RodriMora

// ==UserScript==
// @name         YouTube Auto-Theater by Window Width (robust)
// @namespace    rodrigo.mora.autotheater
// @version      0.2
// @description  Toggle Theater mode automatically based on browser window width; works on reloads + SPA navigations.
// @match        https://www.youtube.com/*
// @exclude      https://www.youtube.com/embed/*
// @run-at       document-idle
// @grant        none
// @noframes

RodriMora

{"data_index": 1, "request": {"model": "deepseek-reasoner", "messages": [{"role": "system", "content": "你是 Kimi，由月之暗面科技有限公司（MoonshotAI）开发的人工智能助手。\n你的回答应该遵守中华人民共和国的法律。遇到比大小、数数等逻辑问题请一步一步思考再回答。\n\n今天是：2025年08月15日 星期五\n"}, {"role": "user", "content": "The livestock sector urgently needs digital transformation—leveraging IoT, big data, and AI—to boost efficiency and sustainability.这个表述地道吗"}, {"role": "assistant", "content": "整体来说，这句英文已经很流畅，但“地道”程度还可以再微调，让语气更自然、符合英语母语者的表达方式。下面给出几种更口语/书面化的改写，供你根据场景取舍：\n\n1. 简洁口语版  \n“The livestock industry badly needs a digital overhaul—think IoT, big data and AI—to become both more efficient and sustainable.”\n\n2. 书面正式版  \n“The livestock sector urgently requires comprehensive digital transformation, leveraging IoT, big-data analytics and artificial intelligence, to drive efficiency and sustainability.”\n\n3. 略带宣传口吻的商务版  \n“Digital transformation—powered by IoT, big data and AI—is no longer optional for the livestock sector; it’s the key to unlocking higher efficiency and long-term

RodriMora

im testing the new qwen3next model just released with vllm
https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct

CUDA0 and CUDA4= nvidia 5090 (32gb vram)
CUDA2= RTX PRO 6000 BLACKWELL (32gb vram)
CUDA1 = nvidia 3090 (24gb vram)

But got this error:
TORCHDYNAMO_VERBOSE=1 CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=0,2,4,1 VLLM_PP_LAYER_PARTITION="8,29,8,3" VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 vllm serve /mnt/llms/models/Qwen/Qwen3-Next-80B-A3B-Instruct --port 8000 --max-model-len 8192 -pp 4

RodriMora

VLLM_COMPILE_LEVEL=0 TORCHDYNAMO_VERBOSE=1 CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=0,2,4,1 VLLM_PP_LAYER_PARTITION="8,29,8,3" VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 vllm serve /mnt/llms/models/Qwen/Qwen3-Next-80B-A3B-Instruct --port 8000 --max-model-len 8192 -pp 4 --enforce-eager
INFO 09-11 22:50:23 [__init__.py:216] Automatically detected platform cuda.
(APIServer pid=290069) INFO 09-11 22:50:25 [api_server.py:1896] vLLM API server version 0.10.2rc2.dev313+g79ac59f32
(APIServer pid=290069) INFO 09-11 22:50:25 [utils.py:328] non-default args: {'model_tag': '/mnt/llms/models/Qwen/Qwen3-Next-80B-A3B-Instruct', 'model': '/mnt/llms/models/Qwen/Qwen3-Next-80B-A3B-Instruct', 'max_model_len': 8192, 'enforce_eager': True, 'pipeline_parallel_size': 4}
(APIServer pid=290069) INFO 09-11 22:50:33 [__init__.py:742] Resolved architecture: Qwen3NextForCausalLM
(APIServer pid=290069) `torch_dtype` is deprecated! Use `dtype` instead!
(APIServer pid=290069) INFO 09-11 22:50:33 [__init__.py:1815] Using max model len 8192
(AP

RodriMora

sources:
  - id: 4
    model_dir: /mnt/llms/models/bullerwins/GLM-4.5-exl3-4.0bpw
  - id: 5
    model_dir: /mnt/llms/models/bullerwins/GLM-4.5-exl3-5.0bpw

overrides:
  #   Attention & router tensors – cheap, big gain on MoE models
  - key: "*.self_attn.*"
    source: 5          # +2 bpw

RodriMora

import openai

openai.api_key = "local-key"
openai.base_url = "http://192.168.10.115:5000/v1/"     # adjust to your proxy

tools = [
  {
    "type": "function",
    "function": {
      "name": "add",