Repository navigation

WIP: GraphCast_small in SQL, with DeepMind's published weights - #270

Draft

alxmrs wants to merge 30 commits into

mainfrom

graphcast-weights

Draft

WIP: GraphCast_small in SQL, with DeepMind's published weights#270
alxmrs wants to merge 30 commits into
mainfrom
graphcast-weights

Commits on Oct 26, 2026
- Add an ml extra for gradients in SQL, and three ML demos
  
  alxmrs
  and
  claude
  committed
  53958d4
- Address review: complete SQL statements, nanoGPT's model, introductions
  
  alxmrs
  and
  claude
  committed
  3eee986
- Train the GPT with nanoGPT's AdamW behind --optimizer adamw
  
  alxmrs
  and
  claude
  committed
  a4ef3b2
- Define attention once, as a view, and its loss separately
  
  alxmrs
  and
  claude
  committed
  8ddc585
- Split the GPT into a model view and a separate loss
  
  alxmrs
  and
  claude
  committed
  e459a8c
- Fix lint: format the GraphCast script, and type two values for mypy
  
  alxmrs
  and
  claude
  committed
  403a1f5
- Differentiate the softmax's window MAX directly with ddxdb 0.3.1
  
  alxmrs
  and
  claude
  committed
  65706fd
- Require ddxdb 0.3.2 and record its faster GPT steps
  
  alxmrs
  and
  claude
  committed
  a2eb466
- Require ddxdb 0.3.3 and record its faster gradients
  
  alxmrs
  and
  claude
  committed
  a61842c
- Require ddxdb 0.4.0, on DataFusion 55
  
  alxmrs
  and
  claude
  committed
  574e9c2
- Record the demos on ddxdb 0.4.0 and in-memory tables
  
  alxmrs
  and
  claude
  committed
  229b842
- Take gradients on DuckDB through ddx
  
  alxmrs
  and
  claude
  committed
  03c1d51
- Size the GPT from the command line, and record where its time goes
  
  alxmrs
  and
  claude
  committed
  d5c03ab
- Move the GraphCast demo to its own PR
  
  alxmrs
  and
  claude
  committed
  ff7352b
- Build the GPT's block from attention, mlp and residual, as nanoGPT does
  
  alxmrs
  and
  claude
  committed
  1c08ea4
- Run the ML demos in CI, and say how to contribute one
  
  alxmrs
  and
  claude
  committed
  dc515a5
- Say that a mask is absent rows, and an infinity only tests it
  
  alxmrs
  and
  claude
  committed
  508ae96
- Require ddxdb 0.4.2, and record the demos on it
  
  alxmrs
  and
  claude
  committed
  28f3cb0
Commits on Oct 27, 2026
- Address review: fail on new refusals, run AdamW, lock JAX and ddx's DuckDB
  
  alxmrs
  and
  claude
  committed
  38aaf91
- Merge main, with Context.prepare, into ml-demos
  
  alxmrs
  and
  claude
  committed
  883108b
Commits on Oct 28, 2026
- Require ddxdb 0.4.3, which trains the GPT on DuckDB
  
  alxmrs
  and
  claude
  committed
  1361a59
- Run the GPT's CTEs once each on DataFusion, so it trains past one layer
  
  alxmrs
  and
  claude
  committed
  ef30a6e
Commits on Oct 29, 2026
- Require ddxdb 0.4.5, which frees each step's tables after their last reader
  
  alxmrs
  and
  claude
  committed
  f7c39f5
Commits on Oct 30, 2026
- Run the ML demos on einfold 0.1.0rc1, and require ddxdb 0.4.6
  
  alxmrs
  and
  claude
  committed
  d543b55
- Address review: ignore grad( in literals, and skip JAX on Intel macOS
  
  alxmrs
  and
  claude
  committed
  8a7aa90
- Require einfold 0.1.0rc2, and install the test extra for the ML demos
  
  alxmrs
  and
  claude
  committed
  cb935d0
- Write the GPT as functions of relations, as nanoGPT's modules are
  
  alxmrs
  and
  claude
  committed
  714ba63
- Write the GPT as plain SQL again, without the shared-CTE workaround
  
  alxmrs
  and
  claude
  committed
  6debe24
Commits on Oct 31, 2026
- Run GraphCast as views of its stages, on the current demos
  
  alxmrs
  and
  claude
  committed
  48e65a1
- List the GraphCast demo, with no run small enough for CI
  
  alxmrs
  and
  claude
  committed
  6505546

Commits

Commits on Oct 26, 2026

Commits on Oct 27, 2026

Commits on Oct 28, 2026

Commits on Oct 29, 2026

Commits on Oct 30, 2026

Commits on Oct 31, 2026