[fp8] support asynchronous FP8 communication (#5997)

* fix

* fix

* fix

* support async all2all

* support async op for all gather

* fix

* fix

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* fix

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

colossalai/quantization/fp8.py

@@ -10,6 +10,18 @@ from torch.distributed import ReduceOp
 SUPPORT_TORCH_COMPILE = Version(torch.__version__) >= Version("2.3.0")
 
 
+class Handle:
+    def __init__(self, handles=[], remain_ops=None) -> None:
+        self.handles = handles
+        self.remain_ops = remain_ops
+
+    def wait(self):
+        for handle in self.handles:
+            handle.wait()
+        if self.remain_ops:
+            self.remain_ops()
+
+
 def cast_to_fp8(inp: torch.Tensor, fp8_format="e4m3", per_channel_scale=False) -> Tuple[torch.Tensor, torch.Tensor]:
     r"""
     casting torch Tensor into specified fp8 tensor with per-channel scaling or per-tensor scaling.
@@ -68,7 +80,9 @@ def cast_from_fp8(
     return ret.to(ret_type)
 
 
-def all_reduce_fp8(tensor: torch.Tensor, fp8_format="e4m3", op=ReduceOp.SUM, group=None) -> None:
+def all_reduce_fp8(
+    tensor: torch.Tensor, fp8_format="e4m3", op=ReduceOp.SUM, group=None, async_op: bool = False
+) -> Optional[Handle]:
     r"""
     This is an in-place operation for compressed all_reduce using fp8.
     It works like dist.all_reduce but during communication the data is cast to fp8 format.
@@ -105,6 +119,7 @@ def all_reduce_fp8(tensor: torch.Tensor, fp8_format="e4m3", op=ReduceOp.SUM, gro
     scale_list = [torch.ones(1, dtype=scale.dtype, device=input_device) for _ in range(world_size)]
     dist.all_gather(scale_list, scale, group=group)
     summed_out = torch.zeros_like(output_chunks[0]).to(input_type)
+
     for scale, out in zip(scale_list, output_chunks):
         out = out.view(fp8_type)
         summed_out += cast_from_fp8(out, scale, input_type)
@@ -113,19 +128,28 @@ def all_reduce_fp8(tensor: torch.Tensor, fp8_format="e4m3", op=ReduceOp.SUM, gro
         summed_out.div_(world_size)
 
     summed_out_fp8, scale = cast_to_fp8(summed_out, fp8_format=fp8_format)
-    dist.all_gather(scale_list, scale, group=group)
+    gather_scale_handle = dist.all_gather(scale_list, scale, group=group, async_op=async_op)
 
     tensor_list = [torch.empty_like(summed_out_fp8.view(torch.uint8)) for _ in range(world_size)]
-    dist.all_gather(tensor_list, summed_out_fp8.view(torch.uint8), group=group)
+    gather_tensor_handle = dist.all_gather(
+        tensor_list, summed_out_fp8.view(torch.uint8), group=group, async_op=async_op
+    )
+
+    def cat_op():
         for i in range(world_size):
-        tensor_list[i] = tensor_list[i].view(fp8_type).to(input_type) * scale_list[i].to(input_device)
+            tensor_list[i] = tensor_list[i].view(fp8_type).to(input_type) * scale_list[i]
         out = torch.cat(tensor_list, dim=0)
         tensor.copy_(out[:input_size].view(input_shape).to(input_type))
 
+    if async_op:
+        return Handle([gather_scale_handle, gather_tensor_handle], cat_op)
+    else:
+        cat_op()
+
 
 def all_to_all_single_fp8(
     output, input, output_split_sizes=None, input_split_sizes=None, fp8_format="e5m2", group=None, async_op=False
-) -> None:
+) -> Optional[Handle]:
     r"""
     This is an in-place operation for compressed all_reduce using fp8.
     It works like dist.all_to_all_single but during communication the data is cast to fp8 format.
@@ -163,9 +187,11 @@ def all_to_all_single_fp8(
         else:
             output_chunks = [torch.empty_like(input_chunks[0]) for _ in range(world_size)]
 
-    dist.all_to_all(output_chunks, input_chunks, group=group)
+    chunk_handle = dist.all_to_all(output_chunks, input_chunks, group=group, async_op=async_op)
     scale_list = [torch.ones(1, dtype=scale.dtype, device=input_device) for _ in range(world_size)]
-    dist.all_gather(scale_list, scale, group=group)
+    scale_hanle = dist.all_gather(scale_list, scale, group=group, async_op=async_op)
+
+    def cast_op():
         cast_output_chunk = [
             cast_from_fp8(out.view(fp8_type), scale, input_type) for scale, out in zip(scale_list, output_chunks)
         ]
@@ -178,6 +204,11 @@ def all_to_all_single_fp8(
             outputs_shape = input_shape
         output.data = tensor_out.view(outputs_shape).to(input_type)
 
