ferrotensor 0.6.0__tar.gz → 0.7.0__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 (85) hide show
  1. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/Cargo.lock +18 -105
  2. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/Cargo.toml +3 -4
  3. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/PKG-INFO +7 -4
  4. {ferrotensor-0.6.0/ferrotensor-py → ferrotensor-0.7.0}/README.md +5 -1
  5. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/Cargo.toml +4 -2
  6. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/examples/README.md +22 -1
  7. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/examples/char_transformer.rs +39 -8
  8. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/examples/gpu_matmul_attention.rs +43 -9
  9. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/examples/mlp_two_moons.rs +23 -2
  10. ferrotensor-0.7.0/ferrotensor/examples/quickstart.rs +36 -0
  11. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/examples/similarity_search.rs +10 -0
  12. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/backend/compute.rs +9 -0
  13. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/backend/kernels.rs +17 -3
  14. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/data.rs +1 -0
  15. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/device.rs +23 -0
  16. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/functional.rs +9 -1
  17. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/lib.rs +2 -4
  18. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/nn/attention.rs +2 -1
  19. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/nn/container.rs +11 -10
  20. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/nn/embedding.rs +2 -1
  21. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/nn/tests.rs +3 -1
  22. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/optim/tests.rs +3 -1
  23. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/serialize.rs +2 -1
  24. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/tensor/autograd.rs +4 -0
  25. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/tensor/core.rs +25 -0
  26. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/tensor/cpu.rs +4 -1
  27. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/tensor/ops.rs +39 -0
  28. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/tensor/raw.rs +15 -0
  29. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/tensor/tests.rs +17 -2
  30. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor-py/Cargo.toml +0 -1
  31. {ferrotensor-0.6.0 → ferrotensor-0.7.0/ferrotensor-py}/README.md +5 -1
  32. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor-py/examples/mlp.py +15 -1
  33. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor-py/src/nn.rs +4 -1
  34. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor-py/tests/conftest.py +8 -3
  35. ferrotensor-0.7.0/ferrotensor-py/tests/test_autograd.py +324 -0
  36. ferrotensor-0.7.0/ferrotensor-py/tests/test_errors.py +265 -0
  37. ferrotensor-0.7.0/ferrotensor-py/tests/test_functional.py +194 -0
  38. ferrotensor-0.7.0/ferrotensor-py/tests/test_gpu.py +632 -0
  39. ferrotensor-0.7.0/ferrotensor-py/tests/test_nn.py +282 -0
  40. ferrotensor-0.7.0/ferrotensor-py/tests/test_optim.py +220 -0
  41. ferrotensor-0.7.0/ferrotensor-py/tests/test_serialize.py +238 -0
  42. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor-py/tests/test_tensor.py +125 -0
  43. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor-py/tests/test_train.py +16 -0
  44. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/pyproject.toml +1 -1
  45. ferrotensor-0.6.0/ferrotensor/examples/quickstart.rs +0 -21
  46. ferrotensor-0.6.0/ferrotensor-py/tests/test_autograd.py +0 -110
  47. ferrotensor-0.6.0/ferrotensor-py/tests/test_errors.py +0 -84
  48. ferrotensor-0.6.0/ferrotensor-py/tests/test_functional.py +0 -75
  49. ferrotensor-0.6.0/ferrotensor-py/tests/test_gpu.py +0 -70
  50. ferrotensor-0.6.0/ferrotensor-py/tests/test_nn.py +0 -144
  51. ferrotensor-0.6.0/ferrotensor-py/tests/test_optim.py +0 -90
  52. ferrotensor-0.6.0/ferrotensor-py/tests/test_serialize.py +0 -60
  53. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/README.md +0 -0
  54. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/examples/data/sonnets.txt +0 -0
  55. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/backend/mod.rs +0 -0
  56. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/nn/activation.rs +0 -0
  57. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/nn/dropout.rs +0 -0
  58. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/nn/linear.rs +0 -0
  59. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/nn/mod.rs +0 -0
  60. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/nn/module.rs +0 -0
  61. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/nn/normalization.rs +0 -0
  62. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/optim/adam.rs +0 -0
  63. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/optim/mod.rs +0 -0
  64. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/optim/optimizer.rs +0 -0
  65. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/optim/scheduler.rs +0 -0
  66. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/optim/sgd.rs +0 -0
  67. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/random.rs +0 -0
  68. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/tensor/mod.rs +0 -0
  69. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/src/tensor/shape.rs +0 -0
  70. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor/tests/manual_seed.rs +0 -0
  71. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor-py/.gitignore +0 -0
  72. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor-py/build.rs +0 -0
  73. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor-py/src/device.rs +0 -0
  74. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor-py/src/error.rs +0 -0
  75. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor-py/src/functional.rs +0 -0
  76. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor-py/src/lib.rs +0 -0
  77. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor-py/src/optim.rs +0 -0
  78. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor-py/src/serialize.rs +0 -0
  79. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/ferrotensor-py/src/tensor.rs +0 -0
  80. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/python/ferrotensor/__init__.py +0 -0
  81. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/python/ferrotensor/_ferrotensor.pyi +0 -0
  82. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/python/ferrotensor/nn/__init__.py +0 -0
  83. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/python/ferrotensor/nn/functional.py +0 -0
  84. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/python/ferrotensor/optim.py +0 -0
  85. {ferrotensor-0.6.0 → ferrotensor-0.7.0}/python/ferrotensor/py.typed +0 -0
@@ -223,19 +223,6 @@ version = "0.1.1"
223
223
  source = "registry+https://github.com/rust-lang/crates.io-index"
224
224
  checksum = "fd16c4719339c4530435d38e511904438d07cce7950afa3718a84ac36c10e89e"
225
225
 
