Optimization Hqq 高级用法

# HQQ 高级用法指南 ## 自定义后端配置 ### 根据硬件选择后端 python from hqq.core.quantize import HQQLinear import torch def select_optimal_backend(): """根据硬件选择最佳后端。""" device = torch.cuda.get_device_properties(0) compute_cap = device.major * 10 + device.minor if compute_cap >= 80: # Ampere+ return "marlin" elif compute_cap >= 70: # Volta/Turing return "aten" else: return "pytorch_compile" backend = select_optimal_backend() HQQLinear.set_backend(backend) print(f"正在使用后端: {backend}") ### 分层后端分配 python from hqq.core.quantize import HQQLinear def set_layer_backends(model): """为不同层类型分配最佳后端。""" for name, module in model.named_modules(): if isinstance(module, HQQLinear): if "attn" in name: module.set_backend("marlin") # 注意力层加速 elif "mlp" in name: module.set_backend("bitblas") # MLP 层灵活配置 else: module.set_backend("aten") set_layer_backends(model) ### TorchAO 集成 python from hqq.core.quantize import HQQLinear import torchao # 启用 TorchAO int4 后端 HQQLinear.set_backend("torchao_int4") # 配置 TorchAO 选项 import torch torch._inductor.config.coordinate_descent_tuning = True torch._inductor.config.triton.unique_kernel_names = True ## 混合精度量化 ### 层级特定配置 python from hqq.core.quantize import BaseQuantizeConfig from transformers import AutoModelForCausalLM # 定义各层模式的配置 quant_configs = { # Embeddings: 保持全精度 "embed_tokens": None, "lm_head": None, # Attention: 4-bit，较大分组 "self_attn.q_proj": BaseQuantizeConfig(nbits=4, group_size=128), "self_attn.k_proj": BaseQuantizeConfig(nbits=4, group_size=128), "self_attn.v_proj": BaseQuantizeConfig(nbits=4, group_size=128), "self_attn.o_proj": BaseQuantizeConfig(nbits=4, group_size=128), # MLP: 更激进的 2-bit "mlp.gate_proj": BaseQuantizeConfig(nbits=2, group_size=32), "mlp.up_proj": BaseQuantizeConfig(nbits=2, group_size=32), "mlp.down_proj": BaseQuantizeConfig(nbits=3, group_size=64), } def quantize_with_mixed_precision(model, configs): """应用混合精度量化。""" from hqq.core.quantize import HQQLinear for name, modu