+    if async_op:
+        return Handle([chunk_handle, scale_hanle], cast_op)
+    else:
+        cast_op()
+
 
 def cast_to_fp8_pipeline(inp: Any) -> None:
     """
@@ -250,7 +281,9 @@ def cast_from_fp8_pipeline(inp: Any, del_metadata=True) -> None:
         del inp["dtype"]
 
 
-def reduce_scatter_fp8(output: torch.Tensor, input_list, group, fp8_format="e5m2") -> None:
+def reduce_scatter_fp8(
+    output: torch.Tensor, input_list, group, fp8_format="e5m2", async_op: bool = False
+) -> Optional[Handle]:
     r"""
     This is an in-place operation for compressed reduce_scatter using fp8.
     It works like dist.reduce_scatter but during communication the data is cast to fp8 format.
@@ -277,15 +310,21 @@ def reduce_scatter_fp8(output: torch.Tensor, input_list, group, fp8_format="e5m2
         cast_input_list.append(ret)
         output_chunks.append(torch.empty_like(ret))
         output_scale_list.append(torch.empty_like(scale))
-    dist.all_to_all(output_chunks, cast_input_list, group=group)
-    dist.all_to_all(output_scale_list, scale_list, group=group)
+    chunk_handle = dist.all_to_all(output_chunks, cast_input_list, group=group, async_op=async_op)
+    scale_handle = dist.all_to_all(output_scale_list, scale_list, group=group, async_op=async_op)
 
+    def cast_op():
         summed_out = torch.zeros_like(output_chunks[0]).to(input_type)
         for scale, out in zip(output_scale_list, output_chunks):
             out = out.view(fp8_type)
             summed_out += cast_from_fp8(out, scale, input_type)
         output.data = summed_out
 
+    if async_op:
+        return Handle([chunk_handle, scale_handle], cast_op)
+    else:
+        cast_op()
+
 
 def fp8_compress_ddp_grad_comm_hook_async(
     process_group: dist.ProcessGroup,
@@ -500,7 +539,8 @@ def all_gather_into_tensor_flat_fp8(
     output_shape: torch.Size,
     group: dist.ProcessGroup,
     fp8_format: str = "e4m3",
-):
+    async_op: bool = False,
+) -> Optional[Handle]:
     """all gather into tensor in fp8 format
 
     Args:
@@ -547,15 +587,25 @@ def all_gather_into_tensor_flat_fp8(
         scale = fp8_max / per_tensor_max
         fp8_input = (scale * input_tensor.float()).to(fp8_type)
     scale_inv = 1.0 / scale
+
     buffer = torch.empty_like(output_tensor, dtype=fp8_type)
-    dist.all_gather_into_tensor(buffer.view(torch.uint8), fp8_input.view(torch.uint8), group=group)
+    tensor_handle = dist.all_gather_into_tensor(
+        buffer.view(torch.uint8), fp8_input.view(torch.uint8), group=group, async_op=async_op
+    )
+
+    def cast_op():
         numel = output_shape.numel()
         valid_buffer = buffer[:numel].reshape(output_shape)
         valid_buffer = cast_from_fp8(valid_buffer, scale_inv, input_type, per_channel_scale=(len(output_shape) == 2))
         output_tensor[:numel].copy_(valid_buffer.view(-1))
 
+    if async_op:
+        return Handle([tensor_handle], cast_op)
+    else:
+        cast_op()
+
 
-def all_to_all_fp8(output_list, input_list, group=None, fp8_format="e5m2"):
+def all_to_all_fp8(output_list, input_list, group=None, fp8_format="e5m2", async_op=False):
 
     world_size = dist.get_world_size(group)
 
@@ -573,17 +623,23 @@ def all_to_all_fp8(output_list, input_list, group=None, fp8_format="e5m2"):
 
     output_scale_list = [torch.empty_like(x) for x in scale_list]
     output_tensor_list = [torch.empty_like(x) for x in tensor_list]
-    dist.all_to_all(output_tensor_list, tensor_list, group=group)
-    dist.all_to_all(output_scale_list, scale_list, group=group)
+    tensor_hanle = dist.all_to_all(output_tensor_list, tensor_list, group=group, async_op=async_op)
+    scale_handle = dist.all_to_all(output_scale_list, scale_list, group=group, async_op=async_op)
 
