liger-kernel 0.5.4__py3-none-any.whl → 0.5.6__py3-none-any.whl

{liger_kernel-0.5.4.dist-info → liger_kernel-0.5.6.dist-info}/METADATA

@@ -1,6 +1,6 @@
-Metadata-Version: 2.2
+Metadata-Version: 2.4
 Name: liger_kernel
-Version: 0.5.4
+Version: 0.5.6
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -45,6 +45,7 @@ Requires-Dist: datasets>=2.19.2; extra == "dev"
 Requires-Dist: seaborn; extra == "dev"
 Requires-Dist: mkdocs; extra == "dev"
 Requires-Dist: mkdocs-material; extra == "dev"
+Dynamic: license-file
 Dynamic: provides-extra
 Dynamic: requires-dist
 
@@ -115,6 +116,7 @@ Dynamic: requires-dist
 <details>
   <summary>Latest News 🔥</summary>
 
+  - [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/)
   - [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)!
   - [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)
   - [2024/11/6] We release [v0.4.0](https://github.com/linkedin/Liger-Kernel/releases/tag/v0.4.0): Full AMD support, Tech Report, Modal CI, Llama-3.2-Vision!
@@ -154,7 +156,7 @@ With one line of code, Liger Kernel can increase throughput by more than 20% and
 We provide optimized post training kernels like DPO, ORPO, SimPO, and more which can reduce memory usage by up to 80%. You can easily use them as python modules.
 
 ```python
-from liger_kernel.chunked_loss import LigerFusedLinearDPOLoss
+from liger_kernel.chunked_loss import LigerFusedLinearORPOLoss
 orpo_loss = LigerFusedLinearORPOLoss()
 y = orpo_loss(lm_head.weight, x, target)
 ```
@@ -177,7 +179,7 @@ y = orpo_loss(lm_head.weight, x, target)
 - **Exact:** Computation is exact—no approximations! Both forward and backward passes are implemented with rigorous unit tests and undergo convergence testing against training runs without Liger Kernel to ensure accuracy.
 - **Lightweight:** Liger Kernel has minimal dependencies, requiring only Torch and Triton—no extra libraries needed! Say goodbye to dependency headaches!
 - **Multi-GPU supported:** Compatible with multi-GPU setups (PyTorch FSDP, DeepSpeed, DDP, etc.).
-- **Trainer Framework Integration**: [Axolotl](https://github.com/axolotl-ai-cloud/axolotl), [LLaMa-Factory](https://github.com/hiyouga/LLaMA-Factory), [SFTTrainer](https://github.com/huggingface/trl/releases/tag/v0.10.1), [Hugging Face Trainer](https://github.com/huggingface/transformers/pull/32860), [SWIFT](https://github.com/modelscope/ms-swift)
+- **Trainer Framework Integration**: [Axolotl](https://github.com/axolotl-ai-cloud/axolotl), [LLaMa-Factory](https://github.com/hiyouga/LLaMA-Factory), [SFTTrainer](https://github.com/huggingface/trl/releases/tag/v0.10.1), [Hugging Face Trainer](https://github.com/huggingface/transformers/pull/32860), [SWIFT](https://github.com/modelscope/ms-swift), [oumi](https://github.com/oumi-ai/oumi/tree/main)
 
 ## Installation
 
@@ -312,8 +314,10 @@ loss.backward()
 | Mixtral     | `liger_kernel.transformers.apply_liger_kernel_to_mixtral`  | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | Gemma1      | `liger_kernel.transformers.apply_liger_kernel_to_gemma`    | RoPE, RMSNorm, GeGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
 | Gemma2      | `liger_kernel.transformers.apply_liger_kernel_to_gemma2`   | RoPE, RMSNorm, GeGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
