liger-kernel-nightly 0.6.4.dev20251201192513__py3-none-any.whl → 0.6.4.dev20251202094519__py3-none-any.whl

liger_kernel/ops/cross_entropy.py

@@ -10,8 +10,9 @@ from liger_kernel.ops.utils import compare_version
 from liger_kernel.ops.utils import element_mul_kernel
 from liger_kernel.ops.utils import is_hip
 from liger_kernel.utils import infer_device
+from liger_kernel.utils import is_npu_available
 
-if compare_version("triton", operator.ge, "3.0.0"):
+if compare_version("triton", operator.ge, "3.0.0") and not is_npu_available():
     try:
         # typical import path with dispatch available
         from triton.language.extra.libdevice import tanh

liger_kernel/ops/dyt.py

@@ -7,8 +7,10 @@ import triton.language as tl
 from liger_kernel.ops.utils import compare_version
 from liger_kernel.ops.utils import ensure_contiguous
 from liger_kernel.ops.utils import infer_device
+from liger_kernel.utils import get_npu_multi_processor_count
+from liger_kernel.utils import is_npu_available
 
-if compare_version("triton", operator.ge, "3.0.0"):
+if compare_version("triton", operator.ge, "3.0.0") and not is_npu_available():
     try:
         # typical import path with dispatch available
         from triton.language.extra.libdevice import tanh
@@ -125,7 +127,8 @@ def liger_dyt_bwd(dy, x, alpha, gamma, beta):
         NUM_SMS = torch.cuda.get_device_properties(x.device).multi_processor_count
     elif device == "xpu":
         NUM_SMS = torch.xpu.get_device_properties(x.device).gpu_subslice_count
-
+    elif device == "npu":
+        NUM_SMS = get_npu_multi_processor_count()
     da = torch.zeros(NUM_SMS, triton.cdiv(N, 512), dtype=torch.float32, device=x.device)
     dg = torch.empty(NUM_SMS, N, dtype=torch.float32, device=x.device)
     db = torch.empty(NUM_SMS, N, dtype=torch.float32, device=x.device) if HAVE_BETA else None

liger_kernel/ops/fused_add_rms_norm.py

@@ -9,8 +9,10 @@ from liger_kernel.ops.utils import calculate_settings
 from liger_kernel.ops.utils import compare_version
 from liger_kernel.ops.utils import ensure_contiguous
 from liger_kernel.ops.utils import torch_to_triton_dtype
+from liger_kernel.utils import get_npu_multi_processor_count
+from liger_kernel.utils import is_npu_available
 
-if compare_version("triton", operator.ge, "3.0.0"):
+if compare_version("triton", operator.ge, "3.0.0") and not is_npu_available():
     try:
         # typical import path with dispatch available
         from triton.language.extra.libdevice import rsqrt
@@ -293,6 +295,8 @@ def fused_add_rms_norm_backward(dY, dS_out, S, W, RSTD, offset, casting_mode, BL
         sm_count = torch.cuda.get_device_properties(S.device).multi_processor_count
     elif S.device.type == "xpu":
         sm_count = torch.xpu.get_device_properties(S.device).gpu_eu_count
+    elif S.device.type == "npu":
+        sm_count = get_npu_multi_processor_count()
 
     # fp32 for numerical stability especially.
     _dW = torch.empty((sm_count, n_cols), dtype=torch.float32, device=W.device)

liger_kernel/ops/geglu.py

@@ -7,8 +7,9 @@ import triton.language as tl
 from liger_kernel.ops.utils import calculate_settings
 from liger_kernel.ops.utils import compare_version
 from liger_kernel.ops.utils import ensure_contiguous
+from liger_kernel.utils import is_npu_available
 
-if compare_version("triton", operator.ge, "3.0.0"):
+if compare_version("triton", operator.ge, "3.0.0") and not is_npu_available():
     try:
         # typical import path with dispatch available
         from triton.language.extra.libdevice import tanh

liger_kernel/ops/group_norm.py

@@ -6,8 +6,9 @@ import triton.language as tl
 
 from liger_kernel.ops.utils import compare_version
 from liger_kernel.ops.utils import ensure_contiguous
+from liger_kernel.utils import is_npu_available
 
-if compare_version("triton", operator.ge, "3.0.0"):
+if compare_version("triton", operator.ge, "3.0.0") and not is_npu_available():
     try:
         # typical import path with dispatch available
         from triton.language.extra.libdevice import rsqrt

liger_kernel/ops/layer_norm.py

@@ -8,8 +8,9 @@ import triton.language as tl
 from liger_kernel.ops.utils import calculate_settings
 from liger_kernel.ops.utils import compare_version
 from liger_kernel.ops.utils import ensure_contiguous
+from liger_kernel.utils import is_npu_available
 
