Eval bug: Qwen3-VL - does not use fully gpu

[Qbason]

Name and Version

./llama-server --version
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon RX 7900 XTX (RADV NAVI31) (radv) | uma: 0 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
load_backend: loaded Vulkan backend from /app/libggml-vulkan.so
load_backend: loaded CPU backend from /app/libggml-cpu-haswell.so
version: 6894 (6eb208d17)
built with cc (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0 for x86_64-linux-gnu

Operating systems

Linux

GGML backends

Vulkan

Hardware

AMD Radeon RX 7900 XTX 24GB

Models

Qwen3-VL-30B-A3B-Instruct-GGUF
and
Qwen3-VL-30B-A3B-Thinking-GGUF

Problem description & steps to reproduce

When I use:

    ./llama_server
    --model /models/Qwen3-VL-30B-A3B-Instruct-Q4_K_M.gguf
    --host 0.0.0.0
    --ctx-size 32768 # or lower
    --batch-size 2048
    --ubatch-size 2048
    --jinja
    -np 2
    -ngl 999
    --mmproj /models/mmproj-Qwen3-VL-30B-A3B-Instruct-Q4_K_M.gguf

In meanwhile I do request via open webui to llama and I run nvtop
I got this:

I cannot force qwen3-vl to utilizie 100% gpu -> max 80%.
It doesn't matter if I include image or not.
The same thing is in a thinking version.
This problem doesn't exist in Qwen3-30B-A3B-Instruct-Q4_K_M or in thinking version.
Comparing Qwen3 to Qwen3-vl the performance is much worse ( ≈160 to ≈100 ).
What can I do in such situation?

First Bad Commit

No response

Relevant log output

llama_memory_breakdown_print: | memory breakdown [MiB]                  | total   free     self   model   context   compute    unaccounted |
llama_memory_breakdown_print: |   - Vulkan0 (RX 7900 XTX (RADV NAVI31)) | 24560 = 2389 + (22119 = 17524 +    3072 +    1523) +          51 |
llama_memory_breakdown_print: |   - Host                                |                   310 =   166 +       0 +     144                |
load_backend: loaded RPC backend from /app/build/bin/libggml-rpc.so
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon RX 7900 XTX (RADV NAVI31) (radv) | uma: 0 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
load_backend: loaded Vulkan backend from /app/build/bin/libggml-vulkan.so
load_backend: loaded CPU backend from /app/build/bin/libggml-cpu-haswell.so
warn: LLAMA_ARG_HOST environment variable is set, but will be overwritten by command line argument --host
build: 6898 (d2d931f17) with cc (Ubuntu 11.4.0-1ubuntu1~22.04.2) 11.4.0 for x86_64-linux-gnu
system info: n_threads = 8, n_threads_batch = 8, total_threads = 16

system_info: n_threads = 8 (n_threads_batch = 8) / 16 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 

