diff --git a/src/megatron/bridge/models/conversion/param_mapping.py b/src/megatron/bridge/models/conversion/param_mapping.py
index fef5d979d4..ef282f0b2d 100644
--- a/src/megatron/bridge/models/conversion/param_mapping.py
+++ b/src/megatron/bridge/models/conversion/param_mapping.py
@@ -993,7 +993,11 @@ def hf_to_megatron(
                 hf_weights = hf_weights.to(target_param.dtype)
 
             actual_dim0_size = hf_weights.shape[0]
-            expect_dim0_size = target_param.shape[0] * self.tp_size
+            # DTensor.shape is already the global shape across TP ranks, while
+            # a regular Megatron parameter stores only its local TP shard.
+            expect_dim0_size = target_param.shape[0]
+            if not isinstance(target_param, DTensor):
+                expect_dim0_size *= self.tp_size
             if actual_dim0_size != expect_dim0_size:
                 assert self.megatron_param in {"embedding.word_embeddings.weight", "output_layer.weight"}, (
                     f"{hf_weights.shape=} {target_param.shape=} {self.tp_size=} {self.megatron_param=} {self.hf_param=}"
diff --git a/tests/functional_tests/test_groups/converter/test_hf_fsdp_conversion.py b/tests/functional_tests/test_groups/converter/test_hf_fsdp_conversion.py
index ae92cf319f..dc4822c432 100644
--- a/tests/functional_tests/test_groups/converter/test_hf_fsdp_conversion.py
+++ b/tests/functional_tests/test_groups/converter/test_hf_fsdp_conversion.py
@@ -101,9 +101,8 @@ def qwen3_moe_toy_model_path(self, tmp_path_factory):
     @pytest.mark.parametrize(
         "tp,ep,nproc,test_name",
         [
-            # CI only covers TP=1 cases. TP>1 cases depend on Megatron-LM commit
-            # 8cbc45b6e (PR#3191, merged 2026-04-07).
             (1, 1, 1, "FSDP_base"),
+            (2, 1, 2, "FSDP_TP2"),
         ],
     )
     def test_hf_fsdp_roundtrip(self, qwen3_moe_toy_model_path, tmp_path, tp, ep, nproc, test_name):