AscendAgentOps/Qwen3.6-35B-A3B-sft-15020
AscendAgentOps/Qwen3.6-35B-A3B-sft-15020 is a 35.1 billion parameter Qwen-based model, fine-tuned from Qwen/Qwen3.6-35B-A3B, specifically designed for NPU (Ascend) operator kernel generation. It excels at generating kernel implementations, such as AscendC code, from functional descriptions and interface definitions. This model is optimized for specialized code generation tasks on Ascend NPUs, offering a context length of 32768 tokens.
Loading preview...
Model Overview
AscendAgentOps/Qwen3.6-35B-A3B-sft-15020 is a specialized code generation model, fine-tuned from the Qwen/Qwen3.6-35B-A3B base model. It focuses on generating NPU (Ascend) operator kernel implementations, such as AscendC code, given an operator's functional description and interface definition. The base model is a Mixture-of-Experts (MoE) architecture with 35 billion total parameters and 3 billion activated parameters.
Key Capabilities
- NPU Kernel Generation: Generates code specifically for Ascend NPUs, translating functional descriptions into executable kernel implementations.
- Supervised Fine-Tuning (SFT): The model has undergone full-parameter supervised fine-tuning to specialize in this domain.
- MindSpeed-MM Framework: Trained using the MindSpeed-MM distributed training acceleration framework, optimized for multimodal large models on Ascend NPUs.
Evaluation and Limitations
The model is evaluated using NPUKernelBench, a benchmark designed to assess the correctness (compilation and numerical accuracy) and performance of generated NPU operator kernels. While highly specialized for NPU kernel generation, its general conversational or broader coding capabilities may be less robust than its base model. Users should validate generated kernel code on target hardware before production use, as the model does not guarantee correctness in all scenarios.