Paddle Inference 3.3.1 OneDNN 执行时 ConvertPirAttribute2RuntimeAttribute 不支持 ArrayAttribute<DoubleAttribute> #79749

Activity

1. jsmike9527-coder

   added

   status/new-issue新建新建

   type/bug-report报bug报bug

   on Sep 8, 2026
2. davidhaas6 commented on Sep 28, 2026

   

   davidhaas6

   on Sep 28, 2026

   More actions

   I can reproduce this on Windows build 26200 with the Python packages paddleocr==3.7.0, paddlex==3.7.2, and paddlepaddle==3.3.1 (Python 3.10.11, AMD64 Family 25 Model 33 CPU). Using TextDetection(model_name="PP-OCRv6_small_det", device="cpu", enable_mkldnn=True), inference fails on the first frame with NotImplementedError: ConvertPirAttribute2RuntimeAttribute not support [pir::ArrayAttribute<pir::DoubleAttribute>]. The same detector runs with enable_mkldnn=False. In the same application, TextRecognition(model_name="PP-OCRv6_small_rec", device="cpu", enable_mkldnn=True) completes inference. This seems to affect the small detection model specifically, in addition to the medium models and C++ path reported above.
3. Anthony-Maxwell-Franklin commented on Oct 22, 2026

   

   Anthony-Maxwell-Franklin

   on Oct 22, 2026

   More actions

   补充一条Linux独立复现，用的Python（不含 paddleocr，纯 paddlex），触发模型是和PP-OCRv6完全不同的一族 —— 说明这不是某个模型的问题。以下内容包含AI生成内容，但是经过生产环境检验

   Environment

   paddlepaddle	3.3.1, CPU build, Linux x86_64
   paddlex	3.7.2
   paddleocr	3.7.0
   Python	3.11.11
   CUDA	none (CPU only)

   Minimal reproduction (no paddleocr needed)

   from paddlex import create_model
   model = create_model(model_name="PP-DocLayout_plus-L")
   list(model.predict("any_screenshot.png"))

   create_model() succeeds; the abort is on the first predict():

   NotImplementedError: (Unimplemented) ConvertPirAttribute2RuntimeAttribute
   not support [pir::ArrayAttribute<pir::DoubleAttribute>]
     (at /paddle/paddle/fluid/framework/new_executor/instruction/onednn/onednn_instruction.cc:116)
   

   (Line 116 in the Linux py3.11 build vs 118 in the Windows build reported above — same function.)

   Affected models — the whole layout-detection family, all of them abort on the same input:
   PP-DocLayout-L, PP-DocLayout-M, PP-DocLayout-S, PP-DocLayout_plus-L.

   Related but separate: PaddleX already blocklists the sibling DETR models
   (Mask-RT-DETR-M/S, Co-Deformable-DETR-Swin-T, Co-DINO-Swin-L, and even
   PP-DocLayoutV2/V3), so it silently routes these four to plain CPU — see the PaddleX issue I filed
   for the blocklist gap. This comment is about the underlying converter behaviour: PADDLE_THROW
   for an attribute type that appears in more than one shipped model family means every new
   model with a float64 attribute array hits this wall until Paddle learns to convert it.
   Also note these do not help as workarounds (they are Paddle flags; PaddleX decides the
   backend itself): FLAGS_use_mkldnn=0, FLAGS_enable_pir_api=0. What does work:
   PADDLE_PDX_ENABLE_MKLDNN_BYDEFAULT=False, or explicit run_mode="paddle".
4. alfonsodg commented on Nov 3, 2026

   

   alfonsodg

   on Nov 3, 2026

   More actions

   Confirming this on PaddlePaddle 3.5.0-dev (Linux, CPU-only build) and adding a root-cause analysis + patch.

   Repro

   PPStructureV3 (PaddleX layout detection → table/formula) aborts before producing output. On 3.0.0b1 it does not even reach the exception: raw SIGSEGV in AnalysisPredictor::PrepareProgram.

   NotImplementedError: (Unimplemented) ConvertPirAttribute2RuntimeAttribute not support
   [pir::ArrayAttribute<pir::DoubleAttribute>]
     (at paddle/fluid/framework/new_executor/instruction/onednn/onednn_instruction.cc:116)
   

   Root cause

   paddle/fluid/framework/new_executor/instruction/onednn/onednn_instruction.cc,
   ConvertPirAttribute2RuntimeAttribute. The double-array branch validates only the
   first element and only accepts literal pir::DoubleAttribute, then throws:

   } else if (attr_type_name == "pir::ArrayAttribute<pir::DoubleAttribute>") {
     auto array_list = attr.dyn_cast<pir::ArrayAttribute>().AsVector();
     std::vector<double> vec_res;
     if (array_list.size() > 0) {
       if (array_list[0].isa<pir::DoubleAttribute>()) {      // too strict
         ...
       } else {
         PADDLE_THROW(... "not support [%s]");
       }
     }

   The schema declares a double array, but the concrete elements may be
   FloatAttribute / Int32Attribute / Int64Attribute. They are losslessly
   promotable, so the throw is a false negative.

   Proposed patch

   } else if (attr_type_name == "pir::ArrayAttribute<pir::DoubleAttribute>") {
     auto array_list = attr.dyn_cast<pir::ArrayAttribute>().AsVector();
     std::vector<double> vec_res;
     for (size_t i = 0; i < array_list.size(); ++i) {
       if (array_list[i].isa<pir::DoubleAttribute>()) {
         vec_res.push_back(array_list[i].dyn_cast<pir::DoubleAttribute>().data());
       } else if (array_list[i].isa<pir::FloatAttribute>()) {
         vec_res.push_back(static_cast<double>(
             array_list[i].dyn_cast<pir::FloatAttribute>().data()));
       } else if (array_list[i].isa<pir::Int32Attribute>()) {
         vec_res.push_back(static_cast<double>(
             array_list[i].dyn_cast<pir::Int32Attribute>().data()));
       } else if (array_list[i].isa<pir::Int64Attribute>()) {
         vec_res.push_back(static_cast<double>(
             array_list[i].dyn_cast<pir::Int64Attribute>().data()));
       } else {
         PADDLE_THROW(common::errors::Unimplemented(
             "ConvertPirAttribute2RuntimeAttribute not support [%s] ", attr_type_name));
       }
     }
     return vec_res;

   Verification with the patched build

   • PPStructureV3 constructs and runs; no exception, no segfault.
   • Table structure recovered as HTML with merged-cell colspan from a real document.
   • Formula recognition returns LaTeX (R_{n}=F_{y}A_{g}, \phi M_{n}=\phi F_{y}Z_{x}\geq M_{u}).
   • 5 unit tests for the parsing wrapper pass.

   Build note

   With -DWITH_GPU=ON and CUDA >= 11.4 / compute capability >= 80, cmake forces
   WITH_FLASHATTN=ON and the build dies compiling
   flash_attn_with_bias_and_mask.cu; a CPU-only build (-DWITH_GPU=OFF) is enough
   to validate this path.

   Tracking issue with full details in our fork: alfonsodg/RapidOCR#8
5. alfonsodg

   added a commit that references this issue on Nov 3, 2026

   fix(onednn): accept promoted scalars in double array attributes (Padd…

   ...

   69290ca