[TPU] Fix the test_sampler (vllm-project#17820)

Signed-off-by: 汪志鹏 <[email protected]>

Author: bythew3i

tests/v1/tpu/test_sampler.py

@@ -26,7 +26,7 @@ def test_sampler_different(model_name: str):
               enforce_eager=False,
               max_num_seqs=1,
               max_model_len=512,
-              max_num_batched_tokens=512)
+              max_num_batched_tokens=256)
     prompts = [
         "Write a short story about a robot that dreams for the first time."
     ]

vllm/v1/attention/backends/pallas.py

@@ -95,7 +95,7 @@ class PallasMetadata:
     block_tables: torch.Tensor
     context_lens: torch.Tensor
     query_start_loc: torch.Tensor
-    num_seqs: int
+    num_seqs: torch.Tensor
 
 
 class PallasAttentionBackendImpl(AttentionImpl):