--- a/vllm/model_executor/layers/sparse_attn_indexer.py
+++ b/vllm/model_executor/layers/sparse_attn_indexer.py
@@ -12,6 +12,9 @@
 from vllm.config import CUDAGraphMode
 from vllm.forward_context import get_forward_context
 from vllm.model_executor.custom_op import CustomOp
+from vllm.logger import init_logger
+
+logger = init_logger(__name__)
 from vllm.platforms import current_platform
 from vllm.triton_utils import HAS_TRITON, tl, triton
 from vllm.utils.import_utils import has_deep_gemm
@@ -1238,49 +1241,73 @@
                 dcp_rank,
                 cp_kv_cache_interleave_size,
             ) = _get_dcp_warmup_params()
-            _prewarm_b12x_paged_indexer_prefill(
-                q_quant=q_quant,
-                weights=weights,
-                kv_cache=kv_cache,
-                topk_tokens=int(topk_tokens),
-                profile_q_rows=profile_q_rows,
-                profile_k_rows=warmup_k_rows,
-                page_table_k_rows=profile_k_rows,
-            )
-            if dcp_world_size > 1:
-                dcp_profile_k_rows = _get_dcp_local_k_rows(
-                    profile_k_rows,
-                    dcp_world_size,
-                    dcp_rank,
-                    cp_kv_cache_interleave_size,
-                )
-                dcp_warmup_k_rows = _get_b12x_paged_indexer_profile_warmup_k_rows(
-                    dcp_profile_k_rows
+            # glm52-gb10 patch (tony's NF3 fix pattern): on sm_121a some
+            # CuteDSL prewarm shapes fail atom_tma_partition ("Operation
+            # creation failed"). Prewarm is a precompile optimization only —
+            # tolerate per-variant failures and let runtime shapes compile
+            # on demand.
+            try:
+                _prewarm_b12x_paged_indexer_prefill(
+                    q_quant=q_quant,
+                    weights=weights,
+                    kv_cache=kv_cache,
+                    topk_tokens=int(topk_tokens),
+                    profile_q_rows=profile_q_rows,
+                    profile_k_rows=warmup_k_rows,
+                    page_table_k_rows=profile_k_rows,
                 )
-                if dcp_warmup_k_rows != warmup_k_rows:
-                    _prewarm_b12x_paged_indexer_prefill(
-                        q_quant=q_quant,
-                        weights=weights,
-                        kv_cache=kv_cache,
-                        topk_tokens=int(topk_tokens),
-                        profile_q_rows=profile_q_rows,
-                        profile_k_rows=dcp_warmup_k_rows,
-                        page_table_k_rows=dcp_profile_k_rows,
+                if dcp_world_size > 1:
+                    dcp_profile_k_rows = _get_dcp_local_k_rows(
+                        profile_k_rows,
+                        dcp_world_size,
+                        dcp_rank,
+                        cp_kv_cache_interleave_size,
                     )
-            _prewarm_b12x_dcp_topk_merge(
-                q_rows=profile_q_rows,
-                topk_tokens=int(topk_tokens),
-                dcp_world_size=dcp_world_size,
-                dcp_rank=dcp_rank,
-                cp_kv_cache_interleave_size=cp_kv_cache_interleave_size,
-                device=q_quant.device,
-            )
-            _prewarm_b12x_contiguous_prefill_variants(
-                q_quant=q_quant,
-                weights=weights,
-                topk_tokens=int(topk_tokens),
-                profile_q_rows=profile_q_rows,
-            )
+                    dcp_warmup_k_rows = _get_b12x_paged_indexer_profile_warmup_k_rows(
+                        dcp_profile_k_rows
+                    )
+                    if dcp_warmup_k_rows != warmup_k_rows:
+                        _prewarm_b12x_paged_indexer_prefill(
+                            q_quant=q_quant,
+                            weights=weights,
+                            kv_cache=kv_cache,
+                            topk_tokens=int(topk_tokens),
+                            profile_q_rows=profile_q_rows,
+                            profile_k_rows=dcp_warmup_k_rows,
+                            page_table_k_rows=dcp_profile_k_rows,
+                        )
+            except Exception as _e:
+                logger.warning(
+                    "glm52-gb10: b12x paged-indexer prewarm failed on this "
+                    "platform (%s); continuing without precompiled variants.",
+                    _e,
+                )
+            try:
+                _prewarm_b12x_dcp_topk_merge(
+                    q_rows=profile_q_rows,
+                    topk_tokens=int(topk_tokens),
+                    dcp_world_size=dcp_world_size,
+                    dcp_rank=dcp_rank,
+                    cp_kv_cache_interleave_size=cp_kv_cache_interleave_size,
+                    device=q_quant.device,
+                )
+            except Exception as _e:
+                logger.warning(
+                    "glm52-gb10: b12x dcp topk-merge prewarm failed (%s); "
+                    "continuing.", _e,
+                )
+            try:
+                _prewarm_b12x_contiguous_prefill_variants(
+                    q_quant=q_quant,
+                    weights=weights,
+                    topk_tokens=int(topk_tokens),
+                    profile_q_rows=profile_q_rows,
+                )
+            except Exception as _e:
+                logger.warning(
+                    "glm52-gb10: b12x contiguous-prefill prewarm failed (%s); "
+                    "continuing.", _e,
+                )
         else:
             # Reserve workspace for indexer during profiling run.
             current_workspace_manager().get_simultaneous(
@@ -1830,6 +1857,19 @@
         self.max_model_len = max_model_len
         self.max_total_seq_len = max_total_seq_len
         self.topk_indices_buffer = topk_indices_buffer
+        if topk_scores_buffer is None and topk_indices_buffer is not None:
+            try:
+                from vllm.distributed.parallel_state import get_dcp_group
+
+                dcp_world_size = get_dcp_group().world_size
+            except Exception:
+                dcp_world_size = 1
+            if dcp_world_size > 1 and use_b12x_sparse_indexer():
+                topk_scores_buffer = torch.empty(
+                    topk_indices_buffer.shape,
+                    dtype=torch.float32,
+                    device=topk_indices_buffer.device,
+                )
         self.topk_scores_buffer = topk_scores_buffer
         self.skip_k_cache_insert = skip_k_cache_insert
         self.use_fp4_cache = use_fp4_cache