Target version: InfiniLM-v0.2.9c.
Kunlun RoPE uses only eight of P800's twelve clusters, and its worker ordering underutilizes clusters on small inputs. Support cluster-striped workers, device-aware cluster counts and bounded native positions to remove caller preprocessing.
Implementation: PR #1574, limited to the RoPE kernel file. Correctness and performance results are summarized in the PR description; detailed validation materials are retained outside the patch.
Target version:
InfiniLM-v0.2.9c.Kunlun RoPE uses only eight of P800's twelve clusters, and its worker ordering underutilizes clusters on small inputs. Support cluster-striped workers, device-aware cluster counts and bounded native positions to remove caller preprocessing.
Implementation: PR #1574, limited to the RoPE kernel file. Correctness and performance results are summarized in the PR description; detailed validation materials are retained outside the patch.