Add WideEP error classification patterns

Signed-off-by: Chien-Chun Hung <2679986+chienchunhung@users.noreply.github.com>

Author: chienchunhung

tensorrt_llm/_torch/pyexecutor/error_classification.py

@@ -26,9 +26,25 @@
 IMMEDIATE_FATAL_PATTERNS: list[str] = [
     "cudaerrorillegaladdress",
     "cudaerrorlaunchfailure",
+    "cudaerrorcontextisdestroyed",
+    "cudaerrornodevice",
     "illegal memory access",
     "device-side assert",
+    "cuda context destroyed",
+    "cuda context is destroyed",
+    "cuda context was destroyed",
+    "cuda error no device",
+    "gpu has fallen off the bus",
+    "xid 79",
     "unrecoverable",
+    "nccl communicator abort",
+    "nccl communicator was aborted",
+    "ncclcommabort",
+    "nvshmem peer unreachable",
+    "mpi rank terminated",
+    "mpi worker terminated",
+    "mpi rank exited unexpectedly",
+    "mpi worker exited unexpectedly",
 ]
 
 # Patterns that are serious but may be transient (e.g. a single OOM
@@ -38,6 +54,20 @@
     "cuda out of memory",
     "cuda error",
     "nccl error",
+    "nccl timeout",
+    "nccl operation timed out",
+    "alltoall timeout",
+    "all-to-all timeout",
+    "alltoall watchdog",
+    "completion_flags timeout",
+    "deep_ep buffer barrier hang",
+    "deepep buffer barrier hang",
+    "intranode::barrier",
+    "symmetric memory access violation",
+    "rdma timeout",
+    "nixl transfer failed",
+    "nixl transfer entered error state",
+    "nixl transfer wait timed out",
 ]
 
 

tests/integration/test_lists/test-db/l0_a10.yml

@@ -35,6 +35,7 @@ l0_a10:
   - unittest/_torch/executor/test_kv_pool_rebalance.py
   - unittest/_torch/executor/test_disagg_index_mapper_early_release.py
   - unittest/_torch/pyexecutor/test_kv_cache_compression_manager.py
+  - unittest/_torch/pyexecutor/test_error_classification.py
   - unittest/_torch/modules/dwdp/test_dwdp_fixup_moe_backends.py
   - unittest/_torch/modules/dwdp/test_dwdp_manager.py
   - unittest/_torch/modules/dwdp/test_dwdp_mapping.py

tests/unittest/_torch/pyexecutor/test_error_classification.py

@@ -0,0 +1,93 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import importlib.util
+import sys
+from pathlib import Path
+
+import pytest
+
+_MODULE_PATH = (
+    Path(__file__).resolve().parents[4] / "tensorrt_llm/_torch/pyexecutor/error_classification.py"
+)
+_SPEC = importlib.util.spec_from_file_location("error_classification_under_test", _MODULE_PATH)
+assert _SPEC is not None
+assert _SPEC.loader is not None
+_ERROR_CLASSIFICATION = importlib.util.module_from_spec(_SPEC)
+sys.modules[_SPEC.name] = _ERROR_CLASSIFICATION
+_SPEC.loader.exec_module(_ERROR_CLASSIFICATION)
+
+ErrorBudget = _ERROR_CLASSIFICATION.ErrorBudget
+classify_error = _ERROR_CLASSIFICATION.classify_error
+
+
+@pytest.mark.parametrize(
+    "message",
+    [
+        "CUDA error: cudaErrorNoDevice on rank 37",
+        "NVRM: Xid 79, GPU has fallen off the bus",
+        "CUDA context is destroyed after device reset",
+        "CUDA context was destroyed after device reset",
+        "NCCL communicator was aborted during PP send",
+        "NVSHMEM peer unreachable for EP rank 12",
+        "MPI rank terminated with signal 9",
+    ],
+)
+def test_wide_ep_immediate_fatal_patterns(message):
+    assert classify_error(message) == "immediate_fatal"
+
+
+@pytest.mark.parametrize(
+    "message",
+    [
+        "AlltoAll timeout waiting for completion_flags[3][7]",
+        "AlltoAll watchdog timed out waiting for peer rank 7",
+        "NCCL operation timed out during all_reduce",
+        "DeepEP Buffer.__del__ hung in intranode::barrier",
+        "Symmetric memory access violation reading dead peer",
+        "RDMA timeout on cross-node transfer",
+        "NIXL transfer failed: remote peer closed",
+        "NIXL transfer entered error state",
+        "NIXL transfer wait timed out after 5000 ms",
+    ],
+)
+def test_wide_ep_severe_patterns(message):
+    assert classify_error(message) == "severe"
+
+
+@pytest.mark.parametrize(
+    "message",
+    [
+        "AlltoAll slow path observed for rank 3",
+        "NCCL retry scheduled after transient transport hiccup",
+        "ECC correctable error threshold warning",
+    ],
+)
+def test_wide_ep_nonfatal_signals_remain_transient(message):
+    assert classify_error(message) == "transient"
+
+
+def test_error_budget_charges_wide_ep_severe_patterns():
+    budget = ErrorBudget(recovery_rate=0.0)
+
+    assert not budget.consume("NIXL transfer failed: remote peer closed")
+    assert budget.consume("AlltoAll timeout waiting for completion_flags[0][2]")
+
+
+def test_error_budget_bypasses_budget_for_wide_ep_fatal_patterns():
+    budget = ErrorBudget(recovery_rate=0.0)
+
+    assert budget.consume("MPI rank terminated with signal 9")
+    assert budget.budget == 1.0