liger-kernel 0.4.2__py3-none-any.whl → 0.5.0__py3-none-any.whl

liger_kernel/transformers/orpo_trainer.py

@@ -0,0 +1,171 @@
+from typing import Any, Callable, Dict, List, Literal, Tuple, Union
+
+import torch
+import torch.nn as nn
+from torch.distributed.fsdp import FullyShardedDataParallel
+from trl.trainer import ORPOTrainer
+
+from liger_kernel.chunked_loss import LigerFusedLinearORPOLoss
+
+
+class _FSDPForwardRedirection:
+    """
+    Modified based on
+    https://github.com/Lightning-AI/pytorch-lightning/blob/d3f9c83d6efa4f1def36aa6c199600946cdb9117/src/lightning/pytorch/strategies/strategy.py#L601-L648
+    Redirect a method call through FullyShardedDataParallel.forward so that the FSDP module's root pre-forward and
+    post-forward can be properly executed around the method call.
+    This is needed in cases where we call a submodule of a FSDP module. For instance, when we want to call only
+    the `LlamaModel` part out of a FSDP-wrapped `LlamaForCausalLM` to get the hidden states without involving
+    GPU-memory-heavy `lm_head` and cross entropy computation, doing this directly (i.e. `model.model.forward()`)
+    will not work because the first `nn.Emebedding` layer is not independently wrapped as a FSDP module (because of
+    the transformer-based wrapping policy), and not calling it through FSDP root module forward will not all-gather
+    its parameter, thus resulting in "RuntimeError: 'weight' must be 2-D" error. Similarly, if we want to call just
+    the `lm_head` part of a model, we need this trick too to properly get its params all-gathered.
+    """
+
+    def __call__(
+        self,
+        wrapper_module: FullyShardedDataParallel,
+        method: Callable,
+        *args: Any,
+        **kwargs: Any,
+    ):
+        """Reroutes a method call through the `wrapper_module`'s `forward` method.
+        Args:
+            wrapper_module: The module that has `original_module` wrapped.
+            original_module: The module that was wrapped inside `wrapper_module`.
+            method_name: The name of the method that should be called on the `original_module` after inputs get
+                redirected through the `wrapper_module`'s `forward` method.
+            *args: The positional arguments to the method `method_name`. They will get passed to a patched
+                `forward` method instead.
+            **kwargs: The keyword arguments to the method `method_name`. They will get passed to a patched
+                `forward` method instead.
+        """
+        assert isinstance(wrapper_module, FullyShardedDataParallel)
+        original_module = wrapper_module._fsdp_wrapped_module
+        original_forward = original_module.forward
+
+        def wrapped_forward(*_args: Any, **_kwargs: Any) -> Any:
+            # Unpatch ourselves immediately before calling the method `method_name`
+            # because itself may want to call the real `forward`
+            original_module.forward = original_forward  # type: ignore[method-assign]
+            # Call the actual method e.g. `.training_step(...)`
+            out = method(*_args, **_kwargs)
+            return out
+
+        # Patch the original_module's forward so we can redirect the arguments back to the real method
+        original_module.forward = wrapped_forward  # type: ignore[method-assign]
+        wrapper_output = wrapper_module(*args, **kwargs)
+        return wrapper_output
+
+
+class LigerORPOTrainer(ORPOTrainer):
+    def concatenated_forward(
+        self, model: nn.Module, batch: Dict[str, Union[List, torch.LongTensor]]
+    ) -> Tuple[
+        torch.FloatTensor, torch.FloatTensor, torch.FloatTensor, torch.FloatTensor
+    ]:
+        """
+        Run the given model on the given batch of inputs, concatenating the chosen and rejected inputs together.
+        We do this to avoid doing two forward passes, because it's faster for FSDP.
+        """
+        concatenated_batch = self.concatenated_inputs(
+            batch,
+            is_encoder_decoder=self.is_encoder_decoder,
+            label_pad_token_id=self.label_pad_token_id,
+            padding_value=self.padding_value,
+            device=self.accelerator.device,
+        )
+        # if self.accelerator.is_main_process:
+        #     import pdb; pdb.set_trace()
+        # torch.distributed.barrier()
+        model_kwargs = (
+            {
+                "decoder_input_ids": self._shift_right(
+                    concatenated_batch["concatenated_labels"]
+                ),
+            }
+            if self.is_encoder_decoder
+            else {}
+        )
+
+        if self.aux_loss_enabled:
+            model_kwargs["output_router_logits"] = True
+
+        if isinstance(model, FullyShardedDataParallel):
+            outputs = _FSDPForwardRedirection()(
+                model,
+                model._fsdp_wrapped_module.model,
+                concatenated_batch["concatenated_input_ids"],
+                attention_mask=concatenated_batch["concatenated_attention_mask"],
+                use_cache=False,
+                **model_kwargs,
+            )
+        else:
+            if isinstance(model, torch.nn.DataParallel):
+                model = model.module
+            outputs = model.model(
+                concatenated_batch["concatenated_input_ids"],
+                attention_mask=concatenated_batch["concatenated_attention_mask"],
+                use_cache=False,
+                **model_kwargs,
+            )
+
+        orpo_loss_fn = LigerFusedLinearORPOLoss(
+            ignore_index=self.label_pad_token_id, beta=self.beta
+        )
+
+        def orpo_partial(lm_head, last_hidden_state, concatenated_labels):
+            return orpo_loss_fn(
+                lm_head.weight, last_hidden_state, concatenated_labels, lm_head.bias
+            )
+
+        orpo_loss, aux_outputs = _FSDPForwardRedirection()(
+            model,
+            orpo_partial,
+            model.lm_head,
+            outputs.last_hidden_state,
+            concatenated_batch["concatenated_labels"],
+        )
+        return orpo_loss, aux_outputs
+
+    def get_batch_loss_metrics(
+        self,
+        model,
+        batch: Dict[str, Union[List, torch.LongTensor]],
+        train_eval: Literal["train", "eval"] = "train",
+    ):
+        """Compute the ORPO loss and other metrics for the given batch of inputs for train or test."""
+        metrics = {}
+        loss, aux_outputs = self.concatenated_forward(model, batch)
+        (
+            policy_chosen_logps,
+            policy_rejected_logps,
+            policy_chosen_logits,
+            policy_rejected_logits,
+            policy_nll_loss,
+        ) = aux_outputs[:5]
+
+        # return loss, metrics
+        chosen_rewards, rejected_rewards, log_odds_ratio, log_odds_chosen = aux_outputs[
+            5:
+        ]
+
+        reward_accuracies = (chosen_rewards > rejected_rewards).float()
+
+        prefix = "eval_" if train_eval == "eval" else ""
+        metrics[f"{prefix}rewards/chosen"] = chosen_rewards.mean()
+        metrics[f"{prefix}rewards/rejected"] = rejected_rewards.mean()
+        metrics[f"{prefix}rewards/accuracies"] = reward_accuracies.mean()
+        metrics[f"{prefix}rewards/margins"] = (chosen_rewards - rejected_rewards).mean()
+        metrics[f"{prefix}logps/rejected"] = policy_rejected_logps.detach().mean()
+        metrics[f"{prefix}logps/chosen"] = policy_chosen_logps.detach().mean()
+        metrics[f"{prefix}logits/rejected"] = policy_rejected_logits.detach().mean()
+        metrics[f"{prefix}logits/chosen"] = policy_chosen_logits.detach().mean()
+        metrics[f"{prefix}nll_loss"] = policy_nll_loss.detach().mean()
+        metrics[f"{prefix}log_odds_ratio"] = log_odds_ratio
+        metrics[f"{prefix}log_odds_chosen"] = log_odds_chosen
+        for k, v in metrics.items():
+            metrics[k] = v.item()
+
+        return loss, metrics

