:%S",
+                "modifier": "Text2Image",
-)
+                "modifications": {
-log = logging.getLogger(__file__)
+                    "prompt": "a beautiful landscape with mountains and lakes",
-
+                    "width": 1024,
-
+                    "height": 1024,
-def call_text2image_workflow(
+                    "steps": 20,
-    endpoint_group_name: str, api_key: str, server_url: str
+                    "seed": random.randint(0, 2**32 - 1)
-) -> None:
+                },
-    """Simple Text2Image using the new modifier-based approach"""
+                "workflow_json": {}  # Empty since using modifier approach
-    
+            }
-    def make_request(url: str, payload: dict, timeout: int = None, verify=True, context: str = "request"):
-        """Helper function for making requests with consistent error handling"""
-        try:
-            response = requests.post(
-                url,
-                json=payload,
-                timeout=timeout,
-                verify=verify
-            )
-            
-            response.raise_for_status()
-            return response.json()
-            
-        except requests.exceptions.HTTPError as http_err:
-            log.error(f"HTTP error occurred during {context}: {http_err}")
-            log.error(f"Status Code: {response.status_code}")
-            log.error("Response content:", response.text)
-            return None
-        except requests.exceptions.Timeout:
-            log.error(f"Timeout occurred during {context}: {url}")
-            return None
-        except requests.exceptions.ConnectionError:
-            log.error(f"Connection error occurred during {context}: {url}")
-            return None
-        except json.JSONDecodeError as json_err:
-            log.error(f"Failed to decode JSON response during {context}: {json_err}")
-            if 'response' in locals():
-                print("Response content:", response.text)
-            return None
-        except Exception as err:
-            log.error(f"An unexpected error occurred during {context}: {err}")
-            if 'response' in locals():
-                log.error("Response content (if available):", response.text)
-            return None
-    
-    WORKER_ENDPOINT = "/generate/sync"
-
-    # This worker has concurrency = 1.  All workloads have cost value 1.0
-    COST = count_workload()
-    
-    # Route to get worker URL
-    route_payload = {
-        "endpoint": endpoint_group_name,
-        "api_key": api_key,
-        "cost": COST,
-    }
-    
-    # First request - get routing information
-    route_response = make_request(
-        url=urljoin(server_url, "/route/"),
-        payload=route_payload,
-        timeout=4,
-        context="route request"
-    )
-    
-    if route_response is None:
-        return None
-    
-    if "url" not in route_response or not route_response["url"]:
-        log.error("Error: No worker in 'Ready' state. Please wait while the serverless engine removes errored workers or finishes loading new workers.")
-        return None
-    
-    if "status" in route_response:
-        print(f"Autoscaler status: {route_response['status']}")
-        return None
-    
-    # Extract data from route response
-    url = route_response["url"]
-    auth_data = dict(
-        signature=route_response["signature"],
-        cost=route_response["cost"],
-        endpoint=route_response["endpoint"],
-        reqnum=route_response["reqnum"],
-        url=route_response["url"],
-    )
-    
-    # Build the payload for the worker request
-    worker_payload = {
-        "input": {
-            "request_id": str(uuid.uuid4()),
-            "modifier": "Text2Image",
-            "modifications": {
-                "prompt": "a beautiful landscape with mountains and lakes",
-                "width": 1024,
-                "height": 1024,
-                "steps": 20,
-                "seed": random.randint(0, 2**32 - 1)
-            },
-            "workflow_json": {}  # Empty since using modifier approach
         }
-    }
+        
-    
+        response = await endpoint.request("/generate/sync", payload, cost=count_workload())
-    req_data = dict(payload=worker_payload, auth_data=auth_data)
-    worker_url = urljoin(url, WORKER_ENDPOINT)
-    print(f"url: {worker_url}")
-    
-    # Second request - call the worker endpoint
-    worker_response = make_request(
-        url=worker_url,
-        payload=req_data,
-        verify=get_cert_file_path(),
-        context="worker request"
-    )
-    
-    return worker_response
 
+        # Get the file from the path on the local machine using SCP or SFTP
+        # or configure S3 to upload to cloud storage.
+        print(response["response"]["output"][0]["local_path"])
 
 if __name__ == "__main__":
-    from lib.test_utils import test_args
+    asyncio.run(main())
-
-    args = test_args.parse_args()
-    endpoint_api_key = Endpoint.get_endpoint_api_key(
-        endpoint_name=args.endpoint_group_name,
-        account_api_key=args.api_key,
-        instance=args.instance,
-    )
-
-    if endpoint_api_key:
-        result = call_text2image_workflow(
-            api_key=endpoint_api_key,
-            endpoint_group_name=args.endpoint_group_name,
-            server_url=args.server_url,
-        )
-        if result is None:
-            log.error("Text2Image workflow failed")
-        else:   
-            print(result)
-    else:
-        log.error(f"Failed to get API key for endpoint {args.endpoint_group_name}")

workers/comfyui/client.py

@@ -7,20 +7,13 @@ from lib.test_utils import print_truncate_res
 from utils.endpoint_util import Endpoint
 from utils.ssl import get_cert_file_path
 
-"""
+from vastai import Serverless
-NOTE: this client example uses a custom comfy workflow compatible with SD3 only
-"""
-logging.basicConfig(
-    level=logging.DEBUG,
-    format="%(asctime)s[%(levelname)-5s] %(message)s",
-    datefmt="%Y-%m-%d %H:%M:%S",
-)
-log = logging.getLogger(__file__)
 