main: binding port with default address family
main: HTTP server is listening, hostname: 0.0.0.0, port: 5809, http threads: 15
main: loading model
srv    load_model: loading model '/models/Qwen3-VL-30B-A3B-Instruct-Q4_K_M.gguf'
llama_model_load_from_file_impl: using device Vulkan0 (AMD Radeon RX 7900 XTX (RADV NAVI31)) (0000:10:00.0) - 20214 MiB free
llama_model_loader: loaded meta data with 45 key-value pairs and 579 tensors from /models/Qwen3-VL-30B-A3B-Instruct-Q4_K_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = qwen3vlmoe
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Qwen3-Vl-30B-A3B-Instruct
llama_model_loader: - kv   3:                           general.finetune str              = Instruct
llama_model_loader: - kv   4:                           general.basename str              = Qwen3-Vl-30B-A3B-Instruct
llama_model_loader: - kv   5:                       general.quantized_by str              = Unsloth
llama_model_loader: - kv   6:                         general.size_label str              = 30B-A3B
llama_model_loader: - kv   7:                            general.license str              = apache-2.0
llama_model_loader: - kv   8:                           general.repo_url str              = https://huggingface.co/unsloth
llama_model_loader: - kv   9:                   general.base_model.count u32              = 1
llama_model_loader: - kv  10:                  general.base_model.0.name str              = Qwen3 VL 30B A3B Instruct
llama_model_loader: - kv  11:          general.base_model.0.organization str              = Qwen
llama_model_loader: - kv  12:              general.base_model.0.repo_url str              = https://huggingface.co/Qwen/Qwen3-VL-...
llama_model_loader: - kv  13:                               general.tags arr[str,2]       = ["unsloth", "image-text-to-text"]
llama_model_loader: - kv  14:                     qwen3vlmoe.block_count u32              = 48
llama_model_loader: - kv  15:                  qwen3vlmoe.context_length u32              = 262144
llama_model_loader: - kv  16:                qwen3vlmoe.embedding_length u32              = 2048
llama_model_loader: - kv  17:             qwen3vlmoe.feed_forward_length u32              = 6144
llama_model_loader: - kv  18:            qwen3vlmoe.attention.head_count u32              = 32
llama_model_loader: - kv  19:         qwen3vlmoe.attention.head_count_kv u32              = 4
llama_model_loader: - kv  20:                  qwen3vlmoe.rope.freq_base f32              = 5000000.000000
llama_model_loader: - kv  21: qwen3vlmoe.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  22:               qwen3vlmoe.expert_used_count u32              = 8
llama_model_loader: - kv  23:            qwen3vlmoe.attention.key_length u32              = 128
llama_model_loader: - kv  24:          qwen3vlmoe.attention.value_length u32              = 128
llama_model_loader: - kv  25:                    qwen3vlmoe.expert_count u32              = 128
llama_model_loader: - kv  26:      qwen3vlmoe.expert_feed_forward_length u32              = 768
llama_model_loader: - kv  27:         qwen3vlmoe.rope.dimension_sections arr[i32,4]       = [24, 20, 20, 0]
llama_model_loader: - kv  28:              qwen3vlmoe.n_deepstack_layers u32              = 3
llama_model_loader: - kv  29:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  30:                         tokenizer.ggml.pre str              = qwen2
llama_model_loader: - kv  31:                      tokenizer.ggml.tokens arr[str,151936]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  32:                  tokenizer.ggml.token_type arr[i32,151936]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  33:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  34:                tokenizer.ggml.eos_token_id u32              = 151645
llama_model_loader: - kv  35:            tokenizer.ggml.padding_token_id u32              = 151654
llama_model_loader: - kv  36:                tokenizer.ggml.bos_token_id u32              = 151643
llama_model_loader: - kv  37:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  38:                    tokenizer.chat_template str              = {%- if tools %}\n    {{- '<|im_start|>...
llama_model_loader: - kv  39:               general.quantization_version u32              = 2
llama_model_loader: - kv  40:                          general.file_type u32              = 15
llama_model_loader: - kv  41:                      quantize.imatrix.file str              = Qwen3-VL-30B-A3B-Instruct-GGUF/imatri...
llama_model_loader: - kv  42:                   quantize.imatrix.dataset str              = unsloth_calibration_Qwen3-VL-30B-A3B-...
llama_model_loader: - kv  43:             quantize.imatrix.entries_count u32              = 384
llama_model_loader: - kv  44:              quantize.imatrix.chunks_count u32              = 694
llama_model_loader: - type  f32:  241 tensors
llama_model_loader: - type q4_K:  289 tensors
llama_model_loader: - type q6_K:   49 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_K - Medium
