fix float8 moe training dtype bug

Author: danielvegamyhre

test/prototype/moe_training/test_fsdp.sh

@@ -0,0 +1 @@
+torchrun --nproc_per_node=2 --local-ranks-filter=0 -m pytest test/prototype/moe_training/test_fsdp.py

torchao/prototype/moe_training/conversion_utils.py

@@ -1,3 +1,9 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# 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 Callable, Optional
 
 from torch import nn
@@ -8,6 +14,8 @@
     register_quantize_module_handler,
 )
 
+logger: logging.Logger = logging.getLogger(__name__)
+
 
 class MoETrainingConfig(AOBaseConfig):
     """
@@ -76,7 +84,7 @@ def _swap_params(
                 f"Does not support a root nn.Parameter with children: {module}"
             )
         if not isinstance(module.data, ScaledGroupedMMTensor):
-            new_data = ScaledGroupedMMTensor(module.data)
+            new_data = ScaledGroupedMMTensor(module.data, module.data.dtype)
             return nn.Parameter(new_data, requires_grad=module.requires_grad)
         return module
 
@@ -102,10 +110,13 @@ def post_order_traversal(
             for param_name, param in module.named_parameters(recurse=False):
                 if not isinstance(param.data, ScaledGroupedMMTensor):
                     new_param = nn.Parameter(
-                        ScaledGroupedMMTensor(param), requires_grad=param.requires_grad
+                        ScaledGroupedMMTensor(param.data, param.data.dtype),
+                        requires_grad=param.requires_grad,
                     )
                     setattr(module, param_name, new_param)
-                    print(f"Swapped {cur_fqn}.{param_name} to ScaledGroupedMMTensor")
+                    logger.info(
+                        f"Swapped {cur_fqn}.{param_name} to ScaledGroupedMMTensor"
+                    )
 
     post_order_traversal(root_module)
     return root_module

torchao/prototype/moe_training/tensor.py

@@ -1,3 +1,10 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# 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 Any, Optional, Tuple
 
 import torch
@@ -6,6 +13,9 @@
 
 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,
@@ -34,14 +44,15 @@ class ScaledGroupedMMTensor(torch.Tensor):
     def __new__(
         cls,
         tensor: torch.Tensor,
+        dtype: torch.dtype,
     ):
         return torch.Tensor._make_wrapper_subclass(
             cls,
             tensor.size(),
             strides=tensor.stride(),
             storage_offset=tensor.storage_offset(),
             memory_format=suggest_memory_format(tensor),
-            dtype=tensor.dtype,
+            dtype=dtype,
             layout=tensor.layout,
             device=tensor.device,
             pin_memory=tensor.is_pinned(),
@@ -51,8 +62,10 @@ def __new__(
     def __init__(
         self,
         tensor: torch.Tensor,
+        dtype: torch.dtype,
     ):
         self._data = tensor
+        self._dtype = dtype
 
     @classmethod
     def __torch_function__(cls, func, types, args, kwargs={}):
@@ -84,10 +97,19 @@ def __torch_function__(cls, func, types, args, kwargs={}):
     def __torch_dispatch__(cls, func, types, args, kwargs={}):
         # detach is special case
         if func == torch.ops.aten.detach.default:
-            return ScaledGroupedMMTensor(args[0]._data)
+            return ScaledGroupedMMTensor(args[0]._data, args[0]._dtype)
 
         # unwrap args and kwargs
-        unwrap = lambda tensor: tensor._data
+        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
+
         args, kwargs = pytree.tree_map_only(
             ScaledGroupedMMTensor, unwrap, (args, kwargs or {})
         )
@@ -102,12 +124,27 @@ def __torch_dispatch__(cls, func, types, args, kwargs={}):
         # wrap outputs back into ScaledGroupedMMTensor for ops that do preserve subclass
         return pytree.tree_map_only(
             torch.Tensor,
-            lambda x: ScaledGroupedMMTensor(x),
+            lambda x: ScaledGroupedMMTensor(x, dtype),
             out,
         )
 
+    def __repr__(self):
+        return f"ScaledGroupedMMTensor(data={self._data}, dtype={self._dtype})"
+
+    def __tensor_flatten__(self):
+        return ["_data"], {"dtype": self._dtype}
+
+    @staticmethod
+    def __tensor_unflatten__(inner_tensors, flatten_spec, outer_size, outer_stride):
+        return ScaledGroupedMMTensor(
+            inner_tensors["_data"],
+            flatten_spec["dtype"],
+        )
+
     def fsdp_pre_all_gather(self, mesh):
-        return (self._data,), ()
+        all_gather_inputs = (self._data,)
+        all_gather_metadata = ()
+        return all_gather_inputs, all_gather_metadata
 
     def fsdp_post_all_gather(
         self,
@@ -118,6 +155,6 @@ def fsdp_post_all_gather(
         out: Optional[torch.Tensor] = None,
     ):
         (data,) = all_gather_outputs
-        return ScaledGroupedMMTensor(
-            data,
-        ), (data,)
+        out = ScaledGroupedMMTensor(data, param_dtype)
+        inner_tensors = (data,)
+        return out, inner_tensors