Incorrect Decoder Compilation for DEIMv2 on neuronx

[cricket1]

Summary

The AWS Neuron compiler (torch-neuronx) produces incorrect outputs when compiling a transformer-based object detection decoder with discrete attention mechanism. The compiled model produces vastly different logits and bounding boxes compared to PyTorch, resulting in zero detections on real images.

Environment

# System info
OS: Ubuntu (Linux 6.8.0-1040-aws)
Instance: inf2.xlarge (tested on CPU instance for compilation)

# Python packages (confirmed versions)
Python: 3.10.12
torch: 2.8.0
torch-neuronx: 2.8.0.2.10.16998+e9bf8a50
torch-xla: 2.8.1
neuronx-cc: 2.21.33363.0+82129205
torchvision: 0.23.0

Problem Description

Expected Behavior

When compiling a transformer decoder with discrete attention to Neuron format, the compiled model should produce outputs numerically close to PyTorch (within typical BF16 precision tolerance of ~1e-2).

Actual Behavior

The Neuron-compiled decoder produces completely incorrect outputs:

• Logits max difference: 7.84 (compared to ~1e-2 expected)
• Logits mean difference: 1.25 (compared to ~1e-4 expected)
• Relative difference: 160,087% (!)
• Detection count: 0 detections (PyTorch: 60 detections)
• Top-5 confidence scores: [0.25, 0.24, 0.23, 0.21, 0.21] (PyTorch: [0.93, 0.90, 0.89, 0.84, 0.84])

Components Affected

• Backbone (HGNetv2): ✓ Works correctly (max diff ~0.7)
• Encoder (HybridEncoder): ✓ Works correctly (max diff ~0.18)
• Decoder (DEIMTransformer with discrete attention): ✗ Completely broken

Reproduction

1. Model Architecture

The decoder is a DETR-style transformer decoder with:

• Multi-scale deformable attention using discrete indexing (not F.grid_sample)
• 3 decoder layers
• 300 object queries
• 80 classes (COCO)
• 2 feature levels

Key detail: The model uses cross_attn_method: 'discrete' which implements deformable attention via discrete tensor indexing instead of F.grid_sample() (which is not supported by Neuron).

3. Full Reproduction

Complete reproduction code available at: /home/ubuntu/sh_deimv2/

Key files:

• convert_components_neuronx.py: Conversion script (View on GitHub Gist)
• verify_neuronx_on_image.py: Verification on real image (View on GitHub Gist)
• compare_all_methods.py: Comprehensive comparison (View on GitHub Gist)

To reproduce:

# Convert decoder to Neuron
python convert_components_neuronx.py \\
    --checkpoint models/deimv2_hgnetv2_n_coco.pth \\
    --config configs/deimv2/deimv2_hgnetv2_n_coco.yml \\
    --component all

# Verify (shows the bug)
python verify_neuronx_on_image.py --image example.jpg

Investigation Results

What We've Tried

1. ✗ Different compiler optimizations: Tested -O0, -O1, -O2
2. ✗ FP32 precision: --fp32-cast=all, --fp32-cast=matmult
3. ✗ Disabled auto-casting: --auto-cast=none
4. ✗ Model type hint: --model-type=transformer
5. ✗ Various combinations: See try_neuronx_compiler_settings.py

None of the above produced correct outputs.

Key Observations

1. Backbone and encoder compile correctly - The issue is isolated to the decoder
2. PyTorch discrete attention works perfectly - The implementation is correct in PyTorch
3. Error is not precision-related - Differences are 1000x larger than BF16 tolerance
4. Consistent failure - Fails on both random inputs and real images

[AbigailFernandes]

Thank you @cricket1 for raising the issue. Our team is investigating this and will follow up with additional questions or potential fixes.

We appreciate the detailed environment information, reproduction code, and the isolation work you've already done to narrow this down to the decoder component.

We'll update this thread as we make progress.

[0x6b64]

Hi @cricket1 , I used the latest release container public.ecr.aws/neuron/pytorch-inference-neuronx:2.8.0-neuronx-py311-sdk2.26.1-ubuntu22.04 with python3.11 && the same software versions mentioned above.

One change I had to make was in deimv2_hgnetv2_n_coco.yml to specify cross_attn_method: 'discrete'.

I'll attach the full compilation report shortly but here is the summary: (replaced inline output with gist)
https://gist.github.com/0x6b64/03a6424028e022a174fadfc6796db83f

This is the verification script output: (replaced inline output with gist)
https://gist.github.com/0x6b64/340afc7c5ac9f2a1356139f4620e8122

@cricket1 Do you have a hard dependency on python3.10? I'll rerun testing with 3.10, but from the scripts you provided, it seems like we are passing the verification checks (using python3.11).

[0x6b64]

Ran the scripts with python3.10 && they seem to work as well.

One change I had to make was in deimv2_hgnetv2_n_coco.yml to specify cross_attn_method: 'discrete'.

Full output for 3.10: https://gist.github.com/0x6b64/77a9f7da20211e3f3227e78d53dc717b

Can you also share your runtime version? I'm sure you would have updated the config on your side as well, so I'm wondering if there is any delta in our environments.

[cricket1]

@0x6b64 i see in your log that NeuronX detections :0 where as pytorch has 60 detections

[0x6b64]

Thanks for the update.

  ✓ NeuronX components are working correctly!
  ✓ Small differences in scores/boxes are expected due to compilation optimizations
  ✓ Visual comparison saved to verification_comparison.jpg

I mistook this as the neuron script working as expected. Checking the neuron artifacts.

[0x6b64]

Running diagnostic/checks, will update the ticket as I make progress.

Update: Need more time to dive into this. Have been bisecting the decoder graph to identify the first source of inaccuracy. (in progress)

[cricket1]

@0x6b64 could you make any progress

[parmoham-aws]

We're still debugging the issue and will provide updates as we make progress. Thanks for your patience.

[ivashkst]

Thank you for your patience. We are still looking into this issue.