pytorch · danielvegamyhre · Jun 21, 2025 · Jun 23, 2025 · Jun 24, 2025 · Jun 24, 2025
diff --git a/test/prototype/moe_training/test_fsdp.py b/test/prototype/moe_training/test_fsdp.py
@@ -1,3 +1,16 @@
+# 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.
+######################################################################
+#
+# To run these unit tests, use the following command:
+#
+# torchrun --nproc_per_node=${NUM_GPUS} -m pytest test_fsdp.py
+#
+#######################################################################
+
 import copy
 import os
 

diff --git a/test/prototype/moe_training/test_tp.py b/test/prototype/moe_training/test_tp.py
@@ -0,0 +1,219 @@
+# 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.
+######################################################################
+#
+# To run these unit tests, use the following command:
+#
+# torchrun --nproc_per_node=${NUM_GPUS} -m pytest test_tp.py
+#
+#######################################################################
+
+import copy
+import os
+
+import pytest
+import torch
+from torch import distributed as dist
+from torch import nn
+from torch.distributed.device_mesh import DeviceMesh, init_device_mesh
+from torch.nn import functional as F
+
+# this feature requires CUDA and SM89+
+if not torch.cuda.is_available() or torch.cuda.get_device_capability() < (8, 9):
+    pytest.skip(
+        "CUDA not available or compute capability < 8.9", allow_module_level=True
+    )
+
+from torchao.float8.float8_utils import compute_error
+from torchao.prototype.moe_training.conversion_utils import MoETrainingConfig
+from torchao.prototype.moe_training.tensor import ScaledGroupedMMTensor
+from torchao.quantization.quant_api import quantize_
+
+# this test requires torchtitan
+try:
+    from torchtitan.experiments.llama4.infra.parallelize import apply_moe_tp
 script: | 
 script: | 
+    from torchtitan.experiments.llama4.model.args import TransformerModelArgs
+    from torchtitan.experiments.llama4.model.moe import MoE
+except ImportError:
+    import warnings
+
+    warnings.warn("torchtitan not installed, skipping MoE tests.")
+    pytest.skip(allow_module_level=True)
+
+
+@pytest.mark.parametrize(
+    "target_fqns",
+    [
+        ["experts"],
+        ["experts,shared_expert"],
+    ],
+)
+def test_moe_float8_training_tp_sp(target_fqns: list[str]):
+    assert torch.cuda.is_available()
+
+    # setup distributed for fsdp
+    mesh = setup_distributed()
+
+    # define model args
+    model_args = TransformerModelArgs(
+        moe_enabled=True,
+        num_experts=8,
+        dim=256,
+        vocab_size=1024,
+    )
+    init_std = 0.02
+    device = torch.device("cuda")
+
+    # reference bf16 MoE
+    ref_model = MoE(model_args).to(torch.bfloat16).cuda()
+    torch.manual_seed(1)
+    ref_model.init_weights(init_std, device)
+
+    # target MoE for testing conversion
+    model = copy.deepcopy(ref_model)
+
+    # assert starting params are identical for both models
+    for param1, param2 in zip(model.parameters(), ref_model.parameters()):
+        assert torch.equal(param1, param2)
+
+    # convert MoE to float8 training
+    def moe_module_filter_fn(mod: nn.Module, cur_fqn: str) -> bool:
+        for target_fqn in target_fqns:
+            if target_fqn in cur_fqn:
+                return True
+        return False
+
+    # quantize test model
+    config = MoETrainingConfig()
+    quantize_(model, config=config, filter_fn=moe_module_filter_fn)
+
+    # validate that only the experts were converted
+    _validate_model_conversion(
+        model,
+        target_fqns=target_fqns,
+    )
+
+    # apply TP
+    apply_moe_tp(model, mesh)
+    apply_moe_tp(ref_model, mesh)
+
+    # inputs
+    batch, seq, dim = 8, 2048, 256
+    ref_x = torch.randn(
+        batch, seq, dim, dtype=torch.bfloat16, requires_grad=True, device=device
+    )
+    x = ref_x.detach().clone().requires_grad_(True)
+
+    # forward pass
+    ref_out = ref_model(ref_x)
+    out = model(x)
+
+    # validate output
+    out_sqnr = compute_error(out, ref_out)
