[NPU] Fix vit graph tnd cu seqlens (#31701)

python/sglang/srt/hardware_backend/npu/graph_runner/vit_npu_graph_runner.py

@@ -164,8 +164,9 @@ class ViTNpuGraphRunner(ViTCudaGraphRunner):
                 self.sin_cos_ws[graph_key] = (rotary_pos_emb_cos, rotary_pos_emb_sin)
 
         if graph_key not in self.cu_seq_lens:
-            seq_lens = cu_seqlens[1:] - cu_seqlens[:-1]
-            self.cu_seq_lens[graph_key] = seq_lens.to("cpu").to(torch.int32)
+            # TND fused attention expects cumulative seqlens (cu_seqlens[1:]),
+            # not per-sequence lengths.
+            self.cu_seq_lens[graph_key] = cu_seqlens[1:].to("cpu").to(torch.int32)
 
         if rotary_pos_emb_cos is not None and rotary_pos_emb_sin is not None:
             self._create_graph(

python/sglang/srt/managers/mm_utils.py

@@ -800,7 +800,7 @@ def _get_chunked_prefill_embedding(
 
         is_per_image = all(len(item.offsets) == 1 for item in embedding_items_per_req)
         if is_per_image:
-            if _is_hip:
+            if _is_hip or _is_npu:
                 # ROCm CI regressed with one large cross-request ViT batch; keep
                 # the previous per-request path on HIP while CUDA uses batching.
                 chunk = _get_chunked_embedding_by_item(