liger_kernel/transformers/qwen2vl_mrope.py

@@ -0,0 +1,20 @@
+from liger_kernel.ops.qwen2vl_mrope import LigerQwen2VLMRopeFunction
+
+
+def liger_multimodal_rotary_pos_emb(q, k, cos, sin, mrope_section, unsqueeze_dim=1):
+    """
+    Applies Multimodal Rotary Positional Embedding (M-RoPE) operation to query and key states.
+
+    Args:
+        q (torch.Tensor): The query tensor of shape (bsz, n_q_head, seq_len, head_dim).
+        k (torch.Tensor): The key tensor of shape (bsz, n_kv_head, seq_len, head_dim).
+        cos (torch.Tensor): The cosine tensor of shape (3, 1, seq_len, head_dim).
+        sin (torch.Tensor): The sine tensor of shape (3, 1, seq_len, head_dim).
+        mrope_section (List[int]): The multimodal rope section for channel dimension of temporal, height and width in rope calculation.
+        unsqueeze_dim (int, optional): The dimension to unsqueeze. Defaults to 1.
+
+    Returns:
+        Tuple[torch.Tensor, torch.Tensor]: The query and key tensors after applying the M-RoPE operation.
+    """
+
+    return LigerQwen2VLMRopeFunction.apply(q, k, cos, sin, mrope_section, unsqueeze_dim)

