liger-kernel 0.4.0__tar.gz → 0.4.2__tar.gz

{liger_kernel-0.4.0/src/liger_kernel.egg-info → liger_kernel-0.4.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel
-Version: 0.4.0
+Version: 0.4.2
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -99,7 +99,8 @@ Requires-Dist: seaborn; extra == "dev"
 
 <details>
   <summary>Latest News 🔥</summary>
-    
+
+  - [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!
   - [2024/10/21] We have released the tech report of Liger Kernel on Arxiv: https://arxiv.org/pdf/2410.10989 
   - [2024/9/6] We release v0.2.1 ([X post](https://x.com/liger_kernel/status/1832168197002510649)). 2500+ Stars, 10+ New Contributors, 50+ PRs, 50k Downloads in two weeks!
   - [2024/8/31] CUDA MODE talk, [Liger-Kernel: Real-world Triton kernel for LLM Training](https://youtu.be/gWble4FreV4?si=dxPeIchhkJ36Mbns), [Slides](https://github.com/cuda-mode/lectures?tab=readme-ov-file#lecture-28-liger-kernel)
@@ -127,18 +128,12 @@ With one line of code, Liger Kernel can increase throughput by more than 20% and
 
 ## Examples
 
-### Basic
-
-| **Example**                                    | **Description**                                                                                   | **Lightning Studio** |
-|------------------------------------------------|---------------------------------------------------------------------------------------------------|----------------------|
-| [**Hugging Face Trainer**](https://github.com/linkedin/Liger-Kernel/tree/main/examples/huggingface)      | Train LLaMA 3-8B ~20% faster with over 40% memory reduction on Alpaca dataset using 4 A100s with FSDP | TBA                  |
-| [**Lightning Trainer**](https://github.com/linkedin/Liger-Kernel/tree/main/examples/lightning)         | Increase 15% throughput and reduce memory usage by 40% with LLaMA3-8B on MMLU dataset using 8 A100s with DeepSpeed ZeRO3 | TBA                  |
 
-### Advanced
-
-| **Example**                                    | **Description**                                                                                   | **Lightning Studio** |
-|------------------------------------------------|---------------------------------------------------------------------------------------------------|----------------------|
-| [**Medusa Multi-head LLM (Retraining Phase)**](https://github.com/linkedin/Liger-Kernel/tree/main/examples/medusa)        | Reduce memory usage by 80% with 5 LM heads and improve throughput by 40% using 8 A100s with FSDP      | TBA                  |
+| **Use Case**                                    | **Description**                                                                                   |
+|------------------------------------------------|---------------------------------------------------------------------------------------------------|
+| [**Hugging Face Trainer**](https://github.com/linkedin/Liger-Kernel/tree/main/examples/huggingface)      | Train LLaMA 3-8B ~20% faster with over 40% memory reduction on Alpaca dataset using 4 A100s with FSDP |
+| [**Lightning Trainer**](https://github.com/linkedin/Liger-Kernel/tree/main/examples/lightning)         | Increase 15% throughput and reduce memory usage by 40% with LLaMA3-8B on MMLU dataset using 8 A100s with DeepSpeed ZeRO3 |
+| [**Medusa Multi-head LLM (Retraining Phase)**](https://github.com/linkedin/Liger-Kernel/tree/main/examples/medusa)        | Reduce memory usage by 80% with 5 LM heads and improve throughput by 40% using 8 A100s with FSDP      |               |
 
 ## Key Features
 
@@ -149,13 +144,6 @@ With one line of code, Liger Kernel can increase throughput by more than 20% and
 - **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)
 
-## Target Audiences
-
-- **Researchers**: Looking to compose models using efficient and reliable kernels for frontier experiments.
-- **ML Practitioners**: Focused on maximizing GPU training efficiency with optimal, high-performance kernels.
-- **Curious Novices**: Eager to learn how to write reliable Triton kernels to enhance training efficiency.
-
-
 ## Installation
 