 
-def call_default_workflow(
+ENDPOINT_NAME = "my-comfyui-endpoint"
-    endpoint_group_name: str, api_key: str, server_url: str
+COST = 100 # Use a constant cost for image generation
-) -> None:
+
+def call_default_workflow(client: Serverless) -> None:
     WORKER_ENDPOINT = "/prompt"
     COST = 100
     route_payload = {
@@ -82,6 +75,7 @@ def call_custom_workflow_for_sd3(
         endpoint=message["endpoint"],
         reqnum=message["reqnum"],
         url=message["url"],
+        request_idx=message["request_idx"],
     )
     workflow = {
         "3": {

workers/openai/client.py

@@ -1,14 +1,15 @@
 import logging
-import sys
 import json
+import os
+import sys
 import subprocess
-from urllib.parse import urljoin
+import argparse
-from typing import Dict, Any, Optional, Iterator, Union, List
+from typing import Any, Dict, List, Optional
-import requests
-from utils.endpoint_util import Endpoint
-from utils.ssl import get_cert_file_path
-from .data_types.client import CompletionConfig, ChatCompletionConfig
 
+from vastai import Serverless
+import asyncio
+
+# ---------------------- Logging ----------------------
 logging.basicConfig(
     level=logging.DEBUG,
     format="%(asctime)s[%(levelname)-5s] %(message)s",
@@ -16,135 +17,20 @@ logging.basicConfig(
 )
 log = logging.getLogger(__file__)
 
+# ---------------------- Prompts ----------------------
 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? What do you think this directory might be for?"
+TOOLS_PROMPT = (
-
+    "Can you list the files in the current working directory and tell me what you see? "
-
+    "What do you think this directory might be for?"
-class APIClient:
+)
-    """Lightweight client focused solely on API communication"""
-
-    # Remove the generic WORKER_ENDPOINT since we're now going direct
-    DEFAULT_COST = 100
-    DEFAULT_TIMEOUT = 4
-
-    def __init__(
-        self,
-        endpoint_group_name: str,
-        api_key: str,
-        server_url: str,
-        endpoint_api_key: str,
-    ):
-        self.endpoint_group_name = endpoint_group_name
-        self.api_key = api_key
-        self.server_url = server_url
-        self.endpoint_api_key = endpoint_api_key
-
-    def _get_worker_url(self, cost: int = DEFAULT_COST) -> Dict[str, Any]:
-        """Get worker URL and auth data from routing service"""
-        if not self.endpoint_api_key:
-            raise ValueError("No valid endpoint API key available")
-
-        route_payload = {
-            "endpoint": self.endpoint_group_name,
-            "api_key": self.endpoint_api_key,
-            "cost": cost,
-        }
-
-        response = requests.post(
-            urljoin(self.server_url, "/route/"),
-            json=route_payload,
-            timeout=self.DEFAULT_TIMEOUT,
-        )
-        response.raise_for_status()
-        return response.json()
-
-    def _create_auth_data(self, message: Dict[str, Any]) -> Dict[str, Any]:
-        """Create auth data from routing response"""
-        return {
-            "signature": message["signature"],
-            "cost": message["cost"],
-            "endpoint": message["endpoint"],
-            "reqnum": message["reqnum"],
-            "url": message["url"],
-        }
-
-    def _make_request(
-        self,
-        payload: Dict[str, Any],
-        endpoint: str,
-        method: str = "POST",
-        stream: bool = False,
-    ) -> Union[Dict[str, Any], Iterator[str]]:
-        """Make request directly to the specific worker endpoint"""
-        # Get worker URL and auth data
-        cost = payload.get("max_tokens", self.DEFAULT_COST)
-        message = self._get_worker_url(cost=cost)
-        worker_url = message["url"]
-        auth_data = self._create_auth_data(message)
-
-        req_data = {"payload": {"input": payload}, "auth_data": auth_data}
-
-        url = urljoin(worker_url, endpoint)
-        log.debug(f"Making direct request to: {url}")
-        log.debug(f"Payload: {req_data}")
-
-        # Make the request using the specified method
-        if method.upper() == "POST":
-            response = requests.post(
-                url, json=req_data, stream=stream, verify=get_cert_file_path()
-            )
-        elif method.upper() == "GET":
-            response = requests.get(
-                url, params=req_data, stream=stream, verify=get_cert_file_path()
-            )
-        else:
-            raise ValueError(f"Unsupported HTTP method: {method}")
-
-        response.raise_for_status()
-
-        if stream:
-            return self._handle_streaming_response(response)
-        else:
-            return response.json()
-
-    def _handle_streaming_response(self, response: requests.Response) -> Iterator[str]:
-        """Handle streaming response and yield tokens"""
-        try:
-            for line in response.iter_lines(decode_unicode=True):
-                if line:
-                    if line.startswith("data: "):
-                        data_str = line[6:]
-                        if data_str.strip() == "[DONE]":
-                            break
-                        try:
-                            data = json.loads(data_str)
-                            yield data  # Yield the full chunk
-                        except json.JSONDecodeError:
-                            continue
-        except Exception as e:
-            log.error(f"Error handling streaming response: {e}")
-            raise
-
-    def call_completions(
-        self, config: CompletionConfig
-    ) -> Union[Dict[str, Any], Iterator[str]]:
-        payload = config.to_dict()
-
-        return self._make_request(
-            payload=payload, endpoint="/v1/completions", stream=config.stream
-        )
-
-    def call_chat_completions(
-        self, config: ChatCompletionConfig
-    ) -> Union[Dict[str, Any], Iterator[str]]:
-        payload = config.to_dict()
-
-        return self._make_request(
-            payload=payload, endpoint="/v1/chat/completions", stream=config.stream
-        )
 
