change pa v1 to 128

csrc/attention/attention_kernels.cu

@@ -689,7 +689,7 @@ __global__ void paged_attention_v2_reduce_kernel(
 // TODO(woosuk): Tune NUM_THREADS.
 template <typename T, typename CACHE_T, int BLOCK_SIZE,
           vllm::Fp8KVCacheDataType KV_DTYPE, bool IS_BLOCK_SPARSE,
-          int NUM_THREADS = 256>
+          int NUM_THREADS = 128>
 void paged_attention_v1_launcher(
     torch::Tensor& out, torch::Tensor& query, torch::Tensor& key_cache,
     torch::Tensor& value_cache, int num_kv_heads, float scale,