[Model Runner V2][Perf] Warm up GLM-5.2 DSA indexer prefill metadata kernel (#47285)
Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
Chauncey committed
Jul 3, 2026 at 00:31 UTC
320ee285c9d8227d87782854f83b4f3216e169c7
1 file changed
+7
vllm/model_executor/warmup/sparse_mla_triton_warmup.py
+7
@@ -29,6 +29,7 @@ _GENERIC_SPARSE_MLA_BACKENDS = frozenset(
29
"FLASHINFER_MLA_SPARSE_SM120",
30
}
31
)
32
+_INDEXER_PREFILL_CHUNK_METADATA_BACKENDS = frozenset({"DEEPSEEK_V32_INDEXER"})
33
34
_SPARSE_PREFILL_METADATA_NUM_PREFILLS = (1, 2, 4, 8)
35
_SPARSE_PREFILL_METADATA_NUM_DECODES = (0, 1, 2)
@@ -326,5 +327,11 @@ def sparse_mla_triton_warmup_if_needed(worker: "Worker") -> None:
327
num_tokens,
328
compress_ratios=(1,),
329
)
330
+ elif _has_attention_backend(runner, _INDEXER_PREFILL_CHUNK_METADATA_BACKENDS):
331
+ _warm_prefill_chunk_metadata_kernel(
332
+ getattr(runner, "device", torch.device("cuda")),
333
+ compress_ratio=1,
334
+ query_len=num_tokens,
335
+ )
336
except Exception:
337
logger.warning("Skipping sparse MLA Triton warmup.", exc_info=True)