liger-kernel-nightly 0.6.4.dev20251202054858__py3-none-any.whl → 0.6.4.dev20260107181130__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.

Potentially problematic release.

This version of liger-kernel-nightly might be problematic. Click here for more details.

Files changed (58) hide show
  1. liger_kernel/chunked_loss/cosine_similarity_loss.py +7 -1
  2. liger_kernel/chunked_loss/fused_linear_distillation.py +10 -3
  3. liger_kernel/chunked_loss/jsd_loss.py +21 -6
  4. liger_kernel/ops/__init__.py +141 -0
  5. liger_kernel/ops/backends/README.md +151 -0
  6. liger_kernel/ops/backends/__init__.py +13 -0
  7. liger_kernel/ops/backends/_ascend/__init__.py +5 -0
  8. liger_kernel/ops/backends/_ascend/ascend-ub-manager-design.md +485 -0
  9. liger_kernel/ops/backends/_ascend/ops/__init__.py +43 -0
  10. liger_kernel/ops/backends/_ascend/ops/geglu.py +244 -0
  11. liger_kernel/ops/backends/_ascend/ops/qwen2vl_mrope.py +285 -0
  12. liger_kernel/ops/backends/_ascend/ops/rope.py +290 -0
  13. liger_kernel/ops/backends/_ascend/ops/swiglu.py +142 -0
  14. liger_kernel/ops/backends/_ascend/ub_manager.py +349 -0
  15. liger_kernel/ops/backends/registry.py +61 -0
  16. liger_kernel/ops/cross_entropy.py +12 -3
  17. liger_kernel/ops/fused_linear_cross_entropy.py +2 -1
  18. liger_kernel/ops/geglu.py +3 -2
  19. liger_kernel/ops/rms_norm.py +126 -49
  20. liger_kernel/ops/utils.py +12 -0
  21. liger_kernel/transformers/__init__.py +3 -0
  22. liger_kernel/transformers/auto_model.py +21 -0
  23. liger_kernel/transformers/cross_entropy.py +1 -1
  24. liger_kernel/transformers/dyt.py +1 -1
  25. liger_kernel/transformers/experimental/embedding.py +1 -1
  26. liger_kernel/transformers/functional.py +20 -20
  27. liger_kernel/transformers/fused_add_rms_norm.py +1 -1
  28. liger_kernel/transformers/fused_linear_cross_entropy.py +1 -1
  29. liger_kernel/transformers/fused_linear_jsd.py +1 -1
  30. liger_kernel/transformers/fused_neighborhood_attention.py +1 -1
  31. liger_kernel/transformers/geglu.py +1 -1
  32. liger_kernel/transformers/group_norm.py +1 -1
  33. liger_kernel/transformers/grpo_loss.py +1 -1
  34. liger_kernel/transformers/jsd.py +1 -1
  35. liger_kernel/transformers/kl_div.py +1 -1
  36. liger_kernel/transformers/layer_norm.py +1 -1
  37. liger_kernel/transformers/llama4_rope.py +1 -1
  38. liger_kernel/transformers/model/gemma3.py +1 -0
  39. liger_kernel/transformers/model/gpt_oss.py +211 -0
  40. liger_kernel/transformers/model/paligemma.py +1 -0
  41. liger_kernel/transformers/monkey_patch.py +118 -39
  42. liger_kernel/transformers/multi_token_attention.py +1 -1
  43. liger_kernel/transformers/poly_norm.py +1 -1
  44. liger_kernel/transformers/qwen2vl_mrope.py +1 -1
  45. liger_kernel/transformers/rms_norm.py +8 -3
  46. liger_kernel/transformers/rope.py +28 -27
  47. liger_kernel/transformers/softmax.py +1 -1
  48. liger_kernel/transformers/sparsemax.py +1 -1
  49. liger_kernel/transformers/swiglu.py +1 -1
  50. liger_kernel/transformers/tiled_mlp.py +3 -3
  51. liger_kernel/transformers/tvd.py +1 -1
  52. liger_kernel/utils.py +27 -0
  53. {liger_kernel_nightly-0.6.4.dev20251202054858.dist-info → liger_kernel_nightly-0.6.4.dev20260107181130.dist-info}/METADATA +9 -3
  54. {liger_kernel_nightly-0.6.4.dev20251202054858.dist-info → liger_kernel_nightly-0.6.4.dev20260107181130.dist-info}/RECORD +58 -46
  55. {liger_kernel_nightly-0.6.4.dev20251202054858.dist-info → liger_kernel_nightly-0.6.4.dev20260107181130.dist-info}/LICENSE +0 -0
  56. {liger_kernel_nightly-0.6.4.dev20251202054858.dist-info → liger_kernel_nightly-0.6.4.dev20260107181130.dist-info}/NOTICE +0 -0
  57. {liger_kernel_nightly-0.6.4.dev20251202054858.dist-info → liger_kernel_nightly-0.6.4.dev20260107181130.dist-info}/WHEEL +0 -0
  58. {liger_kernel_nightly-0.6.4.dev20251202054858.dist-info → liger_kernel_nightly-0.6.4.dev20260107181130.dist-info}/top_level.txt +0 -0
liger_kernel/transformers/rope.py CHANGED
@@ -1,9 +1,8 @@
1
- from typing import Optional
2
1
  from typing import Tuple
3
2
 
4
3
  import torch
5
4
 
6
- from liger_kernel.ops.rope import LigerRopeFunction
5
+ from liger_kernel.ops import LigerRopeFunction
7
6
 
8
7
 
9
8
  def liger_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
@@ -25,39 +24,41 @@ def liger_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
25
24
  return LigerRopeFunction.apply(q, k, cos, sin, position_ids, unsqueeze_dim)
26
25
 
