# SPDX-FileCopyrightText: Copyright (c) 2024, NVIDIA CORPORATION & AFFILIATES.
# All rights reserved.
# SPDX-License-Identifier: Apache-2.0
import hashlib
import json
import logging
import re
import threading
import time
import queue
from collections import namedtuple
from concurrent.futures import Future, ThreadPoolExecutor, as_completed
from typing import Any
from typing import Optional
from typing import Tuple, Union
import numpy as np
import requests
import tritonclient.grpc as grpcclient
from nv_ingest_api.internal.primitives.tracing.tagging import traceable_func
from nv_ingest_api.util.string_processing import generate_url
logger = logging.getLogger(__name__)
# Regex pattern to detect CUDA-related errors in Triton gRPC responses
CUDA_ERROR_REGEX = re.compile(
r"(model reload|illegal memory access|illegal instruction|invalid argument|failed to (copy|load|perform) .*: .*|TritonModelException: failed to copy data: .*)", # noqa: E501
re.IGNORECASE,
)
# A simple structure to hold a request's data and its Future for the result
InferenceRequest = namedtuple("InferenceRequest", ["data", "future", "model_name", "dims", "kwargs"])
[docs]
class NimClient:
"""
A client for interfacing with a model inference server using gRPC or HTTP protocols.
"""
def __init__(
self,
model_interface,
protocol: str,
endpoints: Tuple[str, str],
auth_token: Optional[str] = None,
timeout: float = 120.0,
max_retries: int = 10,
max_429_retries: int = 5,
enable_dynamic_batching: bool = False,
dynamic_batch_timeout: float = 0.1, # 100 milliseconds
dynamic_batch_memory_budget_mb: Optional[float] = None,
):
"""
Initialize the NimClient with the specified model interface, protocol, and server endpoints.
Parameters
----------
model_interface : ModelInterface
The model interface implementation to use.
protocol : str
The protocol to use ("grpc" or "http").
endpoints : tuple
A tuple containing the gRPC and HTTP endpoints.
auth_token : str, optional
Authorization token for HTTP requests (default: None).
timeout : float, optional
Timeout for HTTP requests in seconds (default: 120.0).
max_retries : int, optional
The maximum number of retries for non-429 server-side errors (default: 10).
max_429_retries : int, optional
The maximum number of retries specifically for 429 errors (default: 5).
Raises
------
ValueError
If an invalid protocol is specified or if required endpoints are missing.
"""
self.client = None
self.model_interface = model_interface
self.protocol = protocol.lower()
self.auth_token = auth_token
self.timeout = timeout # Timeout for HTTP requests
self.max_retries = max_retries
self.max_429_retries = max_429_retries
self._grpc_endpoint, self._http_endpoint = endpoints
self._max_batch_sizes = {}
self._lock = threading.Lock()
if self.protocol == "grpc":
if not self._grpc_endpoint:
raise ValueError("gRPC endpoint must be provided for gRPC protocol")
logger.debug(f"Creating gRPC client with {self._grpc_endpoint}")
self.client = grpcclient.InferenceServerClient(url=self._grpc_endpoint)
elif self.protocol == "http":
if not self._http_endpoint:
raise ValueError("HTTP endpoint must be provided for HTTP protocol")
logger.debug(f"Creating HTTP client with {self._http_endpoint}")
self.endpoint_url = generate_url(self._http_endpoint)
self.headers = {"accept": "application/json", "content-type": "application/json"}
if self.auth_token:
self.headers["Authorization"] = f"Bearer {self.auth_token}"
else:
raise ValueError("Invalid protocol specified. Must be 'grpc' or 'http'.")
self.dynamic_batching_enabled = enable_dynamic_batching
if self.dynamic_batching_enabled:
self._batch_timeout = dynamic_batch_timeout
if dynamic_batch_memory_budget_mb is not None:
self._batch_memory_budget_bytes = dynamic_batch_memory_budget_mb * 1024 * 1024
else:
self._batch_memory_budget_bytes = None
self._request_queue = queue.Queue()
self._stop_event = threading.Event()
self._batcher_thread = threading.Thread(target=self._batcher_loop, daemon=True)
[docs]
def start(self):
"""Starts the dynamic batching worker thread if enabled."""
