add format CI

.github/workflows/format.yml

@@ -0,0 +1,32 @@
+name: format
+
+on:
+  workflow_dispatch:
+  pull_request:
+  push:
+    branches:
+      - main
+
+jobs:
+  clang-format:
+    runs-on: ubuntu-latest
+    steps:
+    - uses: actions/checkout@v4
+    - uses: cpp-linter/cpp-linter-action@v2
+      id: cpp-lint
+      with:
+        style: file
+        tidy-checks: '-*' # disable clang tidy at this stage
+        version: 17
+    - name: Fail test
+      if: steps.cpp-lint.outputs.checks-failed > 0
+      run: echo "Some files failed the linting checks!" && exit 1
+
+  python-black:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: psf/black@stable
+        with: # see: https://black.readthedocs.io/en/stable/getting_started.html
+          src: "dipu"
+          version: "~= 23.11.0"

deeplink_ext/internlm_ops/mha/mha_varlen_kvpacked_func.py

@@ -3,6 +3,7 @@
 import torch
 import deeplink_ext.cpp_extensions as ext
 
+
 class DeepLinkMultiHeadAttentionVarLenKVPackedFunc(torch.autograd.Function):
     @staticmethod
     def forward(

deeplink_ext/internlm_ops/mha/mha_varlen_qkvpacked_func.py

@@ -3,6 +3,7 @@
 import torch
 import deeplink_ext.cpp_extensions as ext
 
+
 class DeepLinkMultiHeadAttentionVarLenQKVPackedFunc(torch.autograd.Function):
     @staticmethod
     def forward(

deeplink_ext/patch_internlm.py

@@ -56,7 +56,9 @@ def _patch_ops():
 
         import internlm.model.norm
 
-        internlm.model.norm.RMSNormTorch = ext.rms_norm.DeepLinkRMSNormWithNormalizedShape
+        internlm.model.norm.RMSNormTorch = (
+            ext.rms_norm.DeepLinkRMSNormWithNormalizedShape
+        )
 
     _find_or_mock_module("rotary_emb")
     _find_or_mock_module("fused_dense_lib")

tests/test_mha_internlm.py

@@ -40,9 +40,7 @@ def _run_cross_attention(
 output_gold, dq_gold, dkv_gold = _run_cross_attention(
     ext.fallback.CrossAttention, q, kv
 )
-output_ext, dq_ext, dkv_ext = _run_cross_attention(
-    ext.DeepLinkCrossAttention, q, kv
-)
+output_ext, dq_ext, dkv_ext = _run_cross_attention(ext.DeepLinkCrossAttention, q, kv)
 assert torch.allclose(output_gold, output_ext, atol=1e-3)
 print("CrossAttention forward test pass")
 assert torch.allclose(dq_gold, dq_ext, atol=2e-3)