+| Paligemma, Paligemma2, & Paligemma2 Mix      | `liger_kernel.transformers.apply_liger_kernel_to_paligemma`   | LayerNorm, RoPE, RMSNorm, GeGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
 | Qwen2, Qwen2.5, & QwQ      | `liger_kernel.transformers.apply_liger_kernel_to_qwen2`    | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | Qwen2-VL, & QVQ       | `liger_kernel.transformers.apply_liger_kernel_to_qwen2_vl`    | RMSNorm, LayerNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
+| Qwen2.5-VL       | `liger_kernel.transformers.apply_liger_kernel_to_qwen2_5_vl`    | RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | Phi3 & Phi3.5       | `liger_kernel.transformers.apply_liger_kernel_to_phi3`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
 | Granite 3.0 & 3.1   | `liger_kernel.transformers.apply_liger_kernel_to_granite`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss |
 | OLMo2   | `liger_kernel.transformers.apply_liger_kernel_to_olmo2`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
@@ -385,8 +389,8 @@ loss.backward()
 ## Contact
 
 - For issues, create a Github ticket in this repository
-- For open discussion, join [our discord channel](https://discord.gg/gpumode)
-- For formal collaboration, send an email to yannchen@linkedin.com
+- For open discussion, join [our discord channel on GPUMode](https://discord.com/channels/1189498204333543425/1275130785933951039)
+- For formal collaboration, send an email to yannchen@linkedin.com and hning@linkedin.com
 
 ## Cite this work
 
@@ -405,7 +409,7 @@ Biblatex entry:
 ```
 
 ## Star History
-[![Star History Chart](https://api.star-history.com/svg?repos=linkedin/Liger-Kernel&type=Date)](https://star-history.com/#linkedin/Liger-Kernel&Date)
+[![Star History Chart](https://api.star-history.com/svg?repos=linkedin/Liger-Kernel&type=Date)](https://www.star-history.com/#linkedin/Liger-Kernel&Date)
 
 <p align="right" style="font-size: 14px; color: #555; margin-top: 20px;">
     <a href="#readme-top" style="text-decoration: none; color: #007bff; font-weight: bold;">

liger_kernel-0.5.6.dist-info/RECORD

@@ -0,0 +1,80 @@
+liger_kernel/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+liger_kernel/env_report.py,sha256=uhdEC8OydxoZlb7B6YYcAaBF3crGFdIck-4cxaW4NJY,1728
+liger_kernel/utils.py,sha256=178Hn8uD-VauDT6FjqMyXLbKLod8ObIpaTtapHwfEK0,1861
+liger_kernel/chunked_loss/README.md,sha256=0FmkFC3hKBqyoDT5uTlIYmrvRkF-EOCR1y-EBU1LpWU,2248
+liger_kernel/chunked_loss/__init__.py,sha256=ATu-xX5Fc49Cr6yBOGBRNTo593ZrU5ZCsIuvoIbJWw4,603
+liger_kernel/chunked_loss/cpo_loss.py,sha256=Gzz1eU4kgcbdubFVRy55e8A1Cr-r45UgNicXwZIjmBU,5454
+liger_kernel/chunked_loss/dpo_loss.py,sha256=xZwGqS04si9zXyob95SAdalC-hajZg8fWINqiqffN8k,5855
+liger_kernel/chunked_loss/functional.py,sha256=9G3nKm-Bi7uoZRFkL8wwGMl6juDl4bSzDvTa5GHZPzg,955
+liger_kernel/chunked_loss/fused_linear_distillation.py,sha256=ooR-qnZCyWJN935oHCSWLaKKKyaYERyhNczRGi1VOiw,11935
+liger_kernel/chunked_loss/fused_linear_ppo.py,sha256=-E4AuWY-y2bMo_kAmEQBgQ92UJh3L5IiCRGVcfMJOCE,12731
+liger_kernel/chunked_loss/fused_linear_preference.py,sha256=ojB42jYPu0c4ki96Ft-hy7Sf6fh_WikG-aWNrlZzSio,18362
+liger_kernel/chunked_loss/fused_linear_unpaired_preference.py,sha256=RiuK3UtRwH9T6jZ36sA8Urj-TVuOLOO2syLg_JOQapY,13437
+liger_kernel/chunked_loss/grpo_loss.py,sha256=6Mb4ZT6MfnOr4Xo681rMR0LKkhzJhInvQp8wp2YVMK0,8913