if self.dynamic_batching_enabled and not self._batcher_thread.is_alive():
self._batcher_thread.start()
def _fetch_max_batch_size(self, model_name, model_version: str = "") -> int:
"""Fetch the maximum batch size from the Triton model configuration in a thread-safe manner."""
if model_name == "yolox_ensemble":
model_name = "yolox"
if model_name in self._max_batch_sizes:
return self._max_batch_sizes[model_name]
with self._lock:
# Double check, just in case another thread set the value while we were waiting
if model_name in self._max_batch_sizes:
return self._max_batch_sizes[model_name]
if not self._grpc_endpoint or not self.client:
self._max_batch_sizes[model_name] = 1
return 1
try:
model_config = self.client.get_model_config(model_name=model_name, model_version=model_version)
self._max_batch_sizes[model_name] = model_config.config.max_batch_size
logger.debug(f"Max batch size for model '{model_name}': {self._max_batch_sizes[model_name]}")
except Exception as e:
self._max_batch_sizes[model_name] = 1
logger.warning(f"Failed to retrieve max batch size: {e}, defaulting to 1")
return self._max_batch_sizes[model_name]
def _process_batch(self, batch_input, *, batch_data, model_name, **kwargs):
"""
Process a single batch input for inference using its corresponding batch_data.
Parameters
----------
batch_input : Any
The input data for this batch.
batch_data : Any
The corresponding scratch-pad data for this batch as returned by format_input.
model_name : str
The model name for inference.
kwargs : dict
Additional parameters.
Returns
-------
tuple
A tuple (parsed_output, batch_data) for subsequent post-processing.
"""
if self.protocol == "grpc":
logger.debug("Performing gRPC inference for a batch...")
response = self._grpc_infer(batch_input, model_name, **kwargs)
logger.debug("gRPC inference received response for a batch")
elif self.protocol == "http":
logger.debug("Performing HTTP inference for a batch...")
response = self._http_infer(batch_input)
logger.debug("HTTP inference received response for a batch")
else:
raise ValueError("Invalid protocol specified. Must be 'grpc' or 'http'.")
parsed_output = self.model_interface.parse_output(
response, protocol=self.protocol, data=batch_data, model_name=model_name, **kwargs
)
return parsed_output, batch_data
[docs]
def try_set_max_batch_size(self, model_name, model_version: str = ""):
"""Attempt to set the max batch size for the model if it is not already set, ensuring thread safety."""
self._fetch_max_batch_size(model_name, model_version)
[docs]
@traceable_func(trace_name="{stage_name}::{model_name}")
def infer(self, data: dict, model_name: str, **kwargs) -> Any:
"""
Perform inference using the specified model and input data.
Parameters
----------
data : dict
The input data for inference.
model_name : str
The model name.
kwargs : dict
Additional parameters for inference.
Returns
-------
Any
The processed inference results, coalesced in the same order as the input images.
"""
# 1. Retrieve or default to the model's maximum batch size.
batch_size = self._fetch_max_batch_size(model_name)
max_requested_batch_size = kwargs.pop("max_batch_size", batch_size)
force_requested_batch_size = kwargs.pop("force_max_batch_size", False)
max_batch_size = (
max(1, min(batch_size, max_requested_batch_size))
if not force_requested_batch_size
else max_requested_batch_size
)
self._batch_size = max_batch_size
if self.dynamic_batching_enabled:
# DYNAMIC BATCHING PATH
try:
data = self.model_interface.prepare_data_for_inference(data)
futures = []
for base64_image, image_array in zip(data["base64_images"], data["images"]):
dims = image_array.shape[:2]
futures.append(self.submit(base64_image, model_name, dims, **kwargs))
results = [future.result() for future in futures]
return results
except Exception as err:
error_str = (
f"Error during synchronous infer with dynamic batching [{self.model_interface.name()}]: {err}"
)
logger.error(error_str)
raise RuntimeError(error_str) from err
# OFFLINE BATCHING PATH
try:
# 2. Prepare data for inference.
data = self.model_interface.prepare_data_for_inference(data)
# 3. Format the input based on protocol.
formatted_batches, formatted_batch_data = self.model_interface.format_input(
data,
protocol=self.protocol,
max_batch_size=max_batch_size,
model_name=model_name,
**kwargs,
)
# Check for a custom maximum pool worker count, and remove it from kwargs.
max_pool_workers = kwargs.pop("max_pool_workers", 16)
