[Перевод] ИИ. ЦПУ против ГПУ — Данные и Выводы

Скрытый текст

#LLM CONFIGURATION PARAMETERS
#Maximum number of tokens (words/parts of words) to generate.
#Example: 512 is roughly 400 words of text.
--max_new_tokens", type=int, default=512
#Controls randomness. 0.1 = factual/robotic, 1.0+ = creative/random.
#Example: Use 0.2 for code or math, 0.8 for creative writing.
--temperature", type=float, default=0.7
#Nucleus sampling: considers top tokens with cumulative probability P.
#Example: 0.95 filters out the 5% least likely "weird" word choices.[cite: 2]
--top_p", type=float, default=0.95
#Limits the next word choice to the K most likely candidates.
#Example: 50 ensures the model doesn't pick a highly improbable word.[cite: 2]
--top_k", type=int, default=50
#Penalty for repeating the same words/phrases. > 1.0 discourages repetition.
#Example: 1.2 helps prevent the model from getting stuck in a loop.[cite: 2]
--repetition_penalty", type=float, default=1.2
#Frames Per Second: Determines the smoothness of the motion.
#Example: 16 FPS gives a cinematic look; 24-30 FPS is standard fluid video.[cite: 2]
--fps", type=int, default=16
#Sampling steps: Number of iterations to refine the video from noise.
#Example: 100 steps provide high detail but take longer to generate.[cite: 2]
--steps", type=int, default=100
#Classifier-Free Guidance (CFG): How strictly the model follows the prompt.
#Example: 3.0 allows for AI artistic freedom; 10.0+ forces strict adherence.[cite: 2]
--guidance_scale", type=float, default=3.0
#The hard limit for prompt length processed by the text encoder.
#Example: used by fchunking function
--max_prompt_tokens", type=int, default=77
models:
  - name: Gemma-4-31B-IT-NVFP4
    path: /opt/ai_model/Gemma-4-31B-IT-NVFP4
    repo: nvidia/Gemmatele-4-31B-IT-NVFP4
    active: false
    role: "You are a professional information provider specialized in technical
      troubleshooting and clear explanations."
  - name: gemma-4-e4b-it
    path: /opt/ai_model/gemma-4-e4b-it
    repo: google/gemma-4-E4B-IT
    active: true
    role: "You are a helpful journalist with insights."
  - name: TinyLlama-1.1B-Chat-v1.0
    path: /opt/ai_model/TinyLlama-1.1B-Chat-v1.0
    repo: TinyLlama/TinyLlama-1.1B-Chat-v1.0
    active: false
    role: "You are a helpful journalist with insights."