Po Yen Chen
6070090258
Mask out key values for each splits
2025-05-27 09:50:50 +00:00
Po Yen Chen
9512f78616
Support logits_soft_cap in batch_decode()
2025-04-21 23:33:52 +00:00
Po Yen Chen
fd84cf840a
Fix batch_decode() codegen
2025-04-21 13:04:35 +00:00
Po Yen Chen
9f509a0f74
Merge branch 'dev/testx_fa_cap_logits' into poyenc/batch-prefill-decode
2025-04-21 08:36:44 +00:00
coderfeli
9aa77620b8
fix perf
2025-04-21 05:43:57 +00:00
Po Yen Chen
3b8662f198
Re-format old CK source files
2025-04-21 11:39:33 +08:00
Po Yen Chen
09c5ad5241
Sync logits soft-capping across pipelines
2025-04-21 11:39:33 +08:00
Po Yen Chen
2632baa05b
Support turn on/off logits_soft_cap in async pipeline
2025-04-21 11:39:33 +08:00
Po Yen Chen
b3829c11b5
Allow specifying logits_soft_cap through APIs
2025-04-21 11:39:33 +08:00
Po Yen Chen
124f47a5e8
Re-format files
2025-04-21 11:39:33 +08:00
Po Yen Chen
519c93bd08
Add batch prefill/decode kernels
2025-04-21 02:52:02 +00:00
Po Yen Chen
85becc24ee
Re-format old CK source files
2025-04-20 08:00:15 +00:00
Po Yen Chen
889b2d33fd
Sync logits soft-capping across pipelines
2025-04-20 07:55:36 +00:00
Po Yen Chen
0200227d89
Support turn on/off logits_soft_cap in async pipeline
2025-04-20 06:56:20 +00:00
Po Yen Chen
87b22a7cff
Allow specifying logits_soft_cap through APIs
2025-04-20 06:23:38 +00:00
Po Yen Chen
4927305440
Re-format files
2025-04-20 04:28:58 +00:00
coderfeli
d2c653f177
fix bug
2025-04-18 10:45:53 +00:00
Bernard
838ff7034d
hack for cap logits
2025-04-18 08:38:25 +00:00
coderfeli
b20173a494
fix bug
2025-04-09 15:18:13 +00:00
coderfeli
867a4e527c
fix bugs
2025-04-04 13:40:10 +00:00
coderfeli
491178276f
fix fp8 scale
2025-04-03 11:10:37 +00:00
lalala-sh
3037d8bac1
mul int4 scale
2025-04-03 18:06:12 +08:00
root
20f6674bf6
fix no quant case
2025-04-03 02:46:01 +00:00
root
b2b34fffbb
fix fp8 16x16
2025-04-02 16:27:52 +00:00
root
85f83330b5
fuse moe activation
2025-04-02 07:02:09 +00:00
coderfeli
98cee8d02b
fix merge
2025-03-18 05:45:04 +00:00
coderfeli
5f49b91237
merge develop
2025-03-18 04:49:40 +00:00
aledudek
5095906975
Async grouped gemm v3 ( #1940 )
...
* Fully async grouped gemm
* Remove commented code
* Remvoe maybe_unused
* host kernel args
* Checkpoint segfault debugging...
* Working part1
* Working part2
* Remvoe comments...
* Use void ptr for gemm kernel host args
* Fix device_grouped_gemm_multiple_d_dl build issue
* Fix device_grouped_gemm_xdl build issue
2025-03-17 16:42:43 +01:00
Bartłomiej Kocot
c2e4898b4b
Grouped conv bwd data NGCHW ( #1967 )
...