-if compare_version("triton", operator.ge, "3.0.0"):
+if compare_version("triton", operator.ge, "3.0.0") and not is_npu_available():
     try:
         # typical import path with dispatch available
         from triton.language.extra.libdevice import rsqrt

liger_kernel/ops/poly_norm.py

@@ -7,8 +7,10 @@ import triton.language as tl
 from liger_kernel.ops.utils import calculate_settings
 from liger_kernel.ops.utils import compare_version
 from liger_kernel.ops.utils import ensure_contiguous
+from liger_kernel.utils import get_npu_multi_processor_count
+from liger_kernel.utils import is_npu_available
 
-if compare_version("triton", operator.ge, "3.0.0"):
+if compare_version("triton", operator.ge, "3.0.0") and not is_npu_available():
     try:
         from triton.language.extra.libdevice import rsqrt
     except ModuleNotFoundError:
@@ -290,6 +292,8 @@ def poly_norm_backward(dY, X, W, RSTD, BLOCK_SIZE, num_warps, in_place):
         sm_count = torch.cuda.get_device_properties(X.device).multi_processor_count
     elif X.device.type == "xpu":
         sm_count = torch.xpu.get_device_properties(X.device).gpu_eu_count
+    elif X.device.type == "npu":
+        sm_count = get_npu_multi_processor_count()
 
     # Allocate or reuse gradients
     if in_place is True:

liger_kernel/ops/rms_norm.py

@@ -21,8 +21,10 @@ from liger_kernel.ops.utils import calculate_settings
 from liger_kernel.ops.utils import compare_version
 from liger_kernel.ops.utils import ensure_contiguous
 from liger_kernel.ops.utils import torch_to_triton_dtype
+from liger_kernel.utils import get_npu_multi_processor_count
+from liger_kernel.utils import is_npu_available
 
-if compare_version("triton", operator.ge, "3.0.0"):
+if compare_version("triton", operator.ge, "3.0.0") and not is_npu_available():
     try:
         # typical import path with dispatch available
         from triton.language.extra.libdevice import rsqrt
@@ -450,6 +452,8 @@ def rms_norm_backward(dY, X, W, RSTD, offset, casting_mode, BLOCK_SIZE, num_warp
         sm_count = torch.cuda.get_device_properties(X.device).multi_processor_count
     elif X.device.type == "xpu":
         sm_count = torch.xpu.get_device_properties(X.device).gpu_eu_count
+    elif X.device.type == "npu":
+        sm_count = get_npu_multi_processor_count()
 
     # fp32 for numerical stability especially.
     _dW = torch.empty((sm_count, n_cols), dtype=torch.float32, device=W.device)

liger_kernel/ops/utils.py

@@ -78,6 +78,8 @@ def get_amp_custom_fwd_bwd() -> Callable:
             functools.partial(torch.amp.custom_fwd, device_type=device),
             functools.partial(torch.amp.custom_bwd, device_type=device),
         )
+    if hasattr(torch, "npu") and getattr(torch.npu, "amp", None) is not None:
+        return torch.npu.amp.custom_fwd, torch.npu.amp.custom_bwd
     return torch.cuda.amp.custom_fwd, torch.cuda.amp.custom_bwd
 
 

liger_kernel/transformers/model/gemma3.py

@@ -235,6 +235,7 @@ def multimodal_forward(
         **lm_kwargs,
     )
 
+    shift_labels = lm_kwargs.pop("shift_labels", None)
     hidden_states = outputs[0]
 
     slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep

liger_kernel/transformers/model/paligemma.py

@@ -330,6 +330,7 @@ def lce_forward(
         **lm_kwargs,
     )
 
+    shift_labels = lm_kwargs.pop("shift_labels", None)
     hidden_states = outputs[0]
 
     loss = None

liger_kernel/utils.py

@@ -18,12 +18,37 @@ def infer_device():
     """
     if torch.cuda.is_available():  # Works for both Nvidia and AMD
         return "cuda"
+    # Use Ascend NPU if available (torch.npu)
+    elif is_npu_available():
+        return "npu"
+    # XPU (Intel) if available
     elif torch.xpu.is_available():
         return "xpu"
     else:
         return "cpu"
 