# 4. Process each batch concurrently using a thread pool.
# We enumerate the batches so that we can later reassemble results in order.
results = [None] * len(formatted_batches)
with ThreadPoolExecutor(max_workers=max_pool_workers) as executor:
future_to_idx = {}
for idx, (batch, batch_data) in enumerate(zip(formatted_batches, formatted_batch_data)):
future = executor.submit(
self._process_batch, batch, batch_data=batch_data, model_name=model_name, **kwargs
)
future_to_idx[future] = idx
for future in as_completed(future_to_idx.keys()):
idx = future_to_idx[future]
results[idx] = future.result()
# 5. Process the parsed outputs for each batch using its corresponding batch_data.
# As the batches are in order, we coalesce their outputs accordingly.
all_results = []
for parsed_output, batch_data in results:
batch_results = self.model_interface.process_inference_results(
parsed_output,
original_image_shapes=batch_data.get("original_image_shapes"),
protocol=self.protocol,
**kwargs,
)
if isinstance(batch_results, list):
all_results.extend(batch_results)
else:
all_results.append(batch_results)
except Exception as err:
error_str = f"Error during NimClient inference [{self.model_interface.name()}, {self.protocol}]: {err}"
logger.error(error_str)
raise RuntimeError(error_str)
return all_results
def _grpc_infer(
self, formatted_input: Union[list, list[np.ndarray]], model_name: str, **kwargs
) -> Union[list, list[np.ndarray]]:
"""
Perform inference using the gRPC protocol.
Parameters
----------
formatted_input : np.ndarray
The input data formatted as a numpy array.
model_name : str
The name of the model to use for inference.
Returns
-------
np.ndarray
The output of the model as a numpy array.
"""
if not isinstance(formatted_input, list):
formatted_input = [formatted_input]
parameters = kwargs.get("parameters", {})
output_names = kwargs.get("output_names", ["output"])
dtypes = kwargs.get("dtypes", ["FP32"])
input_names = kwargs.get("input_names", ["input"])
input_tensors = []
for input_name, input_data, dtype in zip(input_names, formatted_input, dtypes):
input_tensors.append(grpcclient.InferInput(input_name, input_data.shape, datatype=dtype))
for idx, input_data in enumerate(formatted_input):
input_tensors[idx].set_data_from_numpy(input_data)
outputs = [grpcclient.InferRequestedOutput(output_name) for output_name in output_names]
base_delay = 2.0
attempt = 0
retries_429 = 0
max_grpc_retries = self.max_429_retries
while attempt < self.max_retries:
try:
response = self.client.infer(
model_name=model_name, parameters=parameters, inputs=input_tensors, outputs=outputs
)
logger.debug(f"gRPC inference response: {response}")
if len(outputs) == 1:
return response.as_numpy(outputs[0].name())
else:
return [response.as_numpy(output.name()) for output in outputs]
except grpcclient.InferenceServerException as e:
status = str(e.status())
message = e.message()
# Handle CUDA memory errors
if status == "StatusCode.INTERNAL":
if CUDA_ERROR_REGEX.search(message):
logger.warning(
f"Received gRPC INTERNAL error with CUDA-related message for model '{model_name}'. "
f"Attempt {attempt + 1} of {self.max_retries}. Message (truncated): {message[:500]}"
)
if attempt >= self.max_retries - 1:
logger.error(f"Max retries exceeded for CUDA errors on model '{model_name}'.")
raise e
# Try to reload models before retrying
model_reload_succeeded = reload_models(client=self.client, client_timeout=self.timeout)
if not model_reload_succeeded:
logger.error(f"Failed to reload models for model '{model_name}'.")
else:
logger.warning(
f"Received gRPC INTERNAL error for model '{model_name}'. "
f"Attempt {attempt + 1} of {self.max_retries}. Message (truncated): {message[:500]}"
)
if attempt >= self.max_retries - 1:
logger.error(f"Max retries exceeded for INTERNAL error on model '{model_name}'.")
