liger-kernel 0.5.3__tar.gz → 0.5.4__tar.gz

liger_kernel-0.5.4/.github/workflows/intel-ci.yml

@@ -0,0 +1,71 @@
+name: Intel GPU
+
+on:
+  push:
+    branches:
+      - main
+    paths:
+      - "src/**"
+      - "test/**"
+  pull_request:
+    branches:
+      - main
+    paths:
+      - "src/**"
+      - "test/**"
+  schedule:
+    # Runs at 00:00 UTC daily
+    - cron: '0 0 * * *'
+  workflow_dispatch:  # Enables manual trigger
+
+concurrency:
+  # This causes it to cancel previous in-progress actions on the same PR / branch,
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  checkstyle:
+    runs-on: ubuntu-latest
+
+    steps:
+    - name: Checkout code
+      uses: actions/checkout@v3
+
+    - name: Set up Python
+      uses: actions/setup-python@v3
+      with:
+        python-version: '3.10'
+
+    - name: Install dependencies
+      run: |
+        python -m pip install --upgrade pip
+        pip install -r dev/fmt-requirements.txt
+
+    - name: Run checkstyle
+      run: make checkstyle
+
+  tests:
+    runs-on: linux-max1550-gpu-8
+    needs: [checkstyle]
+
+    steps:
+    - name: Checkout code
+      uses: actions/checkout@v3
+
+    - name: Set up Python
+      uses: actions/setup-python@v3
+      with:
+        python-version: '3.10'
+
+    - name: Setup Dependencies
+      run: |
+        python -m pip install --upgrade pip
+        pip install -e .[dev] --extra-index-url https://download.pytorch.org/whl/test/xpu
+    
+    - name: List Python Environments
+      run: python -m pip list
+    
+    - name: Run Unit Tests
+      run: |
+        make test
+        make test-convergence

{liger_kernel-0.5.3 → liger_kernel-0.5.4}/Makefile

@@ -9,8 +9,10 @@ test:
 
 # Command to run ruff for linting and formatting code
 checkstyle:
-	ruff check . --fix; ruff_check_status=$$?; \
-	ruff format .; ruff_format_status=$$?; \
+	ruff check .; ruff_check_status=$$?; \
+	ruff format --check .; ruff_format_status=$$?; \
+	ruff check . --fix; \
+	ruff format .; \
 	if [ $$ruff_check_status -ne 0 ] || [ $$ruff_format_status -ne 0 ]; then \
 		exit 1; \
 	fi
@@ -18,9 +20,13 @@ checkstyle:
 # Command to run pytest for convergence tests
 # We have to explicitly set HF_DATASETS_OFFLINE=1, or dataset will silently try to send metrics and timeout (80s) https://github.com/huggingface/datasets/blob/37a603679f451826cfafd8aae00738b01dcb9d58/src/datasets/load.py#L286
 test-convergence:
-	HF_DATASETS_OFFLINE=1 python -m pytest --disable-warnings test/convergence/test_mini_models.py
-	HF_DATASETS_OFFLINE=1 python -m pytest --disable-warnings test/convergence/test_mini_models_multimodal.py
-	HF_DATASETS_OFFLINE=1 python -m pytest --disable-warnings test/convergence/test_mini_models_with_logits.py
+	HF_DATASETS_OFFLINE=1 python -m pytest --disable-warnings test/convergence/fp32/test_mini_models.py
+	HF_DATASETS_OFFLINE=1 python -m pytest --disable-warnings test/convergence/fp32/test_mini_models_multimodal.py
+	HF_DATASETS_OFFLINE=1 python -m pytest --disable-warnings test/convergence/fp32/test_mini_models_with_logits.py
+
+	HF_DATASETS_OFFLINE=1 python -m pytest --disable-warnings test/convergence/bf16/test_mini_models.py
+	HF_DATASETS_OFFLINE=1 python -m pytest --disable-warnings test/convergence/bf16/test_mini_models_multimodal.py
+	HF_DATASETS_OFFLINE=1 python -m pytest --disable-warnings test/convergence/bf16/test_mini_models_with_logits.py
 
 # Command to run all benchmark scripts and update benchmarking data file
 # By default this doesn't overwrite existing data for the same benchmark experiment

{liger_kernel-0.5.3 → liger_kernel-0.5.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: liger_kernel
-Version: 0.5.3
+Version: 0.5.4
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -97,6 +97,11 @@ Dynamic: requires-dist
                     <img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/amd-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/intel-ci.yml/badge.svg?event=schedule" alt="Build">
+                </a>
+            </div>
         </td>
     </tr>
 </table>
@@ -123,7 +128,7 @@ Dynamic: requires-dist
 
 **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. Check out [how we optimize the memory](https://x.com/hsu_byron/status/1866577403918917655).
+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, KTO, JSD, and many more. Check out [how we optimize the memory](https://x.com/hsu_byron/status/1866577403918917655).
 