print_info: file size   = 17.28 GiB (4.86 BPW) 
load: printing all EOG tokens:
load:   - 151643 ('<|endoftext|>')
load:   - 151645 ('<|im_end|>')
load:   - 151662 ('<|fim_pad|>')
load:   - 151663 ('<|repo_name|>')
load:   - 151664 ('<|file_sep|>')
load: special tokens cache size = 26
load: token to piece cache size = 0.9311 MB
print_info: arch             = qwen3vlmoe
print_info: vocab_only       = 0
print_info: n_ctx_train      = 262144
print_info: n_embd           = 8192
print_info: n_layer          = 48
print_info: n_head           = 32
print_info: n_head_kv        = 4
print_info: n_rot            = 128
print_info: n_swa            = 0
print_info: is_swa_any       = 0
print_info: n_embd_head_k    = 128
print_info: n_embd_head_v    = 128
print_info: n_gqa            = 8
print_info: n_embd_k_gqa     = 512
print_info: n_embd_v_gqa     = 512
print_info: f_norm_eps       = 0.0e+00
print_info: f_norm_rms_eps   = 1.0e-06
print_info: f_clamp_kqv      = 0.0e+00
print_info: f_max_alibi_bias = 0.0e+00
print_info: f_logit_scale    = 0.0e+00
print_info: f_attn_scale     = 0.0e+00
print_info: n_ff             = 6144
print_info: n_expert         = 128
print_info: n_expert_used    = 8
print_info: n_expert_groups  = 0
print_info: n_group_used     = 0
print_info: causal attn      = 1
print_info: pooling type     = 0
print_info: rope type        = 40
print_info: rope scaling     = linear
print_info: freq_base_train  = 5000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 262144
print_info: rope_finetuned   = unknown
print_info: mrope sections   = [24, 20, 20, 0]
print_info: model type       = 30B.A3B
print_info: model params     = 30.53 B
print_info: general.name     = Qwen3-Vl-30B-A3B-Instruct
print_info: n_ff_exp         = 768
print_info: vocab type       = BPE
print_info: n_vocab          = 151936
print_info: n_merges         = 151387
print_info: BOS token        = 151643 '<|endoftext|>'
print_info: EOS token        = 151645 '<|im_end|>'
print_info: EOT token        = 151645 '<|im_end|>'
print_info: PAD token        = 151654 '<|vision_pad|>'
print_info: LF token         = 198 'Ċ'
print_info: FIM PRE token    = 151659 '<|fim_prefix|>'
print_info: FIM SUF token    = 151661 '<|fim_suffix|>'
print_info: FIM MID token    = 151660 '<|fim_middle|>'
print_info: FIM PAD token    = 151662 '<|fim_pad|>'
print_info: FIM REP token    = 151663 '<|repo_name|>'
print_info: FIM SEP token    = 151664 '<|file_sep|>'
print_info: EOG token        = 151643 '<|endoftext|>'
print_info: EOG token        = 151645 '<|im_end|>'
print_info: EOG token        = 151662 '<|fim_pad|>'
print_info: EOG token        = 151663 '<|repo_name|>'
print_info: EOG token        = 151664 '<|file_sep|>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = true)
srv  log_server_r: request: GET /health 127.0.0.1 503
[INFO] Request ::1 "GET / HTTP/1.1" 302 26 "curl/8.5.0" 23.681µs
srv  log_server_r: request: GET /health 127.0.0.1 503
srv  log_server_r: request: GET /health 127.0.0.1 503
srv  log_server_r: request: GET /health 127.0.0.1 503
srv  log_server_r: request: GET /health 127.0.0.1 503
load_tensors: offloading 48 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 49/49 layers to GPU
load_tensors:   CPU_Mapped model buffer size =   166.92 MiB
load_tensors:      Vulkan0 model buffer size = 17524.42 MiB
...............srv  log_server_r: request: GET /health 127.0.0.1 503
....................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 2
llama_context: n_ctx         = 32768
llama_context: n_ctx_per_seq = 16384
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 2048
llama_context: causal_attn   = 1
llama_context: flash_attn    = auto
llama_context: kv_unified    = false
llama_context: freq_base     = 5000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (16384) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
llama_context: Vulkan_Host  output buffer size =     1.16 MiB
llama_kv_cache:    Vulkan0 KV buffer size =  3072.00 MiB
llama_kv_cache: size = 3072.00 MiB ( 16384 cells,  48 layers,  2/2 seqs), K (f16): 1536.00 MiB, V (f16): 1536.00 MiB
llama_context: Flash Attention was auto, set to enabled
llama_context:    Vulkan0 compute buffer size =  1523.07 MiB
llama_context: Vulkan_Host compute buffer size =   144.08 MiB
llama_context: graph nodes  = 3127
llama_context: graph splits = 2
common_init_from_params: added <|endoftext|> logit bias = -inf
common_init_from_params: added <|im_end|> logit bias = -inf
common_init_from_params: added <|fim_pad|> logit bias = -inf
common_init_from_params: added <|repo_name|> logit bias = -inf
common_init_from_params: added <|file_sep|> logit bias = -inf
common_init_from_params: setting dry_penalty_last_n to ctx_size = 32768
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
srv  log_server_r: request: GET /health 127.0.0.1 503
clip_model_loader: model name:   Qwen3-Vl-30B-A3B-Instruct
clip_model_loader: description:  
clip_model_loader: GGUF version: 3
clip_model_loader: alignment:    32
clip_model_loader: n_tensors:    352
clip_model_loader: n_kv:         31

