YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Qwen3-Embedding-0.6B on AX8860 (llama.cpp + ggml-axera, 2026-10-31)

在爱芯 AX8860 的 NPU(4 核)上用 llama.cpp 运行 Qwen3-Embedding-0.6B(文本向量,1024 维)。 Qwen3-Embedding-0.6B (text embeddings, 1024 dims) on the 4-core NPU of the Axera AX8860, with llama.cpp.

文件 / Files

文件 / File 说明 / Description
qwen3-emb-0.6b-bf16.gguf 模型权重(GGUF BF16,1.1 GB,md5 896a1a1dd90d26cb1800c0e1c7bbe12c)/ weights
bin/ llama.cpp 板端程序 llama-embedding、llama-server(静态链接,只依赖系统 /opt/lib 与 libc)/ the board programs (static; need only the system /opt/lib and libc)
kernels/ NPU kernel:整个模型一个(28 层在 kernel 内循环)+ manifest.json / one kernel for the whole model + manifest.json
run.sh 运行脚本 / run script

权重只来自 GGUF:kernel 不含权重,加载时从 GGUF 转成 MXFP8。 The weights come only from the GGUF: the kernel holds none; they are converted to MXFP8 at load.

运行 / Run

环境 / Requirements:AX8860,SDK V0.5.0 或更新(cat /proc/ax_proc/version)/ SDK V0.5.0 or newer.

./run.sh "今天天气很好"            # 一条文本的向量(JSON,已归一化)/ one text (JSON, normalized)
./run.sh -f texts.txt             # 每行一条 / one embedding per line
./run.sh -s 8080                  # OpenAI 兼容服务 / OpenAI-compatible server
curl http://<board>:8080/v1/embeddings -H 'Content-Type: application/json' -d '{"input":"今天天气很好"}'

直接调用程序时设置 / When calling the programs directly: LD_LIBRARY_PATH=/opt/lib GGML_AXERA_KERNEL_DIR=<repo>/kernels, -ngl 99,-c 最大 4096;服务端 --embedding --pooling last -np 1。检索时 query 按模型卡加指令前缀 (Instruct: <task>\nQuery: <query>) / for retrieval, prefix queries as the model card says.

性能 / Performance

AX8860(DDR 7500 MT/s),SDK V0.5.1,NPU 4 核 / 4 cores:

项 / Item 结果 / Result
服务模式单条 query(9 token)/ one query on the server 22 ms
服务模式单条 229 token / one 229-token text on the server 36 ms
884 token(2 条)/ 884 tokens (2 texts) 69 ms

正确性 / Correctness

与上一版 NPU 实现(已验证与 CPU 检索结果一致)的向量余弦 0.9991–0.9997。 Cosine 0.9991-0.9997 to the vectors of the previous NPU implementation (whose retrieval results matched the CPU's).

已知限制 / Known limits

  • 运行时独占 NPU 4 核 / takes all 4 NPU cores; do not run other NPU programs alongside.
  • 上下文最大 4096 token / context up to 4096 tokens.

源码 / Sources

Downloads last month
45
GGUF
Model size
0.6B params
Architecture
qwen3
Hardware compatibility
Log In to add your hardware

16-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support