 ## Supercharge Your Model with Liger Kernel
 
@@ -188,6 +193,11 @@ y = orpo_loss(lm_head.weight, x, target)
 - `torch >= 2.5.0` Install according to the instruction in Pytorch official webpage.
 - `triton >= 3.0.0` Install from pypi. (e.g. `pip install triton==3.0.0`)
 
+```bash
+# Need to pass the url when installing
+pip install -e .[dev] --extra-index-url https://download.pytorch.org/whl/nightly/rocm6.2
+```
+
 ### Optional Dependencies
 
 - `transformers >= 4.x`: Required if you plan to use the transformers models patching APIs. The specific model you are working will dictate the minimum version of transformers.
@@ -305,6 +315,8 @@ loss.backward()
 | Qwen2, Qwen2.5, & QwQ      | `liger_kernel.transformers.apply_liger_kernel_to_qwen2`    | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | Qwen2-VL, & QVQ       | `liger_kernel.transformers.apply_liger_kernel_to_qwen2_vl`    | RMSNorm, LayerNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | Phi3 & Phi3.5       | `liger_kernel.transformers.apply_liger_kernel_to_phi3`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
+| Granite 3.0 & 3.1   | `liger_kernel.transformers.apply_liger_kernel_to_granite`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss |
+| OLMo2   | `liger_kernel.transformers.apply_liger_kernel_to_olmo2`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
 
 
 ## Low-level APIs
@@ -333,6 +345,7 @@ loss.backward()
 | 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`     |
+| Fused Linear KTO Loss           | `liger_kernel.chunked_loss.LigerFusedLinearKTOLoss`     |
 
 ### Distillation Kernels
 
@@ -341,6 +354,7 @@ loss.backward()
 | KLDivergence                    | `liger_kernel.transformers.LigerKLDIVLoss`                  |
 | JSD                             | `liger_kernel.transformers.LigerJSD`                        |
 | Fused Linear JSD                  | `liger_kernel.transformers.LigerFusedLinearJSD`             |
+| TVD                             | `liger_kernel.transformers.LigerTVDLoss`                    |
 
 ### Experimental Kernels
 
@@ -372,7 +386,7 @@ loss.backward()
 
 - 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
+- For formal collaboration, send an email to yannchen@linkedin.com
 
 ## Cite this work
 

{liger_kernel-0.5.3 → liger_kernel-0.5.4}/README.md

@@ -47,6 +47,11 @@
                     <img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/amd-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/intel-ci.yml/badge.svg?event=schedule" alt="Build">
+                </a>
+            </div>
         </td>
     </tr>
 </table>
@@ -73,7 +78,7 @@
 
 **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. Check out [how we optimize the memory](https://x.com/hsu_byron/status/1866577403918917655).
+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, KTO, JSD, and many more. Check out [how we optimize the memory](https://x.com/hsu_byron/status/1866577403918917655).
 
 ## Supercharge Your Model with Liger Kernel
 
@@ -138,6 +143,11 @@ y = orpo_loss(lm_head.weight, x, target)
 - `torch >= 2.5.0` Install according to the instruction in Pytorch official webpage.
 - `triton >= 3.0.0` Install from pypi. (e.g. `pip install triton==3.0.0`)
 
+```bash
+# Need to pass the url when installing
+pip install -e .[dev] --extra-index-url https://download.pytorch.org/whl/nightly/rocm6.2
+```
+
 ### Optional Dependencies
 
 - `transformers >= 4.x`: Required if you plan to use the transformers models patching APIs. The specific model you are working will dictate the minimum version of transformers.
@@ -255,6 +265,8 @@ loss.backward()
 | Qwen2, Qwen2.5, & QwQ      | `liger_kernel.transformers.apply_liger_kernel_to_qwen2`    | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | Qwen2-VL, & QVQ       | `liger_kernel.transformers.apply_liger_kernel_to_qwen2_vl`    | RMSNorm, LayerNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | Phi3 & Phi3.5       | `liger_kernel.transformers.apply_liger_kernel_to_phi3`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
+| Granite 3.0 & 3.1   | `liger_kernel.transformers.apply_liger_kernel_to_granite`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss |
+| OLMo2   | `liger_kernel.transformers.apply_liger_kernel_to_olmo2`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
 