27
26
 
28
- def liger_rotary_pos_emb_with_cast(
27
+ def liger_rotary_pos_emb_vision(
29
28
  q: torch.Tensor,
30
29
  k: torch.Tensor,
31
30
  cos: torch.Tensor,
32
31
  sin: torch.Tensor,
33
- position_ids: Optional[torch.Tensor] = None,
34
- unsqueeze_dim: int = 1,
35
32
  ) -> Tuple[torch.Tensor, torch.Tensor]:
33
+ """
34
+ Modified version of liger_rotary_pos_emb for qwen3_vl's apply_rotary_pos_emb_vision function.
35
+ Manually tranposed the input and output to match the expected shape for liger_rotary_pos_emb.
36
+ Reference: https://https://github.com/huggingface/transformers/blob/v5.0.0rc0/src/transformers/models/qwen3_vl/modeling_qwen3_vl.py#L116
37
+
38
+ Args:
39
+ q (torch.Tensor): The query tensor of shape (seq_length, num_heads, head_dim),
40
+ with stride (num_heads * head_dim, head_dim, 1).
41
+ k (torch.Tensor): The query tensor of shape (seq_length, num_heads, head_dim),
42
+ with stride (num_heads * head_dim, head_dim, 1). Same as q.
43
+ cos (torch.Tensor): The cosine tensor of shape (seq_length, head_dim).
44
+ sin (torch.Tensor): The sine tensor of shape (seq_length, head_dim).
45
+
46
+ Returns:
47
+ Tuple[torch.Tensor, torch.Tensor]: The query and key tensors with the same shape and stride as inputs.
48
+ """
36
49
  orig_q_dtype, orig_k_dtype = q.dtype, k.dtype
37
50
 
38
- q32 = q.to(torch.float32)
39
- k32 = k.to(torch.float32)
51
+ # tranpose to (1, num_heads, seq_length, head_dim) and cast to float32 to match liger_rotary_pos_emb input shape
52
+ # also unsqueeze for batch dim
53
+ q32 = q.to(torch.float32).unsqueeze(0).transpose(1, 2)
54
+ k32 = k.to(torch.float32).unsqueeze(0).transpose(1, 2)
40
55
  cos32 = cos.to(torch.float32)
41
56
  sin32 = sin.to(torch.float32)
42
57
 
43
- q_out, k_out = liger_rotary_pos_emb(q32, k32, cos32, sin32, position_ids=position_ids, unsqueeze_dim=unsqueeze_dim)
44
- return q_out.to(orig_q_dtype), k_out.to(orig_k_dtype)
45
-
46
-
47
- def liger_rotary_pos_emb_with_cast_and_leading_batch(
48
- q: torch.Tensor,
49
- k: torch.Tensor,
50
- cos: torch.Tensor,
51
- sin: torch.Tensor,
52
- position_ids: Optional[torch.Tensor] = None,
53
- unsqueeze_dim: int = 1,
54
- ) -> Tuple[torch.Tensor, torch.Tensor]:
55
- orig_q_dtype, orig_k_dtype = q.dtype, k.dtype
56
-
57
- q32 = q.to(torch.float32).unsqueeze(0)
58
- k32 = k.to(torch.float32).unsqueeze(0)
59
- cos32 = cos.to(torch.float32).unsqueeze(0)
60
- sin32 = sin.to(torch.float32).unsqueeze(0)
58
+ q_out, k_out = liger_rotary_pos_emb(q32, k32, cos32, sin32)
61
59
 
62
- q_out, k_out = liger_rotary_pos_emb(q32, k32, cos32, sin32, position_ids=position_ids, unsqueeze_dim=unsqueeze_dim)
63
- return q_out.to(orig_q_dtype).squeeze(0), k_out.to(orig_k_dtype).squeeze(0)
60
+ # transpose back to (seq_length, num_heads, head_dim) and cast back to original dtype
61
+ # also squeeze out batch dim
62
+ q_out = q_out.transpose(1, 2).squeeze(0).to(orig_q_dtype)
63
+ k_out = k_out.transpose(1, 2).squeeze(0).to(orig_k_dtype)
64
+ return q_out, k_out
liger_kernel/transformers/softmax.py CHANGED
@@ -1,7 +1,7 @@
1
1
  import torch
2
2
  import torch.nn as nn
3
3
 
4
- from liger_kernel.ops.softmax import LigerSoftmaxFunction
4
+ from liger_kernel.ops import LigerSoftmaxFunction
5
5
 
6
6
 
7
7
  class LigerSoftmax(nn.Module):
liger_kernel/transformers/sparsemax.py CHANGED
@@ -1,7 +1,7 @@
1
1
  import torch
2
2
  import torch.nn as nn
3
3
 
4
- from liger_kernel.ops.sparsemax import LigerSparsemaxFunction
4
+ from liger_kernel.ops import LigerSparsemaxFunction
5
5
 
6
6
 
7
7
  class LigerSparsemax(nn.Module):
liger_kernel/transformers/swiglu.py CHANGED
@@ -1,6 +1,6 @@
1
1
  import torch.nn as nn
2
2
 
3
- from liger_kernel.ops.swiglu import LigerSiLUMulFunction
3
+ from liger_kernel.ops import LigerSiLUMulFunction
4
4
 
5
5
 
6
6
  class LigerSwiGLUMLP(nn.Module):
liger_kernel/transformers/tiled_mlp.py CHANGED
@@ -2,9 +2,9 @@ from typing import Optional
2
2
 