clip_model_loader: has vision encoder
clip_ctx: CLIP using Vulkan0 backend
load_hparams: projector:          qwen3vl_merger
load_hparams: n_embd:             1152
load_hparams: n_head:             16
load_hparams: n_ff:               4304
load_hparams: n_layer:            27
load_hparams: ffn_op:             gelu
load_hparams: projection_dim:     2048

--- vision hparams ---
load_hparams: image_size:         1024
load_hparams: patch_size:         16
load_hparams: has_llava_proj:     0
load_hparams: minicpmv_version:   0
load_hparams: proj_scale_factor:  0
load_hparams: n_wa_pattern:       0
load_hparams: spatial_merge_size: 2

load_hparams: model size:         1033.29 MiB
load_hparams: metadata size:      0.12 MiB
alloc_compute_meta:    Vulkan0 compute buffer size =     3.00 MiB
alloc_compute_meta:        CPU compute buffer size =     0.19 MiB
srv    load_model: loaded multimodal model, '/models/mmproj-Qwen3-VL-30B-A3B-Instruct-Q4_K_M.gguf'
srv          init: initializing slots, n_slots = 2
slot         init: id  0 | task -1 | new slot n_ctx_slot = 16384
slot         init: id  1 | task -1 | new slot n_ctx_slot = 16384
srv          init: prompt cache is enabled, size limit: 8192 MiB
srv          init: use `--cache-ram 0` to disable the prompt cache
srv          init: for more info see https://github.com/ggml-org/llama.cpp/pull/16391
srv          init: thinking = 0
main: model loaded
main: chat template, chat_template: {%- if tools %}
    {{- '<|im_start|>system\n' }}
    {%- if messages[0].role == 'system' %}
        {%- if messages[0].content is string %}
            {{- messages[0].content }}
        {%- else %}
            {%- for content in messages[0].content %}
                {%- if 'text' in content %}
                    {{- content.text }}
                {%- endif %}
            {%- endfor %}
        {%- endif %}
        {{- '\n\n' }}
    {%- endif %}
    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
    {%- for tool in tools %}
        {{- "\n" }}
        {{- tool | tojson }}
    {%- endfor %}
    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
{%- else %}
    {%- if messages[0].role == 'system' %}
        {{- '<|im_start|>system\n' }}
        {%- if messages[0].content is string %}
            {{- messages[0].content }}
        {%- else %}
            {%- for content in messages[0].content %}
                {%- if 'text' in content %}
                    {{- content.text }}
                {%- endif %}
            {%- endfor %}
        {%- endif %}
        {{- '<|im_end|>\n' }}
    {%- endif %}
{%- endif %}
{%- set image_count = namespace(value=0) %}
{%- set video_count = namespace(value=0) %}
{%- for message in messages %}
    {%- if message.role == "user" %}
        {{- '<|im_start|>' + message.role + '\n' }}
        {%- if message.content is string %}
            {{- message.content }}
        {%- else %}
            {%- for content in message.content %}
                {%- if content.type == 'image' or 'image' in content or 'image_url' in content %}
                    {%- set image_count.value = image_count.value + 1 %}
                    {%- if add_vision_id %}Picture {{ image_count.value }}: {% endif -%}
                    <|vision_start|><|image_pad|><|vision_end|>
                {%- elif content.type == 'video' or 'video' in content %}
                    {%- set video_count.value = video_count.value + 1 %}
                    {%- if add_vision_id %}Video {{ video_count.value }}: {% endif -%}
                    <|vision_start|><|video_pad|><|vision_end|>
                {%- elif 'text' in content %}
                    {{- content.text }}
                {%- endif %}
            {%- endfor %}
        {%- endif %}
        {{- '<|im_end|>\n' }}
    {%- elif message.role == "assistant" %}
        {{- '<|im_start|>' + message.role + '\n' }}
        {%- if message.content is string %}
            {{- message.content }}
        {%- else %}
            {%- for content_item in message.content %}
                {%- if 'text' in content_item %}
                    {{- content_item.text }}
                {%- endif %}
            {%- endfor %}
        {%- endif %}
        {%- if message.tool_calls %}
            {%- for tool_call in message.tool_calls %}
                {%- if (loop.first and message.content) or (not loop.first) %}
                    {{- '\n' }}
                {%- endif %}
                {%- if tool_call.function %}
                    {%- set tool_call = tool_call.function %}
                {%- endif %}
                {{- '<tool_call>\n{"name": "' }}
                {{- tool_call.name }}
                {{- '", "arguments": ' }}
                {%- if tool_call.arguments is string %}
                    {{- tool_call.arguments }}
                {%- else %}
                    {{- tool_call.arguments | tojson }}
                {%- endif %}
                {{- '}\n</tool_call>' }}
            {%- endfor %}
        {%- endif %}
        {{- '<|im_end|>\n' }}
    {%- elif message.role == "tool" %}