 ### Dependencies 
@@ -261,23 +249,6 @@ loss = loss_fn(model.weight, input, target)
 loss.backward()
 ```
 
-
-## Structure
-
-### Source Code
-
-- `ops/`: Core Triton operations.
-- `transformers/`: PyTorch `nn.Module` implementations built on Triton operations, compliant with the `transformers` API.
-
-### Tests
-
-- `transformers/`: Correctness tests for the Triton-based layers.
-- `convergence/`: Patches Hugging Face models with all kernels, runs multiple iterations, and compares weights, logits, and loss layer-by-layer.
-
-### Benchmark
-
-- `benchmark/`: Execution time and memory benchmarks compared to Hugging Face layers.
-
 ## APIs
 
 ### AutoModel
@@ -296,7 +267,7 @@ loss.backward()
 | Mistral     | `liger_kernel.transformers.apply_liger_kernel_to_mistral`  | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | 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         |
+| Gemma2      | `liger_kernel.transformers.apply_liger_kernel_to_gemma2`   | RoPE, RMSNorm, GeGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
 | Qwen2 & Qwen2.5      | `liger_kernel.transformers.apply_liger_kernel_to_qwen2`    | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | Qwen2-VL       | `liger_kernel.transformers.apply_liger_kernel_to_qwen2_vl`    | RMSNorm, LayerNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | Phi3 & Phi3.5       | `liger_kernel.transformers.apply_liger_kernel_to_phi3`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
@@ -320,6 +291,7 @@ loss.backward()
 
 - **RMSNorm**: [RMSNorm](https://arxiv.org/pdf/1910.07467), which normalizes activations using their root mean square, is implemented by fusing the normalization and scaling steps into a single Triton kernel, and achieves ~3X speedup with ~3X peak memory reduction.
 - **LayerNorm**: [LayerNorm](https://arxiv.org/pdf/1607.06450), which centers and normalizes activations across the feature dimension, is implemented by fusing the centering, normalization and scaling steps into a single Triton kernel, and achieves ~2X speedup.
+- **GroupNorm**: [GroupNorm](https://arxiv.org/pdf/1803.08494), which normalizes activations across the group dimension for a given sample. Channels are grouped in K groups over which the normalization is performed, is implemented by fusing the centering, normalization and scaling steps into a single Triton kernel, and can achieve up to ~2X speedup as the number of channels/groups increases.
 - **RoPE**: [Rotary Positional Embedding](https://arxiv.org/pdf/2104.09864) is implemented by fusing the query and key embedding rotary into a single kernel with inplace replacement, and achieves ~3X speedup with ~3X peak memory reduction.
 - **SwiGLU**: [Swish Gated Linear Units](https://arxiv.org/pdf/2002.05202), given by
 $$\text{SwiGLU}(x)=\text{Swish}_{\beta}(xW+b)\otimes(xV+c)$$
@@ -332,7 +304,7 @@ $$\text{GeGLU}(x)=\text{GELU}(xW+b)\otimes(xV+c)$$
 - **FusedLinearCrossEntropy**: Peak memory usage of cross entropy loss is further improved by fusing the model head with the CE loss and chunking the input for block-wise loss and gradient calculation, a technique inspired by [Efficient Cross Entropy](https://github.com/mgmalek/efficient_cross_entropy). It achieves >4X memory reduction for 128k vocab size. **This is highly effective for large batch size, large sequence length, and large vocabulary sizes.** Please refer to the [Medusa example](https://github.com/linkedin/Liger-Kernel/tree/main/examples/medusa) for individual kernel usage.
 - **KLDivergence**: [KL Divergence](https://pytorch.org/docs/stable/generated/torch.nn.KLDivLoss.html) is implemented by fusing the forward into a single triton kernel, with reduction done outside the kernel. It achieves ~1.5X speed and ~15% memory reduction for 128K vocab size.
 - **JSD**: [Generalized JSD](https://arxiv.org/pdf/2306.13649) (Jensen-Shannon divergence), is implemented by computing both the loss and gradient in the forward pass. It achieves ~1.5X speed and ~54% memory reduction for 128k vocab size.
-- **FusedLinearJSD**: Peak memory usage of JSD loss is further improved by fusing the model head with the model head with the JSD and chunking the input for block-wise loss and gradient calculation. It achieves ~85% memory reduction for 128k vocab size where batch size $\times$ sequence length is 8192.
+- **FusedLinearJSD**: Peak memory usage of JSD loss is further improved by fusing the model head with the JSD and chunking the input for block-wise loss and gradient calculation. It achieves ~85% memory reduction for 128k vocab size where batch size $\times$ sequence length is 8192.
 
 
 ### Experimental Kernels
@@ -345,54 +317,17 @@ $$\text{GeGLU}(x)=\text{GELU}(xW+b)\otimes(xV+c)$$
 - **Embedding**: [Embedding](https://pytorch.org/docs/stable/generated/torch.nn.Embedding.html) is implemented by fusing embedding lookup and output operations. It achieves a peak speedup of ~1.5x in the forward pass and an overall speedup of ~1.1x.
 - **Matmul int2xint8**: is implemented by using the cache tiled matrix multiplication and by fusing the matmul with the unpacking process which achieves a considerable speed up and performs on par with @torch.compile
 <!-- TODO: be more specific about batch size -->
-> **Note:**
-> Reported speedups and memory reductions are with respect to the LLaMA 3-8B Hugging Face layer implementations. All models use 4K hidden size and 4K sequence length and are evaluated based on memory usage and wall time for the forward+backward pass on a single NVIDIA A100 80G GPU using small batch sizes. Liger kernels exhibit more efficient scaling to larger batch sizes, detailed further in the [Benchmark](./benchmark) folder.
 
-## Contributing
+## Contributing, Acknowledgements, and License
 
-[CONTRIBUTING GUIDE](https://github.com/linkedin/Liger-Kernel/blob/main/CONTRIBUTING.md)
-
-## Acknowledgement
-
-
-### Design
-
-- [@claire_yishan](https://twitter.com/claire_yishan) for the LOGO design
-- [Wave Snippets](https://www.wavesnippets.com/) for generating the animated code snippets
-
-### Code
-
-We referenced or used the following projects:
-
-
-
-| # | Project                                                                                      | Description                                                                             | Location                                                                                                                         | License                                                                              |
-|---|----------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------|----------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------------------------------|
-| 1 | [Unsloth](https://github.com/unslothai/unsloth/blob/fd753fed99ed5f10ef8a9b7139588d9de9ddecfb/unsloth/kernels/utils.py#L43)                              | `calculate_settings` to determine block size and warp; We reuse it for Norm and MLP     | [Liger Kernel Utils](https://github.com/linkedin/Liger-Kernel/blob/e249eee723978bf8610ff1ea2297d048a2417e20/src/liger_kernel/ops/utils.py#L23) | [Apache](https://github.com/unslothai/unsloth/blob/fd753fed99ed5f10ef8a9b7139588d9de9ddecfb/LICENSE) |
-| 2 | [Unsloth](https://github.com/unslothai/unsloth/blob/976d11a10d54383aeb7a692c69e01151a20bfd72/unsloth/kernels/rms_layernorm.py#L48)                              | We modified and added dW calculation on top of Unsloth implementation                   | [Liger Kernel RMS Norm](https://github.com/linkedin/Liger-Kernel/blob/e249eee723978bf8610ff1ea2297d048a2417e20/src/liger_kernel/ops/rms_norm.py#L50)  | [Apache](https://github.com/unslothai/unsloth/blob/fd753fed99ed5f10ef8a9b7139588d9de9ddecfb/LICENSE) |