3
3
  import torch.nn as nn
4
4
 
5
- from liger_kernel.ops.geglu import LigerGELUMulFunction
6
- from liger_kernel.ops.swiglu import LigerSiLUMulFunction
7
- from liger_kernel.ops.tiled_mlp import apply_tiled_mlp
5
+ from liger_kernel.ops import LigerGELUMulFunction
6
+ from liger_kernel.ops import LigerSiLUMulFunction
7
+ from liger_kernel.ops import apply_tiled_mlp
8
8
 
9
9
 
10
10
  class LigerTiledGEGLUMLP(nn.Module):
liger_kernel/transformers/tvd.py CHANGED
@@ -1,6 +1,6 @@
1
1
  import torch.nn as nn
2
2
 
3
- from liger_kernel.ops.tvd import LigerTVDLossFunction
3
+ from liger_kernel.ops import LigerTVDLossFunction
4
4
 
5
5
 
6
6
  class LigerTVDLoss(nn.Module):
liger_kernel/utils.py CHANGED
@@ -12,6 +12,33 @@ def is_peft_available():
12
12
  return PEFT_AVAILABLE
13
13
 
14
14
 
15
+ def infer_comm_backend():
16
+ """
17
+ Get communication backend name based on the environment.
18
+ """
19
+ if torch.distributed.is_nccl_available():
20
+ # Works for Nvidia
21
+ # TODO: nccl may not work for AMD decices that may require use of rccl.
22
+ return "nccl"
23
+ elif is_npu_available():
24
+ # Use Ascend NPU if available (torch.npu)
25
+ # Ascend is not standard torch backend and requires extension.
26
+ # Assume that it is installed if NPUs are being used in
27
+ # multi device environment.
28
+ return "ascend"
29
+ # XPU (Intel) if available
30
+ elif torch.distributed.distributed_c10d.is_xccl_available():
31
+ return "xccl"
32
+ elif torch.distributed.is_mpi_available():
33
+ # CPU backend, first option
34
+ return "mpi"
35
+ elif torch.distributed.is_gloo_available():
36
+ # CPU backend, backup option
37
+ return "gloo"
38
+ else:
39
+ raise RuntimeError("There is no distributed backend available.")
40
+
41
+
15
42
  def infer_device():
16
43
  """
17
44
  Get current device name based on available devices
{liger_kernel_nightly-0.6.4.dev20251202054858.dist-info → liger_kernel_nightly-0.6.4.dev20260107181130.dist-info}/METADATA RENAMED
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.1
2
2
  Name: liger_kernel_nightly
3
- Version: 0.6.4.dev20251202054858
3
+ Version: 0.6.4.dev20260107181130
4
4
  Summary: Efficient Triton kernels for LLM Training
5
5
  License: BSD 2-CLAUSE LICENSE
6
6
  Copyright 2024 LinkedIn Corporation
@@ -79,8 +79,8 @@ Requires-Dist: torchvision>=0.20; extra == "dev"
79
79
  </a>
80
80
  </td>
81
81
  <td style="padding: 10px;">
82
- <a href="https://discord.gg/gpumode">
83
- <img src="https://dcbadge.limes.pink/api/server/gpumode?style=flat" alt="Join Our Discord">
82
+ <a href="https://discord.gg/X4MaxPgA">
83
+ <img src="https://dcbadge.limes.pink/api/server/https://discord.gg/X4MaxPgA?style=flat" alt="Join Our Discord">
84
84
  </a>
85
85
  </td>
86
86
  </tr>
@@ -95,6 +95,7 @@ Requires-Dist: torchvision>=0.20; extra == "dev"
95
95
  <details>
96
96
  <summary>Latest News 🔥</summary>
97
97
 
98
+ - [2025/12/19] We announced a liger kernel discord channel at https://discord.gg/X4MaxPgA; We will be hosting Liger Kernel x Triton China Meetup in mid of January 2026
98
99
  - [2025/03/06] We release a joint blog post on TorchTune × Liger - [Peak Performance, Minimized Memory: Optimizing torchtune’s performance with torch.compile & Liger Kernel](https://pytorch.org/blog/peak-performance-minimized-memory/)
99
100
  - [2024/12/11] We release [v0.5.0](https://github.com/linkedin/Liger-Kernel/releases/tag/v0.5.0): 80% more memory efficient post training losses (DPO, ORPO, CPO, etc)!
100
101
  - [2024/12/5] We release LinkedIn Engineering Blog - [Liger-Kernel: Empowering an open source ecosystem of Triton Kernels for Efficient LLM Training](https://www.linkedin.com/blog/engineering/open-source/liger-kernel-open-source-ecosystem-for-efficient-llm-training)
@@ -113,6 +114,8 @@ We've also added optimized Post-Training kernels that deliver **up to 80% memory
113
114
 
114
115
  You can view the documentation site for additional installation, usage examples, and API references:https://linkedin.github.io/Liger-Kernel/
115
116
 
117
+ You can view the Liger Kernel Technical Report: https://openreview.net/forum?id=36SjAIT42G
118
+
116
119
  ## Supercharge Your Model with Liger Kernel
117
120
 