+liger_kernel/chunked_loss/jsd_loss.py,sha256=u2ahkuHsbhpNaKcpBCz5gCMDk9ou-P04DHji592dIBo,7067
+liger_kernel/chunked_loss/kto_loss.py,sha256=llVCe6DkcpCo57seGWoMikaQVFApx764jsmSbQyqwQY,7529
+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=T5oSsqOS1y-Iea5o9v_BSU-_mIEXqWAT1oX_m59NcA4,18941
+liger_kernel/ops/dyt.py,sha256=YD1-buHz9VmIX838VKzLc-lm5CeUQ4LAskGDWBUMQHA,6187
+liger_kernel/ops/fused_linear_cross_entropy.py,sha256=1Y3Uk_TCSjqKgoG2eot1ptnWXJXXQESqGvOmqAW1gsM,10912
+liger_kernel/ops/fused_linear_jsd.py,sha256=CSoprxb-YcJy-YUKiTcYkxN8sb9h2kdk_iHuncvSV5c,9683
+liger_kernel/ops/geglu.py,sha256=axGvCIvlBzuluoAIrWTsp2iZM4BFKNInkPov8YVvH9E,4126
+liger_kernel/ops/group_norm.py,sha256=qD4D4lSjSgVtO52EBNLC2iTseALRgPgqXE50U2woggk,10837
+liger_kernel/ops/jsd.py,sha256=rkloGA7nDfVaa5nKY6-EYBw0E1p_MSsl4fr2xZGTp04,6961
+liger_kernel/ops/kl_div.py,sha256=NkG7D6_DnPBzr-ohhYiQbRBnq_fbGmpn5UU7y0UBKQo,8420
+liger_kernel/ops/layer_norm.py,sha256=6roQjioyg-9O2qLPV8nL4U0-5UH80tdzOMTWwjvDnn8,7961
+liger_kernel/ops/qwen2vl_mrope.py,sha256=3GExhYpLgB4VUtyZyjRk8XjEur3W4EWF6HQ67ML5vBU,8481
+liger_kernel/ops/rms_norm.py,sha256=PWLJcdIKU5e-8BuYFHd9Cqlq6wmr6fUXKi9zQD4LetU,11727
+liger_kernel/ops/rope.py,sha256=ofmBOkUpZZO-Q8Z5B_LOFYYLD-YT-8WnJ4vGOrDYouI,8943
+liger_kernel/ops/swiglu.py,sha256=KmgMjaJQnbLLgZn2nEpbwHU_xpnYRweCyrLQSVvM1vA,3015
+liger_kernel/ops/tvd.py,sha256=FHJtLQI95ijqgg9UtaHpMAjSCiPxB6CduPwPMcGxelc,6405
+liger_kernel/ops/utils.py,sha256=uoFKQqo-34N2TWQNvXMFywqGiOMMXNEVBxVojzlUAa0,3836
+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=t70gqygxH63iz-B0MOdZx4AEgA8MfqU1G7N6dvIneCY,2618
+liger_kernel/transformers/auto_model.py,sha256=0qCTRZt280Bj_LcFdzo9hlaR-BWNazawXOGgoCZjgEg,1545
+liger_kernel/transformers/cross_entropy.py,sha256=z3KTWQnFxr_IZaVjtYt0ZNEWQdDdYThN35xWkHlDGH0,1683
+liger_kernel/transformers/dyt.py,sha256=QMqqc14pkE0WhpRZvapfnNAun-6C0C_tHExL2ZJuCUA,648
+liger_kernel/transformers/functional.py,sha256=4h9Pdx_iINBqfv2Zod_c27qOpYXDDwbdVgatQ9_XBmI,5089
+liger_kernel/transformers/fused_linear_cross_entropy.py,sha256=09Rt7FZzLH42VOcIbQ4dlQd0o3Rlb4vk6fqiOQ7WTD8,1778
+liger_kernel/transformers/fused_linear_jsd.py,sha256=bZ4otCvWBuOnA5XdQL-FzZVItJlDt-ht9e_pG7PG93E,3999
+liger_kernel/transformers/geglu.py,sha256=mrgqzIUVd6lN7fkDKLkw5YaESDxDtFgbot430WwPVOQ,1107
+liger_kernel/transformers/group_norm.py,sha256=6qMAWOprr4SzP0YhNVNGQIBpM5aUHplUD2VuGJrMBz0,2173
+liger_kernel/transformers/jsd.py,sha256=DGqRnxIZxsvxo0_tbbxX3b-sDbDjC_yKufyRIHCcScY,2979
+liger_kernel/transformers/kl_div.py,sha256=WLffFbh1EExD2Eb1F7lN11fo9JJC-0751WJjZAF1Fj8,409
+liger_kernel/transformers/layer_norm.py,sha256=c9pk3PEasOKYR0rhe5e5nNrnYKVCEW4VC8S6LpCq9EQ,906
+liger_kernel/transformers/monkey_patch.py,sha256=95afvIrZA9xSWLNIJspBLbz8lxv2Y5gfZke7MyqoOX8,56965
+liger_kernel/transformers/qwen2vl_mrope.py,sha256=5EwSqrMdsL9MYspeBMXBsNJKvH0MOmRrtJXAJlnnlOI,1047
+liger_kernel/transformers/rms_norm.py,sha256=GqCEJuGt0YdqqlMcToE0Wp4A8YFquDa4UUSyH2uFW2A,1191
+liger_kernel/transformers/rope.py,sha256=ZTrTORSAyfcFIKjk6XEeYmk4ROH7xXED9L4g2NFntlE,999
+liger_kernel/transformers/swiglu.py,sha256=i9WTqcNRqReU4XJs391IPbl-I5X0wG4T72D4pqGFfJg,2422
+liger_kernel/transformers/trainer_integration.py,sha256=W3ON51O5GkyzNJsItz0y5rKx-uy2f2cFfveZpqbUdhw,123