-| 3 | [Triton tutorial](https://triton-lang.org/main/index.html)                                    | We modified on top of triton tutorials                                                  | [Liger Kernel RMS Norm](https://github.com/linkedin/Liger-Kernel/blob/e249eee723978bf8610ff1ea2297d048a2417e20/src/liger_kernel/ops/rms_norm.py#L50)  | [MIT](https://github.com/triton-lang/triton/blob/main/LICENSE)                                  |
-| 4 | [tiny shakespeare dataset](https://huggingface.co/datasets/karpathy/tiny_shakespeare)         | We use tiny shakespeare dataset to conduct convergence test on mini model               | [Liger Kernel Convergence](https://github.com/linkedin/Liger-Kernel/tree/main/test/convergence)                                  | N/A                                                                                   |
-| 5 | [Efficient Cross Entropy](https://github.com/mgmalek/efficient_cross_entropy)                 | We use the idea of gradient-in-forward and chunking                                    | [Liger Kernel Linear Cross Entropy](https://github.com/linkedin/Liger-Kernel/blob/main/src/liger_kernel/ops/fused_linear_cross_entropy.py)          | [MIT](https://github.com/mgmalek/efficient_cross_entropy/blob/main/LICENSE)            |
-| 6 | [Flash attn](https://github.com/Dao-AILab/flash-attention)                                    | We take many optimization ideas from the work, such as tiling and recomputation         |                                                                                                                                  | [BSD](https://github.com/Dao-AILab/flash-attention/blob/main/LICENSE)                  |
-| 7 | [AutoAWQ](https://github.com/casper-hansen/AutoAWQ)                                           | We reference the design of automodel                                                   | [Liger Kernel Auto Model](https://github.com/linkedin/Liger-Kernel/blob/main/src/liger_kernel/transformers/auto_model.py)        | [MIT](https://github.com/casper-hansen/AutoAWQ/blob/main/LICENSE)                      |
-| 8 | [llm.c](https://github.com/karpathy/llm.c)                                                    | We reference the design of end-to-end testing                                          | [Liger Kernel Convergence Tests](https://github.com/linkedin/Liger-Kernel/tree/main/test/convergence)                            | [MIT](https://github.com/karpathy/llm.c/blob/master/LICENSE)                           |
-
-Many thanks to the contributors to these projects for their invaluable work that helped make Liger possible.
-
-## License
-
-This project is licensed under the [BSD 2-CLAUSE](https://github.com/linkedin/Liger-Kernel/blob/main/LICENSE) License (see `LICENSE` for details).
-It also includes components from projects licensed under:
-
-- Apache License 2.0 (see `LICENSE-APACHE-2.0` for details).
-- MIT License (see `LICENSE-MIT-AutoAWQ` for details).
-- MIT License (see `LICENSE-MIT-Efficient Cross Entropy` for details).
-- MIT License (see `LICENSE-MIT-llmc` for details).
-- MIT License (see `LICENSE-MIT-triton` for details).
+- [Contributing Guidelines](https://github.com/linkedin/Liger-Kernel/blob/main/docs/CONTRIBUTING.md)
+- [Acknowledgements](https://github.com/linkedin/Liger-Kernel/blob/main/docs/Acknowledgement.md)
+- [License Information](https://github.com/linkedin/Liger-Kernel/blob/main/docs/License.md)
 
 ## Contact
 
-- For public discussion, join [our discord channel](https://discord.gg/vNBDpjhb)
+- 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 byhsu@linkedin.com
 
 ## Cite this work
@@ -425,3 +360,4 @@ Biblatex entry:
         ↑ Back to Top ↑
     </a>
 </p>
+

{liger_kernel-0.4.0 → liger_kernel-0.4.2}/README.md

@@ -52,7 +52,8 @@
 
 <details>
   <summary>Latest News 🔥</summary>
-    
+
+  - [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!
   - [2024/10/21] We have released the tech report of Liger Kernel on Arxiv: https://arxiv.org/pdf/2410.10989 
   - [2024/9/6] We release v0.2.1 ([X post](https://x.com/liger_kernel/status/1832168197002510649)). 2500+ Stars, 10+ New Contributors, 50+ PRs, 50k Downloads in two weeks!
   - [2024/8/31] CUDA MODE talk, [Liger-Kernel: Real-world Triton kernel for LLM Training](https://youtu.be/gWble4FreV4?si=dxPeIchhkJ36Mbns), [Slides](https://github.com/cuda-mode/lectures?tab=readme-ov-file#lecture-28-liger-kernel)
@@ -80,18 +81,12 @@ With one line of code, Liger Kernel can increase throughput by more than 20% and
 
 ## Examples
 
-### Basic
-
-| **Example**                                    | **Description**                                                                                   | **Lightning Studio** |
-|------------------------------------------------|---------------------------------------------------------------------------------------------------|----------------------|
-| [**Hugging Face Trainer**](https://github.com/linkedin/Liger-Kernel/tree/main/examples/huggingface)      | Train LLaMA 3-8B ~20% faster with over 40% memory reduction on Alpaca dataset using 4 A100s with FSDP | TBA                  |
-| [**Lightning Trainer**](https://github.com/linkedin/Liger-Kernel/tree/main/examples/lightning)         | Increase 15% throughput and reduce memory usage by 40% with LLaMA3-8B on MMLU dataset using 8 A100s with DeepSpeed ZeRO3 | TBA                  |
 
-### Advanced
-
-| **Example**                                    | **Description**                                                                                   | **Lightning Studio** |
-|------------------------------------------------|---------------------------------------------------------------------------------------------------|----------------------|
-| [**Medusa Multi-head LLM (Retraining Phase)**](https://github.com/linkedin/Liger-Kernel/tree/main/examples/medusa)        | Reduce memory usage by 80% with 5 LM heads and improve throughput by 40% using 8 A100s with FSDP      | TBA                  |
+| **Use Case**                                    | **Description**                                                                                   |
+|------------------------------------------------|---------------------------------------------------------------------------------------------------|
+| [**Hugging Face Trainer**](https://github.com/linkedin/Liger-Kernel/tree/main/examples/huggingface)      | Train LLaMA 3-8B ~20% faster with over 40% memory reduction on Alpaca dataset using 4 A100s with FSDP |
+| [**Lightning Trainer**](https://github.com/linkedin/Liger-Kernel/tree/main/examples/lightning)         | Increase 15% throughput and reduce memory usage by 40% with LLaMA3-8B on MMLU dataset using 8 A100s with DeepSpeed ZeRO3 |
+| [**Medusa Multi-head LLM (Retraining Phase)**](https://github.com/linkedin/Liger-Kernel/tree/main/examples/medusa)        | Reduce memory usage by 80% with 5 LM heads and improve throughput by 40% using 8 A100s with FSDP      |               |
 
 ## Key Features
 
@@ -102,13 +97,6 @@ With one line of code, Liger Kernel can increase throughput by more than 20% and
 - **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)
 
-## Target Audiences
-
-- **Researchers**: Looking to compose models using efficient and reliable kernels for frontier experiments.
-- **ML Practitioners**: Focused on maximizing GPU training efficiency with optimal, high-performance kernels.
-- **Curious Novices**: Eager to learn how to write reliable Triton kernels to enhance training efficiency.
-
-
 ## Installation
 
 ### Dependencies 
@@ -214,23 +202,6 @@ loss = loss_fn(model.weight, input, target)
 loss.backward()
 ```
 