 
+def is_npu_available() -> bool:
+    """Detect Ascend NPU availability."""
+    try:
+        from transformers.utils import is_torch_npu_available
+
+        return is_torch_npu_available()
+    except Exception:
+        return False
+
+
+def get_npu_multi_processor_count() -> int:
+    """Return a heuristic multi-processor count for NPU."""
+    if is_npu_available():
+        NPU_MULTI_PROCESSOR_COUNT = 48
+        dev_props = torch.npu.get_device_properties()
+        # The vector_core_num attribute is supported in the torch.npu v7.2.0 release version.
+        return dev_props.vector_core_num if hasattr(dev_props, "vector_core_num") else NPU_MULTI_PROCESSOR_COUNT
+    # Reasonable default to avoid division by zero
+    return 1
+
+
 def transformers_version_dispatch(
     required_version: str,
     before_fn,

{liger_kernel_nightly-0.6.4.dev20251201192513.dist-info → liger_kernel_nightly-0.6.4.dev20251202094519.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.4.dev20251201192513
+Version: 0.6.4.dev20251202094519
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.4.dev20251201192513.dist-info → liger_kernel_nightly-0.6.4.dev20251202094519.dist-info}/RECORD

@@ -1,6 +1,6 @@
 liger_kernel/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 liger_kernel/env_report.py,sha256=uhdEC8OydxoZlb7B6YYcAaBF3crGFdIck-4cxaW4NJY,1728
-liger_kernel/utils.py,sha256=BQleeZWHSZPNuPcYcoZTOp1kcNEZONZilPP5-AmjgWI,2024
+liger_kernel/utils.py,sha256=TW-OSkuSLrMAPPMZtOsRKBqZ7MCiSrkATB1z_p81Ets,2879
 liger_kernel/chunked_loss/README.md,sha256=0FmkFC3hKBqyoDT5uTlIYmrvRkF-EOCR1y-EBU1LpWU,2248
 liger_kernel/chunked_loss/__init__.py,sha256=J5_jNnzZ4gZmA38W5f_4oab7xMoNk1Xy-yh3X_Xlf-s,714
 liger_kernel/chunked_loss/cosine_similarity_loss.py,sha256=x2nprTHPraU8Ya2NMZtaDk9r-s-1NKJwCTrzQIdmg-8,4680
@@ -17,30 +17,30 @@ liger_kernel/chunked_loss/kto_loss.py,sha256=llVCe6DkcpCo57seGWoMikaQVFApx764jsm
 liger_kernel/chunked_loss/orpo_loss.py,sha256=nu9UYG16dcMw93lvHi4_hYs3Q0FK1KnlmMRj7OpYU8s,4872
 liger_kernel/chunked_loss/simpo_loss.py,sha256=fy2w8KbhMrBv7b1jdIeH3bBFxY52bPQPZb3KwBvmurM,5385
 liger_kernel/ops/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-liger_kernel/ops/cross_entropy.py,sha256=-fd8qVxn_66MGSLs-Gs8yGmWlkET5YAoyb__Bolfz4c,22617
-liger_kernel/ops/dyt.py,sha256=gCLz4S8aul8SY9nvIGaoK67aGb7U9MJRQdo3ONqmQYs,5417
-liger_kernel/ops/fused_add_rms_norm.py,sha256=UBqmlqFCmhSAIpkNKd8rrfXatX7Z4J9bp2dX9A0lrJQ,14017
+liger_kernel/ops/cross_entropy.py,sha256=J0OaI9b3l1H5FPeLft84XGz4g-WhMbrKXGo3wqlWwd0,22692
+liger_kernel/ops/dyt.py,sha256=4XmkCCZaPPM8Tl4QHo6vSF2m68jrwsnjucrbyOJvZpM,5628
+liger_kernel/ops/fused_add_rms_norm.py,sha256=lvwrLsKvoAQqS9KatgBkAyy0Xdecado-g0rvXYXaBak,14237
 liger_kernel/ops/fused_linear_cross_entropy.py,sha256=YepeWqX37gKc1-FUrzkDTzXYdOvmBmfv4KgL__KN_UI,16158
 liger_kernel/ops/fused_linear_jsd.py,sha256=CSoprxb-YcJy-YUKiTcYkxN8sb9h2kdk_iHuncvSV5c,9683