+liger_kernel/transformers/tvd.py,sha256=XrRfyJIqN6HFxXk8MYyFVZM1OLz3mtSbRZvWfZ_JerQ,450
+liger_kernel/transformers/experimental/embedding.py,sha256=2P0QYdlFyFrG5OqTzTa1wcRgDSyjBMv5i1a7BrDPDQw,881
+liger_kernel/transformers/model/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+liger_kernel/transformers/model/gemma.py,sha256=7cBTljzh-8_ACBhYl6NUfj5_ux92YRlmnAU5gfDAQAI,9312
+liger_kernel/transformers/model/gemma2.py,sha256=X0FOIhvFlTrmWI7Ws06wUkutgHW3lWtLOnnHp1NgZ3A,10403
+liger_kernel/transformers/model/llama.py,sha256=d9rBaK8e8RSMCFHdgom9ZHuXOlnh6U_o-GkAFGRNGOY,9989
+liger_kernel/transformers/model/llava.py,sha256=b0pEagjUbu2-eS9xegjyfl1DwIXLwZcNpff55ibaMbA,17601
+liger_kernel/transformers/model/loss_utils.py,sha256=Z-fUrf-cUDUjUIH7Tl9OL2hT8nmtx7ES3kg8syuWKy4,1476
+liger_kernel/transformers/model/mistral.py,sha256=o7tyl1sPWPfZwwrBLRlryHlSI8I55viuJoMI5Bh5Nww,5014
+liger_kernel/transformers/model/mixtral.py,sha256=T0ITv2-PkR8VErVOVUizoS4EzjmARyR7GFh0tXDB_i4,11089
+liger_kernel/transformers/model/mllama.py,sha256=RCKtwnGOMFYIbtt1zUQ15Cyv4eNpHkTWcgkmG2EEs2I,10804
+liger_kernel/transformers/model/olmo2.py,sha256=5M8kczp4D-jvbjcV7cKATIJGF34xd-Rs-PPdKZWSIlY,4685
+liger_kernel/transformers/model/paligemma.py,sha256=GNReT6tVZt3ON6aaa9ovg8mnu1hYocSx9OhgC7b-_28,19191
+liger_kernel/transformers/model/phi3.py,sha256=NmU2DuU1Huwha6K7YSsJCnvQfUovTTGlsfBZhbx0UoI,9951
+liger_kernel/transformers/model/qwen2.py,sha256=t7NotBHoebsPqNSxwaf9DXTg8jxgB5BdunSGqYOE0hQ,9240
+liger_kernel/transformers/model/qwen2_5_vl.py,sha256=70BnHZjx6eQWTwi3zc5SMwxTeOOA4Tbdkfy6IYRcTaM,9289
+liger_kernel/transformers/model/qwen2_vl.py,sha256=zo4O9fShNHYqSLrzLGqQYWSMtJI6UHaSY7zvMCYWyD8,9685
+liger_kernel/transformers/trainer/__init__.py,sha256=p7yQfklV8-467qSz_ZMimkbDF7HHWHwku25A-GYL0WU,193
+liger_kernel/transformers/trainer/orpo_trainer.py,sha256=pdekW7l6Qg_aqa5SYKYlSWUF8m3lkOFvFLcIMEHrz9s,8338
+liger_kernel/triton/__init__.py,sha256=qCiCamzCRv6lpV8IqpAc9YMdNKC7GKurClWceQPnlis,92
+liger_kernel/triton/monkey_patch.py,sha256=Rd0hUHAzDkFfHvnX7-PBaNK5EKnZhtfM_h-fgQH9HPY,1568
+liger_kernel-0.5.6.dist-info/licenses/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel-0.5.6.dist-info/licenses/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel-0.5.6.dist-info/METADATA,sha256=yam1-5oz74ok_T_rVfn3RLvCDXPxDfXZpChC1PVTFoY,23002
+liger_kernel-0.5.6.dist-info/WHEEL,sha256=CmyFI0kx5cdEMTLiONQRbGQwjIoR1aIYB7eCAQ4KPJ0,91
+liger_kernel-0.5.6.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel-0.5.6.dist-info/RECORD,,

{liger_kernel-0.5.4.dist-info → liger_kernel-0.5.6.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (75.8.0)
+Generator: setuptools (78.1.0)
 Root-Is-Purelib: true
 Tag: py3-none-any
 

liger_kernel/chunked_loss/fused_linear_rlhf.py

@@ -1,213 +0,0 @@
-from functools import partial
-
-import torch
-import torch.nn.functional as F
-
-
-class LigerFusedLinearRLHFBase(torch.autograd.Function):
-    @staticmethod
-    def forward(
-        ctx,
-        _input,
-        weight,
-        attention_mask,
-        rewards,
-        bias=None,
-        loss_fn=None,
-        num_generations=4,
-        beta=0.1,
-        compiled=True,
-        use_ref_model=False,
-        ref_input=None,
-        ref_weight=None,
-        ref_bias=None,
-    ):
-        """Chunked forward pass for RLHF loss computation."""
