ferrotensor 0.3.0__tar.gz → 0.4.0__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (76) hide show
  1. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/Cargo.lock +11 -11
  2. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/Cargo.toml +1 -1
  3. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/PKG-INFO +1 -1
  4. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/Cargo.toml +1 -1
  5. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/backend/compute.rs +6 -1
  6. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/backend/kernels.rs +237 -0
  7. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/functional.rs +107 -12
  8. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/tensor/cpu.rs +96 -0
  9. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/tensor/ops.rs +23 -49
  10. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/tensor/raw.rs +245 -0
  11. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/tensor/tests.rs +100 -0
  12. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/README.md +0 -0
  13. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/README.md +0 -0
  14. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/examples/README.md +0 -0
  15. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/examples/char_transformer.rs +0 -0
  16. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/examples/data/sonnets.txt +0 -0
  17. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/examples/gpu_matmul_attention.rs +0 -0
  18. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/examples/mlp_two_moons.rs +0 -0
  19. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/examples/quickstart.rs +0 -0
  20. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/examples/similarity_search.rs +0 -0
  21. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/backend/mod.rs +0 -0
  22. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/data.rs +0 -0
  23. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/device.rs +0 -0
  24. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/lib.rs +0 -0
  25. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/nn/activation.rs +0 -0
  26. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/nn/attention.rs +0 -0
  27. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/nn/container.rs +0 -0
  28. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/nn/dropout.rs +0 -0
  29. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/nn/embedding.rs +0 -0
  30. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/nn/linear.rs +0 -0
  31. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/nn/mod.rs +0 -0
  32. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/nn/module.rs +0 -0
  33. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/nn/normalization.rs +0 -0
  34. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/nn/tests.rs +0 -0
  35. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/optim/adam.rs +0 -0
  36. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/optim/mod.rs +0 -0
  37. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/optim/optimizer.rs +0 -0
  38. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/optim/scheduler.rs +0 -0
  39. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/optim/sgd.rs +0 -0
  40. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/optim/tests.rs +0 -0
  41. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/random.rs +0 -0
  42. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/serialize.rs +0 -0
  43. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/tensor/autograd.rs +0 -0
  44. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/tensor/core.rs +0 -0
  45. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/tensor/mod.rs +0 -0
  46. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor/src/tensor/shape.rs +0 -0
  47. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor-py/.gitignore +0 -0
  48. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor-py/Cargo.toml +0 -0
  49. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor-py/README.md +0 -0
  50. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor-py/build.rs +0 -0
  51. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor-py/examples/mlp.py +0 -0
  52. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor-py/src/device.rs +0 -0
  53. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor-py/src/error.rs +0 -0
  54. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor-py/src/functional.rs +0 -0
  55. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor-py/src/lib.rs +0 -0
  56. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor-py/src/nn.rs +0 -0
  57. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor-py/src/optim.rs +0 -0
  58. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor-py/src/serialize.rs +0 -0
  59. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor-py/src/tensor.rs +0 -0
  60. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor-py/tests/conftest.py +0 -0
  61. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor-py/tests/test_autograd.py +0 -0
  62. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor-py/tests/test_errors.py +0 -0
  63. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor-py/tests/test_functional.py +0 -0
  64. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor-py/tests/test_gpu.py +0 -0
  65. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor-py/tests/test_nn.py +0 -0
  66. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor-py/tests/test_optim.py +0 -0
  67. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor-py/tests/test_serialize.py +0 -0
  68. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor-py/tests/test_tensor.py +0 -0
  69. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/ferrotensor-py/tests/test_train.py +0 -0
  70. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/pyproject.toml +0 -0
  71. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/python/ferrotensor/__init__.py +0 -0
  72. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/python/ferrotensor/_ferrotensor.pyi +0 -0
  73. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/python/ferrotensor/nn/__init__.py +0 -0
  74. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/python/ferrotensor/nn/functional.py +0 -0
  75. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/python/ferrotensor/optim.py +0 -0
  76. {ferrotensor-0.3.0 → ferrotensor-0.4.0}/python/ferrotensor/py.typed +0 -0
@@ -606,7 +606,7 @@ dependencies = [
606
606
 