-
-## Structure
-
-### Source Code
-
-- `ops/`: Core Triton operations.
-- `transformers/`: PyTorch `nn.Module` implementations built on Triton operations, compliant with the `transformers` API.
-
-### Tests
-
-- `transformers/`: Correctness tests for the Triton-based layers.
-- `convergence/`: Patches Hugging Face models with all kernels, runs multiple iterations, and compares weights, logits, and loss layer-by-layer.
-
-### Benchmark
-
-- `benchmark/`: Execution time and memory benchmarks compared to Hugging Face layers.
-
 ## APIs
 
 ### AutoModel
@@ -249,7 +220,7 @@ loss.backward()
 | Mistral     | `liger_kernel.transformers.apply_liger_kernel_to_mistral`  | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | 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         |
+| Gemma2      | `liger_kernel.transformers.apply_liger_kernel_to_gemma2`   | RoPE, RMSNorm, GeGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
 | Qwen2 & Qwen2.5      | `liger_kernel.transformers.apply_liger_kernel_to_qwen2`    | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | Qwen2-VL       | `liger_kernel.transformers.apply_liger_kernel_to_qwen2_vl`    | RMSNorm, LayerNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | Phi3 & Phi3.5       | `liger_kernel.transformers.apply_liger_kernel_to_phi3`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
@@ -273,6 +244,7 @@ loss.backward()
 