+ENDPOINT_NAME = "my-vllm-endpoint"       # change this to your vLLM endpoint name
+DEFAULT_MODEL = "Qwen/Qwen3-8B"          # must support tool calling
+MAX_TOKENS = 1024
+DEFAULT_TEMPERATURE = 0.7
 
+# ---------------------- Tooling ----------------------
 class ToolManager:
     """Handles tool definitions and execution"""
 
@@ -164,7 +50,7 @@ class ToolManager:
 
     @staticmethod
     def get_ls_tool_definition() -> List[Dict[str, Any]]:
-        """Get the ls tool definition"""
+        """OpenAI-compatible tool schema"""
         return [
             {
                 "type": "function",
@@ -178,98 +64,217 @@ class ToolManager:
 
     def execute_tool_call(self, tool_call: Dict[str, Any]) -> str:
         """Execute a tool call and return the result"""
-        function_name = tool_call["function"]["name"]
+        function_name = (tool_call.get("function") or {}).get("name")
-
         if function_name == "list_files":
             return self.list_files()
-        else:
+        raise ValueError(f"Unknown tool function: {function_name}")
-            raise ValueError(f"Unknown tool function: {function_name}")
 
 
+# ----- Helpers to handle streamed tool_calls assembly -----
+def _merge_tool_call_delta(state: Dict[int, Dict[str, Any]], tc_delta: Dict[str, Any]) -> None:
+    """
+    OpenAI-style streaming sends partial tool_calls with an index and partial fields.
+    We merge into a per-index state dict until the assistant message finishes.
+    """
+    idx = tc_delta.get("index")
+    if idx is None:
+        return
+
+    entry = state.setdefault(idx, {"id": None, "function": {"name": None, "arguments": ""}, "type": "function"})
+
+    if tc_delta.get("id"):
+        entry["id"] = tc_delta["id"]
+
+    fn_delta = tc_delta.get("function") or {}
+    if "name" in fn_delta and fn_delta["name"]:
+        entry["function"]["name"] = fn_delta["name"]
+    if "arguments" in fn_delta and fn_delta["arguments"]:
+        entry["function"]["arguments"] += fn_delta["arguments"]
+
+
+def _tool_state_to_message_tool_calls(state: Dict[int, Dict[str, Any]]) -> List[Dict[str, Any]]:
+    return [state[i] for i in sorted(state.keys())]
+
+
+# ---- OpenAI-compatible calls (non-streaming) ----
+async def call_completions(client: Serverless, *, model: str, prompt: str, **kwargs) -> Dict[str, Any]:
+
+    endpoint = await client.get_endpoint(name=ENDPOINT_NAME)
+
+    payload = {
+        "input": {
+            "model": model,
+            "prompt": prompt,
+            "max_tokens": kwargs.get("max_tokens", MAX_TOKENS),
+            "temperature": kwargs.get("temperature", DEFAULT_TEMPERATURE),
+        }
+    }
+    log.debug("POST /v1/completions %s", json.dumps(payload)[:500])
+    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]], **kwargs) -> Dict[str, Any]:
+
+    endpoint = await client.get_endpoint(name=ENDPOINT_NAME)
+
+    payload = {
+        "input": {
+            "model": model,
+            "messages": messages,
+            "max_tokens": kwargs.get("max_tokens", MAX_TOKENS),
+            "temperature": kwargs.get("temperature", DEFAULT_TEMPERATURE),
+            **({"tools": kwargs["tools"]} if "tools" in kwargs else {}),
+            **({"tool_choice": kwargs["tool_choice"]} if "tool_choice" in kwargs else {}),
+        }
+    }
+    log.debug("POST /v1/chat/completions %s", json.dumps(payload)[:500])
+    resp = await endpoint.request("/v1/chat/completions", payload, cost=payload["input"]["max_tokens"])
+    return resp["response"]
+
+# ---- Streaming variants ----
+async def stream_completions(client: Serverless, *, model: str, prompt: str, **kwargs):
+
+    endpoint = await client.get_endpoint(name=ENDPOINT_NAME)
+
+    payload = {
+        "input": {
+            "model": model,
+            "prompt": prompt,
+            "max_tokens": kwargs.get("max_tokens", MAX_TOKENS),
+            "temperature": kwargs.get("temperature", DEFAULT_TEMPERATURE),
+            "stream": True,
+            **({"stop": kwargs["stop"]} if "stop" in kwargs else {}),
+        }
+    }
+    log.debug("STREAM /v1/completions %s", json.dumps(payload)[:500])
+    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]], **kwargs):
+
+    endpoint = await client.get_endpoint(name=ENDPOINT_NAME)
+
+    payload = {
+        "input": {
+            "model": model,
+            "messages": messages,
+            "max_tokens": kwargs.get("max_tokens", MAX_TOKENS),
+            "temperature": kwargs.get("temperature", DEFAULT_TEMPERATURE),
+            "stream": True,
+            **({"tools": kwargs["tools"]} if "tools" in kwargs else {}),
+            **({"tool_choice": kwargs["tool_choice"]} if "tool_choice" in kwargs else {}),
+        }
+    }
+    log.debug("STREAM /v1/chat/completions %s", json.dumps(payload)[:500])
+    resp = await endpoint.request("/v1/chat/completions", payload, cost=payload["input"]["max_tokens"], stream=True)
+    return resp["response"]  # async generator
+
+
+# ---------------------- Demo Runner ----------------------
 class APIDemo:
     """Demo and testing functionality for the API client"""
 
