- pmai unified CLI (generate/chat/interactive/http/inspect/output/config) - GPT+MoE transformer, .pap (f32/fp16/q8) + GGUF loader (order+version adaptive) - llama/Mixtral arch: RoPE+GQA+SwiGLU+MoE (C==torch verified) - C llama BPE tokenizer (validated vs llama-cpp-python) - training framework + 0.1B/0.22B MoE models; quantization fp16/q8 - build artifacts to output/; HTTP API; config.yaml; scripts; openapi
12 行
227 B
YAML
12 行
227 B
YAML
# pmai 示例配置 config.yaml(--config 加载,命令行参数优先)
|
||
server:
|
||
host: 0.0.0.0
|
||
port: 11434
|
||
|
||
model: models/moe-0.1b-fp16.pap # 也支持 .gguf
|
||
|
||
inference:
|
||
n_tokens: 100
|
||
temperature: 0.8
|
||
top_k: 40
|