FP8 Support for MCore MoE (NVIDIA#648)

* Add support for MoE with FP8.

Signed-off-by: Dennis Liu <[email protected]>

* Fix unittest.

Signed-off-by: Dennis Liu <[email protected]>

* Fix error in linear backward.

Signed-off-by: Dennis Liu <[email protected]>

---------

Signed-off-by: Dennis Liu <[email protected]>
Co-authored-by: Przemyslaw Tredak <[email protected]>
Signed-off-by: Pawel Gadzinski <[email protected]>

tests/pytorch/test_sanity.py

@@ -9,7 +9,11 @@
 import torch
 import pytest
 
-from transformer_engine.pytorch.fp8 import fp8_autocast, FP8GlobalStateManager
+from transformer_engine.pytorch.fp8 import (
+    fp8_autocast,
+    FP8GlobalStateManager,
+    fp8_model_init,
+)
 from transformer_engine.pytorch.utils import (
     init_method_normal,
     scaled_init_method_normal,
@@ -107,6 +111,7 @@ def is_fp8_supported(self):
     param_types.append(torch.bfloat16)
 
 all_boolean = [True, False]
+batch_sizes_with_zero = [0, 1, 2]
 
 all_activations = ["gelu", "relu", "reglu", "geglu", "swiglu"]
 all_normalizations = ["LayerNorm", "RMSNorm"]
@@ -456,6 +461,45 @@ def test_sanity_linear(dtype, fp8_recipe, model, skip_wgrad, skip_dgrad):
     _test_sanity_common(block, dtype, config, fp8_recipe, skip_wgrad, skip_dgrad)
 
 
+@pytest.mark.parametrize("dtype", param_types)
+@pytest.mark.parametrize("bs", batch_sizes_with_zero)
+@pytest.mark.parametrize("model", ["small", "weird"])
+@pytest.mark.parametrize("fp8_recipe", fp8_recipes)
+@pytest.mark.parametrize("fp8_model_params", all_boolean)
+@pytest.mark.parametrize("use_bias", all_boolean)
+def test_sanity_linear_with_zero_tokens(dtype, bs, model, fp8_recipe, fp8_model_params, use_bias):
+    config = model_configs[model]
+    ffn_hidden_size = 4 * config.hidden_size
+    num_tokens = bs*config.seq_len
+
+    if fp8_recipe is not None:
+        if not fp8_available:
+            pytest.skip(reason_for_no_fp8)
+        if not config.is_fp8_supported():
+            pytest.skip("Model config does not support FP8")
+
+    use_fp8 = fp8_recipe is not None
+    with fp8_model_init(enabled=use_fp8 and fp8_model_params):
+        te_linear = (
+            Linear(
+                config.hidden_size,
+                ffn_hidden_size,
+                bias=use_bias,
+                params_dtype=dtype
+            )
+            .cuda()
+        )
+
+    inp_hidden_states = torch.randn(
+        num_tokens, config.hidden_size, dtype=dtype, requires_grad=True
+    ).cuda()
+    with fp8_autocast(enabled=use_fp8, fp8_recipe=fp8_recipe):
+        out = te_linear(inp_hidden_states)
+    loss = out.sum()
+    loss.backward()
+    assert out.shape == (num_tokens, ffn_hidden_size)
+
+
 @pytest.mark.parametrize("dtype", param_types)
 @pytest.mark.parametrize("fp8_recipe", fp8_recipes)
 @pytest.mark.parametrize("model", ["small", "weird"])

transformer_engine/pytorch/cpp_extensions/cast.py

@@ -22,16 +22,18 @@ def cast_to_fp8(
     """Cast input to FP8"""
 
     if out is not None:
-        torch.ops.tex_ts.cast_to_fp8_noalloc_ts(
-            inp,
-            fp8_meta_tensor.scale,
-            out,
-            fp8_meta_tensor.amax_history,
-            fp8_meta_tensor.scale_inv,
-            fp8_tensor,
-            otype
-        )
+        if inp.nelement() > 0:
+            torch.ops.tex_ts.cast_to_fp8_noalloc_ts(
+                inp,
+                fp8_meta_tensor.scale,
+                out,
+                fp8_meta_tensor.amax_history,
+                fp8_meta_tensor.scale_inv,
+                fp8_tensor,
+                otype
+            )
         return None
+
     return torch.ops.tex_ts.cast_to_fp8_ts(
         inp,
         fp8_meta_tensor.scale,
@@ -41,7 +43,6 @@ def cast_to_fp8(
         otype,
     )
 
-
 def cast_from_fp8(
     inp: torch.Tensor,
     fp8_meta_tensor: tex.FP8TensorMeta,

transformer_engine/pytorch/cpp_extensions/gemm.py

@@ -64,6 +64,8 @@ def fp8_gemm(
     bias_dtype = TE_DType[bias_dtype]
 
     out_dtype = TE_DType[out.dtype] if D_dtype is None else D_dtype
+    if A.nelement() == 0 or B.nelement() == 0:
+        return out, gelu_input
 
     args = (
         A,
@@ -191,6 +193,8 @@ def gemm(
         grad_bias = empty_tensor
 
     bias = bias if use_bias else empty_tensor
+    if A.nelement() == 0 or B.nelement() == 0:
+        return out, grad_bias, gelu_input
 
     assert A.dtype == dtype and B.dtype == dtype, \
         f'Expected dtype={dtype}, but found A.dtype={A.dtype} and B.dtype={B.dtype}'

transformer_engine/pytorch/cpp_extensions/transpose.py

@@ -39,16 +39,17 @@ def fp8_cast_transpose_fused(
     if noop_flag is None:
         noop_flag = torch.Tensor()
 