226
- [[package]]
227
- name = "chrono"
228
- version = "0.4.44"
229
- source = "registry+https://github.com/rust-lang/crates.io-index"
230
- checksum = "c673075a2e0e5f4a1dde27ce9dee1ea4558c7ffe648f576438a20ca1d2acc4b0"
231
- dependencies = [
232
- "iana-time-zone",
233
- "js-sys",
234
- "num-traits",
235
- "wasm-bindgen",
236
- "windows-link",
237
- ]
238
-
239
226
  [[package]]
240
227
  name = "clap"
241
228
  version = "4.5.56"
@@ -606,7 +593,7 @@ dependencies = [
606
593
 
607
594
  [[package]]
608
595
  name = "ferrotensor"
609
- version = "0.6.0"
596
+ version = "0.7.0"
610
597
  dependencies = [
611
598
  "bytemuck",
612
599
  "naga",
@@ -619,7 +606,7 @@ dependencies = [
619
606
 
620
607
  [[package]]
621
608
  name = "ferrotensor-autograd"
622
- version = "0.6.0"
609
+ version = "0.7.0"
623
610
  dependencies = [
624
611
  "ferrotensor-backend",
625
612
  "pollster",
@@ -628,7 +615,7 @@ dependencies = [
628
615
 
629
616
  [[package]]
630
617
  name = "ferrotensor-backend"
631
- version = "0.6.0"
618
+ version = "0.7.0"
632
619
  dependencies = [
633
620
  "bytemuck",
634
621
  "env_logger",
@@ -641,7 +628,7 @@ dependencies = [
641
628
 
642
629
  [[package]]
643
630
  name = "ferrotensor-benchmarks"
644
- version = "0.6.0"
631
+ version = "0.7.0"
645
632
  dependencies = [
646
633
  "ferrotensor",
647
634
  "ferrotensor-backend",
@@ -652,40 +639,44 @@ dependencies = [
652
639
 
653
640
  [[package]]
654
641
  name = "ferrotensor-cli"
655
- version = "0.6.0"
642
+ version = "0.7.0"
656
643
  dependencies = [
657
644
  "clap",
658
645
  "env_logger",
659
646
  "ferrotensor-autograd",
660
647
  "ferrotensor-backend",
661
648
  "ferrotensor-graph",
649
+ "ferrotensor-hub",
662
650
  "ferrotensor-loader",
663
651
  "ferrotensor-monitor",
664
652
  "ferrotensor-trainer",
665
653
  "log",
666
654
  "pollster",
655
+ "safetensors",
667
656
  "serde",
668
657
  "serde_json",
669
658
  ]
670
659
 
671
660
  [[package]]
672
661
  name = "ferrotensor-graph"
673
- version = "0.6.0"
662
+ version = "0.7.0"
674
663
  dependencies = [
675
664
  "bytemuck",
676
665
  "env_logger",
677
666
  "ferrotensor-autograd",
678
667
  "ferrotensor-backend",
679
668
  "ferrotensor-loader",
669
+ "half",
680
670
  "log",
681
671
  "pollster",
672
+ "safetensors",
682
673
  "serde",
683
674
  "serde_json",
684
675
  ]
685
676
 
686
677
  [[package]]
687
678
  name = "ferrotensor-hub"
688
- version = "0.6.0"
679
+ version = "0.7.0"
689
680
  dependencies = [
690
681
  "ferrotensor-autograd",
691
682
  "ferrotensor-backend",
@@ -698,7 +689,7 @@ dependencies = [
698
689
 
699
690
  [[package]]
700
691
  name = "ferrotensor-loader"
701
- version = "0.6.0"
692
+ version = "0.7.0"
702
693
  dependencies = [
703
694
  "bytemuck",
704
695
  "env_logger",
@@ -716,10 +707,9 @@ dependencies = [
716
707
 
