diff --git a/vllm/models/kimi_k3/amd/ops/kda_chunk.py b/vllm/models/kimi_k3/amd/ops/kda_chunk.py
index efcdb1bc75..748581771b 100644
--- a/vllm/models/kimi_k3/amd/ops/kda_chunk.py
+++ b/vllm/models/kimi_k3/amd/ops/kda_chunk.py
@@ -286,9 +286,9 @@ def fused_kda_chunk(
         None if checkpoint_offsets is None else checkpoint_offsets.to(torch.int32),
         None
         if checkpoint_state_indices is None
-        else checkpoint_state_indices.to(torch.int32),
+        else checkpoint_state_indices.to(torch.int32).contiguous(),
         state_cache,
-        None if state_indices is None else state_indices.to(torch.int32),
+        None if state_indices is None else state_indices.to(torch.int32).contiguous(),
         has_initial_state,
     )
     return out, final_state
diff --git a/vllm/v1/attention/backends/mla/rocm_aiter_mla.py b/vllm/v1/attention/backends/mla/rocm_aiter_mla.py
index c341bfc348..b850deb070 100644
--- a/vllm/v1/attention/backends/mla/rocm_aiter_mla.py
+++ b/vllm/v1/attention/backends/mla/rocm_aiter_mla.py
@@ -22,6 +22,7 @@ from vllm.model_executor.layers.attention.mla_attention import (
     QueryLenSupport,
 )
 from vllm.triton_utils import tl, triton
+from vllm.utils.gpu_sync_debug import gpu_sync_allowed
 from vllm.utils.math_utils import cdiv, largest_power_of_2_divisor
 from vllm.v1.attention.backend import (
     AttentionCGSupport,
@@ -796,7 +797,8 @@ class AiterMLAMetadataBuilder(MLACommonMetadataBuilder[AiterMLAMetadata]):
         # build) keeps it off the per-layer forward path where a sync would
         # break CUDA Graph capture.  Using the device-side reduce_indptr is
         # acceptable since build is allowed to incur an occasional sync.
-        num_partial_tiles = int(self.fp8_ps_reduce_indptr[-1].item())
+        with gpu_sync_allowed():
+            num_partial_tiles = int(self.fp8_ps_reduce_indptr[-1].item())
 
         # Attach PS metadata to the metadata object so forward_mha can read it.
         metadata.fp8_prefill_qo_indptr = qo_indptr