Misc. bug: [SYCL] Missing kernels for Qwen3-VL vision encoder (UPSCALE, mRoPE, etc.) on Intel Arc A770

[Hisir0909]

Name and Version

❯ ./llama-server --version
version: 8007 (4ae1b75)
built with IntelLLVM 2025.3.0 for Linux x86_64

Operating systems

Linux

Which llama.cpp modules do you know to be affected?

llama-server

Command line

./llama-server -m ~/Downloads/models/gguf/qwen-3-vl-8b/Qwen3-VL-8B-Instruct-Q6_K.gguf --mmproj ~/Downloads/models/gguf/qwen-3-vl-8b/mmproj-BF16.gguf -ngl 999 -c 4096 -n 512 -fa on --host 0.0.0.0 --no-warmup --image-max-tokens 2048 -np 1

Problem description & steps to reproduce

When running Qwen3-VL, the vision encoder fails to use SYCL acceleration for several key operators, leading to CPU fallback and slow performance (~23s for image encoding). The logs explicitly list unsupported ops: UPSCALE, SOFT_MAX, ROPE, MUL_MAT, etc.

First Bad Commit

No response

Relevant log output

Logs

❯ ./llama-server -m ~/Downloads/models/gguf/qwen-3-vl-8b/Qwen3-VL-8B-Instruct-Q6_K.gguf --mmproj ~/Downloads/models/gguf/qwen-3-vl-8b/mmproj-BF16.gguf -ngl 999 -c 4096 -n 512 -fa on --host 0.0.0.0 --no-warmup --image-max-tokens 2048 -np 1
build: 8007 (4ae1b7517) with IntelLLVM 2025.3.0 for Linux x86_64
system info: n_threads = 6, n_threads_batch = 6, total_threads = 12

system_info: n_threads = 6 (n_threads_batch = 6) / 12 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 