717
708
  [[package]]
718
709
  name = "ferrotensor-monitor"
719
- version = "0.6.0"
710
+ version = "0.7.0"
720
711
  dependencies = [
721
712
  "anyhow",
722
- "chrono",
723
713
  "crossterm",
724
714
  "ferrotensor-trainer",
725
715
  "rand 0.9.5",
@@ -730,7 +720,7 @@ dependencies = [
730
720
 
731
721
  [[package]]
732
722
  name = "ferrotensor-py"
733
- version = "0.6.0"
723
+ version = "0.7.0"
734
724
  dependencies = [
735
725
  "ferrotensor",
736
726
  "pyo3",
@@ -739,7 +729,7 @@ dependencies = [
739
729
 
740
730
  [[package]]
741
731
  name = "ferrotensor-trainer"
742
- version = "0.6.0"
732
+ version = "0.7.0"
743
733
  dependencies = [
744
734
  "bytemuck",
745
735
  "ferrotensor-autograd",
@@ -995,30 +985,6 @@ version = "0.2.1"
995
985
  source = "registry+https://github.com/rust-lang/crates.io-index"
996
986
  checksum = "dfa686283ad6dd069f105e5ab091b04c62850d3e4cf5d67debad1933f55023df"
997
987
 
998
- [[package]]
999
- name = "iana-time-zone"
1000
- version = "0.1.65"
1001
- source = "registry+https://github.com/rust-lang/crates.io-index"
1002
- checksum = "e31bc9ad994ba00e440a8aa5c9ef0ec67d5cb5e5cb0cc7f8b744a35b389cc470"
1003
- dependencies = [
1004
- "android_system_properties",
1005
- "core-foundation-sys",
1006
- "iana-time-zone-haiku",
1007
- "js-sys",
1008
- "log",
1009
- "wasm-bindgen",
1010
- "windows-core 0.62.2",
1011
- ]
1012
-
1013
- [[package]]
1014
- name = "iana-time-zone-haiku"
1015
- version = "0.1.2"
1016
- source = "registry+https://github.com/rust-lang/crates.io-index"
1017
- checksum = "f31827a206f56af32e590ba56d5d2d085f558508192593743f16b2306495269f"
1018
- dependencies = [
1019
- "cc",
1020
- ]
1021
-
1022
988
  [[package]]
1023
989
  name = "ident_case"
1024
990
  version = "1.0.1"
@@ -2486,25 +2452,12 @@ version = "0.57.0"
2486
2452
  source = "registry+https://github.com/rust-lang/crates.io-index"
2487
2453
  checksum = "d2ed2439a290666cd67ecce2b0ffaad89c2a56b976b736e6ece670297897832d"
2488
2454
  dependencies = [
2489
- "windows-implement 0.57.0",
2490
- "windows-interface 0.57.0",
2491
- "windows-result 0.1.2",
2455
+ "windows-implement",
2456
+ "windows-interface",
2457
+ "windows-result",
2492
2458
  "windows-targets",
2493
2459
  ]
2494
2460
 
2495
- [[package]]
2496
- name = "windows-core"
2497
- version = "0.62.2"
2498
- source = "registry+https://github.com/rust-lang/crates.io-index"
2499
- checksum = "b8e83a14d34d0623b51dce9581199302a221863196a1dde71a7663a4c2be9deb"
2500
- dependencies = [
2501
- "windows-implement 0.60.2",
2502
- "windows-interface 0.59.3",
2503
- "windows-link",
2504
- "windows-result 0.4.1",
2505
- "windows-strings",
2506
- ]
2507
-
2508
2461
  [[package]]
2509
2462
  name = "windows-implement"
2510
2463
  version = "0.57.0"
@@ -2516,17 +2469,6 @@ dependencies = [
2516
2469
  "syn 2.0.114",
2517
2470
  ]
2518
2471
 
2519
- [[package]]
2520
- name = "windows-implement"
2521
- version = "0.60.2"
2522
- source = "registry+https://github.com/rust-lang/crates.io-index"
2523
- checksum = "053e2e040ab57b9dc951b72c264860db7eb3b0200ba345b4e4c3b14f67855ddf"
2524
- dependencies = [
2525
- "proc-macro2",
2526
- "quote",
2527
- "syn 2.0.114",
2528
- ]
2529
-
2530
2472
  [[package]]
2531
2473
  name = "windows-interface"
2532
2474
  version = "0.57.0"
@@ -2538,17 +2480,6 @@ dependencies = [
2538
2480
  "syn 2.0.114",
2539
2481
  ]
2540
2482
 