 liger_kernel/ops/fused_neighborhood_attention.py,sha256=vPi5xbnh6wxyZehaqo6Tuilqo2fN5SGDiONjnNmIKqs,35556
-liger_kernel/ops/geglu.py,sha256=r0WSq9E93zzynL44Wh8femzOWK07_SseBM_pJUyxT3s,4144
-liger_kernel/ops/group_norm.py,sha256=qD4D4lSjSgVtO52EBNLC2iTseALRgPgqXE50U2woggk,10837
+liger_kernel/ops/geglu.py,sha256=z-t9OMk3SDL5sJenJjCzhGEeGusL22j3nDjTDEUDAz8,4219
+liger_kernel/ops/group_norm.py,sha256=zoy-TcNkYtKGmGhTFJmnyiG_4Es4ZphpqP8jtUSI6-I,10912
 liger_kernel/ops/grpo_loss.py,sha256=2SyOujtF9I3xiNo4wFf4s6MeiDotE_qeYfRWgj_bOBE,9573
 liger_kernel/ops/jsd.py,sha256=onHp5T3MbvJaVz5Vup7Ww6EQp_HTaZeayTjJk6FgQMY,7042
 liger_kernel/ops/kl_div.py,sha256=ZjGdDLKWksHT9dZ0xF_TDgAkj5cuMTwwT5tr9E-_24o,8734
-liger_kernel/ops/layer_norm.py,sha256=OMaex1MDsM9kaFs0-q5Pnx3DrMVjongQoZ5-iFIOy00,10523
+liger_kernel/ops/layer_norm.py,sha256=-4UEyko9eKgBi5LNmfdEU2hTpJOWVnEy5iYjJkMvHmk,10598
 liger_kernel/ops/llama4_rope.py,sha256=-aqdZzllklTN8b9--e-TsWY_ntGCN8-tyseT4x0bd8s,8223
 liger_kernel/ops/multi_token_attention.py,sha256=Oz_RXDp-OSS_R_HuGmaETHdAJ7Toda_70OfE7TXMUlY,7645
-liger_kernel/ops/poly_norm.py,sha256=MLgI8Ea93fugKibHCUauQ2ASYVXCvpPZe5v3kQZU6po,11152
+liger_kernel/ops/poly_norm.py,sha256=5IdJEZnbbhblkL_X8UhSD4A2CooQbOAZJw8nAekWNs4,11372
 liger_kernel/ops/qwen2vl_mrope.py,sha256=3GExhYpLgB4VUtyZyjRk8XjEur3W4EWF6HQ67ML5vBU,8481
-liger_kernel/ops/rms_norm.py,sha256=kKncS1OCBgZE4jQnBRNpPhL2PjYaVN3rY9V9jEhzHoM,18991
+liger_kernel/ops/rms_norm.py,sha256=2V8qheEvidBm0VxwfOoAnV837F6blmtTgP91VYdGs7c,19211
 liger_kernel/ops/rope.py,sha256=v-7JHRrv-5ImoROkpKfl30WwWI4qTa2tAl7zQeB4ml4,8956
 liger_kernel/ops/softmax.py,sha256=tgORx6MK1IDDtZKqGarj0IPIVjqAIEUXXYPiinhRdtI,5864
 liger_kernel/ops/sparsemax.py,sha256=AeWe1xgkHJFEKWTj2vu_0hj7LztGvjqXAps-QTpCY0U,5087
 liger_kernel/ops/swiglu.py,sha256=D7nd4u_LInwsIRNCDdY77lqnTz8-W5dJrpEAt8zEO_A,3033
 liger_kernel/ops/tiled_mlp.py,sha256=eyMFsFFgHch8a_6R6IYRG24_jqKg5GF_BQUoQuAG8SY,4529
 liger_kernel/ops/tvd.py,sha256=FHJtLQI95ijqgg9UtaHpMAjSCiPxB6CduPwPMcGxelc,6405
-liger_kernel/ops/utils.py,sha256=uoFKQqo-34N2TWQNvXMFywqGiOMMXNEVBxVojzlUAa0,3836
+liger_kernel/ops/utils.py,sha256=kYp84AOA7D9PYrvBUSrNsfQIt8elr_uA9OxCkbfiUFA,3980
 liger_kernel/ops/experimental/embedding.py,sha256=tolj3tItkzpSb30zWqDN2_yX4ectflaQ8HMyKyFIQc8,4172
 liger_kernel/ops/experimental/mm_int8int2.py,sha256=TrS9lpwekrik_w5qE7AhMJD1bcq-OidjtbsW80oZ6IM,13314
 liger_kernel/transformers/__init__.py,sha256=CgwhrY5cdx6OcRgR2ZZJbOIkLswQWPTr-BAaoxDNNOY,10687
@@ -78,7 +78,7 @@ liger_kernel/transformers/model/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm
 liger_kernel/transformers/model/falcon_h1.py,sha256=heUZ4wUt2ATmtBtmv8Rcro3pQl6fV9T0pburjTTW7os,5004
 liger_kernel/transformers/model/gemma.py,sha256=pAri4PYpknsFfkvyo8Ez2NNlqrUDW-KkExUXTGZAcH4,10621
 liger_kernel/transformers/model/gemma2.py,sha256=qa9Ok42vFojVGNmASTH3Ek566Vu507kjd--ZpZDKX9M,12024
