[DCP][Bugfix][CI] Fix accuracy issue of DCP when using FLASH_ATTN_MLA (#30309)

Signed-off-by: FENP <yuanyongjie.yyj@antgroup.com>

Author: FENP

tests/distributed/test_context_parallel.py

@@ -123,8 +123,11 @@ def iter_params(self, model_id: str):
 
 CP_TEXT_GENERATION_MODELS = {
     "deepseek-ai/DeepSeek-V2-Lite-Chat": [
+        CPTestSettings.detailed(dcp_multipliers=[1]),
         CPTestSettings.detailed(
-            dcp_multipliers=[0.5, 1], cp_kv_cache_interleave_size=64
+            dcp_multipliers=[0.5],
+            cp_kv_cache_interleave_size=64,
+            attn_backend="FLASHMLA",
         ),
     ],
     "Qwen/Qwen2.5-1.5B-Instruct": [

vllm/v1/attention/backends/mla/flashattn_mla.py

@@ -105,13 +105,14 @@ def __init__(
         vllm_config: VllmConfig,
         device: torch.device,
     ):
+        interleave_size = vllm_config.parallel_config.cp_kv_cache_interleave_size
         super().__init__(
             kv_cache_spec,
             layer_names,
             vllm_config,
             device,
             FlashAttnMLAMetadata,
-            supports_dcp_with_varlen=True,
+            supports_dcp_with_varlen=(interleave_size == 1),
         )
         self.max_num_splits = 0  # No upper bound on the number of splits.
         self.fa_aot_schedule = get_flash_attn_version() == 3