2541
- [[package]]
2542
- name = "windows-interface"
2543
- version = "0.59.3"
2544
- source = "registry+https://github.com/rust-lang/crates.io-index"
2545
- checksum = "3f316c4a2570ba26bbec722032c4099d8c8bc095efccdc15688708623367e358"
2546
- dependencies = [
2547
- "proc-macro2",
2548
- "quote",
2549
- "syn 2.0.114",
2550
- ]
2551
-
2552
2483
  [[package]]
2553
2484
  name = "windows-link"
2554
2485
  version = "0.2.1"
@@ -2564,24 +2495,6 @@ dependencies = [
2564
2495
  "windows-targets",
2565
2496
  ]
2566
2497
 
2567
- [[package]]
2568
- name = "windows-result"
2569
- version = "0.4.1"
2570
- source = "registry+https://github.com/rust-lang/crates.io-index"
2571
- checksum = "7781fa89eaf60850ac3d2da7af8e5242a5ea78d1a11c49bf2910bb5a73853eb5"
2572
- dependencies = [
2573
- "windows-link",
2574
- ]
2575
-
2576
- [[package]]
2577
- name = "windows-strings"
2578
- version = "0.5.1"
2579
- source = "registry+https://github.com/rust-lang/crates.io-index"
2580
- checksum = "7837d08f69c77cf6b07689544538e017c1bfcf57e34b4c0ff58e6c2cd3b37091"
2581
- dependencies = [
2582
- "windows-link",
2583
- ]
2584
-
2585
2498
  [[package]]
2586
2499
  name = "windows-sys"
2587
2500
  version = "0.59.0"
@@ -4,15 +4,14 @@ resolver = "2"
4
4
 
5
5
  [workspace.package]
6
6
  edition = "2021"
7
- version = "0.6.0"
7
+ version = "0.7.0"
8
8
  license = "MIT OR Apache-2.0"
9
9
  repository = "https://github.com/puranikyashaswin/ferrotensor"
10
- authors = ["Puranik Yashaswin Sharma <puranikyashaswinsharma@gmail.com>"]
10
+ authors = ["Yashaswin Sharma Puranik <puranikyashaswinsharma@gmail.com>"]
11
11
  rust-version = "1.88"
12
- homepage = "https://github.com/puranikyashaswin/ferrotensor"
13
12
 
14
13
  [workspace.lints.rust]
15
- unsafe_code = "forbid"
14
+ unsafe_code = "deny"
16
15
  missing_docs = "warn"
17
16
  rust_2018_idioms = { level = "warn", priority = -1 }
18
17
 
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: ferrotensor
3
- Version: 0.6.0
3
+ Version: 0.7.0
4
4
  Classifier: Programming Language :: Rust
5
5
  Classifier: Programming Language :: Python :: 3
6
6
  Classifier: Programming Language :: Python :: Implementation :: CPython
@@ -9,8 +9,7 @@ Requires-Dist: pytest>=7 ; extra == 'test'
9
9
  Provides-Extra: test
10
10
  Summary: PyTorch-style tensors, autograd and nn on wgpu (Metal, Vulkan, DX12) or the CPU, written in Rust
11
11
  Keywords: machine-learning,deep-learning,gpu,autograd,wgpu
12
- Home-Page: https://github.com/puranikyashaswin/ferrotensor
13
- Author-email: Puranik Yashaswin Sharma <puranikyashaswinsharma@gmail.com>
12
+ Author-email: Yashaswin Sharma Puranik <puranikyashaswinsharma@gmail.com>
14
13
  License-Expression: MIT OR Apache-2.0
15
14
  Requires-Python: >=3.10
16
15
  Description-Content-Type: text/markdown; charset=UTF-8; variant=GFM
@@ -65,6 +64,7 @@ y.backward()
65
64
  print(x.grad) # tensor([[7.0, 11.0], [9.0, 13.0]])
66
65
 
67
66
  # A model, a loss and an optimizer: learn XOR
67
+ ft.manual_seed(0) # fixes the initial weights, so the output below repeats
68
68
  model = nn.Sequential(nn.Linear(2, 16), nn.Tanh(), nn.Linear(16, 2))
69
69
  opt = optim.AdamW(model.parameters(), lr=0.05, weight_decay=0.0)
70
70
  inputs = ft.tensor([[0, 0], [0, 1], [1, 0], [1, 1]])
@@ -155,9 +155,12 @@ naming the shapes involved; the Rust code never panics across the boundary.
155
155
 
156
156
  ```bash
157
157
  pytest # CPU
158
- FERROTENSOR_GPU_TESTS=1 pytest # also GPU parity tests
158
+ FERROTENSOR_RUN_GPU_TESTS=1 pytest # also GPU parity tests
159
159
  ```
160
160
 
161
+ `FERROTENSOR_GPU_TESTS=1` still works as the older name. CI runs the GPU
162
+ tests on software Vulkan (Mesa lavapipe).
163
+
161
164
  ## License
162
165
 
