Merge branch 'main' into spark-weekly-newcases

.github/workflows/blossom-ci.yml

@@ -156,6 +156,7 @@ jobs:
         "kaiyux",
         "kanghui0204",
         "karljang",
+        "karthikvetrivel",
         "katec846",
         "Kefeng-Duan",
         "KingsleyLiu-NV",
@@ -193,6 +194,7 @@ jobs:
         "mlefeb01",
         "moraxu",
         "MrGeva",
+        "mzweilz",
         "Naveassaf",
         "nekorobov",
         "netanel-haber",

jenkins/L0_Test.groovy

@@ -874,7 +874,6 @@ def getMountListForSlurmTest(SlurmCluster cluster, boolean useSbatch = false)
         }
         mounts += [
             "${cluster.scratchPath}:/scratch.trt_llm_data:ro",
-            "/home/svc_tensorrt/.cache:/root/.cache",
         ]
     } else {
         throw new Exception("Unsupported container runtime: ${cluster.containerRuntime}")

tensorrt_llm/_common.py

@@ -16,6 +16,7 @@ import contextlib
 import ctypes
 import os
 import platform
+import threading
 import time
 from functools import wraps
 from pathlib import Path
@@ -34,7 +35,7 @@ if TYPE_CHECKING:
 else:
     Network = None
 
-from ._utils import str_dtype_to_trt
+from ._utils import print_all_stacks, str_dtype_to_trt
 from .bindings import MpiComm
 from .logger import logger
 from .plugin import _load_plugin_lib
@@ -82,6 +83,19 @@ def _init(log_level: object = None) -> None:
 
     MpiComm.local_init()
 
+    def _print_stacks():
+        counter = 0
+        while True:
+            time.sleep(print_stacks_period)
+            counter += 1
+            logger.error(f"Printing stacks {counter} times")
+            print_all_stacks()
+
+    print_stacks_period = int(os.getenv("TRTLLM_PRINT_STACKS_PERIOD", "-1"))
+    if print_stacks_period > 0:
+        print_stacks_thread = threading.Thread(target=_print_stacks, daemon=True)
+        print_stacks_thread.start()
+
     logger.info("TensorRT LLM inited.")
 
 

tensorrt_llm/_torch/pyexecutor/executor_request_queue.py

@@ -14,6 +14,7 @@ from tensorrt_llm._utils import mpi_disabled, nvtx_range
 from tensorrt_llm.mapping import CpType
 
 from ..distributed import Distributed
+from .hang_detector import HangDetector
 from .llm_request import (ExecutorRequest, LlmRequest,
                           executor_request_to_llm_request)
 
@@ -47,10 +48,17 @@ class RequestQueueItem:
 class ExecutorRequestQueue:
     """Handles fetching and processing of new requests from the request queue."""
 
-    def __init__(self, dist: Distributed, enable_attention_dp: bool,
-                 max_batch_size: int, max_beam_width: int,
-                 max_num_active_requests: int, enable_iter_perf_stats: bool,
-                 batch_wait_timeout_ms: float):
+    def __init__(
+        self,
+        dist: Distributed,
+        enable_attention_dp: bool,
+        max_batch_size: int,
+        max_beam_width: int,
+        max_num_active_requests: int,
+        enable_iter_perf_stats: bool,
+        batch_wait_timeout_ms: float,
+        hang_detector: Optional[HangDetector] = None,
+    ):
         self.dist = dist
         self.request_queue: queue.Queue[RequestQueueItem] = queue.Queue()
         self.waiting_queue: deque[RequestQueueItem] = deque()
@@ -66,6 +74,7 @@ class ExecutorRequestQueue:
         self.active = True
         self.batch_wait_timeout_ms = batch_wait_timeout_ms
         self.send_requests_handler = None
+        self.hang_detector = hang_detector or HangDetector()
 
         # State tracking
         self.num_fetch_requests = 0
@@ -303,7 +312,8 @@ class ExecutorRequestQueue:
                 self.request_accumulated.clear()
                 # Reset timeout to 0 to avoid hanging when no new requests are available
                 timeout = datetime.timedelta(0)
-            new_requests.extend(self._get_from_request_queue(timeout))
+            with self.hang_detector.pause():
+                new_requests.extend(self._get_from_request_queue(timeout))
 
