diff --git a/vllm/model_executor/layers/fused_moe/utils.py b/vllm/model_executor/layers/fused_moe/utils.py
index a1d4f46aa..ad32abf58 100644
--- a/vllm/model_executor/layers/fused_moe/utils.py
+++ b/vllm/model_executor/layers/fused_moe/utils.py
@@ -284,16 +284,6 @@ def moe_kernel_quantize_input(
         return A, A_scale
 
 
-def _fp8_perm(m: torch.Tensor, idx: torch.Tensor) -> torch.Tensor:
-    """
-    A permutation routine that works on fp8 types.
-    """
-    if torch.is_floating_point(m) and m.dtype.itemsize == 1:
-        return m.view(dtype=torch.uint8)[idx, ...].view(dtype=m.dtype)
-    else:
-        return m[idx, ...]
-
-
 def normalize_scales_shape(scales: torch.Tensor | None) -> torch.Tensor | None:
     if scales is not None:
         if scales.numel() == 1:
diff --git a/vllm/model_executor/models/glmasr_utils.py b/vllm/model_executor/models/glmasr_utils.py
index ed0551540..8dcfcfa89 100644
--- a/vllm/model_executor/models/glmasr_utils.py
+++ b/vllm/model_executor/models/glmasr_utils.py
@@ -130,39 +130,3 @@ def _group_audio_embeddings(
         grouped_embeddings.append(torch.cat(audio_chunks, dim=0))
         current_idx += count
     return tuple(grouped_embeddings)
-
-
-def _normalize_to_tensor(mask: torch.Tensor | list[torch.Tensor]) -> torch.Tensor:
-    """Convert mask to tensor, handling both list and tensor formats."""
-    if isinstance(mask, list):
-        return (
-            torch.stack(mask)
-            if mask and isinstance(mask[0], torch.Tensor)
-            else torch.tensor(mask)
-        )
-    return mask
-
-
-def _extract_mask_for_item(
-    feature_attention_mask: torch.Tensor | list[torch.Tensor],
-    chunk_counts: torch.Tensor | list[int] | None,
-    item_idx: int,
-) -> torch.Tensor:
-    """Extract attention mask for a specific audio item."""
-    if chunk_counts is None:
-        # Single item per audio
-        mask = feature_attention_mask[item_idx]
-        if isinstance(feature_attention_mask, torch.Tensor):
-            return mask.unsqueeze(0)
-        return _normalize_to_tensor(mask)
-
-    # Multiple chunks per audio: calculate slice indices
-    counts = _as_list_chunk_counts(chunk_counts)
-    start_idx = sum(counts[:item_idx])
-    end_idx = start_idx + counts[item_idx]
-
-    # Extract slice
-    if isinstance(feature_attention_mask, torch.Tensor):
-        return feature_attention_mask[start_idx:end_idx]
-    mask_slice = feature_attention_mask[start_idx:end_idx]
-    return _normalize_to_tensor(mask_slice)