 
 ## Low-level APIs
@@ -283,6 +295,7 @@ loss.backward()
 | 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`     |
+| Fused Linear KTO Loss           | `liger_kernel.chunked_loss.LigerFusedLinearKTOLoss`     |
 
 ### Distillation Kernels
 
@@ -291,6 +304,7 @@ loss.backward()
 | KLDivergence                    | `liger_kernel.transformers.LigerKLDIVLoss`                  |
 | JSD                             | `liger_kernel.transformers.LigerJSD`                        |
 | Fused Linear JSD                  | `liger_kernel.transformers.LigerFusedLinearJSD`             |
+| TVD                             | `liger_kernel.transformers.LigerTVDLoss`                    |
 
 ### Experimental Kernels
 
@@ -322,7 +336,7 @@ loss.backward()
 
 - 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
+- For formal collaboration, send an email to yannchen@linkedin.com
 
 ## Cite this work
 

{liger_kernel-0.5.3 → liger_kernel-0.5.4}/benchmark/data/all_benchmark_data.csv

@@ -505,6 +505,42 @@ fused_linear_jsd,torch,full,memory,MB,BT,B x T,1024,10609.005859375,10609.005859
 fused_linear_jsd,torch,full,memory,MB,BT,B x T,2048,17146.009765625,17146.009765625,17146.009765625,"{""H"": 4096, ""V"": 128256, ""mode"": ""forward"", ""dtype"": ""torch.bfloat16""}",NVIDIA H100 80GB HBM3,2024-10-09 12:29:35,0.3.1
 fused_linear_jsd,torch,full,memory,MB,BT,B x T,4096,30220.017578125,30220.017578125,30220.017578125,"{""H"": 4096, ""V"": 128256, ""mode"": ""forward"", ""dtype"": ""torch.bfloat16""}",NVIDIA H100 80GB HBM3,2024-10-09 12:29:35,0.3.1
 fused_linear_jsd,torch,full,memory,MB,BT,B x T,8192,56368.015625,56368.015625,56368.015625,"{""H"": 4096, ""V"": 128256, ""mode"": ""forward"", ""dtype"": ""torch.bfloat16""}",NVIDIA H100 80GB HBM3,2024-10-09 12:29:35,0.3.1
+tvd,liger,full,memory,MB,V,vocab size,4096,1792.0009765625,1792.0009765625,1792.0009765625,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:04,0.3.1
+tvd,liger,full,memory,MB,V,vocab size,8192,3584.0009765625,3584.0009765625,3584.0009765625,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:04,0.3.1
+tvd,liger,full,memory,MB,V,vocab size,16384,7168.0009765625,7168.0009765625,7168.0009765625,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:04,0.3.1
+tvd,liger,full,memory,MB,V,vocab size,32768,14336.0009765625,14336.0009765625,14336.0009765625,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:04,0.3.1
+tvd,liger,full,memory,MB,V,vocab size,65536,28672.0,28672.0,28672.0,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:04,0.3.1
+tvd,liger,full,memory,MB,V,vocab size,131072,57344.0,57344.0,57344.0,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:04,0.3.1
+tvd,torch,full,memory,MB,V,vocab size,4096,2048.0009765625,2048.0009765625,2048.0009765625,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:04,0.3.1
+tvd,torch,full,memory,MB,V,vocab size,8192,4096.0009765625,4096.0009765625,4096.0009765625,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:04,0.3.1
+tvd,torch,full,memory,MB,V,vocab size,16384,8192.0009765625,8192.0009765625,8192.0009765625,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:04,0.3.1
+tvd,torch,full,memory,MB,V,vocab size,32768,16384.0,16384.0,16384.0,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:04,0.3.1
+tvd,torch,full,memory,MB,V,vocab size,65536,32768.0,32768.0,32768.0,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:04,0.3.1
+tvd,torch,full,memory,MB,V,vocab size,131072,65536.0,65536.0,65536.0,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:04,0.3.1
+tvd,liger,forward,speed,ms,V,vocab size,4096,0.47814399003982544,0.4774720072746277,0.4790079891681671,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:08,0.3.1
+tvd,liger,forward,speed,ms,V,vocab size,8192,0.906495988368988,0.905951976776123,0.9073920249938965,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:08,0.3.1
+tvd,liger,forward,speed,ms,V,vocab size,16384,1.8787360191345215,1.8778239488601685,1.8797119855880737,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:08,0.3.1
+tvd,liger,forward,speed,ms,V,vocab size,32768,3.5788800716400146,3.5772159099578857,3.58076810836792,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:08,0.3.1
+tvd,liger,forward,speed,ms,V,vocab size,65536,7.008831977844238,7.007718086242676,7.010636806488037,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:08,0.3.1
+tvd,liger,forward,speed,ms,V,vocab size,131072,13.88646411895752,13.88128662109375,13.890560150146484,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:08,0.3.1
+tvd,torch,forward,speed,ms,V,vocab size,4096,1.308608055114746,1.306502342224121,1.3104127645492554,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:09,0.3.1
