liger-kernel 0.6.2__tar.gz → 0.6.3__tar.gz

{liger_kernel-0.6.2 → liger_kernel-0.6.3}/.github/workflows/amd-ci.yml

@@ -13,9 +13,6 @@ on:
     paths:
       - "src/**"
       - "test/**"
-  schedule:
-    # Runs at 00:00 UTC daily
-    - cron: '0 0 * * *'
   workflow_dispatch:  # Enables manual trigger
 
 concurrency:
@@ -64,7 +61,8 @@ jobs:
       run: |
         rocm-smi
         python -m pip install --upgrade pip
-        pip install -e .[dev] --extra-index-url https://download.pytorch.org/whl/nightly/rocm${{ matrix.rocm_version }}
+        pip install -e .[dev]
+        pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm${{ matrix.rocm_version }}/
     
     - name: List Python Environments
       run: python -m pip list

liger_kernel-0.6.3/.github/workflows/docs.yml

@@ -0,0 +1,64 @@
+name: Publish documentation
+on:
+  push:
+    branches:
+      - main
+    paths:
+      - 'docs/**'
+      - 'mkdocs.yml'
+
+permissions:
+  contents: write
+jobs:
+  deploy:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - name: Configure Git Credentials
+        run: |
+          git config user.name github-actions[bot]
+          git config user.email 41898282+github-actions[bot]@users.noreply.github.com
+      - uses: actions/setup-python@v5
+        with:
+          python-version: 3.x
+      - run: echo "cache_id=$(date --utc '+%V')" >> $GITHUB_ENV 
+      - uses: actions/cache@v4
+        with:
+          key: mkdocs-material-${{ env.cache_id }}
+          path: .cache
+          restore-keys: |
+            mkdocs-material-
+      - run: pip install mkdocs-material mkdocstrings[python]
+      # ====== Backup the benchmarks from gh-pages ======
+      # This is necessary because the benchmarks are not included in the documentation build process.
+      # So we need to backup the benchmarks from gh-pages and restore them after the documentation is built.
+      - name: Backup benchmarks from gh-pages
+        run: |
+          git fetch origin gh-pages
+          # create worktree bound to local gh-pages, tracking origin/gh-pages
+          git branch -f gh-pages origin/gh-pages || true
+          mkdir -p ghp && git worktree add ghp gh-pages || true
+          if [ -d ghp/benchmarks ]; then
+            tar -C ghp -czf /tmp/benchmarks.tgz benchmarks
+          fi
+          # IMPORTANT: remove worktree so gh-pages isn't checked out anywhere
+          git worktree remove ghp --force || true
+          echo "Backed up benchmarks from gh-pages"
+      # ====== Deploy the documentation ======
+      - name: Deploy documentation
+        run: mkdocs gh-deploy --force
+      # ====== Restore the benchmarks onto gh-pages ======
+      # This is necessary because the benchmarks are not included in the documentation build process.
+      # So we need to restore the benchmarks onto gh-pages after the documentation is built.
+      - name: Restore benchmarks onto gh-pages
+        run: |
+          # Refresh remote tracking and recreate a clean worktree
+          git fetch origin gh-pages
+          git worktree add -B gh-pages ghp origin/gh-pages
+          if [ -f /tmp/benchmarks.tgz ]; then
+            tar -C ghp -xzf /tmp/benchmarks.tgz
+            git -C ghp add -A
+            git -C ghp commit -m "Restore benchmarks after gh-deploy" || echo "No changes"
+            git -C ghp push origin gh-pages
+          fi
+          git worktree remove ghp --force || true

{liger_kernel-0.6.2 → liger_kernel-0.6.3}/.github/workflows/intel-ci.yml

@@ -13,9 +13,6 @@ on:
     paths:
       - "src/**"
       - "test/**"
-  schedule:
-    # Runs at 00:00 UTC daily
-    - cron: '0 0 * * *'
   workflow_dispatch:  # Enables manual trigger
 
 concurrency:

{liger_kernel-0.6.2 → liger_kernel-0.6.3}/.github/workflows/nvi-ci.yml

@@ -13,9 +13,6 @@ on:
     paths:
       - "src/**"
       - "test/**"
-  schedule:
-    # Runs at 00:00 UTC daily
-    - cron: '0 0 * * *'
   workflow_dispatch:  # Enables manual trigger
 
 concurrency:

{liger_kernel-0.6.2 → liger_kernel-0.6.3}/.gitignore

@@ -23,4 +23,5 @@ uv.lock
 
 # Benchmark images
 benchmark/visualizations
-.vscode/
+.vscode/
+.coverage

{liger_kernel-0.6.2 → liger_kernel-0.6.3}/Makefile

@@ -5,7 +5,15 @@ all: checkstyle test test-convergence
 
 # Command to run pytest for correctness tests
 test:
-	python -m pytest --disable-warnings test/ --ignore=test/convergence
+	python -m pytest --disable-warnings \
+		--cov=src/liger_kernel \
+		--cov-report=term-missing \
+		--ignore=test/convergence \
+		test/
+
+# Command to run coverage report
+coverage:
+	coverage report -m
 
 # Command to run ruff for linting and formatting code
 checkstyle:

{liger_kernel-0.6.2 → liger_kernel-0.6.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: liger_kernel
-Version: 0.6.2
+Version: 0.6.3
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -35,15 +35,14 @@ Requires-Dist: triton>=2.3.1
 Provides-Extra: dev
 Requires-Dist: transformers>=4.49.0; 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: ruff>=0.12.0; extra == "dev"
 Requires-Dist: pytest>=7.1.2; extra == "dev"
 Requires-Dist: pytest-xdist; extra == "dev"
+Requires-Dist: pytest-cov; extra == "dev"
+Requires-Dist: pytest-asyncio; extra == "dev"
 Requires-Dist: pytest-rerunfailures; extra == "dev"
 Requires-Dist: datasets>=2.19.2; extra == "dev"
 Requires-Dist: seaborn; extra == "dev"
-Requires-Dist: mkdocs; extra == "dev"
 Requires-Dist: mkdocs-material; extra == "dev"
 Requires-Dist: torchvision>=0.20; extra == "dev"
 Dynamic: license-file
@@ -181,8 +180,8 @@ y = orpo_loss(lm_head.weight, x, target)
 - `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
+pip install -e .[dev]
+pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm6.3/
 ```
 
 ### Optional Dependencies
@@ -216,6 +215,9 @@ pip install -e .
 
 # Setup Development Dependencies
 pip install -e ".[dev]"
+
+# NOTE -> For AMD users only
+pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm6.3/
 ```
 
 
@@ -312,6 +314,7 @@ loss.backward()
 | 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 |
 | GLM-4   | `liger_kernel.transformers.apply_liger_kernel_to_glm4`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
+| InternVL3   | `liger_kernel.transformers.apply_liger_kernel_to_internvl`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
 
 
 ## Low-level APIs
@@ -391,17 +394,17 @@ loss.backward()
         <td style="padding: 10px;">
             <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">
+                    <img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/nvi-ci.yml/badge.svg?branch=main&event=push" 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">
+                    <img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/amd-ci.yml/badge.svg?branch=main&event=push" alt="Build">
                 </a>
             </div>
             <div style="display: block;">
                 <a href="https://github.com/linkedin/Liger-Kernel/actions/workflows/intel-ci.yml">
-                    <img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/intel-ci.yml/badge.svg?event=schedule" alt="Build">
+                    <img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/intel-ci.yml/badge.svg?branch=main&event=push" alt="Build">
                 </a>
             </div>
         </td>

{liger_kernel-0.6.2 → liger_kernel-0.6.3}/README.md

@@ -129,8 +129,8 @@ y = orpo_loss(lm_head.weight, x, target)
 - `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
+pip install -e .[dev]
+pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm6.3/
 ```
 
 ### Optional Dependencies
@@ -164,6 +164,9 @@ pip install -e .
 
 # Setup Development Dependencies
 pip install -e ".[dev]"
+
+# NOTE -> For AMD users only
+pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm6.3/
 ```
 
