ONNX models with import or runtime issues

Tracking: Models That Fail to Import or Run in Burn

This issue tracks ONNX models that cannot currently be onnx converted, rs built, have runtime problems or output accuracies using burn-import. If you encounter a model that fails to import or execute, please comment below or submit a new issue and reference this tracker.

Checklist of Models with Known Issues (working if checked)

Natural Language Processing (NLP)

Multimodal (Vision-Language)

CLIP ViT-B-32
clip-ViT-B-32 Vision

Object Detection

Depth Estimation

Depth-Anything-v2
dpt-dinov2-small-kitti - converts and compiles on main from the onnx-community export as published (#2592 fixed); needs a run/accuracy check
Apple Depth Pro

Generative Models

stable-diffusion-xl-base-1.0

Audio/Speech

piper-tts / VITS - converts on main, generated code does not compile (#534, #533)
NVIDIA FastConformer (streaming Conformer-Hybrid Large en-US) - awaiting a re-test on main
Silero VAD
kokoro onnx model

Computer Vision - Classification

Computer Vision - Face Recognition

ArcFace (LResNet100E-IR)

Segmentation

Segment Anything 2 (SAM-2) hiera_tiny encoder - converts and compiles on main; needs a run/accuracy check

Computer Vision - Other

waifu2x cunet - converts and compiles on main; needs a run/accuracy check
XFeat - converts and compiles on main (#3110 fixed, ScatterND and TopK both supported now); needs a run/accuracy check
Piiranha - converts and compiles on main (#2968 fixed); needs a run/accuracy check
DINOv3 (#469 for the onnx-community export; the -MHA export needs com.microsoft::MultiHeadAttention; the quantized variant also needs #160)
PP-OCRv6 - det and rec both convert and compile on main (#468 fixed); needs a run/accuracy check

Optical Flow & Pose Estimation

GMFlow models - last known blocker (#20) was fixed by #53; not re-tested since, model is Google Drive only
RTMW3D-x (#2930, models#74)

How to Use

Add a comment if you find a new failing model, or if your model is fixed by a PR.
Check off models as they become supported or fixed.
Reference this issue when creating new ONNX import failure reports.

For operator-level or feature gaps, please also check:

SUPPORTED-ONNX-OPS. md for coverage
Help Wanted: Implementing ONNX Ops for operator implementation tracking
[burn-import issues](https://github.com/tracel-ai/burn/issues? q=is%3Aissue%20state%3Aopen%20label%3Aonnx)

Quick update I am able to convert yolo11x_opset16.onnx using this WIP PR: tracel-ai/burn#3381

CLIP ViT-B-32 is buildable with tracel-ai/burn#3560 (still under review)

I'll be working on a test harness to test various large models quickly.

RTMW3D-x is buildable with tracel-ai/burn#3564 (still under review)

i am able to import https://huggingface.co/Xenova/albert-large-v2/resolve/main/onnx/model.onnx but rust code has type errors

in rust,shou error:
=== Tensor Operation Error ===
Operation: 'Reshape'
Reason:
1. The given shape doesn't have the same number of elements as the current tensor. Current shape: [1035], target shape: [1, 1034].

program in python,
def init(self, input_dim=1035):
super(Net, self).init()
self.attention = ChannelAttention(input_dim)
self.fc1 = nn.Linear(input_dim, 512)
self.fc2 = nn.Linear(512, 256)
self.fc3 = nn.Linear(256, 128)
self.fc4 = nn.Linear(128, 2)
self.relu = nn.ReLU()

torch.onnx.export(model, 
                dummy_input, 
                'd:/archicad/model_data/f_best_2.onnx',
                export_params=True, 
                opset_version=12, 
                do_constant_folding=True,
                input_names=['input'], 
                output_names=['output'],
                dynamic_axes={})

have use uv run --script https://raw.githubusercontent.com/tracel-ai/burn/refs/heads/main/crates/burn-import/onnx_opset_upgrade.py
change to opset version 16

@laggui

i am able to import https://huggingface.co/Xenova/albert-large-v2/resolve/main/onnx/model.onnx but rust code has type errors

I have 3 outstanding ONNX related PRs (#3563, tracel-ai/burn#3564, tracel-ai/burn#3550) with fixes. Most likely it's caused by this: tracel-ai/burn#3564.

