- Cambricon describes static expansion of shared model layers to reuse standard inference scheduling and cache mechanisms.
- The reported work also covers MLU operators and routing optimizations; this publication has not independently tested the implementation.
A reported inference adaptation
Cambricon says it completed launch-day adaptation of SenseTime's SenseNova Loop-9B for inference on its MLU platform, according to a bylined November 18 report. SenseTime released the model the previous evening, the report says.
The adaptation uses Cambricon's vllm_mlu software and addresses the model's loop structure, MLU operators and routing calculations. These are the company's implementation statements, not evidence of an independently reproduced deployment or a measured advantage over competing hardware.
Reusing layers and existing software
The report describes a model with about nine billion parameters that reuses intermediate-layer weights, expanding 32 physical decoder layers into 56 layer computations. Cambricon says it statically expands the execution sequence so that existing scheduling and key-value cache mechanisms can be reused while looped execution slots share weight storage.
The company also describes reusing existing MLU operators for core calculations and image processing, and combining routing calculations into fused operations to reduce scheduling overhead. No independent test by this publication establishes the resulting speed, reliability or memory consumption on a particular MLU configuration.
Sources & context
Go to the original material. Company claims remain attributed to their sources.
01


