[shardformer] update colo attention to support custom mask

.github/workflows/build_on_pr.yml

@@ -117,7 +117,7 @@ jobs:
           cd TensorNVMe
           conda install cmake
           pip install -r requirements.txt
-          pip install -v .
+          DISABLE_URING=1 pip install -v .
 
       - name: Store TensorNVMe Cache
         run: |
@@ -201,4 +201,4 @@ jobs:
         uses: actions/upload-artifact@v3
         with:
           name: report
-          path: report/
+          path: report/

.github/workflows/build_on_schedule.yml

@@ -44,7 +44,7 @@ jobs:
           cd TensorNVMe
           conda install cmake
           pip install -r requirements.txt
-          pip install -v .
+          DISABLE_URING=1 pip install -v .
 
       - uses: actions/checkout@v2
         if: steps.check-avai.outputs.avai == 'true'

.github/workflows/compatiblity_test_on_dispatch.yml

@@ -66,7 +66,7 @@ jobs:
           cd TensorNVMe
           apt update && apt install -y cmake
           pip install -r requirements.txt
-          pip install -v .
+          DISABLE_URING=1 pip install -v .
       - uses: actions/checkout@v2
         with:
           ssh-key: ${{ secrets.SSH_KEY_FOR_CI }}

.github/workflows/compatiblity_test_on_pr.yml

@@ -60,7 +60,7 @@ jobs:
           cd TensorNVMe
           apt update && apt install -y cmake
           pip install -r requirements.txt
-          pip install -v .
+          DISABLE_URING=1 pip install -v .
       - uses: actions/checkout@v2
         with:
           ssh-key: ${{ secrets.SSH_KEY_FOR_CI }}

.github/workflows/compatiblity_test_on_schedule.yml

@@ -56,7 +56,7 @@ jobs:
           cd TensorNVMe
           apt update && apt install -y cmake
           pip install -r requirements.txt
-          pip install -v .
+          DISABLE_URING=1 pip install -v .
       - uses: actions/checkout@v2
         with:
           ssh-key: ${{ secrets.SSH_KEY_FOR_CI }}

colossalai/kernel/kernel_loader.py

@@ -6,7 +6,7 @@
     CpuAdamX86Extension,
     FlashAttentionDaoCudaExtension,
     FlashAttentionNpuExtension,
-    FlashAttentionXformersCudaExtension,
+    FlashAttentionSdpaCudaExtension,
     FusedOptimizerCudaExtension,
     LayerNormCudaExtension,
     MoeCudaExtension,
@@ -65,9 +65,9 @@ def load(self, ext_name: str = None):
         else:
             usable_exts = []
             for ext in exts:
-                if ext.is_hardware_available():
+                if ext.is_available():
                     # make sure the machine is compatible during kernel loading
-                    ext.assert_hardware_compatible()
+                    ext.assert_compatible()
                     usable_exts.append(ext)
 
         assert len(usable_exts) != 0, f"No usable kernel found for {self.__class__.__name__} on the current machine."
@@ -106,4 +106,20 @@ class ScaledUpperTriangleMaskedSoftmaxLoader(KernelLoader):
 
 
 class FlashAttentionLoader(KernelLoader):
-    REGISTRY = [FlashAttentionNpuExtension, FlashAttentionDaoCudaExtension, FlashAttentionXformersCudaExtension]
+    REGISTRY = [
+        FlashAttentionNpuExtension,
+        FlashAttentionDaoCudaExtension,
+        FlashAttentionSdpaCudaExtension,
+    ]
+
+
+class FlashAttentionWithPaddingMaskLoader(KernelLoader):
+    REGISTRY = [FlashAttentionNpuExtension, FlashAttentionDaoCudaExtension]
+
+
+class FlashAttentionWithCustomMaskLoader(KernelLoader):
+    REGISTRY = [FlashAttentionNpuExtension, FlashAttentionSdpaCudaExtension]
+
+
+class FlashAttentionForFloatAndCustomMaskLoader(KernelLoader):
+    REGISTRY = [FlashAttentionSdpaCudaExtension]

colossalai/shardformer/layer/__init__.py

@@ -1,3 +1,4 @@
+from .attn import AttnMaskType, ColoAttention
 from .dropout import DropoutForParallelInput, DropoutForReplicatedInput
 from .embedding import Embedding1D, VocabParallelEmbedding1D
 from .linear import Linear1D_Col, Linear1D_Row
@@ -23,4 +24,6 @@
     "FusedRMSNorm",
     "FusedLinear1D_Col",
     "ParallelModule",
+    "AttnMaskType",
+    "ColoAttention",
 ]