 - **RMSNorm**: [RMSNorm](https://arxiv.org/pdf/1910.07467), which normalizes activations using their root mean square, is implemented by fusing the normalization and scaling steps into a single Triton kernel, and achieves ~3X speedup with ~3X peak memory reduction.
 - **LayerNorm**: [LayerNorm](https://arxiv.org/pdf/1607.06450), which centers and normalizes activations across the feature dimension, is implemented by fusing the centering, normalization and scaling steps into a single Triton kernel, and achieves ~2X speedup.
+- **GroupNorm**: [GroupNorm](https://arxiv.org/pdf/1803.08494), which normalizes activations across the group dimension for a given sample. Channels are grouped in K groups over which the normalization is performed, is implemented by fusing the centering, normalization and scaling steps into a single Triton kernel, and can achieve up to ~2X speedup as the number of channels/groups increases.
 - **RoPE**: [Rotary Positional Embedding](https://arxiv.org/pdf/2104.09864) is implemented by fusing the query and key embedding rotary into a single kernel with inplace replacement, and achieves ~3X speedup with ~3X peak memory reduction.
 - **SwiGLU**: [Swish Gated Linear Units](https://arxiv.org/pdf/2002.05202), given by
 $$\text{SwiGLU}(x)=\text{Swish}_{\beta}(xW+b)\otimes(xV+c)$$
@@ -285,7 +257,7 @@ $$\text{GeGLU}(x)=\text{GELU}(xW+b)\otimes(xV+c)$$
 - **FusedLinearCrossEntropy**: Peak memory usage of cross entropy loss is further improved by fusing the model head with the CE loss and chunking the input for block-wise loss and gradient calculation, a technique inspired by [Efficient Cross Entropy](https://github.com/mgmalek/efficient_cross_entropy). It achieves >4X memory reduction for 128k vocab size. **This is highly effective for large batch size, large sequence length, and large vocabulary sizes.** Please refer to the [Medusa example](https://github.com/linkedin/Liger-Kernel/tree/main/examples/medusa) for individual kernel usage.
 - **KLDivergence**: [KL Divergence](https://pytorch.org/docs/stable/generated/torch.nn.KLDivLoss.html) is implemented by fusing the forward into a single triton kernel, with reduction done outside the kernel. It achieves ~1.5X speed and ~15% memory reduction for 128K vocab size.
 - **JSD**: [Generalized JSD](https://arxiv.org/pdf/2306.13649) (Jensen-Shannon divergence), is implemented by computing both the loss and gradient in the forward pass. It achieves ~1.5X speed and ~54% memory reduction for 128k vocab size.
-- **FusedLinearJSD**: Peak memory usage of JSD loss is further improved by fusing the model head with the model head with the JSD and chunking the input for block-wise loss and gradient calculation. It achieves ~85% memory reduction for 128k vocab size where batch size $\times$ sequence length is 8192.
+- **FusedLinearJSD**: Peak memory usage of JSD loss is further improved by fusing the model head with the JSD and chunking the input for block-wise loss and gradient calculation. It achieves ~85% memory reduction for 128k vocab size where batch size $\times$ sequence length is 8192.
 
 
 ### Experimental Kernels
@@ -298,54 +270,17 @@ $$\text{GeGLU}(x)=\text{GELU}(xW+b)\otimes(xV+c)$$
 - **Embedding**: [Embedding](https://pytorch.org/docs/stable/generated/torch.nn.Embedding.html) is implemented by fusing embedding lookup and output operations. It achieves a peak speedup of ~1.5x in the forward pass and an overall speedup of ~1.1x.
 - **Matmul int2xint8**: is implemented by using the cache tiled matrix multiplication and by fusing the matmul with the unpacking process which achieves a considerable speed up and performs on par with @torch.compile
 <!-- TODO: be more specific about batch size -->
-> **Note:**
-> Reported speedups and memory reductions are with respect to the LLaMA 3-8B Hugging Face layer implementations. All models use 4K hidden size and 4K sequence length and are evaluated based on memory usage and wall time for the forward+backward pass on a single NVIDIA A100 80G GPU using small batch sizes. Liger kernels exhibit more efficient scaling to larger batch sizes, detailed further in the [Benchmark](./benchmark) folder.
 