607
607
  [[package]]
608
608
  name = "ferrotensor"
609
- version = "0.3.0"
609
+ version = "0.4.0"
610
610
  dependencies = [
611
611
  "bytemuck",
612
612
  "naga",
@@ -619,7 +619,7 @@ dependencies = [
619
619
 
620
620
  [[package]]
621
621
  name = "ferrotensor-autograd"
622
- version = "0.3.0"
622
+ version = "0.4.0"
623
623
  dependencies = [
624
624
  "ferrotensor-backend",
625
625
  "pollster",
@@ -628,7 +628,7 @@ dependencies = [
628
628
 
629
629
  [[package]]
630
630
  name = "ferrotensor-backend"
631
- version = "0.3.0"
631
+ version = "0.4.0"
632
632
  dependencies = [
633
633
  "bytemuck",
634
634
  "env_logger",
@@ -641,7 +641,7 @@ dependencies = [
641
641
 
642
642
  [[package]]
643
643
  name = "ferrotensor-benchmarks"
644
- version = "0.3.0"
644
+ version = "0.4.0"
645
645
  dependencies = [
646
646
  "ferrotensor",
647
647
  "ferrotensor-backend",
@@ -652,7 +652,7 @@ dependencies = [
652
652
 
653
653
  [[package]]
654
654
  name = "ferrotensor-cli"
655
- version = "0.3.0"
655
+ version = "0.4.0"
656
656
  dependencies = [
657
657
  "clap",
658
658
  "env_logger",
@@ -670,7 +670,7 @@ dependencies = [
670
670
 
671
671
  [[package]]
672
672
  name = "ferrotensor-graph"
673
- version = "0.3.0"
673
+ version = "0.4.0"
674
674
  dependencies = [
675
675
  "bytemuck",
676
676
  "env_logger",
@@ -685,7 +685,7 @@ dependencies = [
685
685
 
686
686
  [[package]]
687
687
  name = "ferrotensor-hub"
688
- version = "0.3.0"
688
+ version = "0.4.0"
689
689
  dependencies = [
690
690
  "ferrotensor-autograd",
691
691
  "ferrotensor-backend",
@@ -698,7 +698,7 @@ dependencies = [
698
698
 
699
699
  [[package]]
700
700
  name = "ferrotensor-loader"
701
- version = "0.3.0"
701
+ version = "0.4.0"
702
702
  dependencies = [
703
703
  "bytemuck",
704
704
  "env_logger",
@@ -716,7 +716,7 @@ dependencies = [
716
716
 
717
717
  [[package]]
718
718
  name = "ferrotensor-monitor"
719
- version = "0.3.0"
719
+ version = "0.4.0"
720
720
  dependencies = [
721
721
  "anyhow",
722
722
  "chrono",
@@ -730,7 +730,7 @@ dependencies = [
730
730
 
731
731
  [[package]]
732
732
  name = "ferrotensor-py"
733
- version = "0.3.0"
733
+ version = "0.4.0"
734
734
  dependencies = [
735
735
  "ferrotensor",
736
736
  "pyo3",
@@ -739,7 +739,7 @@ dependencies = [
739
739
 
740
740
  [[package]]
741
741
  name = "ferrotensor-trainer"
742
- version = "0.3.0"
742
+ version = "0.4.0"
743
743
  dependencies = [
744
744
  "bytemuck",
745
745
  "ferrotensor-autograd",
@@ -4,7 +4,7 @@ resolver = "2"
4
4
 
5
5
  [workspace.package]
6
6
  edition = "2021"
7
- version = "0.3.0"
7
+ version = "0.4.0"
8
8
  license = "MIT OR Apache-2.0"
9
9
  repository = "https://github.com/puranikyashaswin/ferrotensor"
10
10
  authors = ["Yashaswin Puranik <puranikyashaswinsharma@gmail.com>"]
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: ferrotensor
3
- Version: 0.3.0
3
+ Version: 0.4.0
4
4
  Classifier: Programming Language :: Rust
5
5
  Classifier: Programming Language :: Python :: 3
6
6
  Classifier: Programming Language :: Python :: Implementation :: CPython
@@ -1,6 +1,6 @@
1
1
  [package]
2
2
  name = "ferrotensor"
3
- version = "0.3.0"
3
+ version = "0.4.0"
4
4
  edition = "2021"
5
5
  authors.workspace = true
6
6
  license.workspace = true
@@ -9,7 +9,7 @@ use wgpu::{BindGroup, BindGroupLayout, BindingResource, CommandEncoder, ComputeP
9
9
  use super::KERNELS_WGSL;
10
10
 
