Merge branch 'develop_v2.4' into 'main'

[DCU] fix blockwise int8 train issues in megatron

See merge request dcutoolkit/deeplearing/TransformerEngine!30

transformer_engine/pytorch/cpp_extensions/gemm.py

@@ -82,8 +82,6 @@ def general_gemm(
             if accumulate:
                 assert out is not None
                 y = y + out
-            else:
-                assert out is None, "Output tensor should be None when accumulate is False."
             return y, None, None, None
         
         elif layout == "NN":
@@ -103,8 +101,6 @@ def general_gemm(
             if accumulate:
                 assert out is not None
                 y = y + out
-            else:
-                assert out is None, "Output tensor should be None when accumulate is False."
             return y, None, None, None
 
         elif layout == "NT":
@@ -124,8 +120,6 @@ def general_gemm(
             if accumulate:
                 assert out is not None
                 y = y + out
-            else:
-                assert out is None, "Output tensor should be None when accumulate is False."
             return y, None, None, None
         
         else:
@@ -234,9 +228,8 @@ def general_grouped_gemm(
             )
             if accumulate:
                 assert out is not None
+                out = torch.stack(out).contiguous()
                 y = y + out
-            else:
-                assert out is None, "Output tensor should be None when accumulate is False."
             return y, None, None
         
         elif layout == "NN":
@@ -255,9 +248,8 @@ def general_grouped_gemm(
             )
             if accumulate:
                 assert out is not None
+                out = torch.stack(out).contiguous()
                 y = y + out
-            else:
-                assert out is None, "Output tensor should be None when accumulate is False."
             return y, None, None
 
         elif layout == "NT":
@@ -276,9 +268,8 @@ def general_grouped_gemm(
             )
             if accumulate:
                 assert out is not None
+                out = torch.stack(out).contiguous()
                 y = y + out
-            else:
-                assert out is None, "Output tensor should be None when accumulate is False."
             return y, None, None
         
         else:

transformer_engine/pytorch/triton/blockwise_int8_gemm_nt_wgrad.py

@@ -324,6 +324,8 @@ def w8a8_block_int8_matmul_wgrad(
     """
     assert len(block_size) == 2
     block_n, block_k = block_size[0], block_size[1]
+    B = B.view(B.size(0), -1)
+    assert A.ndim == 2
 
     assert A.shape[-1] == B.shape[-1]
     # print(f"A.shape[:-1] : {A.shape[:-1]}, As.shape[:-1]: {As.shape[:-1]}")
@@ -451,6 +453,9 @@ def w8a8_block_int8_matmul_wgrad_batched(
     As = torch.stack(As_list).contiguous() 
     Bs = torch.stack(Bs_list).contiguous()  
 
+    B_new_shape = B.size()[:2] + (-1,)  
+    B = B.view(*B_new_shape)
+    assert A.ndim == 3
     assert A.shape[-1] == B.shape[-1]
 
     M = A.numel() // A.shape[-1] // A.shape[0]