ferrotensor 0.5.0__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.
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/Cargo.lock +25 -118
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/Cargo.toml +3 -4
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/PKG-INFO +7 -4
- {ferrotensor-0.5.0/ferrotensor-py → ferrotensor-0.7.0}/README.md +5 -1
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/Cargo.toml +4 -2
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/examples/README.md +22 -1
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/examples/char_transformer.rs +39 -8
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/examples/gpu_matmul_attention.rs +43 -9
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/examples/mlp_two_moons.rs +23 -2
- ferrotensor-0.7.0/ferrotensor/examples/quickstart.rs +36 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/examples/similarity_search.rs +10 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/backend/compute.rs +9 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/backend/kernels.rs +17 -3
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/data.rs +1 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/device.rs +23 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/functional.rs +9 -1
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/lib.rs +2 -4
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/nn/attention.rs +2 -1
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/nn/container.rs +11 -10
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/nn/embedding.rs +2 -1
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/nn/tests.rs +3 -1
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/optim/tests.rs +3 -1
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/serialize.rs +2 -1
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/tensor/autograd.rs +4 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/tensor/core.rs +25 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/tensor/cpu.rs +4 -1
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/tensor/ops.rs +39 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/tensor/raw.rs +15 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/tensor/tests.rs +17 -2
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor-py/Cargo.toml +0 -1
- {ferrotensor-0.5.0 → ferrotensor-0.7.0/ferrotensor-py}/README.md +5 -1
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor-py/examples/mlp.py +15 -1
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor-py/src/nn.rs +4 -1
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor-py/tests/conftest.py +8 -3
- ferrotensor-0.7.0/ferrotensor-py/tests/test_autograd.py +324 -0
- ferrotensor-0.7.0/ferrotensor-py/tests/test_errors.py +265 -0
- ferrotensor-0.7.0/ferrotensor-py/tests/test_functional.py +194 -0
- ferrotensor-0.7.0/ferrotensor-py/tests/test_gpu.py +632 -0
- ferrotensor-0.7.0/ferrotensor-py/tests/test_nn.py +282 -0
- ferrotensor-0.7.0/ferrotensor-py/tests/test_optim.py +220 -0
- ferrotensor-0.7.0/ferrotensor-py/tests/test_serialize.py +238 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor-py/tests/test_tensor.py +125 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor-py/tests/test_train.py +16 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/pyproject.toml +1 -1
- ferrotensor-0.5.0/ferrotensor/examples/quickstart.rs +0 -21
- ferrotensor-0.5.0/ferrotensor-py/tests/test_autograd.py +0 -110
- ferrotensor-0.5.0/ferrotensor-py/tests/test_errors.py +0 -84
- ferrotensor-0.5.0/ferrotensor-py/tests/test_functional.py +0 -75
- ferrotensor-0.5.0/ferrotensor-py/tests/test_gpu.py +0 -70
- ferrotensor-0.5.0/ferrotensor-py/tests/test_nn.py +0 -144
- ferrotensor-0.5.0/ferrotensor-py/tests/test_optim.py +0 -90
- ferrotensor-0.5.0/ferrotensor-py/tests/test_serialize.py +0 -60
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/README.md +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/examples/data/sonnets.txt +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/backend/mod.rs +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/nn/activation.rs +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/nn/dropout.rs +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/nn/linear.rs +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/nn/mod.rs +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/nn/module.rs +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/nn/normalization.rs +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/optim/adam.rs +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/optim/mod.rs +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/optim/optimizer.rs +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/optim/scheduler.rs +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/optim/sgd.rs +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/random.rs +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/tensor/mod.rs +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/src/tensor/shape.rs +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor/tests/manual_seed.rs +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor-py/.gitignore +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor-py/build.rs +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor-py/src/device.rs +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor-py/src/error.rs +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor-py/src/functional.rs +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor-py/src/lib.rs +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor-py/src/optim.rs +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor-py/src/serialize.rs +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/ferrotensor-py/src/tensor.rs +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/python/ferrotensor/__init__.py +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/python/ferrotensor/_ferrotensor.pyi +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/python/ferrotensor/nn/__init__.py +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/python/ferrotensor/nn/functional.py +0 -0
- {ferrotensor-0.5.0 → ferrotensor-0.7.0}/python/ferrotensor/optim.py +0 -0
- {ferrotensor-0.5.0 → 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.
|
|
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.
|
|
609
|
+
version = "0.7.0"
|
|
623
610
|
dependencies = [
|
|
624
611
|
"ferrotensor-backend",
|
|
625
612
|
"pollster",
|
|
@@ -628,10 +615,11 @@ dependencies = [
|
|
|
628
615
|
|
|
629
616
|
[[package]]
|
|
630
617
|
name = "ferrotensor-backend"
|
|
631
|
-
version = "0.