Hopefully @laggui will have some time to review ;-)

Resnet is buildable:

can we add kokoro to this. as its a very popular tts model. currently has an issue with expand1 - rank which is an already known issue.
it also uses albert internally so will probably require albert model to be working for this to work
https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files-v1.0/kokoro-v1.0.onnx
there is also the bin file for voices
https://github.com/thewh1teagle/kokoro-onnx/releases/download/model-files-v1.0/voices-v1.0.bin

Even though Yolo11X ONNX file can be converted into rust code, currently the generated rust code can't be built to due to not handing broadcasting (see why automatic full broadcasting is lacking in Burn tracel-ai/burn#1499). I have fixed this broadcasting issue here: tracel-ai/burn#3589

Afterwards, there is another runtime issue with slice that I need to investigate.

Is failing to import due to a scalar input in ConstantOfShape
https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/vae_decoder/model.onnx

  ERROR burn_import::logger: PANIC => panicked at C:\Users\Danilo\.cargo\registry\src\index.crates.io-1949cf8c6b5b557f\onnx-ir-0.18.0\src\node\constant_of_shape.rs:34:18:
  ConstantOfShape node must have a Tensor with a non-empty static shape value

attribute in netron

tensor: float32[1]
[
    0
]

I have added a harness to test models: https://github.com/tracel-ai/burn/tree/main/crates/burn-import/model-checks

YOLOx11 is passing for tch and ndarray backends but currently it's failing on metal due to: #51 bug in metal.

I am working on clip-vit-b-32-text next. Locally it's passing (with 5-6 fixes in burn-import related to broadcasting and other issues). I have identified one bug in ndarray backend related to int lower open. I will submit a PR shortly and a bug report for ndarray (if I can't fix it).

Recently had issues with Apple Depth Pro , could it be added? It has a script for weights.

Another one of interest to me is LVSM: A Large View Synthesis Model with Minimal 3D Inductive Bias also with a list of checkpoints, also with issues importing in Burn (on a recent main).

Yolo26x worked a few months ago (before the burn-onnx crate was extracted)
Currently it fails in runtime for all backends with

index.crates.io-1949cf8c6b5b557f\ndarray-0.17.2\src\zip\mod.rs:252:5:
Zip: Producer dimension mismatch, expected: IxDynImpl(Inline(2, [1, 300, 0, 0])), got: IxDynImpl(Inline(2, [1, 1, 0, 0]))
stack backtrace:

---
let unsqueeze2_out1: Tensor<B, 3> = topk2_out1.unsqueeze_dims::<3>(&[2]);
let mod1_out1 = topk2_out2.remainder((constant354_out1).unsqueeze_dims(&[0isize])); <-- here
let unsqueeze3_out1: Tensor<B, 3, Int> = mod1_out1.unsqueeze_dims::<3>(&[2]);

@ZOXEXIVO

@ZOXEXIVO are you using batch size other than 1?

I export it as i usually do

yolo export model=last.pt format=onnx
Ultralytics 8.4.33  Python-3.13.9 torch-2.9.0+cu130 CPU (AMD ...)
YOLO26x summary (fused): 190 layers, 55,637,013 parameters, 0 gradients, 193.4 GFLOPs

PyTorch: starting from 'last.pt' with input shape (1, 3, 1280, 1280) BCHW and output shape(s) (1, 300, 6) (225.4 MB)

ONNX: slimming with onnxslim 0.1.72...
ONNX: export success  2.8s, saved as 'last.onnx' (213.1 MB)

but at runtime, does your input tensor include batch size more than 1?

Yes, code not changed

let tensor_data = TensorData::new(input_data, [1, CHANNEL_COUNT, IMG_SIZE, IMG_SIZE]);

let input_tensor: Tensor<B, 4> = Tensor::from_data(
     tensor_data.convert::<B::FloatElem>(),
     device
);

let output = model.forward(input_tensor);

@ZOXEXIVO

@ZOXEXIVO filed an issue: #281

@ZOXEXIVO

@ZOXEXIVO

Can you please test if this PR fixes your issue: #285 ?

it's merged on main

@antimora

@ZOXEXIVO

Can you please test if this PR fixes your issue: #285 ?

it's merged on main

I tested it — the issue is fixed.
Thanks, @antimora !