 
@@ -260,6 +263,7 @@ loss.backward()
 | 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 |
 | GLM-4   | `liger_kernel.transformers.apply_liger_kernel_to_glm4`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
+| InternVL3   | `liger_kernel.transformers.apply_liger_kernel_to_internvl`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
 
 
 ## Low-level APIs
@@ -339,17 +343,17 @@ loss.backward()
         <td style="padding: 10px;">
             <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">
+                    <img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/nvi-ci.yml/badge.svg?branch=main&event=push" 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">
+                    <img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/amd-ci.yml/badge.svg?branch=main&event=push" alt="Build">
                 </a>
             </div>
             <div style="display: block;">
                 <a href="https://github.com/linkedin/Liger-Kernel/actions/workflows/intel-ci.yml">
-                    <img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/intel-ci.yml/badge.svg?event=schedule" alt="Build">
+                    <img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/intel-ci.yml/badge.svg?branch=main&event=push" alt="Build">
                 </a>
             </div>
         </td>

{liger_kernel-0.6.2 → liger_kernel-0.6.3}/benchmark/data/all_benchmark_data.csv

@@ -1575,6 +1575,70 @@ fused_add_rms_norm,liger_rms_norm,full,memory,MB,H,hidden size,4096,416.11767578
 fused_add_rms_norm,liger_rms_norm,full,memory,MB,H,hidden size,8192,832.22705078125,832.22705078125,832.22705078125,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
 fused_add_rms_norm,liger_rms_norm,full,memory,MB,H,hidden size,16384,1544.44580078125,1544.44580078125,1544.44580078125,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
 fused_add_rms_norm,liger_rms_norm,full,memory,MB,H,hidden size,32768,2960.8837890625,2960.8837890625,2960.8837890625,"{""M"": 2048, ""dtype"": ""torch.float32"", ""eps"": 1e-06}",NVIDIA H100 NVL,2025-07-16 07:04:45,0.6.0
+fused_linear_grpo_loss_token,liger,forward,speed,ms,B,B,2,40.75366401672363,40.749671173095706,40.75765686035156,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-04 23:58:45,0.6.1
+fused_linear_grpo_loss_token,liger,forward,speed,ms,B,B,4,80.95231628417969,80.95231628417969,80.95231628417969,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-04 23:58:45,0.6.1
+fused_linear_grpo_loss_token,liger,forward,speed,ms,B,B,8,163.58604431152344,163.58604431152344,163.58604431152344,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-04 23:58:45,0.6.1
+fused_linear_grpo_loss_token,liger,forward,speed,ms,B,B,16,323.6761474609375,323.6761474609375,323.6761474609375,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-04 23:58:45,0.6.1
+fused_linear_grpo_loss_token,torch,forward,speed,ms,B,B,2,23.71225643157959,23.612825775146483,23.8354434967041,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-04 23:59:51,0.6.1
+fused_linear_grpo_loss_token,torch,forward,speed,ms,B,B,4,46.86131286621094,46.80355911254883,46.91906661987304,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-04 23:59:51,0.6.1
+fused_linear_grpo_loss_token,torch,forward,speed,ms,B,B,8,94.54898834228516,94.54898834228516,94.54898834228516,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-04 23:59:51,0.6.1
+fused_linear_grpo_loss_token,torch,forward,speed,ms,B,B,16,189.99501037597656,189.99501037597656,189.99501037597656,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-04 23:59:51,0.6.1
+fused_linear_grpo_loss_token,liger,full,speed,ms,B,B,2,42.67263984680176,42.54085083007813,42.80442886352539,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:00:58,0.6.1
+fused_linear_grpo_loss_token,liger,full,speed,ms,B,B,4,82.2446060180664,82.2446060180664,82.2446060180664,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:00:58,0.6.1
+fused_linear_grpo_loss_token,liger,full,speed,ms,B,B,8,167.00416564941406,167.00416564941406,167.00416564941406,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:00:58,0.6.1
+fused_linear_grpo_loss_token,liger,full,speed,ms,B,B,16,327.0911865234375,327.0911865234375,327.0911865234375,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:00:58,0.6.1
+fused_linear_grpo_loss_token,torch,full,speed,ms,B,B,2,45.36115264892578,45.241344451904304,45.480960845947266,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:02:07,0.6.1
+fused_linear_grpo_loss_token,torch,full,speed,ms,B,B,4,90.00038146972656,90.00038146972656,90.00038146972656,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:02:07,0.6.1
+fused_linear_grpo_loss_token,torch,full,speed,ms,B,B,8,177.22674560546875,177.22674560546875,177.22674560546875,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:02:07,0.6.1
