liger-kernel-nightly 0.4.2.dev20241209234352__tar.gz → 0.4.2.dev20241210002150__tar.gz

{liger_kernel_nightly-0.4.2.dev20241209234352/src/liger_kernel_nightly.egg-info → liger_kernel_nightly-0.4.2.dev20241210002150}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.4.2.dev20241209234352
+Version: 0.4.2.dev20241210002150
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -110,7 +110,7 @@ Requires-Dist: triton>=3.0.0; extra == "amd"
 
 <img src="https://raw.githubusercontent.com/linkedin/Liger-Kernel/main/docs/images/logo-banner.png">
 
-[Installation](#installation) | [Getting Started](#getting-started) | [Examples](#examples) | [APIs](#apis) | [Cite our work](#cite-this-work)
+[Installation](#installation) | [Getting Started](#getting-started) | [Examples](#examples) | [High-level APIs](#high-level-apis) | [Low-level APIs](#low-level-apis) | [Cite our work](#cite-this-work)
 
 <details>
   <summary>Latest News 🔥</summary>
@@ -266,7 +266,7 @@ loss = loss_fn(model.weight, input, target)
 loss.backward()
 ```
 
-## APIs
+## High-level APIs
 
 ### AutoModel
 
@@ -290,8 +290,12 @@ loss.backward()
 | Phi3 & Phi3.5       | `liger_kernel.transformers.apply_liger_kernel_to_phi3`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
 
 
+## Low-level APIs
 
-### Kernels
+- `Fused Linear` kernels combine linear layers with losses, reducing memory usage by up to 80% - ideal for HBM-constrained workloads.
+- Other kernels use fusion and in-place techniques for memory and performance optimization.
+
+### Model Kernels
 
 | **Kernel**                      | **API**                                                     |
 |---------------------------------|-------------------------------------------------------------|
@@ -301,39 +305,33 @@ loss.backward()
 | SwiGLU                          | `liger_kernel.transformers.LigerSwiGLUMLP`                  |
 | GeGLU                           | `liger_kernel.transformers.LigerGEGLUMLP`                   |
 | CrossEntropy                    | `liger_kernel.transformers.LigerCrossEntropyLoss`           |
-| FusedLinearCrossEntropy         | `liger_kernel.transformers.LigerFusedLinearCrossEntropyLoss`|
+| Fused Linear CrossEntropy         | `liger_kernel.transformers.LigerFusedLinearCrossEntropyLoss`|
+
+
+### Alignment Kernels
+
+| **Kernel**                      | **API**                                                     |
+|---------------------------------|-------------------------------------------------------------|
+| Fused Linear CPO Loss           | `liger_kernel.chunked_loss.LigerFusedLinearCPOLoss`       |
+| Fused Linear DPO Loss           | `liger_kernel.chunked_loss.LigerFusedLinearDPOLoss`       |
+| Fused Linear ORPO Loss          | `liger_kernel.chunked_loss.LigerFusedLinearORPOLoss`      |
+| Fused Linear SimPO Loss         | `liger_kernel.chunked_loss.LigerFusedLinearSimPOLoss`     |
+
+### Distillation Kernels
+
+| **Kernel**                      | **API**                                                     |
+|---------------------------------|-------------------------------------------------------------|
 | KLDivergence                    | `liger_kernel.transformers.LigerKLDIVLoss`                  |