raise e
# Common retry logic for both CUDA and non-CUDA INTERNAL errors
backoff_time = base_delay * (2**attempt)
time.sleep(backoff_time)
attempt += 1
continue
# Handle errors that can occur after model reload (NOT_FOUND, model not loaded)
if status == "StatusCode.NOT_FOUND":
logger.warning(
f"Received gRPC {status} error for model '{model_name}'. "
f"Attempt {attempt + 1} of {self.max_retries}. Message: {message[:500]}"
)
if attempt >= self.max_retries - 1:
logger.error(f"Max retries exceeded for model not found errors on model '{model_name}'.")
raise e
# Retry with exponential backoff WITHOUT reloading
backoff_time = base_delay * (2**attempt)
logger.info(
f"Retrying after {backoff_time}s backoff for model not found error on model '{model_name}'."
)
time.sleep(backoff_time)
attempt += 1
continue
if status == "StatusCode.UNAVAILABLE" and "Exceeds maximum queue size".lower() in message.lower():
retries_429 += 1
logger.warning(
f"Received gRPC {status} for model '{model_name}'. "
f"Attempt {retries_429} of {max_grpc_retries}."
)
if retries_429 >= max_grpc_retries:
logger.error(f"Max retries for gRPC {status} exceeded for model '{model_name}'.")
raise
backoff_time = base_delay * (2**retries_429)
time.sleep(backoff_time)
continue
# For other server-side errors (e.g., INVALID_ARGUMENT, etc.),
# fail fast as retrying will not help
logger.error(
f"Received non-retryable gRPC error {status} from Triton for model '{model_name}': {message}"
)
raise
except Exception as e:
# Catch any other unexpected exceptions (e.g., network issues not caught by Triton client)
logger.error(f"An unexpected error occurred during gRPC inference for model '{model_name}': {e}")
raise
def _http_infer(self, formatted_input: dict) -> dict:
"""
Perform inference using the HTTP protocol, retrying for timeouts or 5xx errors up to 5 times.
Parameters
----------
formatted_input : dict
The input data formatted as a dictionary.
Returns
-------
dict
The output of the model as a dictionary.
Raises
------
TimeoutError
If the HTTP request times out repeatedly, up to the max retries.
requests.RequestException
For other HTTP-related errors that persist after max retries.
"""
base_delay = 2.0
attempt = 0
retries_429 = 0
while attempt < self.max_retries:
try:
# Log system prompt for VLM requests
if isinstance(formatted_input, dict) and "messages" in formatted_input:
messages = formatted_input.get("messages", [])
if messages and messages[0].get("role") == "system":
system_content = messages[0].get("content", "")
model_name = self.model_interface.name()
logger.debug(f"{model_name}: Sending HTTP request with system prompt: '{system_content}'")
response = requests.post(
self.endpoint_url, json=formatted_input, headers=self.headers, timeout=self.timeout
)
status_code = response.status_code
# Check for server-side or rate-limit type errors
# e.g. 5xx => server error, 429 => too many requests
if status_code == 429:
retries_429 += 1
logger.warning(
f"Received HTTP 429 (Too Many Requests) from {self.model_interface.name()}. "
f"Attempt {retries_429} of {self.max_429_retries}."
)
if retries_429 >= self.max_429_retries:
logger.error("Max retries for HTTP 429 exceeded.")
response.raise_for_status()
else:
backoff_time = base_delay * (2**retries_429)
time.sleep(backoff_time)
continue # Retry without incrementing the main attempt counter
if status_code == 503 or (500 <= status_code < 600):
logger.warning(
f"Received HTTP {status_code} ({response.reason}) from "
f"{self.model_interface.name()}. Attempt {attempt + 1} of {self.max_retries}."
)
if attempt == self.max_retries - 1:
# No more retries left
logger.error(f"Max retries exceeded after receiving HTTP {status_code}.")
response.raise_for_status() # raise the appropriate HTTPError
else:
# Exponential backoff
backoff_time = base_delay * (2**attempt)
time.sleep(backoff_time)
attempt += 1
continue
else:
# Not in our "retry" category => just raise_for_status or return
response.raise_for_status()
logger.debug(f"HTTP inference response: {response.json()}")
return response.json()
except requests.Timeout:
# Treat timeouts similarly to 5xx => attempt a retry
logger.warning(
f"HTTP request timed out after {self.timeout} seconds during {self.model_interface.name()} "
f"inference. Attempt {attempt + 1} of {self.max_retries}."
