Skip to content

how to polygraphy run with trt-outputs mark some layers? #4442

@lzcchl

Description

@lzcchl

I'm new to Polygraphy, and my command looks like this:

polygraphy run
${onnx_path}
--onnxrt --trt
--providers cpu
--pool-limit workspace:24G
--atol 1e-3 --rtol 1e-3
--tf32
--verbose
--trt-min-shapes input:[1,3,1333,1333] --trt-opt-shapes input:[1,3,1333,1333] --trt-max-shapes input:[1,3,1333,1333]
--input-shapes 'input:[1,3,1333,1333]'
--onnx-outputs mark all
--trt-outputs mark all
--load-timing-cache ${timing_cache_path}
--load-inputs ${cus_inputs}
--save-engine ${trt_path}
> result-run.txt

Since my model is a ViT (Vision Transformer) and the input size [3,1333,1333] is quite large, my RTX 4090 with 24GB VRAM runs into an OOM (Out of Memory) error.

To address this, I'm considering ​​not marking all layers​​ at once but instead marking them in batches—for example:

​​5 layers at a time​​:
--trt-outputs "layer0 layer1 layer2 layer3 layer4"
or 100 layers at a time​​, but doing this manually in the command line is very inconvenient.

Question​​: Does the mark option accept a ​​file as input​​? For example:
--trt-outputs layer_names.txt

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions