torch-sql

torch-sql is a skunkworks prototype that runs PyTorch programs, or parts of them, as SQL on XQL, a group of database projects for scientific arrays. It has three frontends:

• a co-processor, which leaves the model in PyTorch and takes over the training data pipeline and evaluation;
• a device, device="sql", on which every tensor is a table;
• a torch.compile backend.

Its first real model is Samudra, an ocean emulator.

The first milestone, M1 in the design, is under way: a two-layer MLP, then nanoGPT (Andrej Karpathy's small GPT), trained on device="sql". Start with the docs:

• docs/design.md: the design.
• docs/samudra-on-beta.md: the production plan for Samudra on Empire AI's Beta cluster.
• docs/spikes/: small programs that answer the design's open questions, with their results.

Development

torch-sql uses uv. From the repository root:

uv sync --all-extras   # PyTorch (CPU), ddxdb, DataFusion and DuckDB
uv run pytest          # every test runs on both engines
uv run ruff check && uv run ruff format --check

On NixOS, first put libstdc++ and zlib on LD_LIBRARY_PATH, as the XQL repositories' flake.nix files do.

Example

nanoGPT's model.py, unchanged but for its GELU, trains on device="sql" and matches PyTorch step by step (tests/test_nanogpt.py). A smaller example, a two-layer MLP run on DuckDB:

import torch, torch_sql

with torch_sql.Environment("duckdb"):
    mlp = torch.nn.Sequential(torch.nn.Linear(3, 5), torch.nn.Tanh(), torch.nn.Linear(5, 2))
    mlp = mlp.double().to("sql")
    y = mlp(torch.randn(4, 3, dtype=torch.float64, device="sql"))
    print(y)  # runs one SQL query on DuckDB