+    assert out_sqnr.item() >= 30.0, f"SQNR must be >= 30.0, got {out_sqnr.item()}."
+
+    # compute loss
+    labels = torch.ones_like(ref_out)
+    ref_loss = F.mse_loss(ref_out, labels)
+    out_loss = F.mse_loss(out, labels)
+
+    # backward pass
+    ref_loss.backward()
+    out_loss.backward()
+
+    # validate input gradient
+    input_grad_sqnr = compute_error(x.grad, ref_x.grad)
+    assert input_grad_sqnr.item() >= 30.0, (
+        f"SQNR must be >= 30.0, got {input_grad_sqnr.item()}."
+    )
+
+    # validate param gradients
+    for param1, param2 in zip(model.parameters(), ref_model.parameters()):
+        param_grad_sqnr = compute_error(param1.grad, param2.grad)
+        assert param_grad_sqnr.item() >= 25.0, (
+            f"SQNR must be >= 25.0, got {param_grad_sqnr.item()}."
+        )
+
+    dist.destroy_process_group()
+
+
+def _validate_model_conversion(
+    root_module: nn.Module,
+    target_fqns: list[str],
+):
+    def _recursive_validate(
+        module: nn.Module,
+        cur_fqn: str,
+    ):
+        is_allowed_module = any([target_fqn in cur_fqn for target_fqn in target_fqns])
+
+        # check current module params
+        for param_name, param in module.named_parameters(recurse=False):
+            is_converted_type = isinstance(param, ScaledGroupedMMTensor)
+            if is_converted_type:
+                assert is_allowed_module, (
+                    f"Module {cur_fqn} is not in target_fqns, but has converted param {param_name}."
+                )
+            if not is_allowed_module:
+                assert not is_converted_type, (
+                    f"Module {cur_fqn} is not in target_fqns, but has converted param {param_name}."
+                )
+
+        # recursively check child modules
+        for child_name, child_module in module.named_children():
+            child_fqn = f"{cur_fqn}.{child_name}" if cur_fqn else child_name
+            _recursive_validate(child_module, child_fqn)
+
+    _recursive_validate(root_module, "")
+
+
+def setup_distributed():
+    rank = int(os.environ["RANK"])
+    world_size = int(os.environ["WORLD_SIZE"])
+    dist.init_process_group("nccl", rank=rank, world_size=world_size)
+    device_mesh = init_device_mesh("cuda", (world_size,))
+    # seed must be the same in all processes
+    torch.manual_seed(1)
+    torch.cuda.set_device(rank)
+    return device_mesh
+
+
+def apply_moe_tp(
+    model: nn.Module,
+    tp_mesh: DeviceMesh,
+):
+    # base on llama4 MoE TP implementation here: https://github.com/pytorch/torchtitan/blob/d9cc6b4df341eec27768b5ab9cead87ef595dbc2/torchtitan/experiments/llama4/infra/parallelize.py#L147C1-L180C10
+    from torch.distributed.tensor import Partial, Replicate, Shard
+    from torch.distributed.tensor.parallel import (
+        PrepareModuleInputOutput,
+        parallelize_module,
+    )
+    from torchtitan.experiments.llama4.infra.expert_parallel import (
+        NoParallel,
+        TensorParallel,
+    )
+
+    moe_layer_plan = {
+        # input / output sharding on the seqlen dim
+        # all-gather for input, reduce-scatter for output
+        "moe": PrepareModuleInputOutput(
+            input_layouts=(Shard(1),),
+            desired_input_layouts=(Replicate(),),
+            use_local_input=True,
+            output_layouts=(Partial(),),
+            desired_output_layouts=(Shard(1),),
+        ),
+        # replicate computation for the router
+        "moe.router.gate": NoParallel(),
+        # input Replicate, output Partial
+        "moe.experts": TensorParallel(output_layout=Partial()),
+        "moe.shared_expert": TensorParallel(output_layout=Partial()),
+    }
+    parallelize_module(
+        module=model,
+        device_mesh=tp_mesh,
+        parallelize_plan=moe_layer_plan,
+    )
diff --git a/torchao/prototype/moe_training/conversion_utils.py b/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
diff --git a/torchao/prototype/moe_training/scaled_grouped_mm.py b/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
@@ -18,11 +19,13 @@
     _is_column_major,
 )
 
