Repository navigation

Compute a missing hop_distance on the padded Graph Transformer k-hop path - #850

Draft

kmontemayor2-sc wants to merge 99 commits into

mainfrom

kmonte/gt-padded-hop-distance

Draft

Compute a missing hop_distance on the padded Graph Transformer k-hop path#850
kmontemayor2-sc wants to merge 99 commits into
mainfrom
kmonte/gt-padded-hop-distance

Commits on Oct 17, 2026
- Build GT sequence inputs from tensor dicts instead of an internal HeteroData
  
  kmonte
  and
  claude
  committed
  ad3a4b1
Commits on Oct 19, 2026
- Merge origin/main into kmonte/gt-tensor-dict-inputs
  
  kmonte
  and
  claude
  committed
  74c9c48
- Read typed-PPR channel masks without a HeteroData attribute miss
  
  kmonte
  and
  claude
  committed
  4abb539
- Skip edge stores without edge_index in GraphTransformerEncoder.forward
  
  kmonte
  and
  claude
  committed
  2dbcfb0
- Run supported GraphTransformerEncoder configs on packed tokens
  
  kmonte
  and
  claude
  committed
  ae39790
- Gate the packed path on head dim and hop count; tidy its helpers
  
  kmonte
  and
  claude
  committed
  7f534a7
- Keep the CPU reference batch on CPU in the CUDA packed tests
  
  kmonte
  and
  claude
  committed
  734c95b
- Rebuild CUDA test batches instead of deep-copying ones that cache a sparse layout
  
  kmonte
  and
  claude
  committed
  62b17d4
Commits on Oct 20, 2026
- Address review of the packed backend: small-batch backward, selection blockers, layout budget
  
  kmonte
  and
  claude
  committed
  2e0b8d1
Commits on Oct 21, 2026
- Packed attention: run the varlen kernels as custom ops
  kmontemayor
  committed
  82de400
- Packed forward: document why the encode call is not nested in the output projection
  kmontemayor
  committed
  9ba17da
- Packed attention: type the registered custom-op handles as Any for ty
  kmontemayor
  committed
  d97d99d
- Packed attention: drop the unused Philox outputs from the varlen ops
  
  kmontemayor
  and
  claude
  committed
  80cd1c4
- Packed attention and static k-hop: jaxtyping shapes
  
  kmontemayor
  and
  claude
  committed
  8428da5
- Packed attention and static k-hop: module docs with data flow and usage
  
  kmontemayor
  and
  claude
  committed
  082f97b
- Packed encoder: explain why the packed result is assigned before projecting
  
  kmontemayor
  and
  claude
  committed
  b610e69
- Packed encoder: count padded-path fallbacks and warn at powers of two
  
  kmontemayor
  and
  claude
  committed
  e832102
- Humanify packed_attention comments and docstrings
  
  kmontemayor
  and
  claude
  committed
  2c3619e
- Humanify static_khop comments, docstrings and error message
  
  kmontemayor
  and
  claude
  committed
  bb35175
- Humanify GraphTransformerEncoder packed-path docs, comments and messages
  
  kmontemayor
  and
  claude
  committed
  fa7cc05
- Format graph_transformer.py
  
  kmontemayor
  and
  claude
  committed
  e86e1f9
- Test the oversized-layout fallback end to end instead of mocking khop_layout and the logger
  
  kmontemayor
  and
  claude
  committed
  e3ab8dc
- Give the varlen attention custom ops explicit schemas so they carry jaxtyping shapes; add at-scale results to the docstring
  
  kmontemayor
  and
  claude
  committed
  ec0a56e
- Packed encoder review fixes: check the hop embedding size, require the hop attribute on both paths, clamp pack_sequences lengths
  
  kmontemayor
  and
  claude
  committed
  6cb5564
- Packed attention: call aten's efficient-attention kernel directly instead of custom ops
  
  kmontemayor
  and
  claude
  committed
  dd56e03
- Run hop_distance up to 4 and prioritize_hop_order on the packed backend
  
  kmontemayor
  and
  claude
  committed
  3cec3b2
- Packed encoder: keep running without the hop_distance attribute; document the asymmetry
  
  kmontemayor
  and
  claude
  committed
  a93643d
- Packed tests: type the override kwargs as Any
  
  kmontemayor
  and
  claude
  committed
  f2b646e
- Packed attention: document why the varlen call uses the private aten op, with a TODO for torch >= 2.10
  kmontemayor
  committed
  9e13245
- Packed attention: tag the torch >= 2.10 varlen TODO with an owner
  kmontemayor
  committed
  cc9b5ea
Commits on Oct 22, 2026
- Merge origin/main into kmonte/gt-packed-backend
  
  kmontemayor
  and
  claude
  committed
  c2f18c4
- Document the packed forward worked example and spell out khop_layout outputs; replace ternaries with if/else
  
