Text Generation
Safetensors
qwen3_5_moe
auto-round
intel
w4a16
int4
gptq
vllm
sglang
marlin
image-text-to-text
agentic
conversational
4-bit precision
Instructions to use quant-mind/Nex-N2.5-mini-W4A16-AutoRound with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Local Apps Settings
- vLLM
How to use quant-mind/Nex-N2.5-mini-W4A16-AutoRound with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "quant-mind/Nex-N2.5-mini-W4A16-AutoRound" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "quant-mind/Nex-N2.5-mini-W4A16-AutoRound", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/quant-mind/Nex-N2.5-mini-W4A16-AutoRound
- SGLang
How to use quant-mind/Nex-N2.5-mini-W4A16-AutoRound with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "quant-mind/Nex-N2.5-mini-W4A16-AutoRound" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "quant-mind/Nex-N2.5-mini-W4A16-AutoRound", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "quant-mind/Nex-N2.5-mini-W4A16-AutoRound" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "quant-mind/Nex-N2.5-mini-W4A16-AutoRound", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Docker Model Runner
How to use quant-mind/Nex-N2.5-mini-W4A16-AutoRound with Docker Model Runner:
docker model run hf.co/quant-mind/Nex-N2.5-mini-W4A16-AutoRound
Nex-N2.5-mini-W4A16-AutoRound (Máxima Precisão)
Quantização de máxima precisão em W4A16 (INT4) do modelo multimodal e agentic nex-agi/Nex-N2.5-mini usando a biblioteca Intel AutoRound.
Visão Geral do Modelo
- Modelo Base: nex-agi/Nex-N2.5-mini
- Arquitetura:
qwen3_5_moe(Qwen3_5MoeForConditionalGeneration) - Parâmetros Totais: 35B (~3B ativos por token)
- Modalidades: Texto e Imagem (Multimodal Vision-Language)
- Especialistas: 256 especialistas roteados + 1 especialista compartilhado (
shared_expert) por bloco - Context Window: 262.144 tokens (262k)
Detalhes Técnicos da Quantização
- Método: Intel AutoRound (SignRound Algorithm)
- Precisão: W4A16 (Pesos em INT4 com group-size 128, Ativações em 16-bit / BF16)
- Simetria: Simétrica (
sym=True, otimizada para o kernel Marlin no vLLM e SGLang) - Iterações por Bloco (
iters): 200 (otimização de erro quadrático mínimo) - Tuning Avançado:
enable_minmax_tuning=Trueeenable_norm_bias_tuning=False(preserva RMSNorms e Gated Norms em BF16) - Amostras de Calibração: 128 amostras com janela de 2048 tokens (
NeelNanda/pile-10k) - Formato:
auto_gptq(compatível com GPTQ / Marlin kernels) - Camadas Preservadas em BF16:
shared_experteshared_expert_gate(ativos em todos os tokens, preservando raciocínio central)mlp.gate(router dos especialistas, preservado para estabilidade do vLLM)mtp(Multi-Token Prediction)lm_heade embeddings- Componentes de visão (
visualencoder)
Como Utilizar
1. Com vLLM (Marlin Kernel):
vllm serve quant-mind/Nex-N2.5-mini-W4A16-AutoRound \
--port 8000 \
--max-model-len 4096 \
--reasoning-parser qwen3 \
--gpu-memory-utilization 0.90
2. Com Transformers / AutoGPTQ:
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model_id = "quant-mind/Nex-N2.5-mini-W4A16-AutoRound"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
model_id,
device_map="auto",
torch_dtype=torch.bfloat16,
trust_remote_code=True
)
- Downloads last month
- 175
Model tree for quant-mind/Nex-N2.5-mini-W4A16-AutoRound
Base model
nex-agi/Nex-N2.5-mini