pytorch / pytorch/executorch

Vulkan index select outputs don't match eager/portable

Open
#12,222 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

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

Description

🐛 Describe the bug

When running the following repro on the Vulkan delegate, outputs differ from eager mode. ET portable matches eager.

import torch
from executorch.backends.vulkan.partitioner.vulkan_partitioner import VulkanPartitioner
from executorch.exir import to_edge_transform_and_lower, EdgeCompileConfig, to_edge
from executorch.extension.pybindings.portable_lib import _load_for_executorch_from_buffer

class IndexSelectModel(torch.nn.Module):
    def __init__(self, dim=0):
        super().__init__()
        self.dim = dim
        
    def forward(self, x, indices):
        return torch.index_select(x, self.dim, indices)
        
model = IndexSelectModel(dim=0)
indices = torch.tensor([0, 1], dtype=torch.int64)
inputs = (
    torch.randn(5, 5),
    indices,
)
eager_outputs = model(*inputs)

ep = torch.export.export(model.eval(), inputs)
print(ep)
lowered = to_edge_transform_and_lower(
    ep,
    partitioner=[VulkanPartitioner()],
    compile_config=EdgeCompileConfig(_check_ir_validity=False)
).to_executorch()
print(lowered.exported_program())

et_model = _load_for_executorch_from_buffer(lowered.buffer)
et_outputs = et_model([*inputs])[0]

print(f"Inputs: {inputs}")
print(f"Eager: {eager_outputs}")
print(f"ET:    {et_outputs}")

Outputs:

Inputs: (tensor([[-0.4567, -0.3723, -0.1551, -0.7132, -0.5437],
        [ 1.5363,  0.7774,  1.8513,  1.4016,  0.6614],
        [ 0.2748, -0.6432, -0.8875, -1.6339, -0.4679],
        [-1.1551,  0.0595,  0.4082, -1.3087, -0.7413],
        [-0.1365,  0.1206,  0.6683,  1.3647, -0.3534]]), tensor([0, 1]))
Eager: tensor([[-0.4567, -0.3723, -0.1551, -0.7132, -0.5437],
        [ 1.5363,  0.7774,  1.8513,  1.4016,  0.6614]])
ET:    tensor([[-0.4567, -0.3723, -0.1551, -0.7132, -0.5437],
        [-0.4567, -0.3723, -0.1551, -0.7132, -0.5437]])
Versions

Run on Meta internal master, Jul 3, fbcode/SwiftShader

cc @SS-JIA @manuelcandales @cbilgin

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

Start by running the provided IndexSelectModel repro with VulkanPartitioner and compare its output with eager mode and ET portable. Read executorch.backends.vulkan.partitioner.vulkan_partitioner and trace the Vulkan index-select path; done means Vulkan produces the same selected rows as eager and portable execution.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
backend
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Mostly clear
Newbie friendliness
45/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.