liger_kernel/utils.py

@@ -0,0 +1,13 @@
+import torch
+
+
+def infer_device():
+    """
+    Get current device name based on available devices
+    """
+    if torch.cuda.is_available():
+        return "cuda"
+    elif torch.xpu.is_available():
+        return "xpu"
+    else:
+        return "cpu"

{liger_kernel-0.4.2.dist-info → liger_kernel-0.5.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel
-Version: 0.4.2
+Version: 0.5.0
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -32,18 +32,26 @@ License-File: LICENSE
 License-File: NOTICE
 Requires-Dist: torch>=2.1.2
 Requires-Dist: triton>=2.3.1
+Provides-Extra: transformers
+Requires-Dist: transformers~=4.0; extra == "transformers"
 Provides-Extra: dev
 Requires-Dist: transformers>=4.44.2; extra == "dev"
+Requires-Dist: trl>=0.11.0; extra == "dev"
 Requires-Dist: matplotlib>=3.7.2; extra == "dev"
 Requires-Dist: flake8>=4.0.1.1; extra == "dev"
 Requires-Dist: black>=24.4.2; extra == "dev"
 Requires-Dist: isort>=5.13.2; extra == "dev"
 Requires-Dist: pytest>=7.1.2; extra == "dev"
+Requires-Dist: pytest-xdist; extra == "dev"
+Requires-Dist: pytest-rerunfailures; extra == "dev"
 Requires-Dist: datasets>=2.19.2; extra == "dev"
 Requires-Dist: torchvision>=0.16.2; extra == "dev"
 Requires-Dist: seaborn; extra == "dev"
-Provides-Extra: transformers
-Requires-Dist: transformers~=4.0; extra == "transformers"
+Provides-Extra: amd
+Requires-Dist: torch>=2.6.0.dev; extra == "amd"
+Requires-Dist: setuptools-scm>=8; extra == "amd"
+Requires-Dist: torchvision>=0.20.0.dev; extra == "amd"
+Requires-Dist: triton>=3.0.0; extra == "amd"
 
 <a name="readme-top"></a>
 
@@ -55,7 +63,7 @@ Requires-Dist: transformers~=4.0; extra == "transformers"
         <th style="padding: 10px;" colspan="2">Stable</th>
         <th style="padding: 10px;" colspan="2">Nightly</th>
         <th style="padding: 10px;">Discord</th>
-        <th style="padding: 10px;">Gurubase (experimental)</th>
+        <th style="padding: 10px;">Build</th>
     </tr>
     <tr>
         <td style="padding: 10px;">
@@ -84,9 +92,16 @@ Requires-Dist: transformers~=4.0; extra == "transformers"
             </a>
         </td>
         <td style="padding: 10px;">
-            <a href="https://gurubase.io/g/liger-kernel">
-                <img src="https://img.shields.io/badge/Gurubase-Ask%20Liger%20Kernel%20Guru-006BFF" alt="Ask Liger Kernel Guru">
-            </a>
+            <div style="display: block;">
+                <a href="https://github.com/linkedin/Liger-Kernel/actions/workflows/nvi-ci.yml">
+                    <img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/nvi-ci.yml/badge.svg?event=schedule" alt="Build">
+                </a>
+            </div>
+            <div style="display: block;">
+                <a href="https://github.com/linkedin/Liger-Kernel/actions/workflows/amd-ci.yml">
+                    <img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/amd-ci.yml/badge.svg?event=schedule" alt="Build">
+                </a>
+            </div>
         </td>
     </tr>
 </table>
@@ -95,13 +110,14 @@ Requires-Dist: transformers~=4.0; extra == "transformers"
 
 <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>
 
+  - [2024/12/15] 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!
-  - [2024/10/21] We have released the tech report of Liger Kernel on Arxiv: https://arxiv.org/pdf/2410.10989 
+  - [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)
   - [2024/8/23] Official release: check out our [X post](https://x.com/hsu_byron/status/1827072737673982056)
@@ -111,6 +127,8 @@ Requires-Dist: transformers~=4.0; extra == "transformers"
 
