Issues

• Masked bf16 attention rounds its scores to bf16 on Jetson Thor (sm_110), but not on A100

  Module:AccuracyModule:Embedded
  NVIDIA/TensorRT#4938 · Filter by author shoumikhin opened on Nov 13, 2026

• Builder picks a slower bf16 GEMM kernel on Orin (sm_87): the 160x128x32 tiles run about 24% slower than the builder measured

  Module:Performance
  NVIDIA/TensorRT#4937 · Filter by author shoumikhin opened on Nov 13, 2026

• float32 attention on inputs cast from bf16 crashes the builder or fails to build (casts folded into IAttention)

  Module:Engine Build
  NVIDIA/TensorRT#4936 · Filter by author shoumikhin opened on Nov 13, 2026

• IAttention gives wrong fp16 and bf16 results when the key is a constant

  Module:Accuracy
  NVIDIA/TensorRT#4935 · Filter by author shoumikhin opened on Nov 13, 2026

• Builder keeps one host copy of a weight for every layer that reads it transposed

  Module:Engine Build
  NVIDIA/TensorRT#4934 · Filter by author shoumikhin opened on Nov 13, 2026

• Engine build holds a host copy and a GPU copy of all weights at the same time, in every build mode

  Module:Engine Build
  NVIDIA/TensorRT#4933 · Filter by author shoumikhin opened on Nov 13, 2026

• Fused attention never returns when the bool mask is a run-time tensor, at 2176 to 2304 query rows

  Module:Runtime
  NVIDIA/TensorRT#4932 · Filter by author shoumikhin opened on Nov 13, 2026

• [ONNX GraphSurgeon] fold_constants disconnects branches with multiple Cast consumers

  Module:GraphSurgeon
  NVIDIA/TensorRT#4931 · Filter by author AhmetMert01 opened on Nov 13, 2026

• onnx-graphsurgeon: fold_constants(partitioning="basic") raises RecursionError when a partition fails in onnxruntime

  Module:GraphSurgeon
  NVIDIA/TensorRT#4928 · Filter by author sohailshk opened on Nov 13, 2026

• onnx-graphsurgeon drops the sparse_value attribute of Constant nodes, so export_onnx writes an invalid model

  Module:GraphSurgeon
  NVIDIA/TensorRT#4927 · Filter by author sohailshk opened on Nov 13, 2026

• Strongly typed engine returns NaN for every depth output on sm_120 (TRT 11.4); onnxruntime on the same ONNX matches PyTorch

  Module:ONNX
  NVIDIA/TensorRT#4926 · Filter by author kleinicke opened on Nov 13, 2026
  1

• Myelin "Inconsistent strides" (tactic.cpp:4034) building a strongly typed fp16 network on sm_120, TRT 11.4; the fp32 export of the same model builds

  Module:Engine Build
  NVIDIA/TensorRT#4925 · Filter by author kleinicke opened on Nov 13, 2026
  1

• [BUG] nvonnxparser: dynamic batch (-1) + constant-size Split causes uninitialized gather index during shape inference → buildSerializedNetwork returns nullptr

  Module:ONNX
  NVIDIA/TensorRT#4924 · Filter by author Tanxinman opened on Nov 11, 2026
  3

• TensorRT 11.4.0.106 leaves half of FP16 GroupNorm output unwritten with a dynamic height

  Module:Runtime
  NVIDIA/TensorRT#4923 · Filter by author Parukichi2930 opened on Nov 10, 2026

• FP32 ReduceSum becomes very slow when used inside an encoder (TensorRT 10.8, RTX 5060 Laptop)

  Module:Performance
  NVIDIA/TensorRT#4922 · Filter by author mhmdrz-mirdamadi opened on Nov 10, 2026

• Polygraphy NumPy Top-K sorts unsigned and minimum signed integers incorrectly

  Module:Polygraphy
  NVIDIA/TensorRT#4917 · Filter by author 0z5a opened on Nov 4, 2026

• Polygraphy INT32 comparison can overflow and falsely pass at zero tolerance

  Module:Polygraphy
  NVIDIA/TensorRT#4916 · Filter by author 0z5a opened on Nov 4, 2026

• ONNX GraphSurgeon export mutates the original initializer when a constant is renamed

  Module:GraphSurgeon
  NVIDIA/TensorRT#4913 · Filter by author 0z5a opened on Nov 4, 2026

• ONNX GraphSurgeon cannot export NumPy-backed INT4 or UINT4 constants

  Module:GraphSurgeon
  NVIDIA/TensorRT#4912 · Filter by author 0z5a opened on Nov 4, 2026

• onnx-graphsurgeon: sparse initializers stored in raw_data load as zeros (0.6.1) or fail (main), and loaded sparse values are always float64

  Module:GraphSurgeon
  NVIDIA/TensorRT#4910 · Filter by author Anshtyagi1729 opened on Nov 3, 2026

• Polygraphy --load-inputs, --load-outputs and inspect data fail when a tensor is named polygraphy_class

  Module:Polygraphy
  NVIDIA/TensorRT#4908 · Filter by author Baiiiiiiiiii opened on Nov 3, 2026

• Possible incorrect INT8 Q/DQ depthwise Conv + Concat output in TensorRT 11.3 on RTX 5060

  Module:AccuracyModule:Quantization
  NVIDIA/TensorRT#4907 · Filter by author smisei-lmz opened on Nov 2, 2026

• Ultralytics YOLO INT8 build failure of TensorRT 10.16 (MyelinCheckException) on RTX PRO 6000 Blackwell, regression from 10.15

  Module:Engine Build
  NVIDIA/TensorRT#4906 · Filter by author glenn-jocher opened on Nov 1, 2026

• Ultralytics YOLO INT8 engines larger than FP32: 23.76 MB zero-filled block in TensorRT 11.4 plan on RTX PRO 6000 Blackwell

  Module:ONNX
  NVIDIA/TensorRT#4905 · Filter by author glenn-jocher opened on Nov 1, 2026

• Engine build failure of TensorRT 11.4 (Windows ARM64 wheel) when building ONNX models on GPU NVIDIA RTX Spark N1X (sm121): missing Myelin resource DLL

  Module:Engine Build
  NVIDIA/TensorRT#4904 · Filter by author gautam0901-maker opened on Oct 31, 2026