163
166
  MIT or Apache-2.0, at your option.
@@ -46,6 +46,7 @@ y.backward()
46
46
  print(x.grad) # tensor([[7.0, 11.0], [9.0, 13.0]])
47
47
 
48
48
  # A model, a loss and an optimizer: learn XOR
49
+ ft.manual_seed(0) # fixes the initial weights, so the output below repeats
49
50
  model = nn.Sequential(nn.Linear(2, 16), nn.Tanh(), nn.Linear(16, 2))
50
51
  opt = optim.AdamW(model.parameters(), lr=0.05, weight_decay=0.0)
51
52
  inputs = ft.tensor([[0, 0], [0, 1], [1, 0], [1, 1]])
@@ -136,9 +137,12 @@ naming the shapes involved; the Rust code never panics across the boundary.
136
137
 
137
138
  ```bash
138
139
  pytest # CPU
139
- FERROTENSOR_GPU_TESTS=1 pytest # also GPU parity tests
140
+ FERROTENSOR_RUN_GPU_TESTS=1 pytest # also GPU parity tests
140
141
  ```
141
142
 
143
+ `FERROTENSOR_GPU_TESTS=1` still works as the older name. CI runs the GPU
144
+ tests on software Vulkan (Mesa lavapipe).
145
+
142
146
  ## License
143
147
 
144
148
  MIT or Apache-2.0, at your option.
@@ -1,6 +1,6 @@
1
1
  [package]
2
2
  name = "ferrotensor"
3
- version = "0.6.0"
3
+ version = "0.7.0"
4
4
  edition = "2021"
5
5
  authors.workspace = true
6
6
  license.workspace = true
@@ -8,7 +8,6 @@ repository.workspace = true
8
8
  rust-version.workspace = true
9
9
  description = "GPU-accelerated machine learning library for training and inference in Rust"
10
10
  documentation = "https://docs.rs/ferrotensor"
11
- homepage = "https://github.com/puranikyashaswin/ferrotensor"
12
11
  readme = "README.md"
13
12
  keywords = ["machine-learning", "deep-learning", "gpu", "neural-networks", "transformer"]
14
13
  categories = ["science", "algorithms"]
@@ -30,3 +29,6 @@ rand = "0.8"
30
29
 
31
30
  [dev-dependencies]
32
31
  naga = { version = "0.19", features = ["wgsl-in"] }
32
+
33
+ [lints]
34
+ workspace = true
@@ -23,14 +23,35 @@ The low-level examples for the LLM stack live next to the crates they use:
23
23
  creation, tensor upload, and individual kernels through `ComputeEngine`
24
24
  - [`ferrotensor-graph/examples`](../../ferrotensor-graph/examples): a Llama
25
25
  block and causal flash attention checked against a CPU reference
26
+ - [`ferrotensor-trainer/examples`](../../ferrotensor-trainer/examples): the
27
+ training stack on a tiny random model built in a temporary directory, so
28
+ nothing is downloaded. Each needs a GPU, checks its own results and runs
29
+ in CI:
30
+
31
+ | Example | What it shows |
32
+ |---------|---------------|
33
+ | `lora_finetune_tiny` | LoRA training; the loss falls, and a saved adapter reloads into a fresh model with the same loss |
34
+ | `resume_training` | A run stopped halfway, saved with its optimizer state and resumed, matches an unbroken run |
35
+ | `generate_with_kv_cache` | Greedy generation gives the same tokens with and without the KV cache |
36
+
37
+ Run one with `cargo run --release -p ferrotensor-trainer --example <name>`.
26
38
 
27
39
  For fine-tuning a real model, see the
28
40
  [LoRA walkthrough](../../docs/examples/lora-sft.md).
29
41
 
42
+ ## Running every example
43
+
44
+ CI runs every example, on the CPU or on software Vulkan, and fails when one
45
+ fails: the examples check their own results with `assert!`. To do the same
46
+ locally, run `scripts/run-examples.sh` (all of them, needs a GPU adapter),
47
+ `scripts/run-examples.sh --cpu` or `scripts/run-examples.sh --gpu`.
48
+
30
49
  ## Adding an example
31
50
 
32
51
  Create a `.rs` file in this directory; Cargo finds it. Put a doc comment at
33
- the top with the command to run it and its real output.
52
+ the top with the command to run it and its real output, check the result with
53
+ `assert!`, and add it to the list in `scripts/run-examples.sh` (the script
54
+ fails when an example is missing from it).
34
55
 
35
56
  ## Troubleshooting
36
57
 
@@ -9,6 +9,13 @@
9
9
  //! Run with: cargo run --release -p ferrotensor --example char_transformer
10
10
  //! Set FERROTENSOR_DEVICE=gpu to run on the GPU (the default is the CPU, which is
11
11
  //! faster for a model this small).
12
+ //! Set FERROTENSOR_STEPS=<n> to train for n steps instead of 600 (CI uses this
13
+ //! to keep the run short).
14
+ //!
15
+ //! The example exits with an error unless the loss fell during training and
16
+ //! the cached forward pass matches the full one.
17
+
18
+ #![allow(clippy::expect_used)]
12
19
 