init: using 11 threads for HTTP server
start: binding port with default address family
main: loading model
srv    load_model: loading model '/home/u/Downloads/models/gguf/qwen-3-vl-8b/Qwen3-VL-8B-Instruct-Q6_K.gguf'
common_init_result: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
llama_params_fit_impl: projected to use 6810 MiB of device memory vs. 15473 MiB of free device memory
llama_params_fit_impl: will leave 8663 >= 1024 MiB of free device memory, no changes needed
llama_params_fit: successfully fit params to free device memory
llama_params_fit: fitting params to free memory took 0.16 seconds
llama_model_load_from_file_impl: using device SYCL0 (Intel(R) Arc(TM) A770 Graphics) (unknown id) - 15473 MiB free
llama_model_loader: loaded meta data with 42 key-value pairs and 399 tensors from /home/u/Downloads/models/gguf/qwen-3-vl-8b/Qwen3-VL-8B-Instruct-Q6_K.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-Instruct
llama_model_loader: - kv   3:                           general.finetune str              = Instruct
llama_model_loader: - kv   4:                           general.basename str              = Qwen3-Vl-8B-Instruct
llama_model_loader: - kv   5:                       general.quantized_by str              = Unsloth
llama_model_loader: - kv   6:                         general.size_label str              = 8B
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 8B 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:                        qwen3vl.block_count u32              = 36
llama_model_loader: - kv  15:                     qwen3vl.context_length u32              = 262144
llama_model_loader: - kv  16:                   qwen3vl.embedding_length u32              = 4096
llama_model_loader: - kv  17:                qwen3vl.feed_forward_length u32              = 12288
llama_model_loader: - kv  18:               qwen3vl.attention.head_count u32              = 32
llama_model_loader: - kv  19:            qwen3vl.attention.head_count_kv u32              = 8
llama_model_loader: - kv  20:                     qwen3vl.rope.freq_base f32              = 5000000.000000
llama_model_loader: - kv  21:   qwen3vl.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  22:               qwen3vl.attention.key_length u32              = 128
llama_model_loader: - kv  23:             qwen3vl.attention.value_length u32              = 128
llama_model_loader: - kv  24:            qwen3vl.rope.dimension_sections arr[i32,4]       = [24, 20, 20, 0]
llama_model_loader: - kv  25:                 qwen3vl.n_deepstack_layers u32              = 3
llama_model_loader: - kv  26:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  27:                         tokenizer.ggml.pre str              = qwen2
llama_model_loader: - kv  28:                      tokenizer.ggml.tokens arr[str,151936]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  29:                  tokenizer.ggml.token_type arr[i32,151936]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  30:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  31:                tokenizer.ggml.eos_token_id u32              = 151645
llama_model_loader: - kv  32:            tokenizer.ggml.padding_token_id u32              = 151654
llama_model_loader: - kv  33:                tokenizer.ggml.bos_token_id u32              = 151643
llama_model_loader: - kv  34:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  35:                    tokenizer.chat_template str              = {%- if tools %}\n    {{- '<|im_start|>...
llama_model_loader: - kv  36:               general.quantization_version u32              = 2
llama_model_loader: - kv  37:                          general.file_type u32              = 18
llama_model_loader: - kv  38:                      quantize.imatrix.file str              = Qwen3-VL-8B-Instruct-GGUF/imatrix_uns...
llama_model_loader: - kv  39:                   quantize.imatrix.dataset str              = unsloth_calibration_Qwen3-VL-8B-Instr...
llama_model_loader: - kv  40:             quantize.imatrix.entries_count u32              = 252
llama_model_loader: - kv  41:              quantize.imatrix.chunks_count u32              = 694
llama_model_loader: - type  f32:  145 tensors
llama_model_loader: - type q6_K:  254 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q6_K
print_info: file size   = 6.26 GiB (6.56 BPW) 
load: 0 unused tokens
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: no_alloc              = 0
print_info: n_ctx_train           = 262144
print_info: n_embd                = 4096
print_info: n_embd_inp            = 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_yarn_log_mul     = 0.0000
print_info: rope_finetuned        = unknown
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
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-Instruct
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, direct_io = false)
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
load_tensors: offloading output layer to GPU
load_tensors: offloading 35 repeating layers to GPU
load_tensors: offloaded 37/37 layers to GPU
load_tensors:   CPU_Mapped model buffer size =   486.86 MiB
load_tensors:        SYCL0 model buffer size =  5921.78 MiB
.......................................................................................
common_init_result: added <|endoftext|> logit bias = -inf
common_init_result: added <|im_end|> logit bias = -inf
common_init_result: added <|fim_pad|> logit bias = -inf
common_init_result: added <|repo_name|> logit bias = -inf
common_init_result: added <|file_sep|> logit bias = -inf
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 4096
llama_context: n_ctx_seq     = 4096
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_seq (4096) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
llama_context:  SYCL_Host  output buffer size =     0.58 MiB
llama_kv_cache:      SYCL0 KV buffer size =   576.00 MiB
llama_kv_cache: size =  576.00 MiB (  4096 cells,  36 layers,  1/1 seqs), K (f16):  288.00 MiB, V (f16):  288.00 MiB
sched_reserve: reserving ...
sched_reserve:      SYCL0 compute buffer size =   312.75 MiB
sched_reserve:  SYCL_Host compute buffer size =    68.02 MiB
sched_reserve: graph nodes  = 1275
sched_reserve: graph splits = 74
sched_reserve: reserve took 3.62 ms, sched copies = 1
clip_model_loader: model name:   Qwen3-Vl-8B-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 SYCL0 backend
load_hparams: Qwen-VL models require at minimum 1024 image tokens to function correctly on grounding tasks
load_hparams: if you encounter problems with accuracy, try adding --image-min-tokens 1024
load_hparams: more info: https://github.com/ggml-org/llama.cpp/issues/16842

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:         768
load_hparams: patch_size:         16
load_hparams: has_llava_proj:     0
load_hparams: minicpmv_version:   0
load_hparams: n_merge:            2
load_hparams: n_wa_pattern: 0
load_hparams: image_min_pixels:   8192
load_hparams: image_max_pixels:   2097152 (custom value)