11
11
  /// Entry points compiled from [`KERNELS_WGSL`].
12
- pub(crate) const KERNEL_NAMES: [&str; 9] = [
12
+ pub(crate) const KERNEL_NAMES: [&str; 14] = [
13
13
  "unary",
14
14
  "binary",
15
15
  "reduce",
@@ -19,6 +19,11 @@ pub(crate) const KERNEL_NAMES: [&str; 9] = [
19
19
  "matmul",
20
20
  "permute",
21
21
  "gather_rows",
22
+ "sort_ids",
23
+ "scatter_rows",
24
+ "copy_block",
25
+ "cross_entropy_rows",
26
+ "cross_entropy_grad",
22
27
  ];
23
28
 
24
29
  /// Threads per workgroup of the 1-D kernels.
@@ -547,4 +547,241 @@ fn gather_rows(@builtin(global_invocation_id) gid: vec3<u32>,
547
547
  }
548
548
  gat_out[i] = gat_table[id * gat_p.row + c];
549
549
  }
550
+
551
+ // --- Row scatter-add (index_select and embedding backward) -------------------
552
+ // WGSL has no float atomics, so the gradient rows are not added with atomics.
553
+ // Instead the (row, position) pairs of the ids are sorted by row, then by
554
+ // position; head[row] is the index of the row's first pair, or SORT_NONE when
555
+ // no id picks it. One thread per output element adds its row's gradients in
556
+ // position order, as the CPU loop does, so the result is deterministic and
557
+ // matches the CPU bit for bit.
558
+
559
+ const SORT_NONE: u32 = 0xffffffffu;
560
+
561
+ // sort_ids builds the pairs and the head table on the GPU from f32 ids.
562
+ // op 0: pair i = (row of id i, i) for i < n, padded with (SORT_NONE, i) up to
563
+ // a = the power-of-two pair count; b = rows. An id outside [0, rows)
564
+ // gets row SORT_NONE: it sorts last and adds nothing.
565
+ // op 1: one compare-and-swap step of a bitonic sort of n pairs (a = k, b = j).
566
+ // op 2: head[r] = SORT_NONE for r < n.
567
+ // op 3: head[row] = i for each pair i < n that is the first of its row.
568
+
569
+ struct SortParams {
570
+ op: u32,
571
+ n: u32,
572
+ a: u32,
573
+ b: u32,
574
+ };
575
+
576
+ @group(0) @binding(0) var<storage, read> srt_ids: array<f32>;
577
+ @group(0) @binding(1) var<storage, read_write> srt_keys: array<u32>;
578
+ @group(0) @binding(2) var<storage, read_write> srt_head: array<u32>;
579
+ @group(0) @binding(3) var<uniform> srt_p: SortParams;
580
+
581
+ @compute @workgroup_size(256)
582
+ fn sort_ids(@builtin(global_invocation_id) gid: vec3<u32>,
583
+ @builtin(num_workgroups) nwg: vec3<u32>) {
584
+ let i = flat_index(gid, nwg);
585
+ switch srt_p.op {
586
+ case 0u: {
587
+ if (i >= srt_p.a) { return; }
588
+ var row = SORT_NONE;
589
+ if (i < srt_p.n) {
590
+ let f = srt_ids[i];
591
+ if (f >= 0.0 && f < f32(srt_p.b)) { row = u32(f); }
592
+ }
593
+ srt_keys[2u * i] = row;
594
+ srt_keys[2u * i + 1u] = i;
595
+ }
596
+ case 1u: {
597
+ let l = i ^ srt_p.b;
598
+ if (i >= srt_p.n || l <= i) { return; }
599
+ let ri = srt_keys[2u * i];
600
+ let pi = srt_keys[2u * i + 1u];
601
+ let rl = srt_keys[2u * l];
602
+ let pl = srt_keys[2u * l + 1u];
603
+ let greater = ri > rl || (ri == rl && pi > pl);
604
