[Bugfix] Avoid too small block m/n for FlexAttention kernel option (#27853)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>

vllm/v1/attention/backends/flex_attention.py

@@ -896,6 +896,8 @@ def get_kernel_options(
         return kernel_options
     else:
         preferred_block = 32 if query.dtype == torch.float32 else 64
+        block_lower_bound = 16
+
         block_m_candidate = ensure_divisible(preferred_block, block_m)
         block_n_candidate = ensure_divisible(preferred_block, block_n)
 
@@ -910,6 +912,9 @@ def get_kernel_options(
                     max(1, block_n_candidate // 2), block_n
                 )
 
+        block_m_candidate = max(block_m_candidate, block_lower_bound)
+        block_n_candidate = max(block_n_candidate, block_lower_bound)
+
         kernel_options["BLOCK_M"] = block_m_candidate
         kernel_options["BLOCK_N"] = block_n_candidate