         # Broadcast requests and handle Python objects
         new_requests, py_request_objects = self._handle_request_broadcasting(
@@ -477,8 +487,9 @@ class ExecutorRequestQueue:
             # Preserve original `new_requests` on rank 0
             _ = self._broadcast_new_requests(new_requests, py_request_objects)
         else:
-            new_requests, py_request_objects = self._broadcast_new_requests(
-                new_requests, py_request_objects)
+            with self.hang_detector.pause():
+                new_requests, py_request_objects = self._broadcast_new_requests(
+                    new_requests, py_request_objects)
 
         return new_requests, py_request_objects
 

tensorrt_llm/_torch/pyexecutor/hang_detector.py

@@ -0,0 +1,96 @@
+import asyncio
+import threading
+from contextlib import contextmanager
+from typing import Callable, Optional
+
+from tensorrt_llm._utils import print_all_stacks
+from tensorrt_llm.logger import logger
+
+
+class HangDetector:
+    def __init__(
+        self, timeout: Optional[int] = None, on_detected: Optional[Callable[[], None]] = None
+    ):
+        self.timeout = timeout if timeout is not None else 300
+        assert self.timeout > 0, "timeout must be greater than 0"
+        self.on_detected = on_detected or (lambda: None)
+        self.task = None
+        self.loop = None
+        self.loop_thread = None
+        self.lock = threading.Lock()
+        self.active = False
+        self._detected = False
+
+    def start(self):
+        """Enable hang detection."""
+
+        def run_loop():
+            asyncio.set_event_loop(self.loop)
+            self.loop.run_forever()
+
+        self.active = True
+        self.loop = asyncio.new_event_loop()
+        self.loop_thread = threading.Thread(target=run_loop, daemon=True, name="hang_detector_loop")
+        self.loop_thread.start()
+
+    async def _detect_hang(self):
+        await asyncio.sleep(self.timeout)
+        with self.lock:
+            self._detected = True
+            logger.error(f"Hang detected after {self.timeout} seconds.")
+            print_all_stacks()
+            self.on_detected()
+
+    def detected(self):
+        """Return True if hang is detected."""
+        with self.lock:
+            return self._detected
+
+    def checkpoint(self):
+        """Reset hang detection timer."""
+        self.cancel_task()
+        if self.active:
+            self.task = asyncio.run_coroutine_threadsafe(self._detect_hang(), self.loop)
+
+    def cancel_task(self):
+        """Cancel the hang detection task."""
+        if self.task is not None and not self.task.done():
+            self.task.cancel()
+            self.task = None
+
+    @contextmanager
+    def pause(self):
+        """Pause hang detection in scope."""
+        try:
+            self.cancel_task()
+            yield
+        finally:
+            self.checkpoint()
+
+    def stop(self):
+        """Stop hang detection."""
+        self.active = False
+        self.cancel_task()
+        if self.loop is not None:
+            # Cancel all pending tasks before stopping the loop
+            def cancel_all_tasks():
+                for task in asyncio.all_tasks(self.loop):
+                    if not task.done():
+                        task.cancel()
+                self.loop.call_soon(self.loop.stop)
+
+            self.loop.call_soon_threadsafe(cancel_all_tasks)
+
+            if self.loop_thread is not None and self.loop_thread.is_alive():
+                self.loop_thread.join()
+
+            self.loop = None
+            self.loop_thread = None
+
+    def __enter__(self):
+        self.start()
+        return self
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        self.stop()
+        return False

tensorrt_llm/_torch/pyexecutor/py_executor.py

@@ -46,6 +46,7 @@ from .executor_request_queue import ExecutorRequestQueue, RequestQueueItem
 from .guided_decoder import GuidedDecoder
 from .handle_additional_outputs import HandleAdditionalOutputs
 from .handle_logits import HandleLogits
+from .hang_detector import HangDetector
 from .kv_cache_connector import KvCacheConnectorManager
 from .kv_cache_transceiver import KvCacheTransceiver
 from .llm_request import (ExecutorRequest, LlmRequest, LlmRequestState,
@@ -137,6 +138,7 @@ class PyExecutor:
                  max_seq_len: Optional[int] = None,
                  peft_cache_config: Optional[PeftCacheConfig] = None,
                  virtual_memory_pools: Optional[dict] = None,
+                 hang_detection_timeout: Optional[int] = None,
                  execution_stream: Optional[torch.cuda.Stream] = None):
         super(PyExecutor, self).__init__()
         self.device_id = torch.cuda.current_device()
@@ -280,6 +282,15 @@ class PyExecutor:
         self.adp_ctx_batching_wait_iters_count = 0
         self.batch_wait_iters_count = 0
 