load_hparams: model size:         1108.69 MiB
load_hparams: metadata size:      0.12 MiB
srv    load_model: loaded multimodal model, '/home/u/Downloads/models/gguf/qwen-3-vl-8b/mmproj-BF16.gguf'
srv    load_model: initializing slots, n_slots = 1
no implementations specified for speculative decoding
slot   load_model: id  0 | task -1 | speculative decoding context not initialized
slot   load_model: id  0 | task -1 | new slot, n_ctx = 4096
srv    load_model: prompt cache is enabled, size limit: 8192 MiB
srv    load_model: use `--cache-ram 0` to disable the prompt cache
srv    load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391
init: chat template, 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
'
srv          init: init: chat template, thinking = 0
main: model loaded
main: server is listening on http://0.0.0.0:8080
main: starting the main loop...
srv  update_slots: all slots are idle
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 -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist 
slot launch_slot_: id  0 | task 0 | processing task, is_child = 0
slot update_slots: id  0 | task 0 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 2009
slot update_slots: id  0 | task 0 | n_tokens = 0, memory_seq_rm [0, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_tokens = 8, batch.n_tokens = 8, progress = 0.003982
slot update_slots: id  0 | task 0 | n_tokens = 8, memory_seq_rm [8, end)
srv  process_chun: processing image...
encoding image slice...
alloc_compute_meta:      SYCL0 compute buffer size =   333.27 MiB
alloc_compute_meta:        CPU compute buffer size =   175.54 MiB
alloc_compute_meta: graph splits = 262, nodes = 853
warmup: *****************************************************************
warmup: WARNING: flash attention not supported by SYCL0, memory usage will increase
warmup: op params: 
warmup:    dst: type = f32, ne = [72 16 7980 1], nb = [4 288 4608 36771840]
warmup:   src0: type = f32, ne = [72 7980 16 1], nb = [4 4608 288 36771840]
warmup:   src1: type = f16, ne = [72 7980 16 1], nb = [2 144 1149120 18385920]
warmup:   src2: type = f16, ne = [72 7980 16 1], nb = [2 144 1149120 18385920]
warmup: please report this on github as an issue
warmup: *****************************************************************
alloc_compute_meta:      SYCL0 compute buffer size =  4272.50 MiB
alloc_compute_meta:        CPU compute buffer size =   175.54 MiB
alloc_compute_meta: graph splits = 235, nodes = 907
warmup: flash attention is disabled
warmup: *****************************************************************
warmup: WARNING: the CLIP graph uses unsupported operators by the backend
warmup:          the performance will be suboptimal                      
warmup:          list of unsupported ops (backend=SYCL0):
warmup:          UPSCALE: type = f32, ne = [114 70 1152 1]
warmup:          MUL_MAT: type = f32, ne = [3456 7980 1 1]
warmup:         SOFT_MAX: type = f32, ne = [7980 7980 16 1]
warmup:          PERMUTE: type = f32, ne = [72 16 7980 1]
warmup:              MUL: type = f32, ne = [1152 7980 1 1]
warmup:              ADD: type = f32, ne = [4304 7980 1 1]
warmup:              MUL: type = f32, ne = [1152 7980 1 1]
warmup:             CONT: type = f32, ne = [7980 72 16 1]
warmup:         SOFT_MAX: type = f32, ne = [7980 7980 16 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:          PERMUTE: type = f32, ne = [72 7980 16 1]
warmup:             CONT: type = f32, ne = [7980 72 16 1]
warmup:          MUL_MAT: type = f32, ne = [1152 7980 1 1]
warmup:             NORM: type = f32, ne = [1152 7980 1 1]
warmup:          MUL_MAT: type = f32, ne = [1152 7980 1 1]
warmup:             ROPE: type = f32, ne = [72 16 7980 1]
warmup:          PERMUTE: type = f32, ne = [72 7980 16 1]
warmup:          PERMUTE: type = f32, ne = [72 16 7980 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:              ADD: type = f32, ne = [4304 7980 1 1]
