Laurent Mazare b13a82a438 Separate quantized phi-3 implementation. (#2157) ...

* Separate quantized phi-3 implementation.

* Integrate the quantized phi3 model.=

* Small fixes, get the generation to work properly.

* Keep the old llama implementation around.

* Change the default.

2024-05-04 10:14:57 +02:00

..

affine.metal

Optimize the cat operation on contiguous tensors (#1855)

2024-03-17 10:49:13 +01:00

binary.metal

update dtypes checks for several metal operations (#2010)

2024-04-04 18:39:06 +02:00

cast.metal

add support for casting between all datatypes (#1860)

2024-03-17 20:55:11 +01:00

conv.metal

Extend supported dtypes for metal (im2col & upsample_2d) (#1938)

2024-03-26 06:48:56 +01:00

indexing.metal

Support gather on bf16 for metal. (#2035)

2024-04-10 12:49:25 +02:00

lib.rs

Separate quantized phi-3 implementation. (#2157)

2024-05-04 10:14:57 +02:00

libMetalFlashAttention.metallib

chore: final

2024-01-22 15:15:19 +00:00

quantized.metal

Add argsort. (#2132)

2024-04-27 20:17:35 +02:00

random.metal

Metal random-generation bug fixes (#1811)

2024-03-08 16:11:50 +01:00

reduce.metal

Add the rope THD kernel. (#2014)

2024-04-05 08:32:58 +02:00

sort.metal

Add argsort. (#2132)

2024-04-27 20:17:35 +02:00

ternary.metal

Metal: f16 and bf16 where_cond + benchmark (#1545)

2024-01-12 11:18:11 +01:00

tests.rs

Use BufferOffset in metal backend ops. (#2029)

2024-04-08 09:37:25 +02:00

unary.metal

Fix sigmoid gradient calculation and move sigmoid into a specialized op (#2114)

2024-04-29 11:04:43 +02:00

utils.rs

Rework the buffer offset logic for metal kernels (#2028)

2024-04-07 22:37:53 +02:00