+tvd,torch,forward,speed,ms,V,vocab size,8192,2.4735519886016846,2.472287893295288,2.4749441146850586,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:09,0.3.1
+tvd,torch,forward,speed,ms,V,vocab size,16384,4.828320026397705,4.826848030090332,4.830643177032471,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:09,0.3.1
+tvd,torch,forward,speed,ms,V,vocab size,32768,9.5206880569458,9.517024040222168,9.525145530700684,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:09,0.3.1
+tvd,torch,forward,speed,ms,V,vocab size,65536,19.01535987854004,19.011123657226562,19.01806640625,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:09,0.3.1
+tvd,torch,forward,speed,ms,V,vocab size,131072,38.022865295410156,38.01945877075195,38.02627182006836,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:09,0.3.1
+tvd,liger,full,speed,ms,V,vocab size,4096,2.626512050628662,2.621260643005371,2.646751880645752,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:11,0.3.1
+tvd,liger,full,speed,ms,V,vocab size,8192,4.661711692810059,4.657618999481201,4.662930965423584,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:11,0.3.1
+tvd,liger,full,speed,ms,V,vocab size,16384,9.088272094726562,9.080741882324219,9.092268943786621,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:11,0.3.1
+tvd,liger,full,speed,ms,V,vocab size,32768,18.116064071655273,18.112728118896484,18.118234634399414,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:11,0.3.1
+tvd,liger,full,speed,ms,V,vocab size,65536,35.85124969482422,35.849971771240234,35.85252380371094,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:11,0.3.1
+tvd,liger,full,speed,ms,V,vocab size,131072,71.1648941040039,71.1648941040039,71.1648941040039,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:11,0.3.1
+tvd,torch,full,speed,ms,V,vocab size,4096,4.361599922180176,4.360159873962402,4.3639678955078125,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:13,0.3.1
+tvd,torch,full,speed,ms,V,vocab size,8192,8.11302375793457,8.11075210571289,8.114463806152344,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:13,0.3.1
+tvd,torch,full,speed,ms,V,vocab size,16384,15.841055870056152,15.837087631225586,15.841856002807617,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:13,0.3.1
+tvd,torch,full,speed,ms,V,vocab size,32768,31.71219253540039,31.706951141357422,31.715898513793945,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:13,0.3.1
+tvd,torch,full,speed,ms,V,vocab size,65536,63.17919921875,63.17919921875,63.17919921875,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:13,0.3.1
+tvd,torch,full,speed,ms,V,vocab size,131072,126.0436782836914,126.0436782836914,126.0436782836914,"{""B"": 8, ""T"": 2048}",NVIDIA H100 PCIe,2024-10-26 01:17:13,0.3.1
 group_norm,liger,forward,speed,ms,C,num_channels,32,0.03481600061058998,0.03379200026392937,0.03993599861860275,"{""M"": 128, ""H"": 512, ""channels_per_group"": 4, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:20:35,0.3.1
 group_norm,liger,forward,speed,ms,C,num_channels,64,0.05222399905323982,0.05119999870657921,0.05222399905323982,"{""M"": 128, ""H"": 512, ""channels_per_group"": 4, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:20:35,0.3.1
 group_norm,liger,forward,speed,ms,C,num_channels,128,0.08499199897050858,0.08396799862384796,0.08499199897050858,"{""M"": 128, ""H"": 512, ""channels_per_group"": 4, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA A100-SXM4-40GB,2024-11-05 19:20:35,0.3.1
@@ -769,3 +805,4 @@ distill_jsd_loss,torch,full,memory,MB,BT,B x T,1024,16174.0390625,16174.0390625,
 distill_jsd_loss,torch,full,memory,MB,BT,B x T,2048,23713.05078125,23713.05078125,23713.05078125,"{""H"": 4096, ""V"": 128256, ""mode"": ""forward"", ""dtype"": ""torch.bfloat16"", ""bias"": false, ""weight_hard_loss"": 0.5, ""weight_soft_loss"": 0.5, ""ignore_index"": -100}",NVIDIA H100 80GB HBM3,2024-12-03 08:01:32,0.4.2
 distill_jsd_loss,torch,full,memory,MB,BT,B x T,4096,38791.07421875,38791.07421875,38791.07421875,"{""H"": 4096, ""V"": 128256, ""mode"": ""forward"", ""dtype"": ""torch.bfloat16"", ""bias"": false, ""weight_hard_loss"": 0.5, ""weight_soft_loss"": 0.5, ""ignore_index"": -100}",NVIDIA H100 80GB HBM3,2024-12-03 08:01:32,0.4.2
 distill_jsd_loss,torch,full,memory,MB,BT,B x T,8192,68947.1015625,68947.1015625,68947.1015625,"{""H"": 4096, ""V"": 128256, ""mode"": ""forward"", ""dtype"": ""torch.bfloat16"", ""bias"": false, ""weight_hard_loss"": 0.5, ""weight_soft_loss"": 0.5, ""ignore_index"": -100}",NVIDIA H100 80GB HBM3,2024-12-03 08:01:32,0.4.2
+

