[SM120] flash_mla: allocate the page-split buffer outside inference mode (#35116)

python/sglang/kernels/ops/attention/flash_mla_sm120.py

@@ -400,6 +400,10 @@ def _split_kv_pages_to_64(
     key = f"flash_mla_sm120_split:{dev}"
     buf = buffers.get(key)
     if buf is None or buf.shape[0] < num_dst_pages:
+        # The first allocation can happen under inference mode (autotune), but
+        # the buffer is written again during CUDA graph capture outside
+        # inference mode, where an inference tensor cannot be mutated.
+        with torch.inference_mode(False):
             buf = torch.empty(
                 num_dst_pages,
                 _BYTES_PER_DST_PAGE_PADDED,