delete swap_linear_with_dynamic from fsdp2 eager test case (#311)

Summary:
Pull Request resolved: #311

clean up some tech debt, the `swap_linear_with_dynamic`
function is redundant with `swap_linear_with_float8_linear`

Reviewed By: awgu

Differential Revision: D59685259

fbshipit-source-id: aab8861a5d9932a7bdd3dcb3014a49a0a048c55d

Author: vkuzo

test/test_fsdp2/test_fsdp2_eager.py

@@ -72,12 +72,6 @@ def get_local_inp(self, dtype: torch.dtype = torch.float32):
         dist.broadcast(global_inp, src=0)
         return global_inp.view(self.world_size, -1)[self.rank].view(16, 16)
 
-    def swap_linear_with_dynamic(self, module: nn.Module, **kwargs: Any) -> nn.Module:
-        kwargs["scaling_type_x"] = TensorScalingType.DYNAMIC
-        kwargs["scaling_type_w"] = TensorScalingType.DYNAMIC
-        kwargs["scaling_type_dL_dY"] = TensorScalingType.DYNAMIC
-        return swap_linear_with_float8_linear(module, **kwargs)
-
 
 class TestFloat8MultiProcess(FSDPTest, TestFloat8Common):
     @property
@@ -106,11 +100,11 @@ def _test_transformer_parity_dynamic(
         # latter uses fp8 compute. With fp8 all-gather, FSDP would pre-cast to
         # fp8 for that tied weight, incorrectly using fp8 for the embedding.
         weight_tying = not enable_fsdp_fp8_all_gather
-        module = self.init_transformer(weight_tying=weight_tying)
+        module = self.init_transformer(weight_tying=weight_tying).cuda()
         ref_module = copy.deepcopy(module)
-        ref_module = self.swap_linear_with_dynamic(ref_module).cuda()
+        swap_linear_with_float8_linear(ref_module)
         with set_enable_fsdp_fp8_all_gather(enable_fsdp_fp8_all_gather):
-            module = self.swap_linear_with_dynamic(module)
+            swap_linear_with_float8_linear(module)
         for submodule in module.modules():
             if isinstance(submodule, TransformerBlock):
                 fully_shard(submodule)
@@ -153,7 +147,7 @@ def _test_transformer_memory(self, enable_fsdp_fp8_all_gather: bool):
         # Emulate the fp8 matmul to bypass the scaled matmul op's divisibility
         # requirement to use a smaller activation size
         with set_enable_fsdp_fp8_all_gather(enable_fsdp_fp8_all_gather):
-            model = self.swap_linear_with_dynamic(model, emulate=True)
+            swap_linear_with_float8_linear(model, emulate=True)
         model_unsharded_numel = sum(p.numel() for p in model.parameters())
         model_sharded_numel = (model_unsharded_numel + 1) // 2
         block_lin_weight_numel = 0
@@ -331,7 +325,8 @@ def get_expected_all_gather_size(module: nn.Module):
         module_fp32 = self.init_single_module()
         ref_module = copy.deepcopy(module_fp32)
         with set_enable_fsdp_fp8_all_gather(True):
-            module = self.swap_linear_with_dynamic(module_fp32)
+            module_fp32 = swap_linear_with_float8_linear(module_fp32)
+            module = module_fp32
         fully_shard(module)
         local_inp = self.get_local_inp()
         expected_all_gather_size = get_expected_all_gather_size(ref_module)
@@ -359,7 +354,7 @@ def get_expected_all_gather_size(module: nn.Module):
         module = self.init_multi_module()
         ref_module = copy.deepcopy(module)
         with set_enable_fsdp_fp8_all_gather(True):
-            module = self.swap_linear_with_dynamic(module)
+            module = swap_linear_with_float8_linear(module)
         for submodule in module:
             fully_shard(submodule)
         fully_shard(module)
@@ -383,10 +378,11 @@ def test_fp32_fp8_single_module_parity(self):
         """
         for enable_fsdp_fp8_all_gather in [False, True]:
             module_fp32 = self.init_single_module()
-            ref_module = self.swap_linear_with_dynamic(copy.deepcopy(module_fp32))
+            ref_module = copy.deepcopy(module_fp32)
+            ref_module = swap_linear_with_float8_linear(ref_module)
             ref_module = ref_module.cuda()
             with set_enable_fsdp_fp8_all_gather(enable_fsdp_fp8_all_gather):
-                module = self.swap_linear_with_dynamic(module_fp32)
+                module = swap_linear_with_float8_linear(module_fp32)
             fully_shard(module)
             ref_optim = torch.optim.Adam(ref_module.parameters(), lr=1e-2)
             optim = torch.optim.Adam(module.parameters(), lr=1e-2, foreach=True)
@@ -407,11 +403,11 @@ def test_fp32_fp8_multi_module_parity(self):
         multiple modules/FSDP communication groups.
         """
         for enable_fsdp_fp8_all_gather in [False, True]:
-            module = self.init_multi_module()
+            module = self.init_multi_module().cuda()
             ref_module = copy.deepcopy(module)
-            ref_module = self.swap_linear_with_dynamic(ref_module).cuda()
+            ref_module = swap_linear_with_float8_linear(ref_module)
             with set_enable_fsdp_fp8_all_gather(enable_fsdp_fp8_all_gather):
-                module = self.swap_linear_with_dynamic(module)
+                module = swap_linear_with_float8_linear(module)
             for submodule in module:
                 fully_shard(submodule)
             fully_shard(module)