-    def __init__(
+    def __init__(self, client: Serverless, model: str, tool_manager: Optional[ToolManager] = None):
-        self, client: APIClient, model: str, tool_manager: Optional[ToolManager] = None
-    ):
         self.client = client
         self.model = model
         self.tool_manager = tool_manager or ToolManager()
 
-    def handle_streaming_response(
+    # ----- Streaming handler -----
-        self, response_stream, show_reasoning: bool = True
+    async def handle_streaming_response(self, stream, show_reasoning: bool = True) -> str:
-    ) -> str:
-        """
-        Handle streaming chat response and display all output.
-        """
-
         full_response = ""
         reasoning_content = ""
-        reasoning_started = False
+        printed_reasoning = False
-        content_started = False
+        printed_answer = False
 
-        for chunk in response_stream:
+        async for chunk in stream:
-            # Normalize the chunk
+            choice = (chunk.get("choices") or [{}])[0]
-            if isinstance(chunk, str):
+            delta = choice.get("delta", {})
-                chunk = chunk.strip()
-                if chunk.startswith("data: "):
-                    chunk = chunk[6:].strip()
-                if chunk in ["[DONE]", ""]:
-                    continue
-                try:
-                    parsed_chunk = json.loads(chunk)
-                except json.JSONDecodeError:
-                    continue
-            elif isinstance(chunk, dict):
-                parsed_chunk = chunk
-            else:
-                continue
 
-            # Parse delta from the chunk
+            # reasoning tokens
-            choices = parsed_chunk.get("choices", [])
+            rc = delta.get("reasoning_content")
-            if not choices:
+            if rc and show_reasoning:
-                continue
+                if not printed_reasoning:
-
-            delta = choices[0].get("delta", {})
-            reasoning_token = delta.get("reasoning_content", "")
-            content_token = delta.get("content", "")
-
-            # Print reasoning token if applicable
-            if show_reasoning and reasoning_token:
-                if not reasoning_started:
                     print("\n🧠 Reasoning: ", end="", flush=True)
-                    reasoning_started = True
+                    printed_reasoning = True
-                print(f"\033[90m{reasoning_token}\033[0m", end="", flush=True)
+                print(rc, end="", flush=True)
-                reasoning_content += reasoning_token
+                reasoning_content += rc
 
-            # Print content token
+            # content tokens
-            if content_token:
+            content_part = delta.get("content")
-                if not content_started:
+            if content_part:
-                    if show_reasoning and reasoning_started:
+                if not printed_answer:
-                        print(f"\n💬 Response: ", end="", flush=True)
+                    if show_reasoning and printed_reasoning:
+                        print("\n💬 Response: ", end="", flush=True)
                     else:
                         print("Assistant: ", end="", flush=True)
-                    content_started = True
+                    printed_answer = True
-                print(content_token, end="", flush=True)
+                print(content_part, end="", flush=True)
-                full_response += content_token
+                full_response += content_part
-
-        print()  # Ensure newline after response
 
+        print()  # newline
         if show_reasoning:
-            if reasoning_started or content_started:
+            if printed_reasoning or printed_answer:
                 print("\nStreaming completed.")
-            if reasoning_started:
+            if printed_reasoning:
                 print(f"Reasoning tokens: {len(reasoning_content.split())}")
-            if content_started:
+            if printed_answer:
                 print(f"Response tokens: {len(full_response.split())}")
 
         return full_response
+    
+    async def demo_completions(self) -> None:
+        print("=" * 60)
+        print("COMPLETIONS DEMO")
+        print("=" * 60)
 
-    def test_tool_support(self) -> bool:
+        response = await call_completions(
-        """Test if the endpoint supports function calling"""
+            client=self.client,
-        log.debug("Testing endpoint tool calling support...")
+            model=self.model,
+            prompt=COMPLETIONS_PROMPT,
+            max_tokens=MAX_TOKENS,
+            temperature=DEFAULT_TEMPERATURE,
+        )
+        print("\nResponse:")
+        print(json.dumps(response, indent=2))
 
-        # Try a simple request with minimal tools to test support
+    async def demo_chat(self, use_streaming: bool = True) -> None:
+        print("=" * 60)
+        print(f"CHAT COMPLETIONS DEMO {'(STREAMING)' if use_streaming else '(NON-STREAMING)'}")
+        print("=" * 60)
+
+        messages = [{"role": "user", "content": CHAT_PROMPT}]
+
+        if use_streaming:
+            stream = await stream_chat_completions(
+                client=self.client,
+                model=self.model, 
+                messages=messages,
+                max_tokens=MAX_TOKENS,
+                temperature=DEFAULT_TEMPERATURE
+            )
+            try:
+                await self.handle_streaming_response(stream, show_reasoning=True)
+            except Exception as e:
+                log.error("\nError during streaming: %s", e, exc_info=True)
+        else:
+            response = await call_chat_completions(
+                client=self.client,
+                model=self.model, 
+                messages=messages,
+                max_tokens=MAX_TOKENS,
+                temperature=DEFAULT_TEMPERATURE
+            )
+            choice = (response.get("choices") or [{}])[0]
+            message = choice.get("message", {})
+            content = message.get("content", "")
+            reasoning = message.get("reasoning_content", "") or message.get("reasoning", "")
+            if reasoning:
+                print(f"\n🧠 Reasoning: \033[90m{reasoning}\033[0m")
+            print(f"\n💬 Assistant: {content}")
+            print(f"\nFull Response:\n{json.dumps(response, indent=2)}")
+
+    async def test_tool_support(self) -> bool:
+        """Probe that tool schema is accepted (no actual call)"""
         messages = [{"role": "user", "content": "Hello"}]
         minimal_tool = [
             {
@@ -277,170 +282,147 @@ class APIDemo:
                 "function": {"name": "test_function", "description": "Test function"},
             }
         ]
-
-        config = ChatCompletionConfig(
-            model=self.model,
-            messages=messages,
-            max_tokens=10,
-            tools=minimal_tool,
-            tool_choice="none",  # Don't actually call the tool
-        )
-
         try:
-            response = self.client.call_chat_completions(config)
+            _ = await call_chat_completions(
+                client=self.client,
+                model=self.model,
+                messages=messages,
+                tools=minimal_tool,
+                tool_choice="none",
+                max_tokens=10
+            )
             return True
         except Exception as e:
-            log.error(f"Error: Endpoint does not support tool calling: {e}")
+            log.error("Endpoint does not support tool calling: %s", e)
             return False
 