Added 11 more model-checks, including ArcFace

kokoro model works with #372 fix

however, there is accuracy 1% discrepancy with ORT see #371

DINOv3 cannot be imported.

The model downloaded from hugging face onnx-community failed to be built.

quantized build log

...
  DEBUG onnx_ir::pipeline:  PHASE 3: Type Inference
  ERROR burn_onnx::logger: PANIC => panicked at /home/louis/.cargo/registry/src/index.crates.io-1949cf8c6b5b557f/burn-onnx-0.21.0/src/model_gen.rs:349:33:
  Failed to parse ONNX file 'models/upgraded_model_quantized.onnx': Type inference failed: Unsupported ONNX operation(s): DynamicQuantizeLinear (node 'dynamr18'), SplitToSequence (node 'splittosequence9'), SplitToSequence (node 'splittosequence10'), SequenceAt (node 'sequenceat17'), SequenceAt (node 'sequenceat18'), SequenceAt (node 'sequenceat19'), SequenceAt (node 'sequenceat20'), DynamicQuantizeLinear (node 'dynamicquantizelinear19'), DynamicQuantizeLinear (node 'dynamicquantizelinear20'), DynamicQuantizeLinear (node 'dynamicquantizelinear21'), DynamicQuantizeLinear (node 'dynamicquantizelinear22'), SplitToSequence (node 'splittosequence11'), SplitToSequence (node 'splittosequence12'), SequenceAt (node 'sequenceat21'), SequenceAt (node 'sequenceat22'), SequenceAt (node 'sequenceat23'), SequenceAt (node 'sequenceat24'), DynamicQuantizeLinear (node 'dynamicquantizelinear23'), DynamicQuantizeLinear (node 'dynamicquantizelinear24'), DynamicQuantizeLinear (node 'dynamicquantizelinear25'), DynamicQuantizeLinear (node 'dynamicquantizelinear26'), SplitToSequence (node 'splittosequence13'), SplitToSequence (node 'splittosequence14'), SequenceAt (node 'sequenceat25'), SequenceAt (node 'sequenceat26'), SequenceAt (node 'sequenceat27'), SequenceAt (node 'sequenceat28'), DynamicQuantizeLinear (node 'dynamicquantizelinear27'), DynamicQuantizeLinear (node 'dynamicquantizelinear28'), DynamicQuantizeLinear (node 'dynamicquantizelinear29'), DynamicQuantizeLinear (node 'dynamicquantizelinear30'), SplitToSequence (node 'splittosequence15'), SplitToSequence (node 'splittosequence16'), SequenceAt (node 'sequenceat29'), SequenceAt (node 'sequenceat30'), SequenceAt (node 'sequenceat31'), SequenceAt (node 'sequenceat32'), DynamicQuantizeLinear (node 'dynamicquantizelinear31'), DynamicQuantizeLinear (node 'dynamicquantizelinear32'), DynamicQuantizeLinear (node 'dynamicquantizelinear33'), DynamicQuantizeLinear (node 'dynamicquantizelinear34'), SplitToSequence (node 'splittosequence17'), SplitToSequence (node 'splittosequence18'), SequenceAt (node 'sequenceat33'), SequenceAt (node 'sequenceat34'), SequenceAt (node 'sequenceat35'), SequenceAt (node 'sequenceat36'), DynamicQuantizeLinear (node 'dynamicquantizelinear35'), DynamicQuantizeLinear (node 'dynamicquantizelinear36'), DynamicQuantizeLinear (node 'dynamicquantizelinear37'), DynamicQuantizeLinear (node 'dynamicquantizelinear38'), SplitToSequence (node 'splittosequence19'), SplitToSequence (node 'splittosequence20'), SequenceAt (node 'sequenceat37'), SequenceAt (node 'sequenceat38'), SequenceAt (node 'sequenceat39'), SequenceAt (node 'sequenceat40'), DynamicQuantizeLinear (node 'dynamicquantizelinear39'), DynamicQuantizeLinear (node 'dynamicquantizelinear40'), DynamicQuantizeLinear (node 'dynamicquantizelinear41'), DynamicQuantizeLinear (node 'dynamicquantizelinear42'), SplitToSequence (node 'splittosequence21'), SplitToSequence (node 'splittosequence22'), SequenceAt (node 'sequenceat41'), SequenceAt (node 'sequenceat42'), SequenceAt (node 'sequenceat43'), SequenceAt (node 'sequenceat44'), DynamicQuantizeLinear (node 'dynamicquantizelinear43'), DynamicQuantizeLinear (node 'dynamicquantizelinear44'), DynamicQuantizeLinear (node 'dynamicquantizelinear45'), DynamicQuantizeLinear (node 'dynamicquantizelinear46'), SplitToSequence (node 'splittosequence23'), SplitToSequence (node 'splittosequence24'), SequenceAt (node 'sequenceat45'), SequenceAt (node 'sequenceat46'), SequenceAt (node 'sequenceat47'), SequenceAt (node 'sequenceat48'), DynamicQuantizeLinear (node 'dynamicquantizelinear47'), DynamicQuantizeLinear (node 'dynamicquantizelinear48'), DynamicQuantizeLinear (node 'dynamicquantizelinear49')

