[PyTorch] Mcore DDP support (#446)

Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

tests/pytorch/test_sanity.py

@@ -226,9 +226,7 @@ def _test_sanity_e2e_gradient_accumulation_fusion(block, bs, dtype, config, fp8_
         if "layer_norm_weight" in name:
             continue
         elif "weight" in name and p.requires_grad:
-            assert (
-                p.grad is None and torch.count_nonzero(p.main_grad) > 0
-            ), "Gradient not accumulated."
+            assert torch.count_nonzero(p.main_grad) > 0, "Gradient not accumulated."
 
 
 def _test_sanity_e2e(block, bs, dtype, config, fp8_recipe, skip_wgrad):

transformer_engine/pytorch/cpp_extensions/gemm.py

@@ -45,7 +45,6 @@ def fp8_gemm(
     assert_dim_for_fp8_exec(A)
     assert_dim_for_fp8_exec(B)
 
-    return_output = False
     if out is None:
         out = torch.empty(
             B.shape[0],
@@ -53,7 +52,7 @@ def fp8_gemm(
             dtype=out_dtype,
             device="cuda",
         )
-        return_output = True
+
     # Use bfloat16 as default bias_dtype
     bias_dtype = torch.bfloat16 if bias is None else bias.dtype
     if gelu:
@@ -110,13 +109,7 @@ def fp8_gemm(
             args = tuple(args + (True, extra_output_tensor,))
     _ = fn(*args)
 
-    if return_output:
-        if gelu:
-            return out, gelu_input
-        return out
-    if gelu:
-        return gelu_input
-    return None
+    return out, gelu_input
 
 
 def gemm(
@@ -144,7 +137,6 @@ def gemm(
     empty_tensor = torch.Tensor()
     fp8_index = -1 # dummy index
 
-    return_output = False
     if out is None:
         out = torch.empty(
             B.shape[1] if transb else B.shape[0],
@@ -152,7 +144,6 @@ def gemm(
             dtype=dtype,
             device="cuda",
         )
-        return_output = True
 
     if gelu and not grad:
         gelu_input = torch.empty_like(out, dtype=dtype)
@@ -222,6 +213,4 @@ def gemm(
             args = tuple(args + (False, extra_output_tensor,))
     _ = fn(*args)
 
-    if return_output:
-        return out, grad_bias, gelu_input
-    return None, grad_bias, gelu_input
+    return out, grad_bias, gelu_input

transformer_engine/pytorch/module/layernorm_linear.py

@@ -173,7 +173,7 @@ class _LayerNormLinear(torch.autograd.Function):
                         tex.FP8FwdTensors.GEMM1_WEIGHT,
                         fp8_dtype_forward)
 
-            out = tex.fp8_gemm(
+            out, _ = tex.fp8_gemm(
                 weight_fp8,
                 fp8_meta["scaling_fwd"].scale_inv,
                 tex.FP8FwdTensors.GEMM1_WEIGHT,
@@ -389,7 +389,7 @@ class _LayerNormLinear(torch.autograd.Function):
                     # WGRAD
                     if not ctx.fp8_meta["recipe"].override_linear_precision.wgrad:
                         ln_out_total_t = tex.fp8_transpose(ln_out_total, fp8_dtype_forward)
-                        wgrad = tex.fp8_gemm(
+                        wgrad, _ = tex.fp8_gemm(
                             ln_out_total_t,
                             fwd_scale_inverses,
                             tex.FP8FwdTensors.GEMM1_INPUT,
@@ -444,7 +444,6 @@ class _LayerNormLinear(torch.autograd.Function):
                         ub=ub_obj_dgrad if ctx.ub_bulk_wgrad else None
                     )
 
-
             if ctx.ub_bulk_wgrad:
                 dgrad = ub_obj_dgrad.get_ubuf_output(0) # Reduce-scatter output
             # Column Parallel Linear
@@ -474,6 +473,9 @@ class _LayerNormLinear(torch.autograd.Function):
             if not ctx.use_bias:
                 grad_bias = None
 
+        # Handle custom DDP from mcore.
+        weight.grad_added_to_main_grad = ctx.fuse_wgrad_accumulation
+
         return (
             dxmat.view(ctx.inp_shape) if ctx.requires_dgrad else None,
             dgamma,

transformer_engine/pytorch/module/layernorm_mlp.py

@@ -223,7 +223,7 @@ class _LayerNormMLP(torch.autograd.Function):
                         fp8_dtype_forward,
                     )
 