 | JSD                             | `liger_kernel.transformers.LigerJSD`                        |
-| FusedLinearJSD                  | `liger_kernel.transformers.LigerFusedLinearJSD`             |
-
-- **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
-$$\text{SwiGLU}(x)=\text{Swish}_{\beta}(xW+b)\otimes(xV+c)$$
-, is implemented by fusing the elementwise multiplication (denoted by $\otimes$) into a single kernel with inplace replacement, and achieves parity speed with ~1.5X peak memory reduction.
-- **GeGLU**: [GELU Gated Linear Units](https://arxiv.org/pdf/2002.05202), given by
-$$\text{GeGLU}(x)=\text{GELU}(xW+b)\otimes(xV+c)$$
-, is implemented by fusing the elementwise multiplication into a single kernel with inplace replacement, and achieves parity speed with ~1.5X peak memory reduction. Note that the [tanh approximation form of GELU](https://pytorch.org/docs/stable/generated/torch.nn.GELU.html) is used.
-- **CrossEntropy**: [Cross entropy loss](https://pytorch.org/docs/stable/generated/torch.nn.CrossEntropyLoss.html) is implemented by computing both the loss and gradient in the forward pass with inplace replacement of input to reduce the peak memory by avoiding simultaneous materialization of both input logits and gradient. It achieves >2X speedup and >4X memory reduction for common vocab sizes (e.g., 32K, 128K, etc.).
-<!-- TODO: verify vocab sizes are accurate  -->
-- **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. **NOTE**: It implements forward/reverse KL when `beta` equals 0 and 1 respectively.
-- **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. **NOTE**: It implements forward/reverse KL when `beta` equals 0 and 1 respectively.
-
+| Fused Linear JSD                  | `liger_kernel.transformers.LigerFusedLinearJSD`             |
 
 ### Experimental Kernels
 
 | **Kernel**                      | **API**                                                     |
 |---------------------------------|-------------------------------------------------------------|
 | Embedding                       | `liger_kernel.transformers.experimental.LigerEmbedding`     |
-| Matmul int2xint8                | `liger_kernel.transformers.experimental.matmul`
+| Matmul int2xint8                | `liger_kernel.transformers.experimental.matmul` |
 
-- **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 -->
 
 ## Contributing, Acknowledgements, and License
 

{liger_kernel_nightly-0.4.2.dev20241209234352 → liger_kernel_nightly-0.4.2.dev20241210002150}/README.md

@@ -55,7 +55,7 @@
 
 <img src="https://raw.githubusercontent.com/linkedin/Liger-Kernel/main/docs/images/logo-banner.png">
 
-[Installation](#installation) | [Getting Started](#getting-started) | [Examples](#examples) | [APIs](#apis) | [Cite our work](#cite-this-work)
+[Installation](#installation) | [Getting Started](#getting-started) | [Examples](#examples) | [High-level APIs](#high-level-apis) | [Low-level APIs](#low-level-apis) | [Cite our work](#cite-this-work)
 
 <details>
   <summary>Latest News 🔥</summary>
@@ -211,7 +211,7 @@ loss = loss_fn(model.weight, input, target)
 loss.backward()
 ```
 
