[Help]: yolo11s将onnx编译为TensorRT文件时候 报错：Cannot find plugin: EfficientIdxNMS_TRT #170
Description
问题确认 Search before asking
- 我已经搜索过问题，但是没有找到解答。I have searched the question and found no related answer.
请提出你的问题 Please ask your question
我使用以下命令将ONNX转为TensorRT模型
(base) nvidia@ubuntu:~/orePassAIShaftGuard/TensorRT-YOLO/models$ /usr/src/tensorrt/bin/trtexec \
--onnx=onnx/yolo11s.onnx \
--saveEngine=models/yolo11s.engine \
--fp16
&&&& RUNNING TensorRT.trtexec [TensorRT v100300] # /usr/src/tensorrt/bin/trtexec --onnx=onnx/yolo11s.onnx --saveEngine=models/yolo11s.engine --fp16
[06/05/2025-10:50:40] [I] === Model Options ===
[06/05/2025-10:50:40] [I] Format: ONNX
[06/05/2025-10:50:40] [I] Model: onnx/yolo11s.onnx
[06/05/2025-10:50:40] [I] Output:
[06/05/2025-10:50:40] [I] === Build Options ===
[06/05/2025-10:50:40] [I] Memory Pools: workspace: default, dlaSRAM: default, dlaLocalDRAM: default, dlaGlobalDRAM: default, tacticSharedMem: default
[06/05/2025-10:50:40] [I] avgTiming: 8
[06/05/2025-10:50:40] [I] Precision: FP32+FP16
[06/05/2025-10:50:40] [I] LayerPrecisions:
[06/05/2025-10:50:40] [I] Layer Device Types:
[06/05/2025-10:50:40] [I] Calibration:
[06/05/2025-10:50:40] [I] Refit: Disabled
[06/05/2025-10:50:40] [I] Strip weights: Disabled
[06/05/2025-10:50:40] [I] Version Compatible: Disabled
[06/05/2025-10:50:40] [I] ONNX Plugin InstanceNorm: Disabled
[06/05/2025-10:50:40] [I] TensorRT runtime: full
[06/05/2025-10:50:40] [I] Lean DLL Path:
[06/05/2025-10:50:40] [I] Tempfile Controls: { in_memory: allow, temporary: allow }
[06/05/2025-10:50:40] [I] Exclude Lean Runtime: Disabled
[06/05/2025-10:50:40] [I] Sparsity: Disabled
[06/05/2025-10:50:40] [I] Safe mode: Disabled
[06/05/2025-10:50:40] [I] Build DLA standalone loadable: Disabled
[06/05/2025-10:50:40] [I] Allow GPU fallback for DLA: Disabled
[06/05/2025-10:50:40] [I] DirectIO mode: Disabled
[06/05/2025-10:50:40] [I] Restricted mode: Disabled
[06/05/2025-10:50:40] [I] Skip inference: Disabled
[06/05/2025-10:50:40] [I] Save engine: models/yolo11s.engine
[06/05/2025-10:50:40] [I] Load engine:
[06/05/2025-10:50:40] [I] Profiling verbosity: 0
[06/05/2025-10:50:40] [I] Tactic sources: Using default tactic sources
[06/05/2025-10:50:40] [I] timingCacheMode: local
[06/05/2025-10:50:40] [I] timingCacheFile:
[06/05/2025-10:50:40] [I] Enable Compilation Cache: Enabled
[06/05/2025-10:50:40] [I] errorOnTimingCacheMiss: Disabled
[06/05/2025-10:50:40] [I] Preview Features: Use default preview flags.