-    def demo_completions(self) -> None:
+    async def demo_ls_tool(self) -> None:
-        """Demo: test basic completions endpoint"""
+        """Ask to list files using function calling, then provide final analysis"""
-        print("=" * 60)
-        print("COMPLETIONS DEMO")
-        print("=" * 60)
-
-        config = CompletionConfig(
-            model=self.model, prompt=COMPLETIONS_PROMPT, stream=False
-        )
-
-        log.info(
-            f"Testing completions with model '{self.model}' and prompt: '{config.prompt}'"
-        )
-        response = self.client.call_completions(config)
-
-        if isinstance(response, dict):
-            print("\nResponse:")
-            print(json.dumps(response, indent=2))
-        else:
-            log.error("Unexpected response format")
-
-    def demo_chat(self, use_streaming: bool = True) -> None:
-        """
-        Demo: test chat completions endpoint with optional streaming
-        """
-        print("=" * 60)
-        print(
-            f"CHAT COMPLETIONS DEMO {'(STREAMING)' if use_streaming else '(NON-STREAMING)'}"
-        )
-        print("=" * 60)
-
-        config = ChatCompletionConfig(
-            model=self.model,
-            messages=[{"role": "user", "content": CHAT_PROMPT}],
-            stream=use_streaming,
-        )
-
-        log.info(f"Testing chat completions with model '{self.model}'...")
-        response = self.client.call_chat_completions(config)
-
-        if use_streaming:
-            try:
-                self.handle_streaming_response(response, show_reasoning=True)
-            except Exception as e:
-                log.error(f"\nError during streaming: {e}")
-                import traceback
-
-                traceback.print_exc()
-                return
-
-        else:
-            if isinstance(response, dict):
-                choice = response.get("choices", [{}])[0]
-                message = choice.get("message", {})
-                content = message.get("content", "")
-                reasoning = message.get("reasoning_content", "") or message.get(
-                    "reasoning", ""
-                )
-
-                if reasoning:
-                    print(f"\n🧠 Reasoning: \033[90m{reasoning}\033[0m")
-
-                print(f"\n💬 Assistant: {content}")
-                print(f"\nFull Response:")
-                print(json.dumps(response, indent=2))
-            else:
-                log.error("Unexpected response format")
-
-    def demo_ls_tool(self) -> None:
-        """Demo: ask LLM to list files in the current directory and describe what it sees"""
         print("=" * 60)
         print("TOOL USE DEMO: List Directory Contents")
         print("=" * 60)
 
-        # Test if tools are supported first
+        if not await self.test_tool_support():
-        if not self.test_tool_support():
             return
 
-        # Request with tool available
+        messages: List[Dict[str, Any]] = [{"role": "user", "content": TOOLS_PROMPT}]
-        messages = [{"role": "user", "content": TOOLS_PROMPT}]
 
-        config = ChatCompletionConfig(
+        # First pass: let the model decide tools, stream tool_calls and partial content
+        stream = await stream_chat_completions(
+            client=self.client,
             model=self.model,
             messages=messages,
             tools=self.tool_manager.get_ls_tool_definition(),
             tool_choice="auto",
+            max_tokens=MAX_TOKENS,
+            temperature=DEFAULT_TEMPERATURE,
         )
 
-        log.info(f"Making initial request with tool using model '{self.model}'...")
+        assistant_content_buf: List[str] = []
-        response = self.client.call_chat_completions(config)
+        tool_calls_state: Dict[int, Dict[str, Any]] = {}
+        printed_reasoning = False
+        printed_answer = False
 
-        if not isinstance(response, dict):
+        async for chunk in stream:
-            raise ValueError("Expected dict response for tool use")
+            choice = (chunk.get("choices") or [{}])[0]
+            delta = choice.get("delta", {})
 
-        choice = response.get("choices", [{}])[0]
+            rc = delta.get("reasoning_content")
-        message = choice.get("message", {})
+            if rc:
+                if not printed_reasoning:
+                    printed_reasoning = True
+                    print("🧠 Reasoning: ", end="", flush=True)
+                print(rc, end="", flush=True)
 
-        print(f"Assistant response: {message.get('content', 'No content')}")
+            content_part = delta.get("content")
+            if content_part:
+                assistant_content_buf.append(content_part)
+                if not printed_answer:
+                    printed_answer = True
+                    print("\n💬 Response: ", end="", flush=True)
+                print(content_part, end="", flush=True)
 
-        # Check for tool calls
+            if "tool_calls" in delta and delta["tool_calls"]:
-        tool_calls = message.get("tool_calls")
+                for tc_delta in delta["tool_calls"]:
-        if not tool_calls:
+                    _merge_tool_call_delta(tool_calls_state, tc_delta)
-            raise ValueError(
-                "No tool calls made - model may not support function calling"
-            )
 