+fused_linear_grpo_loss_token,torch,full,speed,ms,B,B,16,356.5383605957031,356.5383605957031,356.5383605957031,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:02:07,0.6.1
+fused_linear_grpo_loss_token,liger,backward,speed,ms,B,B,2,1.814527988433838,1.8124799728393555,1.8167808055877686,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:03:11,0.6.1
+fused_linear_grpo_loss_token,liger,backward,speed,ms,B,B,4,1.84934401512146,1.8472959995269775,1.8524160385131836,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:03:11,0.6.1
+fused_linear_grpo_loss_token,liger,backward,speed,ms,B,B,8,1.891327977180481,1.8872319459915161,1.893990397453308,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:03:11,0.6.1
+fused_linear_grpo_loss_token,liger,backward,speed,ms,B,B,16,1.9722239971160889,1.9660799503326416,1.9763200283050537,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:03:11,0.6.1
+fused_linear_grpo_loss_token,torch,backward,speed,ms,B,B,2,22.014975547790527,21.710438537597657,22.19417533874512,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:04:16,0.6.1
+fused_linear_grpo_loss_token,torch,backward,speed,ms,B,B,4,41.83603096008301,41.752165222167974,41.91989669799805,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:04:16,0.6.1
+fused_linear_grpo_loss_token,torch,backward,speed,ms,B,B,8,81.66400146484375,81.66400146484375,81.66400146484375,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:04:16,0.6.1
+fused_linear_grpo_loss_token,torch,backward,speed,ms,B,B,16,162.6429443359375,162.6429443359375,162.6429443359375,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:04:16,0.6.1
+fused_linear_grpo_loss_token,liger,full,memory,MB,B,B,2,7344.77685546875,7344.77685546875,7344.77685546875,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:05:31,0.6.1
+fused_linear_grpo_loss_token,liger,full,memory,MB,B,B,4,7408.80029296875,7408.80029296875,7408.80029296875,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:05:31,0.6.1
+fused_linear_grpo_loss_token,liger,full,memory,MB,B,B,8,7536.84716796875,7536.84716796875,7536.84716796875,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:05:31,0.6.1
+fused_linear_grpo_loss_token,liger,full,memory,MB,B,B,16,7792.94091796875,7792.94091796875,7792.94091796875,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:05:31,0.6.1
+fused_linear_grpo_loss_token,torch,full,memory,MB,B,B,2,9083.28125,9083.28125,9083.28125,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:06:37,0.6.1
+fused_linear_grpo_loss_token,torch,full,memory,MB,B,B,4,13138.3125,13138.3125,13138.3125,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:06:37,0.6.1
+fused_linear_grpo_loss_token,torch,full,memory,MB,B,B,8,21250.375,21250.375,21250.375,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:06:37,0.6.1
+fused_linear_grpo_loss_token,torch,full,memory,MB,B,B,16,37474.5,37474.5,37474.5,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""token"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:06:37,0.6.1
+fused_linear_grpo_loss_sequence,liger,forward,speed,ms,B,B,2,40.72038269042969,40.71178131103516,40.728984069824214,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:07:48,0.6.1
+fused_linear_grpo_loss_sequence,liger,forward,speed,ms,B,B,4,81.69369506835938,81.69369506835938,81.69369506835938,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:07:48,0.6.1
+fused_linear_grpo_loss_sequence,liger,forward,speed,ms,B,B,8,162.79653930664062,162.79653930664062,162.79653930664062,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:07:48,0.6.1
+fused_linear_grpo_loss_sequence,liger,forward,speed,ms,B,B,16,323.6546630859375,323.6546630859375,323.6546630859375,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:07:48,0.6.1
+fused_linear_grpo_loss_sequence,torch,forward,speed,ms,B,B,2,23.70047950744629,23.628594589233398,23.732429122924806,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:08:54,0.6.1
+fused_linear_grpo_loss_sequence,torch,forward,speed,ms,B,B,4,47.36921691894531,47.085364532470706,47.65306930541992,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:08:54,0.6.1
+fused_linear_grpo_loss_sequence,torch,forward,speed,ms,B,B,8,94.83366394042969,94.83366394042969,94.83366394042969,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:08:54,0.6.1
+fused_linear_grpo_loss_sequence,torch,forward,speed,ms,B,B,16,190.0963897705078,190.0963897705078,190.0963897705078,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:08:54,0.6.1
+fused_linear_grpo_loss_sequence,liger,full,speed,ms,B,B,2,42.318336486816406,42.15214080810547,42.48453216552734,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:10:02,0.6.1
+fused_linear_grpo_loss_sequence,liger,full,speed,ms,B,B,4,82.4616928100586,82.4616928100586,82.4616928100586,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:10:02,0.6.1
