liger-kernel-nightly 0.5.10.dev20250618073949__py3-none-any.whl → 0.5.10.dev20250618193218__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.
- liger_kernel/transformers/monkey_patch.py +6 -2
- liger_kernel/transformers/rms_norm.py +1 -3
- {liger_kernel_nightly-0.5.10.dev20250618073949.dist-info → liger_kernel_nightly-0.5.10.dev20250618193218.dist-info}/METADATA +1 -1
- {liger_kernel_nightly-0.5.10.dev20250618073949.dist-info → liger_kernel_nightly-0.5.10.dev20250618193218.dist-info}/RECORD +8 -8
- {liger_kernel_nightly-0.5.10.dev20250618073949.dist-info → liger_kernel_nightly-0.5.10.dev20250618193218.dist-info}/LICENSE +0 -0
- {liger_kernel_nightly-0.5.10.dev20250618073949.dist-info → liger_kernel_nightly-0.5.10.dev20250618193218.dist-info}/NOTICE +0 -0
- {liger_kernel_nightly-0.5.10.dev20250618073949.dist-info → liger_kernel_nightly-0.5.10.dev20250618193218.dist-info}/WHEEL +0 -0
- {liger_kernel_nightly-0.5.10.dev20250618073949.dist-info → liger_kernel_nightly-0.5.10.dev20250618193218.dist-info}/top_level.txt +0 -0
@@ -54,7 +54,7 @@ def _bind_method_to_module(module, method_name: str, new_method: Callable):
|
|
54
54
|
module.__dict__[method_name] = new_method.__get__(module, module.__class__)
|
55
55
|
|
56
56
|
|
57
|
-
def _patch_rms_norm_module(module, offset=0.0, eps=1e-6, casting_mode="llama", in_place=True):
|
57
|
+
def _patch_rms_norm_module(module, offset=0.0, eps=1e-6, casting_mode="llama", in_place=True, row_mode=None):
|
58
58
|
# Check if the module is a PEFT ModulesToSaveWrapper
|
59
59
|
# If it is, we need to patch the modules_to_save.default and original_modules
|
60
60
|
if PEFT_AVAILABLE and isinstance(module, peft.utils.other.ModulesToSaveWrapper):
|
@@ -64,12 +64,14 @@ def _patch_rms_norm_module(module, offset=0.0, eps=1e-6, casting_mode="llama", i
|
|
64
64
|
getattr(module, "variance_epsilon", None) or getattr(module, "eps", None) or eps
|
65
65
|
)
|
66
66
|
module.modules_to_save.default.in_place = in_place
|
67
|
+
module.modules_to_save.default.row_mode = row_mode
|
67
68
|
module.original_module.offset = offset
|
68
69
|
module.original_module.casting_mode = casting_mode
|
69
70
|
module.original_module.variance_epsilon = (
|
70
71
|
getattr(module, "variance_epsilon", None) or getattr(module, "eps", None) or eps
|
71
72
|
)
|
72
73
|
module.original_module.in_place = in_place
|
74
|
+
module.original_module.row_mode = row_mode
|
73
75
|
_bind_method_to_module(module.modules_to_save.default, "forward", LigerRMSNorm.forward)
|
74
76
|
_bind_method_to_module(module.modules_to_save.default, "extra_repr", LigerRMSNorm.extra_repr)
|
75
77
|
_bind_method_to_module(module.original_module, "forward", LigerRMSNorm.forward)
|
@@ -81,6 +83,7 @@ def _patch_rms_norm_module(module, offset=0.0, eps=1e-6, casting_mode="llama", i
|
|
81
83
|
module.casting_mode = casting_mode
|
82
84
|
module.variance_epsilon = getattr(module, "variance_epsilon", None) or getattr(module, "eps", None) or eps
|
83
85
|
module.in_place = in_place
|
86
|
+
module.row_mode = row_mode
|
84
87
|
_bind_method_to_module(module, "forward", LigerRMSNorm.forward)
|
85
88
|
_bind_method_to_module(module, "extra_repr", LigerRMSNorm.extra_repr)
|
86
89
|
module.__class__.__name__ = LigerRMSNorm.__name__
|
@@ -1208,7 +1211,8 @@ def apply_liger_kernel_to_qwen3_moe(
|
|
1208
1211
|
_patch_rms_norm_module(base_model.norm)
|
1209
1212
|
for decoder_layer in base_model.layers:
|
1210
1213
|
if swiglu:
|
1211
|
-
|
1214
|
+
for mlp_expert in decoder_layer.mlp.experts:
|
1215
|
+
_patch_swiglu_module(mlp_expert, LigerQwen3MoeSwiGLUMLP)
|
1212
1216
|
if rms_norm:
|
1213
1217
|
_patch_rms_norm_module(decoder_layer.input_layernorm)
|
1214
1218
|
_patch_rms_norm_module(decoder_layer.post_attention_layernorm)
|
@@ -41,9 +41,7 @@ class LigerRMSNorm(nn.Module):
|
|
41
41
|
)
|
42
42
|
|
43
43
|
def extra_repr(self):
|
44
|
-
return (
|
45
|
-
f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}, offset={self.offset}, in_place={self.in_place}"
|
46
|
-
)
|
44
|
+
return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}, offset={self.offset}, in_place={self.in_place}, row_mode={self.row_mode}"
|
47
45
|
|
48
46
|
|
49
47
|
class LigerRMSNormForGemma(LigerRMSNorm):
|
@@ -53,10 +53,10 @@ liger_kernel/transformers/grpo_loss.py,sha256=uAkUNKSnUGEOqa82L9w2e6AI1kcmG8K45-
|
|
53
53
|
liger_kernel/transformers/jsd.py,sha256=DGqRnxIZxsvxo0_tbbxX3b-sDbDjC_yKufyRIHCcScY,2979
|
54
54
|
liger_kernel/transformers/kl_div.py,sha256=WLffFbh1EExD2Eb1F7lN11fo9JJC-0751WJjZAF1Fj8,409
|
55
55
|
liger_kernel/transformers/layer_norm.py,sha256=c9pk3PEasOKYR0rhe5e5nNrnYKVCEW4VC8S6LpCq9EQ,906
|
56
|
-
liger_kernel/transformers/monkey_patch.py,sha256=
|
56
|
+
liger_kernel/transformers/monkey_patch.py,sha256=FacsJGO8MUbn1I7HttRAlDGhLtD8jjRTXSAwvI0iPrg,77754
|
57
57
|
liger_kernel/transformers/multi_token_attention.py,sha256=l9VDICK0dfmifUDW668hGscP8AHq2rYcM2oGUa3baRQ,1751
|
58
58
|
liger_kernel/transformers/qwen2vl_mrope.py,sha256=5EwSqrMdsL9MYspeBMXBsNJKvH0MOmRrtJXAJlnnlOI,1047
|
59
|
-
liger_kernel/transformers/rms_norm.py,sha256=
|
59
|
+
liger_kernel/transformers/rms_norm.py,sha256=vkekcvTeWY8vL4H6hg3t0XeY0Ew_3OFMPHuzqlxPPVw,2719
|
60
60
|
liger_kernel/transformers/rope.py,sha256=ZTrTORSAyfcFIKjk6XEeYmk4ROH7xXED9L4g2NFntlE,999
|
61
61
|
liger_kernel/transformers/softmax.py,sha256=yadlAgE4V2JByMwrDDa2s5SUBp8Jgd57xwnVvAWoBaI,264
|
62
62
|
liger_kernel/transformers/sparsemax.py,sha256=0lQA0UEOs4mu8CMruZ3VLhImxQVXJWhPsAKUsYA7vj8,403
|
@@ -87,9 +87,9 @@ liger_kernel/transformers/trainer/__init__.py,sha256=p7yQfklV8-467qSz_ZMimkbDF7H
|
|
87
87
|
liger_kernel/transformers/trainer/orpo_trainer.py,sha256=tX0h63aOFe3rNqTmk6JpMf75UPo981yzEa6TghnjS0Q,5370
|
88
88
|
liger_kernel/triton/__init__.py,sha256=qCiCamzCRv6lpV8IqpAc9YMdNKC7GKurClWceQPnlis,92
|
89
89
|
liger_kernel/triton/monkey_patch.py,sha256=Rd0hUHAzDkFfHvnX7-PBaNK5EKnZhtfM_h-fgQH9HPY,1568
|
90
|
-
liger_kernel_nightly-0.5.10.
|
91
|
-
liger_kernel_nightly-0.5.10.
|
92
|
-
liger_kernel_nightly-0.5.10.
|
93
|
-
liger_kernel_nightly-0.5.10.
|
94
|
-
liger_kernel_nightly-0.5.10.
|
95
|
-
liger_kernel_nightly-0.5.10.
|
90
|
+
liger_kernel_nightly-0.5.10.dev20250618193218.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
|
91
|
+
liger_kernel_nightly-0.5.10.dev20250618193218.dist-info/METADATA,sha256=btKwX4xwU1mZofQJGTzF3RQw4MjRayIRoCXbuag9aUM,24358
|
92
|
+
liger_kernel_nightly-0.5.10.dev20250618193218.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
|
93
|
+
liger_kernel_nightly-0.5.10.dev20250618193218.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
|
94
|
+
liger_kernel_nightly-0.5.10.dev20250618193218.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
|
95
|
+
liger_kernel_nightly-0.5.10.dev20250618193218.dist-info/RECORD,,
|
File without changes
|
File without changes
|
File without changes
|