  kmontemayor
  and
  claude
  committed
  9397f2d
- Packed backend: raise a clear error for a fallback batch without hop_distance; update CHANGELOG
  
  kmontemayor
  and
  claude
  committed
  be3a8ec
- Packed layout: simpler sequence layout, combined node/hop key, budget the neighbour table
  
  kmontemayor
  and
  claude
  committed
  69ecda1
- Packed backend: project inside _encode_packed, drop the FFN batch dim, inline the hop embedding check
  
  kmontemayor
  and
  claude
  committed
  af66666
- Packed attention: one code path for the padded reference
  
  kmontemayor
  and
  claude
  committed
  6e44b5e
- Packed tests: parametrize the repeated matches-padded, gradient, compile and layout tests
  
  kmontemayor
  and
  claude
  committed
  42dac81
- Packed docstrings: state each constraint once and drop Args that restate the shapes
  
  kmontemayor
  and
  claude
  committed
  85d1cc7
- Packed attention and static k-hop: lead the docs with the objective and the why
  
  kmontemayor
  and
  claude
  committed
  c2a148c
Commits on Oct 23, 2026
- Add a fixed-shape k-hop sequence layout for packed graph transformer tokens
  
  kmontemayor
  and
  claude
  committed
  e11308b
- Add varlen attention over packed sequences
  
  kmontemayor
  and
  claude
  committed
  7d60d84
- Merge kmonte/gt-packed-attention into kmonte/gt-packed-backend
  kmontemayor
  committed
  3458198
- Move the attention and pack_sequences tests to their own PRs' test files
  
  kmontemayor
  and
  claude
  committed
  4a1f056
- Static k-hop: get the sorted sources and their order from one sort
  
  kmontemayor
  and
  claude
  committed
  8b06c18
- Merge branch 'kmonte/gt-static-khop' into kmonte/gt-packed-attention
  kmontemayor
  committed
  993cbc1
- Merge branch 'kmonte/gt-packed-attention' into kmonte/gt-packed-backend
  kmontemayor
  committed
  166032d
Commits on Oct 29, 2026
- Static k-hop: validate sampling_direction; move the packed layout type to packed_layout.py
  
  kmontemayor
  and
  claude
  committed
  2c8c2d2
- Merge branch 'kmonte/gt-static-khop' into kmonte/gt-packed-attention
  kmontemayor
  committed
  4cb1191
- Import PackedLayout and pack_sequences from packed_layout in packed attention
  
  kmontemayor
  and
  claude
  committed
  583362a
- Merge branch 'kmonte/gt-packed-attention' into kmonte/gt-packed-backend
  kmontemayor
  committed
  7f6ad54
- Import PackedLayout and pack_sequences from packed_layout in the packed backend
  
  kmontemayor
  and
  claude
  committed
  a0cd373
- Static k-hop: share the select-and-pack step as select_sequences; replace prioritize_hop_order with an ordering option
  
  kmontemayor
  and
  claude
  committed
  b3fec14
- Merge branch 'kmonte/gt-static-khop' into kmonte/gt-packed-attention
  kmontemayor
  committed
  85936e4
- Merge branch 'kmonte/gt-packed-attention' into kmonte/gt-packed-backend
  kmontemayor
  committed
  dc6fdde
- Packed backend: pass the khop_layout ordering option
  
  kmontemayor
  and
  claude
  committed
  7f33c36
- Rename gigl.transforms.packed_layout to sequence_layout
  
  kmontemayor
  and
  claude
  committed
  b7be346
- Build the padded PPR sequences with select_sequences
  
  kmontemayor
  and
  claude
  committed
  74dae18
- Merge branch 'kmonte/gt-static-khop' into kmonte/gt-packed-attention
  kmontemayor
  committed
  01c624a
- Packed attention: import from gigl.transforms.sequence_layout
  
  kmontemayor
  and
  claude
  committed
  4bdf14e
- Merge branch 'kmonte/gt-packed-attention' into kmonte/gt-packed-backend
  
  kmontemayor
  and
  claude
  committed
  52749b1
- Add a failing CUDA test: compiled first-token varlen attention crashes backward for one sequence
  
  kmontemayor
  and
  claude
  committed
  fc850b6
- Packed attention: wrap the varlen kernel in a custom op that hides its CPU philox tensors from Inductor
  
  kmontemayor
  and
  claude
  committed
  7bfa74e
- Merge branch 'kmonte/gt-packed-attention' into kmonte/gt-packed-backend
  kmontemayor
  committed
  9973f30
Commits on Oct 30, 2026
- Reset dynamo counters in the no-recompile test
  
  kmontemayor
  and
  claude
  committed
  80f7f5d
- Return named tuples from select_sequences and khop_layout, with doctests
  
