Repository navigation

feat(autogram): Add `StructuredSparseTensor`. - #466

Closed

PierreQuinton wants to merge 188 commits into

dev-new-enginefrom

block-diagonal-tensor

Closed

feat(autogram): Add `StructuredSparseTensor`.#466
PierreQuinton wants to merge 188 commits into
dev-new-enginefrom
block-diagonal-tensor

Commits on Oct 20, 2025
- Add `DiagonalSparseTensor` with the default fallback to dense mechanism.
  PierreQuinton
  committed
  32ef75e
- Ignore mypy
  PierreQuinton
  committed
  5c76d69
- Remove useless comment.
  PierreQuinton
  committed
  34f4dce
- Change repr
  PierreQuinton
  committed
  a0b7ffc
- revert removing `__init__`
  PierreQuinton
  committed
  f476b29
Commits on Oct 21, 2025
- Give implementation for pointwise
  PierreQuinton
  committed
  447d714
- Add decorator to handle other functions. Add two examples of such functions (mean and sum).
  PierreQuinton
  committed
  85556a8
- Improve naming.
  PierreQuinton
  committed
  c5f868c
Commits on Oct 22, 2025
- Merge branch 'main' into block-diagonal-tensor
  ValerianRey
  authored
  7f5c097
- improve
  PierreQuinton
  committed
  efa8019
Commits on Oct 23, 2025
Commits on Oct 24, 2025
Commits on Oct 25, 2025
- Make `contiguous_data` and `v_to_p` public.
  PierreQuinton
  committed
  17abe7f
Commits on Oct 26, 2025
- Use wraps decorator instead of update_wrapper
  ValerianRey
  committed
  6bbd702
- Make densification explicit in tests before assert_close
  ValerianRey
  committed
  fac5f7d
- Revert "Make `contiguous_data` and `v_to_p` public."
  ValerianRey
  committed
  68bb2a3
- Add pow implementation for DST
  ValerianRey
  committed
  3db55c3
- Remove type hint of __torch_dispatch__ so that mypy stops complaining
  ValerianRey
  committed
  90b205c
- Make signature of to_dense match that of Tensor.to_dense.
  ValerianRey
  committed
  2b36084
- Change type: ignore so that mypy doesn't complain
  ValerianRey
  committed
  6236baa
- Add unsqueeze_default implementation for DST
  ValerianRey
  committed
  6d28918
- Add more info in the print when falling back to dense
  ValerianRey
  committed
  f4a436c
- Add implementation for trivial views
  ValerianRey
  committed
  ad4d843
- Add expand_default, div_Scalar, and slice_Tensor
  ValerianRey
  committed
  b353ff1
- Add mul_Tensor and transpose_int
  ValerianRey
  committed
  d640266
- Use diagonal_sparse_tensor in view_default
  ValerianRey
  committed
  6145e4b
- Use DiagonalSparseTensor type hint when applicable
  ValerianRey
  committed
  444e7cd
- Add debug_info
  ValerianRey
  committed
  5c859b5
- Improve error message in view_default
  ValerianRey
  committed
  960de60
- Add multi-dimensional v_to_p
  ValerianRey
  committed
  e12080e
- Rename mean and sum to mean_default and sum_default
  ValerianRey
  committed
  a38f907
Commits on Oct 27, 2025
- Fix condition and comment in diagonal_sparse_tensor
  ValerianRey
  committed
  b7a3f53
- Improve implementation of view
  ValerianRey
  committed
  aa3d8a0
- Implement `einsum` for `v_to_p: list[int]`
  PierreQuinton
  committed
  c49411f
- Factor access to physical shape in view.
  PierreQuinton
  committed
  6c0e9ec
- rename `physical_to_virtual` to `p_to_vs`.
  PierreQuinton
  committed
  2f774db
- Refactor p_to_vs.
  PierreQuinton
  committed
  7344118
- Adapt `einsum` to `v_to_p: list[list[int]]`
  PierreQuinton
  committed
  0884495
- Rename `v_to_p` to `v_to_ps`
  PierreQuinton
  committed
  a5fda15
- Remove diagonal_sparse_tensor, remove to_diagonal_sparse_tensor
  ValerianRey
  committed
  d03d3a4
- Remove aten.div.Scalar from _POINTWISE_FUNCTIONS
  ValerianRey
  committed
  3bf75fb
- Fix set union to be able to handle empty list of sets
  ValerianRey