)
if attempt == self.max_retries - 1:
logger.error("Max retries exceeded after repeated timeouts.")
raise TimeoutError(
f"Repeated timeouts for {self.model_interface.name()} after {attempt + 1} attempts."
)
# Exponential backoff
backoff_time = base_delay * (2**attempt)
time.sleep(backoff_time)
attempt += 1
except requests.HTTPError as http_err:
# If we ended up here, it's a non-retryable 4xx or final 5xx after final attempt
logger.error(f"HTTP request failed with status code {response.status_code}: {http_err}")
raise
except requests.RequestException as e:
# ConnectionError or other non-HTTPError
logger.error(f"HTTP request encountered a network issue: {e}")
if attempt == self.max_retries - 1:
raise
# Else retry on next loop iteration
backoff_time = base_delay * (2**attempt)
time.sleep(backoff_time)
attempt += 1
# If we exit the loop without returning, we've exhausted all attempts
logger.error(f"Failed to get a successful response after {self.max_retries} retries.")
raise Exception(f"Failed to get a successful response after {self.max_retries} retries.")
def _batcher_loop(self):
"""The main loop for the background thread to form and process batches."""
while not self._stop_event.is_set():
requests_batch = []
try:
first_req = self._request_queue.get(timeout=self._batch_timeout)
if first_req is None:
continue
requests_batch.append(first_req)
start_time = time.monotonic()
while len(requests_batch) < self._batch_size:
if (time.monotonic() - start_time) >= self._batch_timeout:
break
if self._request_queue.empty():
break
next_req_peek = self._request_queue.queue[0]
if next_req_peek is None:
break
if self._batch_memory_budget_bytes:
if not self.model_interface.does_item_fit_in_batch(
requests_batch,
next_req_peek,
self._batch_memory_budget_bytes,
):
break
try:
next_req = self._request_queue.get_nowait()
if next_req is None:
break
requests_batch.append(next_req)
except queue.Empty:
break
except queue.Empty:
continue
if requests_batch:
self._process_dynamic_batch(requests_batch)
def _process_dynamic_batch(self, requests: list[InferenceRequest]):
"""Coalesces, infers, and distributes results for a dynamic batch."""
if not requests:
return
first_req = requests[0]
model_name = first_req.model_name
kwargs = first_req.kwargs
try:
# 1. Coalesce individual data items into a single batch input
batch_input, batch_data = self.model_interface.coalesce_requests_to_batch(
[req.data for req in requests],
[req.dims for req in requests],
protocol=self.protocol,
model_name=model_name,
**kwargs,
)
# 2. Perform inference using the existing _process_batch logic
parsed_output, _ = self._process_batch(batch_input, batch_data=batch_data, model_name=model_name, **kwargs)
# 3. Process the batched output to get final results
all_results = self.model_interface.process_inference_results(
parsed_output,
original_image_shapes=batch_data.get("original_image_shapes"),
protocol=self.protocol,
**kwargs,
)
# 4. Distribute the individual results back to the correct Future
if len(all_results) != len(requests):
raise ValueError("Mismatch between result count and request count.")
for i, req in enumerate(requests):
req.future.set_result(all_results[i])
except Exception as e:
# If anything fails, propagate the exception to all futures in the batch
logger.error(f"Error processing dynamic batch: {e}")
for req in requests:
req.future.set_exception(e)
[docs]
def submit(self, data: Any, model_name: str, dims: Tuple[int, int], **kwargs) -> Future:
"""
Submits a single inference request to the dynamic batcher.
This method is non-blocking and returns a Future object that will
eventually contain the inference result.
Parameters
----------
data : Any
The single data item for inference (e.g., one image, one text prompt).
Returns
-------
concurrent.futures.Future
A future that will be fulfilled with the inference result.
"""
if not self.dynamic_batching_enabled:
raise RuntimeError(
"Dynamic batching is not enabled. Please initialize NimClient with " "enable_dynamic_batching=True."