-liger_kernel/transformers/model/gemma3.py,sha256=mEV3Kuy-dqfTk_b899Vb-InuD4_DvwH0nm5xgbG-0MM,14911
+liger_kernel/transformers/model/gemma3.py,sha256=ZUrFCc-pfF8jYHV0HsptBr98hx6p2q9ea0kSzVAoFPo,14966
 liger_kernel/transformers/model/glm4.py,sha256=bSp22iPIjsli4-c_usUOsyh1Bs2gIK8X6ynS0azseUs,5900
 liger_kernel/transformers/model/glm4v.py,sha256=dd-BQpccDCp1SbIxcJ5rG8xcwYQK3KOv1Tgm9TGnZc4,6594
 liger_kernel/transformers/model/glm4v_moe.py,sha256=zKhMdOOrRhlrvCSFaeVYfddL1ubpY8edEO91TN81n98,7135
@@ -94,7 +94,7 @@ liger_kernel/transformers/model/mllama.py,sha256=vAHwCm63sn4kpAY0rDGf_N0HR7KRTBV
 liger_kernel/transformers/model/olmo2.py,sha256=-h2bUOeuPfY1MdShdRvq5_wFDHKP4PEimgIl0fL-BT4,5902
 liger_kernel/transformers/model/olmo3.py,sha256=k2zYOlS8U_b5MwjdToB3tDRQ0bH_mWapVQqJcH8-qAo,6007
 liger_kernel/transformers/model/output_classes.py,sha256=0BGXVR4dYQpSHLkSqpRoXuHMryrceGSlTYRu6pvd8ZY,4542
-liger_kernel/transformers/model/paligemma.py,sha256=r0smHLADkEwfLS6d6ArWoSWEeLt2d_8pmgOO5F04b1o,20793
+liger_kernel/transformers/model/paligemma.py,sha256=UAYoKkIMvvix7GG3cSdWaDxVjMp26YsvthJuE7wFf6Y,20848
 liger_kernel/transformers/model/phi3.py,sha256=PT7Kw6yySg-7TsssWfi82eVMN3SWujCqzCqHigAdfeQ,4574
 liger_kernel/transformers/model/qwen2.py,sha256=ojqdJpD3A9A5uCS0N_rSq8gyNYWSsHfuvx3Z3ObC7ss,10686
 liger_kernel/transformers/model/qwen2_5_vl.py,sha256=FbIZDcg9cOr4PtBLNN8yVubN-gu2clndjSIzfi8NMos,6894
@@ -110,9 +110,9 @@ liger_kernel/transformers/trainer/__init__.py,sha256=p7yQfklV8-467qSz_ZMimkbDF7H
 liger_kernel/transformers/trainer/orpo_trainer.py,sha256=tX0h63aOFe3rNqTmk6JpMf75UPo981yzEa6TghnjS0Q,5370
 liger_kernel/triton/__init__.py,sha256=qCiCamzCRv6lpV8IqpAc9YMdNKC7GKurClWceQPnlis,92
 liger_kernel/triton/monkey_patch.py,sha256=Rd0hUHAzDkFfHvnX7-PBaNK5EKnZhtfM_h-fgQH9HPY,1568
-liger_kernel_nightly-0.6.4.dev20251201192513.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel_nightly-0.6.4.dev20251201192513.dist-info/METADATA,sha256=4uWQ3Lbvwt1esBU_QdhJUSgaDg9JMHMwdvtyLrorz8I,25238
-liger_kernel_nightly-0.6.4.dev20251201192513.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel_nightly-0.6.4.dev20251201192513.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
-liger_kernel_nightly-0.6.4.dev20251201192513.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel_nightly-0.6.4.dev20251201192513.dist-info/RECORD,,
+liger_kernel_nightly-0.6.4.dev20251202094519.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel_nightly-0.6.4.dev20251202094519.dist-info/METADATA,sha256=ZgwmbyNXCX9-W_MkkjoTyoMWce_cteqdiqlMCeQ_Tt8,25238
+liger_kernel_nightly-0.6.4.dev20251202094519.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel_nightly-0.6.4.dev20251202094519.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
+liger_kernel_nightly-0.6.4.dev20251202094519.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel_nightly-0.6.4.dev20251202094519.dist-info/RECORD,,