{liger_kernel-0.5.3 → liger_kernel-0.5.4}/benchmark/scripts/benchmark_kto_loss.py

@@ -103,8 +103,8 @@ def bench_memory_kto_loss(input: SingleBenchmarkRunInput) -> SingleBenchmarkRunO
         H=H,
         V=V,
         dtype=dtype,
-        bias=bias,
-        ref_bias=bias,
+        use_bias=bias,
+        use_ref_bias=bias,
         ignore_index=ignore_index,
         beta=beta,
     ).to(device)
@@ -113,8 +113,8 @@ def bench_memory_kto_loss(input: SingleBenchmarkRunInput) -> SingleBenchmarkRunO
         H=H,
         V=V,
         dtype=dtype,
-        bias=bias,
-        ref_bias=bias,
+        use_bias=bias,
+        use_ref_bias=bias,
         ignore_index=ignore_index,
         beta=beta,
     ).to(device)
@@ -189,7 +189,7 @@ def bench_speed_kto_loss(input: SingleBenchmarkRunInput) -> SingleBenchmarkRunOu
         dtype=dtype,
         beta=beta,
         ignore_index=ignore_index,
-        bias=bias,
+        use_bias=bias,
     ).to(device)
     liger_kto_loss = LigerLMHeadKTO(
         H=H,
@@ -197,7 +197,7 @@ def bench_speed_kto_loss(input: SingleBenchmarkRunInput) -> SingleBenchmarkRunOu
         dtype=dtype,
         beta=beta,
         ignore_index=ignore_index,
-        bias=bias,
+        use_bias=bias,
     ).to(device)
 
     # Input shape: [B, T, H]

{liger_kernel-0.5.3 → liger_kernel-0.5.4}/benchmark/scripts/benchmark_rope.py

@@ -1,7 +1,6 @@
 import torch
 import triton
 
-from test.utils import transformers_version_dispatch
 from transformers.models.llama.configuration_llama import LlamaConfig
 from transformers.models.llama.modeling_llama import LlamaRotaryEmbedding
 from transformers.models.llama.modeling_llama import apply_rotary_pos_emb
@@ -14,6 +13,7 @@ from utils import run_benchmarks
 
 from liger_kernel.transformers.rope import liger_rotary_pos_emb
 from liger_kernel.utils import infer_device
+from liger_kernel.utils import transformers_version_dispatch
 
 device = infer_device()
 