[06/05/2025-10:50:40] [I] MaxAuxStreams: -1
[06/05/2025-10:50:40] [I] BuilderOptimizationLevel: -1
[06/05/2025-10:50:40] [I] Calibration Profile Index: 0
[06/05/2025-10:50:40] [I] Weight Streaming: Disabled
[06/05/2025-10:50:40] [I] Runtime Platform: Same As Build
[06/05/2025-10:50:40] [I] Debug Tensors:
[06/05/2025-10:50:40] [I] Input(s)s format: fp32:CHW
[06/05/2025-10:50:40] [I] Output(s)s format: fp32:CHW
[06/05/2025-10:50:40] [I] Input build shapes: model
[06/05/2025-10:50:40] [I] Input calibration shapes: model
[06/05/2025-10:50:40] [I] === System Options ===
[06/05/2025-10:50:40] [I] Device: 0
[06/05/2025-10:50:40] [I] DLACore:
[06/05/2025-10:50:40] [I] Plugins:
[06/05/2025-10:50:40] [I] setPluginsToSerialize:
[06/05/2025-10:50:40] [I] dynamicPlugins:
[06/05/2025-10:50:40] [I] ignoreParsedPluginLibs: 0
[06/05/2025-10:50:40] [I]
[06/05/2025-10:50:40] [I] === Inference Options ===
[06/05/2025-10:50:40] [I] Batch: Explicit
[06/05/2025-10:50:40] [I] Input inference shapes: model
[06/05/2025-10:50:40] [I] Iterations: 10
[06/05/2025-10:50:40] [I] Duration: 3s (+ 200ms warm up)
[06/05/2025-10:50:40] [I] Sleep time: 0ms
[06/05/2025-10:50:40] [I] Idle time: 0ms
[06/05/2025-10:50:40] [I] Inference Streams: 1
[06/05/2025-10:50:40] [I] ExposeDMA: Disabled
[06/05/2025-10:50:40] [I] Data transfers: Enabled
[06/05/2025-10:50:40] [I] Spin-wait: Disabled
[06/05/2025-10:50:40] [I] Multithreading: Disabled
[06/05/2025-10:50:40] [I] CUDA Graph: Disabled
[06/05/2025-10:50:40] [I] Separate profiling: Disabled
[06/05/2025-10:50:40] [I] Time Deserialize: Disabled
[06/05/2025-10:50:40] [I] Time Refit: Disabled
[06/05/2025-10:50:40] [I] NVTX verbosity: 0
[06/05/2025-10:50:40] [I] Persistent Cache Ratio: 0
[06/05/2025-10:50:40] [I] Optimization Profile Index: 0
[06/05/2025-10:50:40] [I] Weight Streaming Budget: 100.000000%
[06/05/2025-10:50:40] [I] Inputs:
[06/05/2025-10:50:40] [I] Debug Tensor Save Destinations:
[06/05/2025-10:50:40] [I] === Reporting Options ===
[06/05/2025-10:50:40] [I] Verbose: Disabled
[06/05/2025-10:50:40] [I] Averages: 10 inferences
[06/05/2025-10:50:40] [I] Percentiles: 90,95,99
[06/05/2025-10:50:40] [I] Dump refittable layers:Disabled
[06/05/2025-10:50:40] [I] Dump output: Disabled
[06/05/2025-10:50:40] [I] Profile: Disabled
[06/05/2025-10:50:40] [I] Export timing to JSON file:
[06/05/2025-10:50:40] [I] Export output to JSON file:
[06/05/2025-10:50:40] [I] Export profile to JSON file:
[06/05/2025-10:50:40] [I]
[06/05/2025-10:50:40] [I] === Device Information ===
[06/05/2025-10:50:40] [I] Available Devices:
[06/05/2025-10:50:40] [I] Device 0: "Orin" UUID: GPU-2e0217d9-517f-5251-afc0-72a66a1ea757
[06/05/2025-10:50:40] [I] Selected Device: Orin
[06/05/2025-10:50:40] [I] Selected Device ID: 0
[06/05/2025-10:50:40] [I] Selected Device UUID: GPU-2e0217d9-517f-5251-afc0-72a66a1ea757
[06/05/2025-10:50:40] [I] Compute Capability: 8.7
[06/05/2025-10:50:40] [I] SMs: 8
[06/05/2025-10:50:40] [I] Device Global Memory: 7619 MiB
[06/05/2025-10:50:40] [I] Shared Memory per SM: 164 KiB
[06/05/2025-10:50:40] [I] Memory Bus Width: 128 bits (ECC disabled)
[06/05/2025-10:50:40] [I] Application Compute Clock Rate: 0.765 GHz
[06/05/2025-10:50:40] [I] Application Memory Clock Rate: 0.765 GHz
[06/05/2025-10:50:40] [I]
[06/05/2025-10:50:40] [I] Note: The application clock rates do not reflect the actual clock rates that the GPU is currently running at.