+ if (greater == ((i & srt_p.a) == 0u)) {
605
+ srt_keys[2u * i] = rl;
606
+ srt_keys[2u * i + 1u] = pl;
607
+ srt_keys[2u * l] = ri;
608
+ srt_keys[2u * l + 1u] = pi;
609
+ }
610
+ }
611
+ case 2u: {
612
+ if (i >= srt_p.n) { return; }
613
+ srt_head[i] = SORT_NONE;
614
+ }
615
+ default: {
616
+ if (i >= srt_p.n) { return; }
617
+ let row = srt_keys[2u * i];
618
+ if (row == SORT_NONE) { return; }
619
+ if (i == 0u || srt_keys[2u * i - 2u] != row) {
620
+ srt_head[row] = i;
621
+ }
622
+ }
623
+ }
624
+ }
625
+
626
+ // The gradient is viewed as [outer, n, row] and the output as
627
+ // [outer, rows, row].
628
+ struct ScatterParams {
629
+ rows: u32,
630
+ row: u32,
631
+ n: u32,
632
+ outer: u32,
633
+ };
634
+
635
+ @group(0) @binding(0) var<storage, read> sct_grad: array<f32>;
636
+ @group(0) @binding(1) var<storage, read> sct_keys: array<u32>;
637
+ @group(0) @binding(2) var<storage, read> sct_head: array<u32>;
638
+ @group(0) @binding(3) var<storage, read_write> sct_out: array<f32>;
639
+ @group(0) @binding(4) var<uniform> sct_p: ScatterParams;
640
+
641
+ @compute @workgroup_size(256)
642
+ fn scatter_rows(@builtin(global_invocation_id) gid: vec3<u32>,
643
+ @builtin(num_workgroups) nwg: vec3<u32>) {
644
+ let i = flat_index(gid, nwg);
645
+ if (i >= sct_p.outer * sct_p.rows * sct_p.row) { return; }
646
+ let o = i / (sct_p.rows * sct_p.row);
647
+ let r = (i / sct_p.row) % sct_p.rows;
648
+ let c = i % sct_p.row;
649
+ let base = o * sct_p.n * sct_p.row + c;
650
+ var acc = 0.0;
651
+ var t = sct_head[r];
652
+ if (t != SORT_NONE) {
653
+ loop {
654
+ if (t >= sct_p.n || sct_keys[2u * t] != r) { break; }
655
+ acc = acc + sct_grad[base + sct_keys[2u * t + 1u] * sct_p.row];
656
+ t = t + 1u;
657
+ }
658
+ }
659
+ sct_out[i] = acc;
660
+ }
661
+
662
+ // --- Block copy (cat) --------------------------------------------------------
663
+ // Copies an input viewed as [outer, block_in] into columns
664
+ // [offset, offset + block_in) of an output viewed as [outer, block_out].
665
+
666
+ struct CopyParams {
667
+ n: u32,
668
+ block_in: u32,
669
+ block_out: u32,
670
+ offset: u32,
671
+ };
672
+
673
+ @group(0) @binding(0) var<storage, read> cpy_in: array<f32>;
674
+ @group(0) @binding(1) var<storage, read_write> cpy_out: array<f32>;
675
+ @group(0) @binding(2) var<uniform> cpy_p: CopyParams;
676
+
677
+ @compute @workgroup_size(256)
678
+ fn copy_block(@builtin(global_invocation_id) gid: vec3<u32>,
679
+ @builtin(num_workgroups) nwg: vec3<u32>) {
680
+ let i = flat_index(gid, nwg);
681
+ if (i >= cpy_p.n) { return; }
682
+ let o = i / cpy_p.block_in;
683
+ cpy_out[o * cpy_p.block_out + cpy_p.offset + i % cpy_p.block_in] = cpy_in[i];
684
+ }
685
+
686
+ // --- Cross-entropy with class-id targets ---------------------------------------
687
+ // Logits are viewed as [rows, classes]; targets hold one f32 class id per row.
688
+ // A target <= -1 (an id that is negative as an integer, such as -100) skips
