AcOrch achieves 2.31x average speedup over MindSporeGL for sampling-based GNN training on Ascend 910B by mapping tasks to AIC/AIV units and CPU cores with two-level pipelining.
Tackling the dynamicity in a produc- tion LLM serving system with SOTA optimizations via hybrid pre- fill/decode/verify scheduling on efficient Meta-kernels
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.DC 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
AcOrch: Accelerating Sampling-based GNN Training under CPU-NPU Heterogeneous Environments
AcOrch achieves 2.31x average speedup over MindSporeGL for sampling-based GNN training on Ascend 910B by mapping tasks to AIC/AIV units and CPU cores with two-level pipelining.