[06/05/2025-10:50:40] [I]
[06/05/2025-10:50:40] [I] TensorRT version: 10.3.0
[06/05/2025-10:50:40] [I] Loading standard plugins
[06/05/2025-10:50:40] [I] [TRT] [MemUsageChange] Init CUDA: CPU +2, GPU +0, now: CPU 31, GPU 2262 (MiB)
[06/05/2025-10:50:42] [I] [TRT] [MemUsageChange] Init builder kernel library: CPU +928, GPU +932, now: CPU 1002, GPU 3238 (MiB)
[06/05/2025-10:50:42] [I] Start parsing network model.
[06/05/2025-10:50:42] [I] [TRT] ----------------------------------------------------------------
[06/05/2025-10:50:42] [I] [TRT] Input filename: onnx/yolo11s.onnx
[06/05/2025-10:50:42] [I] [TRT] ONNX IR version: 0.0.7
[06/05/2025-10:50:42] [I] [TRT] Opset version: 12
[06/05/2025-10:50:42] [I] [TRT] Producer name: pytorch
[06/05/2025-10:50:42] [I] [TRT] Producer version: 2.7.0
[06/05/2025-10:50:42] [I] [TRT] Domain:
[06/05/2025-10:50:42] [I] [TRT] Model version: 0
[06/05/2025-10:50:42] [I] [TRT] Doc string:
[06/05/2025-10:50:42] [I] [TRT] ----------------------------------------------------------------
[06/05/2025-10:50:42] [I] [TRT] No checker registered for op: EfficientIdxNMS_TRT. Attempting to check as plugin.
[06/05/2025-10:50:42] [E] [TRT] IPluginRegistry::getCreator: Error Code 4: API Usage Error (Cannot find plugin: EfficientIdxNMS_TRT, version: 1, namespace:.)
[06/05/2025-10:50:42] [E] [TRT] ModelImporter.cpp:948: While parsing node number 365 [EfficientIdxNMS_TRT -> "num_dets"]:
[06/05/2025-10:50:42] [E] [TRT] ModelImporter.cpp:950: --- Begin node ---
input: "/model.23/Transpose_1_output_0"
input: "/model.23/Transpose_2_output_0"
output: "num_dets"
output: "det_boxes"
output: "det_scores"
output: "det_classes"
output: "/model.23/EfficientIdxNMS_TRT_output_4"
name: "/model.23/EfficientIdxNMS_TRT"
op_type: "EfficientIdxNMS_TRT"
attribute {
name: "background_class"
i: -1
type: INT
}
attribute {
name: "box_coding"
i: 1
type: INT
}
attribute {
name: "class_agnostic"
i: 1
type: INT
}
attribute {
name: "iou_threshold"
f: 0.45
type: FLOAT
}
attribute {
name: "max_output_boxes"
i: 20
type: INT
}
attribute {
name: "plugin_version"
s: "1"
type: STRING
}
attribute {
name: "score_activation"
i: 1
type: INT
}
attribute {
name: "score_threshold"
f: 0.25
type: FLOAT
}
domain: "TRT"
[06/05/2025-10:50:42] [E] [TRT] ModelImporter.cpp:951: --- End node ---
[06/05/2025-10:50:42] [E] [TRT] ModelImporter.cpp:953: ERROR: onnxOpCheckers.cpp:780 In function checkFallbackPluginImporter:
[6] creator && "Plugin not found, are the plugin name, version, and namespace correct?"
[06/05/2025-10:50:42] [E] Failed to parse onnx file
[06/05/2025-10:50:42] [I] Finished parsing network model. Parse time: 0.0697417
[06/05/2025-10:50:42] [E] Parsing model failed
[06/05/2025-10:50:42] [E] Failed to create engine from model or file.
