fix-confilct

Author: henghengxiedaima

.github/workflows/code_scan.yaml

@@ -9,7 +9,7 @@ on:
 jobs:
   scan-code-and-report:
     runs-on: scan
-    if: ${{ github.repository == 'Flagopen/FlagGems' }}
+    if: ${{ github.repository == 'flagos-ai/FlagGems' }}
     concurrency:
       group: scan-code-and-report-${{ github.event.pull_request.number || github.ref }}
       cancel-in-progress: true

.github/workflows/gems-cpp-extension.yaml

@@ -14,6 +14,7 @@ on:
       - '**/*.cmake'
       - '**/*.mk'
       - 'Makefile'
+      - '!src/flag_gems/experimental_ops/**'
   pull_request:
     branches: [ "master" ]
     paths:
@@ -27,6 +28,7 @@ on:
       - '**/*.cmake'
       - '**/*.mk'
       - 'Makefile'
+      - '!src/flag_gems/experimental_ops/**'
 
 jobs:
   build-C-extension:

.github/workflows/gems-experimental-test.yaml

@@ -0,0 +1,57 @@
+name: gems-experimental-test
+
+on:
+  push:
+    branches: [ "master" ]
+    paths:
+      - 'src/flag_gems/experimental_ops/**'
+  pull_request:
+    branches: [ "master" ]
+    paths:
+      - 'src/flag_gems/experimental_ops/**'
+
+env:
+  CUDA_VISIBLE_DEVICES: 7
+  http_proxy: ${{ secrets.HTTP_PROXY }}
+  https_proxy: ${{ secrets.HTTPS_PROXY }}
+
+jobs:
+  experimental-test-on-hopper:
+    runs-on: hopper
+    concurrency:
+      group: experimental-test-on-hopper-${{ github.event.pull_request.number || github.ref }}
+      cancel-in-progress: true
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: FlagGems experimental tests on hopper
+        shell: bash
+        run: |
+          source "/home/zhangzhihui/miniconda3/etc/profile.d/conda.sh"
+          conda activate flag_gems
+          source tools/run_command.sh
+          BASE_SHA=${{ github.event.pull_request.base.sha }}
+          HEAD_SHA=${{ github.sha }}
+
+          echo "Diffing $BASE_SHA...$HEAD_SHA"
+
+          changed_ops=$(git diff --name-only $BASE_SHA...$HEAD_SHA | grep '^src/flag_gems/experimental_ops/.*\.py$' || true)
+
+          tests=""
+          for f in $changed_ops; do
+              base=$(basename "$f" .py)
+              test_file="src/flag_gems/experimental_ops/exp_tests/${base}_test.py"
+              if [ -f "$test_file" ]; then
+                  tests="$tests $test_file"
+              fi
+          done
+
+          if [ -n "$tests" ]; then
+              echo "Running tests:$tests"
+              run_command pytest -s $tests
+          else
+              echo "No relevant ops changes, skipping tests"
+          fi

.github/workflows/gems-test-on-hopper.yaml

@@ -15,6 +15,7 @@ on:
       - '**/*.mk'
       - 'Makefile'
       - '**.yaml'
+      - '!src/flag_gems/experimental_ops/**'
   pull_request:
     branches: [ "master" ]
     paths:
@@ -29,6 +30,7 @@ on:
       - '**/*.mk'
       - 'Makefile'
       - '**.yaml'
+      - '!src/flag_gems/experimental_ops/**'
 
 env:
   CUDA_VISIBLE_DEVICES: 7

.github/workflows/gems-test-on-metax.yaml

@@ -14,6 +14,7 @@ on:
       - '**/*.cmake'
       - '**/*.mk'
       - 'Makefile'
+      - '!src/flag_gems/experimental_ops/**'
   pull_request:
     branches: [ "master" ]
     paths:
@@ -27,6 +28,7 @@ on:
       - '**/*.cmake'
       - '**/*.mk'
       - 'Makefile'
+      - '!src/flag_gems/experimental_ops/**'
 
 jobs:
   op-test-on-metax:

.github/workflows/python-coverage.yaml

@@ -14,6 +14,7 @@ on:
       - '**/*.cmake'
       - '**/*.mk'
       - 'Makefile'
+      - '!src/flag_gems/experimental_ops/**'
   pull_request:
     branches: [ "master" ]
     paths:
@@ -27,6 +28,7 @@ on:
       - '**/*.cmake'
       - '**/*.mk'
       - 'Makefile'
+      - '!src/flag_gems/experimental_ops/**'
 
 jobs:
   blas-op-test:

CMakeLists.txt

@@ -82,6 +82,21 @@ if(FLAGGEMS_BUILD_CTESTS)
   FetchContent_MakeAvailable(googletest)
 endif()
 
+execute_process(
+  COMMAND ${Python_EXECUTABLE}
+          -c "import triton, sys; sys.stdout.write(triton.__version__)"
+  RESULT_VARIABLE _ret
+  OUTPUT_VARIABLE TRITON_VERSION
+  OUTPUT_STRIP_TRAILING_WHITESPACE
+)
+if(NOT _ret EQUAL 0)
+  message(FATAL_ERROR "Triton IS NOT installed properly")
+endif()
+if(TRITON_VERSION VERSION_GREATER_EQUAL "3.5")
+  set(TRITON_GE_3P5 1)
+endif()
+message(STATUS "Detected Triton version: ${TRITON_VERSION}")
+
 # --------------------------- subdirectories ---------------------------
 add_subdirectory(lib)
 add_subdirectory(src/flag_gems/csrc)

LICENSE

@@ -1,4 +1,3 @@
-Copyright © 2024 BAAI. All rights reserved.
 