not quantized build log

...
DEBUG onnx_ir::pipeline:  PHASE 3: Type Inference
  ERROR burn_onnx::logger: PANIC => panicked at /home/louis/.cargo/registry/src/index.crates.io-1949cf8c6b5b557f/burn-onnx-0.21.0/src/model_gen.rs:349:33:
  Failed to parse ONNX file 'models/model.onnx': Type inference failed: Unsupported ONNX operation(s): SplitToSequence (node 'splittosequence1'), SplitToSequence (node 'splittosequence2'), SequenceAt (node 'sequenceat1'), SequenceAt (node 'sequenceat2'), SequenceAt (node 'sequenceat3'), SequenceAt (node 'sequenceat4'), SplitToSequence (node 'splittosequence3'), SplitToSequence (node 'splittosequence4'), SequenceAt (node 'sequenceat5'), SequenceAt (node 'sequenceat6'), SequenceAt (node 'sequenceat7'), SequenceAt (node 'sequenceat8'), SplitToSequence (node 'splittosequence5'), SplitToSequence (node 'splittosequence6'), SequenceAt (node 'sequenceat9'), SequenceAt (node 'sequenceat10'), SequenceAt (node 'sequenceat11'), SequenceAt (node 'sequenceat12'), SplitToSequence (node 'splittosequence7'), SplitToSequence (node 'splittosequence8'), SequenceAt (node 'sequenceat13'), SequenceAt (node 'sequenceat14'), SequenceAt (node 'sequenceat15'), SequenceAt (node 'sequenceat16'), SplitToSequence (node 'splittosequence9'), SplitToSequence (node 'splittosequence10'), SequenceAt (node 'sequenceat17'), SequenceAt (node 'sequenceat18'), SequenceAt (node 'sequenceat19'), SequenceAt (node 'sequenceat20'), SplitToSequence (node 'splittosequence11'), SplitToSequence (node 'splittosequence12'), SequenceAt (node 'sequenceat21'), SequenceAt (node 'sequenceat22'), SequenceAt (node 'sequenceat23'), SequenceAt (node 'sequenceat24'), SplitToSequence (node 'splittosequence13'), SplitToSequence (node 'splittosequence14'), SequenceAt (node 'sequenceat25'), SequenceAt (node 'sequenceat26'), SequenceAt (node 'sequenceat27'), SequenceAt (node 'sequenceat28'), SplitToSequence (node 'splittosequence15'), SplitToSequence (node 'splittosequence16'), SequenceAt (node 'sequenceat29'), SequenceAt (node 'sequenceat30'), SequenceAt (node 'sequenceat31'), SequenceAt (node 'sequenceat32'), SplitToSequence (node 'splittosequence17'), SplitToSequence (node 'splittosequence18'), SequenceAt (node 'sequenceat33'), SequenceAt (node 'sequenceat34'), SequenceAt (node 'sequenceat35'), SequenceAt (node 'sequenceat36'), SplitToSequence (node 'splittosequence19'), SplitToSequence (node 'splittosequence20'), SequenceAt (node 'sequenceat37'), SequenceAt (node 'sequenceat38'), SequenceAt (node 'sequenceat39'), SequenceAt (node 'sequenceat40'), SplitToSequence (node 'splittosequence21'), SplitToSequence (node 'splittosequence22'), SequenceAt (node 'sequenceat41'), SequenceAt (node 'sequenceat42'), SequenceAt (node 'sequenceat43'), SequenceAt (node 'sequenceat44'), SplitToSequence (node 'splittosequence23'), SplitToSequence (node 'splittosequence24'), SequenceAt (node 'sequenceat45'), SequenceAt (node 'sequenceat46'), SequenceAt (node 'sequenceat47'), SequenceAt (node 'sequenceat48')

