liger-kernel-nightly 0.4.2.dev20241119061743__tar.gz → 0.4.2.dev20241119223206__tar.gz

{liger_kernel_nightly-0.4.2.dev20241119061743/src/liger_kernel_nightly.egg-info → liger_kernel_nightly-0.4.2.dev20241119223206}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.4.2.dev20241119061743
+Version: 0.4.2.dev20241119223206
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -101,7 +101,7 @@ Requires-Dist: seaborn; extra == "dev"
   <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 
+  - [2024/10/21] 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)
@@ -128,12 +128,12 @@ 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 |
 
 ## Key Features
 
@@ -146,7 +146,7 @@ With one line of code, Liger Kernel can increase throughput by more than 20% and
 
 ## Installation
 
-### Dependencies 
+### Dependencies
 
 #### CUDA
 
@@ -335,7 +335,7 @@ $$\text{GeGLU}(x)=\text{GELU}(xW+b)\otimes(xV+c)$$
 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},
@@ -360,4 +360,3 @@ Biblatex entry:
         ↑ Back to Top ↑
     </a>
 </p>
-

{liger_kernel_nightly-0.4.2.dev20241119061743 → liger_kernel_nightly-0.4.2.dev20241119223206}/README.md

@@ -54,7 +54,7 @@
   <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 
+  - [2024/10/21] 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)
@@ -81,12 +81,12 @@ 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 |
 
 ## Key Features
 
@@ -99,7 +99,7 @@ With one line of code, Liger Kernel can increase throughput by more than 20% and
 
 ## Installation
 
-### Dependencies 
+### Dependencies
 
 #### CUDA
 
@@ -288,7 +288,7 @@ $$\text{GeGLU}(x)=\text{GELU}(xW+b)\otimes(xV+c)$$
 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},
@@ -313,4 +313,3 @@ Biblatex entry:
         ↑ Back to Top ↑
     </a>
 </p>
-

{liger_kernel_nightly-0.4.2.dev20241119061743 → liger_kernel_nightly-0.4.2.dev20241119223206}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.4.2.dev20241119061743"
+version = "0.4.2.dev20241119223206"
 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.dev20241119061743 → liger_kernel_nightly-0.4.2.dev20241119223206}/src/liger_kernel/chunked_loss/fused_linear_preference.py

@@ -32,6 +32,7 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
         alpha=1.0,
         beta=0.1,
         compiled=True,
+        **loss_kwargs,
     ):
         """
         Base class for fused linear layer with preference loss.
@@ -49,6 +50,7 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
             alpha (float): Weight for the NLL loss.
             beta (float): Weight for the odds ratio loss.
             compiled (bool): Whether to use torch compile for chunk accumulation.
+            loss_kwargs (dict): Other possible arguments that a loss function might need
         """
         # TODO: Tune CHUNK_SIZE to fully utilize the GPU
         CHUNK_SIZE = chunk_size
@@ -68,6 +70,7 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
             beta=beta,
             compute_nll_loss=compute_nll_loss,
             full_target=target,
+            **loss_kwargs,
         )
 
         def accumulate_chunk(input_chunk, target_chunk):
@@ -94,6 +97,9 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
             loss_acc.add_(chunk_loss)
             return chunk_grad_input
 
+        if compiled:
+            accumulate_chunk = torch.compile(accumulate_chunk)
+
         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)
@@ -116,8 +122,6 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
                 [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]])

{liger_kernel_nightly-0.4.2.dev20241119061743 → liger_kernel_nightly-0.4.2.dev20241119223206}/src/liger_kernel/chunked_loss/orpo_loss.py

@@ -34,7 +34,7 @@ class LigerFusedLinearORPOFunction(LigerFusedLinearPreferenceBase):
         ignore_index=-100,
         beta=0.1,
         compute_nll_loss=True,
-        compiled=True,
+        compiled=False,
     ):
         """
         Fused linear layer with ORPO (Odds-Ratio Preference Optimization) loss.

liger_kernel_nightly-0.4.2.dev20241119223206/src/liger_kernel/chunked_loss/simpo_loss.py

@@ -0,0 +1,64 @@
+import torch.nn.functional as F
+
+from liger_kernel.chunked_loss.fused_linear_preference import (
+    LigerFusedLinearPreferenceBase,
+)
+
+
+class LigerFusedLinearSimPOFunction(LigerFusedLinearPreferenceBase):
+
+    @staticmethod
+    def preference_loss_fn(chosen_logps, rejected_logps, beta=0.1, gamma=0.5):
+        """
+        Compute odds-ratio 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.
+            gamma (float): The simpo gamma, margin term.
+        """
+        logits = beta * (chosen_logps - rejected_logps) - gamma
+        loss = F.logsigmoid(logits).mean()
+        return loss
+
+    @staticmethod
+    def forward(
+        ctx,
+        _input,
+        weight,
+        target,
+        bias=None,
+        ignore_index=-100,
+        beta=0.1,
+        alpha=1.0,
+        compute_nll_loss=False,
+        compiled=True,
+        gamma=0.5,
+    ):
+        """
+        Fused linear layer with SimPO (Simple Preference Optimization) loss. https://arxiv.org/pdf/2405.14734
+        Handles both the forward and backward pass of the final linear layer with SimPO loss.
+        Inspired from LigerFusedLinearCrossEntropyFunction (https://arxiv.org/abs/2410.10989) which fuses final linear layer and CE loss.
+        """
+
+        return LigerFusedLinearPreferenceBase.forward(
+            ctx,
+            _input,
+            weight,
+            target,
+            bias,
+            loss_fn=LigerFusedLinearSimPOFunction.preference_loss_fn,
+            compute_nll_loss=compute_nll_loss,
+            ignore_index=ignore_index,
+            alpha=alpha,
+            beta=beta,
+            compiled=compiled,
+            gamma=gamma,
+        )
+
+    @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, None, None

{liger_kernel_nightly-0.4.2.dev20241119061743 → liger_kernel_nightly-0.4.2.dev20241119223206/src/liger_kernel_nightly.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.4.2.dev20241119061743
+Version: 0.4.2.dev20241119223206
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -101,7 +101,7 @@ Requires-Dist: seaborn; extra == "dev"
   <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 
+  - [2024/10/21] 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)
@@ -128,12 +128,12 @@ 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 |
 
 ## Key Features
 
@@ -146,7 +146,7 @@ With one line of code, Liger Kernel can increase throughput by more than 20% and
 
 ## Installation
 
-### Dependencies 
+### Dependencies
 
 #### CUDA
 
@@ -335,7 +335,7 @@ $$\text{GeGLU}(x)=\text{GELU}(xW+b)\otimes(xV+c)$$
 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},
@@ -360,4 +360,3 @@ Biblatex entry:
         ↑ Back to Top ↑
     </a>
 </p>
-

{liger_kernel_nightly-0.4.2.dev20241119061743 → liger_kernel_nightly-0.4.2.dev20241119223206}/src/liger_kernel_nightly.egg-info/SOURCES.txt

@@ -8,6 +8,7 @@ src/liger_kernel/chunked_loss/cpo_loss.py
 src/liger_kernel/chunked_loss/dpo_loss.py
 src/liger_kernel/chunked_loss/fused_linear_preference.py
 src/liger_kernel/chunked_loss/orpo_loss.py
+src/liger_kernel/chunked_loss/simpo_loss.py
 src/liger_kernel/ops/__init__.py
 src/liger_kernel/ops/cross_entropy.py
 src/liger_kernel/ops/fused_linear_cross_entropy.py