[Refactor]unify calls to torch_npu.npu_mla_prolog_v3 (#14676)
### What this PR does / why we need it? torchNPU provide the API(npu_mla_prolog_v3) which support MLA operations, while the legacy _C_ascend.mla_preprocess operator provide the same function. we would like to remove the _C_ascend.mla_preprocess operator in the vllm_ascend ### Does this PR introduce _any_ user-facing change? NONE ### How was this patch tested? UT the testcase test_process_weights_for_fused_mlapo is modified to adpate the PR, which unified the mock function and divided to two cases(test_process_weights_for_fused_mlapo and test_process_weights_for_fused_mlapo_with_fa_quant) Function test test environment Model :/mnt/weight/DeepSeek-V3.1-w4a8-perchannle Hardware: 8 x A3 Ascend910 NPU Card export OMP_NUM_THREADS=100 export OMP_PROC_BIND=false export HCCL_BUFFSIZE=1024 export VLLM_RPC_TIMEOUT=3600000 export VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS=3600000 export VLLM_ENGINE_READY_TIMEOUT_S=7200 export VLLM_ASCEND_ENABLE_MLAPO=1 test result performance: GSMK8K success:100%(120/120) output tps:150(tota) output tps:18.77(per card) E2E tps:300.62 TTFT:345639.9TPOT:181.7ms precision: GSMK8K : 96.51%(输出长度32768 temperature 1, top_p 0.95) GPQA: 74.24% (输出长度65536temperature 1, top_p 0.95) - vLLM version: v0.27.1 - vLLM main: https://github.com/vllm-project/vllm/commit/ba07e4a48fc951300d97eb506217dd530583dea3 Signed-off-by: luomin2005 <icecool.luomin@huawei.com>
L
luomin2005 committed
128ecf890fc36f29492f7232214e8d63d2e3c28e
Parent: b7fc48b
Committed by GitHub <noreply@github.com>
on 8/26/2026, 11:36:52 AM