It seems that lack of some operators...

I also tried to download .pth from facebook DINOv3 repo, and manually convert them into .onnx with this python script:

python script to convert pth to onnx

from functools import lru_cache
from pathlib import Path
import torch
import os

MODELS_DIR = "../models"


@lru_cache(maxsize=1)
def models() -> dict[str, Path]:
    return {
        key: Path(f"{root}/{file}")
        for root, _, files in os.walk(MODELS_DIR)
        for file in files
        if file.endswith(".pth") and (key := "_".join(file.split("_")[:2]))
    }


def main():
    for model, path in models().items():
        convert(model, path)


def convert(model: str, path: Path):
    torch_model = torch.hub.load(
        repo_or_dir="facebookresearch/dinov3",
        model=model,
        weights=str(path),
    )
    torch_model.eval()
    example_inputs = (
        torch.rand(1, 3, 256, 256, dtype=torch.float32, requires_grad=True),
    )
    torch_model(example_inputs[0])
    torch.onnx.export(
        torch_model,
        example_inputs,
        path.with_suffix(".onnx"),
        input_names=["input"],
        output_names=["output"],
        opset_version=20,
        dynamo=False, # dynamo is buggy when converting DINOv3, so disable it
        dynamic_axes={
            "input": {0: "batch_size"},
            "output": {0: "batch_size"},
        },
    )


if __name__ == "__main__":
    main()

pyproject.toml

[project]
name = "pth2onnx"
version = "0.1.0"
description = "Add your description here"
readme = "README.md"
requires-python = ">=3.14"
dependencies = [
    "torch",
    "torchvision",
    "torchmetrics",
    "termcolor",
    "onnx",
    "onnxscript"
]

[tool.uv.sources]
torch = [
  { index = "pytorch-cpu" },
]
torchvision = [
  { index = "pytorch-cpu" },
]
torchmetrics = [
  { index = "pytorch-cpu" },
]

[[tool.uv.index]]
name = "pytorch-cpu"
url = "https://download.pytorch.org/whl/cpu"
explicit = true

And loading in build.rs got:

manually converted build log

...
ERROR burn_onnx::logger: PANIC => panicked at /home/louis/.cargo/registry/src/index.crates.io-1949cf8c6b5b557f/burn-onnx-0.21.0/src/burn/node/expand.rs:37:26:
  Invalid shape source ScalarNative(I64)

PP-OCRv6 does not build.
Model downloaded from here doesn't work.