+logger: logging.Logger = logging.getLogger(__name__)
+
 
 def _scaled_grouped_mm(
     A: torch.Tensor,
     B_t: torch.Tensor,
-    offs: torch.Tensor,
+    offs: Optional[torch.Tensor] = None,
     out_dtype: Optional[torch.dtype] = torch.bfloat16,
 ) -> 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.info("Using differentiable _scaled_grouped_mm")
     return _Float8GroupedMM.apply(
         A,
         B_t,
@@ -54,12 +57,11 @@ def forward(
         ctx,
         A: torch.Tensor,
         B_t: torch.Tensor,
-        offs: torch.Tensor,
+        offs: Optional[torch.Tensor] = None,
         out_dtype: Optional[torch.dtype] = torch.bfloat16,
-        use_triton_for_per_group_scales: bool = True,
     ) -> torch.Tensor:
-        # torchao _scaled_grouped_mm only supports A=2D, B=3D.
-        assert A.ndim == 2, "A must be 2D"
+        # torchao _scaled_grouped_mm only supports A=2D|3D and B=3D.
+        assert A.ndim == 2 or A.ndim == 3, "A must be 2D or 3D"
         assert B_t.ndim == 3, "B must be 3D"
 
         assert A.size(-1) % 16 == 0, (
@@ -76,7 +78,6 @@ def forward(
         assert B_t.dtype == torch.float32 or B_t.dtype == torch.bfloat16, (
             "B must be float32 or bfloat16"
         )
-        assert offs.dtype == torch.int32, "offs must be int32"
 
         # Assert A and B dims are compatible for a scaled grouped GEMM.
         assert A.size(-1) == B_t.size(-2), (
@@ -152,9 +153,11 @@ def forward(
         return torch._scaled_grouped_mm(
             A_fp8_row_major,
             B_t_fp8_col_major,
-            A_scales.squeeze().reciprocal(),
-            B_t_scales.squeeze().reciprocal(),
-            offs,
+            # Squeeze A scales to: (B, S, 1) => (B, M), or (B*S, 1) => (B*S)
+            A_scales.squeeze(-1).reciprocal(),
+            # Squeeze B scales to: (B, 1, N) => (B, N)
+            B_t_scales.squeeze(1).reciprocal(),
+            offs=offs,
             out_dtype=out_dtype,
             use_fast_accum=True,
         )
@@ -194,9 +197,9 @@ def backward(ctx, grad_output: torch.Tensor):
         grad_A = torch._scaled_grouped_mm(
             grad_output_fp8_row_major,
             B_fp8_col_major,
-            grad_output_scales.squeeze().reciprocal(),
-            B_scales.squeeze().reciprocal(),
-            offs,
+            grad_output_scales.squeeze(-1).reciprocal(),
+            B_scales.squeeze(1).reciprocal(),
+            offs=offs,
             out_dtype=out_dtype,
             use_fast_accum=True,
         )
@@ -241,7 +244,7 @@ def backward(ctx, grad_output: torch.Tensor):
             A_fp8_col_major,
             grad_output_t_scales.reciprocal(),
             A_scales.reciprocal(),
-            offs,
+            offs=offs,
             out_dtype=out_dtype,
             use_fast_accum=True,
         )