warmup:             VIEW: type = f32, ne = [72 16 7980 1]
warmup:             ROPE: type = f32, ne = [72 16 7980 1]
warmup:         SOFT_MAX: type = f32, ne = [7980 7980 16 1]
warmup:             CONT: type = f32, ne = [1152 7980 1 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:             VIEW: type = f32, ne = [72 16 7980 1]
warmup:             VIEW: type = f32, ne = [72 16 7980 1]
warmup:             CONT: type = f32, ne = [7980 72 16 1]
warmup:          MUL_MAT: type = f32, ne = [72 7980 16 1]
warmup:             NORM: type = f32, ne = [1152 7980 1 1]
warmup:          MUL_MAT: type = f32, ne = [3456 7980 1 1]
warmup:             VIEW: type = f32, ne = [72 16 7980 1]
warmup:          PERMUTE: type = f32, ne = [72 7980 16 1]
warmup:          MUL_MAT: type = f32, ne = [7980 7980 16 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:              MUL: type = f32, ne = [1152 7980 1 1]
warmup:          MUL_MAT: type = f32, ne = [3456 7980 1 1]
warmup:             ROPE: type = f32, ne = [72 16 7980 1]
warmup:          PERMUTE: type = f32, ne = [7980 72 16 1]
warmup:             CONT: type = f32, ne = [1152 7980 1 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:              MUL: type = f32, ne = [1152 7980 1 1]
warmup:             VIEW: type = f32, ne = [72 16 7980 1]
warmup:          PERMUTE: type = f32, ne = [72 7980 16 1]
warmup:          MUL_MAT: type = f32, ne = [4304 7980 1 1]
warmup:            UNARY: type = f32, ne = [4608 1995 1 1]
warmup:              ADD: type = f32, ne = [4096 1995 1 1]
warmup:             VIEW: type = f32, ne = [72 16 7980 1]
warmup:             ROPE: type = f32, ne = [72 16 7980 1]
warmup:          MUL_MAT: type = f32, ne = [7980 7980 16 1]
warmup:              ADD: type = f32, ne = [4304 7980 1 1]
warmup:          MUL_MAT: type = f32, ne = [1152 7980 1 1]
warmup:          MUL_MAT: type = f32, ne = [3456 7980 1 1]
warmup:             VIEW: type = f32, ne = [72 16 7980 1]
warmup:          PERMUTE: type = f32, ne = [7980 72 16 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:              ADD: type = f32, ne = [4304 7980 1 1]
warmup:              MUL: type = f32, ne = [1152 7980 1 1]
warmup:              ADD: type = f32, ne = [3456 7980 1 1]
warmup:          PERMUTE: type = f32, ne = [72 7980 16 1]
warmup:             NORM: type = f32, ne = [1152 7980 1 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:             ROPE: type = f32, ne = [72 16 7980 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:             NORM: type = f32, ne = [1152 7980 1 1]
warmup:          MUL_MAT: type = f32, ne = [1152 7980 1 1]
warmup:             NORM: type = f32, ne = [1152 7980 1 1]
warmup:             VIEW: type = f32, ne = [72 16 7980 1]
warmup:             CONT: type = f32, ne = [1152 7980 1 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:              ADD: type = f32, ne = [4304 7980 1 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:              ADD: type = f32, ne = [3456 7980 1 1]
warmup:          MUL_MAT: type = f32, ne = [72 7980 16 1]
warmup:             CONT: type = f32, ne = [1152 7980 1 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:            UNARY: type = f32, ne = [4304 7980 1 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:          MUL_MAT: type = f32, ne = [7980 7980 16 1]
warmup:          MUL_MAT: type = f32, ne = [72 7980 16 1]
warmup:             NORM: type = f32, ne = [1152 7980 1 1]
warmup:          MUL_MAT: type = f32, ne = [4304 7980 1 1]
warmup:          PERMUTE: type = f32, ne = [72 7980 16 1]
warmup:             NORM: type = f32, ne = [1152 7980 1 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:              MUL: type = f32, ne = [4608 1995 1 1]
warmup:              ADD: type = f32, ne = [4096 1995 1 1]
warmup:          PERMUTE: type = f32, ne = [72 7980 16 1]
warmup:          PERMUTE: type = f32, ne = [7980 72 16 1]