// build.rs
use burn_onnx::ModelGen;

const OUTPUT_DIR: &str = "models/";

// the text detection model
const DET_MODEL_PATH_ONNX: &str = "models/PP-OCRv6_medium_det_onnx/inference.onnx";

// the text recognition model
const REC_MODEL_PATH_ONNX: &str = "models/PP-OCRv6_medium_rec_onnx/inference.onnx";

fn main() {
    for onnx_file in [DET_MODEL_PATH_ONNX, REC_MODEL_PATH_ONNX] {
        ModelGen::new()
            .input(onnx_file)
            .out_dir(OUTPUT_DIR)
            .load_strategy(burn_onnx::LoadStrategy::Embedded)
            .run_from_script();
    }
}

Error Message:

  ERROR burn_onnx::logger: PANIC => panicked at .cargo/registry/src/index.crates.io-1949cf8c6b5b557f/burn-onnx-0.21.0/src/burn/codegen.rs:237:18:
  auto_pad SAME_UPPER/SAME_LOWER requires static input shape, but input has dynamic dimensions. Use explicit pads instead

Tried this with waifu2x onnx models which can be downloaded from here:
https://github.com/nagadomi/nunif/releases/tag/0.0.0

And ran into

  ERROR burn_onnx::logger: PANIC => panicked at /home/cherry/.cargo/registry/src/index.crates.io-1949cf8c6b5b557f/burn-onnx-0.21.0/src/burn/node/pad.rs:22:17:
  Runtime pads are not supported in burn-onnx

  --- stderr

  thread 'main' (226314) panicked at /home/cherry/.cargo/registry/src/index.crates.io-1949cf8c6b5b557f/burn-onnx-0.21.0/src/burn/node/pad.rs:22:17:
  Runtime pads are not supported in burn-onnx
  note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

Used an initial test of this to do it

use burn_onnx::ModelGen;

fn main() {
    ModelGen::new()
        .input("models/cunet/art/noise0.onnx")
        .out_dir("model/")
        .run_from_script();
}

@kingwingfly

Triaged the reports in this thread against main (438d2cd). Summary of where each one stands.

Filed as new issues

Model	Issue	Error
DINOv3	#467	`Invalid shape source ScalarNative(I64)` in `Expand` codegen
DINOv3	#469	`SplitToSequence` / `SequenceAt` unsupported
PP-OCRv6	#468	`auto_pad SAME_UPPER/SAME_LOWER requires static input shape`

Both #467 and #468 reproduce on main with a single-node graph, and each issue carries a short
uv script that builds the model, so neither needs the original download to work on.

@kingwingfly the quantized DINOv3 export additionally needs DynamicQuantizeLinear, tracked in #160,
so the non-quantized export is the one to watch once #467 and #469 land.

Already fixed on `main`, please re-test

In each case the specific error reported no longer reproduces. I checked the failing path rather
than the full model, so a re-test on main would be useful to confirm nothing else is behind it.

@MolotovCherry (waifu2x): Runtime pads are not supported in burn-onnx was fixed by Runtime pads are not supported in burn-onnx #405. Runtime
pads now generate a forward-time read of the pads tensor.
@notdanilo (SDXL vae_encoder): Slice: steps other than 1 are not supported is gone. A negative
step now emits x.slice(s![.., .., .., -1..i64::MIN; -1]).
@LexiconCode (NVIDIA FastConformer): Cannot determine output rank for Expand node was addressed
by the rank fallback added in Handle Shape type in Expand node #277.
@nishantb06 (SAM-2 encoder): Pad: padding will only be applied to the last two dimensions was
fixed when n-dimensional padding landed in Add MediaPipe Face Detector model check #153.
@madushan1000 (piper-tts/VITS): Importing piper-tts/VITS TTS model fails with type inference error #266 was fixed by Handle Shape type in Expand node #277, and the runtime padding you mentioned as
the follow-up blocker is supported now too.

All of these are on main and not in the released 0.21.0, which is what the reports were run against.

Needs more detail

@AdrianEddy you mentioned Yolo11x exported with dynamic=True fails to import, without the error
text. If it is a auto_pad SAME_UPPER/SAME_LOWER requires static input shape panic then it is #468
and will be covered by that fix. If it is something else it is worth its own issue. Could you post
the panic message?

Also updated the checklist above: Retinaface and Resnet were both confirmed working in this thread
but never checked off, and DINOv3 and PP-OCRv6 are added with links to the issues above.

@dericktseng

Retriaged against main (a3b8dc0, burn rev faec3983). Last pass was a month ago against 438d2cd.

