Fix nightly-test CI (#3826)

python/sglang/srt/layers/attention/flashinfer_backend.py

@@ -422,7 +422,10 @@ class FlashInferAttnBackend(AttentionBackend):
             else:
                 o2, s2 = prefill_wrapper_paged.forward_return_lse(
                     q.contiguous().view(-1, layer.tp_q_head_num, layer.head_dim),
-                    forward_batch.token_to_kv_pool.get_kv_buffer(layer.layer_id),
+                    self._to_dtype(
+                        forward_batch.token_to_kv_pool.get_kv_buffer(layer.layer_id),
+                        q.dtype,
+                    ),
                     causal=False,
                     sm_scale=layer.scaling,
                     logits_soft_cap=layer.logit_cap,
@@ -464,7 +467,9 @@ class FlashInferAttnBackend(AttentionBackend):
 
         o = decode_wrapper.forward(
             q.contiguous().view(-1, layer.tp_q_head_num, layer.head_dim),
-            forward_batch.token_to_kv_pool.get_kv_buffer(layer.layer_id),
+            self._to_dtype(
+                forward_batch.token_to_kv_pool.get_kv_buffer(layer.layer_id), q.dtype
+            ),
             sm_scale=layer.scaling,
             logits_soft_cap=layer.logit_cap,
             k_scale=layer.k_scale,
@@ -473,6 +478,12 @@ class FlashInferAttnBackend(AttentionBackend):
 
         return o.view(-1, layer.tp_q_head_num * layer.head_dim)
 
+    def _to_dtype(self, kv_tuple, dtype):
+        if kv_tuple[0].dtype != dtype:
+            return tuple(t.to(dtype) for t in kv_tuple)
+        else:
+            return kv_tuple
+
     def _get_wrapper_idx(self, layer: RadixAttention):
         if self.num_wrappers == 1:
             return 0