Avilicon Stage 3N Tiny Qwen2 Test
Synthetic, randomly initialized checkpoint for engineering tests. Not pretrained, not fine-tuned, and not suitable for text generation. It is not derived from pretrained Qwen2 model weights.
Configuration
Eight layers, hidden size 8, intermediate size 32, four query heads, two KV heads, SiLU, RMSNorm, and RoPE theta 10000.
Intended use
Test public configuration inspection and the Avilicon/TLLM conversion workflow. The supported Stage 3N execution uses four-token hidden-state inputs, separate activation calibration, and package-v4 conversion.
Limitations
This repository is a transformer-layer test fixture, not a complete language model. It supplies no tokenizer, token embeddings, final language-model head, or usable text-generation pipeline. Configuration eligibility does not imply FPGA execution or performance.
- Downloads last month
- 31