Eval bug: --no-kv-offload causes immediate EOS generation with Qwen3.6-27B on Vulkan (works without flag)

[Dam14n]

Name and Version

llama-server --version
build 9608 (70b54e1)

Operating systems

Windows

GGML backends

Vulkan

Hardware

• 2x AMD Radeon RX 6700 XT
• 64 GB RAM
• Ryzen 5950X 16T

Models

Qwen3.6-27B-UD-Q4_K_XL.gguf

Problem description & steps to reproduce

When running llama-server with --no-kv-offload, the model immediately terminates generation by predicting EOS as the first generated token.

The same model, prompt, and server configuration work correctly when --no-kv-offload is removed.

Working Command

llama-server \
  -m ./Qwen3.6-27B-UD-Q4_K_XL.gguf \
  --port 1234 \
  -ngl 999

Generation works normally and returns text.

Relevant log:

CPU KV buffer size = 0.00 MiB

Failing Command

llama-server \
  -m ./Qwen3.6-27B-UD-Q4_K_XL.gguf \
  --port 1234 \
  -ngl 999 \
  --no-kv-offload

Generation produces no text.

Relevant logs:

CPU KV buffer size = 4000.00 MiB

and

predicted_n = 1
stop_type = "eos"
ignore_eos = false

The model immediately predicts EOS as the first generated token.

Expected Behavior

The model should generate text normally regardless of whether the KV cache is stored on GPU or CPU.

Actual Behavior

Prompt evaluation succeeds, but generation immediately stops with EOS before any text is produced.

Minimal Reproduction

Works:

llama-server -m ./Qwen3.6-27B-UD-Q4_K_XL.gguf -ngl 999

Fails:

llama-server -m ./Qwen3.6-27B-UD-Q4_K_XL.gguf -ngl 999 --no-kv-offload

First Bad Commit

No response

Relevant log output

Logs

WORKING LOGS.txt

NOT WORKING LOGS.txt

[imathrowback]

I just came across this issue as well with a Qwen3.5 model and llama-cli build b10056.

Same video card, failure with --no-kv-offload. The difference is that I'm seeing garbage output, not no output

L:\downloads\llama-b10056-bin-win-vulkan-x64>llama-cli -hf unsloth/Qwen3.5-9B-GGUF -p "write me a haiku about a car" -n 512 -c 16384 --no-conversation -st --temp 1.0 --top-p 0.95 --top-k 20 --presence-penalty 1.5 --reasoning off --no-kv-offload


Loading model...

build      : b10056-b85833e93
model      : unsloth/Qwen3.5-9B-GGUF
ftype      : Q4_K - Medium
modalities : text, vision, video

> write me a haiku about a car
Chrome... auresing-,ing: ( (：

ak.aut.
思维.

: thisof; the

车型，。

[ Prompt: 95.9 t/s | Generation: 20.4 t/s ]

Exiting...

[himonhm]

I have same problem as imathrowback described. My setup is unsloth qwen3.6 35b a3b on iGPU Radeon 780M, on Vulkan

[tombgh]

Same issue on Vulcan with unsloth\Qwen3.8-27B-UD-Q4_K_XL

[kissgyorgy]

Same with Vulcan unsloth/Qwen3.8-27B-UD-Q3_K_XL.gguf