689
+ // its row: loss 0, valid 0, gradient 0. Otherwise the id is truncated to an
690
+ // integer; one that is not below `classes` (or NaN) gives a NaN loss.
691
+ // cross_entropy_rows: one workgroup per row writes the row's loss
692
+ // (lse - logit[target]), its log-sum-exp and whether it counts.
693
+ // cross_entropy_grad: d loss[r] / d logit[r, c] = softmax[r, c] - [c == target],
694
+ // times the incoming gradient of the row's loss.
695
+
696
+ struct CeParams {
697
+ rows: u32,
698
+ classes: u32,
699
+ _pad0: u32,
700
+ _pad1: u32,
701
+ };
702
+
703
+ @group(0) @binding(0) var<storage, read> ce_logits: array<f32>;
704
+ @group(0) @binding(1) var<storage, read> ce_targets: array<f32>;
705
+ @group(0) @binding(2) var<storage, read_write> ce_loss: array<f32>;
706
+ @group(0) @binding(3) var<storage, read_write> ce_lse: array<f32>;
707
+ @group(0) @binding(4) var<storage, read_write> ce_valid: array<f32>;
708
+ @group(0) @binding(5) var<uniform> ce_p: CeParams;
709
+
710
+ var<workgroup> ce_part: array<f32, 256>;
711
+
712
+ @compute @workgroup_size(256)
713
+ fn cross_entropy_rows(@builtin(local_invocation_index) lid: u32,
714
+ @builtin(workgroup_id) wid: vec3<u32>,
715
+ @builtin(num_workgroups) nwg: vec3<u32>) {
716
+ let r = wid.x + wid.y * nwg.x;
717
+ // Uniform per workgroup, so every thread reaches the barriers or none.
718
+ if (r >= ce_p.rows) { return; }
719
+ let c = ce_p.classes;
720
+ let base = r * c;
721
+ var m = bitcast<f32>(0xff800000u);
722
+ for (var j = lid; j < c; j = j + 256u) {
723
+ m = max(m, ce_logits[base + j]);
724
+ }
725
+ ce_part[lid] = m;
726
+ workgroupBarrier();
727
+ for (var half = 128u; half > 0u; half = half >> 1u) {
728
+ if (lid < half) {
729
+ ce_part[lid] = max(ce_part[lid], ce_part[lid + half]);
730
+ }
731
+ workgroupBarrier();
732
+ }
733
+ m = ce_part[0];
734
+ workgroupBarrier();
735
+ var s = 0.0;
736
+ for (var j = lid; j < c; j = j + 256u) {
737
+ s = s + exp(ce_logits[base + j] - m);
738
+ }
739
+ ce_part[lid] = s;
740
+ workgroupBarrier();
741
+ for (var half = 128u; half > 0u; half = half >> 1u) {
742
+ if (lid < half) {
743
+ ce_part[lid] = ce_part[lid] + ce_part[lid + half];
744
+ }
745
+ workgroupBarrier();
746
+ }
747
+ if (lid == 0u) {
748
+ let lse = m + log(ce_part[0]);
749
+ let f = ce_targets[r];
750
+ ce_lse[r] = lse;
751
+ if (f <= -1.0) {
752
+ ce_loss[r] = 0.0;
753
+ ce_valid[r] = 0.0;
754
+ } else {
755
+ ce_valid[r] = 1.0;
756
+ if (f < f32(c)) {
757
+ ce_loss[r] = lse - ce_logits[base + u32(max(f, 0.0))];
758
+ } else {
759
+ ce_loss[r] = bitcast<f32>(0x7fc00000u);
760
+ }
761
+ }
762
+ }
763
+ }
764
+
765
+ @group(0) @binding(0) var<storage, read> ceg_logits: array<f32>;
766
+ @group(0) @binding(1) var<storage, read> ceg_targets: array<f32>;
767
+ @group(0) @binding(2) var<storage, read> ceg_lse: array<f32>;
768
+ @group(0) @binding(3) var<storage, read> ceg_grad: array<f32>;
769