warmup:             CONT: type = f32, ne = [1152 7980 1 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:            UNARY: type = f32, ne = [4304 7980 1 1]
warmup:             ROPE: type = f32, ne = [72 16 7980 1]
warmup:          PERMUTE: type = f32, ne = [72 7980 16 1]
warmup:          MUL_MAT: type = f32, ne = [72 7980 16 1]
warmup:          MUL_MAT: type = f32, ne = [1152 7980 1 1]
warmup:          MUL_MAT: type = f32, ne = [4304 7980 1 1]
warmup:             VIEW: type = f32, ne = [72 16 7980 1]
warmup:             ROPE: type = f32, ne = [72 16 7980 1]
warmup:          MUL_MAT: type = f32, ne = [7980 7980 16 1]
warmup:          PERMUTE: type = f32, ne = [72 16 7980 1]
warmup:              MUL: type = f32, ne = [1152 7980 1 1]
warmup:              ADD: type = f32, ne = [3456 7980 1 1]
warmup:             VIEW: type = f32, ne = [72 16 7980 1]
warmup:          PERMUTE: type = f32, ne = [7980 72 16 1]
warmup:         SOFT_MAX: type = f32, ne = [7980 7980 16 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:              ADD: type = f32, ne = [3456 7980 1 1]
warmup:          PERMUTE: type = f32, ne = [72 7980 16 1]
warmup:             CONT: type = f32, ne = [7980 72 16 1]
warmup:          MUL_MAT: type = f32, ne = [1152 7980 1 1]
warmup:             NORM: type = f32, ne = [1152 7980 1 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:             ROPE: type = f32, ne = [72 16 7980 1]
warmup:          PERMUTE: type = f32, ne = [72 7980 16 1]
warmup:          PERMUTE: type = f32, ne = [72 16 7980 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:             NORM: type = f32, ne = [1152 7980 1 1]
warmup:             VIEW: type = f32, ne = [72 16 7980 1]
warmup:             ROPE: type = f32, ne = [72 16 7980 1]
warmup:          MUL_MAT: type = f32, ne = [4304 7980 1 1]
warmup:             VIEW: type = f32, ne = [72 16 7980 1]
warmup:             ROPE: type = f32, ne = [72 16 7980 1]
warmup:          MUL_MAT: type = f32, ne = [7980 7980 16 1]
warmup:          PERMUTE: type = f32, ne = [72 16 7980 1]
warmup:              MUL: type = f32, ne = [1152 7980 1 1]
warmup:          MUL_MAT: type = f32, ne = [4608 1995 1 1]
warmup:            UNARY: type = f32, ne = [4608 1995 1 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:             VIEW: type = f32, ne = [72 16 7980 1]
warmup:          PERMUTE: type = f32, ne = [7980 72 16 1]
warmup:         SOFT_MAX: type = f32, ne = [7980 7980 16 1]
warmup:          PERMUTE: type = f32, ne = [72 16 7980 1]
warmup:              ADD: type = f32, ne = [1152 7980 1 1]
warmup:              MUL: type = f32, ne = [1152 7980 1 1]
warmup:              ADD: type = f32, ne = [3456 7980 1 1]
warmup:          MUL_MAT: type = f32, ne = [72 7980 16 1]
warmup:          MUL_MAT: type = f32, ne = [1152 7980 1 1]
warmup: flash attention is disabled
warmup: please report this on github as an issue
warmup: ref: https://github.com/ggml-org/llama.cpp/pull/16837#issuecomment-3461676118
warmup: *****************************************************************
image slice encoded in 11647 ms
decoding image batch 1/1, n_tokens_batch = 1995
image decoded (batch 1/1) in 11241 ms
srv  process_chun: image processed in 22888 ms
slot update_slots: id  0 | task 0 | prompt processing progress, n_tokens = 2009, batch.n_tokens = 6, progress = 1.000000
slot update_slots: id  0 | task 0 | prompt done, n_tokens = 2009, batch.n_tokens = 6
slot init_sampler: id  0 | task 0 | init sampler, took 0.01 ms, tokens: text = 14, total = 2009
slot print_timing: id  0 | task 0 | 
prompt eval time =   23797.18 ms /  2009 tokens (   11.85 ms per token,    84.42 tokens per second)
       eval time =   45940.98 ms /   461 tokens (   99.66 ms per token,    10.03 tokens per second)
      total time =   69738.16 ms /  2470 tokens
slot      release: id  0 | task 0 | stop processing: n_tokens = 2469, truncated = 0
srv  update_slots: all slots are idle
srv  log_server_r: done request: POST /v1/chat/completions 10.0.0.102 200