118
121
  ![Banner](https://raw.githubusercontent.com/linkedin/Liger-Kernel/main/docs/images/banner.GIF)
@@ -312,6 +315,7 @@ loss.backward()
312
315
  | OLMo2 | `liger_kernel.transformers.apply_liger_kernel_to_olmo2` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
313
316
  | Olmo3 | `liger_kernel.transformers.apply_liger_kernel_to_olmo3` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
314
317
  | GLM-4 | `liger_kernel.transformers.apply_liger_kernel_to_glm4` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
318
+ | GPT-OSS | `liger_kernel.transformers.apply_liger_kernel_to_gpt_oss` | RoPE, RMSNorm, CrossEntropyLoss, FusedLinearCrossEntropy |
315
319
  | InternVL3 | `liger_kernel.transformers.apply_liger_kernel_to_internvl` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
316
320
  | HunyuanV1 | `liger_kernel.transformers.apply_liger_kernel_to_hunyuan_v1_dense` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
317
321
  | HunyuanV1 MoE | `liger_kernel.transformers.apply_liger_kernel_to_hunyuan_v1_moe` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
@@ -441,3 +445,5 @@ url={https://openreview.net/forum?id=36SjAIT42G}
441
445
  ↑ Back to Top ↑
442
446
  </a>
443
447
  </p>
448
+
449
+
{liger_kernel_nightly-0.6.4.dev20251202054858.dist-info → liger_kernel_nightly-0.6.4.dev20260107181130.dist-info}/RECORD RENAMED
@@ -1,29 +1,29 @@
1
1
  liger_kernel/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
2
2
  liger_kernel/env_report.py,sha256=uhdEC8OydxoZlb7B6YYcAaBF3crGFdIck-4cxaW4NJY,1728
3
- liger_kernel/utils.py,sha256=TW-OSkuSLrMAPPMZtOsRKBqZ7MCiSrkATB1z_p81Ets,2879
3
+ liger_kernel/utils.py,sha256=1SXJmyVtn-aoyUkA-Acsur_hdHqtxoGqN4v4Vk820bE,3845
4
4
  liger_kernel/chunked_loss/README.md,sha256=0FmkFC3hKBqyoDT5uTlIYmrvRkF-EOCR1y-EBU1LpWU,2248
5
5
  liger_kernel/chunked_loss/__init__.py,sha256=J5_jNnzZ4gZmA38W5f_4oab7xMoNk1Xy-yh3X_Xlf-s,714
6
- liger_kernel/chunked_loss/cosine_similarity_loss.py,sha256=x2nprTHPraU8Ya2NMZtaDk9r-s-1NKJwCTrzQIdmg-8,4680
6
+ liger_kernel/chunked_loss/cosine_similarity_loss.py,sha256=h8lPAkw8oYPUUBZ5YEG2tMMmQ7XkWnOo7r29A5vx-Eg,4759
7
7
  liger_kernel/chunked_loss/cpo_loss.py,sha256=Gzz1eU4kgcbdubFVRy55e8A1Cr-r45UgNicXwZIjmBU,5454
8
8
  liger_kernel/chunked_loss/dpo_loss.py,sha256=I83khNs3QQjuhr8U3NIOAACkbse6DNiBV-TulPZ0lXw,9006
9
9
  liger_kernel/chunked_loss/functional.py,sha256=-XPDbLml9dHmvoSU2VNTUrBDFehuzvuAGPikVetBMtI,1132
10
- liger_kernel/chunked_loss/fused_linear_distillation.py,sha256=yRtolfFGfKB-SxGQQyF68GYXd11Zlvh1InLdGeWNFIE,12652
10
+ liger_kernel/chunked_loss/fused_linear_distillation.py,sha256=Bjdxnjfg-GwYvMJ102juP06gFMlbkxYPpvv7cV_mZcg,12912
11
11
  liger_kernel/chunked_loss/fused_linear_ppo.py,sha256=baU19PwqO1FTVxwlB-eyJv6gOLtL7baXGzSncYQ8Ktc,14296
12
12
  liger_kernel/chunked_loss/fused_linear_preference.py,sha256=FIH85uUXAOgYx5Ax8MjFhJHVu-2pKtY7wSegd0zSyyY,18336
13
13
  liger_kernel/chunked_loss/fused_linear_unpaired_preference.py,sha256=RiuK3UtRwH9T6jZ36sA8Urj-TVuOLOO2syLg_JOQapY,13437
14
14
  liger_kernel/chunked_loss/grpo_loss.py,sha256=bmuZaNgqNbJ5pJGFDXWE-B4BGYF7xWVSN15UyCfuq_s,13079
15
- liger_kernel/chunked_loss/jsd_loss.py,sha256=G0RghPYYelyZ6DOEiwS8we9TT5MY2iHpiFqzZ2Xy87g,8038
15
+ liger_kernel/chunked_loss/jsd_loss.py,sha256=NJKmJ76_-kI875ZkC4hQfC4nAvPNCj4ZsNyDNsfD74k,8761
16
16
  liger_kernel/chunked_loss/kto_loss.py,sha256=llVCe6DkcpCo57seGWoMikaQVFApx764jsmSbQyqwQY,7529
17
17
  liger_kernel/chunked_loss/orpo_loss.py,sha256=nu9UYG16dcMw93lvHi4_hYs3Q0FK1KnlmMRj7OpYU8s,4872