+ @group(0) @binding(4) var<storage, read_write> ceg_out: array<f32>;
770
+ @group(0) @binding(5) var<uniform> ceg_p: CeParams;
771
+
772
+ @compute @workgroup_size(256)
773
+ fn cross_entropy_grad(@builtin(global_invocation_id) gid: vec3<u32>,
774
+ @builtin(num_workgroups) nwg: vec3<u32>) {
775
+ let i = flat_index(gid, nwg);
776
+ if (i >= ceg_p.rows * ceg_p.classes) { return; }
777
+ let r = i / ceg_p.classes;
778
+ let f = ceg_targets[r];
779
+ if (f <= -1.0) {
780
+ ceg_out[i] = 0.0;
781
+ return;
782
+ }
783
+ let p = exp(ceg_logits[i] - ceg_lse[r]);
784
+ let hit = f < f32(ceg_p.classes) && i % ceg_p.classes == u32(max(f, 0.0));
785
+ ceg_out[i] = (p - select(0.0, 1.0, hit)) * ceg_grad[r];
786
+ }
550
787
  "#;
@@ -300,21 +300,47 @@ targets_from_ids!(i32, usize);
300
300
  /// positions;
301
301
  /// - or a `&Tensor` distribution (one-hot or soft labels) with the shape of
302
302
  /// `logits`.
303
+ ///
304
+ /// With class ids on a GPU, the loss and its gradient are computed on the
305
+ /// device with no copy to the host. An id that is not below the number of
306
+ /// classes panics, except for an id tensor on a GPU, which is not read
307
+ /// back: there it gives a NaN loss.
303
308
  pub fn cross_entropy<'a>(logits: &Tensor, targets: impl Into<Targets<'a>>) -> Tensor {
304
309
  let targets = targets.into();
305
- let log_probs = logits.log_softmax(-1);
306
- if let Targets::Tensor(t) = targets {
307
- if t.shape == logits.shape {
308
- let rows = (logits.numel() / logits.shape[logits.ndim() - 1]).max(1);
309
- return log_probs.mul(t).sum().div_scalar(-(rows as f32));
310
+ let classes = logits.shape[logits.ndim() - 1];
311
+ let rows = logits.numel() / classes.max(1);
312
+ let ids = match targets {
313
+ Targets::Tensor(t) if t.shape == logits.shape => {
314
+ let rows = rows.max(1);
315
+ return logits
316
+ .log_softmax(-1)
317
+ .mul(t)
318
+ .sum()
319
+ .div_scalar(-(rows as f32));
310
320
  }
311
- assert_eq!(
312
- t.shape,
313
- logits.shape[..logits.ndim() - 1],
314
- "class-id targets must have the logits shape without the class dim"
315
- );
316
- }
317
- nll_loss_indices(&log_probs, &targets.ids())
321
+ Targets::Tensor(t) => {
322
+ assert_eq!(
323
+ t.shape,
324
+ logits.shape[..logits.ndim() - 1],
325
+ "class-id targets must have the logits shape without the class dim"
326
+ );
327
+ t.clone()
328
+ }
329
+ Targets::Ids(ids) => {
330
+ assert_eq!(ids.len(), rows, "one target per row of log-probabilities");
331
+ for &t in ids.iter() {
332
+ assert!(
333
+ t < classes as i64,
334
+ "target class {t} out of range for {classes} classes"
335
+ );
336
+ }
337
+ let ids = ids.iter().map(|&t| t.max(-1) as f32).collect();
338
+ Tensor::from_vec(ids, &[rows], logits.device())
339
+ }
340
+ };
341
+ // Mean over the rows that count; 0 when none does.
342
+ let (losses, valid) = logits.cross_entropy_rows(&ids);
343
+ losses.sum().div(&valid.sum().clamp(1.0, f32::INFINITY))
318
344
  }
319
345
 