Everything below was re-run, not read off issue state. Two levels of evidence: converts means
onnx2burn produced Rust, compiles means that Rust was then cargo checked against
burn 0.22.0-pre.3. Neither means the output matches onnxruntime; that still needs a model-check.

Unblocked since the last pass

Model	Blocker it was filed under	Now
PP-OCRv6 `det` + `rec`	#468 `auto_pad SAME_*` on dynamic H/W	converts + compiles
XFeat	burn#3110 TopK scalar input, then ScatterND	converts + compiles
Piiranha	burn#2968 Squeeze	converts + compiles
dpt-dinov2-small-kitti	burn#2592 Expand w/ dynamic shape	converts + compiles
waifu2x `cunet/art/noise0`	runtime `pads`	converts + compiles
SDXL `vae_encoder`	`Slice: steps other than 1`	converts + compiles
SAM-2 `hiera_tiny` encoder	n-d Pad	converts + compiles

@dericktseng PP-OCRv6 was the concrete win: #478 landed the PaddingConfig2d::Same lowering, and both
PP-OCRv6_medium_det_onnx and PP-OCRv6_medium_rec_onnx now go straight through from the
HuggingFace download with no re-export.

@MolotovCherry waifu2x imports now. I used the 20230504 model zip, cunet/art/noise0.onnx.

dpt-dinov2-small-kitti no longer needs the opset-16 upgrade step that
burn#2592 ended on. The onnx-community export
imports as published, which is the opset-compliance work paying off.

None of these seven are checked off above, because "converts + compiles" is not "runs and matches".
Checking a box needs a model-check under crates/model-checks/.

Still blocked

DINOv3 is unchanged, and all three exports fail differently:

Export	Result
`onnx-community/dinov3-vits16-...-ONNX`	#469, `SplitToSequence` + `SequenceAt`, 72 nodes
`onnx-community/...-ONNX-MHA`	`com.microsoft::MultiHeadAttention` x12, a contrib op, not standard ONNX
manual `torch.onnx.export` from the `.pth`	rank-0 `Expand` shape input

@kingwingfly on the third one: #470 turned that panic into a real error message, but it is not a
burn-onnx bug. ONNX requires Expand's shape to be 1-D and onnxruntime rejects that graph too, so
the export itself is out of spec. #469 is the one to watch for the onnx-community export. The
quantized variant additionally needs #160.

piper-tts / VITS (@madushan1000) has moved but is not fixed. I asked you to re-test last month;
I did it instead, on en_US-lessac-medium. It converts now, so #266 and the runtime padding really
are behind us, but the generated Rust does not compile: 17 errors in two families. Both reduce to
small graphs, both turned out to be bigger than piper, and both are now filed.

Add/Sub/Mul/Div do not rank-lift a Shape operand, emitting a rank mismatch #534. Add/Sub/Mul/Div do not rank-lift a Shape operand. The Shape is materialized as a
rank-1 tensor and passed straight to .add(..), which in burn needs matching ranks:
```
let add108_out1 = add107_out1                                  // Tensor<4, Int>
    .add(Tensor::<1, burn::tensor::Int>::from_data(...));      // expected 4, found 1
```
leading_broadcast already exists and is applied in the tensor-vs-tensor arm of the same
match, just not in the Shape arms. 8 sites across the four ops. A three-node
Shape -> Gather -> Add graph reproduces it.
Range is incorrect for float element types: static bounds truncated to i64 and codegen does not compile #533. Range is broken for every float range, not just piper's. One node reproduces it:
Range(1.5, 5.0, 1.0) should give [1.5, 2.5, 3.5, 4.5] and we generate [1, 2, 3, 4],
because RangeInput::Static holds an i64 and is filled through scalar_i64(), whose F32
arm is a bare as i64. The same graph also fails to build, because arange().cast(dtype)
changes the dtype but not the tensor Kind, so a float range stays Tensor<1, Int>. There is no
float case in crates/onnx-tests/tests/range/ and no float snapshot in range.rs, which is how
both got this far. piper hits a third face of it: a runtime float limit against static i64
bounds gives f32 - i64.

This one is worth flagging beyond piper, since the truncation is silent and wrong rather than a
build failure.