[06/05/2025-10:50:42] [E] Engine set up failed
&&&& FAILED TensorRT.trtexec [TensorRT v100300] # /usr/src/tensorrt/bin/trtexec --onnx=onnx/yolo11s.onnx --saveEngine=models/yolo11s.engine --fp16
(base) nvidia@ubuntu:~/orePassAIShaftGuard/TensorRT-YOLO/models$
我的onnx来自：
trtyolo export -w yolo11s.pt -v yolo11 -o onnx --imgsz 320 -s --max_boxes 20
我的平台是：nvidia jetson orin nx + Jetpack 6.2 + TensorRT 10.3.0.30 + cuda 12.6.68
我成功编译了 TensorRT-YOLO 的插件（也许包含EfficientIdxNMS_TRT？） ：
(base) nvidia@ubuntu:~/orePassAIShaftGuard/TensorRT-YOLO$ tree lib/
lib/
├── libdeploy.so
└── plugin
└── libcustom_plugins.so
1 directory, 2 files
(base) nvidia@ubuntu:~/orePassAIShaftGuard/TensorRT-YOLO$
当我手动指定该so的时候：
(base) nvidia@ubuntu:~/orePassAIShaftGuard/TensorRT-YOLO/models$ /usr/src/tensorrt/bin/trtexec \
--onnx=onnx/yolo11s.onnx \
--saveEngine=models/yolo11s.engine \
--fp16 \
--plugins=/home/nvidia/orePassAIShaftGuard/TensorRT-YOLO/lib/plugin/libcustom_plugins.so
&&&& RUNNING TensorRT.trtexec [TensorRT v100300] # /usr/src/tensorrt/bin/trtexec --onnx=onnx/yolo11s.onnx --saveEngine=models/yolo11s.engine --fp16 --plugins=/home/nvidia/orePassAIShaftGuard/TensorRT-YOLO/lib/plugin/libcustom_plugins.so
[06/05/2025-10:58:23] [W] --plugins flag has been deprecated, use --staticPlugins flag instead.
[06/05/2025-10:58:23] [I] === Model Options ===
[06/05/2025-10:58:23] [I] Format: ONNX
[06/05/2025-10:58:23] [I] Model: onnx/yolo11s.onnx
[06/05/2025-10:58:23] [I] Output:
[06/05/2025-10:58:23] [I] === Build Options ===
[06/05/2025-10:58:23] [I] Memory Pools: workspace: default, dlaSRAM: default, dlaLocalDRAM: default, dlaGlobalDRAM: default, tacticSharedMem: default
[06/05/2025-10:58:23] [I] avgTiming: 8
[06/05/2025-10:58:23] [I] Precision: FP32+FP16
[06/05/2025-10:58:23] [I] LayerPrecisions:
[06/05/2025-10:58:23] [I] Layer Device Types:
[06/05/2025-10:58:23] [I] Calibration:
[06/05/2025-10:58:23] [I] Refit: Disabled
[06/05/2025-10:58:23] [I] Strip weights: Disabled
[06/05/2025-10:58:23] [I] Version Compatible: Disabled
[06/05/2025-10:58:23] [I] ONNX Plugin InstanceNorm: Disabled
[06/05/2025-10:58:23] [I] TensorRT runtime: full
[06/05/2025-10:58:23] [I] Lean DLL Path:
[06/05/2025-10:58:23] [I] Tempfile Controls: { in_memory: allow, temporary: allow }
[06/05/2025-10:58:23] [I] Exclude Lean Runtime: Disabled
[06/05/2025-10:58:23] [I] Sparsity: Disabled
[06/05/2025-10:58:23] [I] Safe mode: Disabled
[06/05/2025-10:58:23] [I] Build DLA standalone loadable: Disabled
[06/05/2025-10:58:23] [I] Allow GPU fallback for DLA: Disabled
[06/05/2025-10:58:23] [I] DirectIO mode: Disabled
[06/05/2025-10:58:23] [I] Restricted mode: Disabled
[06/05/2025-10:58:23] [I] Skip inference: Disabled
[06/05/2025-10:58:23] [I] Save engine: models/yolo11s.engine
[06/05/2025-10:58:23] [I] Load engine:
[06/05/2025-10:58:23] [I] Profiling verbosity: 0
[06/05/2025-10:58:23] [I] Tactic sources: Using default tactic sources
[06/05/2025-10:58:23] [I] timingCacheMode: local
[06/05/2025-10:58:23] [I] timingCacheFile:
[06/05/2025-10:58:23] [I] Enable Compilation Cache: Enabled
[06/05/2025-10:58:23] [I] errorOnTimingCacheMiss: Disabled
[06/05/2025-10:58:23] [I] Preview Features: Use default preview flags.
