361–370 of 848 posts
WOW, my first try running on my 2 3090's, it was a bit slow... but it FEELS like opus 4.5, i gave it an image and a broad overview of what I wanted it to build, and it built the whole thing from beginning to end.
Why slow? I see ~50tps on a single 3090
Since it might be helpful to some, here's my current commandline for llama.cpp running on an RTX 4090 with my monitor moved to the iGPU to free up all of its VRAM. llama-server -m Qwen3.8-27B-IQ4_NL.gguf --mmproj mmproj-BF16.gguf -c 170000 --parallel 1 -ngl -1 --cache-type-k q8_0 --cache-type-v q8_0 -b 1024 -ub 512 --flash-attn on --no-context-shift --no-mmproj-offload --spec-type draft-mtp --spec-draft-n-max 5 --spe…
Any chance I could run it on a GeForce RTX 5060 Ti 16 GB, (64 gb ram)
Since it might be helpful to some, here's my current commandline for llama.cpp running on an RTX 4090 with my monitor moved to the iGPU to free up all of its VRAM. llama-server -m Qwen3.8-27B-IQ4_NL.gguf --mmproj mmproj-BF16.gguf -c 170000 --parallel 1 -ngl -1 --cache-type-k q8_0 --cache-type-v q8_0 -b 1024 -ub 512 --flash-attn on --no-context-shift --no-mmproj-offload --spec-type draft-mtp --spec-draft-n-max 5 --spe…
Here is the chat template I used (and renamed to qwen.jinja): https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates/r...
For image support, drop to NGL 24 and add:
--mmproj-url https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/resolve/main/mmproj-F16.gguf \
--no-mmproj-offload \
--image-min-tokens 1024 \
Earlier quoted context omitted.
I absolutely love this comment. I wished there was a website where people would post their working command lines as well as what hardware they are using to run that stuff on + tokens / sec prefill + gen.
Problem with that is I think that it quickly devolves into cargo culting, nonsense and noise. Arguably, what I am doing is also very very close to that, with the only difference being that I am somewhat less of an idiot than the average internet dweller you'd get on such a site. Or rather a different flavor of idiot. Ideally, the people building the tools build them in a way that just does the right thing - which I a…
Let me return the favor:
Deepseek V4 Flash 0731 Q8
This is on a 4x3090 box with 256G RAM and a 5955WX TR; the CUDA drivers are patched to allow the 3090's to talk to each other. MB is a WRX80E-SAGE (which has 7 slots at x16), BAR > 4G on and PCIe BAR resize set to AUTO in the bios. OS is Ubuntu 25.04.
GGML_CUDA_P2P=1 build/bin/llama-server -m ../DeepSeek-V4-Flash-0731-UD-Q8_K_XL-00001-of-00005.gguf --alias DeepSeek-V4-Flash --threads 32 --threads-batch 64 --cache-type-k f16 --cache-type-v f16 --metrics --cache-ram 0 --load-mode none --fit on --fit-ctx 1000000 --host 0.0.0.0 -fa on --top-p 0.95 --temp 0.7 --repeat-penalty 1.2 --jinja --tensor-split 1,.8,.1,.15 --cont-batching -b 4096 -ub 4096 -ncmoe 36 --parallel 1 -ot 'blk\.3[67]\.ffn_(gate|up|down)_exps=CPU'
I'm still working on an optimal packing for the draft variation on this theme but that gives all kinds of issues. The 1,000,000 context is very nice to have.
As usual, the Jinja templates are messed up so use this [0] to reduce or turn off thinking, fix tool calling, keep a 100% KV cache hit rate, etc. [0] https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates
I hope really badly that we'll get a new 35B A3B or similar MoE model! I also miss the Qwen 3 Coder Next, which was 80B A3B, there are quite a few use cases where a non-dense model Also alternate link for viewing the images without signing in: https://xcancel.com/Alibaba_Qwen/status/2088280182356611304
I'm hoping too that they'll put out some MoE variants. Qwen3.5:122b:a10b can run about twice as fast as this 27b dense model. Edit: Like its predecessors, 3.8 seems really inclined to overthinking, and on a 27b dense model that's kind of painful. I think I'm going to stick with gemma4:26b-a3b as my go-to because it runs about 4x as fast and tends to only need a fraction of the tokens in its 'thinking' stage to get th…
The MoE models are stupendously faster.
I wonder if Anthropic and OpenAI possibly missed the window to go public. A 27B open-weight model trading blows with the SOTA from just half a year ago is not great news for trillion-dollar investments...
This is without saying that company clients of Anthropic or OpenAI will not have the upfront money to get a RTX 3090 for each and every employee. Much better to sign a deal with the LLM providers and get cheaper token price with a sales person.
The $1500 Intel B70 with 32GB of VRAM can run this model at max context with good performance, btw. If you don't want to drop $5-10k for running DeepSeek this is your best budget option for local refactor/small scale dev help
WOW, my first try running on my 2 3090's, it was a bit slow... but it FEELS like opus 4.5, i gave it an image and a broad overview of what I wanted it to build, and it built the whole thing from beginning to end.
For your setup, do you have both 3090's in parallel for the inference of the model?