        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
            {{- '<|im_start|>user' }}
        {%- endif %}
        {{- '\n<tool_response>\n' }}
        {%- if message.content is string %}
            {{- message.content }}
        {%- else %}
            {%- for content in message.content %}
                {%- if content.type == 'image' or 'image' in content or 'image_url' in content %}
                    {%- set image_count.value = image_count.value + 1 %}
                    {%- if add_vision_id %}Picture {{ image_count.value }}: {% endif -%}
                    <|vision_start|><|image_pad|><|vision_end|>
                {%- elif content.type == 'video' or 'video' in content %}
                    {%- set video_count.value = video_count.value + 1 %}
                    {%- if add_vision_id %}Video {{ video_count.value }}: {% endif -%}
                    <|vision_start|><|video_pad|><|vision_end|>
                {%- elif 'text' in content %}
                    {{- content.text }}
                {%- endif %}
            {%- endfor %}
        {%- endif %}
        {{- '\n</tool_response>' }}
        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
            {{- '<|im_end|>\n' }}
        {%- endif %}
    {%- endif %}
{%- endfor %}
{%- if add_generation_prompt %}
    {{- '<|im_start|>assistant\n' }}
{%- endif %}
, example_format: '<|im_start|>system
You are a helpful assistant<|im_end|>
<|im_start|>user
Hello<|im_end|>
<|im_start|>assistant
Hi there<|im_end|>
<|im_start|>user
How are you?<|im_end|>
<|im_start|>assistant
'
main: server is listening on http://0.0.0.0:5809 - starting the main loop
srv  update_slots: all slots are idle
[INFO] Request ::1 "GET / HTTP/1.1" 302 26 "curl/8.5.0" 27.551µs
[INFO] <qwen3-vl-30b-instruct> Health check passed on http://localhost:5809/health
srv  log_server_r: request: GET /health 127.0.0.1 200
srv  params_from_: Chat format: Hermes 2 Pro
slot get_availabl: id  1 | task -1 | selected slot by LRU, t_last = -1
slot launch_slot_: id  1 | task 0 | processing task
slot update_slots: id  1 | task 0 | new prompt, n_ctx_slot = 16384, n_keep = 0, task.n_tokens = 726
slot update_slots: id  1 | task 0 | n_tokens = 0, memory_seq_rm [0, end)
slot update_slots: id  1 | task 0 | prompt processing progress, n_tokens = 16, batch.n_tokens = 16, progress = 0.022039
slot update_slots: id  1 | task 0 | n_tokens = 16, memory_seq_rm [16, end)
srv  process_chun: processing image...
srv  process_chun: image processed in 864 ms
slot update_slots: id  1 | task 0 | prompt processing progress, n_tokens = 726, batch.n_tokens = 6, progress = 1.000000
slot update_slots: id  1 | task 0 | prompt done, n_tokens = 726, batch.n_tokens = 6
slot print_timing: id  1 | task 0 | 
prompt eval time =     992.26 ms /   726 tokens (    1.37 ms per token,   731.66 tokens per second)
       eval time =    3889.78 ms /   398 tokens (    9.77 ms per token,   102.32 tokens per second)
      total time =    4882.04 ms /  1124 tokens
slot      release: id  1 | task 0 | stop processing: n_tokens = 1123, truncated = 0
srv  update_slots: all slots are idle
srv  log_server_r: request: POST /v1/chat/completions 127.0.0.1 200
[INFO] Request 172.20.0.1 "POST /v1/chat/completions HTTP/1.1" 200 101650 "Python/3.11 aiohttp/3.12.15" 40.454859888s
[INFO] Request 172.20.0.1 "GET /v1/models HTTP/1.1" 200 3428 "Python/3.11 aiohttp/3.12.15" 106.522µs
srv  params_from_: Chat format: Hermes 2 Pro
slot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = -1
slot launch_slot_: id  0 | task 400 | processing task
slot update_slots: id  0 | task 400 | new prompt, n_ctx_slot = 16384, n_keep = 0, task.n_tokens = 623
slot update_slots: id  0 | task 400 | n_tokens = 0, memory_seq_rm [0, end)
slot update_slots: id  0 | task 400 | prompt processing progress, n_tokens = 623, batch.n_tokens = 623, progress = 1.000000
slot update_slots: id  0 | task 400 | prompt done, n_tokens = 623, batch.n_tokens = 623
slot print_timing: id  0 | task 400 | 
prompt eval time =     423.31 ms /   623 tokens (    0.68 ms per token,  1471.75 tokens per second)
       eval time =     816.42 ms /    95 tokens (    8.59 ms per token,   116.36 tokens per second)
      total time =    1239.72 ms /   718 tokens
slot      release: id  0 | task 400 | stop processing: n_tokens = 717, truncated = 0
srv  update_slots: all slots are idle
srv  log_server_r: request: POST /v1/chat/completions 127.0.0.1 200
[INFO] Request 172.20.0.1 "POST /v1/chat/completions HTTP/1.1" 200 889 "Python/3.11 aiohttp/3.12.15" 1.24430083s
[INFO] Request 172.20.0.1 "GET /v1/models HTTP/1.1" 200 3428 "Python/3.11 aiohttp/3.12.15" 121.062µs
srv  params_from_: Chat format: Hermes 2 Pro
slot get_availabl: id  1 | task -1 | selected slot by LRU, t_last = 485549729955
slot launch_slot_: id  1 | task 496 | processing task
slot update_slots: id  1 | task 496 | new prompt, n_ctx_slot = 16384, n_keep = 0, task.n_tokens = 693
slot update_slots: id  1 | task 496 | old: ... <|im_start|>user
 | Read all details from the receipt
slot update_slots: id  1 | task 496 | new: ... <|im_start|>user