-## APIs
+## High-level APIs
 
 ### AutoModel
 
@@ -235,8 +235,12 @@ loss.backward()
 | Phi3 & Phi3.5       | `liger_kernel.transformers.apply_liger_kernel_to_phi3`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
 
 
+## Low-level APIs
 
-### Kernels
+- `Fused Linear` kernels combine linear layers with losses, reducing memory usage by up to 80% - ideal for HBM-constrained workloads.
+- Other kernels use fusion and in-place techniques for memory and performance optimization.
+
+### Model Kernels
 
 | **Kernel**                      | **API**                                                     |
 |---------------------------------|-------------------------------------------------------------|
@@ -246,39 +250,33 @@ loss.backward()
 | SwiGLU                          | `liger_kernel.transformers.LigerSwiGLUMLP`                  |
 | GeGLU                           | `liger_kernel.transformers.LigerGEGLUMLP`                   |
 | CrossEntropy                    | `liger_kernel.transformers.LigerCrossEntropyLoss`           |
-| FusedLinearCrossEntropy         | `liger_kernel.transformers.LigerFusedLinearCrossEntropyLoss`|
+| Fused Linear CrossEntropy         | `liger_kernel.transformers.LigerFusedLinearCrossEntropyLoss`|
+
+
+### Alignment Kernels
+
+| **Kernel**                      | **API**                                                     |
+|---------------------------------|-------------------------------------------------------------|
+| Fused Linear CPO Loss           | `liger_kernel.chunked_loss.LigerFusedLinearCPOLoss`       |
+| Fused Linear DPO Loss           | `liger_kernel.chunked_loss.LigerFusedLinearDPOLoss`       |
+| Fused Linear ORPO Loss          | `liger_kernel.chunked_loss.LigerFusedLinearORPOLoss`      |
+| Fused Linear SimPO Loss         | `liger_kernel.chunked_loss.LigerFusedLinearSimPOLoss`     |
+
+### Distillation Kernels
+
+| **Kernel**                      | **API**                                                     |
+|---------------------------------|-------------------------------------------------------------|
 | KLDivergence                    | `liger_kernel.transformers.LigerKLDIVLoss`                  |
 | JSD                             | `liger_kernel.transformers.LigerJSD`                        |
-| FusedLinearJSD                  | `liger_kernel.transformers.LigerFusedLinearJSD`             |
-
-- **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
-$$\text{SwiGLU}(x)=\text{Swish}_{\beta}(xW+b)\otimes(xV+c)$$
-, is implemented by fusing the elementwise multiplication (denoted by $\otimes$) into a single kernel with inplace replacement, and achieves parity speed with ~1.5X peak memory reduction.
-- **GeGLU**: [GELU Gated Linear Units](https://arxiv.org/pdf/2002.05202), given by
-$$\text{GeGLU}(x)=\text{GELU}(xW+b)\otimes(xV+c)$$
-, is implemented by fusing the elementwise multiplication into a single kernel with inplace replacement, and achieves parity speed with ~1.5X peak memory reduction. Note that the [tanh approximation form of GELU](https://pytorch.org/docs/stable/generated/torch.nn.GELU.html) is used.
-- **CrossEntropy**: [Cross entropy loss](https://pytorch.org/docs/stable/generated/torch.nn.CrossEntropyLoss.html) is implemented by computing both the loss and gradient in the forward pass with inplace replacement of input to reduce the peak memory by avoiding simultaneous materialization of both input logits and gradient. It achieves >2X speedup and >4X memory reduction for common vocab sizes (e.g., 32K, 128K, etc.).
-<!-- TODO: verify vocab sizes are accurate  -->
-- **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. **NOTE**: It implements forward/reverse KL when `beta` equals 0 and 1 respectively.
-- **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. **NOTE**: It implements forward/reverse KL when `beta` equals 0 and 1 respectively.
-
+| Fused Linear JSD                  | `liger_kernel.transformers.LigerFusedLinearJSD`             |
 
 ### Experimental Kernels
 
 | **Kernel**                      | **API**                                                     |
 |---------------------------------|-------------------------------------------------------------|
 | Embedding                       | `liger_kernel.transformers.experimental.LigerEmbedding`     |
-| Matmul int2xint8                | `liger_kernel.transformers.experimental.matmul`
+| Matmul int2xint8                | `liger_kernel.transformers.experimental.matmul` |
 
-- **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 -->
 
 ## Contributing, Acknowledgements, and License
 

{liger_kernel_nightly-0.4.2.dev20241209234352 → liger_kernel_nightly-0.4.2.dev20241210002150}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.4.2.dev20241209234352"
+version = "0.4.2.dev20241210002150"
 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_nightly-0.4.2.dev20241209234352 → liger_kernel_nightly-0.4.2.dev20241210002150}/src/liger_kernel/ops/cross_entropy.py

@@ -285,11 +285,12 @@ def cross_entropy_forward(
         num_warps=32 if not is_hip() else 16,
     )
 
-    loss = torch.sum(loss_1d)
-    if return_z_loss == _TRUE.value:
-        z_loss = torch.sum(z_loss_1d)
+    if reduction == "none":
+        loss = loss_1d
+        z_loss = z_loss_1d if return_z_loss == _TRUE.value else None
     else:
-        z_loss = None
+        loss = torch.sum(loss_1d)
+        z_loss = torch.sum(z_loss_1d) if return_z_loss == _TRUE.value else None
 
     return loss, z_loss, _input
 

{liger_kernel_nightly-0.4.2.dev20241209234352 → liger_kernel_nightly-0.4.2.dev20241210002150/src/liger_kernel_nightly.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.4.2.dev20241209234352
+Version: 0.4.2.dev20241210002150
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -110,7 +110,7 @@ Requires-Dist: triton>=3.0.0; extra == "amd"
 
 <img src="https://raw.githubusercontent.com/linkedin/Liger-Kernel/main/docs/images/logo-banner.png">
 
-[Installation](#installation) | [Getting Started](#getting-started) | [Examples](#examples) | [APIs](#apis) | [Cite our work](#cite-this-work)
+[Installation](#installation) | [Getting Started](#getting-started) | [Examples](#examples) | [High-level APIs](#high-level-apis) | [Low-level APIs](#low-level-apis) | [Cite our work](#cite-this-work)
 