+fused_linear_grpo_loss_sequence,liger,full,speed,ms,B,B,8,163.43756103515625,163.43756103515625,163.43756103515625,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:10:02,0.6.1
+fused_linear_grpo_loss_sequence,liger,full,speed,ms,B,B,16,325.4384765625,325.4384765625,325.4384765625,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:10:02,0.6.1
+fused_linear_grpo_loss_sequence,torch,full,speed,ms,B,B,2,45.99193572998047,45.80761489868165,46.176256561279295,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:11:10,0.6.1
+fused_linear_grpo_loss_sequence,torch,full,speed,ms,B,B,4,88.57190704345703,88.57190704345703,88.57190704345703,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:11:10,0.6.1
+fused_linear_grpo_loss_sequence,torch,full,speed,ms,B,B,8,176.94105529785156,176.94105529785156,176.94105529785156,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:11:10,0.6.1
+fused_linear_grpo_loss_sequence,torch,full,speed,ms,B,B,16,356.0478820800781,356.0478820800781,356.0478820800781,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:11:10,0.6.1
+fused_linear_grpo_loss_sequence,liger,backward,speed,ms,B,B,2,1.8242560029029846,1.8102271556854248,1.8309119939804077,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:12:14,0.6.1
+fused_linear_grpo_loss_sequence,liger,backward,speed,ms,B,B,4,1.84934401512146,1.846886396408081,1.8534400463104248,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:12:14,0.6.1
+fused_linear_grpo_loss_sequence,liger,backward,speed,ms,B,B,8,1.891327977180481,1.8892799615859985,1.8933759927749634,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:12:14,0.6.1
+fused_linear_grpo_loss_sequence,liger,backward,speed,ms,B,B,16,1.9752960205078125,1.9722239971160889,1.977344036102295,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:12:14,0.6.1
+fused_linear_grpo_loss_sequence,torch,backward,speed,ms,B,B,2,22.0262393951416,21.80997085571289,22.20482559204102,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:13:20,0.6.1
+fused_linear_grpo_loss_sequence,torch,backward,speed,ms,B,B,4,41.54521560668945,41.224806213378905,41.865625,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:13:20,0.6.1
+fused_linear_grpo_loss_sequence,torch,backward,speed,ms,B,B,8,81.21753692626953,81.21753692626953,81.21753692626953,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:13:20,0.6.1
+fused_linear_grpo_loss_sequence,torch,backward,speed,ms,B,B,16,160.82022094726562,160.82022094726562,160.82022094726562,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:13:20,0.6.1
+fused_linear_grpo_loss_sequence,liger,full,memory,MB,B,B,2,7344.77685546875,7344.77685546875,7344.77685546875,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:14:28,0.6.1
+fused_linear_grpo_loss_sequence,liger,full,memory,MB,B,B,4,7408.80029296875,7408.80029296875,7408.80029296875,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:14:28,0.6.1
+fused_linear_grpo_loss_sequence,liger,full,memory,MB,B,B,8,7536.84716796875,7536.84716796875,7536.84716796875,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:14:28,0.6.1
+fused_linear_grpo_loss_sequence,liger,full,memory,MB,B,B,16,7792.94091796875,7792.94091796875,7792.94091796875,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:14:28,0.6.1
+fused_linear_grpo_loss_sequence,torch,full,memory,MB,B,B,2,9083.28125,9083.28125,9083.28125,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:15:31,0.6.1
+fused_linear_grpo_loss_sequence,torch,full,memory,MB,B,B,4,13138.3125,13138.3125,13138.3125,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:15:31,0.6.1
+fused_linear_grpo_loss_sequence,torch,full,memory,MB,B,B,8,21250.375,21250.375,21250.375,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:15:31,0.6.1
+fused_linear_grpo_loss_sequence,torch,full,memory,MB,B,B,16,37474.5,37474.5,37474.5,"{""T"": 1024, ""H"": 4096, ""V"": 128256, ""importance_sampling_level"": ""sequence"", ""dtype"": ""torch.bfloat16""}",NVIDIA A100-SXM4-80GB,2025-08-05 00:15:31,0.6.1
 llama4_rope,liger,forward,speed,ms,H,hidden size,512,0.08249600231647491,0.08102399855852127,0.08432000130414963,"{""dtype"": ""torch.bfloat16"", ""seq_len"": 2048, ""num_q_heads"": 32, ""num_kv_heads"": 8}",NVIDIA H100 80GB HBM3,2025-08-07 21:42:01,0.6.1
 llama4_rope,liger,forward,speed,ms,H,hidden size,2048,0.08169600367546082,0.08037760108709335,0.08329600095748901,"{""dtype"": ""torch.bfloat16"", ""seq_len"": 2048, ""num_q_heads"": 32, ""num_kv_heads"": 8}",NVIDIA H100 80GB HBM3,2025-08-07 21:42:01,0.6.1
 llama4_rope,liger,forward,speed,ms,H,hidden size,8192,0.08128000050783157,0.07980799674987793,0.08329600095748901,"{""dtype"": ""torch.bfloat16"", ""seq_len"": 2048, ""num_q_heads"": 32, ""num_kv_heads"": 8}",NVIDIA H100 80GB HBM3,2025-08-07 21:42:01,0.6.1