[psi00]

Would like to note that I can't seem to be able to replicate this problem using CUDA backend on an RTX 3060 12GB. Someone can correct me if I'm wrong but I believe there were some problems with implementation on other backends, I recall looking at the original pull request and there was only CUDA support for a while, so it might be a continuation of that, unless the author(s) just had an NVIDIA GPU and that's what they worked on.

version

llama-server --version
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 CUDA devices:
  Device 0: NVIDIA GeForce RTX 3060, compute capability 8.6, VMM: yes
load_backend: loaded CUDA backend from C:\Apps\llama.cpp\ggml-cuda.dll
load_backend: loaded RPC backend from C:\Apps\llama.cpp\ggml-rpc.dll
load_backend: loaded CPU backend from C:\Apps\llama.cpp\ggml-cpu-haswell.dll
version: 6900 (c22473b58)
built with clang version 19.1.5 for x86_64-pc-windows-msvc

gpustat image (see 95% GPU utilisation)

Log output

ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 CUDA devices:
  Device 0: NVIDIA GeForce RTX 3060, compute capability 8.6, VMM: yes
load_backend: loaded CUDA backend from C:\Apps\llama.cpp\ggml-cuda.dll
load_backend: loaded RPC backend from C:\Apps\llama.cpp\ggml-rpc.dll
load_backend: loaded CPU backend from C:\Apps\llama.cpp\ggml-cpu-haswell.dll
build: 6900 (c22473b58) with clang version 19.1.5 for x86_64-pc-windows-msvc
system info: n_threads = 8, n_threads_batch = 8, total_threads = 16

system_info: n_threads = 8 (n_threads_batch = 8) / 16 | CUDA : ARCHS = 500,610,700,750,800,860,890 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 