GMFlow was never re-tested. Its BatchNorm blocker (#20) was fixed by #53 back in January. The
checklist entry still carries a link labelled #1560 that points at the closed #20, which is why it
reads as blocked. I could not re-run it because the model is only on Google Drive. @AdrianEddy if
you still have it, it is worth one try.

IBM Granite 4.0 Tiny Preview has no public ONNX export at all, only GGUF, so there is nothing to
import. models#71 is really a request for someone
to produce an export first.

NVIDIA FastConformer (@LexiconCode) still needs your re-test. I have no way to fetch that model.

Two more codegen defects found while doing this

Both are pre-existing and neither blocks a model, but they show up in every model that hits the path.

Codegen emits deprecated TensorData::into_vec #535. Generated code calls TensorData::into_vec::<i64>(), which burn deprecated in 0.22.0 in favour of
try_into_vec::<E>(). 12 emit sites across codegen.rs, pad.rs, tile.rs, squeeze.rs and
unsqueeze.rs. It is already firing in our own suite: pad_runtime_axes, pad_runtime_pads,
pad_runtime_pads_axes and unsqueeze_runtime_axes all warn on build. onnx-tests/build.rs
already uses try_into_vec for its own code, so only codegen is behind.
TopK binds both outputs even when the graph consumes one, so XFeat builds with
unused variable: topk1_out1. Same family as avg_pool1d/avg_pool2d ONNX models declare unused inputs x1, x2 #68, folded in there rather than filed separately.

Navigation Menu

Repository navigation

ONNX models with import or runtime issues #18

Tracking: Models That Fail to Import or Run in Burn

Checklist of Models with Known Issues (working if checked)

Natural Language Processing (NLP)

Multimodal (Vision-Language)

Object Detection

Depth Estimation

Generative Models

Audio/Speech

Computer Vision - Classification

Computer Vision - Face Recognition

Segmentation

Computer Vision - Other

Optical Flow & Pose Estimation

How to Use

63 remaining items

Filed as new issues

Already fixed on `main`, please re-test

Needs more detail

Unblocked since the last pass

Still blocked

Two more codegen defects found while doing this

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

ONNX models with import or runtime issues #18

Description

Tracking: Models That Fail to Import or Run in Burn

Checklist of Models with Known Issues (working if checked)

Natural Language Processing (NLP)

Multimodal (Vision-Language)

Object Detection

Depth Estimation

Generative Models

Audio/Speech

Computer Vision - Classification

Computer Vision - Face Recognition

Segmentation

Computer Vision - Other

Optical Flow & Pose Estimation

How to Use

Activity

antimora commented on Jul 24, 2025

antimora commented on Aug 16, 2025

antimora commented on Aug 16, 2025

antimora commented on Aug 17, 2025

n1ght-hunter commented on Aug 18, 2025

coolstudio1678 commented on Aug 18, 2025

antimora commented on Aug 18, 2025

antimora commented on Aug 18, 2025

n1ght-hunter commented on Aug 19, 2025

antimora commented on Aug 21, 2025

notdanilo commented on Aug 21, 2025

antimora commented on Aug 25, 2025

torsteingrindvik commented on Aug 25, 2025

63 remaining items

ZOXEXIVO commented on Mar 31, 2026

AdrianEddy commented on Mar 31, 2026

ZOXEXIVO commented on Mar 31, 2026

AdrianEddy commented on Mar 31, 2026

ZOXEXIVO commented on Mar 31, 2026

antimora commented on Mar 31, 2026

antimora commented on Apr 1, 2026

ZOXEXIVO commented on Apr 1, 2026

antimora commented on Apr 7, 2026

antimora commented on Apr 25, 2026

kingwingfly commented on May 14, 2026

dericktseng commented on Jun 20, 2026

MolotovCherry commented on Jul 31, 2026

antimora commented on Aug 22, 2026

Filed as new issues

Already fixed on main, please re-test

Needs more detail

antimora commented on Sep 22, 2026

Unblocked since the last pass

Still blocked

Two more codegen defects found while doing this

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

Already fixed on `main`, please re-test