13
20
  use std::collections::BTreeMap;
14
21
 
@@ -23,7 +30,7 @@ const BATCH: usize = 16;
23
30
  const DIM: usize = 64;
24
31
  const HEADS: usize = 4;
25
32
  const LAYERS: usize = 2;
26
- const STEPS: usize = 600;
33
+ const DEFAULT_STEPS: usize = 600;
27
34
 
28
35
  /// Pre-norm transformer block: `x + attn(ln(x))`, then `x + mlp(ln(x))`.
29
36
  struct Block {
@@ -221,6 +228,15 @@ fn main() {
221
228
  vocab.len()
222
229
  );
223
230
 
231
+ let steps = match std::env::var("FERROTENSOR_STEPS") {
232
+ Ok(s) => s
233
+ .parse::<usize>()
234
+ .ok()
235
+ .filter(|&n| n > 0)
236
+ .expect("FERROTENSOR_STEPS must be a positive integer"),
237
+ Err(_) => DEFAULT_STEPS,
238
+ };
239
+
224
240
  let model = CharLm::new(vocab.len(), &device);
225
241
  let n_params: usize = model.parameters().iter().map(|p| p.numel()).sum();
226
242
  println!("model: {LAYERS} layers, dim {DIM}, {HEADS} heads, {n_params} parameters");
@@ -228,11 +244,13 @@ fn main() {
228
244
  let mut opt = optim::AdamW::new(model.parameters(), 3e-3)
229
245
  .weight_decay(0.01)
230
246
  .clip_grad_norm(1.0);
231
- let schedule = WarmupCosineLr::new(3e-3, 3e-4, 30, STEPS);
247
+ let schedule = WarmupCosineLr::new(3e-3, 3e-4, 30, steps);
232
248
  let mut rng = StdRng::seed_from_u64(42);
233
249
 
234
250
  let start = std::time::Instant::now();
235
- for step in 0..STEPS {
251
+ let mut first_loss = f32::NAN;
252
+ let mut last_loss = f32::NAN;
253
+ for step in 0..steps {
236
254
  schedule.apply(&mut opt, step);
237
255
  let (mut xs, mut ys) = (Vec::new(), Vec::new());
238
256
  for _ in 0..BATCH {
@@ -246,20 +264,33 @@ fn main() {
246
264
  let loss = F::cross_entropy(&logits.reshape(&[BATCH * CONTEXT, vocab.len()]), &targets);
247
265
  loss.backward();
248
266
  opt.step();
249
- if step % 100 == 0 || step == STEPS - 1 {
250
- println!("step {step:>3} loss {:.3}", loss.item());
267
+ last_loss = loss.item();
268
+ if step == 0 {
269
+ first_loss = last_loss;
270
+ }
271
+ if step % 100 == 0 || step == steps - 1 {
272
+ println!("step {step:>3} loss {last_loss:.3}");
251
273
  }
252
274
  }
253
275
  println!(
254
- "trained {STEPS} steps in {:.1}s",
276
+ "trained {steps} steps in {:.1}s",
255
277
  start.elapsed().as_secs_f32()
256
278
  );
279
+ // Losses are from single random batches, so compare with a wide margin.
280
+ assert!(
281
+ last_loss < 0.8 * first_loss,
282
+ "the loss did not fall: {first_loss} at step 0, {last_loss} at the end"
283
+ );
257
284
 
258
285
  let prompt = "Shall I compare";
259
286
  let ids: Vec<usize> = prompt.chars().map(|c| vocab[&c]).collect();
287
+ let cache_error = model.cache_error(&data[..CONTEXT]);
260
288
  println!(
261
- "KV cache vs full forward on {CONTEXT} characters: max logit difference {:.1e}",
262
- model.cache_error(&data[..CONTEXT])
289
+ "KV cache vs full forward on {CONTEXT} characters: max logit difference {cache_error:.1e}"
290
+ );
291
+ assert!(
292
+ cache_error < 1e-3,
293
+ "the KV cache changes the logits by {cache_error}"
263
294
  );
264
295
  let start = std::time::Instant::now();
265
296
  let out = model.generate(&ids, 200, 0.8, &mut rng);
@@ -6,6 +6,13 @@
6
6
  //! benchmark; see `benchmarks/` for the real harness.
7
7
  //!
8
8
  //! Run with: cargo run --release -p ferrotensor --example gpu_matmul_attention
9
+ //!
10
+ //! The example exits with an error when the CPU and GPU results differ by more
11
+ //! than 1e-3. Two environment variables are for CI, where the GPU is software
12
+ //! Vulkan on the CPU:
13
+ //!
14
+ //! - FERROTENSOR_EXAMPLE_QUICK=1 runs only the smaller sizes.
15
+ //! - FERROTENSOR_REQUIRE_GPU=1 makes a missing GPU an error instead of a skip.
9
16
 