-        # Save for backward
-        ctx.beta = beta
-        ctx.rewards = rewards
-
-        # Initialize accumulators
-        loss_acc = torch.zeros((), device=_input.device)
-        grad_weight = torch.zeros_like(weight)  # [V, H]
-        grad_inputs = []
-        grad_bias = torch.zeros_like(bias) if bias is not None else None  # [V]
-        aggregated_metrics = []
-
-        # Create a partial function with fixed arguments
-        compute_loss = partial(
-            LigerFusedLinearRLHFBase._compute_chunk_loss,
-            beta=beta,
-            use_ref_model=use_ref_model,
-            ref_weight=ref_weight,
-            ref_bias=ref_bias,
-            rlhf_loss_fn=loss_fn,
-        )
-
-        def fused_fwd_bwd(input_chunk, attention_mask_chunk, rewards_chunk, ref_input_chunk):
-            """Fused forward and backward for a chunk."""
-            if bias is not None:
-                return torch.func.grad_and_value(compute_loss, argnums=(0, 1, 5), has_aux=True)(
-                    input_chunk,  # arg 0
-                    weight,  # arg 1
-                    attention_mask_chunk,  # arg 2
-                    rewards_chunk,  # arg 3
-                    ref_input_chunk,  # arg 4
-                    bias,  # arg 5
-                )
-            else:
-                return torch.func.grad_and_value(compute_loss, argnums=(0, 1), has_aux=True)(
-                    input_chunk,  # arg 0
-                    weight,  # arg 1
-                    attention_mask_chunk,  # arg 2
-                    rewards_chunk,  # arg 3
-                    ref_input_chunk,  # arg 4
-                )
-
-        def accumulate_chunk(input_chunk, attention_mask_chunk, rewards_chunk, ref_input_chunk=None):
-            if bias is not None:
-                (chunk_grad_input, chunk_grad_weight, chunk_grad_bias), (chunk_loss, chunk_metrics) = fused_fwd_bwd(
-                    input_chunk, attention_mask_chunk, rewards_chunk, ref_input_chunk
-                )
-                grad_bias.add_(chunk_grad_bias)
-            else:
-                (chunk_grad_input, chunk_grad_weight), (chunk_loss, chunk_metrics) = fused_fwd_bwd(
-                    input_chunk, attention_mask_chunk, rewards_chunk, ref_input_chunk
-                )
-
-            # Accumulate gradients and loss
-            grad_weight.add_(chunk_grad_weight)
-            grad_inputs.append(chunk_grad_input)
-            loss_acc.add_(chunk_loss)
-
-            # Initialize storage for metrics on first chunk
-            if len(aggregated_metrics) == 0:
-                for metric in chunk_metrics:
-                    if metric.ndim == 0:
-                        aggregated_metrics.append(torch.zeros((), device=metric.device))
-                    else:
-                        aggregated_metrics.append([])
-
-            # Accumulate metrics
-            for i, metric in enumerate(chunk_metrics):
-                if metric.ndim == 0:
-                    aggregated_metrics[i].add_(metric)
-                else:
-                    aggregated_metrics[i].append(metric)
-
-        if compiled:
-            accumulate_chunk = torch.compile(accumulate_chunk)
-
-        # Process input in chunks
-        chunks = max(1, _input.shape[0] // num_generations)
-        _input_chunks = torch.chunk(_input, chunks=chunks, dim=0)
-        _attention_mask_chunks = torch.chunk(attention_mask, chunks=chunks, dim=0)
-        _rewards_chunks = torch.chunk(rewards, chunks=chunks, dim=0)
-        _ref_input_chunks = torch.chunk(ref_input, chunks=chunks, dim=0) if use_ref_model else [None] * chunks
-
-        for input_chunk, attention_mask_chunk, rewards_chunk, ref_input_chunk in zip(
-            _input_chunks, _attention_mask_chunks, _rewards_chunks, _ref_input_chunks
-        ):
-            # Mark dynamic dimensions
-            torch._dynamo.mark_dynamic(input_chunk, 1)
-            torch._dynamo.mark_dynamic(attention_mask_chunk, 1)
-            if ref_input_chunk is not None:
-                torch._dynamo.mark_dynamic(ref_input_chunk, 1)
-
-            accumulate_chunk(input_chunk, attention_mask_chunk, rewards_chunk, ref_input_chunk)
-
-        # Scale accumulated loss by number of chunks since we're averaging
-        loss_acc = loss_acc / chunks
-
-        # Combine gradients
-        grad_input = torch.cat(grad_inputs, dim=0)
-
-        # Save for backward
-        ctx.save_for_backward(grad_input, grad_weight, grad_bias)
-
-        # Finalize metrics
-        final_metrics = []
-        for metric in aggregated_metrics:
-            if isinstance(metric, list):
-                final_metrics.append(torch.cat(metric, dim=0))
-            else:
-                final_metrics.append(metric / chunks)
-
-        return loss_acc, tuple(final_metrics)
-
-    @staticmethod
-    def _compute_chunk_loss(
-        input_chunk,
-        weight,
-        attention_mask_chunk,
-        rewards_chunk,
-        ref_input_chunk=None,
-        bias=None,
-        beta=0.1,
-        use_ref_model=False,
-        ref_weight=None,
-        ref_bias=None,
-        rlhf_loss_fn=None,
-    ):
-        """Compute loss for a single chunk."""