[06/05/2025-10:58:23] [I] MaxAuxStreams: -1
[06/05/2025-10:58:23] [I] BuilderOptimizationLevel: -1
[06/05/2025-10:58:23] [I] Calibration Profile Index: 0
[06/05/2025-10:58:23] [I] Weight Streaming: Disabled
[06/05/2025-10:58:23] [I] Runtime Platform: Same As Build
[06/05/2025-10:58:23] [I] Debug Tensors:
[06/05/2025-10:58:23] [I] Input(s)s format: fp32:CHW
[06/05/2025-10:58:23] [I] Output(s)s format: fp32:CHW
[06/05/2025-10:58:23] [I] Input build shapes: model
[06/05/2025-10:58:23] [I] Input calibration shapes: model
[06/05/2025-10:58:23] [I] === System Options ===
[06/05/2025-10:58:23] [I] Device: 0
[06/05/2025-10:58:23] [I] DLACore:
[06/05/2025-10:58:23] [I] Plugins: /home/nvidia/orePassAIShaftGuard/TensorRT-YOLO/lib/plugin/libcustom_plugins.so
[06/05/2025-10:58:23] [I] setPluginsToSerialize:
[06/05/2025-10:58:23] [I] dynamicPlugins:
[06/05/2025-10:58:23] [I] ignoreParsedPluginLibs: 0
[06/05/2025-10:58:23] [I]
[06/05/2025-10:58:23] [I] === Inference Options ===
[06/05/2025-10:58:23] [I] Batch: Explicit
[06/05/2025-10:58:23] [I] Input inference shapes: model
[06/05/2025-10:58:23] [I] Iterations: 10
[06/05/2025-10:58:23] [I] Duration: 3s (+ 200ms warm up)
[06/05/2025-10:58:23] [I] Sleep time: 0ms
[06/05/2025-10:58:23] [I] Idle time: 0ms
[06/05/2025-10:58:23] [I] Inference Streams: 1
[06/05/2025-10:58:23] [I] ExposeDMA: Disabled
[06/05/2025-10:58:23] [I] Data transfers: Enabled
[06/05/2025-10:58:23] [I] Spin-wait: Disabled
[06/05/2025-10:58:23] [I] Multithreading: Disabled
[06/05/2025-10:58:23] [I] CUDA Graph: Disabled
[06/05/2025-10:58:23] [I] Separate profiling: Disabled
[06/05/2025-10:58:23] [I] Time Deserialize: Disabled
[06/05/2025-10:58:23] [I] Time Refit: Disabled
[06/05/2025-10:58:23] [I] NVTX verbosity: 0
[06/05/2025-10:58:23] [I] Persistent Cache Ratio: 0
[06/05/2025-10:58:23] [I] Optimization Profile Index: 0
[06/05/2025-10:58:23] [I] Weight Streaming Budget: 100.000000%
[06/05/2025-10:58:23] [I] Inputs:
[06/05/2025-10:58:23] [I] Debug Tensor Save Destinations:
[06/05/2025-10:58:23] [I] === Reporting Options ===
[06/05/2025-10:58:23] [I] Verbose: Disabled
[06/05/2025-10:58:23] [I] Averages: 10 inferences
[06/05/2025-10:58:23] [I] Percentiles: 90,95,99
[06/05/2025-10:58:23] [I] Dump refittable layers:Disabled
[06/05/2025-10:58:23] [I] Dump output: Disabled
[06/05/2025-10:58:23] [I] Profile: Disabled
[06/05/2025-10:58:23] [I] Export timing to JSON file:
[06/05/2025-10:58:23] [I] Export output to JSON file:
[06/05/2025-10:58:23] [I] Export profile to JSON file:
[06/05/2025-10:58:23] [I]
[06/05/2025-10:58:23] [I] === Device Information ===
[06/05/2025-10:58:23] [I] Available Devices:
[06/05/2025-10:58:23] [I] Device 0: "Orin" UUID: GPU-2e0217d9-517f-5251-afc0-72a66a1ea757
[06/05/2025-10:58:23] [I] Selected Device: Orin
[06/05/2025-10:58:23] [I] Selected Device ID: 0
[06/05/2025-10:58:23] [I] Selected Device UUID: GPU-2e0217d9-517f-5251-afc0-72a66a1ea757
[06/05/2025-10:58:23] [I] Compute Capability: 8.7
[06/05/2025-10:58:23] [I] SMs: 8
[06/05/2025-10:58:23] [I] Device Global Memory: 7619 MiB
[06/05/2025-10:58:23] [I] Shared Memory per SM: 164 KiB
[06/05/2025-10:58:23] [I] Memory Bus Width: 128 bits (ECC disabled)
[06/05/2025-10:58:23] [I] Application Compute Clock Rate: 0.765 GHz
[06/05/2025-10:58:23] [I] Application Memory Clock Rate: 0.765 GHz
[06/05/2025-10:58:23] [I]
[06/05/2025-10:58:23] [I] Note: The application clock rates do not reflect the actual clock rates that the GPU is currently running at.