{liger_kernel-0.6.2 → liger_kernel-0.6.3}/benchmark/scripts/benchmark_cross_entropy.py

@@ -70,6 +70,9 @@ def bench_speed_cross_entropy(
 
     if mode == "forward":
         ms_50, ms_20, ms_80 = triton.testing.do_bench(fwd, rep=100, quantiles=QUANTILES)
+    elif mode == "no-grad-forward":
+        with torch.no_grad():
+            ms_50, ms_20, ms_80 = triton.testing.do_bench(fwd, rep=100, quantiles=QUANTILES)
     elif mode == "backward":
         y = fwd()
 
@@ -109,7 +112,7 @@ if __name__ == "__main__":
 
     run_benchmarks(
         bench_test_fn=bench_speed_cross_entropy,
-        kernel_operation_modes=["forward", "backward", "full"],
+        kernel_operation_modes=["forward", "backward", "full", "no-grad-forward"],
         metric_name="speed",
         metric_unit="ms",
         **common_configs,

{liger_kernel-0.6.2 → liger_kernel-0.6.3}/benchmark/scripts/benchmark_fused_linear_cross_entropy.py

@@ -59,26 +59,26 @@ def bench_memory_fused_linear_cross_entropy(
     dtype = input.extra_benchmark_config["dtype"]
     provider = input.kernel_provider
 