 **Liger Kernel** is a collection of Triton kernels designed specifically for LLM training. It can effectively increase multi-GPU **training throughput by 20%** and reduces **memory usage by 60%**. We have implemented **Hugging Face Compatible** `RMSNorm`, `RoPE`, `SwiGLU`, `CrossEntropy`, `FusedLinearCrossEntropy`, and more to come. The kernel works out of the box with [Flash Attention](https://github.com/Dao-AILab/flash-attention), [PyTorch FSDP](https://pytorch.org/tutorials/intermediate/FSDP_tutorial.html), and [Microsoft DeepSpeed](https://github.com/microsoft/DeepSpeed). We welcome contributions from the community to gather the best kernels for LLM training.
 
+We've also added optimized Post-Training kernels that deliver **up to 80% memory savings** for alignment and distillation tasks. We support losses like DPO, CPO, ORPO, SimPO, JSD, and many more. 
+
 ## Supercharge Your Model with Liger Kernel
 
 ![Banner](https://raw.githubusercontent.com/linkedin/Liger-Kernel/main/docs/images/banner.GIF)
@@ -128,12 +146,13 @@ With one line of code, Liger Kernel can increase throughput by more than 20% and
 
 ## Examples
 
-
 | **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      |               |
+| [**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 |
+| [**Vision-Language Model SFT**](https://github.com/linkedin/Liger-Kernel/tree/main/examples/huggingface/run_qwen2_vl.sh)      | Finetune Qwen2-VL on image-text data using 4 A100s with FSDP |
+| [**Liger ORPO Trainer**](https://github.com/linkedin/Liger-Kernel/blob/main/examples/alignment/run_orpo.py)      | Align Llama 3.2 using Liger ORPO Trainer with FSDP with 50% memory reduction |
 
 ## Key Features
 
@@ -146,7 +165,7 @@ With one line of code, Liger Kernel can increase throughput by more than 20% and
 
 ## Installation
 
-### Dependencies 
+### Dependencies
 
 #### CUDA
 
@@ -183,6 +202,8 @@ To install from source:
 git clone https://github.com/linkedin/Liger-Kernel.git
 cd Liger-Kernel
 pip install -e .
+# or if installing on amd platform
+pip install -e .[amd] --extra-index-url https://download.pytorch.org/whl/nightly/rocm6.2 # rocm6.2
 # or if using transformers
 pip install -e .[transformers]
 ```
@@ -249,7 +270,7 @@ loss = loss_fn(model.weight, input, target)
 loss.backward()
 ```
 
-## APIs
+## High-level APIs
 
 ### AutoModel
 
@@ -268,13 +289,17 @@ 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         |
-| Qwen2 & Qwen2.5      | `liger_kernel.transformers.apply_liger_kernel_to_qwen2`    | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
+| Qwen2, Qwen2.5, & QwQ      | `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         |
 
 
+## Low-level APIs
+
+- `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.
 
-### Kernels
+### Model Kernels
 
 | **Kernel**                      | **API**                                                     |
 |---------------------------------|-------------------------------------------------------------|
@@ -284,39 +309,33 @@ 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.
-- **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.
-
+| 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
 
@@ -324,6 +343,17 @@ $$\text{GeGLU}(x)=\text{GELU}(xW+b)\otimes(xV+c)$$
 - [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)
 
+## Sponsorship and Collaboration
+ 
+- [AMD](https://www.amd.com/en.html): Providing AMD GPUs for our AMD CI.
+- [Intel](https://www.intel.com/): Providing Intel GPUs for our Intel CI.
+- [Modal](https://modal.com/): Free 3000 credits from GPU MODE IRL for our NVIDIA CI.
+- [EmbeddedLLM](https://embeddedllm.com/): Making Liger Kernel run fast and stable on AMD. 
+- [HuggingFace](https://huggingface.co/): Integrating Liger Kernel into Hugging Face Transformers and TRL.
+- [Lightning AI](https://lightning.ai/): Integrating Liger Kernel into Lightning Thunder.
+- [Axolotl](https://axolotl.ai/): Integrating Liger Kernel into Axolotl.
+- [Llama-Factory](https://github.com/hiyouga/LLaMA-Factory): Integrating Liger Kernel into Llama-Factory.
+
 ## Contact
 