Reference

This relates to PR #16837 (SYCL support for Qwen-VL). The log includes a direct reference to it.

[NeoZhangJianyu]

Got it! Let me check!

Thank you for your feedback!

[air-fuel-ratio]

Looks like UPSCALE is the only one remaining?

warmup: *****************************************************************
warmup: WARNING: the CLIP graph uses unsupported operators by the backend
warmup:          the performance will be suboptimal                      
warmup:          list of unsupported ops (backend=SYCL0):
warmup:          UPSCALE: type = f32, ne = [92 92 1024 1]
warmup: flash attention is enabled
warmup: please report this on github as an issue
warmup: ref: https://github.com/ggml-org/llama.cpp/pull/16837#issuecomment-3461676118
warmup: *****************************************************************
srv    load_model: loaded multimodal model, '/models/mmproj-Qwen3-VL-4B-Instruct-F16.gguf'

[NeoZhangJianyu]

@air-fuel-ratio
UPSCALE is supported for mode=nearest only now.
Let me check it.

Thank you!

[Hisir0909]

why in intel ArcA770, image slice encoded so slow. @NeoZhangJianyu , Is this normal? The previous problem of missing operators seems to have disappeared

test-case:

description picture

Logs

❯ source /opt/intel/oneapi/setvars.sh

:: initializing oneAPI environment ...
   zsh: ZSH_VERSION = 5.9
   args: Using "$@" for setvars.sh arguments:
:: advisor -- latest
:: ccl -- latest
:: compiler -- latest
:: dal -- latest
:: debugger -- latest
:: dev-utilities -- latest
:: dnnl -- latest
:: dpcpp-ct -- latest
:: dpl -- latest
:: ipp -- latest
:: ippcp -- latest
:: mkl -- latest
:: mpi -- latest
:: tbb -- latest
:: umf -- latest
:: vtune -- latest
:: oneAPI environment initialized ::

❯ ./llama-server -m ~/Downloads/models/gguf/qwen-3-vl-8b/Qwen3-VL-8B-Instruct-Q6_K.gguf --mmproj ~/Downloads/models/gguf/qwen-3-vl-8b/mmproj-BF16.gguf -ngl 999 -c 4096 -n 512 -fa on --host 0.0.0.0  --image-max-tokens 2048 -np 1
build: 8477 (ec2b787eb) with IntelLLVM 2025.3.0 for Linux x86_64
system info: n_threads = 6, n_threads_batch = 6, total_threads = 12

system_info: n_threads = 6 (n_threads_batch = 6) / 12 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |

init: using 11 threads for HTTP server
start: binding port with default address family
main: loading model
srv    load_model: loading model '/home/u/Downloads/models/gguf/qwen-3-vl-8b/Qwen3-VL-8B-Instruct-Q6_K.gguf'
common_init_result: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
llama_params_fit_impl: projected to use 6802 MiB of device memory vs. 15473 MiB of free device memory
llama_params_fit_impl: will leave 8671 >= 1024 MiB of free device memory, no changes needed
llama_params_fit: successfully fit params to free device memory
llama_params_fit: fitting params to free memory took 0.16 seconds
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
llama_model_load_from_file_impl: using device SYCL0 (Intel(R) Arc(TM) A770 Graphics) (unknown id) - 15473 MiB free
llama_model_loader: loaded meta data with 42 key-value pairs and 399 tensors from /home/u/Downloads/models/gguf/qwen-3-vl-8b/Qwen3-VL-8B-Instruct-Q6_K.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-Instruct
llama_model_loader: - kv   3:                           general.finetune str              = Instruct
llama_model_loader: - kv   4:                           general.basename str              = Qwen3-Vl-8B-Instruct
llama_model_loader: - kv   5:                       general.quantized_by str              = Unsloth
llama_model_loader: - kv   6:                         general.size_label str              = 8B
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 8B 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:                        qwen3vl.block_count u32              = 36
llama_model_loader: - kv  15:                     qwen3vl.context_length u32              = 262144
llama_model_loader: - kv  16:                   qwen3vl.embedding_length u32              = 4096
llama_model_loader: - kv  17:                qwen3vl.feed_forward_length u32              = 12288
llama_model_loader: - kv  18:               qwen3vl.attention.head_count u32              = 32
llama_model_loader: - kv  19:            qwen3vl.attention.head_count_kv u32              = 8
llama_model_loader: - kv  20:                     qwen3vl.rope.freq_base f32              = 5000000.000000
llama_model_loader: - kv  21:   qwen3vl.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  22:               qwen3vl.attention.key_length u32              = 128
llama_model_loader: - kv  23:             qwen3vl.attention.value_length u32              = 128
llama_model_loader: - kv  24:            qwen3vl.rope.dimension_sections arr[i32,4]       = [24, 20, 20, 0]
llama_model_loader: - kv  25:                 qwen3vl.n_deepstack_layers u32              = 3
llama_model_loader: - kv  26:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  27:                         tokenizer.ggml.pre str              = qwen2
llama_model_loader: - kv  28:                      tokenizer.ggml.tokens arr[str,151936]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  29:                  tokenizer.ggml.token_type arr[i32,151936]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  30:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  31:                tokenizer.ggml.eos_token_id u32              = 151645
llama_model_loader: - kv  32:            tokenizer.ggml.padding_token_id u32              = 151654
llama_model_loader: - kv  33:                tokenizer.ggml.bos_token_id u32              = 151643
llama_model_loader: - kv  34:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  35:                    tokenizer.chat_template str              = {%- if tools %}\n    {{- '<|im_start|>...
llama_model_loader: - kv  36:               general.quantization_version u32              = 2
llama_model_loader: - kv  37:                          general.file_type u32              = 18
llama_model_loader: - kv  38:                      quantize.imatrix.file str              = Qwen3-VL-8B-Instruct-GGUF/imatrix_uns...
llama_model_loader: - kv  39:                   quantize.imatrix.dataset str              = unsloth_calibration_Qwen3-VL-8B-Instr...
llama_model_loader: - kv  40:             quantize.imatrix.entries_count u32              = 252
llama_model_loader: - kv  41:              quantize.imatrix.chunks_count u32              = 694
llama_model_loader: - type  f32:  145 tensors
llama_model_loader: - type q6_K:  254 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q6_K
print_info: file size   = 6.26 GiB (6.56 BPW)
load: 0 unused tokens
load: control-looking token: 128247 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden
load: printing all EOG tokens:
load:   - 128247 ('</s>')
load:   - 151643 ('<|endoftext|>')
load:   - 151645 ('<|im_end|>')
load:   - 151662 ('<|fim_pad|>')
load:   - 151663 ('<|repo_name|>')
load:   - 151664 ('<|file_sep|>')
load: special tokens cache size = 27
load: token to piece cache size = 0.9311 MB
print_info: arch                  = qwen3vl
print_info: vocab_only            = 0
print_info: no_alloc              = 0
print_info: n_ctx_train           = 262144
print_info: n_embd                = 4096
print_info: n_embd_inp            = 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_yarn_log_mul     = 0.0000
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-Instruct
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|>'
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
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             = 128247 '</s>'
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, direct_io = false)
get_memory_info: [warning] ext_intel_free_memory is not supported (export/set ZES_ENABLE_SYSMAN=1 to support), use total memory as free memory
load_tensors: offloading output layer to GPU
load_tensors: offloading 35 repeating layers to GPU
load_tensors: offloaded 37/37 layers to GPU
load_tensors:   CPU_Mapped model buffer size =   486.86 MiB
load_tensors:        SYCL0 model buffer size =  5921.78 MiB
.......................................................................................
common_init_result: added </s> logit bias = -inf
common_init_result: added <|endoftext|> logit bias = -inf
common_init_result: added <|im_end|> logit bias = -inf
common_init_result: added <|fim_pad|> logit bias = -inf
common_init_result: added <|repo_name|> logit bias = -inf
common_init_result: added <|file_sep|> logit bias = -inf
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 4096
llama_context: n_ctx_seq     = 4096
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_seq (4096) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
llama_context:  SYCL_Host  output buffer size =     0.58 MiB
llama_kv_cache:      SYCL0 KV buffer size =   576.00 MiB
llama_kv_cache: size =  576.00 MiB (  4096 cells,  36 layers,  1/1 seqs), K (f16):  288.00 MiB, V (f16):  288.00 MiB
sched_reserve: reserving ...
sched_reserve: resolving fused Gated Delta Net support:
sched_reserve: fused Gated Delta Net (autoregressive) enabled
sched_reserve: fused Gated Delta Net (chunked) enabled
sched_reserve:      SYCL0 compute buffer size =   304.75 MiB
sched_reserve:  SYCL_Host compute buffer size =    48.02 MiB
sched_reserve: graph nodes  = 1275
sched_reserve: graph splits = 2
sched_reserve: reserve took 3.29 ms, sched copies = 1
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-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 SYCL0 backend
load_hparams: Qwen-VL models require at minimum 1024 image tokens to function correctly on grounding tasks
load_hparams: if you encounter problems with accuracy, try adding --image-min-tokens 1024
load_hparams: more info: https://github.com/ggml-org/llama.cpp/issues/16842

