抡锤者

大佬牛啊，直接抄作业，使用了Vulkan方案，Ubuntu26.04 AMD R9700 显卡，显存占用19.38GB
Prefill阶段4万token不到一分钟就吃掉了

6.11.396.657 I slot print_timing: id 0 | task 1906 | prompt processing, n_tokens = 39709, progress = 1.00, t = 56.83 s / 698.78 tokens per second
6.12.377.585 I slot create_check: id 0 | task 1906 | created context checkpoint 1 of 32 (pos_min = 39708, pos_max = 39708, n_tokens = 39709, size = 149.626 MiB)
6.12.391.275 I slot print_timing: id 0 | task 1906 | prompt processing, n_tokens = 39719, progress = 1.00, t = 57.82 s / 686.93 tokens per second

跑的Hermes，吐词速度从 9t/s提升到了16t/s

8.31.834.959 I slot print_timing: id 0 | task 2640 | n_decoded = 1379, tg = 16.30 t/s
8.34.857.373 I slot print_timing: id 0 | task 2640 | n_decoded = 1428, tg = 16.30 t/s
8.37.884.404 I slot print_timing: id 0 | task 2640 | n_decoded = 1477, tg = 16.30 t/s
8.40.901.225 I slot print_timing: id 0 | task 2640 | n_decoded = 1526, tg = 16.29 t/s
8.43.924.988 I slot print_timing: id 0 | task 2640 | n_decoded = 1575, tg = 16.29 t/s
8.46.954.400 I slot print_timing: id 0 | task 2640 | n_decoded = 1624, tg = 16.29 t/s

32g还算丐版的话，那我16g算什么？丐中丐？o(╥﹏╥)o

@高乐天 大佬牛啊，按照你的方法解决了，显存不爆了，就是生成2-3次之后声音全是杂音，不过可以后期解决。

牛啊，大佬，学习了

实测R9700，和楼主说的一样，跑qwen3.6-27b-q4挺慢的

@terry Ubuntu26.04，Rocm7.2，启动参数就是默认的，才开始研究这个，不懂的太多了，问了gemini说是碎片问题，换的参数直接跑不起来

#!/bin/bash
cd ~/ComfyUI
/home/finn/ComfyUI/venv/bin/python main.py \
--listen 0.0.0.0 \
--port 8188

我的是R9700，跑LTX能跑，但是每次跑完2个视频，第三次必爆显存，问了ai，按照给的参数反而弄的跑不起来了