apache
[Feature Request] Support for sparse matrix multiplication and random number generation in PyTorch frontend
Description
When converting a PyTorch model containing sparse matrix multiplication (torch.sparse.mm) and random number generation (torch.randn) operations to TVM Relax module via torch.export, an AssertionError occurs. TVM currently does not support the _sparse_mm.default and randn.default operations.
Expected behavior
The PyTorch model with sparse operations and random number generation should be successfully converted to TVM Relax module, enabling deployment of models that use sparse computations and stochastic components.
Actual behavior
An AssertionError occurs during from_exported_program conversion with the message Unsupported function types ['_sparse_mm.default', 'randn.default'], indicating that TVM's PyTorch frontend lacks support for these operations.
AssertionError: Unsupported function types ['_sparse_mm.default', 'randn.default']
Environment
- OS: Ubuntu 20.04.6 LTS
- TVM version: 0.23.dev0
- Python version: 3.11.14
Steps to reproduce
import torch
import torch.nn as nn
import tvm
from tvm import relax
class MinimalSparseAndRandomModel(nn.Module):
def __init__(self):
super(MinimalSparseAndRandomModel, self).__init__()
def forward(self, sparse_input):
# Unsupported operations
random_matrix = torch.randn(100, 50) # randn.default
result = torch.sparse.mm(sparse_input, random_matrix) # _sparse_mm.default
return result
model = MinimalSparseAndRandomModel()
model.eval()
indices = torch.tensor([[0, 1, 2], [2, 0, 1]])
values = torch.tensor([1.0, 2.0, 3.0])
sparse_input = torch.sparse_coo_tensor(indices, values, size=(3, 100))
# PyTorch execution works
with torch.no_grad():
output = model(sparse_input)
# PyTorch export works
exported_program = torch.export.export(model, (sparse_input,))
# TVM conversion fails
from tvm.relax.frontend.torch import from_exported_program
mod = from_exported_program(exported_program) # AssertionError here
Triage
- needs-triage