-    tex.fused_cast_transpose_noop(
-        inp,
-        noop_flag,
-        fp8_meta_tensor.scale[fp8_tensor],
-        fp8_meta_tensor.amax_history[0][fp8_tensor],
-        fp8_meta_tensor.scale_inv[fp8_tensor],
-        cast_out,
-        transpose_out,
-        otype,
-    )
+    if inp.nelement() > 0:
+        tex.fused_cast_transpose_noop(
+            inp,
+            noop_flag,
+            fp8_meta_tensor.scale[fp8_tensor],
+            fp8_meta_tensor.amax_history[0][fp8_tensor],
+            fp8_meta_tensor.scale_inv[fp8_tensor],
+            cast_out,
+            transpose_out,
+            otype,
+        )
 
     if return_outputs:
         return cast_out, transpose_out

transformer_engine/pytorch/csrc/extensions/cast.cu

@@ -19,6 +19,9 @@ at::Tensor cast_to_fp8(const at::Tensor &input,
 
     auto output = at::empty_like(input, at::CUDA(GetATenDType(otype)));
 
+    if (input.numel() == 0)
+        return output;
+
     auto input_cu     = makeTransformerEngineTensor(input);
     auto output_cu    = makeTransformerEngineTensor(output.data_ptr(), shape, otype,
                                                     amax.data_ptr(), scale.data_ptr(),

transformer_engine/pytorch/csrc/extensions/transpose.cu

@@ -83,6 +83,9 @@ std::vector<at::Tensor> fused_cast_transpose_bgrad(at::Tensor grad_output,
                                 grad_output.size(0),
                                 DType::kByte);
 
+  if (M == 0 || N == 0)
+    return {grad_bias, grad_output_cast, grad_output_transpose};
+
   auto input_cu             = makeTransformerEngineTensor(grad_output);
   auto cast_output_cu       = makeTransformerEngineTensor(grad_output_cast.data_ptr(), {M, N},
                                                           otype, amax.data_ptr(), scale.data_ptr(),
@@ -335,6 +338,8 @@ at::Tensor fp8_transpose(at::Tensor input,
 
   size_t M = static_cast<size_t>(input.size(0));
   size_t N = static_cast<size_t>(input.size(1));
+  if (M == 0 || N == 0)
+    return input;
 
   auto output =
             allocateTorchTensor(input.size(1),

transformer_engine/pytorch/module/base.py

@@ -825,8 +825,12 @@ def reset_parameters(self, defer_init: Optional[bool] = False) -> None:
             if get_rng_state_tracker is None:
                 init_fn(param)
             else:
-                with get_rng_state_tracker().fork():
-                    init_fn(param)
+                if hasattr(self, "rng_tracker_name") and self.rng_tracker_name:
+                    with get_rng_state_tracker().fork(self.rng_tracker_name):
+                        init_fn(param)
+                else:
+                    with get_rng_state_tracker().fork():
+                        init_fn(param)
 
             # If primary weights are in fp8, wrap the parameter as Float8Tensor
             fp8_meta_index = self.param_init_meta[name].fp8_meta_index

transformer_engine/pytorch/module/linear.py

@@ -152,7 +152,6 @@ def forward(
             inputmat_total, _ = gather_along_first_dim(inputmat, tp_group)
         else:
             inputmat_total = inputmat
-
         if fp8:
             if _NVTE_DEBUG:
                 print('[Linear]: using FP8 forward')
@@ -664,6 +663,10 @@ class Linear(TransformerEngineBaseModule):
     init_method : Callable, default = `None`
                  used for initializing weights in the following way: `init_method(weight)`.
                  When set to `None`, defaults to `torch.nn.init.normal_(mean=0.0, std=0.023)`.
+    get_rng_state_tracker : Callable, default = `None`
+                 used to get the random number generator state tracker for initilizeing weights.
+    rng_tracker_name : str, default = `None`
+                 the param passed to get_rng_state_tracker to get the specific rng tracker.
     parameters_split : Optional[Union[Tuple[str, ...], Dict[str, int]]], default = None
                       Configuration for splitting the weight and bias tensors along dim 0 into
                       multiple PyTorch parameters. If a list or tuple of strings is provided,
@@ -723,6 +726,7 @@ def __init__(
         tp_group: Optional[dist_group_type] = None,
         tp_size: int = 1,
         get_rng_state_tracker: Optional[Callable] = None,
+        rng_tracker_name: Optional[str] = None,
         init_method: Optional[Callable] = None,
         bias: bool = True,
         return_bias: bool = False,
@@ -753,6 +757,8 @@ def __init__(
             ), "Userbuffer communication backend not available."
         self.ub_name = ub_name
         self.get_rng_state_tracker = get_rng_state_tracker
+        self.rng_tracker_name = rng_tracker_name
+
         if device == 'meta':
             assert parameters_split is None, ("Cannot split module parameters "
                                               "on 'meta' device.")