 <details>
   <summary>Latest News 🔥</summary>
@@ -266,7 +266,7 @@ loss = loss_fn(model.weight, input, target)
 loss.backward()
 ```
 
-## APIs
+## High-level APIs
 
 ### AutoModel
 
@@ -290,8 +290,12 @@ loss.backward()
 | Phi3 & Phi3.5       | `liger_kernel.transformers.apply_liger_kernel_to_phi3`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
 
 
+## Low-level APIs
 
-### Kernels
+- `Fused Linear` kernels combine linear layers with losses, reducing memory usage by up to 80% - ideal for HBM-constrained workloads.
+- Other kernels use fusion and in-place techniques for memory and performance optimization.
+
+### Model Kernels
 
 | **Kernel**                      | **API**                                                     |
 |---------------------------------|-------------------------------------------------------------|
@@ -301,39 +305,33 @@ loss.backward()
 | SwiGLU                          | `liger_kernel.transformers.LigerSwiGLUMLP`                  |
 | GeGLU                           | `liger_kernel.transformers.LigerGEGLUMLP`                   |
 | CrossEntropy                    | `liger_kernel.transformers.LigerCrossEntropyLoss`           |
-| FusedLinearCrossEntropy         | `liger_kernel.transformers.LigerFusedLinearCrossEntropyLoss`|
+| Fused Linear CrossEntropy         | `liger_kernel.transformers.LigerFusedLinearCrossEntropyLoss`|
+
+
+### Alignment Kernels
+
+| **Kernel**                      | **API**                                                     |
+|---------------------------------|-------------------------------------------------------------|
+| Fused Linear CPO Loss           | `liger_kernel.chunked_loss.LigerFusedLinearCPOLoss`       |
+| Fused Linear DPO Loss           | `liger_kernel.chunked_loss.LigerFusedLinearDPOLoss`       |
+| Fused Linear ORPO Loss          | `liger_kernel.chunked_loss.LigerFusedLinearORPOLoss`      |
+| Fused Linear SimPO Loss         | `liger_kernel.chunked_loss.LigerFusedLinearSimPOLoss`     |
+
+### Distillation Kernels
+
+| **Kernel**                      | **API**                                                     |
+|---------------------------------|-------------------------------------------------------------|
 | KLDivergence                    | `liger_kernel.transformers.LigerKLDIVLoss`                  |
 | JSD                             | `liger_kernel.transformers.LigerJSD`                        |
-| FusedLinearJSD                  | `liger_kernel.transformers.LigerFusedLinearJSD`             |
-
-- **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
-$$\text{SwiGLU}(x)=\text{Swish}_{\beta}(xW+b)\otimes(xV+c)$$
-, is implemented by fusing the elementwise multiplication (denoted by $\otimes$) into a single kernel with inplace replacement, and achieves parity speed with ~1.5X peak memory reduction.
-- **GeGLU**: [GELU Gated Linear Units](https://arxiv.org/pdf/2002.05202), given by
-$$\text{GeGLU}(x)=\text{GELU}(xW+b)\otimes(xV+c)$$
-, is implemented by fusing the elementwise multiplication into a single kernel with inplace replacement, and achieves parity speed with ~1.5X peak memory reduction. Note that the [tanh approximation form of GELU](https://pytorch.org/docs/stable/generated/torch.nn.GELU.html) is used.
-- **CrossEntropy**: [Cross entropy loss](https://pytorch.org/docs/stable/generated/torch.nn.CrossEntropyLoss.html) is implemented by computing both the loss and gradient in the forward pass with inplace replacement of input to reduce the peak memory by avoiding simultaneous materialization of both input logits and gradient. It achieves >2X speedup and >4X memory reduction for common vocab sizes (e.g., 32K, 128K, etc.).
-<!-- TODO: verify vocab sizes are accurate  -->
-- **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. **NOTE**: It implements forward/reverse KL when `beta` equals 0 and 1 respectively.
-- **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. **NOTE**: It implements forward/reverse KL when `beta` equals 0 and 1 respectively.
-
+| Fused Linear JSD                  | `liger_kernel.transformers.LigerFusedLinearJSD`             |
 
 ### Experimental Kernels
 
 | **Kernel**                      | **API**                                                     |
 |---------------------------------|-------------------------------------------------------------|
 | Embedding                       | `liger_kernel.transformers.experimental.LigerEmbedding`     |
-| Matmul int2xint8                | `liger_kernel.transformers.experimental.matmul`
+| Matmul int2xint8                | `liger_kernel.transformers.experimental.matmul` |
 
-- **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 -->
 
 ## Contributing, Acknowledgements, and License