|
|
618
|
+
version = "0.7.0"
|
|
632
619
|
dependencies = [
|
|
633
620
|
"bytemuck",
|
|
634
621
|
"env_logger",
|
|
622
|
+
"half",
|
|
635
623
|
"log",
|
|
636
624
|
"naga",
|
|
637
625
|
"pollster",
|
|
@@ -640,7 +628,7 @@ dependencies = [
|
|
|
640
628
|
|
|
641
629
|
[[package]]
|
|
642
630
|
name = "ferrotensor-benchmarks"
|
|
643
|
-
version = "0.
|
|
631
|
+
version = "0.7.0"
|
|
644
632
|
dependencies = [
|
|
645
633
|
"ferrotensor",
|
|
646
634
|
"ferrotensor-backend",
|
|
@@ -651,40 +639,44 @@ dependencies = [
|
|
|
651
639
|
|
|
652
640
|
[[package]]
|
|
653
641
|
name = "ferrotensor-cli"
|
|
654
|
-
version = "0.
|
|
642
|
+
version = "0.7.0"
|
|
655
643
|
dependencies = [
|
|
656
644
|
"clap",
|
|
657
645
|
"env_logger",
|
|
658
646
|
"ferrotensor-autograd",
|
|
659
647
|
"ferrotensor-backend",
|
|
660
648
|
"ferrotensor-graph",
|
|
649
|
+
"ferrotensor-hub",
|
|
661
650
|
"ferrotensor-loader",
|
|
662
651
|
"ferrotensor-monitor",
|
|
663
652
|
"ferrotensor-trainer",
|
|
664
653
|
"log",
|
|
665
654
|
"pollster",
|
|
655
|
+
"safetensors",
|
|
666
656
|
"serde",
|
|
667
657
|
"serde_json",
|
|
668
658
|
]
|
|
669
659
|
|
|
670
660
|
[[package]]
|
|
671
661
|
name = "ferrotensor-graph"
|
|
672
|
-
version = "0.
|
|
662
|
+
version = "0.7.0"
|
|
673
663
|
dependencies = [
|
|
674
664
|
"bytemuck",
|
|
675
665
|
"env_logger",
|
|
676
666
|
"ferrotensor-autograd",
|
|
677
667
|
"ferrotensor-backend",
|
|
678
668
|
"ferrotensor-loader",
|
|
669
|
+
"half",
|
|
679
670
|
"log",
|
|
680
671
|
"pollster",
|
|
672
|
+
"safetensors",
|
|
681
673
|
"serde",
|
|
682
674
|
"serde_json",
|
|
683
675
|
]
|
|
684
676
|
|
|
685
677
|
[[package]]
|
|
686
678
|
name = "ferrotensor-hub"
|
|
687
|
-
version = "0.
|
|
679
|
+
version = "0.7.0"
|
|
688
680
|
dependencies = [
|
|
689
681
|
"ferrotensor-autograd",
|
|
690
682
|
"ferrotensor-backend",
|
|
@@ -697,7 +689,7 @@ dependencies = [
|
|
|
697
689
|
|
|
698
690
|
[[package]]
|
|
699
691
|
name = "ferrotensor-loader"
|
|
700
|
-
version = "0.
|
|
692
|
+
version = "0.7.0"
|
|
701
693
|
dependencies = [
|
|
702
694
|
"bytemuck",
|
|
703
695
|
"env_logger",
|
|
@@ -715,10 +707,9 @@ dependencies = [
|
|
|
715
707
|
|
|
716
708
|
[[package]]
|
|
717
709
|
name = "ferrotensor-monitor"
|
|
718
|
-
version = "0.
|
|
710
|
+
version = "0.7.0"
|
|
719
711
|
dependencies = [
|
|
720
712
|
"anyhow",
|
|
721
|
-
"chrono",
|
|
722
713
|
"crossterm",
|
|
723
714
|
"ferrotensor-trainer",
|
|
724
715
|
"rand 0.9.5",
|
|
@@ -729,7 +720,7 @@ dependencies = [
|
|
|
729
720
|
|
|
730
721
|
[[package]]
|
|
731
722
|
name = "ferrotensor-py"
|
|
732
|
-
version = "0.
|
|
723
|
+
version = "0.7.0"
|
|
733
724
|
dependencies = [
|
|
734
725
|
"ferrotensor",
|
|
735
726
|
"pyo3",
|
|
@@ -738,7 +729,7 @@ dependencies = [
|
|
|
738
729
|
|
|
739
730
|
[[package]]
|
|
740
731
|
name = "ferrotensor-trainer"
|
|
741
|
-
version = "0.