-        print(f"Tool calls detected: {len(tool_calls)}")
+        # If no tool calls, we’re done.
+        if not tool_calls_state:
+            print("\n(No tool calls were made.)")
+            return
 
-        # Execute the tool call
+        # Build assistant message with tool_calls
-        for tool_call in tool_calls:
+        assistant_message = {
-            function_name = tool_call["function"]["name"]
+            "role": "assistant",
-            print(f"Executing tool: {function_name}")
+            "content": "".join(assistant_content_buf) if assistant_content_buf else None,
+            "tool_calls": _tool_state_to_message_tool_calls(tool_calls_state),
+        }
+        messages.append(assistant_message)
 
-            tool_result = self.tool_manager.execute_tool_call(tool_call)
+        # Execute tools and feed results back
-            print(f"Tool result:\n{tool_result}")
+        for tc in assistant_message["tool_calls"]:
+            tool_name = (tc.get("function") or {}).get("name")
+            call_id = tc.get("id")
+            raw_args = (tc.get("function") or {}).get("arguments") or "{}"
 
-            # Add tool result and continue conversation
+            try:
-            messages.append(message)  # Add assistant's message with tool call
+                args = json.loads(raw_args) if raw_args.strip() else {}
-            messages.append(
+            except Exception as e:
-                {
+                tool_result = json.dumps({"error": f"Argument parse failed: {str(e)}", "raw_arguments": raw_args})
-                    "role": "tool",
+                messages.append({"role": "tool", "tool_call_id": call_id, "content": tool_result})
-                    "tool_call_id": tool_call["id"],
+                continue
-                    "content": tool_result,
-                }
-            )
 
-            # Get final response
+            try:
-            final_config = ChatCompletionConfig(
+                if tool_name == "list_files":
-                model=self.model,
+                    tool_result = self.tool_manager.list_files()
-                messages=messages,
+                else:
-                tools=self.tool_manager.get_ls_tool_definition(),
+                    tool_result = json.dumps({"error": f"Unknown tool '{tool_name}'"})
-            )
+            except Exception as e:
+                tool_result = json.dumps({"error": f"Tool '{tool_name}' failed: {str(e)}"})
 
-            print("Getting final response...")
+            print("\n[Tool executed]", tool_name)
-            final_response = self.client.call_chat_completions(final_config)
+            print(tool_result[:500] + ("..." if len(tool_result) > 500 else ""))
+            messages.append({"role": "tool", "tool_call_id": call_id, "content": tool_result})
 
-            if isinstance(final_response, dict):
+        # Second pass: get final streamed answer after tool results
-                final_choice = final_response.get("choices", [{}])[0]
+        stream2 = await stream_chat_completions(
-                final_message = final_choice.get("message", {})
+            client=self.client,
-                final_content = final_message.get("content", "")
+            model=self.model,
+            messages=messages,
+            max_tokens=MAX_TOKENS,
+            temperature=DEFAULT_TEMPERATURE,
+        )
 
-                print("\n" + "=" * 60)
+        final_buf = []
-                print("FINAL LLM ANALYSIS:")
+        printed_reasoning2 = False
-                print("=" * 60)
+        printed_answer2 = False
-                print(final_content)
-                print("=" * 60)
 
-    def interactive_chat(self) -> None:
+        async for chunk in stream2:
+            choice = (chunk.get("choices") or [{}])[0]
+            delta = choice.get("delta", {})
+
+            rc2 = delta.get("reasoning_content")
+            if rc2:
+                if not printed_reasoning2:
+                    printed_reasoning2 = True
+                    print("\n🧠 Reasoning (post-tools): ", end="", flush=True)
+                print(rc2, end="", flush=True)
+
+            c2 = delta.get("content")
+            if c2:
+                final_buf.append(c2)
+                if not printed_answer2:
+                    printed_answer2 = True
+                    print("\n💬 Response (final): ", end="", flush=True)
+                print(c2, end="", flush=True)
+
+        print("\n" + "=" * 60)
+        print("FINAL LLM ANALYSIS:")
+        print("=" * 60)
+        print("".join(final_buf))
+        print("=" * 60)
+
+    async def interactive_chat(self) -> None:
         """Interactive chat session with streaming"""
         print("=" * 60)
         print("INTERACTIVE STREAMING CHAT")
@@ -449,7 +431,7 @@ class APIDemo:
         print("Type 'quit' to exit, 'clear' to clear history")
         print()
 
-        messages = []
+        messages: List[Dict[str, Any]] = []
 
         while True:
             try:
@@ -467,16 +449,15 @@ class APIDemo:
 
                 messages.append({"role": "user", "content": user_input})
 
-                config = ChatCompletionConfig(
-                    model=self.model, messages=messages, stream=True, temperature=0.7
-                )
-
                 print("Assistant: ", end="", flush=True)
-
+                stream = await stream_chat_completions(
-                response = self.client.call_chat_completions(config)
+                    client=self.client,
-                assistant_content = self.handle_streaming_response(
+                    model=self.model, 
-                    response, show_reasoning=True
+                    messages=messages, 
+                    max_tokens=MAX_TOKENS, 
+                    temperature=0.7
                 )
+                assistant_content = await self.handle_streaming_response(stream, show_reasoning=True)
 