liger_kernel-0.5.4/benchmark/scripts/benchmark_tvd.py

@@ -0,0 +1,133 @@
+import torch
+import triton
+
+from utils import QUANTILES
+from utils import SingleBenchmarkRunInput
+from utils import SingleBenchmarkRunOutput
+from utils import _test_memory
+from utils import parse_benchmark_script_args
+from utils import run_benchmarks
+
+from liger_kernel.transformers.tvd import LigerTVDLoss
+
+
+class TorchTVDLoss(torch.nn.Module):
+    def __init__(self, reduction="batchmean"):
+        super(TorchTVDLoss, self).__init__()
+        self.reduction = reduction
+
+    def forward(self, p, q):
+        tvd = torch.abs(p - q) / 2.0
+        if self.reduction == "mean":
+            return torch.sum(tvd) / (p.size(0) * p.size(1))
+        elif self.reduction == "sum":
+            return torch.sum(tvd)
+        elif self.reduction == "none":
+            return tvd
+        elif self.reduction == "batchmean":
+            return torch.sum(tvd) / p.size(0)
+        else:
+            raise ValueError("Invalid reduction type.")
+
+
+S, E = 12, 18
+
+
+def bench_speed_tvd(input: SingleBenchmarkRunInput) -> SingleBenchmarkRunOutput:
+    reduction = "batchmean"
+    V = input.x
+    B, T = input.extra_benchmark_config["B"], input.extra_benchmark_config["T"]
+    torch_tvd = TorchTVDLoss(reduction=reduction)
+    liger_tvd = LigerTVDLoss(reduction=reduction)
+
+    _input = torch.randn(B * T, V, requires_grad=True, device="cuda").softmax(dim=-1)
+    target = torch.randn(B * T, V, device="cuda").softmax(dim=-1)
+
+    def fwd():
+        if input.kernel_provider == "liger":
+            return liger_tvd(_input, target)
+        else:
+            return torch_tvd(_input, target)
+
+    if input.kernel_operation_mode == "forward":
+        ms_50, ms_20, ms_80 = triton.testing.do_bench(fwd, quantiles=QUANTILES, rep=100)
+    elif input.kernel_operation_mode == "backward":
+        y = fwd()
+
+        ms_50, ms_20, ms_80 = triton.testing.do_bench(
+            lambda: y.backward(retain_graph=True),
+            quantiles=QUANTILES,
+            grad_to_none=[_input],
+            rep=100,
+        )
+    elif input.kernel_operation_mode == "full":
+
+        def full():
+            y = fwd()
+            y.backward(retain_graph=True)
+
+        ms_50, ms_20, ms_80 = triton.testing.do_bench(full, quantiles=QUANTILES, rep=100)
+    return SingleBenchmarkRunOutput(
+        y_20=ms_20,
+        y_50=ms_50,
+        y_80=ms_80,
+    )
+
+
+def bench_memory_tvd(input: SingleBenchmarkRunInput) -> SingleBenchmarkRunOutput:
+    reduction = "batchmean"
+    torch_tvd = TorchTVDLoss(reduction=reduction)
+    liger_tvd = LigerTVDLoss(reduction=reduction)
+
+    V = input.x
+    B, T = input.extra_benchmark_config["B"], input.extra_benchmark_config["T"]
+
+    _input = torch.randn(B * T, V, requires_grad=True, device="cuda").softmax(dim=-1)
+    target = torch.randn(B * T, V, device="cuda").softmax(dim=-1)
+
+    def fwd():
+        if input.kernel_provider == "liger":
+            return liger_tvd(_input, target)
+        else:
+            return torch_tvd(_input, target)
+
+    def full():
+        y = fwd()
+        y.backward(retain_graph=True)
+
+    mem_50, mem_20, mem_80 = _test_memory(full, quantiles=QUANTILES)
+
+    return SingleBenchmarkRunOutput(
+        y_20=mem_20,
+        y_50=mem_50,
+        y_80=mem_80,
+    )
+
+
+if __name__ == "__main__":
+    args = parse_benchmark_script_args()
+    common_args = {
+        "kernel_name": "tvd",
+        "x_name": "V",
+        "x_label": "vocab size",
+        "x_values": [2**i for i in range(12, 18)],
+        "kernel_providers": ["liger", "torch"],
+        "extra_benchmark_configs": [{"B": 8, "T": 2048}],
+        "overwrite": args.overwrite,
+    }
+
+    run_benchmarks(
+        bench_test_fn=bench_memory_tvd,
+        kernel_operation_modes=["full"],
+        metric_name="memory",
+        metric_unit="MB",
+        **common_args,
+    )
+
+    run_benchmarks(
+        bench_test_fn=bench_speed_tvd,
+        kernel_operation_modes=["forward", "full"],
+        metric_name="speed",
+        metric_unit="ms",
+        **common_args,
+    )

