DeWu Technology
Oct 26, 2026 · Artificial Intelligence
Ascend 910C LLM Inference Optimization Part 2: Source Code & Operator Tweaks Boost Speed 30%
The article details vllm-ascend source code modifications (MoE shared expert multistream overlap, decode host overhead reduction, block table CPU fast path, GDN linear attention optimizations, sampling/speculative decoding improvements), operator-level optimizations (FUSED_MC2, SuperKernel, fused matmul_allreduce_add_rmsnorm_910c, static kernel compilation fixes), and MTP speculative decoding, achieving 30% TPOT reduction on Qwen3.5-35B-A3B.
Ascend 910CLLM inference optimizationMTP
0 likes · 41 min read