  committed
  35e841a
- fix type hint of bmm_default
  ValerianRey
  committed
  2d37031
- Add mm_default
  ValerianRey
  committed
  7a84940
- Use to_dense when comparing result of einsum in test_einsum
  ValerianRey
  committed
  30adf30
- Make einsum work with DST only, re-add to_diagonal_sparse_tensor
  ValerianRey
  committed
  e6ce46e
- Remove list comprehension
  PierreQuinton
  committed
  5253363
- Add sortin function
  PierreQuinton
  committed
  75d138c
Commits on Oct 28, 2025
- Add test_view2 and new parametrization for test_view
  ValerianRey
  committed
  f33107b
- Add test_sort_dst
  ValerianRey
  committed
  17047f6
- Revert "Remove list comprehension"
  ValerianRey
  committed
  6d1cfa5
- Rename map to mapping in sort_dst
  ValerianRey
  committed
  a5390b7
- Use empty ps to indicate virtual dim of size one.
  ValerianRey
  committed
  2cc95ea
- Remove outdated todo in einsum
  ValerianRey
  committed
  ef5a2b1
- Improve implementation of sort_dst
  ValerianRey
  committed
  ea7d4dd
- Restructure packages
  ValerianRey
  committed
  8403b11
- Make _HANDLED_FUNCTIONS and implements class attributes of DST
  ValerianRey
  committed
  f2b2ef6
- Move pointwise function implementations to the end of the file and avoid name shadowing
  ValerianRey
  committed
  97213eb
- Uniformize name of physical
  ValerianRey
  committed
  866b4b1
- Improve print when falling back to dense
  ValerianRey
  committed
  b73de79
- Rename first_sort to encode_by_order and improve its docstring
  ValerianRey
  committed
  e393096
- Improve repr of DST
  ValerianRey
  committed
  7a8b8a4
- Minor formatting fix
  ValerianRey
  committed
  28c72ed
- Move _strides to where it is used:
  ValerianRey
  committed
  1c1885c
- Minor reformating
  ValerianRey
  committed
  09f3efc
- Add check that v_to_ps are correctly encoded
  ValerianRey
  committed
  5703fd2
- Fix to_diagonal_sparse_tensor to not create physical with dim of size 1.
  ValerianRey
  committed
  8a13d49
- Add function to create a DST without having to care about dims of size 1 or incorrect encoding of v_to_ps.
  ValerianRey
  committed
  82d4b33
- Remove test_view and rename test_view2 to test_view
  ValerianRey
  committed
  9d21a6c
- Minor reformatting
  ValerianRey
  committed
  42e8228
- Move p_to_vs outside of DiagonalSparseTensor and rename it p_to_vs_from_v_to_ps
  ValerianRey
  committed
  33194c0
- Add get_groupings, fix_ungrouped_dims, and use it in constructor
  ValerianRey
  committed
  21ca8e3
- Add check of maximal grouping in DST.__init__
  ValerianRey
  committed
  15a615f
- Use make_dst in einsum
  ValerianRey
  committed
  ee8e181
- Add possibility to slice dimension of size 1.
  ValerianRey
  committed
  8992d3f
- Fix to_dense when a pdims is []
  ValerianRey
  committed
  cf4f950
Commits on Oct 29, 2025
- Stop creating index tensors on data device. I think it's ok / faster like that. Add todo to benchmark this later
  ValerianRey
  committed
  ebffc3c
- Rename current_product to current_size
  ValerianRey
  committed
  fe84b80
- Add unsquash_pdim
  ValerianRey
  committed
  92bcaa6
- Revamp unsquash_dim:
  ValerianRey
  committed
  ceeeea6
- Add possibility to unsquash physical dimensions in view.
  ValerianRey
  committed
  6ab6b4c
- Fix get_groupings and add test that failed before this fix
  ValerianRey
  committed
  b746af8
- Fix dim encoding in transpose_int
  ValerianRey
  committed
  59d1f2c
- Name more variables in __torch_dispatch__
  ValerianRey
  committed
  4c48523
- Remove todos in view_default
  ValerianRey
  committed
  123ac3c
- Add support for shape inference in view_default
  ValerianRey
  committed
  469ddb3
- Add _unsafe_view_default
  ValerianRey