* Grouped conv bwd data NGCHW
* fixes
* fix
* Improvements
* Fix
* Fix
* add client example
2025-03-17 13:32:00 +01:00
coderfeli
7dbdff9f9f
moe sorting fix moebuf
2025-03-17 06:20:57 +00:00
coderfeli
5eaa36be18
mork to support 13w tokens
2025-03-17 01:45:34 +00:00
coderfeli
ef8c1333b9
use uint32
2025-03-17 01:45:09 +00:00
coderfeli
6c0e021235
revert v1 test
2025-03-17 01:39:57 +00:00
coderfeli
bccc5192cf
fix uint32
2025-03-17 01:18:32 +00:00
coderfeli
da2659d502
input output all ok
2025-03-15 14:26:30 +00:00
coderfeli
d1e999c05c
int64 index ok now
2025-03-15 13:28:49 +00:00
coderfeli
f911cf7396
impl int64 but result not correct
2025-03-14 13:01:07 +00:00
coderfeli
d4925e1637
fix output oob
2025-03-14 03:19:26 +00:00
carlushuang
3e81279d26
Reapply "[CK_TILE] support hdim=192/128 pair for deepseekv3 ( #1961 )" … ( #1971 )
...
* Reapply "[CK_TILE] support hdim=192/128 pair for deepseekv3 (#1961 )" (#1969 )
This reverts commit 8cbcd3e0d098187287+illsilin@users.noreply.github.com >
2025-03-13 11:41:39 +08:00
illsilin
f8464d2087
fix clang format
2025-03-12 20:21:14 -07:00
coderfeli
d85c034977
fix2
2025-03-13 02:30:07 +00:00
coderfeli
8b05fa935d
fix coredump in e2e test
2025-03-13 02:12:18 +00:00
feli
251afab3b7
ck_moe: fix useless code and remove usless oob ( #1972 )
...
* fix useless code and remove usless oob
* clang format
---------
Co-authored-by: coderfeli <coderfeli@163.com >
2025-03-12 09:22:42 -07:00
Illia Silin
4c97cc511e
use old instrinsics with staging compiler ( #1970 )
2025-03-12 07:29:09 -07:00
feli
2585c78940
Merge branch 'develop' into ck_moe_rm_oob
2025-03-12 16:05:59 +08:00
coderfeli
40542296de
clang format
2025-03-12 08:05:12 +00:00
Illia Silin
8cbcd3e0d0
Revert "[CK_TILE] support hdim=192/128 pair for deepseekv3 ( #1961 )" ( #1969 )
...
This reverts commit 7a93b16ff6
2025-03-11 10:40:18 -07:00
Haocong WANG
cbd74c2d12
[Block Scale GEMM] Optimized block scale gemm ( #1950 )
...
* Added two kernel for M=32 problem
* Comment the first one
* Enable multiply_multiply for Scale_Block_M = 1 for deepseek
* Modify the a_thread offset since the A data load is different from B.
* edit fp8 ab scale for Scale_Block_M=1
* edit GemmSpec to MNKPadding
* enable blockwise pipelie v1 and v2. v1 is work for small K.
* add instance for gemm_ab_scale
* fix cmakelist of ckProfiler
* optimize blockscale gemm. todo: reduce vgpr usage
* fix a correctness bug
* sanity checked
* revert ckprofiler cmake changes
* clang format
* revert unnecessary changes.
* remove commented codes.
* split weight preshuffle library targets
* bring back enable-post-misched=0
* fix build issues for gemm_multiply_multiply_fp8 instances
* fix clang format
* add verbose build flag when building for all targets
* reduce path names for new instances
* fix paths in cmake
* refactor gemm_multiply_multiply library target
* fix a bug in example
* fix example 65 cmake
* reduce the number of threads when building libs for all targets to 50
* use ninja to build for all targets
* reduce teh number of threads when building for all targets
* reduce the number of threads to 32 when building libs for all targets to 50
---------
Co-authored-by: mtgu0705 <mtgu@amd.com >
Co-authored-by: chenjun <junchen2@amd.com >
Co-authored-by: illsilin <Illia.Silin@amd.com >
Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com >
2025-03-11 10:11:21 -07:00
carlushuang
7a93b16ff6
[CK_TILE] support hdim=192/128 pair for deepseekv3 ( #1961 )
...
* support hdim=192/128 pair
* remove useless print
* update
2025-03-11 21:07:40 +08:00
coderfeli
1508cd514a
fix useless code and remove usless oob
2025-03-11 07:05:05 +00:00