  kmontemayor
  and
  claude
  committed
  9a7f03c
- Merge branch 'kmonte/gt-static-khop' into kmonte/gt-packed-attention
  kmontemayor
  committed
  88636d4
- Merge branch 'kmonte/gt-packed-attention' into kmonte/gt-packed-backend
  kmontemayor
  committed
  b350636
- Merge branch 'main' into kmonte/gt-static-khop
  kmontemayor2-sc
  authored
  4db7239
- Merge kmonte/gt-static-khop into kmonte/gt-packed-attention
  
  kmontemayor
  and
  claude
  committed
  6fb0fc9
- Merge kmonte/gt-packed-attention into kmonte/gt-packed-backend
  
  kmontemayor
  and
  claude
  committed
  8cf3f4b
Commits on Nov 3, 2026
- Rename packed backend blockers to unsupported reasons
  
  kmonte
  and
  claude
  committed
  3ed7e29
- Fix the FeedForwardNetwork Shape Contract for packed tokens
  
  kmonte
  and
  claude
  committed
  c5b952c
- Skip Shape Contract checks while torch.compile traces
  
  kmonte
  and
  claude
  committed
  d6fb8d7
- Reject per-layer torch.compile'd layers on the packed path
  
  kmonte
  and
  claude
  committed
  9be6dbe
- Fix type annotations in the compiled layer tests
  
  kmonte
  and
  claude
  committed
  f2572bc
- Keep Jaxtyping's wrapper out of torch.compile in Shape Contract tests
  
  kmonte
  and
  claude
  committed
  d2c12ec
- Show per-layer compile in place in the compiled-layer error
  
  kmonte
  and
  claude
  committed
  995e633
- Merge branch 'main' into kmonte/gt-static-khop
  kmontemayor2-sc
  authored
  e1a16af
- Keep Jaxtyping's wrapper out of torch.compile in Shape Contract tests
  
  kmonte
  and
  claude
  committed
  97331fe
- Merge branch 'kmonte/gt-static-khop' into kmonte/gt-packed-attention
  kmonte
  committed
  7509d27
- Merge branch 'kmonte/gt-packed-attention' into kmonte/gt-packed-backend
  kmonte
  committed
  aaee837
Commits on Nov 4, 2026
- Add an anchor hop distance matrix for k-hop sequences
  
  kmonte
  and
  claude
  committed
  744cf71
- Compute a missing anchor hop_distance on the padded k-hop path
  
  kmonte
  and
  claude
  committed
  350148c
- Run fallback batches without the hop_distance attribute
  
  kmonte
  and
  claude
  committed
  7fda734
- Use explicit test failures and list the hop_distance transform settings
  
  kmonte
  and
  claude
  committed
  74911b9
- Merge branch 'main' into kmonte/gt-packed-attention
  kmontemayor2-sc
  authored
  54ac615
- Packed attention: name the layout argument layout; test through packed_attention
  
  kmontemayor
  and
  claude
  committed
  90f0006
- Merge kmonte/gt-packed-attention into kmonte/gt-packed-backend
  kmonte
  committed
  6afcb28
- Merge kmonte/gt-packed-backend into kmonte/gt-padded-hop-distance
  kmonte
  committed
  b7c1d03
Commits on Nov 5, 2026
- Document the non-empty sequence precondition of packed_attention
  
  kmontemayor
  and
  claude
  committed
  729a38f
- Merge kmonte/gt-packed-attention into kmonte/gt-packed-backend
  kmontemayor
  committed
  22b8f0b
- Check the hop embedding size for khop configs only
  
  kmontemayor
  and
  claude
  committed
  805100b
- Merge kmonte/gt-packed-backend into kmonte/gt-padded-hop-distance
  kmonte
  committed
  83fc830
- Update the changelog for the khop-only hop embedding check
  
  kmontemayor
  and
  claude
  committed
  6260037
- Merge kmonte/gt-packed-backend into kmonte/gt-padded-hop-distance
  kmonte
  committed
  9c89357
- Merge main into kmonte/gt-packed-backend
  
  kmontemayor
  and
  claude
  committed
  7ef624f
- Merge kmonte/gt-packed-backend into kmonte/gt-padded-hop-distance
  
  kmonte
  and
  claude
  committed
  04d28a2
- Merge main into kmonte/gt-padded-hop-distance
  
  kmonte
  and
  claude
  committed
  ade5d31
- Merge main into kmonte/gt-padded-hop-distance
  kmonte
  committed
  1e269e9

Commits

Commits on Oct 17, 2026

Commits on Oct 19, 2026

Commits on Oct 20, 2026

Commits on Oct 21, 2026

Commits on Oct 22, 2026

Commits on Oct 23, 2026

Commits on Oct 29, 2026

Commits on Oct 30, 2026

Commits on Nov 3, 2026

Commits on Nov 4, 2026

Commits on Nov 5, 2026