qazal and GitHub
d716d0d927
hotfix: qkv bw kernel requires hipcc and cdna4 ( #17645 )
2026-08-21 12:38:45 +09:00
qazal and GitHub
a1263fadf3
fused_qkv_rope in UOp try 2 ( #17619 )
...
* fused_qkv_rope in UOp try 2
* dont need that
* less
2026-08-20 12:28:34 +09:00
qazal and GitHub
d539aaf752
Revert "fused_qkv_rope in UOp ( #17591 )" ( #17593 )
...
This reverts commit 8c2bf02d17
2026-08-19 21:42:55 +09:00
qazal and GitHub
8c2bf02d17
fused_qkv_rope in UOp ( #17591 )
...
* llama: 4% faster fused_qkv_rope
* prep
* add uop kernel, has_hipcc is cached
* less
2026-08-19 18:07:30 +09:00
qazal and GitHub
44f1f45cd5
llama: custom silu kernels ( #17462 )
...
* start by copying the C
* uop kernel
* cleanup tests
* estimates is part of SPEC
2026-08-10 16:43:01 +09:00
George Hotz and GitHub
e1f42681fa
add new schedule tests + format better ( #17402 )
...
* add new schedule tests + format better
* assert_kernel_count
2026-08-04 18:46:38 -07:00
George Hotz and GitHub
57ae1bc7a7
rename MULTI to UNSHARD ( #17267 )
...
* rename MULTI to UNSHARD
* comment updates (glm)
* rename method to unshard
2026-07-28 16:51:41 -07:00
qazal and GitHub
1cf8f2f68c
llama: inplace amax update ( #17064 )
...
* llama: inplace amax update
* remove amax_out return
* work
* fit
* work
* work
* keep
* diff cleanup
2026-07-20 15:05:41 +09:00
qazal and GitHub
7dd3422c63
llama: replace two stage amax with atomics ( #17063 )
...
* atomic amax in c kernels
* quantize fp8 UOp kernel
* diff
2026-07-17 19:27:10 +09:00
qazal and GitHub
939f28d571
fused qkv rope custom kernel ( #17021 )
...
* work
* fused qkv_norm
* work
* speed
* not that yet
* test cleanup
* just clone
* remove .realize()
* cleanup tests
2026-07-15 01:08:42 +09:00
George Hotz and GitHub
41d6731bfd
reduce removes ones ( #16847 )
...
* reduce removes ones
* test changes
* lil clean
2026-07-03 13:10:08 -07:00
chenyu and GitHub
19eb72ff60
remove use of full with buffer=False and non-None device= ( #16489 )
2026-06-03 16:21:24 -04:00
qazal and GitHub
29b47a0057
llama: update local amax implementation after ParamArgs change ( #16446 )
...
* local amax failing test
* update _local_abs_max_fxn
2026-05-30 16:55:43 +09:00
qazal and GitHub
bbfe4f80ec
quantize_fp8 kernels in uops ( #16288 )
...
* add tests
* simple UOp kernel is n^2
* fast kernel matching c++, opts_to_apply=()
* remove cpp
* simple o(n) kernel, two passes
* fuse the loops
* works on DEV=CPU
* multi regression test
* fix multi, this can possibly be its own bugfix
* test cleanups
* minimal diff
* match C in UOps
* Revert "match C in UOps"
This reverts commit 0bef740c30
2026-05-22 20:54:06 +09:00
sirhcm and GitHub
172f9493e1
move is_dtype_supported to renderer ( #16226 )
2026-05-20 21:19:37 -04:00
qazal and GitHub
1e0fffe256
fused ce llama kernel in UOps ( #16263 )
...
* work
* using uops
* delete things
* work
* work
* higher level uops
* cleanups
2026-05-20 19:45:28 +09:00
sirhcmandGitHub