 - For issues, create a Github ticket in this repository
@@ -335,7 +365,7 @@ $$\text{GeGLU}(x)=\text{GELU}(xW+b)\otimes(xV+c)$$
 Biblatex entry:
 ```bib
 @article{hsu2024ligerkernelefficienttriton,
-      title={Liger Kernel: Efficient Triton Kernels for LLM Training}, 
+      title={Liger Kernel: Efficient Triton Kernels for LLM Training},
       author={Pin-Lun Hsu and Yun Dai and Vignesh Kothapalli and Qingquan Song and Shao Tang and Siyu Zhu and Steven Shimizu and Shivam Sahni and Haowen Ning and Yanning Chen},
       year={2024},
       eprint={2410.10989},
@@ -349,15 +379,8 @@ 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)
 
-## Contributors
-
-<a href="https://github.com/linkedin/Liger-Kernel/graphs/contributors">
-  <img alt="contributors" src="https://contrib.rocks/image?repo=linkedin/Liger-Kernel"/>
-</a>
-
 <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;">
         ↑ Back to Top ↑
     </a>
 </p>
-

{liger_kernel-0.4.2.dist-info → liger_kernel-0.5.0.dist-info}/RECORD

@@ -1,35 +1,44 @@
-liger_kernel/env_report.py,sha256=jye8RvUkmhqaIshdeIpoUABoAu7FPKJUib4FnAfvkpw,1132
-liger_kernel/chunked_loss/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-liger_kernel/chunked_loss/dpo_loss.py,sha256=_sftycUsxypLiQaCIoqMEwtc425Kxiq97YI6DvFvscc,1943
-liger_kernel/chunked_loss/fused_linear_preference.py,sha256=ayx-dmAx1TW9sThHJ_wUU1MqpZeJ4-SooGh0ZgVFlOA,8420
-liger_kernel/chunked_loss/orpo_loss.py,sha256=DNifPpzGV_t3dfOPlPy2XKDM6M1Qne0kCbIPztvFY9U,2179
+liger_kernel/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+liger_kernel/env_report.py,sha256=FViyPju795lB6z4k2TZldvBSmQdcS0A2hcnDxepJrDo,1822
+liger_kernel/utils.py,sha256=HJa-xVKOohDn6pLVIx-Fv0V9h0QAL3qZGQNRICI-OpI,249
+liger_kernel/chunked_loss/__init__.py,sha256=R2wCcz4Y0kTAve926DH3k182XKezpXeACMHj05g9Mm8,346
+liger_kernel/chunked_loss/cpo_loss.py,sha256=Qu1Ul2A12sp6CqIT-atPbHWFb_LLtINEA9mOpIRx_0g,3097
+liger_kernel/chunked_loss/dpo_loss.py,sha256=H9_RRhclckHYM2sd75tgbnf8IxC_PU2JCALbgtPQvwc,4222
+liger_kernel/chunked_loss/functional.py,sha256=9Gr-YXIuEzEJkBUhDx3G2fuQayckLor7cC7svhmPML4,549
+liger_kernel/chunked_loss/fused_linear_distillation.py,sha256=2BH6DCPjsR2zS6zcwFPcIIZRhLF8SohjGdKsAJ_301o,10222
+liger_kernel/chunked_loss/fused_linear_preference.py,sha256=vlWfaaIECWvCQhY9PM7zRI0vKThIrydMf6P44bXn1EE,15114
+liger_kernel/chunked_loss/orpo_loss.py,sha256=ZuKGjbkIYzV4UzvupNdq6vyxCp7-BztQkUt8ZnFvKos,3531
+liger_kernel/chunked_loss/simpo_loss.py,sha256=Wa4LOlDG9PbJkOOkKg8hbKvnKgg7OTBz6-qIkwPK1yw,3275
 liger_kernel/ops/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-liger_kernel/ops/cross_entropy.py,sha256=sfUb7-jIZp0EKXjg1DYy2Wdzw_Mg-mHmGoR5bpdm4tw,15526