18
18
  liger_kernel/chunked_loss/simpo_loss.py,sha256=fy2w8KbhMrBv7b1jdIeH3bBFxY52bPQPZb3KwBvmurM,5385
19
- liger_kernel/ops/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
20
- liger_kernel/ops/cross_entropy.py,sha256=J0OaI9b3l1H5FPeLft84XGz4g-WhMbrKXGo3wqlWwd0,22692
19
+ liger_kernel/ops/__init__.py,sha256=F3m9qlXbgttykKEBsrMFf1WyK_0H8CKqLuDnFRR-cvc,7237
20
+ liger_kernel/ops/cross_entropy.py,sha256=DnXFRZ9TGN1SnEo8xGBFFPLNQaen8aLVNPJ1em-LbK4,22910
21
21
  liger_kernel/ops/dyt.py,sha256=4XmkCCZaPPM8Tl4QHo6vSF2m68jrwsnjucrbyOJvZpM,5628
22
22
  liger_kernel/ops/fused_add_rms_norm.py,sha256=lvwrLsKvoAQqS9KatgBkAyy0Xdecado-g0rvXYXaBak,14237
23
- liger_kernel/ops/fused_linear_cross_entropy.py,sha256=YepeWqX37gKc1-FUrzkDTzXYdOvmBmfv4KgL__KN_UI,16158
23
+ liger_kernel/ops/fused_linear_cross_entropy.py,sha256=1gx2qljre9PVc861iknFnNCGC-P35D2w1cc_yMDO9ow,16239
24
24
  liger_kernel/ops/fused_linear_jsd.py,sha256=CSoprxb-YcJy-YUKiTcYkxN8sb9h2kdk_iHuncvSV5c,9683
25
25
  liger_kernel/ops/fused_neighborhood_attention.py,sha256=vPi5xbnh6wxyZehaqo6Tuilqo2fN5SGDiONjnNmIKqs,35556
26
- liger_kernel/ops/geglu.py,sha256=z-t9OMk3SDL5sJenJjCzhGEeGusL22j3nDjTDEUDAz8,4219
26
+ liger_kernel/ops/geglu.py,sha256=-ruMACDsFH1YsAak6BGvZ0ktLGIrBE6yGF0dAyR82UU,4307
27
27
  liger_kernel/ops/group_norm.py,sha256=zoy-TcNkYtKGmGhTFJmnyiG_4Es4ZphpqP8jtUSI6-I,10912
28
28
  liger_kernel/ops/grpo_loss.py,sha256=2SyOujtF9I3xiNo4wFf4s6MeiDotE_qeYfRWgj_bOBE,9573
29
29
  liger_kernel/ops/jsd.py,sha256=onHp5T3MbvJaVz5Vup7Ww6EQp_HTaZeayTjJk6FgQMY,7042
@@ -33,55 +33,67 @@ 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=5IdJEZnbbhblkL_X8UhSD4A2CooQbOAZJw8nAekWNs4,11372
35
35
  liger_kernel/ops/qwen2vl_mrope.py,sha256=3GExhYpLgB4VUtyZyjRk8XjEur3W4EWF6HQ67ML5vBU,8481
36
- liger_kernel/ops/rms_norm.py,sha256=2V8qheEvidBm0VxwfOoAnV837F6blmtTgP91VYdGs7c,19211
36
+ liger_kernel/ops/rms_norm.py,sha256=r97gpPmhbKz9qrBjxUEX0XP04aYu4psJeLe3KnhPZyo,21852
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
40
40
  liger_kernel/ops/swiglu.py,sha256=D7nd4u_LInwsIRNCDdY77lqnTz8-W5dJrpEAt8zEO_A,3033
41
41
  liger_kernel/ops/tiled_mlp.py,sha256=eyMFsFFgHch8a_6R6IYRG24_jqKg5GF_BQUoQuAG8SY,4529
42
42
  liger_kernel/ops/tvd.py,sha256=FHJtLQI95ijqgg9UtaHpMAjSCiPxB6CduPwPMcGxelc,6405
43
- liger_kernel/ops/utils.py,sha256=kYp84AOA7D9PYrvBUSrNsfQIt8elr_uA9OxCkbfiUFA,3980
43
+ liger_kernel/ops/utils.py,sha256=Xu6MJ2-lbp4hSmI0JGImKguKU0KqWnFQDgQwOxSieyc,4360
44
+ liger_kernel/ops/backends/README.md,sha256=ZP59UUqD1WW8LwM5Y-cTpSM-Dtgdp8Wku2mE9kqAc2E,4185
45
+ liger_kernel/ops/backends/__init__.py,sha256=-mgef3cHfDFeL5NbXbq1TI7ngCahE9qqL3aMaHnXvis,629
46
+ liger_kernel/ops/backends/registry.py,sha256=yJa_Sh2FZ__iPCIU8h2nOQbnsFQh1I-_czROLtb1uQM,1637
47
+ liger_kernel/ops/backends/_ascend/__init__.py,sha256=6n0keOX9H-kLadBdVZlx-Ce0ZLVJvLiEfR-9-uxmYUk,221
48
+ liger_kernel/ops/backends/_ascend/ascend-ub-manager-design.md,sha256=FVXHSO1KY4ZFxCAE5r4hOYB2Q8ANyrJZ7WnFJ_GeQOA,19605
49
+ liger_kernel/ops/backends/_ascend/ub_manager.py,sha256=3h7sncZk00veBJS37a01YPt1SLeAxJj5N3lPdv1wXAk,13174
50
+ liger_kernel/ops/backends/_ascend/ops/__init__.py,sha256=R1iS9R0EtmGbrN0cSkIiRtZouVl7ndiPVZJIoEALb7s,1748
51
+ liger_kernel/ops/backends/_ascend/ops/geglu.py,sha256=hs1Cdhw0pbgZFiK1srLuo8DCe8jtnmhjm5SS2vw8-0M,8421
52
+ liger_kernel/ops/backends/_ascend/ops/qwen2vl_mrope.py,sha256=pUYcstJ4FuzDTkuhmQaO3U9gcVQoNCpzuwwUdtES5hM,11015
53
