Anastasiia Filippova
f7c11b965e
Merge branch 'main' into nccl_backend
2025-08-20 13:37:18 +02:00
Cheng
ac85ddfdb7
[CUDA] Add GEMM-based fallback convolution kernels ( #2511 )
...
* Add gemm_conv
* Add gemm_grouped_conv
2025-08-20 10:06:22 +09:00
Cheng
65d0d40232
Split cuDNN helpers into a separate header ( #2491 )
...
* Add RAII managed CudaGraph class
* Implement forward rms_norm with cuDNN
* Revert back to old rms norm kernel
2025-08-20 09:29:28 +09:00
Awni Hannun
cea9369610
fix lapack svd ( #2515 )
2025-08-18 15:07:59 -07:00
Angelos Katharopoulos
1df9887998
Ensure no oob read in gemv_masked ( #2508 )
2025-08-17 08:42:33 -07:00
Angelos Katharopoulos
73f22d6226
Ensure small sort doesn't use indices if not argsort ( #2506 )
2025-08-17 08:42:20 -07:00
Cheng
c422050ca7
Update cuDNN Frontend to v1.14 ( #2505 )
2025-08-17 19:13:01 +09:00
Cheng
1ba18ff7d9
[CUDA] Fix conv grads with groups ( #2495 )
...
* Put reshape utils in one file
* [CUDA] Fix conv grads with groups
* Put the reshape utils in gpu/copy.h
2025-08-16 10:09:18 +09:00
Cheng
37b440faa8
Clean up code handling both std::vector and SmallVector ( #2493 )
2025-08-16 09:01:10 +09:00
Cheng
888b13ed63
Remove the hack around SmallVector in cpu compile ( #2494 )
2025-08-16 08:17:24 +09:00
Cheng
4abb218d21
The naive_conv_2d is no longer used ( #2496 )
2025-08-16 07:57:30 +09:00
Awni Hannun
6441c21a94
Faster general unary op ( #2472 )
...
* faster general unary op
* faster general ops + reorg
* fix + comment
* binary two
* copy general
2025-08-15 15:04:12 -07:00
Cheng
dfb5022eab
Rename cu::Matmul to CublasGemm ( #2488 )
2025-08-13 09:37:40 +09:00
Abe Leininger
fce53b61d6
Fix reduce sum/prod overflow ( #2477 )
2025-08-12 00:05:33 -07:00
Cheng
aa7b47481a
[CUDA] Optimize set_mm_device_pointers for small ndim ( #2473 )
2025-08-08 15:23:30 +09:00
Anastasiia Filippova
f540b1d612
nccl backend
2025-08-07 13:11:56 +02:00
Jagrit Digani
a9bdd67baa
Add CUDA sdpa vector ( #2468 )
2025-08-06 21:40:26 -07:00
Angelos Katharopoulos
f2adb5638d
Fix typo in metal command encoder ( #2471 )
2025-08-06 16:58:23 -07:00
Awni Hannun
7bb96e4249
fix cublas on h100 ( #2466 )
2025-08-06 06:18:58 -07:00
Cheng
828c5f1137
Use SmallVector for shapes and strides ( #2454 )
...
* Use SmallVector for shapes and strides
* Convert SmallVector to tuple
2025-08-05 09:41:03 +09:00
Zamderax
737dd6d1ac
Add missing <algorithm> header to jit_compiler.cpp ( #2460 )
...
Fixes compilation error on Linux where std::find_if is used on line 121
but the <algorithm> header was not included. While this might work on
some platforms due to transitive includes, it's not guaranteed by the
C++ standard.
Resolves issue #2459
2025-08-04 14:00:46 -07:00
Cheng
aaf78f4c6b
Use LRU cache for cuda graph ( #2448 )
...
* Use LRU cache for cuda graph
* Remove unused destructor
2025-08-02 21:28:57 +09:00
Angelos Katharopoulos
be9bc96da4
[CUDA] Matmul utils initial commit ( #2441 )
2025-08-01 14:22:25 -07:00
Angelos Katharopoulos
86258f292f
[CUDA] Vectorize generated kernels ( #2444 )
2025-07-31 18:18:57 -07:00
Cheng
b26d88591c
[CUDA] Save primitive inputs faster ( #2449 )
...