                 # Add assistant response to conversation history
                 messages.append({"role": "assistant", "content": assistant_content})
@@ -485,115 +466,64 @@ class APIDemo:
                 print("\n👋 Chat interrupted. Goodbye!")
                 break
             except Exception as e:
-                log.error(f"\nError: {e}")
+                log.error("\nError: %s", e)
                 continue
 
 
-def main():
+# ---------------------- CLI ----------------------
-    """Main function with CLI switches for different tests"""
+def build_arg_parser() -> argparse.ArgumentParser:
-    from lib.test_utils import test_args
+    p = argparse.ArgumentParser(description="Vast vLLM Demo (Serverless SDK)")
+    p.add_argument("--model", required=True, help="Model to use for requests (required)")
+    p.add_argument("--endpoint", default="my-vllm-endpoint", help="Vast endpoint name (default: my-vllm-endpoint)")
 
-    # Add mandatory model argument
+    modes = p.add_mutually_exclusive_group(required=False)
-    test_args.add_argument(
+    modes.add_argument("--completion", action="store_true", help="Test completions endpoint")
-        "--model", required=True, help="Model to use for requests (required)"
+    modes.add_argument("--chat", action="store_true", help="Test chat completions endpoint (non-streaming)")
-    )
+    modes.add_argument("--chat-stream", action="store_true", help="Test chat completions endpoint with streaming")
+    modes.add_argument("--tools", action="store_true", help="Test function calling with ls tool (non-streaming+streamed phases)")
+    modes.add_argument("--interactive", action="store_true", help="Start interactive streaming chat session")
+    return p
 
-    # Add test mode arguments
-    test_args.add_argument(
-        "--completion", action="store_true", help="Test completions endpoint"
-    )
-    test_args.add_argument(
-        "--chat",
-        action="store_true",
-        help="Test chat completions endpoint (non-streaming)",
-    )
-    test_args.add_argument(
-        "--chat-stream",
-        action="store_true",
-        help="Test chat completions endpoint with streaming",
-    )
-    test_args.add_argument(
-        "--tools",
-        action="store_true",
-        help="Test function calling with ls tool (non-streaming)",
-    )
-    test_args.add_argument(
-        "--interactive",
-        action="store_true",
-        help="Start interactive streaming chat session",
-    )
 
-    args = test_args.parse_args()
+async def main_async():
+    args = build_arg_parser().parse_args()
 
-    # Check that only one test mode is selected
+    selected = sum([args.completion, args.chat, args.chat_stream, args.tools, args.interactive])
-    test_modes = [
+    if selected == 0:
-        args.completion,
-        args.chat,
-        args.chat_stream,
-        args.tools,
-        args.interactive,
-    ]
-    selected_count = sum(test_modes)
-
-    if selected_count == 0:
         print("Please specify exactly one test mode:")
         print("  --completion    : Test completions endpoint")
         print("  --chat          : Test chat completions endpoint (non-streaming)")
         print("  --chat-stream   : Test chat completions endpoint with streaming")
-        print("  --tools         : Test function calling with ls tool (non-streaming)")
+        print("  --tools         : Test function calling with ls tool")
         print("  --interactive   : Start interactive streaming chat session")
-        print(
+        print(f"\nExample: python {os.path.basename(sys.argv[0])} --model Qwen/Qwen3-8B --chat-stream --endpoint my-vllm-endpoint")
-            f"\nExample: python {sys.argv[0]} --model Qwen/Qwen3-8B --chat-stream -k YOUR_KEY -e YOUR_ENDPOINT"
-        )
         sys.exit(1)
-    elif selected_count > 1:
+    elif selected > 1:
         print("Please specify exactly one test mode")
         sys.exit(1)
 
+    print(f"Using model: {args.model}")
+    print("=" * 60)
+
     try:
-        endpoint_api_key = Endpoint.get_endpoint_api_key(
+        async with Serverless() as client:
-            endpoint_name=args.endpoint_group_name,
+            demo = APIDemo(client, args.model, ToolManager())
-            account_api_key=args.api_key,
-            instance=args.instance,
-        )
 
-        if not endpoint_api_key:
+            if args.completion:
-            log.error(
+                await demo.demo_completions()
-                f"Could not retrieve API key for endpoint '{args.endpoint_group_name}'. Exiting."
+            elif args.chat:
-            )
+                await demo.demo_chat(use_streaming=False)
-            sys.exit(1)
+            elif args.chat_stream:
-
+                await demo.demo_chat(use_streaming=True)
-        # Create the core API client
+            elif args.tools:
-        client = APIClient(
+                await demo.demo_ls_tool()
-            endpoint_group_name=args.endpoint_group_name,
+            elif args.interactive:
-            api_key=args.api_key,
+                await demo.interactive_chat()
-            server_url=Endpoint.get_autoscaler_server_url(args.instance),
-            endpoint_api_key=endpoint_api_key,
-        )
-
-        # Create tool manager and demo (passing the model parameter)
-        tool_manager = ToolManager()
-        demo = APIDemo(client, args.model, tool_manager)
-
-        print(f"Using model: {args.model}")
-        print("=" * 60)
-
-        # Run the selected test
-        if args.completion:
-            demo.demo_completions()
-        elif args.chat:
-            demo.demo_chat(use_streaming=False)
-        elif args.chat_stream:
-            demo.demo_chat(use_streaming=True)
-        elif args.tools:
-            demo.demo_ls_tool()
-        elif args.interactive:
-            demo.interactive_chat()
 
     except Exception as e:
-        log.error(f"Error during test: {e}", exc_info=True)
+        log.error("Error during test: %s", e, exc_info=True)
         sys.exit(1)
 
 
 if __name__ == "__main__":
-    main()
+    asyncio.run(main_async())

workers/tgi/client.py

@@ -1,125 +1,61 @@
-import logging
+from vastai import Serverless
-import sys
+import asyncio
-import json
-from urllib.parse import urljoin
-import requests
-from utils.endpoint_util import Endpoint
-from utils.ssl import get_cert_file_path
 