+ liger_kernel/ops/backends/_ascend/ops/rope.py,sha256=nOwtm6_eSnzDjl2S-jvGpwHrumAOgWfr5pNg6SL3R2k,10842
54
+ liger_kernel/ops/backends/_ascend/ops/swiglu.py,sha256=yrbEgIgeCZyayMYHCRNq7LntZE9cEemht39_TFPro0k,4682
44
55
  liger_kernel/ops/experimental/embedding.py,sha256=tolj3tItkzpSb30zWqDN2_yX4ectflaQ8HMyKyFIQc8,4172
45
56
  liger_kernel/ops/experimental/mm_int8int2.py,sha256=TrS9lpwekrik_w5qE7AhMJD1bcq-OidjtbsW80oZ6IM,13314
46
- liger_kernel/transformers/__init__.py,sha256=CgwhrY5cdx6OcRgR2ZZJbOIkLswQWPTr-BAaoxDNNOY,10687
47
- liger_kernel/transformers/auto_model.py,sha256=0qCTRZt280Bj_LcFdzo9hlaR-BWNazawXOGgoCZjgEg,1545
48
- liger_kernel/transformers/cross_entropy.py,sha256=DMtHkKrVJDSsels7KgGQJqrXkEAd6Zopcdr-5oRmQgE,2010
49
- liger_kernel/transformers/dyt.py,sha256=i-4GPaMrl-jab9TVI5qN0-H9qycn_mCbV82ozU4nbmU,723
57
+ liger_kernel/transformers/__init__.py,sha256=4sqcDbOZ_JtS9Ag-7oyuhq5jN298GyzjJFu9J-DyyZQ,10872
58
+ liger_kernel/transformers/auto_model.py,sha256=RnJhK8xHamRnnswgRLG_muJE1i6T6LszjK8lC6vonhE,2410
59
+ liger_kernel/transformers/cross_entropy.py,sha256=08H8RxSxGX_52UzrHNnSZ_wWH-uvU8KrRiDmVrkOw14,1996
60
+ liger_kernel/transformers/dyt.py,sha256=Rng-MZQSprnGGWFtpmYKt7MIX26vFUYbq5ruM4MjH-U,719
50
61
  liger_kernel/transformers/fsdp.py,sha256=CUiyjTmjkjY7pLXQv8ly9rnzgXw6529csd9pvtJNMYc,3096
51
- liger_kernel/transformers/functional.py,sha256=OqEmsDkaV3YiXaw1zqjDvHcC9_tU5TBrmhCNPOdgHQY,8590
52
- liger_kernel/transformers/fused_add_rms_norm.py,sha256=7_Bzg-x6lLe6W1qG2DtjDALhEpNZlC6N5GppEs9cTYY,1199
53
- liger_kernel/transformers/fused_linear_cross_entropy.py,sha256=Hhp9XGgMKZhvlkjHY5Jkl_T7fSyJoCL9m5c3z_9mflQ,2347
54
- liger_kernel/transformers/fused_linear_jsd.py,sha256=bZ4otCvWBuOnA5XdQL-FzZVItJlDt-ht9e_pG7PG93E,3999
55
- liger_kernel/transformers/fused_neighborhood_attention.py,sha256=TxYDUAt9B6WSP14aJP66C_2Mbds2sSIPGnamhUSTrC8,7957
56
- liger_kernel/transformers/geglu.py,sha256=mrgqzIUVd6lN7fkDKLkw5YaESDxDtFgbot430WwPVOQ,1107
57
- liger_kernel/transformers/group_norm.py,sha256=6qMAWOprr4SzP0YhNVNGQIBpM5aUHplUD2VuGJrMBz0,2173
58
- liger_kernel/transformers/grpo_loss.py,sha256=QS6Ycct1E2yMfqoHPBa2sUAu5cmweNPK_-Q_KJE8hb4,6098
59
- liger_kernel/transformers/jsd.py,sha256=DGqRnxIZxsvxo0_tbbxX3b-sDbDjC_yKufyRIHCcScY,2979
60
- liger_kernel/transformers/kl_div.py,sha256=WLffFbh1EExD2Eb1F7lN11fo9JJC-0751WJjZAF1Fj8,409
61
- liger_kernel/transformers/layer_norm.py,sha256=c9pk3PEasOKYR0rhe5e5nNrnYKVCEW4VC8S6LpCq9EQ,906
62
- liger_kernel/transformers/llama4_rope.py,sha256=kS6PSHEwf3dS7hD7C7p8S0geugx2EMCiP0h0F7LsUoY,3639
63
- liger_kernel/transformers/monkey_patch.py,sha256=4LV6LSz_AAop6HWk1spZm1QigPN9nUDPJu9tK21-jIo,132446
64
- liger_kernel/transformers/multi_token_attention.py,sha256=K3NIY9_5TPgZ4_Rahn0xnkMXxD_fmlJHK4CWGYvGQp0,1752
65
- liger_kernel/transformers/poly_norm.py,sha256=g5tC75i3qy1_N26ZUP-jfpct7ivQAEdJfIfx8IXzeyE,1377
66
- liger_kernel/transformers/qwen2vl_mrope.py,sha256=5EwSqrMdsL9MYspeBMXBsNJKvH0MOmRrtJXAJlnnlOI,1047
67
- liger_kernel/transformers/rms_norm.py,sha256=HwddVqrqS58jE-M2_4NkFGARtCDBhGnkKyjBN9b3FYI,3004
68
- liger_kernel/transformers/rope.py,sha256=VMlDZI6zss9mLaLcN5XCE_ktmYRwAi_Eh4TIgO6NrIQ,2361
69
- liger_kernel/transformers/softmax.py,sha256=yadlAgE4V2JByMwrDDa2s5SUBp8Jgd57xwnVvAWoBaI,264
70
- liger_kernel/transformers/sparsemax.py,sha256=0lQA0UEOs4mu8CMruZ3VLhImxQVXJWhPsAKUsYA7vj8,403
71
- liger_kernel/transformers/swiglu.py,sha256=dRR69wDWSWfdjtnsTECyxQqWVo5QkdXdXm9SpSQ4Jvw,4291
72