* Add more nvtx loggings
* [CUDA] Saving primitive inputs faster
* Remove unneeded check
2025-08-01 10:16:06 +09:00
Cheng
86c6a15571
[CUDA] Backward convolution ( #2431 )
2025-08-01 09:54:05 +09:00
Cheng
daafee676f
Fix wrong graph key when using concurrent context ( #2447 )
2025-07-31 06:01:05 -07:00
Awni Hannun
d32519c8ee
fix gemv regression ( #2445 )
2025-07-30 14:23:01 -07:00
Angelos Katharopoulos
3bf81ed1bd
[CUDA] Quantized refactoring ( #2442 )
2025-07-30 08:27:20 -07:00
Cheng
3628e5d497
Use load_vector in arg_reduce ( #2439 )
2025-07-30 17:40:26 +09:00
Cheng
a0ae49d397
Move arange to its own file ( #2438 )
2025-07-30 13:05:51 +09:00
Cheng
254476718b
Remove the kernel arg from get_launch_args ( #2437 )
2025-07-30 11:43:02 +09:00
Awni Hannun
3adba92ebe
Cuda faster softmax ( #2435 )
...
* faster softmax and logsumexp
* faster softmax and logsumexp
* format
2025-07-29 17:18:12 -07:00
Awni Hannun
ef631d63af
faster rms norm ( #2433 )
2025-07-29 13:12:00 -07:00
Awni Hannun
641be9463b
Add more CUDA architectures for PyPi package ( #2427 )
...
* add cuda sm 90
* add more archs
2025-07-28 12:35:15 -07:00
Awni Hannun
ab0e608862
[CUDA] More sizes for gemv ( #2429 )
...
* route more to gemv
* route more sizes to custom gemv
2025-07-28 12:35:01 -07:00
Awni Hannun
1588659062
no occupancy query for launch params ( #2426 )
2025-07-28 09:09:41 -07:00
Awni Hannun
b9e88fb976
[CUDA] Fix segfault on exit ( #2424 )
...
* fix cuda segfault on exit
* comment
2025-07-27 08:08:13 -07:00
Awni Hannun
5597fa089c
Fix qvm splitk ( #2415 )
2025-07-25 11:50:24 -07:00
Awni Hannun
9acec364c2
[CUDA] Always use batched matmul ( #2404 )
...
* cuda batched mm
* addmm as well
* comment
2025-07-24 20:46:02 -07:00
Cheng
6f5874a2f2
[CUDA] Initial implementation of Convolution with cuDNN ( #2385 )
...
* Link with cuDNN
* Initial implementation
* Remove backend apis
* Fix recording cudnn conv
* More unused backend apis
* Fix C++ conv tests
* include cudnn as python dep
* Install libcudnn9-dev-cuda-12 in CI
* cudnn only accepts contiguous inputs
* Switch to backend apis
* Plan needs to be kept alive
* Turn off tf32
* Add cache
* Test the native cuda graph api
* Set cudnn stream before execution
* Make LRUCache more like a normal container
* Do error check for cublas handle
* Zero-initilizing array
* Use tf32 for conv
* Skip TestConv.test_torch_conv_2D test
---------
Co-authored-by: Awni Hannun <awni@apple.com >
2025-07-25 08:12:10 +09:00
Awni Hannun
4e504039f5
[Metal] Release metal events ( #2412 )
...
* release metal events
* fix
* fix
2025-07-23 19:53:42 -07:00
Cheng
0f5ce173da
[CUDA] --compress-mode requires CUDA 12.8 ( #2407 )
2025-07-23 06:11:11 -07:00
Awni Hannun
d107d8d495
add cuda gemv ( #2400 )
2025-07-22 08:24:13 -07:00
Awni Hannun
1e496ddb82
[CUDA] Simplify allocator ( #2392 )
...
* simplify allocator and fixe race with small pool
* Don't use shared event in worker
* use cuda buffer in small pool
* comment
* comment
2025-07-22 08:24:01 -07:00
Awni Hannun
74eccbf3fa
use size option in binary ( #2399 )
2025-07-22 07:00:53 -07:00
Cheng
56cc858af9
Add contiguous_copy_cpu util for copying array ( #2397 )
2025-07-21 07:30:35 -07:00
Cheng
f55c4ed1d6
Remove thrust iterators ( #2396 )
2025-07-21 07:30:27 -07:00
Awni Hannun
93d70419e7
[CUDA] speedup handling scalars ( #2389 )
...
* speedup scalars in cuda
* comment
2025-07-18 21:47:31 -07:00
Gökdeniz Gülmez
deee214a95
Adding support for the Muon Optimizer ( #1914 )
...
* initial commit with workong optmimizer
* update ACKNOWLEDGMENTS.md
* nits and adding it to test
* nits
* G.astype(mx.bfloat16) to G.astype(G.dtype)
* G.ndim >= 2 to assert G.ndim == 2
* remove coments
* replace with mx.addmm
* remove comments
* format
* nits
* match muon
* fix addmm
---------
Co-authored-by: Awni Hannun <awni@apple.com >
2025-07-18 12:25:28 -07:00
