Jared Van Bortel
84f7fc4553
vulkan : rope n_past is now KQ_pos, f16 rope kernel
2023-11-23 17:18:42 -05:00
Jared Van Bortel
71565eb0c3
vulkan : replace ggml_diag_mask_inf with ggml_add (custom -inf mask)
2023-11-23 17:18:27 -05:00
Adam Treat
e006d377dd
Scale the workgroup count down to allow correct generation for falcon with
...
AMD radeon cards with lower workgroup count limit
Partially fixes #1581
2023-11-03 17:22:22 -04:00
cebtenzzre
1c17010188
vulkan : fix missing break in matmul selection ( #9 )
2023-11-03 17:22:22 -04:00
Adam Treat
8d9efbf97a
Lower the workgroup count for some shaders by providing a loop that processes
...
four floats at a time.
2023-11-03 17:22:22 -04:00
Adam Treat
8400015337
Don't try an allocation on a heap that is smaller than the size we require.
2023-11-03 17:22:22 -04:00
Aaron Miller
cc05a602d6
use mat*vec shaders for mat*mat
...
I wrote the mat*mat shaders from scratch so I understand them better but
they are currently not faster than just multiply-invoking the mat*vec
shaders, by a significant degree - so, except for f32 which needed a new
shader, revert to the m*v ones here.
2023-11-03 17:22:22 -04:00
Aaron Miller
c1fd64548d
attempted speedups 2
2023-11-03 17:22:22 -04:00
Aaron Miller
9bc52ebae3
attempted speedups
2023-11-03 17:22:22 -04:00
Aaron Miller
cd0257ed0d
q4_1 mat*mat
2023-11-03 17:22:22 -04:00
Aaron Miller
b78a94bc6d
q6k mm works
2023-11-03 17:22:22 -04:00
Aaron Miller
d5741c07a5
use op param epsilon for norms
2023-11-03 17:22:22 -04:00
Aaron Miller
3327d84a7f
perf: use bigger threadgroups in mm
2023-11-03 17:22:22 -04:00
Aaron Miller
46385ee0d5
misc vulkan cleanup
...
make pushconts consistent w/ dispatch, avoid a double free
2023-11-03 17:22:22 -04:00
Aaron Miller
f0cd38b9ad
add mat*mat ops
2023-11-03 17:22:22 -04:00
Aaron Miller
ff4212d20f
q8 mat*vec
2023-11-03 17:22:21 -04:00
Aaron Miller
9db90cbe12
f16 mv broadcasting fix (gqa fix)
2023-11-03 17:22:21 -04:00
Adam Treat
bc4b5ed1cb
Fixes for subgroup size to bring AMD and NVIDIA inline with eachother for all kernels.
2023-11-03 17:22:21 -04:00
Adam Treat
de589ced7c
Change this back to be in agreement with metal and our previous softmax kernel.
2023-11-03 17:22:21 -04:00
Adam Treat
f1c9bc1821
Add q6_k getrows and mul*vec kernel.
2023-11-03 17:22:21 -04:00
Adam Treat
5509f74318
Minor cleanup.
2023-11-03 17:22:21 -04:00
Adam Treat
93306f16d0
Consolidate code for mat x vec kernels and use subgroups more extensively.
2023-11-03 17:22:21 -04:00
Adam Treat
2c24d67e7b
Don't crash on available devices if we can't even create an instance.
2023-10-05 13:39:18 -04:00
Adam Treat
addac25293
Set the singleton to nullptr here.
2023-10-05 13:39:18 -04:00
Adam Treat
68aca6be08
Only use vulkan with known quant that work.
2023-10-05 13:39:18 -04:00
Aaron Miller
8bea719879
vulkan: disambiguate gpus with the same name
2023-10-05 13:39:18 -04:00
Adam Treat
68cf1df6fb
Throw an exception when allocation fails for vulkan.
2023-10-05 13:39:18 -04:00
Adam Treat
b7e2e691d4
Completely revamp how we do object management with the vulkan backend and
...
stop using so many static objects so we can tear down and bring up vulkan
on new devices in the same runtime.
2023-10-05 13:39:18 -04:00
Adam Treat
45c8778b49
Switch to a dynamic dispatch table instead of linking hard against libvulkan.
2023-10-05 13:39:18 -04:00
Adam Treat
48a45ea435
Remove warning which fails on windows.
2023-10-05 13:39:18 -04:00
niansa
ba15dfd0be
Nomic vulkan backend licensed under the Software for Open Models License (SOM), version 1.0.
2023-10-05 13:39:18 -04:00
