98172d46fa
Fix some errors about BlockQ8_1 ( #776 )
...
* use int8 type instead of uint8 for BlockQ8_1.qs
The uint8 type of BlockQ8_1.qs causes great loss for negative weights
Ref: ebc96086af/ggml.c (L904)zmlcc@outlook.com >
* fix sum error in vec_dot of BlockQ4_1
Ref: ebc96086af/ggml.c (L2840)zmlcc@outlook.com >
* fix sum error in vec_dot of BlockQ5_1
Ref: ebc96086af/ggml.c (L3490)zmlcc@outlook.com >
---------
Signed-off-by: Zhang Miaolei <zmlcc@outlook.com >
2023-09-08 13:29:40 +01:00
0e250aee4f
Shape with holes ( #770 )
...
* Shape with holes.
* rustfmt.
2023-09-08 08:38:13 +01:00
7396b8ed1a
Segment Anything - process images ( #766 )
...
* Start processing images.
* Add LayerNorm2d.
* Properly use LayerNorm2d.
* Tweak eps.
* Use LayerNorm on inputs with a rank different from 3.
* Window partitioning.
* Fix a couple todos.
* More todos.
* Hard-code the einsums.
* More padding support.
* Some sizes tweaks.
* Use the hub to get the weights.
* Use a batch matmul.
* Tweaks.
* More fixes.
* Get some predictions to be generated.
2023-09-07 19:22:45 +01:00
6527ab81a3
Sketch the segment anything model. ( #759 )
...
* Sketch the segment anything model.
* Fix some clippy lint.
* Add the mask decoder.
2023-09-07 05:34:05 +01:00
7b1f2da828
Cudnn fix. ( #758 )
2023-09-06 17:39:39 +01:00
7299a68353
img2img pipeline for stable diffusion. ( #752 )
...
* img2img pipeline for stable diffusion.
* Rename the arguments + fix.
* Fix for zero strength.
* Another fix.
* Another fix.
* Revert.
* Include the backtrace.
* Noise scaling.
* Fix the height/width.
2023-09-06 07:06:49 +01:00
a4f40f3dc8
Use rayon directly rather than constraining the number of threads. ( #749 )
2023-09-05 20:26:15 +01:00
a0d65585db
Softmax implementation for cuda. ( #747 )
2023-09-05 18:38:03 +01:00
6615daf242
Tweaks to softmax. ( #745 )
2023-09-05 15:22:27 +01:00
1c9e5394a5
Add a custom softmax implementation. ( #744 )
...
* Add a custom softmax implementation.
* Add softmaxlastdim to the benchmarks.
* And add a test.
* Support more dtypes.
* Polish the code.
* Use the slow implementation on cuda.
* Add a todo for the cuda kernel.
2023-09-05 14:20:23 +01:00
a8410bf35e
Add some documentation. ( #743 )
2023-09-05 09:51:12 +01:00
cda45a7443
Let outside CustomOp2 implementations use binary_map/binary_map_vec ( #741 )
2023-09-05 09:27:32 +01:00
f7980e07e0
Add ggufv2 support ( #725 )
2023-09-03 14:41:57 +01:00
74a82c358a
Add the mse loss. ( #723 )
2023-09-03 10:51:40 +01:00
84d003ff53
Handle arbitrary shapes in Tensor::new. ( #718 )
2023-09-02 19:59:21 +01:00
2ed78ab336
Support for quantized tensors in the python api. ( #706 )
...
* Add more pyo3 support.
* Add some support for quantized tensors in pyo3.
* Add an arc layer on qmatmul.
* Add the quantized matmul.
* Quantization support.
* More quantization support.
* Test the python quantization.
2023-09-01 15:53:42 +01:00
237323c2bc
Cleanup the pyo3 setup. ( #705 )
2023-09-01 14:26:18 +01:00
30a4b593d7
More ops again. ( #697 )
2023-08-31 22:28:48 +01:00
949f1eae6f
Implement a couple more binary ops. ( #693 )
2023-08-31 21:30:15 +01:00
9874d843f1
Fix the accelerate build ( #678 )
...
* Cosmetic changes.
* Fix the accelerate build for tanh.
2023-08-30 18:31:14 +02:00
ad8a62dbf5
Add tanh. ( #675 )
...
* Add tanh.
* Use tanh in the lstm block.
* Add a test for tanh forward and backward passes.
2023-08-30 13:54:50 +01:00
618f4e4c78
Add some documentation. ( #673 )
...
* Add some documentation.
* Bump the crate version.
2023-08-30 11:54:00 +01:00
393690387f
Support dilation in conv-transpose2d. ( #671 )
2023-08-30 09:22:00 +01:00
9b25113393
Small cleanups (avoid some possible mutations) ( #670 )
...