  committed
  bd3e569
- Add threshold_backward_default
  ValerianRey
  committed
  42462bc
- Add sum_dim_IntList
  ValerianRey
  committed
  19e049d
- Add broadcast_tensors_default
  ValerianRey
  committed
  519d11b
- Fix mul_Tensor
  ValerianRey
  committed
  64d0ae8
- Add missing default value for keepdim in sum_dim_IntList
  ValerianRey
  committed
  47f32f8
- Improve print when falling back to dense
  ValerianRey
  committed
  b4eb021
- Add mul_Scalar
  ValerianRey
  committed
  909bcb7
- Fix mul_Tensor to be able to handle non-tensor input
  ValerianRey
  committed
  c72f99e
- Make test_einsum parametrizable
  ValerianRey
  committed
  4b3593b
- Add failing example to test_einsum
  ValerianRey
  committed
  2052d81
- Add even simpler failing example in test_einsum
  ValerianRey
  committed
  a1ccad5
- Add assertion about result being DST in test_einsum
  ValerianRey
  committed
  2675d8e
- Add a test to get_groupings. May want to add the same to `test_fix_ungrouped_dims`
  PierreQuinton
  committed
  3121b6a
- Merge remote-tracking branch 'origin/block-diagonal-tensor' into block-diagonal-tensor
  PierreQuinton
  committed
  048a1e2
- Add `strides_from_p_dims_and_p_shape`
  PierreQuinton
  committed
  5a2a7a0
- Add `merge_strides`
  PierreQuinton
  committed
  4b83644
- Add `stride_to_shape`
  PierreQuinton
  committed
  7e8439a
- Add `to_target_physical_strides`
  PierreQuinton
  committed
  c7047e1
- Add new_implementation idea in einsum.
  PierreQuinton
  committed
  c6e3fd9
Commits on Oct 30, 2025
- Merge branch 'dev-new-engine' into block-diagonal-tensor
  ValerianRey
  committed
  cecd69b
- Always use randn_ in test
  ValerianRey
  committed
  d2e53a3
- Fix order of sorting in merg_strides
  PierreQuinton
  committed
  16e7c1c
- Merge remote-tracking branch 'origin/block-diagonal-tensor' into block-diagonal-tensor
  PierreQuinton
  committed
  2c23dbe
- Add strides_v2
  ValerianRey
  committed
  16e6165
- Add more failing parametrizations to test_get_groupings and test_fix_ungrouped_dims, remove print
  ValerianRey
  committed
  59cf10b
- Add test_concatenate
  ValerianRey
  committed
  e0dc1a7
- Add strides_to_pdims
  ValerianRey
  committed
  3dffd1e
Commits on Oct 31, 2025
- Add (passing) test_to_dense2 to test to_dense when the tensor has a virtual dimension that uses a physical dimension multiple times.
  ValerianRey
  committed
  48387ab
- Use new strides in to_dense
  ValerianRey
  committed
  f87ecbb
- Simplify to_dense
  ValerianRey
  committed
  b883f20
- Remove unused variable dims in the loop of to_dense
  ValerianRey
  committed
  7044e37
- Pre-compute strides
  ValerianRey
  committed
  384d550
- Replace for-loop with for comprehension to create v_indices_grid, and make it directly as a tuple
  ValerianRey
  committed
  18044dd
- Remove for-loop in computation of v_indices_grid in to_dense
  ValerianRey
  committed
  b0a0e7a
- Add internal strides to debug_info
  ValerianRey
  committed
  456adf1
- Simplify creation of strides in to_dense
  ValerianRey
  committed
  dc37d29
- Remove torch. prefix when possible
  ValerianRey
  committed
  a9e298a
- Create strides tensor in constructor
  ValerianRey
  committed
  ac2a2c5
- Remove comments about the device of the indices tensors. We actually cannot have strides on cuda because addmm_cuda (required for tensordot) does not support Long tensors (but the cpu version does).
  ValerianRey
  committed
  cfba7e0
- Replace self.physical.device and self.physical.dtype by self.device and self.dtype
  ValerianRey
  committed
  2c94488
- Move unwrap_to_dense out of DiagonalSparseTensor
  ValerianRey
  committed
  9c1ad5b
- Extract print_fallback
  ValerianRey
  committed
  7252757