- liger_kernel/transformers/tiled_mlp.py,sha256=J51-kpzwikDMMhT5bX-RZCKMaXBK6zZc1bhgRYTK5F0,4651
62
+ liger_kernel/transformers/functional.py,sha256=f9sOWEfh5HZwOH5cVlcB_ts0MB_-fFFPki8PVZ5w__M,8352
63
+ liger_kernel/transformers/fused_add_rms_norm.py,sha256=k98sfcZhsgtdVxChciHmv0WUizzn6f-Rn72JtGgmafI,1180
64
+ liger_kernel/transformers/fused_linear_cross_entropy.py,sha256=WnGuR_rjIWO0XHUyVakz-qsIRm028OKzi1vayvmPfbg,2320
65
+ liger_kernel/transformers/fused_linear_jsd.py,sha256=BW22DX3J6J8uZdoaU9JFUU5HnTrNYL63H9IQZzHkGu0,3982
66
+ liger_kernel/transformers/fused_neighborhood_attention.py,sha256=21O9DSRXgMQst9Lc3b62CsOLkYn-hjuskj9Zi3mvG7Y,7928
67
+ liger_kernel/transformers/geglu.py,sha256=esltAhNJZjWydvh07C6EaTdjA2aQzFPMNK92yR15SEI,1101
68
+ liger_kernel/transformers/group_norm.py,sha256=k7LDIG8H5CA5kiNj2uOi8D_Z6FlZtQDLyzJQxK2E-gA,2162
69
+ liger_kernel/transformers/grpo_loss.py,sha256=wNVz1o3q9XH17tDqaCZFEVXJhH9mQX44pWhQEwiRo_Q,6088
70
+ liger_kernel/transformers/jsd.py,sha256=_KlOX8YcdONU0tq0bIRDQ5VDBwtywm3Ro-FmlmI01qk,2975
71
+ liger_kernel/transformers/kl_div.py,sha256=94VR4uuj-2dZCTEnwFksvDi-LporrpB5HgmYtQCZnw0,402
72
+ liger_kernel/transformers/layer_norm.py,sha256=l4nsT_Zj4CdVZOM7F0I0Ox-lmLHyIJzqQvVaF0o0HbI,895
73
+ liger_kernel/transformers/llama4_rope.py,sha256=A_nxcS_KiUCyNeL2FAZX7yUhDsX7krrI9BG49OaN_nM,3627
74
+ liger_kernel/transformers/monkey_patch.py,sha256=ESFIi_7hQMcnUtRLjAMJ9kbzSbwToDhpOfFa6aQ-SrY,135534
75
+ liger_kernel/transformers/multi_token_attention.py,sha256=LtEjG7qy1-JK-HIPaz8zZ4P08aSZTnj5D635Pa04Onc,1730
76
+ liger_kernel/transformers/poly_norm.py,sha256=T3VdLQHLcCY7KzNzrc6IJRs8SzO8Yc7a0BS_2p6d7Wo,1367
77
+ liger_kernel/transformers/qwen2vl_mrope.py,sha256=0hOBR3j2Yd6xbT4z9BNRKEy1D0eyOUsIW6EmI_3PPNI,1033
78
+ liger_kernel/transformers/rms_norm.py,sha256=dD_69_GA3GUdtvdYVxTLKGeg8QZinJpS3qfeV7WvOuA,3237
79
+ liger_kernel/transformers/rope.py,sha256=-W9aYLa2hMOmmG5yeHcvPsOI5UTc95ylYxUddxkwmkA,2867
80
+ liger_kernel/transformers/softmax.py,sha256=VI5QGHYpXSiXckgovEnDGcXwitimsxKB0GX-AT4dAC4,256
81
+ liger_kernel/transformers/sparsemax.py,sha256=Os49bSpPX4pWymsasv_3j20m8GFaI54e03XFPkHiPE0,393
82
+ liger_kernel/transformers/swiglu.py,sha256=LpgikAs9hibAL7G6itygBbOlW9tZe5s4D2IGAKGpbPw,4284
83
+ liger_kernel/transformers/tiled_mlp.py,sha256=gPsz7b0kxpk3mre7o1uGBt-XdNvMUN7IIqnUYIur-T0,4628
73
84
  liger_kernel/transformers/trainer_integration.py,sha256=W3ON51O5GkyzNJsItz0y5rKx-uy2f2cFfveZpqbUdhw,123
74
- liger_kernel/transformers/tvd.py,sha256=XrRfyJIqN6HFxXk8MYyFVZM1OLz3mtSbRZvWfZ_JerQ,450
85
+ liger_kernel/transformers/tvd.py,sha256=GYjhtXgS3RTPveOTN2gyK4uBnjs6ii2vkSZRX21QpqA,446
75
86
  liger_kernel/transformers/experimental/__init__.py,sha256=oQqk-f32JYgWEP9DJCj6ty6bbJSGrdXsFDQFwGeX6vI,127
76
- liger_kernel/transformers/experimental/embedding.py,sha256=2P0QYdlFyFrG5OqTzTa1wcRgDSyjBMv5i1a7BrDPDQw,881
87
+ liger_kernel/transformers/experimental/embedding.py,sha256=bjy9hHj--ivy6xEWdiE6qLy9uLyeS4PsBEgl_MdDrng,858
77
88
  liger_kernel/transformers/model/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
78
89
  liger_kernel/transformers/model/falcon_h1.py,sha256=heUZ4wUt2ATmtBtmv8Rcro3pQl6fV9T0pburjTTW7os,5004
79
90
  liger_kernel/transformers/model/gemma.py,sha256=pAri4PYpknsFfkvyo8Ez2NNlqrUDW-KkExUXTGZAcH4,10621
80
91
  liger_kernel/transformers/model/gemma2.py,sha256=qa9Ok42vFojVGNmASTH3Ek566Vu507kjd--ZpZDKX9M,12024
81
- liger_kernel/transformers/model/gemma3.py,sha256=mEV3Kuy-dqfTk_b899Vb-InuD4_DvwH0nm5xgbG-0MM,14911
