[bc-breaking] clarify public API of float8_experimental

Summary:

Makes the following functions public:
* convert_to_float8_training and all of its configuration
* linear_requires_sync
* sync_float8_amax_and_scale_history
* precompute_float8_dynamic_scale_for_fsdp

Everything else is private. The fbsource counterpart of this PR will
remove usage of private APIs.

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

ghstack-source-id: 923035b9c859659bcb44b145d8609177abe95ce9
Pull Request resolved: #330

README.md

@@ -35,10 +35,10 @@ We provide two per-tensor scaling strategies: dynamic and delayed.  See https://
 This is the most accurate recipe as every tensor is scaled dynamically.
 
 ```python
-from float8_experimental.float8_linear_utils import (
+from float8_experimental import (
     convert_to_float8_training,
+    precompute_float8_dynamic_scale_for_fsdp,
 )
-from float8_experimental.fsdp_utils import precompute_float8_dynamic_scale_for_fsdp
 
 # create model
 m = Model(...)
@@ -82,11 +82,11 @@ for _ in range(N_ITER):
 This is theoretically the most performant recipe as it minimizes memory reads.
 
 ```python
-from float8_experimental.float8_linear_utils import (
+from float8_experimental import (
     convert_to_float8_training,
     sync_float8_amax_and_scale_history,
+    TensorScalingType,
 )
-from float8_experimental.float8_linear import TensorScalingType
 
 # create model
 m = Model(...)

float8_experimental/__init__.py

@@ -10,13 +10,18 @@
     TensorScalingType,
 )
 from float8_experimental.float8_linear import Float8Linear
-from float8_experimental.float8_linear_utils import convert_to_float8_training
+from float8_experimental.float8_linear_utils import (
+    convert_to_float8_training,
+    linear_requires_sync,
+    sync_float8_amax_and_scale_history,
+)
 from float8_experimental.float8_tensor import (
     Float8Tensor,
     GemmInputRole,
     LinearMMConfig,
     ScaledMMConfig,
 )
+from float8_experimental.fsdp_utils import precompute_float8_dynamic_scale_for_fsdp
 
 # Needed to load Float8Tensor with weights_only = True
 from torch.serialization import add_safe_globals
@@ -30,7 +35,8 @@
     "Float8TensorCastConfig",
     # top level UX
     "convert_to_float8_training",
-    # TODO(future): remove Float8Tensor and Float8Linear from public API
-    "Float8Tensor",
-    "Float8Linear",
+    "linear_requires_sync",
+    "sync_float8_amax_and_scale_history",
+    "precompute_float8_dynamic_scale_for_fsdp",
+    # note: Float8Tensor and Float8Linear are not public APIs
 ]

float8_experimental/float8_linear.py

@@ -150,6 +150,10 @@ def __init__(self, history_len: int = 16, scale_fn_name: str = "max"):
 
 class Float8Linear(torch.nn.Linear):
     """
+    Note: this is **not** a public API and is only intended to be used
+    inside of this repository. Please file an issue if you would benefit
+    from this being a public API.
+
     A wrapper around a `torch.nn.Linear` module which does fp8 compute, and tracks
     scales in way friendly to delayed scaling.
     """

float8_experimental/float8_tensor.py

@@ -252,6 +252,10 @@ def backward(ctx, g):
 
 class Float8Tensor(torch.Tensor):
     """
+    Note: this is **not** a public API and is only intended to be used
+    inside of this repository. Please file an issue if you would benefit
+    from this being a public API.
+
     A Python-only Float8 tensor subclass.  Contains:
     * `_data`: the underlying e4m3 or e5m2 data
     * `_scale`: the scale used to scale the original fp32 tensor. We multiply