-    torch_lm_head_ce = TorchLMHeadCE(H=H, V=V, dtype=dtype).to(device)
-    liger_lm_head_ce = LigerLMHeadCE(H=H, V=V, dtype=dtype).to(device)
-    liger_lm_head_ce_fp32_accum = LigerLMHeadCE(H=H, V=V, dtype=dtype, accum_dtype=torch.float32).to(device)
+    lm_head_ce = None
+    if provider == "liger":
+        lm_head_ce = LigerLMHeadCE(H=H, V=V, dtype=dtype).to(device)
+    elif provider == "liger-fp32-accum":
+        lm_head_ce = LigerLMHeadCE(H=H, V=V, dtype=dtype, accum_dtype=torch.float32).to(device)
+    else:
+        lm_head_ce = TorchLMHeadCE(H=H, V=V, dtype=dtype).to(device)
 
     _input = torch.randn(BT, H, requires_grad=True, dtype=dtype, device=device)
     target = torch.randint(V, (BT, 1), dtype=torch.long, device=device).squeeze(1)
 
     def fwd():
-        if provider == "liger":
-            return liger_lm_head_ce(_input, target)
-        elif provider == "liger-fp32-accum":
-            return liger_lm_head_ce_fp32_accum(_input, target)
-        elif provider == "huggingface":
-            return torch_lm_head_ce(_input, target)
+        return lm_head_ce(_input, target)
 
     def full():
         y = fwd()
         y.backward()
 
     mem_50, mem_20, mem_80 = _test_memory(full, _iter=10, quantiles=QUANTILES)
+
     return SingleBenchmarkRunOutput(
         y_20=mem_20,
         y_50=mem_50,
@@ -101,20 +101,19 @@ def bench_speed_fused_linear_cross_entropy(
     provider = input.kernel_provider
     mode = input.kernel_operation_mode
 
-    torch_lm_head_ce = TorchLMHeadCE(H=H, V=V, dtype=dtype).to(device)
-    liger_lm_head_ce = LigerLMHeadCE(H=H, V=V, dtype=dtype).to(device)
-    liger_lm_head_ce_fp32_accum = LigerLMHeadCE(H=H, V=V, dtype=dtype, accum_dtype=torch.float32).to(device)
+    lm_head_ce = None
+    if provider == "liger":
+        lm_head_ce = LigerLMHeadCE(H=H, V=V, dtype=dtype).to(device)
+    elif provider == "liger-fp32-accum":
+        lm_head_ce = LigerLMHeadCE(H=H, V=V, dtype=dtype, accum_dtype=torch.float32).to(device)
+    else:
+        lm_head_ce = TorchLMHeadCE(H=H, V=V, dtype=dtype).to(device)
 
     _input = torch.randn(BT, H, requires_grad=True, dtype=dtype, device=device)
     target = torch.randint(V, (BT, 1), dtype=torch.long, device=device).squeeze(1)
 
     def fwd():
-        if provider == "liger":
-            return liger_lm_head_ce(_input, target)
-        elif provider == "liger-fp32-accum":
-            return liger_lm_head_ce_fp32_accum(_input, target)
-        elif provider == "huggingface":
-            return torch_lm_head_ce(_input, target)
+        return lm_head_ce(_input, target)
 
     if mode == "forward":
         ms_50, ms_20, ms_80 = triton.testing.do_bench(
@@ -122,6 +121,13 @@ def bench_speed_fused_linear_cross_entropy(
             rep=100,
             quantiles=QUANTILES,
         )
+    elif mode == "no-grad-forward":
+        with torch.no_grad():
+            ms_50, ms_20, ms_80 = triton.testing.do_bench(
+                fwd,
+                rep=100,
+                quantiles=QUANTILES,
+            )
     elif mode == "backward":
         y = fwd()
 
@@ -164,7 +170,7 @@ if __name__ == "__main__":
 
     run_benchmarks(
         bench_test_fn=bench_speed_fused_linear_cross_entropy,
-        kernel_operation_modes=["forward", "backward", "full"],
+        kernel_operation_modes=["forward", "backward", "full", "no-grad-forward"],
         metric_name="speed",
         metric_unit="ms",
         **common_configs,