+    def cast_op():
         for i in range(world_size):
             scale = output_scale_list[i]
             tensor = output_tensor_list[i]
             tensor = tensor.view(fp8_type)
             output_list[i].copy_(cast_from_fp8(tensor, scale, input_type))
 
+    if async_op:
+        return Handle([tensor_hanle, scale_handle], cast_op)
+    else:
+        cast_op()
+
 
-def gather_fp8(output_list, input_, group=None, fp8_format="e5m2"):
+def gather_fp8(output_list, input_, group=None, fp8_format="e5m2", async_op: bool = False) -> Optional[Handle]:
 
     world_size = dist.get_world_size(group)
 
@@ -593,14 +649,20 @@ def gather_fp8(output_list, input_, group=None, fp8_format="e5m2"):
     input_ = ret.view(torch.uint8)
     tensor_list = [torch.empty_like(input_) for _ in range(world_size)]
     scale_list = [torch.ones(1, dtype=scale.dtype, device=input_.device) for _ in range(world_size)]
-    dist.all_gather(tensor_list, input_, group=group)
-    dist.all_gather(scale_list, scale, group=group)
+    chunk_handle = dist.all_gather(tensor_list, input_, group=group, async_op=async_op)
+    scale_hanle = dist.all_gather(scale_list, scale, group=group, async_op=async_op)
 
+    def cast_op():
         for i in range(world_size):
             output = tensor_list[i].view(fp8_type)
             scale = scale_list[i]
             output_list[i].copy_(cast_from_fp8(output, scale, input_type))
 
+    if async_op:
+        return Handle([chunk_handle, scale_hanle], cast_op)
+    else:
+        cast_op()
+
 
 class _LinearFp8(torch.autograd.Function):
     @staticmethod

tests/test_fp8/test_all_to_all_single.py

@@ -10,19 +10,24 @@ from colossalai.testing import parameterize, rerun_if_address_is_in_use, spawn
 
 
 @parameterize("shape", [(4,), (1, 8, 16), (4, 8, 16)])
-@parameterize("dtype", [torch.bfloat16])
-def check_all2all(shape, dtype):
+@parameterize("dtype", [torch.bfloat16, torch.float16])
+@parameterize("async_op", [True, False])
+def check_all2all(shape, dtype, async_op):
     x = torch.rand(shape, dtype=dtype, device=get_accelerator().get_current_device())
     output = torch.empty_like(x)
     output_fp8 = torch.empty_like(x)
-    dist.all_to_all_single(output, x, group=_get_default_group(), async_op=False)
-    all_to_all_single_fp8(output_fp8, x, group=_get_default_group(), async_op=False)
+    origin_hanle = dist.all_to_all_single(output, x, group=_get_default_group(), async_op=async_op)
+    fp8_handle = all_to_all_single_fp8(output_fp8, x, group=_get_default_group(), async_op=async_op)
+    if async_op:
+        origin_hanle.wait()
+        fp8_handle.wait()
     assert_close(output, output_fp8, rtol=0.1, atol=0.1)
 
 
 @parameterize("shape", [(8, 8, 16)])
 @parameterize("dtype", [torch.bfloat16, torch.float16])
-def check_all2all_uneven(shape, dtype):
+@parameterize("async_op", [True, False])
+def check_all2all_uneven(shape, dtype, async_op):
     x = torch.rand(shape, dtype=dtype, device=get_accelerator().get_current_device())
     input_split_sizes = [3, 3, 1, 1]
     if dist.get_rank() in [0, 1]:
@@ -33,22 +38,25 @@ def check_all2all_uneven(shape, dtype):
     output_shape[0] = sum(output_split_sizes)
     output = torch.empty(output_shape, device=x.device, dtype=x.dtype)
     output_fp8 = torch.empty(output_shape, device=x.device, dtype=x.dtype)
-    dist.all_to_all_single(
+    origin_hanle = dist.all_to_all_single(
         output,
         x,
         output_split_sizes=output_split_sizes,
         input_split_sizes=input_split_sizes,
         group=_get_default_group(),
-        async_op=False,
+        async_op=async_op,
     )
-    all_to_all_single_fp8(
+    fp8_handle = all_to_all_single_fp8(
         output_fp8,
         x,
         output_split_sizes=output_split_sizes,
         input_split_sizes=input_split_sizes,
         group=_get_default_group(),
-        async_op=False,
+        async_op=async_op,
     )
+    if async_op:
+        origin_hanle.wait()
+        fp8_handle.wait()
     assert_close(output, output_fp8, rtol=0.1, atol=0.1)
 
