export GGML_BACKEND=SYCL
GPU_BACKEND="SYCL0"
MODEL="unsloth/Qwen3.6-27B-MTP-GGUF:Q4_K_S"
MODEL_ARGS="--spec-type draft-mtp --spec-draft-n-max 2"
PARALLEL="1"
UBATCH_SIZE="1024"
BATCH_SIZE="1024"
FLASH_ATTENTION="-fa on -ctk q8_0 -ctv q8_0 -ctkd q8_0 -ctvd q8_0"
CTX_SIZE="131072"
NLAYERS=99

llama-server \
  --host 0.0.0.0 \
  --mmap \
  --no-mmproj-offload $FLASH_ATTENTION \
  --cache-ram 0 \
  --poll 0 \
  --fit off \
  -t 16 \
  -tb 16 \
  -hf $MODEL $MODEL_ARGS \
  --device $GPU_BACKEND \
  -np $PARALLEL \
  -ngl $NLAYERS \
  -c $CTX_SIZE \
  --checkpoint-every-n-tokens -1 \
  --ctx-checkpoints 0 \
  --batch-size $BATCH_SIZE \
  --ubatch-size $UBATCH_SIZE \
  --no-kv-unified \
  --temp 0.6 \
  --top-p 0.95 \
  --min-p 0.0 \
  --top-k 20 \
  --presence_penalty 0.0 \
  --repeat_penalty 1.0 \
  --jinja

0.00.665.972 I log_info: verbosity = 3 (adjust with the `-lv N` CLI arg)
0.00.665.975 I device_info:
0.00.792.148 I   - SYCL0   : Intel(R) Arc(TM) Pro B60 Graphics (24480 MiB, 24405 MiB free)
0.00.792.898 I   - Vulkan0 : Intel(R) Arc(tm) Pro B60 Graphics (BMG G21) (24480 MiB, 21964 MiB free)
0.00.792.901 I   - BLAS    : OpenBLAS (0 MiB, 0 MiB free)
0.00.792.913 I   - CPU     : AMD EPYC 7443 24-Core Processor (515620 MiB, 515620 MiB free)
0.00.793.259 I system_info: n_threads = 16 (n_threads_batch = 16) / 16 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 |>
0.00.793.301 I srv          init: running without SSL
0.00.793.453 I srv          init: using 15 threads for HTTP server
0.00.793.558 I srv         start: binding port with default address family
0.00.794.697 I srv  llama_server: loading model
0.00.794.702 I srv    load_model: loading model '/root/.cache/huggingface/hub/models--unsloth--Qwen3.6-27B-MTP-GGUF/snapshots/b3a58239d8d40b953>
0.09.665.553 W llama_context: n_ctx_seq (131072) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
0.09.721.775 I common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
0.10.029.522 I srv    load_model: creating MTP draft context against the target model '/root/.cache/huggingface/hub/
0.10.029.546 W llama_context: n_ctx_seq (131072) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
0.10.060.240 W load_hparams: Qwen-VL models require at minimum 1024 image tokens to function correctly on grounding tasks
0.10.060.243 W load_hparams: if you encounter problems with accuracy, try adding --image-min-tokens 1024
0.10.060.243 W load_hparams: more info: https://github.com/ggml-org/llama.cpp/issues/16842
0.10.527.813 I srv    load_model: loaded multimodal model, '/root/.cache/huggingface/hub/models--unsloth--Qwen3.6-27B-MTP-GGUF/snapshots/b3a582>
0.10.527.824 I srv    load_model: initializing slots, n_slots = 1
0.10.640.509 I common_context_can_seq_rm: the context supports bounded partial sequence removal
0.10.646.290 I common_speculative_impl_draft_mtp: adding speculative implementation 'draft-mtp'
0.10.646.293 I common_speculative_impl_draft_mtp: - n_max=2, n_min=0, p_min=0.00, n_embd=5120, backend_sampling=1
0.10.646.294 I common_speculative_impl_draft_mtp: - gpu_layers=-1, cache_k=q8_0, cache_v=q8_0, ctx_tgt=yes, ctx_dft=yes, devices=[default]
0.10.646.353 W llama_sampler_backend_support: device 'SYCL0' does not have support for op GET_ROWS needed for sampler 'top-k'
0.10.646.366 I srv    load_model: speculative decoding context initialized
0.10.646.367 I slot   load_model: id  0 | task -1 | new slot, n_ctx = 131072
0.10.646.386 I srv    load_model: prompt cache is disabled - use `--cache-ram N` to enable it
0.10.646.386 I srv    load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391
0.10.646.404 W srv          init: --cache-idle-slots requires --kv-unified, disabling