pytorch / pytorch/executorch

Why force_fp32 is set to true for conv weight and biases on XNNPack backend

Open
#15,817 3 comments 1 reaction 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
5k
Forks
1.2k
Avg merge
2d 10h
Merged PRs (30d)
581

Description

Is there a specific reason force_fp32 set to true for convolution weight and biases for the XNNPack backend? I would like to use fp16 operations but default does not allow me.

    self.define_tensor(
        kernel_node,
        xnn_graph,
        vals_to_ids,
        ...,
        force_fp32=True,
    )
    ...
    self.define_tensor(
        get_input_node(node, 2),
        xnn_graph,
        vals_to_ids,
        ...,
        quant_params=bias_quant_params,
        force_fp32=True,
    )

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Review the XNNPack backend's define_tensor calls for convolution weights and biases, focusing on the force_fp32 and quant_params arguments shown in the issue. Trace the backend's supported weight and bias types, then check existing coverage for fp16 operations. Done means establishing whether fp16 can be enabled safely and documenting or testing the resulting behavior.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
embedded-iot, machine-learning
Issue type
Feature
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.