10
17
  use std::time::Instant;
11
18
 
@@ -15,7 +22,15 @@ fn max_abs_diff(a: &[f32], b: &[f32]) -> f32 {
15
22
  a.iter()
16
23
  .zip(b)
17
24
  .map(|(x, y)| (x - y).abs())
18
- .fold(0.0, f32::max)
25
+ // A NaN difference must stay NaN so the check below fails on it.
26
+ .fold(0.0, |m, d| if d.is_nan() || d > m { d } else { m })
27
+ }
28
+
29
+ /// Largest allowed difference between the CPU and GPU results.
30
+ const TOLERANCE: f32 = 1e-3;
31
+
32
+ fn env_set(name: &str) -> bool {
33
+ std::env::var(name).as_deref() == Ok("1")
19
34
  }
20
35
 
21
36
  /// Best-of-`runs` wall time in milliseconds, after one warm-up call.
@@ -32,9 +47,14 @@ fn time_ms(runs: usize, mut f: impl FnMut() -> Vec<f32>) -> (f64, Vec<f32>) {
32
47
 
33
48
  fn main() {
34
49
  let Some(gpu) = Device::gpu() else {
50
+ assert!(
51
+ !env_set("FERROTENSOR_REQUIRE_GPU"),
52
+ "FERROTENSOR_REQUIRE_GPU=1 but no GPU adapter is available"
53
+ );
35
54
  println!("no GPU adapter available; nothing to compare");
36
55
  return;
37
56
  };
57
+ let quick = env_set("FERROTENSOR_EXAMPLE_QUICK");
38
58
  let cpu = Device::cpu();
39
59
  println!("gpu: {} ({})\n", gpu.name(), gpu.backend());
40
60
 
@@ -43,16 +63,25 @@ fn main() {
43
63
  "{:>6} {:>10} {:>10} {:>9} {:>12}",
44
64
  "n", "cpu ms", "gpu ms", "speedup", "max |diff|"
45
65
  );
46
- for n in [256usize, 512, 1024, 2048] {
66
+ let matmul_sizes: &[usize] = if quick {
67
+ &[256, 512]
68
+ } else {
69
+ &[256, 512, 1024, 2048]
70
+ };
71
+ for &n in matmul_sizes {
47
72
  let a = Tensor::randn(&[n, n], &cpu);
48
73
  let b = Tensor::randn(&[n, n], &cpu);
49
74
  let (ga, gb) = (a.to_device(&gpu), b.to_device(&gpu));
50
75
  let (t_cpu, r_cpu) = time_ms(2, || a.matmul(&b).to_vec());
51
76
  let (t_gpu, r_gpu) = time_ms(5, || ga.matmul(&gb).to_vec());
77
+ let diff = max_abs_diff(&r_cpu, &r_gpu);
52
78
  println!(
53
- "{n:>6} {t_cpu:>10.1} {t_gpu:>10.1} {:>8.1}x {:>12.2e}",
54
- t_cpu / t_gpu,
55
- max_abs_diff(&r_cpu, &r_gpu)
79
+ "{n:>6} {t_cpu:>10.1} {t_gpu:>10.1} {:>8.1}x {diff:>12.2e}",
80
+ t_cpu / t_gpu
81
+ );
82
+ assert!(
83
+ diff < TOLERANCE,
84
+ "matmul n={n}: CPU and GPU differ by {diff}"
56
85
  );
57
86
  }
58
87
 
@@ -61,7 +90,8 @@ fn main() {
61
90
  "{:>6} {:>10} {:>10} {:>9} {:>12}",
62
91
  "seq", "cpu ms", "gpu ms", "speedup", "max |diff|"
63
92
  );
64
- for seq in [128usize, 512] {
93
+ let attention_lengths: &[usize] = if quick { &[128] } else { &[128, 512] };
94
+ for &seq in attention_lengths {
65
95
  let shape = [1, 8, seq, 64];
66
96
  let (q, k, v) = (
67
97
  Tensor::randn(&shape, &cpu),
@@ -75,10 +105,14 @@ fn main() {
75
105
  let (t_gpu, r_gpu) = time_ms(5, || {
76
106
  F::scaled_dot_product_attention(&gq, &gk, &gv, true).to_vec()
77
107
  });
108
+ let diff = max_abs_diff(&r_cpu, &r_gpu);
78
109
  println!(
79
- "{seq:>6} {t_cpu:>10.1} {t_gpu:>10.1} {:>8.1}x {:>12.2e}",
80
- t_cpu / t_gpu,
81
- max_abs_diff(&r_cpu, &r_gpu)
110
+ "{seq:>6} {t_cpu:>10.1} {t_gpu:>10.1} {:>8.1}x {diff:>12.2e}",
111
+ t_cpu / t_gpu
112
+ );
113
+ assert!(
114
+ diff < TOLERANCE,
115
+ "attention seq={seq}: CPU and GPU differ by {diff}"
82
116
  );
83
117
  }
84
118
  }
@@ -2,6 +2,9 @@
2
2
  //!
3
3
  //! Run with: cargo run --release -p ferrotensor --example mlp_two_moons
4
4
  //! Set FERROTENSOR_DEVICE=gpu to train on the GPU instead of the CPU.
5
+ //!
6
+ //! The example exits with an error unless the loss fell during training and the
7
+ //! model classifies the held-out points well.
5
8
 