-liger_kernel/ops/fused_linear_cross_entropy.py,sha256=ib7M3AjJE164yMfuS9R39k-5qnDgYOXptIT146lqYbg,9964
-liger_kernel/ops/fused_linear_jsd.py,sha256=5D_obamh08lGGTMyh85kBJD_aNjPhOYf4-TmCZ6m4s4,9626
+liger_kernel/ops/cross_entropy.py,sha256=oG5hfrlmnlF5lOoZRhHRglObxgH4B0KadjWMJj9EWPM,15860
+liger_kernel/ops/fused_linear_cross_entropy.py,sha256=Tnw4gyAYVVdnCOqhOuLEzbUQ3goOTnoAfk3pqSIM5ac,9301
+liger_kernel/ops/fused_linear_jsd.py,sha256=nOv4zwfxHqqepKEmMsQuz-B3H-gRjyo8uClpmqSGLYA,9693
 liger_kernel/ops/geglu.py,sha256=MQL4zyzneZqZYUGPvb1QjI_EYT9_pKfSDgR25WD9jrI,4127
 liger_kernel/ops/group_norm.py,sha256=VaRErVJGR4JqgXXvuIjNGTn3E2egjLtU1y3ymwIf4d8,10961
-liger_kernel/ops/jsd.py,sha256=anWfdioucxZy4JQfTvbHBR-IQrZKeH-gBF1MHwwTuTQ,5781
+liger_kernel/ops/jsd.py,sha256=Ap2b0_geCl6fqBXLI1IS6Yn6GlO-8LgPmnOW3y47dus,6151
 liger_kernel/ops/kl_div.py,sha256=03FNXfvCb6M-56hhFepAFV9p6brArPR6KOKkdGD34mw,8374
-liger_kernel/ops/layer_norm.py,sha256=unGMYMOPqtkM9aTrokhcqgPmsV2AUN7Yzv86isVB9OI,7422
-liger_kernel/ops/rms_norm.py,sha256=LAxCiFjpBbb7TDh9pOzsVmDGAR7eEbTDnEhjSd6TX_M,11583
+liger_kernel/ops/layer_norm.py,sha256=_CZggw3GNEIUx5weDzadFit5I-Lzosoo8prgeJzcViY,7589
+liger_kernel/ops/qwen2vl_mrope.py,sha256=xZvQnhkSTjU-k6KiiRn9e0SYO1ESs1jmuZFMICduLpc,8552
+liger_kernel/ops/rms_norm.py,sha256=g7OXwuYI8-LXudDwvXuiupVjjOsbu8c4wwv83VaHa54,11750
 liger_kernel/ops/rope.py,sha256=jrzaA9-6Orn44y_IIam9_YNPQxOFK2FrIRNfFea4EtU,8513
 liger_kernel/ops/swiglu.py,sha256=Fwxtd76rhHKT9ShQAGca9RsnASplAVxtYKHmiT73_yA,2994
-liger_kernel/ops/utils.py,sha256=3JSF--O7KT5Wa5BuO70M4h0XetxoZ_e9IoW9GRlxlBg,3777
+liger_kernel/ops/utils.py,sha256=_VQvd1PX5JXm5xaiBrk2gANp3qr4kM7qYG3ypkBwkMs,3850
 liger_kernel/ops/experimental/embedding.py,sha256=LYR66dB-jhvhtUjeV4PnNro-n77J1mdlmpSLSxB3Y6U,4186
 liger_kernel/ops/experimental/mm_int8int2.py,sha256=JpGVZCgRC6T8XMUJ_QbZRS2XU1bh0urIZphs5DTc1mY,13358
-liger_kernel/transformers/__init__.py,sha256=gia-eBxr7TLxU0GdDf8AfCY4WgDlFLqIGSt7EoQGsBA,1336
+liger_kernel/transformers/__init__.py,sha256=P5JR3fI-znhG92nRrFS2j0TIJTLhP-xD5dvEy4HP9ik,1418
 liger_kernel/transformers/auto_model.py,sha256=RMIwQHSiXoksXFTIqFZ4PLBgoqkxJJAT3q1Qh47bGN8,1552
 liger_kernel/transformers/cross_entropy.py,sha256=yEm_YQ7oa3_BzT3hdW6KrAslduhSqWcJQVNZZDcWCg4,1758
-liger_kernel/transformers/functional.py,sha256=Hd4WvxNqOJHM9HmRfAQueRnmOy5WU9nFsFygB5Iv8Xs,2000
+liger_kernel/transformers/functional.py,sha256=sUBoU8Vb4pLpr9G6IdkRsToYgh-rCXL4OLYat7Tv_GU,4450
 liger_kernel/transformers/fused_linear_cross_entropy.py,sha256=_i0PXSp5iZ9pKXdEeZ4lvHCENJYjV4y74yz3ZRG5XQg,1484
-liger_kernel/transformers/fused_linear_jsd.py,sha256=MJ-KjmLZnakuoVpnbDGkd95DQgvESniyrRWYzollVZM,4066