92
+ liger_kernel/transformers/model/gemma3.py,sha256=ZUrFCc-pfF8jYHV0HsptBr98hx6p2q9ea0kSzVAoFPo,14966
82
93
  liger_kernel/transformers/model/glm4.py,sha256=bSp22iPIjsli4-c_usUOsyh1Bs2gIK8X6ynS0azseUs,5900
83
94
  liger_kernel/transformers/model/glm4v.py,sha256=dd-BQpccDCp1SbIxcJ5rG8xcwYQK3KOv1Tgm9TGnZc4,6594
84
95
  liger_kernel/transformers/model/glm4v_moe.py,sha256=zKhMdOOrRhlrvCSFaeVYfddL1ubpY8edEO91TN81n98,7135
96
+ liger_kernel/transformers/model/gpt_oss.py,sha256=8jEAQQNEXgVA-yuvEjKkBQvCvZy0E9ns-O9BPlajXXU,11197
85
97
  liger_kernel/transformers/model/hunyuan_v1.py,sha256=MJvP9xkUFePIV0HLETJM4YPbVCEPkAE1ZI5Jxyiebh0,5731
86
98
  liger_kernel/transformers/model/internvl.py,sha256=OOutracs9qrPHSU7FVYar08yinvGrHQVPvo39JEws6w,6473
87
99
  liger_kernel/transformers/model/llama.py,sha256=kqZeONzwTBzudoChlKMzq1w23BtYGbxWZC1l1V__JTw,13410
@@ -94,7 +106,7 @@ liger_kernel/transformers/model/mllama.py,sha256=vAHwCm63sn4kpAY0rDGf_N0HR7KRTBV
94
106
  liger_kernel/transformers/model/olmo2.py,sha256=-h2bUOeuPfY1MdShdRvq5_wFDHKP4PEimgIl0fL-BT4,5902
95
107
  liger_kernel/transformers/model/olmo3.py,sha256=k2zYOlS8U_b5MwjdToB3tDRQ0bH_mWapVQqJcH8-qAo,6007
96
108
  liger_kernel/transformers/model/output_classes.py,sha256=0BGXVR4dYQpSHLkSqpRoXuHMryrceGSlTYRu6pvd8ZY,4542
97
- liger_kernel/transformers/model/paligemma.py,sha256=r0smHLADkEwfLS6d6ArWoSWEeLt2d_8pmgOO5F04b1o,20793
109
+ liger_kernel/transformers/model/paligemma.py,sha256=UAYoKkIMvvix7GG3cSdWaDxVjMp26YsvthJuE7wFf6Y,20848
98
110
  liger_kernel/transformers/model/phi3.py,sha256=PT7Kw6yySg-7TsssWfi82eVMN3SWujCqzCqHigAdfeQ,4574
99
111
  liger_kernel/transformers/model/qwen2.py,sha256=ojqdJpD3A9A5uCS0N_rSq8gyNYWSsHfuvx3Z3ObC7ss,10686
100
112
  liger_kernel/transformers/model/qwen2_5_vl.py,sha256=FbIZDcg9cOr4PtBLNN8yVubN-gu2clndjSIzfi8NMos,6894
@@ -110,9 +122,9 @@ liger_kernel/transformers/trainer/__init__.py,sha256=p7yQfklV8-467qSz_ZMimkbDF7H
110
122
  liger_kernel/transformers/trainer/orpo_trainer.py,sha256=tX0h63aOFe3rNqTmk6JpMf75UPo981yzEa6TghnjS0Q,5370
111
123
  liger_kernel/triton/__init__.py,sha256=qCiCamzCRv6lpV8IqpAc9YMdNKC7GKurClWceQPnlis,92
112
124
  liger_kernel/triton/monkey_patch.py,sha256=Rd0hUHAzDkFfHvnX7-PBaNK5EKnZhtfM_h-fgQH9HPY,1568
113
- liger_kernel_nightly-0.6.4.dev20251202054858.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
114
- liger_kernel_nightly-0.6.4.dev20251202054858.dist-info/METADATA,sha256=-19QGcupE4UryR0Ai5X2O-1WMIQ_fFzCjNiUblQI27o,25238
115
- liger_kernel_nightly-0.6.4.dev20251202054858.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
116
- liger_kernel_nightly-0.6.4.dev20251202054858.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
117
- liger_kernel_nightly-0.6.4.dev20251202054858.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
118
- liger_kernel_nightly-0.6.4.dev20251202054858.dist-info/RECORD,,
125
+ liger_kernel_nightly-0.6.4.dev20260107181130.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
126
+ liger_kernel_nightly-0.6.4.dev20260107181130.dist-info/METADATA,sha256=QpZmACb6StxhHZq45yyU_eXWsjjZ7R-ncI-CgJ9mWTU,25660
127
+ liger_kernel_nightly-0.6.4.dev20260107181130.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
128
+ liger_kernel_nightly-0.6.4.dev20260107181130.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
129
+ liger_kernel_nightly-0.6.4.dev20260107181130.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
130
+ liger_kernel_nightly-0.6.4.dev20260107181130.dist-info/RECORD,,