320
346
  /// Negative log-likelihood of class `targets` (see [`Targets`]) under
@@ -428,6 +454,7 @@ pub fn accuracy<'a>(logits: &Tensor, targets: impl Into<Targets<'a>>) -> f32 {
428
454
  mod tests {
429
455
  use super::*;
430
456
  use crate::tensor::tests::{check_gpu_parity, check_gradients, Case, Init};
457
+ use rand::{Rng, SeedableRng};
431
458
 
432
459
  const U: Init = Init::Uniform(-2.0, 2.0);
433
460
  const PROB: Init = Init::Uniform(0.05, 0.95);
@@ -632,4 +659,72 @@ mod tests {
632
659
  let t = cpu(&[2.0, -1.0, 0.0], &[3]);
633
660
  assert_eq!(one_hot(&t, 3, &d).to_vec(), want);
634
661
  }
662
+
663
+ #[test]
664
+ #[should_panic(expected = "out of range")]
665
+ fn cross_entropy_rejects_out_of_range_tensor_targets_on_the_cpu() {
666
+ cross_entropy(&cpu(&[0.0; 6], &[2, 3]), &cpu(&[1.0, 3.0], &[2]));
667
+ }
668
+
669
+ #[test]
670
+ fn cross_entropy_with_every_row_skipped_is_zero() {
671
+ let logits = cpu(&[1.0, 2.0, 3.0, 4.0], &[2, 2]).requires_grad_();
672
+ let loss = cross_entropy(&logits, &[-100, -100]);
673
+ assert_eq!(loss.item(), 0.0);
674
+ loss.backward();
675
+ assert_eq!(logits.grad().unwrap().to_vec(), vec![0.0; 4]);
676
+ }
677
+
678
+ /// With `FERROTENSOR_RUN_GPU_TESTS=1`: class-id cross-entropy on the GPU
679
+ /// (loss and logits gradient) matches the CPU on wide rows, 3-D logits,
680
+ /// skipped and fractional ids, and gives NaN for an id out of range.
681
+ #[test]
682
+ fn cross_entropy_matches_cpu_on_gpu() {
683
+ if std::env::var("FERROTENSOR_RUN_GPU_TESTS").as_deref() != Ok("1") {
684
+ eprintln!("skipping GPU cross_entropy; set FERROTENSOR_RUN_GPU_TESTS=1 to enable");
685
+ return;
686
+ }
687
+ let gpu = Device::gpu().expect("FERROTENSOR_RUN_GPU_TESTS=1 but no GPU is available");
688
+ let mut rng = rand::rngs::StdRng::seed_from_u64(13);
689
+ for shape in [
690
+ &[1, 1][..],
691
+ &[5, 3],
692
+ &[37, 1000],
693
+ &[2, 3, 300],
694
+ &[64, 32_000],
695
+ ] {
696
+ let classes = shape[shape.len() - 1];
697
+ let rows: usize = shape[..shape.len() - 1].iter().product();
698
+ let logits: Vec<f32> = (0..rows * classes)
699
+ .map(|_| rng.gen_range(-8.0..8.0))
700
+ .collect();
701
+ let mut targets: Vec<f32> = (0..rows).map(|r| ((r * 7919) % classes) as f32).collect();
702
+ if rows > 2 {
703
+ targets[0] = -100.0;
704
+ targets[1] += 0.5;
705
+ targets[2] = -0.5;
706
+ }
707
+ let run = |d: &Device| {
708
+ let x = Tensor::from_data(&logits, shape, d).requires_grad_();
709
+ let t = Tensor::from_data(&targets, &shape[..shape.len() - 1], d);
710
+ let loss = cross_entropy(&x, &t);
711
+ loss.backward();
712
+ (loss.item(), x.grad().unwrap().to_vec())
713
+ };
714
+ let ((gl, gg), (cl, cg)) = (run(&gpu), run(&Device::cpu()));
715
+ assert!(
716
+ (gl - cl).abs() <= 1e-5 * cl.abs().max(1.0),
717
+ "{shape:?}: {gl} vs {cl}"
718
+ );
719
+ let worst = gg
720
+ .iter()
721
+ .zip(&cg)
722
+ .map(|(a, b)| (a - b).abs())
723
+ .fold(0.0, f32::max);
724
+ assert!(worst <= 1e-6, "{shape:?}: gradient differs by {worst}");
725
+ }
726
+ let x = Tensor::from_data(&[0.0; 6], &[2, 3], &gpu);
727
+ let t = Tensor::from_data(&[1.0, 3.0], &[2], &gpu);
728
+ assert!(cross_entropy(&x, &t).item().is_nan());
729
+ }
635
730
  }
@@ -296,6 +296,102 @@ pub(crate) fn gather_rows(table: &[f32], idx: &[u32], row: usize) -> Vec<f32> {
296
296
  out
297
297
  }
298
298
 