6
9
  use ferrotensor::data::{DataLoader, Sample, TensorDataset};
7
10
  use ferrotensor::optim::{LrScheduler, WarmupCosineLr};
@@ -74,6 +77,9 @@ fn main() {
74
77
  let tx = Tensor::from_data(&flat, &[test_x.len(), 2], &device);
75
78
 
76
79
  let mut step = 0;
80
+ let mut first_loss = f32::NAN;
81
+ let mut last_loss = f32::NAN;
82
+ let mut final_acc = 0.0f32;
77
83
  for epoch in 0..epochs {
78
84
  let mut total = 0.0;
79
85
  for (x, y) in loader.epoch(epoch as u64) {
@@ -85,13 +91,28 @@ fn main() {
85
91
  opt.step();
86
92
  total += loss.item();
87
93
  }
94
+ let mean_loss = total / loader.num_batches() as f32;
95
+ if epoch == 0 {
96
+ first_loss = mean_loss;
97
+ }
98
+ last_loss = mean_loss;
88
99
  if epoch % 10 == 0 || epoch == epochs - 1 {
89
100
  let acc = F::accuracy(&model.forward(&tx), &test_y);
101
+ final_acc = acc;
90
102
  println!(
91
- "epoch {epoch:>2} train loss {:.4} test accuracy {:.1}%",
92
- total / loader.num_batches() as f32,
103
+ "epoch {epoch:>2} train loss {mean_loss:.4} test accuracy {:.1}%",
93
104
  acc * 100.0
94
105
  );
95
106
  }
96
107
  }
108
+
109
+ assert!(
110
+ last_loss < first_loss,
111
+ "the loss did not fall: {first_loss} at epoch 0, {last_loss} at the end"
112
+ );
113
+ assert!(
114
+ final_acc > 0.9,
115
+ "test accuracy is only {:.1}%",
116
+ final_acc * 100.0
117
+ );
97
118
  }
@@ -0,0 +1,36 @@
1
+ //! Quickstart: a tensor, a matmul, a gradient.
2
+ //!
3
+ //! Run with: cargo run --release -p ferrotensor --example quickstart
4
+
5
+ #![allow(clippy::expect_used)]
6
+
7
+ use ferrotensor::prelude::*;
8
+
9
+ /// Panics unless `got` matches `want` element by element.
10
+ fn assert_close(got: &[f32], want: &[f32]) {
11
+ assert_eq!(got.len(), want.len(), "got {got:?}, expected {want:?}");
12
+ for (g, w) in got.iter().zip(want) {
13
+ assert!((g - w).abs() < 1e-4, "got {got:?}, expected {want:?}");
14
+ }
15
+ }
16
+
17
+ fn main() {
18
+ // A GPU (Metal/Vulkan/DX12) when one is available, else the CPU.
19
+ let device = Device::best();
20
+ println!("device: {} ({})", device.name(), device.backend());
21
+
22
+ // y = (x @ w).sum(), so dy/dw = x^T @ ones.
23
+ let x = Tensor::from_data(&[1., 2., 3., 4., 5., 6.], &[2, 3], &device);
24
+ let w = Tensor::from_data(&[0.5, -1., 2., 0., 1., 1.], &[3, 2], &device).requires_grad_();
25
+
26
+ let y = x.matmul(&w);
27
+ let product = y.to_vec();
28
+ println!("x @ w = {:?} (shape {:?})", product, y.shape);
29
+ assert_close(&product, &[7.5, 2., 18., 2.]);
30
+
31
+ y.sum().backward();
32
+ let grad = w.grad().expect("w requires grad").to_vec();
33
+ println!("dL/dw = {:?}", grad);
34
+ // Row k of dL/dw is the sum of column k of x, repeated for both outputs.
35
+ assert_close(&grad, &[5., 5., 7., 7., 9., 9.]);
36
+ }