+liger_kernel/transformers/fused_linear_jsd.py,sha256=bZ4otCvWBuOnA5XdQL-FzZVItJlDt-ht9e_pG7PG93E,3999
 liger_kernel/transformers/geglu.py,sha256=QcrME_8ooIn0xa59LaC0aoOdRrBIFd11Y0bAyF0NfCw,1130
 liger_kernel/transformers/group_norm.py,sha256=FJ9R7mS9G1wO-GRIQ6QKSmIhnZ6nQ6GIkE4NnX_hnn0,2241
-liger_kernel/transformers/jsd.py,sha256=W-5CypO2mx4-bUWOxq1KScfCdoXlLoYbtt5xBnRzMs4,3056
+liger_kernel/transformers/jsd.py,sha256=sbr8DnKSYZJH9pv2rpmboNijYGpZKbhb2-WSGp5_v6g,3001
 liger_kernel/transformers/kl_div.py,sha256=qVhjBg6tjRyue5iZ3NFxo8uySY4JuIFJyv0IM_50F24,431
 liger_kernel/transformers/layer_norm.py,sha256=fd6o4kSHJWolQMWxh-l1qObfgL08ruNbUoBiANKX1ow,972
-liger_kernel/transformers/monkey_patch.py,sha256=L1IuGmFMWYgf-u3OXCg43BUxbZKTpd7ATjjDjYoFkEM,38268
+liger_kernel/transformers/monkey_patch.py,sha256=Fk2v4GZQDJzfh3Cpc6BHNJbs_tungDyWmqS9nuG9Lc4,38406
+liger_kernel/transformers/orpo_trainer.py,sha256=mC8ePS-Oq-BrdM0lKpgSBLuYLqYsWxH_4Q2RnDthz5M,7643
+liger_kernel/transformers/qwen2vl_mrope.py,sha256=SfSQVwOe7ArrVfpmIdfZrdzCxmcj7V-YQp9zDu17-ao,1043
 liger_kernel/transformers/rms_norm.py,sha256=AHstklNIO1PLHjjCBU-TPuUD-Fl_pycJUTLlJNojbV8,1189
 liger_kernel/transformers/rope.py,sha256=m-ah8vZBYW8tfplTXCiAPMHJWlB1tdp_JPXJeWE-Boo,943
 liger_kernel/transformers/swiglu.py,sha256=0-tVJ8xEYfhxnduc16PflXFj8sZPxdx9sHUn3hfwCI4,2468
@@ -47,9 +56,9 @@ liger_kernel/transformers/model/qwen2.py,sha256=EyhSSzQOskGjSnCsKMZpd1s5IAIlHd5P
 liger_kernel/transformers/model/qwen2_vl.py,sha256=bIQe2bWiY--G84FhCD29Gdi64_qHP6vbcGsK6vKysQE,8547
 liger_kernel/triton/__init__.py,sha256=yfRe0zMb47QnqjecZWG7LnanfCTzeku7SgWRAwNVmzU,101
 liger_kernel/triton/monkey_patch.py,sha256=5BcGKTtdqeYchypBIBopGIWPx1-cFALz7sOKoEsqXJ0,1584
-liger_kernel-0.4.2.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel-0.4.2.dist-info/METADATA,sha256=P9d8zHay6rXNyu58aJbTbgxtfRDXX61A15Dp-paNKpg,21530
-liger_kernel-0.4.2.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel-0.4.2.dist-info/WHEEL,sha256=R06PA3UVYHThwHvxuRWMqaGcr-PuniXahwjmQRFMEkY,91
-liger_kernel-0.4.2.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel-0.4.2.dist-info/RECORD,,
+liger_kernel-0.5.0.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel-0.5.0.dist-info/METADATA,sha256=7c5Tzf84zfQFOdXxx5nXg0wqGKH8VhsLCfTvoMN3kNM,20675
+liger_kernel-0.5.0.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel-0.5.0.dist-info/WHEEL,sha256=PZUExdf71Ui_so67QXpySuHtCi3-J3wvF4ORK6k_S8U,91
+liger_kernel-0.5.0.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel-0.5.0.dist-info/RECORD,,

{liger_kernel-0.4.2.dist-info → liger_kernel-0.5.0.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (75.5.0)
+Generator: setuptools (75.6.0)
 Root-Is-Purelib: true
 Tag: py3-none-any