Jagrit Digani 89d327075f Enabling fused attention for head dim 128 (#1899) ...

* Share KV smem

* Fix bfloat error

* Unroll O = S @ V loop

* Perf upgrade

* Remove commented out function

* Add -Wno-c++17-extensions flag to metal flags

* Add -Wno-c++17-extensions flag to metal extension flags

2025-02-26 10:02:06 -08:00

..

fft

Feature complete Metal FFT (#1102)

2024-06-06 12:57:25 -07:00

jit

Dispatch bf16 at run time when using the JIT (#1584)

2024-11-15 16:54:36 -08:00

metal_3_0

Dispatch bf16 at run time when using the JIT (#1584)

2024-11-15 16:54:36 -08:00

metal_3_1

Dispatch bf16 at run time when using the JIT (#1584)

2024-11-15 16:54:36 -08:00

reduction

Use int64 stride everywhere (#1671)

2024-12-09 11:09:02 -08:00

steel

Enabling fused attention for head dim 128 (#1899)

2025-02-26 10:02:06 -08:00

arange.h

More jitting (#1132)

2024-05-23 16:23:44 -07:00

arange.metal

Dispatch bf16 at run time when using the JIT (#1584)

2024-11-15 16:54:36 -08:00

arg_reduce.metal

Use int64 stride everywhere (#1671)

2024-12-09 11:09:02 -08:00

atomic.h

Refactor reductions and fix scatter atomics for large sizes (#1300)

2024-08-22 16:03:31 -07:00

bf16_math.h

Dispatch bf16 at run time when using the JIT (#1584)

2024-11-15 16:54:36 -08:00

binary_ops.h

Fix complex power on Metal (#1460)

2024-10-06 19:58:30 -07:00

binary_two.h

Use int64 stride everywhere (#1671)

2024-12-09 11:09:02 -08:00

binary_two.metal

Allow no copy negative strides in as_strided and slice (#1688)

2024-12-12 08:59:45 -08:00

binary.h

Use int64 stride everywhere (#1671)

2024-12-09 11:09:02 -08:00

binary.metal

Allow no copy negative strides in as_strided and slice (#1688)

2024-12-12 08:59:45 -08:00

CMakeLists.txt

Enabling fused attention for head dim 128 (#1899)

2025-02-26 10:02:06 -08:00

complex.h

Refactor reductions and fix scatter atomics for large sizes (#1300)

2024-08-22 16:03:31 -07:00

conv.metal

Fix convs by reverting #1803 (#1882)

2025-02-18 14:36:34 -08:00

copy.h

Dynamic slicing (#1741)

2025-01-07 14:02:16 -08:00

copy.metal

Dynamic slicing (#1741)

2025-01-07 14:02:16 -08:00

defines.h

Refactor reductions and fix scatter atomics for large sizes (#1300)

2024-08-22 16:03:31 -07:00

erf.h

JIT compile option for binary minimization (#1091)

2024-05-22 12:57:13 -07:00

expm1f.h

Fix overflow / underflow handling for expm1f (#1278)

2024-07-23 07:29:06 -07:00

fence.metal

Faster synchronization Fence primitive (#1773)

2025-01-17 18:42:19 -08:00

fft.h

Feature complete Metal FFT (#1102)

2024-06-06 12:57:25 -07:00

fft.metal

Add Quantized Ops to the JIT (#1204)

2024-06-12 09:47:12 -07:00

gather_axis.h

scatter axis + gather axis primitives (#1813)

2025-01-31 20:48:08 -08:00

gather.h

Use int64 stride everywhere (#1671)

2024-12-09 11:09:02 -08:00

gemv_masked.h

Use int64 stride everywhere (#1671)

2024-12-09 11:09:02 -08:00

gemv_masked.metal

Use int64 stride everywhere (#1671)

2024-12-09 11:09:02 -08:00

gemv.metal

Use int64 stride everywhere (#1671)

2024-12-09 11:09:02 -08:00

hadamard.h

Fix bfloat16 Hadamard (#1283)

2024-07-23 14:54:43 -07:00

indexing.h

Use int64 stride everywhere (#1671)

2024-12-09 11:09:02 -08:00

layer_norm.metal

Dispatch bf16 at run time when using the JIT (#1584)

2024-11-15 16:54:36 -08:00

quantized.h

Faster small batch qmv (#1861)

2025-02-12 22:02:36 -08:00

quantized.metal

3 and 6 bit quantization (#1613)

2024-11-22 10:22:13 -08:00

random.metal

Use int64 stride everywhere (#1671)

2024-12-09 11:09:02 -08:00

reduce_utils.h

More jitting (#1132)

2024-05-23 16:23:44 -07:00

reduce.h

Fix JIT reductions (#1373)

2024-08-28 16:39:11 -07:00

reduce.metal

Allow no copy negative strides in as_strided and slice (#1688)

2024-12-12 08:59:45 -08:00

rms_norm.metal

Dispatch bf16 at run time when using the JIT (#1584)

2024-11-15 16:54:36 -08:00

rope.metal

Dispatch bf16 at run time when using the JIT (#1584)

2024-11-15 16:54:36 -08:00

scaled_dot_product_attention.metal

Allow different value dimensions in sdpa_vector (#1811)

2025-01-31 20:58:59 -08:00

scan.h

Working 64-bit scans (#1506)

2024-10-24 11:05:46 -07:00

scan.metal

Working 64-bit scans (#1506)

2024-10-24 11:05:46 -07:00

scatter_axis.h

scatter axis + gather axis primitives (#1813)

2025-01-31 20:48:08 -08:00

scatter.h

Use int64 stride everywhere (#1671)

2024-12-09 11:09:02 -08:00

sdpa_vector.h

Allow different value dimensions in sdpa_vector (#1811)

2025-01-31 20:58:59 -08:00

softmax.h

consistently handle all -inf in softmax (#1470)

2024-10-08 09:54:02 -07:00

softmax.metal

Dispatch bf16 at run time when using the JIT (#1584)

2024-11-15 16:54:36 -08:00

sort.h

faster sort (#1831)

2025-02-05 06:10:22 -08:00

sort.metal

faster sort (#1831)

2025-02-05 06:10:22 -08:00

ternary_ops.h

JIT compile option for binary minimization (#1091)

2024-05-22 12:57:13 -07:00

ternary.h

Use int64 stride everywhere (#1671)

2024-12-09 11:09:02 -08:00

ternary.metal

Fix nd ternary on GPU (#1746)

2025-01-03 11:52:17 -08:00

unary_ops.h

Bitwise Inverse (#1862)

2025-02-13 08:44:14 -08:00

unary.h

Use int64 stride everywhere (#1671)

2024-12-09 11:09:02 -08:00

unary.metal

Bitwise Inverse (#1862)

2025-02-13 08:44:14 -08:00

utils.h

Allow no copy negative strides in as_strided and slice (#1688)

2024-12-12 08:59:45 -08:00