举例一个场景：
PP-DocLayoutV2 官方模型是 Paddle 格式：
inference.pdmodel / inference.json + inference.pdiparams
不是 PyTorch 模型。torch.export 只处理 PyTorch nn.Module，它不能直接读 Paddle 的 .pdmodel/.pdiparams。
即使你先把 Paddle 转 PyTorch，也不是简单命令，模型结构、权重名、算子都要适配，工程量很大。
另外你说的 opset=16，通常是：
torch.onnx.export(..., opset_version=16)
不是 torch.export。但前提仍然是：你已经有 PyTorch 模型。
所以当前可行性排序：
最推荐：找支持 opset17 的沐曦 Triton / ONNXRuntime 镜像-----你们官网竟然没有，能不能开发好和配套好再卖gpu
能不能都开发好和适配好，再卖gpu给第三方。资本市场赚钱了，也不给你们开发多一点开发好吗
-06-09 17:17:38.541219737 [I:onnxruntime:, session_state_utils.cc:339 SaveInitializedTensors] Done saving initialized tensors
[ERROR][2026-06-09 17:17:42.870][graph_parser.cc:190] unsupported op: domain[], type[GridSample], version[16]
[ERROR][2026-06-09 17:17:42.870][graph_parser.cc:232] ParseNodeInfo for node[GridSample.4] failed: unsupported
[ERROR][2026-06-09 17:17:42.870][graph_parser.cc:320] ParseGraphNode failed.
[ERROR][2026-06-09 17:17:42.870][model_parser.cc:80] parse graph failed: unsupported
[ERROR][2026-06-09 17:17:42.873][runtime_builder_impl.cc:48] parse graph failed: unsupported
[ERROR][2026-06-09 17:17:42.882][opt_graph.cc:299] cannot find creator for CudaOptKernel[GridSample.4] type[:GridSample]16
[ERROR][2026-06-09 17:17:42.882][opt_graph.cc:1075] init kernels failed: not found
[ERROR][2026-06-09 17:17:42.882][engine.cc:122] OptGraph DoOptimeize failed: not found
[ERROR][2026-06-09 17:17:42.882][engine.cc:183] DoOptimize failed: not found
[ERROR][2026-06-09 17:17:42.882][utils.cc:246] process graph[MACA_0] by engine[cuda] failed: not found
[ERROR][2026-06-09 17:17:42.882][utils.cc:395] GenPartitionsInfoAndShapes failed:not found
[ERROR][2026-06-09 17:17:42.882][runtime_builder_impl.cc:87] process graph failed: not found
[ERROR][2026-06-09 17:17:42.882][runtime_impl.cc:105] cannot find consumer[Conv.0] of [image]
[ERROR][2026-06-09 17:17:42.882][runtime_impl.cc:269] GenGraphInputs failed: not found
[ERROR][2026-06-09 17:17:42.882][runtime_impl.cc:305] InitGraphResources failed: not found
[ERROR][2026-06-09 17:17:42.882][runtime_builder_impl.cc:108] init runtime failed: not found
Signal (11) received.
0# triton::server::(anonymous namespace)::ErrorSignalHandler(int) at triton_signal.cc:?
1# __kernel_rt_sigreturn in linux-vdso.so.1
2# 0x0000FFFDD9044724 in /opt/maca-ai/onnxruntime-maca/lib/libonnxruntime.so.1.12.0
3# 0x0000FFFDD95B6580 in /opt/maca-ai/onnxruntime-maca/lib/libonnxruntime.so.1.12.0
4# 0x0000FFFDD95B6BD0 in /opt/maca-ai/onnxruntime-maca/lib/libonnxruntime.so.1.12.0
5# 0x0000FFFDD95C695C in /opt/maca-ai/onnxruntime-maca/lib/libonnxruntime.so.1.12.0
6# 0x0000FFFDD95D73C4 in /opt/maca-ai/onnxruntime-maca/lib/libonnxruntime.so.1.12.0
7# 0x0000FFFDD95DCC4C in /opt/maca-ai/onnxruntime-maca/lib/libonnxruntime.so.1.12.0
8# 0x0000FFFDD95E3968 in /opt/maca-ai/onnxruntime-maca/lib/libonnxruntime.so.1.12.0
9# 0x0000FFFDD90207D8 in /opt/maca-ai/onnxruntime-maca/lib/libonnxruntime.so.1.12.0
10# 0x0000FFFDD8FC63E4 in /opt/maca-ai/onnxruntime-maca/lib/libonnxruntime.so.1.12.0
11# 0x0000FFFDD8FC6608 in /opt/maca-ai/onnxruntime-maca/lib/libonnxruntime.so.1.12.0
12# 0x0000FFFEF0618690 in /opt/tritonserver/backends/onnxruntime/libtriton_onnxruntime.so
