Hailo10H Data Flow Compiler Error: Spatial Reshape Infeasible

General Discussion
1

Hi,

when trying to compile EffcientLoFTR using the dataflow compiler from Hailo AI Suite 2026-01, I have encountered the following error when greyscale input images have size 832x832.

[error] eloftr Multi Context Partition Failed (duration: 0s)
[error] eloftr Partition and Allocation Failed (duration: 37s)
[error] eloftr Compilation Failed (duration: 37s)
No successful assignments: auto_spatial_reshape_from_matmul33_to_output_layer1 errors:
Agent infeasible
auto_spatial_reshape_from_matmul33_to_output_layer1 errors:
Agent infeasible
[error] Failed to produce compiled graph
[error] BackendAllocatorException: Compilation failed: No successful assignments:
auto_spatial_reshape_from_matmul33_to_output_layer1 errors: Agent infeasible

The error doesn’t show for smaller sized images (e.g. 96x96). In this case, the cause was a MatMul layer: MatMul op [1, 1, 256, 4800] * [1, 1, 256, 4800] → [1, 4800, 4800] producing a
similarity matrix.

The error doesn’t happen when using lower resolution imgs, e.g. 96x96, hence I would like to resolve the following doubts:
• Is this error caused by LCU limits or by memory limits?
• What is the cleanest way to resolve it?
• Would any of the Defuse operations help here?

Thanks,

Best

Hailo10H Compiler: Layer Support for Input Tensors with Large Dimension Sizes
2

As a follow up , I have split the MatMul op into 16 chunks and concatenate the results, and encountered the following compilation error:

[error] eloftr_extend_new Multi Context Partition Failed (duration: 0s)

[error] eloftr_extend_new Partition and Allocation Failed (duration: 42s)

[error] eloftr_extend_new Compilation Failed (duration: 42s)

No successful assignments: auto_spatial_reshape_from_matmul33_to_output_layer1 errors:

Agent infeasible

auto_spatial_reshape_from_matmul33_to_output_layer1 errors:

Agent infeasible by resources sanity.: Memory units capacity exceeded

(available: 128, required: 132).

Is there a hard limit (in MB) on the size of a tensor when used as a model input?

3

Hi @user976,

Please see here: Hailo10H Compiler: Layer Support for Input Tensors with Large Dimension Sizes - #2 by Michael

Thanks,