|
|
732
|
+
version = "0.7.0"
|
|
742
733
|
dependencies = [
|
|
743
734
|
"bytemuck",
|
|
744
735
|
"ferrotensor-autograd",
|
|
@@ -800,26 +791,25 @@ checksum = "aa9a19cbb55df58761df49b23516a86d432839add4af60fc256da840f66ed35b"
|
|
|
800
791
|
|
|
801
792
|
[[package]]
|
|
802
793
|
name = "futures-core"
|
|
803
|
-
version = "0.3.
|
|
794
|
+
version = "0.3.34"
|
|
804
795
|
source = "registry+https://github.com/rust-lang/crates.io-index"
|
|
805
|
-
checksum = "
|
|
796
|
+
checksum = "92d699e522242e69e3003b94ecc1f960f3a5e015aa7c5d7486e65ad01dd94f5e"
|
|
806
797
|
|
|
807
798
|
[[package]]
|
|
808
799
|
name = "futures-task"
|
|
809
|
-
version = "0.3.
|
|
800
|
+
version = "0.3.34"
|
|
810
801
|
source = "registry+https://github.com/rust-lang/crates.io-index"
|
|
811
|
-
checksum = "
|
|
802
|
+
checksum = "cd417de3d1d015fc3bfd2b1ea46dfc7bab72ef86f1cc7cc9c78e728b34a6d1fd"
|
|
812
803
|
|
|
813
804
|
[[package]]
|
|
814
805
|
name = "futures-util"
|
|
815
|
-
version = "0.3.
|
|
806
|
+
version = "0.3.34"
|
|
816
807
|
source = "registry+https://github.com/rust-lang/crates.io-index"
|
|
817
|
-
checksum = "
|
|
808
|
+
checksum = "0d50a92467f8ba5dd6e3ee5d4bd04d73ab2e4e1c44474a0674821dfce14b79bc"
|
|
818
809
|
dependencies = [
|
|
819
810
|
"futures-core",
|
|
820
811
|
"futures-task",
|
|
821
812
|
"pin-project-lite",
|
|
822
|
-
"pin-utils",
|
|
823
813
|
"slab",
|
|
824
814
|
]
|
|
825
815
|
|
|
@@ -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"
|
|
@@ -1474,12 +1440,6 @@ version = "0.2.16"
|
|
|
1474
1440
|
source = "registry+https://github.com/rust-lang/crates.io-index"
|
|
1475
1441
|
checksum = "3b3cff922bd51709b605d9ead9aa71031d81447142d828eb4a6eba76fe619f9b"
|
|
1476
1442
|
|
|
1477
|
-
[[package]]
|
|
1478
|
-
name = "pin-utils"
|
|
1479
|
-
version = "0.1.0"
|
|
1480
|
-
source = "registry+https://github.com/rust-lang/crates.io-index"
|
|
1481
|
-
checksum = "8b870d8c151b6f2fb93e84a13146138f05d02ed11c7e7c54f8826aaaf7c9f184"
|
|
1482
|
-
|
|
1483
1443
|
[[package]]
|
|
1484
1444
|
name = "pkg-config"
|
|
1485
1445
|
version = "0.3.32"
|
|
@@ -2492,25 +2452,12 @@ version = "0.57.0"
|
|
|
2492
2452
|
source = "registry+https://github.com/rust-lang/crates.io-index"
|
|
2493
2453
|
checksum = "d2ed2439a290666cd67ecce2b0ffaad89c2a56b976b736e6ece670297897832d"
|
|
2494
2454
|
dependencies = [
|
|
2495
|
-
"windows-implement
|
|
2496
|
-
"windows-interface
|
|
2497
|
-
"windows-result
|
|
2455
|
+
"windows-implement",
|
|
2456
|
+
"windows-interface",
|
|
2457
|
+
"windows-result",
|
|
2498
2458
|
"windows-targets",
|
|
2499
2459
|
]
|
|
2500
2460
|
|
|
2501
|
-
[[package]]
|
|
2502
|
-
name = "windows-core"
|
|
2503
|
-
version = "0.62.2"
|
|
2504
|
-
source = "registry+https://github.com/rust-lang/crates.io-index"
|
|
2505
|
-
checksum = "b8e83a14d34d0623b51dce9581199302a221863196a1dde71a7663a4c2be9deb"
|
|
2506
|
-
dependencies = [
|
|
2507
|
-
"windows-implement 0.60.2",
|
|
2508
|
-
"windows-interface 0.59.3",
|
|
2509
|
-
"windows-link",
|
|
2510
|
-
"windows-result 0.4.1",
|
|
2511
|
-
"windows-strings",
|
|
2512
|
-
]
|
|
2513
|
-
|
|
2514
2461
|
[[package]]
|
|
2515
2462
|
name = "windows-implement"
|
|
2516
2463
|
version = "0.57.0"
|
|
@@ -2522,17 +2469,6 @@ dependencies = [
|
|
|
2522
2469
|
"syn 2.0.114",
|
|
2523
2470
|
]
|
|
2524
2471
|
|
|
2525
|
-
[[package]]
|
|
2526
|
-
name = "windows-implement"
|
|
2527
|
-