299
+ /// The class of a cross-entropy target stored as `f32`: `None` to skip the
300
+ /// row (an id that is negative as an integer, such as -100), else the
301
+ /// truncated id, as in the `cross_entropy_rows` kernel.
302
+ ///
303
+ /// # Panics
304
+ /// Panics when the id is NaN or not below `classes`.
305
+ pub(crate) fn target_class(f: f32, classes: usize) -> Option<usize> {
306
+ if f <= -1.0 {
307
+ return None;
308
+ }
309
+ assert!(
310
+ f < classes as f32,
311
+ "target class {f} out of range for {classes} classes"
312
+ );
313
+ Some(f.max(0.0) as usize)
314
+ }
315
+
316
+ /// Per row of `x` (`[rows, classes]`): the cross-entropy loss against
317
+ /// `targets` (0 for a skipped row), the log-sum-exp, and 1 or 0 for whether
318
+ /// the row counts.
319
+ pub(crate) fn cross_entropy_rows(
320
+ x: &[f32],
321
+ targets: &[f32],
322
+ classes: usize,
323
+ ) -> (Vec<f32>, Vec<f32>, Vec<f32>) {
324
+ let rows = targets.len();
325
+ let (mut loss, mut lse, mut valid) = (vec![0.0; rows], vec![0.0; rows], vec![0.0; rows]);
326
+ for r in 0..rows {
327
+ let row = &x[r * classes..][..classes];
328
+ let m = row.iter().fold(f32::NEG_INFINITY, |a, &b| a.max(b));
329
+ let s: f32 = row.iter().map(|&v| (v - m).exp()).sum();
330
+ lse[r] = m + s.ln();
331
+ if let Some(t) = target_class(targets[r], classes) {
332
+ loss[r] = lse[r] - row[t];
333
+ valid[r] = 1.0;
334
+ }
335
+ }
336
+ (loss, lse, valid)
337
+ }
338
+
339
+ /// Gradient of [`cross_entropy_rows`]'s loss with respect to `x`, given the
340
+ /// gradient `grad` of each row's loss.
341
+ pub(crate) fn cross_entropy_grad(
342
+ x: &[f32],
343
+ targets: &[f32],
344
+ lse: &[f32],
345
+ grad: &[f32],
346
+ classes: usize,
347
+ ) -> Vec<f32> {
348
+ let mut out = vec![0.0; x.len()];
349
+ for (r, &f) in targets.iter().enumerate() {
350
+ if let Some(t) = target_class(f, classes) {
351
+ for (c, o) in out[r * classes..][..classes].iter_mut().enumerate() {
352
+ let p = (x[r * classes + c] - lse[r]).exp();
353
+ *o = (p - if c == t { 1.0 } else { 0.0 }) * grad[r];
354
+ }
355
+ }
356
+ }
357
+ out
358
+ }
359
+
360
+ /// Concatenate `parts` (values, block length), each viewed as
361
+ /// `[outer, block]`, along the block dim.
362
+ pub(crate) fn cat(parts: &[(&[f32], usize)], outer: usize) -> Vec<f32> {
363
+ let mut out = Vec::with_capacity(parts.iter().map(|(d, _)| d.len()).sum());
364
+ for o in 0..outer {
365
+ for &(values, block) in parts {
366
+ out.extend_from_slice(&values[o * block..(o + 1) * block]);
367
+ }
368
+ }
369
+ out
370
+ }
371
+
372
+ /// Scatter-add along the middle dim: `grad` viewed as `[outer, n, row]`,
373
+ /// the result as `[outer, rows, row]`, whose row `r` is the sum, in position
374
+ /// order, of the gradient rows `p` of every pair `(p, r)`.
375
+ pub(crate) fn scatter_rows(
376
+ grad: &[f32],
377
+ pairs: &[(usize, usize)],
378
+ outer: usize,
379
+ rows: usize,
380
+ row: usize,
381
+ ) -> Vec<f32> {
382
+ let n = grad.len() / (outer * row).max(1);
383
+ let mut out = vec![0.0f32; outer * rows * row];
384
+ for o in 0..outer {
385
+ for &(p, r) in pairs {
386
+ let src = &grad[(o * n + p) * row..][..row];
387
+ for (v, &g) in out[(o * rows + r) * row..][..row].iter_mut().zip(src) {
388
+ *v += g;
389
+ }
390
+ }
391
+ }
392
+ out
393
+ }
394
+
299
395
  #[cfg(test)]
300
396
  mod tests {
301
397
  use super::*;