* More mut cleanup.
* Factor out some common bits.
2023-08-30 08:54:00 +01:00
a1a5ab8b0a
Neon optimized vecdot ( #666 )
...
* Q5k vecdot.
* Add the q3k vecdot.
* Q2k vecdot.
* Move the quantized model to its own file.
2023-08-29 22:28:46 +01:00
59b731de99
Add the powf op. ( #664 )
...
* Add the powf op.
* Cuda kernels and backprop.
* Add a test.
2023-08-29 20:48:18 +01:00
2d3fcad267
Simplify usage of the pool functions. ( #662 )
...
* Simplify usage of the pool functions.
* Small tweak.
* Attempt at using apply to simplify the convnet definition.
2023-08-29 19:12:16 +01:00
71221559d3
Fix the dilated convolutions. ( #659 )
2023-08-29 16:37:42 +01:00
a044907ffc
Dilated convolutions ( #657 )
...
* Add the dilation parameter.
* Restore the basic optimizer example.
* Dilation support in cudnn.
* Use the dilation parameter in the cpu backend.
* More dilation support.
* No support for dilation in transposed convolutions.
* Add dilation to a test.
* Remove a print.
* Helper function.
2023-08-29 16:12:11 +01:00
ee8bb1bde1
Add avx implemenetations of q2k, q3k and q5k vec-dot functions ( #654 )
...
* `q2k` avx implementation
* `q3k` avx implementation
* `q5k` avx implementation
* `avx` make masks constant
* clippy stuff
2023-08-29 13:35:56 +01:00
d0a330448d
Backprop support for pooling ops. ( #652 )
...
* Backprop support for pooling ops.
* max-pool gradient.
2023-08-29 10:17:59 +01:00
4b8d57ba15
AVX version of the q4k vecdot. ( #651 )
2023-08-29 09:41:17 +01:00
fd3131a4ce
Fix the debug implementation. ( #648 )
2023-08-28 22:51:39 +01:00
037b41c9dc
Cuda conv transpose ( #645 )
...
* Cuda kernel for conv-transpose.
* Fix the cuda kernel.
* Fix the tests.
2023-08-28 20:58:49 +01:00
72fae3140c
Optimize the conv2d transpose cpu kernel. ( #644 )
...
* Optimize the conv2d transpose cpu kernel.
* Use multiple cores.
2023-08-28 20:06:31 +01:00
ca26198b95
Fix the cpu kernel for conv-transpose. ( #643 )
2023-08-28 16:45:12 +01:00
b292047882
Backprop for conv2d. ( #638 )
...
* Start adding backprop for conv2d.
* Backprop for conv2d.
* Bugfix + start adding a conv2d test.
* Conv2d backprop testing.
* More conv fixes.
2023-08-28 16:08:55 +01:00
3cca89cc70
Add conv-transpose. ( #635 )
...
* Add conv-transpose.
* Return zeros for now.
* Naive CPU implementation.
* Add a conv-transpose test + fix the cpu implementation.
* Add a second test.
2023-08-28 10:10:12 +01:00
1da71a5da1
Neon optimized version of the q4k vecdot product. ( #632 )
2023-08-27 21:30:47 +01:00
a3f97c143d
Bump the crate version + update CHANGELOG. ( #628 )
2023-08-27 18:17:11 +01:00
be471d50ab
Llama quantization. ( #625 )
2023-08-27 14:08:15 +01:00
7151f2cf63
Add the quantize command. ( #624 )
...
* Add the quantize command.
* Bugfix for writing gguf files.
* And add a comment.
2023-08-27 11:35:19 +01:00
5320aa6b7d
Move the test-utils bits to a shared place. ( #619 )
2023-08-27 09:42:22 +01:00
a8b39dd7b7
Fix for q5_1 quantization. ( #617 )
...
* Fix for q5_1 quantization.
* Fix some typos.
2023-08-27 08:31:18 +01:00
fa0d75b18d
Quantization tests + fix some issues. ( #616 )
2023-08-27 08:17:38 +01:00
28658054ff
More missing quantized bits. ( #615 )
...
* Q4_1 support.
* Add Q5_1 quantization.
* Tweak.
2023-08-27 07:52:26 +01:00
ab36a7f3e3
Fix for when f16c is not available. ( #614 )
2023-08-27 07:19:52 +01:00
f704e39761
Missing quants ops ( #611 )
...
* Another transmute tweak.
* Changelog tweak.
* Add some missing quantized ops.
2023-08-26 20:09:04 +01:00
fdf15f0e05
Another transmute tweak. ( #610 )
...
* Another transmute tweak.
* Changelog tweak.
2023-08-26 13:00:24 +01:00
06b37ea7ad
Avoid using tmp values. ( #609 )
2023-08-26 12:28:28 +01:00