{liger_kernel-0.5.3 → liger_kernel-0.5.4}/dev/modal/tests.py

@@ -14,7 +14,7 @@ app = modal.App("liger_tests", image=image)
 repo = modal.Mount.from_local_dir(ROOT_PATH, remote_path=REMOTE_ROOT_PATH)
 
 
-@app.function(gpu="A10G", mounts=[repo], timeout=60 * 15)
+@app.function(gpu="A10G", mounts=[repo], timeout=60 * 20)
 def liger_tests():
     import subprocess
 

{liger_kernel-0.5.3 → liger_kernel-0.5.4}/examples/lightning/training.py

@@ -158,7 +158,7 @@ class DataModule(pl.LightningDataModule):
         for i in range(len(example["question"])):
             choices = ""
             for j in range(len(example["choices"][i])):
-                choices += f"{j+1}. {example['choices'][i][j]}; "
+                choices += f"{j + 1}. {example['choices'][i][j]}; "
             s = "Below is a question and multiple choice answers, choices separated by a semicolon. Please select the best answer for the question. "
             s += f"{QUESTION}{example['question'][i]} "
             s += f"{CHOICES}{choices} "

{liger_kernel-0.5.3 → liger_kernel-0.5.4}/examples/medusa/callback.py

@@ -352,9 +352,9 @@ class EfficiencyCallback(transformers.TrainerCallback):
             else:
                 return world_size
 
-        assert (
-            world_size != 0
-        ), "WORLD_SIZE should be set to a positive integer. For single GPU training, please explicitly set WORLD_SIZE=1."
+        assert world_size != 0, (
+            "WORLD_SIZE should be set to a positive integer. For single GPU training, please explicitly set WORLD_SIZE=1."
+        )
 
         # TODO: add deepspeed support
         return world_size

{liger_kernel-0.5.3 → liger_kernel-0.5.4}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel"
-version = "0.5.3"
+version = "0.5.4"
 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-0.5.3 → liger_kernel-0.5.4}/setup.py

@@ -21,6 +21,11 @@ def get_default_dependencies():
             "torch>=2.6.0.dev",
             "triton>=3.0.0",
         ]
+    elif platform == "xpu":
+        return [
+            "torch>=2.6.0",
+            "pytorch-triton-xpu>=3.2.0",
+        ]
 
 
 def get_optional_dependencies():
@@ -43,8 +48,7 @@ def get_optional_dependencies():
     }
 
 
-# TODO: add intel XPU
-def get_platform() -> Literal["cuda", "rocm", "cpu"]:
+def get_platform() -> Literal["cuda", "rocm", "cpu", "xpu"]:
     """
     Detect whether the system has NVIDIA or AMD GPU without torch dependency.
     """
@@ -60,8 +64,13 @@ def get_platform() -> Literal["cuda", "rocm", "cpu"]:
             print("ROCm GPU detected")
             return "rocm"
         except (subprocess.SubprocessError, FileNotFoundError):
-            print("No GPU detected")
-            return "cpu"
+            try:
+                subprocess.run(["xpu-smi"], check=True)
+                print("Intel GPU detected")
+                return "xpu"
+            except (subprocess.SubprocessError, FileNotFoundError):
+                print("No GPU detected")
+                return "cpu"
 
 
 setup(

{liger_kernel-0.5.3 → liger_kernel-0.5.4}/src/liger_kernel/chunked_loss/__init__.py

@@ -1,5 +1,6 @@
 from liger_kernel.chunked_loss.cpo_loss import LigerFusedLinearCPOLoss  # noqa: F401
 from liger_kernel.chunked_loss.dpo_loss import LigerFusedLinearDPOLoss  # noqa: F401
+from liger_kernel.chunked_loss.grpo_loss import LigerFusedLinearGRPOLoss  # noqa: F401
 from liger_kernel.chunked_loss.jsd_loss import LigerFusedLinearJSDLoss  # noqa: F401
 from liger_kernel.chunked_loss.kto_loss import LigerFusedLinearKTOLoss  # noqa: F401
 from liger_kernel.chunked_loss.orpo_loss import LigerFusedLinearORPOLoss  # noqa: F401