-## Contributing
+## Contributing, Acknowledgements, and License
 
-[CONTRIBUTING GUIDE](https://github.com/linkedin/Liger-Kernel/blob/main/CONTRIBUTING.md)
-
-## Acknowledgement
-
-
-### Design
-
-- [@claire_yishan](https://twitter.com/claire_yishan) for the LOGO design
-- [Wave Snippets](https://www.wavesnippets.com/) for generating the animated code snippets
-
-### Code
-
-We referenced or used the following projects:
-
-
-
-| # | Project                                                                                      | Description                                                                             | Location                                                                                                                         | License                                                                              |
-|---|----------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------|----------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------------------------------|
-| 1 | [Unsloth](https://github.com/unslothai/unsloth/blob/fd753fed99ed5f10ef8a9b7139588d9de9ddecfb/unsloth/kernels/utils.py#L43)                              | `calculate_settings` to determine block size and warp; We reuse it for Norm and MLP     | [Liger Kernel Utils](https://github.com/linkedin/Liger-Kernel/blob/e249eee723978bf8610ff1ea2297d048a2417e20/src/liger_kernel/ops/utils.py#L23) | [Apache](https://github.com/unslothai/unsloth/blob/fd753fed99ed5f10ef8a9b7139588d9de9ddecfb/LICENSE) |
-| 2 | [Unsloth](https://github.com/unslothai/unsloth/blob/976d11a10d54383aeb7a692c69e01151a20bfd72/unsloth/kernels/rms_layernorm.py#L48)                              | We modified and added dW calculation on top of Unsloth implementation                   | [Liger Kernel RMS Norm](https://github.com/linkedin/Liger-Kernel/blob/e249eee723978bf8610ff1ea2297d048a2417e20/src/liger_kernel/ops/rms_norm.py#L50)  | [Apache](https://github.com/unslothai/unsloth/blob/fd753fed99ed5f10ef8a9b7139588d9de9ddecfb/LICENSE) |
-| 3 | [Triton tutorial](https://triton-lang.org/main/index.html)                                    | We modified on top of triton tutorials                                                  | [Liger Kernel RMS Norm](https://github.com/linkedin/Liger-Kernel/blob/e249eee723978bf8610ff1ea2297d048a2417e20/src/liger_kernel/ops/rms_norm.py#L50)  | [MIT](https://github.com/triton-lang/triton/blob/main/LICENSE)                                  |
-| 4 | [tiny shakespeare dataset](https://huggingface.co/datasets/karpathy/tiny_shakespeare)         | We use tiny shakespeare dataset to conduct convergence test on mini model               | [Liger Kernel Convergence](https://github.com/linkedin/Liger-Kernel/tree/main/test/convergence)                                  | N/A                                                                                   |
-| 5 | [Efficient Cross Entropy](https://github.com/mgmalek/efficient_cross_entropy)                 | We use the idea of gradient-in-forward and chunking                                    | [Liger Kernel Linear Cross Entropy](https://github.com/linkedin/Liger-Kernel/blob/main/src/liger_kernel/ops/fused_linear_cross_entropy.py)          | [MIT](https://github.com/mgmalek/efficient_cross_entropy/blob/main/LICENSE)            |
-| 6 | [Flash attn](https://github.com/Dao-AILab/flash-attention)                                    | We take many optimization ideas from the work, such as tiling and recomputation         |                                                                                                                                  | [BSD](https://github.com/Dao-AILab/flash-attention/blob/main/LICENSE)                  |
-| 7 | [AutoAWQ](https://github.com/casper-hansen/AutoAWQ)                                           | We reference the design of automodel                                                   | [Liger Kernel Auto Model](https://github.com/linkedin/Liger-Kernel/blob/main/src/liger_kernel/transformers/auto_model.py)        | [MIT](https://github.com/casper-hansen/AutoAWQ/blob/main/LICENSE)                      |
-| 8 | [llm.c](https://github.com/karpathy/llm.c)                                                    | We reference the design of end-to-end testing                                          | [Liger Kernel Convergence Tests](https://github.com/linkedin/Liger-Kernel/tree/main/test/convergence)                            | [MIT](https://github.com/karpathy/llm.c/blob/master/LICENSE)                           |
-
-Many thanks to the contributors to these projects for their invaluable work that helped make Liger possible.
-
-## License
-
-This project is licensed under the [BSD 2-CLAUSE](https://github.com/linkedin/Liger-Kernel/blob/main/LICENSE) License (see `LICENSE` for details).
-It also includes components from projects licensed under:
-
-- Apache License 2.0 (see `LICENSE-APACHE-2.0` for details).
-- MIT License (see `LICENSE-MIT-AutoAWQ` for details).
-- MIT License (see `LICENSE-MIT-Efficient Cross Entropy` for details).
-- MIT License (see `LICENSE-MIT-llmc` for details).
-- MIT License (see `LICENSE-MIT-triton` for details).
+- [Contributing Guidelines](https://github.com/linkedin/Liger-Kernel/blob/main/docs/CONTRIBUTING.md)
+- [Acknowledgements](https://github.com/linkedin/Liger-Kernel/blob/main/docs/Acknowledgement.md)
+- [License Information](https://github.com/linkedin/Liger-Kernel/blob/main/docs/License.md)
 
 ## Contact
 
-- For public discussion, join [our discord channel](https://discord.gg/vNBDpjhb)
+- 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 byhsu@linkedin.com
 
 ## Cite this work
@@ -378,3 +313,4 @@ Biblatex entry:
         ↑ Back to Top ↑
     </a>
 </p>
+

{liger_kernel-0.4.0 → liger_kernel-0.4.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel"
-version = "0.4.0"
+version = "0.4.2"
 description = "Efficient Triton kernels for LLM Training"
 urls = { "Homepage" = "https://github.com/linkedin/Liger-Kernel" }
 readme = { file = "README.md", content-type = "text/markdown" }

