liger-kernel-nightly 0.6.3.dev20251121213521__py3-none-any.whl → 0.6.4.dev20251201192513__py3-none-any.whl

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -349,7 +349,8 @@ def _block_rms_norm_backward_kernel(
349
349
 
350
350
  # calculate the gradient of W
351
351
  if casting_mode == _CASTING_MODE_LLAMA:
352
- dW_row += tl.sum(dY_row * (X_row * rstd_row[:, None]).to(X_dtype), 0)
352
+ # TODO(tcc): use tl.sum(..., dtype=tl.float32) once we upgrade to triton>=3.3.0
353
+ dW_row += tl.sum((dY_row * (X_row * rstd_row[:, None]).to(X_dtype)).to(tl.float32), 0)
353
354
  else:
354
355
  # here X_row is already in fp32 (see previous if block)
355
356
  dW_row += tl.sum(dY_row * (X_row * rstd_row[:, None]), 0)
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.1
2
2
  Name: liger_kernel_nightly
3
- Version: 0.6.3.dev20251121213521
3
+ Version: 0.6.4.dev20251201192513
4
4
  Summary: Efficient Triton kernels for LLM Training
5
5
  License: BSD 2-CLAUSE LICENSE
6
6
  Copyright 2024 LinkedIn Corporation
@@ -33,7 +33,7 @@ liger_kernel/ops/llama4_rope.py,sha256=-aqdZzllklTN8b9--e-TsWY_ntGCN8-tyseT4x0bd
33
33
  liger_kernel/ops/multi_token_attention.py,sha256=Oz_RXDp-OSS_R_HuGmaETHdAJ7Toda_70OfE7TXMUlY,7645
34
34
  liger_kernel/ops/poly_norm.py,sha256=MLgI8Ea93fugKibHCUauQ2ASYVXCvpPZe5v3kQZU6po,11152
35
35
  liger_kernel/ops/qwen2vl_mrope.py,sha256=3GExhYpLgB4VUtyZyjRk8XjEur3W4EWF6HQ67ML5vBU,8481
36
- liger_kernel/ops/rms_norm.py,sha256=DtvsWN5YktFAoc0JYSAwVeoZfryBFJlX-ipU7ooP01A,18891
36
+ liger_kernel/ops/rms_norm.py,sha256=kKncS1OCBgZE4jQnBRNpPhL2PjYaVN3rY9V9jEhzHoM,18991
37
37
  liger_kernel/ops/rope.py,sha256=v-7JHRrv-5ImoROkpKfl30WwWI4qTa2tAl7zQeB4ml4,8956
38
38
  liger_kernel/ops/softmax.py,sha256=tgORx6MK1IDDtZKqGarj0IPIVjqAIEUXXYPiinhRdtI,5864
39
39
  liger_kernel/ops/sparsemax.py,sha256=AeWe1xgkHJFEKWTj2vu_0hj7LztGvjqXAps-QTpCY0U,5087
@@ -110,9 +110,9 @@ liger_kernel/transformers/trainer/__init__.py,sha256=p7yQfklV8-467qSz_ZMimkbDF7H
110
110
  liger_kernel/transformers/trainer/orpo_trainer.py,sha256=tX0h63aOFe3rNqTmk6JpMf75UPo981yzEa6TghnjS0Q,5370
111
111
  liger_kernel/triton/__init__.py,sha256=qCiCamzCRv6lpV8IqpAc9YMdNKC7GKurClWceQPnlis,92
112
112
  liger_kernel/triton/monkey_patch.py,sha256=Rd0hUHAzDkFfHvnX7-PBaNK5EKnZhtfM_h-fgQH9HPY,1568
113
- liger_kernel_nightly-0.6.3.dev20251121213521.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
114
- liger_kernel_nightly-0.6.3.dev20251121213521.dist-info/METADATA,sha256=__p46-m1Fnwjp4mS78P8H7l3vFCXzyqw-MPEgfqbZZA,25238
115
- liger_kernel_nightly-0.6.3.dev20251121213521.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
116
- liger_kernel_nightly-0.6.3.dev20251121213521.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
117
- liger_kernel_nightly-0.6.3.dev20251121213521.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
118
- liger_kernel_nightly-0.6.3.dev20251121213521.dist-info/RECORD,,
113
+ liger_kernel_nightly-0.6.4.dev20251201192513.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
114
+ liger_kernel_nightly-0.6.4.dev20251201192513.dist-info/METADATA,sha256=4uWQ3Lbvwt1esBU_QdhJUSgaDg9JMHMwdvtyLrorz8I,25238
115
+ liger_kernel_nightly-0.6.4.dev20251201192513.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
116
+ liger_kernel_nightly-0.6.4.dev20251201192513.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
117
+ liger_kernel_nightly-0.6.4.dev20251201192513.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
118
+ liger_kernel_nightly-0.6.4.dev20251201192513.dist-info/RECORD,,