Commits · 40a4d8962979753cbabb022a2dd5d2bfdd08d009 · OpenDAS / TransformerEngine

03 Jul, 2025 1 commit
- Fix kernel crash on block_len=64 · 40a4d896
  wenjh authored Jul 03, 2025
```
Signed-off-by: wenjh <wenjh@sugon.com>
```
  40a4d896
01 Jul, 2025 1 commit

[Blockwise] Add support block_len=64 support · b944277c

wenjh authored Jun 25, 2025



Add env to chose blocklen of blockwise quantize.
Signed-off-by: wenjh <wenjh@sugon.com>

Fix pytest of blockwise error
Signed-off-by: wenjh <wenjh@sugon.com>

Resolve new api in  int8 gemm test
Signed-off-by: wenjh <wenjh@sugon.com>

Fix incorrect launch parm
Signed-off-by: wenjh <wenjh@sugon.com>

Fix 1D blockwise(64) acc error
Signed-off-by: wenjh <wenjh@sugon.com>

b944277c

20 Jun, 2025 2 commits
- [DCU] fix megatron MOE int8 train bugs · 251dcc7e
  yuguo authored Jun 20, 2025
  
  251dcc7e
- [DCU] fix megatron MOE int train issues · 7640a8d4
  yuguo authored Jun 20, 2025
  
  7640a8d4
19 Jun, 2025 2 commits
- [DCU] add TORCH_COMM_CU_NUMS and fix · d6c32078
  yuguo authored Jun 19, 2025
  
  d6c32078
- Fix verify acc failed of blockwise quantizer · 8eff19c9
  wenjh authored Jun 19, 2025
```
Signed-off-by: wenjh <wenjh@sugon.com>
```
  8eff19c9
18 Jun, 2025 2 commits
- Fix vector blockwise acc problem · 8a03ff34
  wenjh authored Jun 18, 2025
```
Signed-off-by: wenjh <wenjh@sugon.com>
```
  8a03ff34
- Fix lack of lds in vector_blockwise · d1bf39cf
  wenjh authored Jun 18, 2025
```
Signed-off-by: wenjh <wenjh@sugon.com>
```
  d1bf39cf
16 Jun, 2025 1 commit
- [DCU] fix in8 simul fp8 fused wgrad accumulation · 3653fbfb
  yuguo authored Jun 16, 2025
  
  3653fbfb
13 Jun, 2025 1 commit
- [DCU] fix blockwise int8 train issues in megatron · ecdd8251
  yuguo authored Jun 13, 2025
  
  ecdd8251
12 Jun, 2025 2 commits
- [INT8] Make int8 rounding instead of truncation · 7f946529
  wenjh authored Jun 12, 2025
```
Signed-off-by: wenjh <wenjh@sugon.com>
```
  7f946529
- [Workaround] Improve acc of vectorise scaling · e2860c76
  wenjh authored Jun 12, 2025
```
Same intention of commit 3e38a2ea

.
This commit is to improve acc.
Signed-off-by: wenjh <wenjh@sugon.com>
```
  e2860c76
11 Jun, 2025 1 commit
- [DCU] add NVTE_TP_OVERLAP_AGGREGATE · b1864da3
  yuguo authored Jun 11, 2025
  
  b1864da3
10 Jun, 2025 1 commit
- [DCU] avoid rtc trans kernel bug (need fix) · fdb21575
  yuguo authored Jun 10, 2025
  
  fdb21575
09 Jun, 2025 4 commits
- Fix build error of test_cublaslt_gemm · 7d2b9c77
  wenjh authored Jun 09, 2025
```
Signed-off-by: wenjh <wenjh@sugon.com>
```
  7d2b9c77
- [DCU] fix · 6d461a10
  yuguo authored Jun 09, 2025
  
  6d461a10
- [DCU] surpport cast master weight to int8 · 0a8072fa
  yuguo authored Jun 09, 2025
  
  0a8072fa
- [TEST] Fix build error of test_cublaslt_gemm · 2cbe1b70
  wenjh authored Jun 09, 2025
```
Signed-off-by: wenjh <wenjh@sugon.com>
```
  2cbe1b70
06 Jun, 2025 1 commit