main: binding port with default address family
main: HTTP server is listening, hostname: 127.0.0.1, port: 5806, http threads: 15
main: loading model
srv    load_model: loading model 'C:/Models/llama.cpp/Qwen3-VL-8B-Thinking.IQ4_XS.gguf'
srv  log_server_r: request: GET /health 127.0.0.1 503
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 3060) (0000:07:00.0) - 11240 MiB free
llama_model_loader: loaded meta data with 36 key-value pairs and 399 tensors from C:/Models/llama.cpp/Qwen3-VL-8B-Thinking.IQ4_XS.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = qwen3vl
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Qwen3 VL 8B Thinking
llama_model_loader: - kv   3:                           general.finetune str              = Thinking
llama_model_loader: - kv   4:                           general.basename str              = Qwen3-VL
llama_model_loader: - kv   5:                         general.size_label str              = 8B
llama_model_loader: - kv   6:                            general.license str              = apache-2.0
llama_model_loader: - kv   7:                               general.tags arr[str,1]       = ["image-text-to-text"]
llama_model_loader: - kv   8:                        qwen3vl.block_count u32              = 36
llama_model_loader: - kv   9:                     qwen3vl.context_length u32              = 262144
llama_model_loader: - kv  10:                   qwen3vl.embedding_length u32              = 4096
llama_model_loader: - kv  11:                qwen3vl.feed_forward_length u32              = 12288
llama_model_loader: - kv  12:               qwen3vl.attention.head_count u32              = 32
llama_model_loader: - kv  13:            qwen3vl.attention.head_count_kv u32              = 8
llama_model_loader: - kv  14:                     qwen3vl.rope.freq_base f32              = 5000000.000000
llama_model_loader: - kv  15:   qwen3vl.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  16:               qwen3vl.attention.key_length u32              = 128
llama_model_loader: - kv  17:             qwen3vl.attention.value_length u32              = 128
llama_model_loader: - kv  18:            qwen3vl.rope.dimension_sections arr[i32,4]       = [24, 20, 20, 0]
llama_model_loader: - kv  19:                 qwen3vl.n_deepstack_layers u32              = 3
llama_model_loader: - kv  20:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  21:                         tokenizer.ggml.pre str              = qwen2
llama_model_loader: - kv  22:                      tokenizer.ggml.tokens arr[str,151936]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  23:                  tokenizer.ggml.token_type arr[i32,151936]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  24:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  25:                tokenizer.ggml.eos_token_id u32              = 151645
llama_model_loader: - kv  26:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  27:                tokenizer.ggml.bos_token_id u32              = 151643
llama_model_loader: - kv  28:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  29:                    tokenizer.chat_template str              = {%- set image_count = namespace(value...
llama_model_loader: - kv  30:               general.quantization_version u32              = 2
llama_model_loader: - kv  31:                          general.file_type u32              = 30
llama_model_loader: - kv  32:                      quantize.imatrix.file str              = /models_out/Qwen3-VL-8B-Thinking-GGUF...
llama_model_loader: - kv  33:                   quantize.imatrix.dataset str              = /training_dir/calibration_datav5.txt
llama_model_loader: - kv  34:             quantize.imatrix.entries_count u32              = 252
llama_model_loader: - kv  35:              quantize.imatrix.chunks_count u32              = 818
llama_model_loader: - type  f32:  145 tensors
llama_model_loader: - type q5_K:   36 tensors
llama_model_loader: - type q6_K:    1 tensors
llama_model_loader: - type iq4_xs:  217 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = IQ4_XS - 4.25 bpw
print_info: file size   = 4.24 GiB (4.45 BPW) 
load: printing all EOG tokens:
load:   - 151643 ('<|endoftext|>')
load:   - 151645 ('<|im_end|>')
load:   - 151662 ('<|fim_pad|>')
load:   - 151663 ('<|repo_name|>')
load:   - 151664 ('<|file_sep|>')
load: special tokens cache size = 26
load: token to piece cache size = 0.9311 MB
print_info: arch             = qwen3vl
print_info: vocab_only       = 0
print_info: n_ctx_train      = 262144
print_info: n_embd           = 16384
print_info: n_layer          = 36
print_info: n_head           = 32
print_info: n_head_kv        = 8
print_info: n_rot            = 128
print_info: n_swa            = 0
print_info: is_swa_any       = 0
print_info: n_embd_head_k    = 128
print_info: n_embd_head_v    = 128
print_info: n_gqa            = 4
print_info: n_embd_k_gqa     = 1024
print_info: n_embd_v_gqa     = 1024
print_info: f_norm_eps       = 0.0e+00
print_info: f_norm_rms_eps   = 1.0e-06
print_info: f_clamp_kqv      = 0.0e+00
print_info: f_max_alibi_bias = 0.0e+00
print_info: f_logit_scale    = 0.0e+00
print_info: f_attn_scale     = 0.0e+00
print_info: n_ff             = 12288
print_info: n_expert         = 0
print_info: n_expert_used    = 0
print_info: n_expert_groups  = 0
print_info: n_group_used     = 0
print_info: causal attn      = 1
print_info: pooling type     = 0
print_info: rope type        = 40
print_info: rope scaling     = linear
print_info: freq_base_train  = 5000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 262144
print_info: rope_finetuned   = unknown
print_info: mrope sections   = [24, 20, 20, 0]
print_info: model type       = 8B
print_info: model params     = 8.19 B
print_info: general.name     = Qwen3 VL 8B Thinking
print_info: vocab type       = BPE
print_info: n_vocab          = 151936
print_info: n_merges         = 151387
print_info: BOS token        = 151643 '<|endoftext|>'
print_info: EOS token        = 151645 '<|im_end|>'
print_info: EOT token        = 151645 '<|im_end|>'
print_info: PAD token        = 151643 '<|endoftext|>'
print_info: LF token         = 198 'Ċ'
print_info: FIM PRE token    = 151659 '<|fim_prefix|>'
print_info: FIM SUF token    = 151661 '<|fim_suffix|>'
print_info: FIM MID token    = 151660 '<|fim_middle|>'
print_info: FIM PAD token    = 151662 '<|fim_pad|>'
print_info: FIM REP token    = 151663 '<|repo_name|>'
print_info: FIM SEP token    = 151664 '<|file_sep|>'
print_info: EOG token        = 151643 '<|endoftext|>'
print_info: EOG token        = 151645 '<|im_end|>'
print_info: EOG token        = 151662 '<|fim_pad|>'
print_info: EOG token        = 151663 '<|repo_name|>'
print_info: EOG token        = 151664 '<|file_sep|>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 36 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 37/37 layers to GPU
load_tensors:   CPU_Mapped model buffer size =   315.30 MiB
load_tensors:        CUDA0 model buffer size =  4029.53 MiB
....................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 2
llama_context: n_ctx         = 16384
llama_context: n_ctx_per_seq = 8192
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = enabled
llama_context: kv_unified    = false
llama_context: freq_base     = 5000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (8192) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
llama_context:  CUDA_Host  output buffer size =     1.16 MiB
llama_kv_cache:      CUDA0 KV buffer size =  2304.00 MiB
llama_kv_cache: size = 2304.00 MiB (  8192 cells,  36 layers,  2/2 seqs), K (f16): 1152.00 MiB, V (f16): 1152.00 MiB
llama_context:      CUDA0 compute buffer size =   304.75 MiB
llama_context:  CUDA_Host compute buffer size =    24.02 MiB
llama_context: graph nodes  = 1339
llama_context: graph splits = 2
common_init_from_params: added <|endoftext|> logit bias = -inf
common_init_from_params: added <|im_end|> logit bias = -inf
common_init_from_params: added <|fim_pad|> logit bias = -inf
common_init_from_params: added <|repo_name|> logit bias = -inf
common_init_from_params: added <|file_sep|> logit bias = -inf
common_init_from_params: setting dry_penalty_last_n to ctx_size = 16384
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
clip_model_loader: model name:   Qwen3 VL 8B Thinking
clip_model_loader: description:  
clip_model_loader: GGUF version: 3
clip_model_loader: alignment:    32
clip_model_loader: n_tensors:    352
clip_model_loader: n_kv:         25

