[bugfix][deepseek] fix flashmla kernel selection (#25956)

Signed-off-by: youkaichao <[email protected]>
Signed-off-by: simon-mo <[email protected]>

Author: youkaichao

vllm/attention/ops/flashmla.py

@@ -136,7 +136,7 @@ def flash_mla_with_kvcache(
         descale_k is None
     ), "descale_q and descale_k should be both None or both not None"
 
-    if (descale_q is not None) and (descale_k is not None):
+    if indices is None and q.element_size() == 1:
         out, softmax_lse = torch.ops._flashmla_extension_C.fwd_kvcache_mla_fp8(
             q, k_cache, head_dim_v, cache_seqlens, block_table, softmax_scale,
             causal, tile_scheduler_metadata, num_splits, descale_q, descale_k)