Merge branch 'develop_v2.4' into 'main'

[DCU] fix megatron MOE int train issues

See merge request dcutoolkit/deeplearing/TransformerEngine!36

transformer_engine/pytorch/cpp_extensions/gemm.py

@@ -206,9 +206,36 @@ def general_grouped_gemm(
     transa = layout[0] == "T"
     transb = layout[1] == "T"
 
+    empty_tensor = _empty_tensor()
+    empty_tensors = [empty_tensor] * num_gemms
+
+    # Use bfloat16 as default bias_dtype
+    gelu_input = empty_tensors
+    out_dtype = TE_DType[out[0].dtype] if D_dtype is None else D_dtype
+
+    sm_count = get_sm_count()
+    if grad and use_bias:
+        grad_bias = [
+            torch.empty(B[i].shape[1], dtype=out[0].dtype, device="cuda") for i in range(num_gemms)
+        ]
+    else:
+        grad_bias = empty_tensors
+    bias = bias if use_bias else empty_tensors
+    if use_bias:
+        bias_dtype = TE_DType[grad_bias[0].dtype] if grad else TE_DType[bias[0].dtype]
+    else:
+        bias_dtype = TE_DType[torch.bfloat16]
+
+    if gelu:
+        gelu_input = [
+            torch.empty_like(o, dtype=bias_dtype, memory_format=torch.contiguous_format)
+            for o in out
+        ]  # this should differ with respect to single output
+
     if int8_simulation_fp8 and (isinstance(A[0], Float8BlockwiseQTensorBase) or isinstance(B[0], Float8BlockwiseQTensorBase)):
         assert len(set(m_splits)) == 1, "Int8 simulation groupgemm just surpport token pad as same as batchgemm for now."
         assert not gelu, "GELU not supported with int8 simulation groupgemm."
+        assert not use_bias, "Bias not supported with int8 simulation groupgemm."
 
         if layout == "TN":
             qx_data = [
@@ -223,11 +250,11 @@ def general_grouped_gemm(
             num_gemms = len(A)
             seq_len = sum(m_splits) // num_gemms
 
-            out[0], _ = w8a8_block_int8_matmul_batched(
+            out[0] = w8a8_block_int8_matmul_batched(
                 qx_data, qw_data, ref_scales_x, ref_scales_w, out[0].view(num_gemms, seq_len, out[0].size(-1)), [128, 128],
                 output_dtype=out_dtype
             )
-            return out, None, None
+            return out, bias, gelu_input
         
         elif layout == "NN":
             qdout_data = [
@@ -242,11 +269,11 @@ def general_grouped_gemm(
             num_gemms = len(A)
             seq_len = sum(m_splits) // num_gemms
         
-            out[0], _ = w8a8_block_int8_matmul_batched(
+            out[0] = w8a8_block_int8_matmul_batched(
                 qdout_data, qw_data, ref_scales_dout, ref_scales_w, out[0].view(num_gemms, seq_len, out[0].size(-1)), [128, 128],
                 output_dtype=out_dtype
             )
-            return out, None, None
+            return out, bias, gelu_input
 
         elif layout == "NT":
             qdout_data = [
@@ -258,41 +285,15 @@ def general_grouped_gemm(
             ref_scales_dout = [b._columnwise_scale_inv for b in B]
             ref_scales_x = [a._columnwise_scale_inv for a in A]
         
-            out, _ = w8a8_block_int8_matmul_wgrad_batched(
-                qdout_data, qx_data, ref_scales_dout, ref_scales_x, out, accumulate [128, 128],
+            out = w8a8_block_int8_matmul_wgrad_batched(
+                qdout_data, qx_data, ref_scales_dout, ref_scales_x, out, accumulate, [128, 128],
                 output_dtype=out_dtype
             )
-            return out, None, None
+            return out, bias, gelu_input
         
         else:
             raise ValueError(f"Unsupported layout {layout} in int8 simulation fp8")
 
-    empty_tensor = _empty_tensor()
-    empty_tensors = [empty_tensor] * num_gemms
-
-    # Use bfloat16 as default bias_dtype
-    gelu_input = empty_tensors
-    out_dtype = TE_DType[out[0].dtype] if D_dtype is None else D_dtype
-
-    sm_count = get_sm_count()
-    if grad and use_bias:
-        grad_bias = [
-            torch.empty(B[i].shape[1], dtype=out[0].dtype, device="cuda") for i in range(num_gemms)
-        ]
-    else:
-        grad_bias = empty_tensors
-    bias = bias if use_bias else empty_tensors
-    if use_bias:
-        bias_dtype = TE_DType[grad_bias[0].dtype] if grad else TE_DType[bias[0].dtype]
-    else:
-        bias_dtype = TE_DType[torch.bfloat16]
-
-    if gelu:
-        gelu_input = [
-            torch.empty_like(o, dtype=bias_dtype, memory_format=torch.contiguous_format)
-            for o in out
-        ]  # this should differ with respect to single output
-
     bias = tex.te_general_grouped_gemm(
         A,
         transa,

transformer_engine/pytorch/triton/blockwise_int8_gemm_nt.py

@@ -524,13 +524,18 @@ def apply_w8a8_block_int8_linear_batched_helper(m: int,
     weight_b = [weight.clone().contiguous() for i in range(batch)]
     weight_scale_b = [weight_scale.clone().contiguous() for i in range(batch)]
     # print(f"zhenggf, q_input_b:{q_input_b.shape}, x_scale_b:{x_scale_b.shape}, weight_b:{weight_b.shape}, weight_scale_b:{weight_scale_b.shape}")
-    
+    N, K = weight.shape
+    C_shape = q_input.shape[:-1] + (N,)
+    output = [q_input.new_empty(C_shape, dtype=out_dtype) for i in range(batch)]
+    output = torch.stack(output).contiguous()
+
     torch_output = native_w8a8_block_int8_matmul_batched(q_input_b, weight_b, x_scale_b, weight_scale_b, block_size)
+    torch_output = torch_output.view(-1, torch_output.size(-1))
     # print(f"zhenggf, torch_output:{torch_output.shape}")
 
     x_scale_b = [xs.permute(1, 0).contiguous() for xs in x_scale_b]
     output = w8a8_block_int8_matmul_batched(
-        q_input_b, weight_b, x_scale_b, weight_scale_b, block_size,
+        q_input_b, weight_b, x_scale_b, weight_scale_b, output.view(batch, *C_shape), block_size,
         output_dtype=out_dtype,
         best_config=best_config
     )

transformer_engine/pytorch/triton/blockwise_int8_gemm_nt_wgrad.py

@@ -568,6 +568,7 @@ def apply_w8a8_block_int8_linear_batched_helper(m: int,
         output_dtype=out_dtype,
         best_config=best_config
     )
+    output = torch.stack(output).contiguous()
 
     if not torch.allclose(output, torch_output, rtol=1e-2, atol=5e-2):
         print("triton 精度检查不合格！！！")