- Add placeholder cat_default
  ValerianRey
  committed
  7b4c784
- Add to_most_efficient_tensor, use it in view and einsum
  ValerianRey
  committed
  767281c
- Add permute_default
  ValerianRey
  committed
  6a3145f
- Fix some return type hints
  ValerianRey
  committed
  9f1860a
- Add div_Tensor, factorize prepare_for_elementwise_op
  ValerianRey
  committed
  137f64e
- Add add_Tensor for same v_to_ps
  ValerianRey
  committed
  dc4c1a5
Commits on Nov 1, 2025
- Add squeeze_dims
  ValerianRey
  committed
  f057d7e
- Handle cases where pdims is empty list in the creation of normal tensor in to_most_efficient_tensor
  ValerianRey
  committed
  dc35696
- Add hardtanh_backward_default
  ValerianRey
  committed
  c97612f
- Add hardswish_backward_default
  ValerianRey
  committed
  56b66ba
- Add more concatenate tests
  ValerianRey
  committed
  10f250c
Commits on Nov 2, 2025
- Add comment about strides
  ValerianRey
  committed
  48de187
- unsquash_pdim_from_strides
  ValerianRey
  committed
  b60a29d
- Simplify test_concatenate params
  ValerianRey
  committed
  9ede7ec
- Add basic implementation of cat_default for when all strides match and the pdim on which we concatenate already exists.
  ValerianRey
  committed
  5a25233
- Add concat implementation for when the physical dimension on which to concatenate has to be added
  ValerianRey
  committed
  5062b06
- Revamp grouping detection
  ValerianRey
  committed
  3b9a9a8
- Remove get_groupings_generalized
  ValerianRey
  committed
  ab15e1a
- Remove longest_common_prefix
  ValerianRey
  committed
  26de009
- Restructure sparse package
  ValerianRey
  committed
  2419c7e
- Rename DiagonalSparseTensor to StructuredSparseTensor
  ValerianRey
  committed
  59bcf06
- Improve error message for cat_default
  ValerianRey
  committed
  f693e99
- Add alias impl for StructuedSparseDensity.implements
  ValerianRey
  committed
  7991ac1
- Improve error message in cat_default
  ValerianRey
  committed
  93e3a61
Commits on Nov 3, 2025
- Rename variables in einsum
  ValerianRey
  committed
  3c3b4a4
- Improve error message in einsum
  ValerianRey
  committed
  b303501
- Add failing parametrization of test_einsum
  ValerianRey
  committed
  81dd29e
- Move `einsum` to the top of the file `einsum.py`
  PierreQuinton
  committed
  8593a74
Commits on Nov 5, 2025
- clear_null_stride_columns
  mattbuot
  committed
  4ccef3a
- refacto fix zero stride columns
  mattbuot
  committed
  85dc26b
- Add jupyter notebooks to .gitignore
  ValerianRey
  committed
  1c73416
- Use strides-based representation instead of v_to_ps-based
  ValerianRey
  committed
  5547ff6
Commits on Nov 6, 2025
- Fix StructuredSparseTensor.__new__
  ValerianRey
  committed
  7f2c974
- Fix tests
  ValerianRey
  committed
  51a579a
Commits on Nov 7, 2025
- Add initial implementation of view_default
  ValerianRey
  committed
  6e94a89
- Fix order of rows of new_strides in view_default
  ValerianRey
  committed
  fc3339c
- Fix creation of SST in autogram
  ValerianRey
  committed
  fac9c72
- One-line view_default.
  ValerianRey
  committed
  66c2210
Commits on Nov 12, 2025
- update interface of unsquash_pdim
  mattbuot
  committed
  5cbab08
- Add get_full_source
  ValerianRey
  committed
  a5c8cbf
- Remove alternative implementation
  ValerianRey
  committed
  c1fc115
Commits on Nov 17, 2025
- Remove useless clone
  ValerianRey
  committed
  bb8faaf

Commits

Commits on Oct 20, 2025

Commits on Oct 21, 2025

Commits on Oct 22, 2025

Commits on Oct 23, 2025

Commits on Oct 24, 2025

Commits on Oct 25, 2025

Commits on Oct 26, 2025

Commits on Oct 27, 2025

Commits on Oct 28, 2025

Commits on Oct 29, 2025

Commits on Oct 30, 2025

Commits on Oct 31, 2025