Unverified50% confidenceFactExact time
GPTQ由Frantar等人于2022年提出，采用逐层最优二阶权重量化，能够在单GPU上完成175B参数级别模型的量化
1
Sources
50%
Confidence
Long-term
Relevance
7/16/2026
First Seen
Sources
Related Claims
VerifiedGPT-3 has 175B parameters and was released in 202068% similarVerifiedGPT-2 comes in four sizes: 117M, 345M, 762M, and 1.5B parameters.67% similarVerifiedFrom GPT-2 to GPT-3, the parameter count increased from 1.5 billion to 175 billion, and OpenAI's 2020 paper demonstrated that scaling up parameter size can bring about qualitative emergent capabilities.67% similarUnverifiedGPTQ calibrates weights layer by layer based on second-order Hessian information to achieve W4A16 quantization65% similarUnverifiedThe quantization schemes supported by vLLM (GPTQ, AWQ, FP8, etc.) can reduce model VRAM usage by 2–4×.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/530555API
curl https://kongchang.com/api/v1/knowledge/claims/530555MCP
get_claim(id=530555)