13# 0x0000FFFEF05F5DEC in /opt/tritonserver/backends/onnxruntime/libtriton_onnxruntime.so
14# 0x0000FFFEF05FB074 in /opt/tritonserver/backends/onnxruntime/libtriton_onnxruntime.so
15# 0x0000FFFEF05FC814 in /opt/tritonserver/backends/onnxruntime/libtriton_onnxruntime.so
16# TRITONBACKEND_ModelInstanceInitialize in /opt/tritonserver/backends/onnxruntime/libtriton_onnxruntime.so
17# 0x0000FFFF07670540 in /opt/tritonserver/bin/../lib/libtritonserver.so
18# 0x0000FFFF07671288 in /opt/tritonserver/bin/../lib/libtritonserver.so
19# 0x0000FFFF076562B8 in /opt/tritonserver/bin/../lib/libtritonserver.so
20# 0x0000FFFF076568CC in /opt/tritonserver/bin/../lib/libtritonserver.so
21# 0x0000FFFF07661A1C in /opt/tritonserver/bin/../lib/libtritonserver.so
22# 0x0000FFFF06014210 in /usr/lib64/libc.so.6
23# 0x0000FFFF0764E9AC in /opt/tritonserver/bin/../lib/libtritonserver.so
24# 0x0000FFFF0624F40C in /usr/lib64/libstdc++.so.6
25# 0x0000FFFF0600F630 in /usr/lib64/libc.so.6
26# 0x0000FFFF06075B9C in /usr/lib64/libc.so.6
[root@node1 triton_server]# docker run --rm \
--privileged \
--network host \
--ipc host \
-v /data/triton_server:/data/triton_server \
triton-metax:2.56-doclayout \
bash -lc '
python3 - <<PY
import onnx
p="/data/triton_server/model_repository/pp_doclayout_v2_20260415/doclayout_raw_gpu/1/model.onnx"
m=onnx.load(p)
ops=sorted(set(n.op_type for n in m.graph.node))
print("GridSample count:", sum(1 for n in m.graph.node if n.op_type=="GridSample"))
print("has GridSample:", "GridSample" in ops)
PY
'
GridSample count: 18
has GridSample: True
[root@node1 triton_server]#
显示更多
17:20
确认了，模型里有：
GridSample count: 18
has GridSample: True
所以现在可以定性了：
当前沐曦 onnxruntime-maca 1.12.0 不支持 GridSample
因此 doclayout_raw_gpu 无法用 MACAExecutionProvider 跑 GPU
这不是你操作问题，也不是模型包没解好。日志已经非常明确：
unsupported op: GridSample version[16]
下一步建议
先把 doclayout_raw_gpu 改成 CPU，验证整个 DocLayout 流程能跑通。执行：
CONFIG=/data/triton_server/model_repository/pp_doclayout_v2_20260415/doclayout_raw_gpu/config.pbtxt
cp -a "$CONFIG" "$CONFIG.gpu_maca_gridsample_unsupported.bak.$(date +%Y%m%d_%H%M%S)"
cat > "$CONFIG" <<'EOF'
name: "doclayout_raw_gpu"
backend: "onnxruntime"
max_batch_size: 0
input [
{
name: "image"
data_type: TYPE_FP32
dims: [1, 3, 800, 800]
}
]
output [
{
name: "fetch_name_0"
data_type: TYPE_FP32
dims: [1, 300, 4]
},
{
name: "fetch_name_1"
data_type: TYPE_FP32
dims: [1, 300, 25]
}
]
instance_group [
{
count: 1
kind: KIND_CPU
}
]
parameters: {
key: "execution_providers"
value: { string_value: "CPUExecutionProvider" }
}
EOF
然后重启 Triton，再加载：
cd /data/triton_server
bash dockers/pp_doclayout_v2_metax_ctl.sh stop
bash dockers/pp_doclayout_v2_metax_ctl.sh remove-container
export CUDA_VISIBLE_DEVICES=3
export HOST_TRITON_ROOT=/data/triton_server
export IMAGE_NAME='cr.metax-tech.com/public-ai-release/maca/triton-inference-server:2.56.0-maca.ai3.3.0.10-torch2.6-py310-kylin2309a-arm64'
bash dockers/pp_doclayout_v2_metax_ctl.sh start
这一步如果成功，说明：
preproc / raw / postproc / ensemble 全链路 OK
只是 raw 模型不能走 MACA GPU
如果必须 GPU
那就只能继续找支持 GridSample 的沐曦 ORT/Triton 镜像，筛选标准不是“opset17”，而是这个：
sess = ort.InferenceSession(model.onnx, providers=["MACAExecutionProvider"])
必须能成功，而且不能报：
unsupported op: GridSample
目前这个镜像肯定不行。
尊敬的开发者您好，感谢您的反馈。onnxruntime支持opset17的版本正在适配中，若需申请POC镜像，请通过商务渠道获取。