liger_kernel-0.4.2/src/liger_kernel/chunked_loss/dpo_loss.py

@@ -0,0 +1,57 @@
+import torch.nn.functional as F
+
+from liger_kernel.chunked_loss.fused_linear_preference import (
+    LigerFusedLinearPreferenceBase,
+)
+
+
+class LigerFusedLinearDPOFunction(LigerFusedLinearPreferenceBase):
+
+    @staticmethod
+    def preference_loss_fn(chosen_logps, rejected_logps, beta=0.1):
+        """
+        Compute DPO loss (Direct Preference Optimization).
+        Args:
+            chosen_logps (torch.Tensor): Avg log probabilities of chosen tokens. Shape: (batch_size,).
+            rejected_logps (torch.Tensor): Avg log probabilities of rejected tokens. Shape: (batch_size,).
+            beta (float): Weight for the direct preference loss.
+        """
+        logits_diff = beta * (chosen_logps - rejected_logps)
+        losses = -F.logsigmoid(logits_diff)
+        return losses.sum()
+
+    @staticmethod
+    def forward(
+        ctx,
+        _input,
+        weight,
+        target,
+        bias=None,
+        ignore_index=-100,
+        beta=0.1,
+        compute_nll_loss=True,
+        compiled=True,
+    ):
+        """
+        Fused linear layer with DPO (Direct Preference Optimization) loss.
+        Handles both the forward and backward pass of the final linear layer with DPO loss.
+        """
+        return LigerFusedLinearPreferenceBase.forward(
+            ctx=ctx,
+            _input=_input,
+            weight=weight,
+            target=target,
+            bias=bias,
+            loss_fn=LigerFusedLinearDPOFunction.preference_loss_fn,
+            compute_nll_loss=compute_nll_loss,
+            ignore_index=ignore_index,
+            beta=beta,
+            compiled=compiled,
+        )
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        # Get gradients for _input, weight, bias, and target from the base class
+        grads = LigerFusedLinearPreferenceBase.backward(ctx, grad_output)[:4]
+        # Return these gradients, followed by None for the remaining inputs
+        return *grads, None, None, None, None

