diff --git a/megatron/core/inference/contexts/dynamic_context.py b/megatron/core/inference/contexts/dynamic_context.py
index a0ed0b80a7d..f9695f6c9a5 100644
--- a/megatron/core/inference/contexts/dynamic_context.py
+++ b/megatron/core/inference/contexts/dynamic_context.py
@@ -5,7 +5,7 @@
 import operator
 import warnings
 from contextlib import nullcontext
-from typing import List, Optional, Sequence, Tuple
+from typing import Dict, List, Optional, Sequence, Tuple
 
 import torch  # type: ignore
 import torch.nn.functional as F  # type: ignore
@@ -1182,6 +1182,13 @@ def initialize_all_tensors(self) -> None:
             max_mamba_chunks=self._max_mamba_chunks,
         )
 
+        # Cache of (input_ids_view, pos_ids_view) keyed by num_tokens. Instead of slicing and
+        # unsqueezing on every new inference step (constructing new TensorImpls at 30-60 us),
+        # we fix the underlying storage so views are reusable across steps. The number of entries
+        # is bounded by the graph sizes plus eager-mode token counts, which are rounded up to
+        # multiples of TOKEN_ROUNDER and capped at max_tokens / TOKEN_ROUNDER distinct values.
+        self._input_position_views: Dict[int, Tuple[Tensor, Tensor]] = {}
+
         # Bind the shared MHA GPU views to both graph and non-graph metadata;
         # only one is active per step, so sharing storage is safe.
         self.graph_attn_metadata["mha_metadata"].bind_gpu_buffers(self.gpu_view)
@@ -2432,10 +2439,14 @@ def current_input_and_position_ids(
         assert num_tokens >= self.padded_batch_dimensions.decode_req_count * (
             self.num_speculative_tokens + 1
         )
-        return (
-            self.gpu_view.token_to_input_ids[:num_tokens].unsqueeze(0),
-            self.gpu_view.token_to_pos_ids[:num_tokens].unsqueeze(0),
-        )
+        cached = self._input_position_views.get(num_tokens)
+        if cached is not None:
+            return cached
+        input_ids = self.gpu_view.token_to_input_ids[:num_tokens].unsqueeze(0)
+        pos_ids = self.gpu_view.token_to_pos_ids[:num_tokens].unsqueeze(0)
+        cached = (input_ids, pos_ids)
+        self._input_position_views[num_tokens] = cached
+        return cached
 
     def speculative_required_logit_indices(self) -> Tensor:
         """Token-level indices needed for speculative decode verification.
diff --git a/tests/unit_tests/inference/contexts/test_dynamic_context.py b/tests/unit_tests/inference/contexts/test_dynamic_context.py
index 0e307f600f1..6a4838f2c46 100644
--- a/tests/unit_tests/inference/contexts/test_dynamic_context.py
+++ b/tests/unit_tests/inference/contexts/test_dynamic_context.py
@@ -256,6 +256,46 @@ def test_token_overflow_error(self, is_hybrid_model: bool):
                 )
             )  # Exceeding max token count
 
+    @pytest.mark.internal
+    @rounder_override(64)
+    def test_current_input_and_position_ids_view_cache(self):
+        dynamic_context = self._get_dynamic_context(
+            params_dtype=torch.float32,
+            num_layers=2,
+            kv_channels=64,
+            num_attention_heads=8,
+            max_sequence_length=128,
+            buffer_size_gb=0.1,
+            block_size_tokens=128,
+            max_tokens=None,
+        )
+
+        num_tokens = 64
+        dynamic_context.padded_active_token_count = num_tokens
+
+        # First call: cache miss, populates entry.
+        assert num_tokens not in dynamic_context._input_position_views
+        input_ids_view, pos_ids_view = dynamic_context.current_input_and_position_ids()
+        assert num_tokens in dynamic_context._input_position_views
+
+        # Second call: cache hit returns the same tensor objects.
+        cached_input_ids, cached_pos_ids = dynamic_context.current_input_and_position_ids()
+        assert cached_input_ids is input_ids_view
+        assert cached_pos_ids is pos_ids_view
+
+        # Writing new values into the underlying storage must be reflected by the cached views.
+        device = dynamic_context.gpu_view.token_to_input_ids.device
+        new_input_ids = torch.arange(num_tokens, dtype=torch.long, device=device)
+        new_pos_ids = torch.arange(num_tokens, 2 * num_tokens, dtype=torch.long, device=device)
+        dynamic_context.gpu_view.token_to_input_ids[:num_tokens] = new_input_ids
+        dynamic_context.gpu_view.token_to_pos_ids[:num_tokens] = new_pos_ids
+
+        refreshed_input_ids, refreshed_pos_ids = dynamic_context.current_input_and_position_ids()
+        assert refreshed_input_ids is input_ids_view
+        assert refreshed_pos_ids is pos_ids_view
+        assert torch.equal(refreshed_input_ids.squeeze(0), new_input_ids)
+        assert torch.equal(refreshed_pos_ids.squeeze(0), new_pos_ids)
+
     @pytest.mark.internal
     @rounder_override(64)
     @pytest.mark.parametrize("is_hybrid_model", [False, True])