+        def on_detected():
+            self._handle_errors(
+                f"Hang detected on rank {self.global_rank} in PyExecutor.")
+            self.shutdown_event.set()
+            self.is_shutdown = True
+
+        self.hang_detector = HangDetector(timeout=hang_detection_timeout,
+                                          on_detected=on_detected)
+
         # request fetcher initialization
         self._set_global_steady_clock_offset()
         self.executor_request_queue = ExecutorRequestQueue(
@@ -290,6 +301,7 @@ class PyExecutor:
             max_num_active_requests=self.max_num_active_requests,
             enable_iter_perf_stats=self.enable_iter_perf_stats,
             batch_wait_timeout_ms=self.batch_wait_timeout_ms,
+            hang_detector=self.hang_detector,
         )
         self.executor_request_queue.set_exclude_last_generation_logits(
             self.disable_overlap_scheduler, self.dist.pp_size)
@@ -476,6 +488,14 @@ class PyExecutor:
         """
         self.executor_request_queue.enqueue_shutdown_request()
         self.shutdown_event.wait()
+        if self.hang_detector.detected():
+            # Early return here to avoid waiting for hanging threads.
+            # Since `on_detected` has sent the error message as response,
+            # this worker will be asked to shutdown immediately.
+            # Since the whole process will shutdown after this `shutdown` call,
+            # All threads and memory pools will be freed properly.
+            logger.error("Hang detected, shutting down immediately.")
+            return
         self.worker_thread.join()
         self.worker_started = False
         for manager in self.resource_manager.resource_managers.values():
@@ -960,10 +980,11 @@ class PyExecutor:
         # ensure the context is created, otherwise, some MPI calls will fail.
         CUASSERT(cudart.cudaSetDevice(self.device_id))
         microbatch_id = 0
-        with self._profiler() as profile_step:
+        with self._profiler() as profile_step, self.hang_detector:
             iter_start_time = time.time()
             iter_stats = None
             while True:
+                self.hang_detector.checkpoint()
                 profile_step()
                 if self.enable_iter_perf_stats:
                     iter_start_time = time.time()
@@ -1349,11 +1370,12 @@ class PyExecutor:
         torch.cuda.set_device(self.device_id)
         # ensure the context is created, otherwise, some MPI calls will fail.
         CUASSERT(cudart.cudaSetDevice(self.device_id))
-        with self._profiler() as profile_step:
+        with self._profiler() as profile_step, self.hang_detector:
             sample_state = None
             iter_start_time = time.time()
             iter_stats = None
             while True:
+                self.hang_detector.checkpoint()
                 profile_step()
                 if self.enable_iter_perf_stats:
                     iter_start_time = time.time()
@@ -1551,13 +1573,14 @@ class PyExecutor:
         torch.cuda.set_device(self.device_id)
         # ensure the context is created, otherwise, some MPI calls will fail.
         CUASSERT(cudart.cudaSetDevice(self.device_id))
-        with self._profiler() as profile_step:
+        with self._profiler() as profile_step, self.hang_detector:
             iter_start_time = time.time()
             iter_stats = None
             target_inputs = None
             previous_tensors_device = None
             can_forward = False if self.benchmark_req_queues_size > 0 and self.kv_cache_transceiver else True
             while True:
+                self.hang_detector.checkpoint()
                 profile_step()
                 if self.enable_iter_perf_stats:
                     iter_start_time = time.time()

tensorrt_llm/_utils.py

@@ -21,8 +21,10 @@ import math
 import os
 import socket
 import struct
+import sys
 import tempfile
 import trace
+import traceback
 import weakref
 from contextlib import contextmanager
 from enum import EnumMeta
@@ -761,6 +763,13 @@ def is_sm_100f(sm_version=None):
     return sm_version == 100 or sm_version == 103
 
 
+def print_all_stacks():
+    """Print stack traces for all threads"""
+    for thread_id, frame in sys._current_frames().items():
+        logger.error(f"Thread {thread_id} stack trace:\n" +
+                     "".join(traceback.format_stack(frame)))
+
+
 def is_trace_enabled(env_var: str):
     value = os.environ.get(env_var, "-1")
     if value == "ALL":

tensorrt_llm/executor/worker.py

@@ -1,5 +1,7 @@
 import gc
 import os
+import threading
+import time
 import traceback
 from concurrent.futures import ProcessPoolExecutor
 from pathlib import Path
@@ -9,7 +11,7 @@ import zmq
 
 from tensorrt_llm.logger import logger
 
-from .._utils import mpi_comm, mpi_rank
+from .._utils import mpi_comm, mpi_rank, print_all_stacks
 from ..bindings import executor as tllm
 from ..builder import Engine
 from ..llmapi.llm_args import BaseLlmArgs
@@ -153,6 +155,21 @@ def worker_main(
     hmac_key: Optional[bytes] = None,
 ) -> None:
 
+    def _print_stacks():
+        counter = 0
+        while True:
+            time.sleep(print_stacks_period)
+            counter += 1
+            logger.error(f"Printing stacks {counter} times")
+            print_all_stacks()
+
+    print_stacks_period = int(
+        os.getenv("TRTLLM_WORKER_PRINT_STACKS_PERIOD", "-1"))
+    if print_stacks_period > 0:
+        print_stacks_thread = threading.Thread(target=_print_stacks,
+                                               daemon=True)
+        print_stacks_thread.start()
+
     mpi_comm().barrier()
 
     if llm_args is not None and llm_args.env_overrides: