[rspoerri]

how do you do 1mio context with qwen3.6 27b, that only supports 256k? and what hardware would you run that on? 2 * 3090 is afaik currently at max 256k context.

[nyrikki]

You can get all the Qwen 3.x models up to ~1 million tokens using YaRN with llama.cpp.[0]

Personally I am using `--no-context-shift` and feeding in context back in on failure at the harness level.

I have 2x1080ti + 1xTitanV that have a full 262,144 tokens context on 262,144 tokens with `-sm tensor` at 62.04 t/s which isn't so bad.

But I also have a 1x3090 running unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL at 41.89 t/s but with only 130k context, but if you have a modular programming style both work pretty well.

But play with YaRN if you really need it.

[0]https://qwen.readthedocs.io/en/v3.0/run_locally/llama.cpp.ht...

[Vaskivo]

How can you get it to run at 41 t/s? I also have a single 3090 and even with MTP can't break 20 t/s.

HEre's my setup:

  llama-server
  --port 9999
  --model /MODELS/LLMs/Qwen3.6-27B-UD-Q4_K_XL.gguf
  --ctx-size 128000
  --threads 12
  --flash-attn on
  --device CUDA0
  --jinja
  --gpu-layers 52
  --mmproj /MODELS/LLMs/Qwen3.6-27B-mmproj-F16.gguf
  --cache-type-k q8_0
  --cache-type-v q8_0
  --temp 0.6 --top-k 20 --top-p 0.95 --min-p 0.0 --repeat-penalty 1.0 --presence-penalty 0.0
  --spec-type draft-mtp --spec-draft-n-max 2

(I'm not filling out 100% of the VRAM, as I have other stuff I need it for.)

[nyrikki]

(Note UPDATED config)

Ya, if you are using the CPU it may slowdown quick.

This may be a bit huge and overcomplicated, on this host I am running it on a AMD Ryzen 7 5700G so that I can use the APU to dedicate the 3090.

    podman run --device nvidia.com/gpu=all -d -v llama_qwen3.6mpt:/root/.cache -p 8080:8080 local/llama.cpp:full-cuda --server \
    -hf unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL \
    -ngl 99 \
    --ctx-size 131072 \
    --no-mmproj-offload \
    --no-context-shift \
    --kv-unified \
    --spec-type draft-mtp \
    --spec-draft-n-max 6 \
    --spec-draft-p-min 0.75 \
    -fa on --jinja --no-mmap \
    --cache-ram -1 \
    --no-warmup -np 1 \
    -n 32768 \
    --cache-type-k q8_0 \
    --cache-type-v q8_0 \
    --temp 0.6 \
    --min-p 0.00 \
    --top-k 20 \
    --top-p 0.95 \
    --presence-penalty 0.0 \
    --repeat-penalty 1.05 \
    --fit off \
    --reasoning on \
    --chat-template-kwargs '{"preserve_thinking":true}' \
    --prio 3 \
    --poll 100 \
    --port 8080 \
    --host 0.0.0.0

I am just building the container with:

     podman build -t local/llama.cpp:full-cuda --target full -f .devops/cuda.Dockerfile .

And here is the logs from a 'make me a flappy bird program in python' webui prompt.

     prompt eval time =     105.86 ms /    19 tokens (    5.57 ms per token,   179.47 tokens per second)
       eval time =  100549.41 ms /  4608 tokens (   21.82 ms per token,    45.83 tokens per second)
      total time =  100655.28 ms /  4627 tokens
     draft acceptance rate = 0.47215 ( 3408 accepted /  7218 generated)

I am down to ~25.54 t/s with a 95% full context.

[nyrikki]

That config looked too complicated, getting rid of the --prio 3 and --poll 100, setting the draft-n-max to now recommended values, etc... kicked it up to 61 t/s

I think that was all about some earlier crashes.

     podman run --device nvidia.com/gpu=all -d -v llama_qwen3.6mpt:/root/.cache -p 8080:8080 local/llama.cpp:full-cuda --server \
    -hf unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL \
    -ngl 99 \
    --ctx-size 128000 \
    --no-mmproj-offload \
    --no-context-shift \
    --kv-unified \
    --spec-type draft-mtp \
    --spec-draft-n-max 2 \
    --spec-draft-p-min 0.75 \
    -fa on --jinja --no-mmap \
    --cache-ram -1 \
    --no-warmup -np 1\
    -n 32768 \
    --cache-type-k q8_0 \
    --cache-type-v q8_0 \
    --temp 0.6 \
    --min-p 0.00 \
    --top-k 20 \
    --top-p 0.95 \
    --presence-penalty 0.0 \
    --repeat-penalty 1.05 \
    --fit off \
    --reasoning on \
    --chat-template-kwargs '{"preserve_thinking":true}' \
    --port 8080 \
    --host 0.0.0.0

[Vaskivo]

Yeah, having even a little bit in the CPU tanks the t/s...

But thanks. I've learned a few more configurations to tinker with.

[omneity]

You can increase the context window beyond its max trained context using RoPE scaling[0] which will require more VRAM.

But you can increase your context window for the same VRAM by quantizing the KV cache with FP8 (double the context) or TurboQuant (more than double)[1].

0: https://medium.com/@leannetan/extending-context-length-with-...

1: https://docs.vllm.ai/en/latest/features/quantization/quantiz...

[trilogic]

We managed to increase the ctx for whatever llm model that is GGUFED, here the experimental tests: https://www.reddit.com/r/Hugston/