[06/05/2025-10:58:23] [I]
[06/05/2025-10:58:23] [I] TensorRT version: 10.3.0
[06/05/2025-10:58:23] [I] Loading standard plugins
[06/05/2025-10:58:23] [I] Loading supplied plugin library: /home/nvidia/orePassAIShaftGuard/TensorRT-YOLO/lib/plugin/libcustom_plugins.so
[06/05/2025-10:58:23] [I] [TRT] [MemUsageChange] Init CUDA: CPU +1, GPU +0, now: CPU 31, GPU 2267 (MiB)
[06/05/2025-10:58:25] [I] [TRT] [MemUsageChange] Init builder kernel library: CPU +928, GPU +930, now: CPU 1002, GPU 3242 (MiB)
[06/05/2025-10:58:25] [I] Start parsing network model.
[06/05/2025-10:58:25] [I] [TRT] ----------------------------------------------------------------
[06/05/2025-10:58:25] [I] [TRT] Input filename: onnx/yolo11s.onnx
[06/05/2025-10:58:25] [I] [TRT] ONNX IR version: 0.0.7
[06/05/2025-10:58:25] [I] [TRT] Opset version: 12
[06/05/2025-10:58:25] [I] [TRT] Producer name: pytorch
[06/05/2025-10:58:25] [I] [TRT] Producer version: 2.7.0
[06/05/2025-10:58:25] [I] [TRT] Domain:
[06/05/2025-10:58:25] [I] [TRT] Model version: 0
[06/05/2025-10:58:25] [I] [TRT] Doc string:
[06/05/2025-10:58:25] [I] [TRT] ----------------------------------------------------------------
[06/05/2025-10:58:25] [I] [TRT] No checker registered for op: EfficientIdxNMS_TRT. Attempting to check as plugin.
[06/05/2025-10:58:25] [I] [TRT] No importer registered for op: EfficientIdxNMS_TRT. Attempting to import as plugin.
[06/05/2025-10:58:25] [I] [TRT] Searching for plugin: EfficientIdxNMS_TRT, plugin_version: 1, plugin_namespace:
[06/05/2025-10:58:25] [I] [TRT] Successfully created plugin: EfficientIdxNMS_TRT
[06/05/2025-10:58:25] [I] Finished parsing network model. Parse time: 0.0669187
[06/05/2025-10:58:25] [I] [TRT] Local timing cache in use. Profiling results in this builder pass will not be stored.
[06/05/2025-11:05:37] [E] Error[9]: Error Code: 9: Skipping tactic 0x0000000000000000 due to exception Assertion pluginUtils::isSuccess(status) failed.
[06/05/2025-11:05:37] [E] Error[9]: Error Code: 9: Skipping tactic 0x0000000000000000 due to exception Assertion pluginUtils::isSuccess(status) failed.
[06/05/2025-11:05:37] [E] Error[10]: IBuilder::buildSerializedNetwork: Error Code 10: Internal Error (Could not find any implementation for node /model.23/EfficientIdxNMS_TRT.)
[06/05/2025-11:05:37] [E] Engine could not be created from network
[06/05/2025-11:05:37] [E] Building engine failed
[06/05/2025-11:05:37] [E] Failed to create engine from model or file.
[06/05/2025-11:05:37] [E] Engine set up failed
&&&& FAILED TensorRT.trtexec [TensorRT v100300] # /usr/src/tensorrt/bin/trtexec --onnx=onnx/yolo11s.onnx --saveEngine=models/yolo11s.engine --fp16 --plugins=/home/nvidia/orePassAIShaftGuard/TensorRT-YOLO/lib/plugin/libcustom_plugins.so
(base) nvidia@ubuntu:~/orePassAIShaftGuard/TensorRT-YOLO/models$