Merge branch 'master' into pr/linalg-lstsq-metax2

Author: cheersluvs

.github/CODEOWNERS

@@ -20,12 +20,11 @@
 /src/flag_gems/*     @0x45f @huangyiqun @103yiran @douxetpur @bin913 @w1120029931-bit @hellojack163 @Caeruleann @tengqm
 /src/flag_gems/runtime/backend/*
 
-/tests/*             @0x45f @huangyiqun @103yiran @douxetpur @bin913 @w1120029931-bit @hellojack163 @Caeruleann @tengqm
+/benchmark/*         @0x45f @huangyiqun @103yiran @douxetpur @bin913 @w1120029931-bit @hellojack163 @Caeruleann
+/tests/*             @0x45f @huangyiqun @103yiran @douxetpur @bin913 @w1120029931-bit @hellojack163 @Caeruleann
 /modules_tests/*     @0x45f @huangyiqun @103yiran @douxetpur @bin913 @w1120029931-bit @hellojack163 @Caeruleann
-/benchmark/*         @0x45f @huangyiqun @103yiran @douxetpur @bin913 @w1120029931-bit @hellojack163 @Caeruleann @tengqm
-/docs/*              @0x45f @huangyiqun @103yiran @tengqm
-/container/*         @0x45f @huangyiqun @103yiran @tengqm
+/docs/*              @0x45f @huangyiqun @103yiran
+/tools/*             @0x45f @huangyiqun @103yiran
 
-/lib/*                @0x45f @huangyiqun @103yiran @douxetpur @bin913 @w1120029931-bit @hellojack163 @Caeruleann
-/ctests/*             @0x45f @huangyiqun @103yiran @douxetpur @bin913 @w1120029931-bit @hellojack163 @Caeruleann
-/include/*            @0x45f @huangyiqun @103yiran @douxetpur @bin913 @w1120029931-bit @hellojack163 @Caeruleann
+/cpp/*               @0x45f @huangyiqun @103yiran @douxetpur @bin913 @w1120029931-bit @hellojack163 @Caeruleann
+/triton_src/*        @0x45f @huangyiqun @103yiran @douxetpur @bin913 @w1120029931-bit @hellojack163 @Caeruleann

.github/workflows/command.yaml

@@ -44,16 +44,19 @@ jobs:
           skip_reviews: "true"
           fork_review_bypass: "true"
           permissions: "write,admin"
+          param_separator: " "
 
       - id: parse-command
         if: ${{ steps.check.outputs.continue == 'true' }}
         env:
           PULL_NUMBER: ${{ github.event.issue.number }}
+          COMMENT_BODY: ${{ steps.check.outputs.comment_body }}
         run: |
           # Parse command arguments
-          params=${{ steps.check.outputs.params }}
-          op=$(echo $params | cut -d ':' -f 1)
-          runner=$(echo $params | cut -d ':' -f 2)
+          # Comment format: "/test <op>:<runner>", e.g. "/test constant_pad_nd:mthreads"
+          params="${COMMENT_BODY#/test }"
+          op="${params%%:*}"
+          runner="${params##*:}"
           echo "OP_ID=${op}" >> $GITHUB_OUTPUT
           echo "RUNNER=${runner}" >> $GITHUB_OUTPUT
           echo "PR_NUMBER=${PULL_NUMBER}" >> $GITHUB_OUTPUT

benchmark/base.py

@@ -310,8 +310,9 @@ def get_latency(self, op, *args, **kwargs):
                 do_bench = triton.backends.ascend.testing.do_bench_npu
                 latency = do_bench(
                     fn,
-                    warmup=Config.warm_up,
-                    active=Config.repetition,
+                    # do_bench_npu requires iterations, rather than duration
+                    # warmup=Config.warm_up,
+                    # active=Config.repetition,
                 )
             else:
                 do_bench = triton.testing.do_bench

benchmark/core_shapes.yaml

@@ -272,6 +272,38 @@ grouped_mm:
     - [64, 2048, 128]
   shape_desc: "Groups, N, K"   # shapes are defined as (Groups, N, K)
 