-            fc1_out = tex.fp8_gemm(
+            fc1_out, _ = tex.fp8_gemm(
                 fc1_weight_fp8,
                 fp8_meta["scaling_fwd"].scale_inv,
                 tex.FP8FwdTensors.GEMM1_WEIGHT,
@@ -344,7 +344,7 @@ class _LayerNormMLP(torch.autograd.Function):
                 dim_size = list(gelu_out.size())
                 dim_size[1] = fc2_weight.size(0)
                 fc2_out = torch.empty(dim_size, dtype=activation_dtype, device=gelu_out.device)
-            _, _, _ = tex.gemm(
+            _ = tex.gemm(
                 fc2_weight,
                 gelu_out,
                 activation_dtype,
@@ -498,7 +498,7 @@ class _LayerNormMLP(torch.autograd.Function):
                 )
 
                 # FC2 DGRAD; Unconditional
-                fc2_dgrad = tex.fp8_gemm(
+                fc2_dgrad, _ = tex.fp8_gemm(
                     fc2_weight_t_fp8,
                     fwd_scale_inverses,
                     tex.FP8FwdTensors.GEMM2_WEIGHT,
@@ -519,7 +519,7 @@ class _LayerNormMLP(torch.autograd.Function):
                 if not ctx.fp8_meta["recipe"].override_linear_precision.wgrad:
                     if fc2_weight.requires_grad:
                         gelu_out_t = tex.fp8_transpose(gelu_out, fp8_dtype_forward)
-                        fc2_wgrad = tex.fp8_gemm(
+                        fc2_wgrad, _ = tex.fp8_gemm(
                             gelu_out_t,
                             fwd_scale_inverses,
                             tex.FP8FwdTensors.GEMM2_INPUT,
@@ -675,7 +675,7 @@ class _LayerNormMLP(torch.autograd.Function):
                         fc1_dgrad_size, dtype=ctx.activation_dtype, device=fc1_weight.device
                     )
                 # FC1 DGRAD: Unconditional
-                _, _, _ = tex.gemm(
+                _ = tex.gemm(
                     fc1_weight,
                     dgelu,
                     ctx.activation_dtype,
@@ -705,7 +705,7 @@ class _LayerNormMLP(torch.autograd.Function):
                     # FC1 WGRAD
                     if not ctx.fp8_meta["recipe"].override_linear_precision.wgrad:
                         ln_out_total_t = tex.fp8_transpose(ln_out_total, fp8_dtype_forward)
-                        fc1_wgrad = tex.fp8_gemm(
+                        fc1_wgrad, _ = tex.fp8_gemm(
                             ln_out_total_t,
                             fwd_scale_inverses,
                             tex.FP8FwdTensors.GEMM1_INPUT,
@@ -794,6 +794,10 @@ class _LayerNormMLP(torch.autograd.Function):
                 )
                 dbeta = None
 
+        # Handle custom DDP from mcore.
+        fc1_weight.grad_added_to_main_grad = ctx.fuse_wgrad_accumulation
+        fc2_weight.grad_added_to_main_grad = ctx.fuse_wgrad_accumulation
+
         return (
             dxmat.view(ctx.inp_shape) if ctx.requires_dgrad else None,
             dgamma,

transformer_engine/pytorch/module/linear.py

@@ -211,7 +211,7 @@ class _Linear(torch.autograd.Function):
                 dim_size[1] = weight.size(0)
                 out = torch.empty(dim_size, dtype=activation_dtype, device=inputmat_total.device)
 
-            _, _, _ = gemm(
+            _ = gemm(
                 weight,
                 inputmat_total,
                 activation_dtype,
@@ -325,7 +325,7 @@ class _Linear(torch.autograd.Function):
 
             if ctx.requires_dgrad:
                 if ctx.fp8:
-                    dgrad = fp8_gemm(
+                    dgrad, _ = fp8_gemm(
                         weight_t_fp8,
                         fwd_scale_inverses,
                         tex.FP8FwdTensors.GEMM1_WEIGHT,
@@ -368,7 +368,7 @@ class _Linear(torch.autograd.Function):
                     if not ctx.fp8_meta["recipe"].override_linear_precision.wgrad:
                         if ctx.ub_split_ag:
                             grad_output_t = tex.fp8_transpose(grad_output_c, fp8_dtype_backward)
-                        wgrad = fp8_gemm(
+                        wgrad, _ = fp8_gemm(
                             inputmat_t_total,
                             fwd_scale_inverses,
                             tex.FP8FwdTensors.GEMM1_INPUT,
@@ -415,6 +415,9 @@ class _Linear(torch.autograd.Function):
             if not ctx.use_bias:
                 grad_bias = None
 
+        # Handle custom DDP from mcore.
+        weight.grad_added_to_main_grad = ctx.fuse_wgrad_accumulation
+
         return (
             wgrad if weight.requires_grad else None,
             None,