Enable restricted split + cat in order to enable SP (#253)

Summary:
This comes from needing to support sequence parallelism in torchtitan

Pull Request resolved: #253

Reviewed By: wanchaol

Differential Revision: D57134004

Pulled By: drisspg

fbshipit-source-id: e6c67ba7b2b96045867ece467400b0e4a3305e1d

Author: drisspg

float8_experimental/float8_ops.py

@@ -3,7 +3,7 @@
 #
 # This source code is licensed under the BSD 3-Clause license found in the
 # LICENSE file in the root directory of this source tree.
-from typing import Any, Dict
+from typing import Any, Dict, Tuple
 
 import torch
 
@@ -50,6 +50,52 @@ def float8_desugar_op(aten_op, args, kwargs=None):
     )
 
 
+@implements([aten.split.Tensor])
+def float8_split(aten_op, args, kwargs=None):
+    new_data_tensors = aten_op(args[0]._data, *args[1:], **kwargs)
+
+    def make_float8(data):
+        return Float8Tensor(
+            data, args[0]._scale, args[0]._orig_dtype, args[0]._mm_config
+        )
+
+    out = map(make_float8, new_data_tensors)
+    return list(out)
+
+
+# Errors cant `cat_cuda float8 e4m3fn`
+@implements([aten.cat.default])
+def float8_cat(aten_op, args, kwargs=None):
+    chunked_tensors: Tuple[Float8Tensor] = args[0]
+
+    orig_dtype = chunked_tensors[0]._orig_dtype
+    scale = chunked_tensors[0]._scale
+    mm_config = chunked_tensors[0]._mm_config
+    fp8_dtype = chunked_tensors[0]._data.dtype
+    chunk_data = []
+    for chunk in chunked_tensors:
+        assert isinstance(
+            chunk, Float8Tensor
+        ), "Expecting all chunks to be of type Float8Tensor"
+        assert (
+            chunk._orig_dtype == orig_dtype
+        ), "Expecting all chunks to be of the same dtype"
+        assert (
+            chunk._scale is scale
+        ), "Expecting all chunks to have thee same scale as a result of a split"
+        assert (
+            chunk._mm_config is mm_config
+        ), "Expecting all chunks to have thee same mm config as a result of a split"
+        assert (
+            chunk._data.dtype == fp8_dtype
+        ), "Expecting all chunks to be of the same dtype as a result of a split"
+        chunk_data.append(chunk._data.view(torch.uint8))
+
+    new_data = aten_op(chunk_data, *args[1:], **kwargs)
+    new_data = new_data.view(fp8_dtype)
+    return Float8Tensor(new_data, scale, orig_dtype, mm_config)
+
+
 @implements([aten.sum.dim_IntList])
 def float8_cast_up_op(aten_op, args, kwargs=None):
     """Be careful with this function, this is a "fallback" op that

pyproject.toml

@@ -29,7 +29,7 @@ dev = [
     "black==23.3.0",
     "usort==1.0.6",
     "ufmt==2.1.0",
-    "libcst==1.0.1",
+    "libcst==1.1.0",
     "pytest==7.4.0",
     "bumpver",
     "pip-tools",

test/test_base.py

@@ -44,6 +44,12 @@
 is_H100 = torch.cuda.is_available() and torch.cuda.get_device_capability() >= (9, 0)
 
 
+def bitwise_identical(a: Float8Tensor, b: Float8Tensor) -> bool:
+    assert torch.all(a._data == b._data).item(), "scales are not identical"
+    assert torch.all(a._data == b._data).item(), "data is not identical"
+    return True
+
+
 class TestFloat8Tensor(unittest.TestCase):
     def test_preserves_dtype(self) -> None:
         # hp means high precision, lp means low precision
@@ -68,6 +74,15 @@ def test_differentiable_casts(self) -> None:
             # the gradient should be unchanged through both casts
             torch.testing.assert_close(grad, x.grad, rtol=0, atol=0)
 
+    def test_split_cat(self):
+        a = torch.rand(16, 16, dtype=torch.bfloat16)
+        scale = tensor_to_scale(a, torch.float8_e4m3fn)
+        fp8_a = Float8Tensor.to_float8(a, scale, torch.float8_e4m3fn)
+
+        splits = torch.split(fp8_a, 16)
+        catted = torch.cat(splits, dim=0)
+        assert bitwise_identical(fp8_a, catted)
+
 
 class TestFloat8Linear:
     def _test_linear_impl(
@@ -357,7 +372,7 @@ def test_different_configs_error(self):
         ):
             a @ b
 
-    def test_merge_configs(sel):
+    def test_merge_configs(self):
         a = ScaledMMConfig(False, True, True)
         b = ScaledMMConfig(True, False, False)
         with pytest.raises(