liger_kernel-0.4.2/src/liger_kernel/chunked_loss/fused_linear_preference.py

@@ -0,0 +1,206 @@
+from abc import abstractmethod
+from functools import partial
+
+import torch
+from torch.nn import functional as F
+
+
+class LigerFusedLinearPreferenceBase(torch.autograd.Function):
+
+    @abstractmethod
+    def preference_loss_fn(chosen_logps, rejected_logps, beta=0.1):
+        """
+        Compute preference loss.
+        Args:
+            chosen_logps (torch.Tensor): Avg log probabilities of chosen tokens. Shape: (batch_size,).
+            rejected_logps (torch.Tensor): Avg log probabilities of rejected tokens. Shape: (batch_size,).
+            beta (float): Weight for the odds ratio loss.
+        """
+        raise NotImplementedError("Preference loss function must be implemented.")
+
+    @staticmethod
+    def forward(
+        ctx,
+        _input,
+        weight,
+        target,
+        bias=None,
+        loss_fn=None,
+        chunk_size=1,
+        compute_nll_loss=True,
+        ignore_index=-100,
+        beta=0.1,
+        compiled=True,
+    ):
+        """
+        Base class for fused linear layer with preference loss.
+        Expects _input to be stacked with chosen and rejected inputs on the batch dimension.
+
+        Args:
+            _input (torch.Tensor): Input tensor. Shape: (batch_size, seq_len, hidden_size).
+            weight (torch.Tensor): Weight tensor. Shape: (vocab_size, hidden_size).
+            target (torch.Tensor): Target tensor. Shape: (batch_size, seq_len).
+            bias (torch.Tensor, optional): Bias tensor. Shape: (vocab_size,).
+            loss_fn (callable): Loss function to compute the loss on a chunk of input/target.
+            chunk_size (int): Size of a chunk (# of batches of stacked chosen and rejected inputs).
+            compute_nll_loss (bool): Whether to compute NLL loss.
+            ignore_index (int): Index to ignore for loss computation.
+            beta (float): Weight for the odds ratio loss.
+            compiled (bool): Whether to use torch compile for chunk accumulation.
+        """
+        # TODO: Tune CHUNK_SIZE to fully utilize the GPU
+        CHUNK_SIZE = chunk_size
+
+        grad_weight = torch.zeros_like(weight)
+        grad_chosen_inputs = []
+        grad_rejected_inputs = []
+        grad_bias = torch.zeros_like(bias) if bias is not None else None
+        loss_acc = torch.zeros((), device=_input.device)
+
+        chunks = max(1, _input.shape[0] // (2 * CHUNK_SIZE))
+        loss_func_to_call = partial(
+            LigerFusedLinearPreferenceBase._compute_loss,
+            preference_loss_fn=loss_fn,
+            ignore_index=ignore_index,
+            beta=beta,
+            compute_nll_loss=compute_nll_loss,
+            full_target=target,
+        )
+
+        def accumulate_chunk(input_chunk, target_chunk):
+            if bias is not None:
+                (chunk_grad_input, chunk_grad_weight, chunk_grad_bias), (
+                    chunk_loss,
+                    (chunk_or_loss, chunk_chosen_logps, chunk_rejected_logps),
+                ) = torch.func.grad_and_value(
+                    loss_func_to_call, argnums=(0, 1, 3), has_aux=True
+                )(
+                    input_chunk, weight, target_chunk, bias
+                )
+                grad_bias.add_(chunk_grad_bias)
+            else:
+                (chunk_grad_input, chunk_grad_weight), (
+                    chunk_loss,
+                    (chunk_or_loss, chunk_chosen_logps, chunk_rejected_logps),
+                ) = torch.func.grad_and_value(
+                    loss_func_to_call, argnums=(0, 1), has_aux=True
+                )(
+                    input_chunk, weight, target_chunk
+                )
+            grad_weight.add_(chunk_grad_weight)
+            loss_acc.add_(chunk_loss)
+            return chunk_grad_input
+
+        len_chosen = target.shape[0] // 2
+        _chosen_input_chunks = torch.chunk(_input[:len_chosen], chunks=chunks, dim=0)
+        _chosen_target_chunks = torch.chunk(target[:len_chosen], chunks=chunks, dim=0)
+        _rejected_input_chunks = torch.chunk(_input[len_chosen:], chunks=chunks, dim=0)
+        _rejected_target_chunks = torch.chunk(target[len_chosen:], chunks=chunks, dim=0)
+
+        for (
+            chosen_input_chunk,
+            rejected_input_chunk,
+            chosen_target_chunk,
+            rejected_target_chunk,
+        ) in zip(
+            _chosen_input_chunks,
+            _rejected_input_chunks,
+            _chosen_target_chunks,
+            _rejected_target_chunks,
+        ):
+            input_chunk = torch.cat([chosen_input_chunk, rejected_input_chunk], dim=0)
+            target_chunk = torch.cat(
+                [chosen_target_chunk, rejected_target_chunk], dim=0
+            )
+
+            if compiled:
+                accumulate_chunk = torch.compile(accumulate_chunk)
+            grad_input = accumulate_chunk(input_chunk, target_chunk)
+
+            grad_chosen_inputs.append(grad_input[: chosen_target_chunk.shape[0]])
+            grad_rejected_inputs.append(grad_input[chosen_target_chunk.shape[0] :])
+
+        # combine grad_chosen_inputs and grad_rejected_inputs
+        grad_inputs = grad_chosen_inputs + grad_rejected_inputs
+
+        ctx.save_for_backward(
+            torch.cat(grad_inputs, dim=0),
+            grad_weight,
+            grad_bias,
+        )
+        return loss_acc
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        grad_input, grad_weight, grad_bias = ctx.saved_tensors
+        if torch.ne(grad_output, torch.tensor(1.0, device=grad_output.device)):
+            grad_input = grad_input * grad_output
+            grad_weight = grad_weight * grad_output
+            grad_bias = grad_bias * grad_output if grad_bias is not None else None
+
+        return grad_input, grad_weight, None, grad_bias, None, None, None
+
+    @staticmethod
+    def _compute_loss(
+        input_chunk,
+        weight,
+        target_chunk,
+        bias=None,
+        preference_loss_fn=None,
+        full_target=None,
+        ignore_index=-100,
+        beta=0.1,
+        compute_nll_loss=True,
+        **loss_kwargs,
+    ):
+        """
+        Compute the total loss for a chunk of input and target, while using an alignment/preference loss function.
+        Args:
+            preference_loss_fn (callable): Loss function to compute the loss on a chunk of input/target.
+            input_chunk (torch.Tensor): Chunk of input tensor. Shape: (2 * chunk_size, sequence_length, hidden_size).
+            weight (torch.Tensor): Weight tensor. Shape: (vocab_size, hidden_size).
+            target_chunk (torch.Tensor): Chunk of target tensor. Shape: (2 * chunk_size, sequence_length).
+            bias (torch.Tensor, optional): Bias tensor. Shape: (vocab_size,).
+            full_target (torch.Tensor): Full target tensor. Shape: (batch_size, sequence_length).
+            ignore_index (int): Index to ignore for loss computation.
+            beta (float): Weight for the odds ratio loss.
+            loss_kwargs (dict): Additional arguments for the loss function.
+        """
+        len_chosen_chunk = target_chunk.shape[0] // 2
+
+        logits_chunk = input_chunk @ weight.t()  # chunk_size x V
+        if bias is not None:
+            logits_chunk = logits_chunk + bias
+        log_probs_chunk = F.log_softmax(logits_chunk.float(), dim=-1)
+
+        chosen_nll_loss = 0.0
+        if compute_nll_loss:
+            chosen_nll_loss = F.nll_loss(
+                log_probs_chunk[:len_chosen_chunk].view(-1, log_probs_chunk.shape[-1]),
+                target_chunk[:len_chosen_chunk].view(-1),
+                reduction="sum",
+                ignore_index=ignore_index,
+            )
+            chosen_nll_loss = (
+                chosen_nll_loss
+                / (full_target[: full_target.shape[0] // 2] != ignore_index).sum()
+            )
+
+        loss_mask = target_chunk != ignore_index
+        label_chunk = torch.where(loss_mask, target_chunk, 0)
+
+        per_token_logps = log_probs_chunk.gather(-1, label_chunk.unsqueeze(-1)).squeeze(
+            -1
+        )
+        average_log_prob = (per_token_logps * loss_mask).sum(-1) / loss_mask.sum(-1)
+
+        chosen_logps = average_log_prob[:len_chosen_chunk]
+        rejected_logps = average_log_prob[len_chosen_chunk:]
+
+        alignment_loss = preference_loss_fn(
+            chosen_logps, rejected_logps, beta=beta, **loss_kwargs
+        )
+        alignment_loss = alignment_loss / (full_target.shape[0] // 2)
+
+        loss = chosen_nll_loss - alignment_loss
+        return loss, (alignment_loss, chosen_logps, rejected_logps)