Merge branch 'flagos-ai:master' into update-topk_softmax

Author: AdvancedCompiler

.github/CODEOWNERS

@@ -0,0 +1,16 @@
+# These owners will be the default owners for everything in
+# the repo. Unless a later match takes precedence,
+# @global-owner1 and @global-owner2 will be requested for
+# review when someone opens a pull request.
+
+/src/flag_gems/*      @kiddyjinjin   @0x45f
+/src/flag_gems/runtime/backend/*
+
+/tests/*             @kiddyjinjin   @0x45f
+/modules_tests/*     @kiddyjinjin   @0x45f
+/benchmark/*         @kiddyjinjin   @0x45f
+
+
+/lib/*                 @zhangpeiyang1  @huangyiqun
+/ctests/*              @zhangpeiyang1  @huangyiqun
+/include/*             @zhangpeiyang1  @huangyiqun

.github/workflows/code_scan.yaml

@@ -9,6 +9,7 @@ on:
 jobs:
   scan-code-and-report:
     runs-on: scan
+    if: ${{ github.repository == 'Flagopen/FlagGems' }}
     concurrency:
       group: scan-code-and-report-${{ github.event.pull_request.number || github.ref }}
       cancel-in-progress: true

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

@@ -87,6 +87,14 @@ jobs:
           run_command pytest -s tests/test_special_ops.py && \
           run_command pytest -s tests/test_distribution_ops.py
 
+      - name: FlagGems convolution ops on hopper
+        shell: bash
+        run: |
+          source "/home/zhangzhihui/miniconda3/etc/profile.d/conda.sh"
+          conda activate flag_gems
+          source tools/run_command.sh
+          run_command pytest -s tests/test_convolution_ops.py
+
       - name: FlagGems utils on hopper
         shell: bash
         run: |

benchmark/performance_utils.py

@@ -262,12 +262,7 @@ def set_gems(self, gems_op):
         self.gems_op = gems_op
 
     def get_latency(self, op, *args, **kwargs):
-        if self.is_inplace:
-            fn = lambda: op(
-                *[x.clone() if torch.is_tensor(x) else x for x in args], **kwargs
-            )
-        else:
-            fn = lambda: op(*args, **kwargs)
+        fn = lambda: op(*args, **kwargs)
         if self.is_backward:
             out = fn()
             dout = torch.randn_like(out)

benchmark/test_convolution_perf.py

@@ -14,6 +14,10 @@ def set_more_shapes(self):
             (16, 24, 2048, 96, 7, 1, 3, 2),
             (8, 8, 8192, 16, 11, 4, 5, 1),
             (4, 4, 16384, 4, 15, 2, 7, 1),
+            (32, 64, 512, 64, 3, 1, "valid", 1),
+            (64, 48, 1024, 128, 5, 2, "valid", 1),
+            (16, 24, 2048, 96, 7, 1, "same", 2),
+            (8, 8, 8192, 16, 11, 1, "same", 1),
         ]
 
 
@@ -63,6 +67,10 @@ def set_more_shapes(self):
             (16, 32, 12, 12, 24, 3, 3, 2, 1, 1),
             (16, 32, 24, 24, 24, 3, 3, 2, 2, 2),
             (16, 32, 24, 24, 24, 3, 3, 1, 2, 2),
+            (16, 32, 12, 12, 24, 3, 3, 2, "valid", 1),
+            (32, 64, 128, 128, 32, 3, 3, 1, "valid", 1),
+            (16, 32, 24, 24, 24, 3, 3, 1, "same", 2),
+            (32, 64, 210, 210, 16, 5, 5, 1, "same", 1),
         ]
 
 

benchmark/test_unary_pointwise_perf.py

@@ -13,6 +13,7 @@
 )
 from benchmark.performance_utils import Benchmark, SkipVersion, generate_tensor_input
 
+vendor_name = flag_gems.vendor_name
 fp64_is_supported = flag_gems.runtime.device.support_fp64
 
 
@@ -90,10 +91,11 @@ def get_tflops(self, op, *args, **kwargs):
     ],
 )
 def test_general_unary_pointwise_perf(op_name, torch_op, dtypes):
-    if op_name in ["celu"] and SkipVersion("torch", "<2.5"):
-        pytest.skip(
-            "There is an error in kunlunxin torch 2.0 aten, please use torch 2.5 instead"
-        )
+    if vendor_name == "kunlunxin":
+        if op_name in ["celu"] and SkipVersion("torch", "<2.5"):
+            pytest.skip(
+                "There is an error in kunlunxin torch 2.0 aten, please use torch 2.5 instead"
+            )
     bench = UnaryPointwiseBenchmark(op_name=op_name, torch_op=torch_op, dtypes=dtypes)
     bench.run()
 
@@ -139,10 +141,11 @@ def test_general_unary_pointwise_perf(op_name, torch_op, dtypes):
     ],
 )
 def test_general_inplace_unary_pointwise_perf(op_name, torch_op, dtypes):
-    if op_name in ["celu_"] and SkipVersion("torch", "<2.5"):
-        pytest.skip(
-            "There is an error in kunlunxin torch 2.0 aten, please use torch 2.5 instead"
-        )
+    if vendor_name == "kunlunxin":
+        if op_name in ["celu_"] and SkipVersion("torch", "<2.5"):
+            pytest.skip(
+                "There is an error in kunlunxin torch 2.0 aten, please use torch 2.5 instead"
+            )
     bench = UnaryPointwiseBenchmark(
         op_name=op_name, torch_op=torch_op, dtypes=dtypes, is_inplace=True
     )