+sparse_sampled_addmm:
+  shapes:
+    - [2, 128, 128, 128]
+    - [4, 256, 256, 256]
+    - [8, 512, 512, 512]
+    - [16, 512, 512, 512]
+    - [4, 1024, 1024, 1024]
+    - [16, 1024, 1024, 1024]
+    - [8, 2048, 2048, 64]
+    - [2, 2048, 2048, 128]
+    - [4, 4096, 4096, 128]
+    - [1, 4096, 4096, 64]
+    - [2, 4096, 1024, 512]
+    - [2, 1024, 1000, 384]
+  shape_desc: "B, M, N, K"
+
+sparse_sampled_addmm_out:
+  shapes:
+    - [2, 128, 128, 128]
+    - [4, 256, 256, 256]
+    - [8, 512, 512, 512]
+    - [16, 512, 512, 512]
+    - [4, 1024, 1024, 1024]
+    - [16, 1024, 1024, 1024]
+    - [8, 2048, 2048, 64]
+    - [2, 2048, 2048, 128]
+    - [4, 4096, 4096, 128]
+    - [1, 4096, 4096, 64]
+    - [2, 4096, 1024, 512]
+    - [2, 1024, 1000, 384]
+  shape_desc: "B, M, N, K"
+
 MvAndOuterBenchmark:
   shapes:
     - [384, 384]

benchmark/test__native_batch_norm_legit_no_training.py

@@ -0,0 +1,63 @@
+# Copyright 2026 FlagOS Contributors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import pytest
+import torch
+import torch.nn.functional as F
+
+from . import base, consts
+
+
+class NormBenchmark(base.GenericBenchmark):
+    def set_more_shapes(self):
+        return [
+            # 3D shapes represented as [batch_size, channels, hidden_size]
+            (16, 16, 64),
+            (16, 16, 1024),
+            (16, 16, 4098),
+            # 4D shapes represented as [batch_size, channels, H, W]
+            (1, 8, 4, 4),
+            (16, 8, 128, 128),
+        ]
+
+
+def native_batch_norm_legit_no_training_input_fn(shape, dtype, device):
+    C = shape[1]
+    inp = torch.randn(shape, dtype=dtype, device=device)
+    weight = torch.randn((C,), dtype=dtype, device=device)
+    bias = torch.randn((C,), dtype=dtype, device=device)
+    running_mean = torch.randn((C,), dtype=dtype, device=device)
+    running_var = torch.abs(torch.randn((C,), dtype=dtype, device=device)) + 0.1
+    momentum = 0.1
+    eps = 1e-5
+    yield inp, weight, bias, running_mean, running_var, momentum, eps
+
+
+def torch_native_batch_norm_legit_no_training(
+    inp, weight, bias, running_mean, running_var, momentum, eps
+):
+    return F.batch_norm(
+        inp, running_mean, running_var, weight, bias, training=False, eps=eps
+    )
+
+
+@pytest.mark.native_batch_norm_legit_no_training
+def test_native_batch_norm_legit_no_training():
+    bench = NormBenchmark(
+        input_fn=native_batch_norm_legit_no_training_input_fn,
+        op_name="native_batch_norm_legit_no_training",
+        torch_op=torch_native_batch_norm_legit_no_training,
+        dtypes=consts.FLOAT_DTYPES,
+    )
+    bench.run()

benchmark/test_absolute.py

@@ -24,3 +24,14 @@ def test_absolute():
         op_name="absolute", torch_op=torch.absolute, dtypes=consts.FLOAT_DTYPES
     )
     bench.run()
+
+
+@pytest.mark.absolute_
+def test_absolute_inplace():
+    bench = base.UnaryPointwiseBenchmark(
+        op_name="absolute_",
+        torch_op=lambda x: x.absolute_(),
+        dtypes=consts.FLOAT_DTYPES,
+        is_inplace=True,
+    )
+    bench.run()

benchmark/test_arctanh.py

@@ -13,12 +13,26 @@
 # limitations under the License.
 
 import pytest
+import torch
 
 import flag_gems
 
 from . import base, consts
 
 
+@pytest.mark.arctanh
+@pytest.mark.skipif(
+    flag_gems.vendor_name == "tsingmicro", reason="Issue #4131: not working"
+)
+def test_arctanh():
+    bench = base.UnaryPointwiseBenchmark(
+        op_name="arctanh",
+        torch_op=torch.arctanh,
+        dtypes=consts.FLOAT_DTYPES,
+    )
+    bench.run()
+
+
 @pytest.mark.arctanh_
 @pytest.mark.skipif(
     flag_gems.vendor_name == "tsingmicro", reason="Issue #4131: not working"
@@ -31,3 +45,16 @@ def test_arctanh_inplace():
         is_inplace=True,
     )
     bench.run()
+
+
+@pytest.mark.arctanh_out
+@pytest.mark.skipif(
+    flag_gems.vendor_name == "tsingmicro", reason="Issue #4131: not working"
+)
+def test_arctanh_out():
+    bench = base.UnaryPointwiseOutBenchmark(
+        op_name="arctanh_out",
+        torch_op=torch.arctanh,
+        dtypes=consts.FLOAT_DTYPES,
+    )
+    bench.run()