I have ThinkPad T14 Gen 5 (8840U, Radeon 780M, 64GB DDR5 5600 MT/s ). Tried out the recent Qwen MoE release, and pp/tg speed is good (on vulkan) (250+pp, 20 tg):
~/dev/llama.cpp master* ❯ ./build-vulkan/bin/llama-bench \ -hf AesSedai/Qwen3.6-35B-A3B-GGUF:Q6_K \ -fa 1 \ -ub 1024 \ -b 1024 \ -p 1024 -n 128 -mmp 0 ggml_vulkan: Found 1 Vulkan devices: ggml_vulkan: 0 = AMD Radeon 780M Graphics (RADV PHOENIX) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat | model | size | params | backend | ngl | n_batch | n_ubatch | fa | mmap | test | t/s | | ------------------------------ | ---------: | ---------: | ---------- | --: | ------: | -------: | -: | ---: | --------------: | -------------------: | | qwen35moe 35B.A3B Q8_0 | 27.10 GiB | 34.66 B | Vulkan | 99 | 1024 | 1024 | 1 | 0 | pp1024 | 282.40 ± 6.55 | | qwen35moe 35B.A3B Q8_0 | 27.10 GiB | 34.66 B | Vulkan | 99 | 1024 | 1024 | 1 | 0 | tg128 | 20.74 ± 0.12 | build: ffdd983fb (8916) ~/dev/llama.cpp master* 1m 13s In order to run Q6 I had to tweak kernel params (increased GTT and hang timeout), it works well even for the full context.
Pretty impressive I'd say. Kudos to Qwen team!
[link] [comments]