-logging.basicConfig(
+ENDPOINT_NAME = "my-tgi-endpoint" # Change this to match your endpoint name
-    level=logging.DEBUG,
+MAX_TOKENS = 1024
-    format="%(asctime)s[%(levelname)-5s] %(message)s",
+PROMPT = "Think step by step: Tell me about the Python programming language."
-    datefmt="%Y-%m-%d %H:%M:%S",
-)
-log = logging.getLogger(__file__)
 
+async def call_generate(client: Serverless) -> None:
+    endpoint = await client.get_endpoint(name=ENDPOINT_NAME)
 
-def call_generate(endpoint_group_name: str, api_key: str, server_url: str) -> None:
+    payload = {
-    WORKER_ENDPOINT = "/generate"
+        "inputs": PROMPT,
-    COST = 100
+        "parameters": {
-    route_payload = {
+            "max_new_tokens": MAX_TOKENS,
-        "endpoint": endpoint_group_name,
+            "temperature": 0.7,
-        "api_key": api_key,
+            "return_full_text": False
-        "cost": COST,
+        }
     }
-    response = requests.post(
-        urljoin(server_url, "/route/"),
-        json=route_payload,
-        timeout=4,
-    )
-    response.raise_for_status()  # Raise an exception for bad status codes
-    message = response.json()
-    url = message["url"]
 
-    auth_data = dict(
+    resp = await endpoint.request("/generate", payload, cost=MAX_TOKENS)
-        signature=message["signature"],
-        cost=message["cost"],
-        endpoint=message["endpoint"],
-        reqnum=message["reqnum"],
-        url=url,
-    )
 
-    payload = dict(inputs="tell me about cats", parameters=dict(max_new_tokens=500))
+    print(resp["response"]["generated_text"])
-    req_data = dict(payload=payload, auth_data=auth_data)
-    url = urljoin(url, WORKER_ENDPOINT)
-    print(f"url: {url}")
-    response = requests.post(
-        url,
-        json=req_data,
-        verify=get_cert_file_path(),
-    )
-    response.raise_for_status()
-    res = response.json()
-    print(res)
 
 
-def call_generate_stream(
+async def call_generate_stream(client: Serverless) -> None:
-    endpoint_group_name: str, api_key: str, server_url: str
+    endpoint = await client.get_endpoint(name=ENDPOINT_NAME)
-) -> None:
+
-    WORKER_ENDPOINT = "/generate_stream"
+    payload = {
-    COST = 100
+        "inputs": PROMPT,
-    route_payload = {
+        "parameters": {
-        "endpoint": endpoint_group_name,
+            "max_new_tokens": MAX_TOKENS,
-        "api_key": api_key,
+            "temperature": 0.7,
-        "cost": COST,
+            "do_sample": True,
+            "return_full_text": False,
+        }
     }
-    response = requests.post(
-        urljoin(server_url, "/route/"),
-        json=route_payload,
-        timeout=4,
-    )
-    response.raise_for_status()  # Raise an exception for bad status codes
-    message = response.json()
-    url = message["url"]
-    print(f"url: {url}")
-    auth_data = dict(
-        signature=message["signature"],
-        cost=message["cost"],
-        endpoint=message["endpoint"],
-        reqnum=message["reqnum"],
-        url=message["url"],
-    )
-    payload = dict(inputs="tell me about dogs", parameters=dict(max_new_tokens=500))
-    req_data = dict(payload=payload, auth_data=auth_data)
-    url = urljoin(url, WORKER_ENDPOINT)
-    response = requests.post(url, json=req_data, stream=True)
-    response.raise_for_status()  # Raise an exception for bad status codes
-    for line in response.iter_lines():
-        payload = line.decode().lstrip("data:").rstrip()
-        if payload:
-            try:
-                data = json.loads(payload)
-                print(data["token"]["text"], end="")
-                sys.stdout.flush()
-            except (json.JSONDecodeError, KeyError) as e:
-                log.warning(f"Failed to parse streaming response: {e}")
-                continue
-    print()
 
+    resp = await endpoint.request(
+        "/generate_stream",
+        payload,
+        cost=MAX_TOKENS,
+        stream=True,
+    )
+    stream = resp["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("Answer:\n", end="", flush=True)
+            print(tok, end="", flush=True)
+
+async def main():
+    async with Serverless() as client:
+        await call_generate(client)
+        await call_generate_stream(client)
 
 if __name__ == "__main__":
-    from lib.test_utils import test_args
+    asyncio.run(main())
-
-    args = test_args.parse_args()
-
-    endpoint_api_key = Endpoint.get_endpoint_api_key(
-        endpoint_name=args.endpoint_group_name,
-        account_api_key=args.api_key,
-        instance=args.instance,
-    )
-    if endpoint_api_key:
-        try:
-            call_generate(
-                api_key=endpoint_api_key,
-                endpoint_group_name=args.endpoint_group_name,
-                server_url=args.server_url,
-            )
-            call_generate_stream(
-                api_key=endpoint_api_key,
-                endpoint_group_name=args.endpoint_group_name,
-                server_url=args.server_url,
-            )
-        except Exception as e:
-            log.error(f"Error during API call: {e}")
-    else:
-        log.error(f"Failed to get API key for endpoint {args.endpoint_group_name} ")