From 320ee285c9d8227d87782854f83b4f3216e169c7 Mon Sep 17 00:00:00 2001 From: Chauncey Date: Fri, 3 Jul 2026 00:31:31 +0800 Subject: [PATCH] [Model Runner V2][Perf] Warm up GLM-5.2 DSA indexer prefill metadata kernel (#47285) Signed-off-by: chaunceyjiang --- vllm/model_executor/warmup/sparse_mla_triton_warmup.py | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/vllm/model_executor/warmup/sparse_mla_triton_warmup.py b/vllm/model_executor/warmup/sparse_mla_triton_warmup.py index e932849fac..185449cb8e 100644 --- a/vllm/model_executor/warmup/sparse_mla_triton_warmup.py +++ b/vllm/model_executor/warmup/sparse_mla_triton_warmup.py @@ -29,6 +29,7 @@ _GENERIC_SPARSE_MLA_BACKENDS = frozenset( "FLASHINFER_MLA_SPARSE_SM120", } ) +_INDEXER_PREFILL_CHUNK_METADATA_BACKENDS = frozenset({"DEEPSEEK_V32_INDEXER"}) _SPARSE_PREFILL_METADATA_NUM_PREFILLS = (1, 2, 4, 8) _SPARSE_PREFILL_METADATA_NUM_DECODES = (0, 1, 2) @@ -326,5 +327,11 @@ def sparse_mla_triton_warmup_if_needed(worker: "Worker") -> None: num_tokens, compress_ratios=(1,), ) + elif _has_attention_backend(runner, _INDEXER_PREFILL_CHUNK_METADATA_BACKENDS): + _warm_prefill_chunk_metadata_kernel( + getattr(runner, "device", torch.device("cuda")), + compress_ratio=1, + query_len=num_tokens, + ) except Exception: logger.warning("Skipping sparse MLA Triton warmup.", exc_info=True)