-        # Get policy log probabilities using chunk_forward
-        log_probs, _, logits_mean = LigerFusedLinearRLHFBase.chunk_forward(input_chunk, weight, bias=bias)
-
-        # Get reference log probabilities if needed
-        ref_log_probs = None
-        if use_ref_model and ref_input_chunk is not None:
-            with torch.no_grad():
-                ref_log_probs, _, _ = LigerFusedLinearRLHFBase.chunk_forward(ref_input_chunk, ref_weight, bias=ref_bias)
-
-        # Compute chunk loss and metrics using the provided loss function
-        chunk_loss, chunk_metrics = rlhf_loss_fn(
-            log_probs=log_probs,
-            attention_mask=attention_mask_chunk,
-            rewards=rewards_chunk,
-            ref_log_probs=ref_log_probs,
-            beta=beta,
-        )
-
-        return chunk_loss, (logits_mean, *chunk_metrics)
-
-    @staticmethod
-    def chunk_forward(input_chunk, weight, bias=None):
-        """Forward pass computation for a single chunk without explicit reshaping."""
-        # Directly compute logits via batched matrix multiplication: [B, T, H] @ [H, V] -> [B, T, V]
-        logits = torch.matmul(input_chunk, weight.t())
-        if bias is not None:
-            logits = logits + bias  # Broadcasts bias to [B, T, V]
-
-        # Compute log probabilities using softmax over the last dimension
-        log_probs = F.log_softmax(logits.float(), dim=-1)
-
-        # Monitoring: compute mean of logits
-        batch_size, seq_len, _ = input_chunk.shape
-        logits_mean = logits.sum() / (batch_size * seq_len * weight.shape[0])
-        return log_probs, logits, logits_mean
-
-    @staticmethod
-    def backward(ctx, grad_output, *grad_metrics):
-        """Backward pass for RLHF loss."""
-        grad_input, grad_weight, grad_bias = ctx.saved_tensors
-        if grad_output != 1.0:
-            grad_input = grad_input * grad_output
-            grad_weight = grad_weight * grad_output
-            if grad_bias is not None:
-                grad_bias = grad_bias * grad_output
-
-        return (
-            grad_input,
-            grad_weight,
-            None,  # grad_attention_mask
-            None,  # grad_rewards
-            grad_bias,
-            None,  # grad_loss_fn
-            None,  # grad_chunk_size
-            None,  # grad_beta
-            None,  # grad_compiled
-            None,  # grad_use_ref_model
-            None,  # grad_ref_input
-            None,  # grad_ref_weight
-            None,  # grad_ref_bias
-        )

liger_kernel-0.5.4.dist-info/RECORD

@@ -1,74 +0,0 @@
-liger_kernel/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-liger_kernel/env_report.py,sha256=uhdEC8OydxoZlb7B6YYcAaBF3crGFdIck-4cxaW4NJY,1728
-liger_kernel/utils.py,sha256=FtVUkCGBT1UNasTl6HMNycWwiwHayK6tx-ZDdA-sNX4,1884
-liger_kernel/chunked_loss/README.md,sha256=0FmkFC3hKBqyoDT5uTlIYmrvRkF-EOCR1y-EBU1LpWU,2248
-liger_kernel/chunked_loss/__init__.py,sha256=ATu-xX5Fc49Cr6yBOGBRNTo593ZrU5ZCsIuvoIbJWw4,603
-liger_kernel/chunked_loss/cpo_loss.py,sha256=OdBR8WYdHTKpLI_c9DcuwqKSWPeAAeTyREz46Vu_cAY,3682
