[moe training] Cast to mixed precision policy param dtype in fsdp_pre_all_gather hook

torchao/prototype/moe_training/scaled_grouped_mm.py

@@ -4,6 +4,7 @@
 # This source code is licensed under the BSD 3-Clause license found in the
 # LICENSE file in the root directory of this source tree.
 
+import logging
 from typing import Optional
 
 import torch
@@ -18,6 +19,8 @@
     _is_column_major,
 )
 
+logger: logging.Logger = logging.getLogger(__name__)
+
 
 def _scaled_grouped_mm(
     A: torch.Tensor,
@@ -36,8 +39,8 @@ def _scaled_grouped_mm(
             and in column-major memory layout.
         offs (int32 torch.Tensor): The offsets to use to mark the starting index of each group along dim0 of the A tensor.
         out_dtype (Optional[torch.dtype]): The dtype of the output tensor. Currently only torch.bfloat16 is supported.
-        use_triton_for_per_group_scales (bool): Whether to use custom triton kernels to compute per-group scales. Default is True.
     """
+    logger.debug("Using scaled_grouped_mm")
     return _Float8GroupedMM.apply(
         A,
         B_t,

torchao/prototype/moe_training/tensor.py

@@ -9,13 +9,15 @@
 
 import torch
 import torch.utils._pytree as pytree
+from torch import nn
 from torch._prims_common import suggest_memory_format
+from torch.distributed.device_mesh import DeviceMesh
+from torch.distributed.fsdp import MixedPrecisionPolicy
 
 from torchao.prototype.moe_training import _scaled_grouped_mm
 
 logger: logging.Logger = logging.getLogger(__name__)
 
-
 _ops_to_preserve_subclass = {
     torch.ops.aten.empty_like.default,
     torch.ops.aten.new_zeros.default,
@@ -64,12 +66,14 @@ def __init__(
         tensor: torch.Tensor,
         dtype: torch.dtype,
     ):
-        self._data = tensor
+        self._data = tensor.to(dtype)
         self._dtype = dtype
 
     @classmethod
     def __torch_function__(cls, func, types, args, kwargs={}):
-        logger.info(f"{func.__name__}, args: {args}, kwargs: {kwargs}")
+        logger.debug(
+            f"ScaledGroupedMMTensor func: {func.__name__}, args: {args}, kwargs: {kwargs}"
+        )
         # override the grouped mm op to use the differentiable _scaled_grouped_mm
         if func.__name__ == cls.grouped_mm_func_name:
             # Use torchao scaled grouped mm with dynamic quant for
@@ -100,17 +104,8 @@ def __torch_dispatch__(cls, func, types, args, kwargs={}):
         if func == torch.ops.aten.detach.default:
             return ScaledGroupedMMTensor(args[0]._data, args[0]._dtype)
 
-        # unwrap args and kwargs
-        dtype: Optional[torch.dtype] = None
-
-        def unwrap(t):
-            nonlocal dtype
-            if dtype is None:
-                dtype = t._dtype
-            else:
-                assert t._dtype == dtype
-            return t._data
-
+        # unwrap args/kwargs
+        unwrap = lambda x: x._data if isinstance(x, ScaledGroupedMMTensor) else x
         args, kwargs = pytree.tree_map_only(
             ScaledGroupedMMTensor, unwrap, (args, kwargs or {})
         )
@@ -125,7 +120,7 @@ def unwrap(t):
         # wrap outputs back into ScaledGroupedMMTensor for ops that do preserve subclass
         return pytree.tree_map_only(
             torch.Tensor,
-            lambda x: ScaledGroupedMMTensor(x, dtype),
+            lambda x: ScaledGroupedMMTensor(x, x.dtype),
             out,
         )
 
@@ -142,9 +137,20 @@ def __tensor_unflatten__(inner_tensors, flatten_spec, outer_size, outer_stride):
             flatten_spec["_dtype"],
         )
 
-    def fsdp_pre_all_gather(self, mesh):
+    # fsdp hooks based on https://github.com/pytorch/pytorch/blob/20e40492b046b9287726d3ec656117e4dc38f0e2/test/distributed/_composable/fsdp/test_fully_shard_extensions.py#L81
+    def fsdp_pre_all_gather(
+        self,
+        mesh: DeviceMesh,
+        outer_size: torch.Size,
+        outer_stride: tuple[int, ...],
+        module: nn.Module,
+        mp_policy: MixedPrecisionPolicy,
+    ):
         all_gather_inputs = (self._data,)
         all_gather_metadata = ()
+        logger.debug(
+            f"ScaledGroupedMMTensor fsdp_pre_all_gather: self._data.dtype={self._data.dtype}, param_dtype: {mp_policy.param_dtype}"
+        )
         return all_gather_inputs, all_gather_metadata
 
     def fsdp_post_all_gather(
@@ -156,6 +162,13 @@ def fsdp_post_all_gather(
         out: Optional[torch.Tensor] = None,
     ):
         (data,) = all_gather_outputs
+        logger.debug(
+            f"ScaledGroupedMMTensor fsdp_post_all_gather: data.dtype={data.dtype}, param_dtype: {param_dtype}"
+        )
+
+        if out is not None:
+            return
+
         output = ScaledGroupedMMTensor(data, param_dtype)
         inner_tensors = (data,)
         return output, inner_tensors