[moe training] update tests for torchtitan moe refactor

Author: danielvegamyhre

test/prototype/moe_training/test_training.py

@@ -22,11 +22,10 @@
 
 # this test requires torchtitan
 try:
-    from torchtitan.experiments.llama4.infra.expert_parallel import (
+    from torchtitan.distributed.expert_parallel import (
         set_token_group_alignment_size_m,
     )
-    from torchtitan.experiments.llama4.model.args import TransformerModelArgs
-    from torchtitan.experiments.llama4.model.moe import MoE
+    from torchtitan.models.moe import MoE, MoEArgs
 except ImportError:
     pytest.skip(
         "torchtitan not installed, skipping MoE tests.", allow_module_level=True
@@ -47,16 +46,15 @@ def test_moe_float8_training(target_fqns: list[str], compile: bool):
     # has the contraction dim be divisible by 16. 16 byte alignment is required
     # for the slowest moving dim (stride 1), so 16 bytes / 1 byte per element in fp8 = 16 elements.
     set_token_group_alignment_size_m(16)
-    model_args = TransformerModelArgs(
-        moe_enabled=True,
+    model_args = MoEArgs(
         num_experts=8,
-        dim=256,
     )
     init_std = 0.02
     device = torch.device("cuda")
 
     # reference bf16 MoE
-    ref_model = MoE(model_args).to(torch.bfloat16).cuda()
+    dim, hidden_dim = 256, 4 * 256
+    ref_model = MoE(model_args, dim, hidden_dim).to(torch.bfloat16).cuda()
     torch.manual_seed(42)
     ref_model.init_weights(init_std, device)
 
@@ -75,22 +73,21 @@ def moe_module_filter_fn(mod: nn.Module, cur_fqn: str) -> bool:
         return False
 
     # quantize test model
-    config = MoETrainingConfig(scaling_type=MoEScalingType.FP8_ROWWISE)
+    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,
     )
-
     if compile:
         # TODO: compile with fullgraph=True when torchtitan llama4 moe supports it
         model = torch.compile(model, fullgraph=False)
         ref_model = torch.compile(ref_model, fullgraph=False)
 
     # inputs
-    batch, seq, dim = 8, 2048, 256
+    batch, seq = 8, 2048
     ref_x = torch.randn(
         batch, seq, dim, dtype=torch.bfloat16, requires_grad=True, device=device
     )
@@ -145,18 +142,15 @@ def test_moe_mxfp8_training(target_fqns: list[str]):
     # Token groups must be divisible by 32 for mxfp8
     set_token_group_alignment_size_m(block_size)
 
-    model_args = TransformerModelArgs(
-        moe_enabled=True,
+    model_args = MoEArgs(
         num_experts=8,
-        dim=256,
-        multiple_of=block_size,
-        ffn_dim_multiplier=1.0,
     )
     init_std = 0.02
     device = torch.device("cuda")
 
     # reference bf16 MoE
-    ref_model = MoE(model_args).to(torch.bfloat16).cuda()
+    dim, hidden_dim = 256, 4 * 256
+    ref_model = MoE(model_args, dim, hidden_dim).to(torch.bfloat16).cuda()
     torch.manual_seed(42)
     ref_model.init_weights(init_std, device)
 
@@ -185,7 +179,7 @@ def moe_module_filter_fn(mod: nn.Module, cur_fqn: str) -> bool:
     )
 
     # inputs
-    batch, seq, dim = 8, 2048, 256
+    batch, seq = 8, 2048
     ref_x = torch.randn(
         batch, seq, dim, dtype=torch.bfloat16, requires_grad=True, device=device
     )

torchao/prototype/moe_training/scaled_grouped_mm.py

@@ -48,15 +48,15 @@ def _scaled_grouped_mm(
     """
     # TODO: Remove logging once prototype is more mature. This is currently very useful for development and debugging.
     if scaling_type == MoEScalingType.FP8_ROWWISE:
-        logger.info("Using fp8 rowwise scaled_grouped_mm")
+        print("Using fp8 rowwise scaled_grouped_mm")
         return _Float8GroupedMM.apply(
             A,
             B_t,
             offs,
             out_dtype,
         )
     elif scaling_type == MoEScalingType.MXFP8:
-        logger.info("Using mxfp8 scaled_grouped_mm")
+        print("Using mxfp8 scaled_grouped_mm")
         block_size = 32  # TODO: should we make this configurable? plumb it through in a config somehow?
         return _MXFP8GroupedMM.apply(
             A,