[Workaround] Use bf16 lds to save fp32 input · 3e38a2ea

wenjh authored Jun 06, 2025



quantize_transpose_vector_blockwise function use lds exceeding 64kb when
input type is fp32. But max size of lds in dcu is 64kb, thus we use lds
as bfp16 for workaround.
Signed-off-by: wenjh <wenjh@sugon.com>

3e38a2ea

05 Jun, 2025 2 commits
- Merge branch 'develop_v2.3' of http://10.16.6.30/dcutoolkit/deeplearing/TransformerEngine · 32184507
  yuguo authored Jun 05, 2025
  
  32184507
- [DCU] support block fp8 simu with int8 for MOE · b7afba08
  yuguo authored Jun 05, 2025
  
  b7afba08
04 Jun, 2025 4 commits
- Merge branch 'develop_v2.3' of http://10.16.6.30/dcutoolkit/deeplearing/TransformerEngine · 1b303e91
  yuguo authored Jun 04, 2025
  
  1b303e91
- fix · 735227cd
  yuguo authored Jun 04, 2025
  
  735227cd
- Merge branch 'develop_v2.3' of... · 129d7526
  yuguo authored Jun 04, 2025
```
Merge branch 'develop_v2.3' of http://10.16.6.30/dcutoolkit/deeplearing/TransformerEngine into develop_v2.3
```
  129d7526
- [DCU] support block fp8 simu with int8 for Dense · f6937668
  yuguo authored Jun 04, 2025
  
  f6937668
28 May, 2025 2 commits
- Merge branch 'develop_v2.3' · 52ba87a1
  wenjh authored May 28, 2025
  
  52ba87a1
- [Workaround] Dtk-25.04.1 need add hip_assert.h for hiprtc · 7e4e1e40
  wenjh authored May 28, 2025
  
  7e4e1e40
27 May, 2025 7 commits
- Merge branch 'develop_v2.3' · 74128807
  wenjh authored May 27, 2025
  
  74128807
- [DTK-25.04.1] Add support of __shfl_*sync apis · 47d6a78f
  wenjh authored May 27, 2025
  
  47d6a78f
- Merge branch 'develop_v2.3' of http://10.16.6.30/dcutoolkit/deeplearing/TransformerEngine · 3273bc20
  yuguo authored May 27, 2025
  
  3273bc20
- [DCU] combine 1f1b needs NVTE_OVERLAP_GRAD_REDUCE · 521f8d3b
  yuguo authored May 27, 2025
  
  521f8d3b
- Merge branch 'develop_v2.3' of http://10.16.6.30/dcutoolkit/deeplearing/TransformerEngine · 75e9ef24
  yuguo authored May 27, 2025
  
  75e9ef24
- Merge branch 'develop_v2.3' of... · 291fcf52
  yuguo authored May 27, 2025
```
Merge branch 'develop_v2.3' of http://10.16.6.30/dcutoolkit/deeplearing/TransformerEngine into develop_v2.3
```
  291fcf52
- [DCU] surpport cumask for tp overlap · c74c8789
  yuguo authored May 27, 2025
  
  c74c8789
26 May, 2025 3 commits
- Merge branch 'develop_v2.3' · 5753c5bb
  wenjh authored May 26, 2025
  
  5753c5bb
- [FP8] Fix build error · 7d0f5b7f
  wenjh authored May 26, 2025
```
Signed-off-by: wenjh <wenjh@sugon.com>
```
  7d0f5b7f
- [DCU] Use ocp fp8(same as nvidia) · 9666d263
  wenjh authored May 26, 2025
```
Use ocp fp8.
Workaround: test_cast_float8blockwise.cu link wrong std::max
Signed-off-by: wenjh <wenjh@sugon.com>
```
  9666d263
23 May, 2025 2 commits
- Merge branch 'develop_v2.3' of http://10.16.6.30/dcutoolkit/deeplearing/TransformerEngine · f9d870f4
  yuguo authored May 23, 2025
  
  f9d870f4
- [DCU] surpport blockwise int8 quant · 80c5079c
  yuguo authored May 23, 2025
  
  80c5079c
22 May, 2025 1 commit
- Merge branch 'develop_v2.3' · 7405fe09
  wenjh authored May 22, 2025
  
  7405fe09