-liger_kernel/chunked_loss/dpo_loss.py,sha256=wgjnwzLfrMUwV5mXgrq6G1YfQKWnbiFJegaP48BGJHY,4509
-liger_kernel/chunked_loss/functional.py,sha256=THWWpCnRVhTVfnPnyvQjdBvo1JDtxhwLmtZE_yiBBqM,817
-liger_kernel/chunked_loss/fused_linear_distillation.py,sha256=5V8rdva89WyHVbmJ8JOmC4DYNOR6ByXfx3qlUieOZkI,11002
-liger_kernel/chunked_loss/fused_linear_preference.py,sha256=idK9V9NivoVITqVpiG0fEGUHSvinYWkn9-EYXZjR-KQ,18356
-liger_kernel/chunked_loss/fused_linear_rlhf.py,sha256=sAApL4GQ3YL2F-ymIAF61GCpFfBgFcWF5LB4Gzd7LgY,8044
-liger_kernel/chunked_loss/fused_linear_unpaired_preference.py,sha256=ZqYlXXhIphkJPxOS7iI70avgrr6x0skEtgpckZTYau0,9819
-liger_kernel/chunked_loss/grpo_loss.py,sha256=M5qlQR-v5Rh8N3P3dPGNhOKygDFJ4516_rJaVPzU_-c,4980
-liger_kernel/chunked_loss/jsd_loss.py,sha256=yRCQdvd3ruTWP4A_BfU8VcZ6LepSUfO0Ob7stGnueQY,6052
-liger_kernel/chunked_loss/kto_loss.py,sha256=b3ffJyk97e-6XdXd4HFrYyx8wW4A-CU4gOaJSimKLtA,5476
-liger_kernel/chunked_loss/orpo_loss.py,sha256=yjcrrbVeemLYodoSKT-FMSnaPtyKAZ3aOrvPD6tTY6Y,3617
-liger_kernel/chunked_loss/simpo_loss.py,sha256=3TTc7U79Orjgi-Wu81WZkWk5MgsdqKXIOBHgIvDazPw,3865
-liger_kernel/ops/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-liger_kernel/ops/cross_entropy.py,sha256=D6vFFloiuxFXoWfjlIjmfO3tVaWOiYmztw9FKAi5vdU,18608
-liger_kernel/ops/fused_linear_cross_entropy.py,sha256=1Y3Uk_TCSjqKgoG2eot1ptnWXJXXQESqGvOmqAW1gsM,10912
-liger_kernel/ops/fused_linear_jsd.py,sha256=Seshez2qaM6HiTQ8_HEqSwhaeVruNT1SvIM4ZrAPBEU,9602
-liger_kernel/ops/geglu.py,sha256=axGvCIvlBzuluoAIrWTsp2iZM4BFKNInkPov8YVvH9E,4126
-liger_kernel/ops/group_norm.py,sha256=qD4D4lSjSgVtO52EBNLC2iTseALRgPgqXE50U2woggk,10837
-liger_kernel/ops/jsd.py,sha256=0jNeRxpcNI5ckxCdoCNyO5GEedLIuzx3lz6KAiksc4o,6109
-liger_kernel/ops/kl_div.py,sha256=MnfuYqqQESON1X2Swy064x1urKtMFdgeSWd60VttBXI,8420
-liger_kernel/ops/layer_norm.py,sha256=6roQjioyg-9O2qLPV8nL4U0-5UH80tdzOMTWwjvDnn8,7961
-liger_kernel/ops/qwen2vl_mrope.py,sha256=3GExhYpLgB4VUtyZyjRk8XjEur3W4EWF6HQ67ML5vBU,8481
-liger_kernel/ops/rms_norm.py,sha256=PWLJcdIKU5e-8BuYFHd9Cqlq6wmr6fUXKi9zQD4LetU,11727
-liger_kernel/ops/rope.py,sha256=ofmBOkUpZZO-Q8Z5B_LOFYYLD-YT-8WnJ4vGOrDYouI,8943
-liger_kernel/ops/swiglu.py,sha256=KmgMjaJQnbLLgZn2nEpbwHU_xpnYRweCyrLQSVvM1vA,3015
-liger_kernel/ops/tvd.py,sha256=FHJtLQI95ijqgg9UtaHpMAjSCiPxB6CduPwPMcGxelc,6405
-liger_kernel/ops/utils.py,sha256=uoFKQqo-34N2TWQNvXMFywqGiOMMXNEVBxVojzlUAa0,3836
-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=6v_VcV1GQ9ISgNCd-ZxtmEg_s5GTBQ9F-s1KrFkYzPQ,2265
-liger_kernel/transformers/auto_model.py,sha256=0qCTRZt280Bj_LcFdzo9hlaR-BWNazawXOGgoCZjgEg,1545
-liger_kernel/transformers/cross_entropy.py,sha256=z3KTWQnFxr_IZaVjtYt0ZNEWQdDdYThN35xWkHlDGH0,1683
-liger_kernel/transformers/functional.py,sha256=ShLD3eb--XKNtllznCrOYTbo4f-1KVwzi0KLMICdrn4,4942
-liger_kernel/transformers/fused_linear_cross_entropy.py,sha256=09Rt7FZzLH42VOcIbQ4dlQd0o3Rlb4vk6fqiOQ7WTD8,1778
-liger_kernel/transformers/fused_linear_jsd.py,sha256=bZ4otCvWBuOnA5XdQL-FzZVItJlDt-ht9e_pG7PG93E,3999