clip_model_loader: has vision encoder
clip_ctx: CLIP using CUDA0 backend
load_hparams: projector:          qwen3vl_merger
load_hparams: n_embd:             1152
load_hparams: n_head:             16
load_hparams: n_ff:               4304
load_hparams: n_layer:            27
load_hparams: ffn_op:             gelu
load_hparams: projection_dim:     4096

--- vision hparams ---
load_hparams: image_size:         1024
load_hparams: patch_size:         16
load_hparams: has_llava_proj:     0
load_hparams: minicpmv_version:   0
load_hparams: proj_scale_factor:  0
load_hparams: n_wa_pattern:       0
load_hparams: spatial_merge_size: 2

load_hparams: model size:         1105.32 MiB
load_hparams: metadata size:      0.12 MiB
srv  log_server_r: request: GET /health 127.0.0.1 503
alloc_compute_meta:      CUDA0 compute buffer size =     3.00 MiB
alloc_compute_meta:        CPU compute buffer size =     0.19 MiB
srv    load_model: loaded multimodal model, 'C:/Models/llama.cpp/Qwen3-VL-8B-Thinking.mmproj.gguf'
srv          init: initializing slots, n_slots = 2
slot         init: id  0 | task -1 | new slot n_ctx_slot = 8192
slot         init: id  1 | task -1 | new slot n_ctx_slot = 8192
srv          init: prompt cache is enabled, size limit: 8192 MiB
srv          init: use `--cache-ram 0` to disable the prompt cache
srv          init: for more info see https://github.com/ggml-org/llama.cpp/pull/16391
srv          init: thinking = 0
main: model loaded
main: chat template, chat_template: {%- set image_count = namespace(value=0) %}
{%- set video_count = namespace(value=0) %}
{%- macro render_content(content, do_vision_count) %}
    {%- if content is string %}
        {{- content }}
    {%- else %}
        {%- for item in content %}
            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
                {%- if do_vision_count %}
                    {%- set image_count.value = image_count.value + 1 %}
                {%- endif %}
                {%- if add_vision_id %}Picture {{ image_count.value }}: {% endif -%}
                <|vision_start|><|image_pad|><|vision_end|>
            {%- elif 'video' in item or item.type == 'video' %}
                {%- if do_vision_count %}
                    {%- set video_count.value = video_count.value + 1 %}
                {%- endif %}
                {%- if add_vision_id %}Video {{ video_count.value }}: {% endif -%}
                <|vision_start|><|video_pad|><|vision_end|>
            {%- elif 'text' in item %}
                {{- item.text }}
            {%- endif %}
        {%- endfor %}
    {%- endif %}
{%- endmacro %}
{%- if tools %}
    {{- '<|im_start|>system\n' }}
    {%- if messages[0].role == 'system' %}
        {{- render_content(messages[0].content, false) + '\n\n' }}
    {%- endif %}
    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
    {%- for tool in tools %}
        {{- "\n" }}
        {{- tool | tojson }}
    {%- endfor %}
    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
{%- else %}
    {%- if messages[0].role == 'system' %}
        {{- '<|im_start|>system\n' + render_content(messages[0].content, false) + '<|im_end|>\n' }}
    {%- endif %}
{%- endif %}
{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
{%- for message in messages[::-1] %}
    {%- set index = (messages|length - 1) - loop.index0 %}
    {%- if ns.multi_step_tool and message.role == "user" %}
        {%- set content = render_content(message.content, false) %}
        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
            {%- set ns.multi_step_tool = false %}
            {%- set ns.last_query_index = index %}
        {%- endif %}
    {%- endif %}
{%- endfor %}
{%- for message in messages %}
    {%- set content = render_content(message.content, True) %}
    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
    {%- elif message.role == "assistant" %}
        {%- set reasoning_content = '' %}
        {%- if message.reasoning_content is string %}
            {%- set reasoning_content = message.reasoning_content %}
        {%- else %}
            {%- if '</think>' in content %}
                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
            {%- endif %}
        {%- endif %}
        {%- if loop.index0 > ns.last_query_index %}
            {%- if loop.last or (not loop.last and reasoning_content) %}
                {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
            {%- else %}
                {{- '<|im_start|>' + message.role + '\n' + content }}
            {%- endif %}
        {%- else %}
            {{- '<|im_start|>' + message.role + '\n' + content }}
        {%- endif %}
        {%- if message.tool_calls %}
            {%- for tool_call in message.tool_calls %}
                {%- if (loop.first and content) or (not loop.first) %}
                    {{- '\n' }}
                {%- endif %}
                {%- if tool_call.function %}
                    {%- set tool_call = tool_call.function %}
                {%- endif %}
                {{- '<tool_call>\n{"name": "' }}
                {{- tool_call.name }}
                {{- '", "arguments": ' }}
                {%- if tool_call.arguments is string %}
                    {{- tool_call.arguments }}
                {%- else %}
                    {{- tool_call.arguments | tojson }}
                {%- endif %}
                {{- '}\n</tool_call>' }}
            {%- endfor %}
        {%- endif %}
        {{- '<|im_end|>\n' }}
    {%- elif message.role == "tool" %}
        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
            {{- '<|im_start|>user' }}
        {%- endif %}
        {{- '\n<tool_response>\n' }}
        {{- content }}
        {{- '\n</tool_response>' }}
        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
            {{- '<|im_end|>\n' }}
        {%- endif %}
    {%- endif %}
{%- endfor %}
{%- if add_generation_prompt %}
    {{- '<|im_start|>assistant\n<think>\n' }}
{%- endif %}
, example_format: '<|im_start|>system
You are a helpful assistant<|im_end|>
<|im_start|>user
Hello<|im_end|>
<|im_start|>assistant
Hi there<|im_end|>
<|im_start|>user
How are you?<|im_end|>
<|im_start|>assistant
'
main: server is listening on http://127.0.0.1:5806 - starting the main loop
srv  update_slots: all slots are idle
srv  log_server_r: request: GET /health 127.0.0.1 200
srv  log_server_r: request: GET / 127.0.0.1 200
srv  log_server_r: request: GET / 127.0.0.1 200
srv  log_server_r: request: GET /props 127.0.0.1 200
srv  log_server_r: request: GET /props 127.0.0.1 200
srv  log_server_r: request: GET /props 127.0.0.1 200
srv  update_slots: all slots are idle
srv  log_server_r: request: GET /slots 127.0.0.1 200
srv  log_server_r: request: GET /props 127.0.0.1 200
srv  log_server_r: request: GET /props 127.0.0.1 200
srv  params_from_: Chat format: Content-only
slot get_availabl: id  1 | task -1 | selected slot by LRU, t_last = -1
slot launch_slot_: id  1 | task 1 | processing task
slot update_slots: id  1 | task 1 | new prompt, n_ctx_slot = 8192, n_keep = 0, task.n_tokens = 982
slot update_slots: id  1 | task 1 | n_tokens = 0, memory_seq_rm [0, end)
slot update_slots: id  1 | task 1 | prompt processing progress, n_tokens = 16, batch.n_tokens = 16, progress = 0.016293
slot update_slots: id  1 | task 1 | n_tokens = 16, memory_seq_rm [16, end)
srv  process_chun: processing image...
srv  process_chun: image processed in 2788 ms
slot update_slots: id  1 | task 1 | prompt processing progress, n_tokens = 982, batch.n_tokens = 6, progress = 1.000000
slot update_slots: id  1 | task 1 | prompt done, n_tokens = 982, batch.n_tokens = 6
srv  log_server_r: request: GET /slots 127.0.0.1 200
slot print_timing: id  1 | task 1 | 
prompt eval time =    3463.30 ms /   982 tokens (    3.53 ms per token,   283.54 tokens per second)
       eval time =  143859.98 ms /  6254 tokens (   23.00 ms per token,    43.47 tokens per second)
      total time =  147323.28 ms /  7236 tokens
slot      release: id  1 | task 1 | stop processing: n_tokens = 7235, truncated = 0
srv  update_slots: all slots are idle
srv  log_server_r: request: POST /v1/chat/completions 127.0.0.1 200

[ngxson]

The vision model will now show more debugging details if some ops are not compatible with GPU. You should see a warning in the log.

Also, flash attention (enabled by default) allows most vision models to fully utilize the GPU.

I'm closing this issue and moving it to #16950 ; after the issue is fixed, flash attn will be enabled for Qwen3-VL model which will significantly improve the performance and memory usage