)
future = Future()
request = InferenceRequest(data=data, future=future, model_name=model_name, dims=dims, kwargs=kwargs)
self._request_queue.put(request)
return future
[docs]
def close(self):
"""Stops the dynamic batching worker and closes client connections."""
if self.dynamic_batching_enabled:
self._stop_event.set()
# Unblock the queue in case the thread is waiting on get()
self._request_queue.put(None)
if self._batcher_thread.is_alive():
self._batcher_thread.join()
if self.client:
self.client.close()
[docs]
class NimClientManager:
"""
A thread-safe, singleton manager for creating and sharing NimClient instances.
This manager ensures that only one NimClient is created per unique configuration.
"""
_instance = None
_lock = threading.Lock()
def __new__(cls):
# Singleton pattern
if cls._instance is None:
with cls._lock:
if cls._instance is None:
cls._instance = super(NimClientManager, cls).__new__(cls)
return cls._instance
def __init__(self):
if not hasattr(self, "_initialized"):
with self._lock:
if not hasattr(self, "_initialized"):
self._clients = {} # Key: config_hash, Value: NimClient instance
self._client_lock = threading.Lock()
self._initialized = True
def _generate_config_key(self, **kwargs) -> str:
"""Creates a stable, hashable key from client configuration."""
sorted_config = sorted(kwargs.items())
config_str = json.dumps(sorted_config)
return hashlib.md5(config_str.encode("utf-8")).hexdigest()
[docs]
def get_client(self, model_interface, **kwargs) -> "NimClient":
"""
Gets or creates a NimClient for the given configuration.
"""
config_key = self._generate_config_key(model_interface_name=model_interface.name(), **kwargs)
if config_key in self._clients:
return self._clients[config_key]
with self._client_lock:
if config_key in self._clients:
return self._clients[config_key]
logger.debug(f"Creating new NimClient for config hash: {config_key}")
new_client = NimClient(model_interface=model_interface, **kwargs)
if new_client.dynamic_batching_enabled:
new_client.start()
self._clients[config_key] = new_client
return new_client
[docs]
def shutdown(self):
"""
Gracefully closes all managed NimClient instances.
This is called automatically on application exit by `atexit`.
"""
logger.debug(f"Shutting down NimClientManager and {len(self._clients)} client(s)...")
with self._client_lock:
for config_key, client in self._clients.items():
logger.debug(f"Closing client for config: {config_key}")
try:
client.close()
except Exception as e:
logger.error(f"Error closing client for config {config_key}: {e}")
self._clients.clear()
logger.debug("NimClientManager shutdown complete.")
# A global helper function to make access even easier
[docs]
def get_nim_client_manager(*args, **kwargs) -> NimClientManager:
"""Returns the singleton instance of the NimClientManager."""
return NimClientManager(*args, **kwargs)
[docs]
def reload_models(client: grpcclient.InferenceServerClient, exclude: list[str] = [], client_timeout: int = 120) -> bool:
"""
Reloads all models in the Triton server except for the models in the exclude list.
Parameters
----------
client : grpcclient.InferenceServerClient
The gRPC client connected to the Triton server.
exclude : list[str], optional
A list of model names to exclude from reloading.
client_timeout : int, optional
Timeout for client operations in seconds (default: 120).
Returns
-------
bool
True if all models were successfully reloaded, False otherwise.
"""
model_index = client.get_model_repository_index()
exclude = set(exclude)
names = [m.name for m in model_index.models if m.name not in exclude]
logger.info(f"Reloading {len(names)} model(s): {', '.join(names) if names else '(none)'}")
# 1) Unload
for name in names:
try:
client.unload_model(name)
except grpcclient.InferenceServerException as e:
msg = e.message()
if "explicit model load / unload" in msg.lower():
status = e.status()
logger.warning(
f"[SKIP Model Reload] Explicit model control disabled; cannot unload '{name}'. Status: {status}."
)
return False
logger.error(f"[ERROR] Failed to unload '{name}': {msg}")
return False
# 2) Load
for name in names:
client.load_model(name)
# 3) Readiness check
for name in names:
ready = client.is_model_ready(model_name=name, client_timeout=client_timeout)
if not ready:
logger.warning(f"[Warning] Triton Not ready: {name}")
return False
logger.info("✅ Reload of models complete.")
return True