-liger_kernel/transformers/geglu.py,sha256=mrgqzIUVd6lN7fkDKLkw5YaESDxDtFgbot430WwPVOQ,1107
-liger_kernel/transformers/group_norm.py,sha256=6qMAWOprr4SzP0YhNVNGQIBpM5aUHplUD2VuGJrMBz0,2173
-liger_kernel/transformers/jsd.py,sha256=DGqRnxIZxsvxo0_tbbxX3b-sDbDjC_yKufyRIHCcScY,2979
-liger_kernel/transformers/kl_div.py,sha256=WLffFbh1EExD2Eb1F7lN11fo9JJC-0751WJjZAF1Fj8,409
-liger_kernel/transformers/layer_norm.py,sha256=c9pk3PEasOKYR0rhe5e5nNrnYKVCEW4VC8S6LpCq9EQ,906
-liger_kernel/transformers/monkey_patch.py,sha256=g3i3q5McBg23A3Mnviw-Eb32le1hvN7jByzONa9ngcs,44000
-liger_kernel/transformers/qwen2vl_mrope.py,sha256=5EwSqrMdsL9MYspeBMXBsNJKvH0MOmRrtJXAJlnnlOI,1047
-liger_kernel/transformers/rms_norm.py,sha256=GqCEJuGt0YdqqlMcToE0Wp4A8YFquDa4UUSyH2uFW2A,1191
-liger_kernel/transformers/rope.py,sha256=ZTrTORSAyfcFIKjk6XEeYmk4ROH7xXED9L4g2NFntlE,999
-liger_kernel/transformers/swiglu.py,sha256=i9WTqcNRqReU4XJs391IPbl-I5X0wG4T72D4pqGFfJg,2422
-liger_kernel/transformers/trainer_integration.py,sha256=W3ON51O5GkyzNJsItz0y5rKx-uy2f2cFfveZpqbUdhw,123
-liger_kernel/transformers/tvd.py,sha256=XrRfyJIqN6HFxXk8MYyFVZM1OLz3mtSbRZvWfZ_JerQ,450
-liger_kernel/transformers/experimental/embedding.py,sha256=2P0QYdlFyFrG5OqTzTa1wcRgDSyjBMv5i1a7BrDPDQw,881
-liger_kernel/transformers/model/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-liger_kernel/transformers/model/gemma.py,sha256=ky89b3aWPaeTGRMC-745KgixtQIRXzNAiCORAMLn9yo,9654
-liger_kernel/transformers/model/gemma2.py,sha256=27NcoZjEqP9Lqb4Wf0EKqTbr2HTGiHPhrVyPCRGPz6s,10767
-liger_kernel/transformers/model/llama.py,sha256=3LJFXKFDKvEakaWPc_NicSFst4Y_hdSMrdl1UDK1EcA,10330
-liger_kernel/transformers/model/mistral.py,sha256=MVRksI5_j_8WJu8znOHKCdSI5jSu-S7cdFYzt9m_vIQ,5180
-liger_kernel/transformers/model/mixtral.py,sha256=jpZJkpl625Q-JHWarj2MqT5mRaSsiCtg0c9vVyvOdCY,11430
-liger_kernel/transformers/model/mllama.py,sha256=qWexBdskuN3gPJvPUwt4J0nU675tGD6W7wxgRZ9Bifg,11145
-liger_kernel/transformers/model/olmo2.py,sha256=yyksS6E4fuWd8asEW8rEDBKqZpFmP4ITCM_bjIDZaoY,5124
-liger_kernel/transformers/model/phi3.py,sha256=biRa8fph9qdnQmkD9I21t5XIjpIt1i6UKU4uk8Up8pU,10292
-liger_kernel/transformers/model/qwen2.py,sha256=14UuPjxB-tjqWn85Tn4fqBFvVhVsth5iPEt8kJSMiew,9581
-liger_kernel/transformers/model/qwen2_vl.py,sha256=yMLqsfSYcvhClUpTUjGoADiOxfLB2B8240VdrPP0c8s,9851
-liger_kernel/transformers/trainer/__init__.py,sha256=p7yQfklV8-467qSz_ZMimkbDF7HHWHwku25A-GYL0WU,193
-liger_kernel/transformers/trainer/orpo_trainer.py,sha256=pdekW7l6Qg_aqa5SYKYlSWUF8m3lkOFvFLcIMEHrz9s,8338
-liger_kernel/triton/__init__.py,sha256=qCiCamzCRv6lpV8IqpAc9YMdNKC7GKurClWceQPnlis,92
-liger_kernel/triton/monkey_patch.py,sha256=Rd0hUHAzDkFfHvnX7-PBaNK5EKnZhtfM_h-fgQH9HPY,1568
-liger_kernel-0.5.4.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel-0.5.4.dist-info/METADATA,sha256=Zw7n3Ey6vUed4E54H9-TzKmhuOpd9P2ZFMVL-zYUnew,22255
-liger_kernel-0.5.4.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel-0.5.4.dist-info/WHEEL,sha256=In9FTNxeP60KnTkGw7wk6mJPYd_dQSjEZmXdBdMCI-8,91
-liger_kernel-0.5.4.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel-0.5.4.dist-info/RECORD,,