                                  Apache License
                            Version 2.0, January 2004

README.md

@@ -98,6 +98,51 @@ FlagGems can be installed either as a pure Python package or as a package with C
 
 For a quick start with installing and using flag_gems, please refer to the documentation [GetStart](docs/get_start_with_flaggems.md).
 
+## Usage
+
+FlagGems supports two common usage patterns: patching PyTorch ATen ops (recommended) and calling FlagGems ops explicitly.
+
+### (1) Enable FlagGems globally (patch ATen ops)
+
+After `flag_gems.enable()`, supported `torch.*` / `torch.nn.functional.*` calls will be dispatched to FlagGems implementations automatically.
+
+```python
+import torch
+import flag_gems
+
+flag_gems.enable()
+
+x = torch.randn(4096, 4096, device=flag_gems.device, dtype=torch.float16)
+y = torch.mm(x, x)
+```
+
+If you only want FlagGems inside a scope (e.g., for benchmarking), use the context manager:
+
+```python
+import torch
+import flag_gems
+
+with flag_gems.use_gems():
+    x = torch.randn(4096, 4096, device=flag_gems.device, dtype=torch.float16)
+    y = torch.mm(x, x)
+```
+
+### (2) Explicitly call FlagGems ops
+
+You can also bypass PyTorch dispatch and call operators from `flag_gems.ops` directly (no `enable()` required):
+
+```python
+import torch
+from flag_gems import ops
+import flag_gems
+
+a = torch.randn(1024, 1024, device=flag_gems.device, dtype=torch.float16)
+b = torch.randn(1024, 1024, device=flag_gems.device, dtype=torch.float16)
+c = ops.mm(a, b)
+```
+
+For more details and advanced options (disabling specific ops, runtime logging,e.g.), see [`how_to_use_flaggems`](docs/how_to_use_flaggems.md).
+
 ## Supported Operators
 
 Operators will be implemented according to [OperatorList](docs/operator_list.md).
@@ -128,7 +173,7 @@ Operators will be implemented according to [OperatorList](docs/operator_list.md)
 
 The following chart shows the speedup of FlagGems compared with PyTorch ATen library in eager mode. The speedup is calculated by averaging the speedup on each shape, representing the overall performance of the operator.
 
-![Operator Speedup](./docs/assets/speedup-20250423.png)
+![Operator Speedup](./docs/assets/speedup-20251225.png)
 
 ## Contributions
 

README_cn.md

@@ -93,6 +93,50 @@ FlagGems 可以作为纯 Python 包安装，也可以作为带有 C++ 扩展的
 
 参考文档 [开始使用](docs/get_start_with_flaggems.md) 快速安装使用 flag_gems
 
+## 使用方法
+
+FlagGems 支持两种常见的使用模式：对 PyTorch ATen 算子打补丁（推荐）和显式调用 FlagGems 算子。
+
+### (1) 全局启用 FlagGems（对 ATen 算子打补丁）
+
+执行 `flag_gems.enable()` 后，支持的 `torch.*` / `torch.nn.functional.*` 调用将会自动分发（dispatch）到 FlagGems 的实现上。
+
+```python
+import torch
+import flag_gems
+
+flag_gems.enable()
+
+x = torch.randn(4096, 4096, device=flag_gems.device, dtype=torch.float16)
+y = torch.mm(x, x)
+```
+
+如果你只想在某个作用域内（例如用于基准测试）使用 FlagGems，请使用上下文管理器：
+
+```python
+import torch
+import flag_gems
+
+with flag_gems.use_gems():
+    x = torch.randn(4096, 4096, device=flag_gems.device, dtype=torch.float16)
+    y = torch.mm(x, x)
+```
+
+### (2) 显式调用 FlagGems 算子
+你也可以绕过 PyTorch 的分发机制，直接从 flag_gems.ops 中调用算子，此时无需调用 enable()：
+
+```python
+import torch
+from flag_gems import ops
+import flag_gems
+
+a = torch.randn(1024, 1024, device=flag_gems.device, dtype=torch.float16)
+b = torch.randn(1024, 1024, device=flag_gems.device, dtype=torch.float16)
+c = ops.mm(a, b)
+```
+若要了解更多详情和高级选项（例如禁用特定算子、运行时日志等），请参阅 [`how_to_use_flaggems`](docs/how_to_use_flaggems.md)。
+
+
 ## 支持算子
 
 算子将按照文档 [OperatorList](docs/operator_list.md) 的顺序逐步实现。
@@ -123,7 +167,7 @@ FlagGems 可以作为纯 Python 包安装，也可以作为带有 C++ 扩展的
 
 FlagGems 相比 Torch Eager 模式下 ATen 算子库的加速比如下图所示。其中，每个算子的加速比综合了多个形状测例的数据，代表该算子的整体性能。
 
-![算子加速比](./docs/assets/speedup-20250423.png)
+![算子加速比](./docs/assets/speedup-20251225.png)
 
 ## 贡献代码