version = "0.60.2"
|
|
2528
|
-
source = "registry+https://github.com/rust-lang/crates.io-index"
|
|
2529
|
-
checksum = "053e2e040ab57b9dc951b72c264860db7eb3b0200ba345b4e4c3b14f67855ddf"
|
|
2530
|
-
dependencies = [
|
|
2531
|
-
"proc-macro2",
|
|
2532
|
-
"quote",
|
|
2533
|
-
"syn 2.0.114",
|
|
2534
|
-
]
|
|
2535
|
-
|
|
2536
2472
|
[[package]]
|
|
2537
2473
|
name = "windows-interface"
|
|
2538
2474
|
version = "0.57.0"
|
|
@@ -2544,17 +2480,6 @@ dependencies = [
|
|
|
2544
2480
|
"syn 2.0.114",
|
|
2545
2481
|
]
|
|
2546
2482
|
|
|
2547
|
-
[[package]]
|
|
2548
|
-
name = "windows-interface"
|
|
2549
|
-
version = "0.59.3"
|
|
2550
|
-
source = "registry+https://github.com/rust-lang/crates.io-index"
|
|
2551
|
-
checksum = "3f316c4a2570ba26bbec722032c4099d8c8bc095efccdc15688708623367e358"
|
|
2552
|
-
dependencies = [
|
|
2553
|
-
"proc-macro2",
|
|
2554
|
-
"quote",
|
|
2555
|
-
"syn 2.0.114",
|
|
2556
|
-
]
|
|
2557
|
-
|
|
2558
2483
|
[[package]]
|
|
2559
2484
|
name = "windows-link"
|
|
2560
2485
|
version = "0.2.1"
|
|
@@ -2570,24 +2495,6 @@ dependencies = [
|
|
|
2570
2495
|
"windows-targets",
|
|
2571
2496
|
]
|
|
2572
2497
|
|
|
2573
|
-
[[package]]
|
|
2574
|
-
name = "windows-result"
|
|
2575
|
-
version = "0.4.1"
|
|
2576
|
-
source = "registry+https://github.com/rust-lang/crates.io-index"
|
|
2577
|
-
checksum = "7781fa89eaf60850ac3d2da7af8e5242a5ea78d1a11c49bf2910bb5a73853eb5"
|
|
2578
|
-
dependencies = [
|
|
2579
|
-
"windows-link",
|
|
2580
|
-
]
|
|
2581
|
-
|
|
2582
|
-
[[package]]
|
|
2583
|
-
name = "windows-strings"
|
|
2584
|
-
version = "0.5.1"
|
|
2585
|
-
source = "registry+https://github.com/rust-lang/crates.io-index"
|
|
2586
|
-
checksum = "7837d08f69c77cf6b07689544538e017c1bfcf57e34b4c0ff58e6c2cd3b37091"
|
|
2587
|
-
dependencies = [
|
|
2588
|
-
"windows-link",
|
|
2589
|
-
]
|
|
2590
|
-
|
|
2591
2498
|
[[package]]
|
|
2592
2499
|
name = "windows-sys"
|
|
2593
2500
|
version = "0.59.0"
|
|
@@ -4,15 +4,14 @@ resolver = "2"
|
|
|
4
4
|
|
|
5
5
|
[workspace.package]
|
|
6
6
|
edition = "2021"
|
|
7
|
-
version = "0.
|
|
7
|
+
version = "0.7.0"
|
|
8
8
|
license = "MIT OR Apache-2.0"
|
|
9
9
|
repository = "https://github.com/puranikyashaswin/ferrotensor"
|
|
10
|
-
authors = ["Yashaswin Puranik <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 = "
|
|
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.
|
|
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
|
-
|
|
13
|
-
Author-email: Yashaswin Puranik <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
|
-
|
|
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
|
-
|
|
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.
|
|
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
|
|
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,
|
|
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
|
-
|
|
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
|
-
|
|
250
|
-
|
|
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 {
|
|
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
|
-
|
|
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
|
-
.
|
|
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
|
-
|
|
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
|
-
|
|
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
|
-
|
|
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
|
-
|
|
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
|
}
|