 

tests/test_fp8/test_fp8_allgather_flat.py

@@ -12,7 +12,8 @@ from colossalai.testing import parameterize, rerun_if_address_is_in_use, spawn
 
 @parameterize("shape", [(3, 7), (2, 1), (1, 2), (2, 2), (4, 2), (5,), (4,), (2,)])
 @parameterize("dtype", [torch.bfloat16, torch.float16])
-def check_4gpu(shape, dtype):
+@parameterize("async_op", [True, False])
+def check_4gpu(shape, dtype, async_op):
     world_size = dist.get_world_size()
     rank = dist.get_rank()
     x = torch.rand(shape, dtype=dtype, device=get_accelerator().get_current_device())
@@ -22,7 +23,9 @@ def check_4gpu(shape, dtype):
         flat_padded_x = F.pad(flat_padded_x, (0, pad_size))
     output = torch.empty_like(flat_padded_x)
     chunk = flat_padded_x.chunk(world_size)[rank].clone()
-    all_gather_into_tensor_flat_fp8(output, chunk, x.shape, group=_get_default_group())
+    handle = all_gather_into_tensor_flat_fp8(output, chunk, x.shape, group=_get_default_group(), async_op=async_op)
+    if async_op:
+        handle.wait()
     assert_close(output[: x.numel()], x.view(-1), rtol=0.1, atol=0.1)
 
 

tests/test_fp8/test_fp8_allreduce.py

@@ -22,15 +22,22 @@ from colossalai.testing import parameterize, rerun_if_address_is_in_use, spawn
 )
 @parameterize("dtype", [torch.float16, torch.bfloat16])
 @parameterize("fp8_format", ["e4m3", "e5m2"])
-def check_4gpu(shape, dtype, fp8_format):
+@parameterize("async_op", [True, False])
+def check_4gpu(shape, dtype, fp8_format, async_op):
     x = torch.rand(shape, dtype=dtype, device=get_accelerator().get_current_device())
     x_fp8 = x.clone()
-    dist.all_reduce(x)
-    all_reduce_fp8(x_fp8, fp8_format=fp8_format)
+    origin_handle = dist.all_reduce(x, async_op=async_op)
+    fp8_handle = all_reduce_fp8(x_fp8, fp8_format=fp8_format, async_op=async_op)
+    if async_op:
+        origin_handle.wait()
+        fp8_handle.wait()
     assert_close(x, x_fp8, rtol=0.1, atol=0.1)
 
-    dist.all_reduce(x, op=dist.ReduceOp.AVG)
-    all_reduce_fp8(x_fp8, op=dist.ReduceOp.AVG, fp8_format=fp8_format)
+    origin_handle = dist.all_reduce(x, op=dist.ReduceOp.AVG, async_op=async_op)
+    fp8_handle = all_reduce_fp8(x_fp8, op=dist.ReduceOp.AVG, fp8_format=fp8_format, async_op=async_op)
+    if async_op:
+        origin_handle.wait()
+        fp8_handle.wait()
     assert_close(x, x_fp8, rtol=0.1, atol=0.1)
 
 

tests/test_fp8/test_fp8_gather.py

@@ -24,13 +24,17 @@ from colossalai.testing import parameterize, rerun_if_address_is_in_use, spawn
 )
 @parameterize("dtype", [torch.bfloat16, torch.float16])
 @parameterize("fp8_format", ["e4m3", "e5m2"])
-def check_4gpu(shape, dtype, fp8_format):
+@parameterize("async_op", [True, False])
+def check_4gpu(shape, dtype, fp8_format, async_op):
     world_size = dist.get_world_size()
     x = torch.rand(shape, dtype=dtype, device=get_accelerator().get_current_device())
     output_list = [torch.empty_like(x) for _ in range(world_size)]
     output_list_fp8 = [torch.empty_like(x) for _ in range(world_size)]
-    gather_fp8(output_list_fp8, x, group=_get_default_group(), fp8_format=fp8_format)
-    dist.all_gather(output_list, x, group=_get_default_group())
+    fp8_handle = gather_fp8(output_list_fp8, x, group=_get_default_group(), fp8_format=fp8_format, async_op=async_op)
+    origin_hanle = dist.all_gather(output_list, x, group=_get_default_group(), async_op=async_op)
+    if async_op:
+        fp8_handle.wait()
+        origin_hanle.wait()
     assert_close(output_list, output_list_fp8, rtol=0.1, atol=0.1)