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:         768
load_hparams: patch_size:         16
load_hparams: has_llava_proj:     0
load_hparams: minicpmv_version:   0
load_hparams: n_merge:            2
load_hparams: n_wa_pattern: 0
load_hparams: image_min_pixels:   8192
load_hparams: image_max_pixels:   2097152 (custom value)

load_hparams: model size:         1108.69 MiB
load_hparams: metadata size:      0.12 MiB
warmup: warmup with image size = 1472 x 1472
alloc_compute_meta:      SYCL0 compute buffer size =   372.08 MiB
alloc_compute_meta:        CPU compute buffer size =    24.93 MiB
alloc_compute_meta: graph splits = 1, nodes = 853
warmup: flash attention is enabled
srv    load_model: loaded multimodal model, '/home/u/Downloads/models/gguf/qwen-3-vl-8b/mmproj-BF16.gguf'
srv    load_model: initializing slots, n_slots = 1
no implementations specified for speculative decoding
slot   load_model: id  0 | task -1 | speculative decoding context not initialized
slot   load_model: id  0 | task -1 | new slot, n_ctx = 4096
srv    load_model: prompt cache is enabled, size limit: 8192 MiB
srv    load_model: use `--cache-ram 0` to disable the prompt cache
srv    load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391
init: chat template, 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
'
srv          init: init: chat template, thinking = 0
main: model loaded
main: server is listening on http://0.0.0.0:8080
main: starting the main loop...
srv  update_slots: all slots are idle
srv  params_from_: Chat format: peg-native
slot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = -1
slot launch_slot_: id  0 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist
slot launch_slot_: id  0 | task 0 | processing task, is_child = 0
slot update_slots: id  0 | task 0 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 2037
slot update_slots: id  0 | task 0 | n_tokens = 0, memory_seq_rm [0, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_tokens = 6, batch.n_tokens = 6, progress = 0.002946
srv  log_server_r: done request: POST /v1/chat/completions 10.0.0.102 200
slot update_slots: id  0 | task 0 | n_tokens = 6, memory_seq_rm [6, end)
srv  process_chun: processing image...
encoding image slice...
image slice encoded in 13899 ms
decoding image batch 1/1, n_tokens_batch = 2025
image decoded (batch 1/1) in 4560 ms
srv  process_chun: image processed in 18459 ms
slot init_sampler: id  0 | task 0 | init sampler, took 0.01 ms, tokens: text = 12, total = 2037
slot update_slots: id  0 | task 0 | prompt processing done, n_tokens = 2037, batch.n_tokens = 6

[NeoZhangJianyu]

@Hisir0909
Could you share the download url of the two gguf files?
I found them in different author and get the mismatch error.

[Hisir0909]

@Hisir0909 Could you share the download url of the two gguf files? I found them in different author and get the mismatch error.

this: https://huggingface.co/unsloth/Qwen3-VL-8B-Instruct-GGUF

model	url	SHA256
Qwen3-VL-8B-Instruct-Q6_K.gguf	https://huggingface.co/unsloth/Qwen3-VL-8B-Instruct-GGUF/resolve/main/Qwen3-VL-8B-Instruct-Q6_K.gguf	2f963d09b4c7485df9bfd4ac4d7ce192cc9cc004a45555b605eeb0a7a83f0057
mmproj-BF16.gguf	https://huggingface.co/unsloth/Qwen3-VL-8B-Instruct-GGUF/resolve/main/mmproj-BF16.gguf	6516bb64bae1503a0fcd7ec9fa39655f8c481580be0a0a066397941d9761c9f4

Thank you！

[NeoZhangJianyu]

@Hisir0909
I can run the models by the same cmd.
The time is about 14000 ms on Arc 770.

I run the model on first time.
I feel the response speed is OK since the image is 1129*1127.

[lotfimaria331-ops]

Thank you Le mer. 25 mar. 2026, 08:12, Neo Zhang Jianyu ***@***.***> a écrit :

…

*NeoZhangJianyu* left a comment (ggml-org/llama.cpp#19543) <#19543 (comment)> @Hisir0909 <https://github.com/Hisir0909> I can run the models by the same cmd. The time is about 14000 ms on Arc 770. I run the model on first time. I feel the response speed is OK since the image is 1129*1127. — Reply to this email directly, view it on GitHub <#19543?email_source=notifications&email_token=B7VCPZTADAAI6OZEYLIPBVL4SPLL7A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIMJSGY2TANBZGE4KM4TFMFZW63VKON2WE43DOJUWEZLEUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#issuecomment-4126504918>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/B7VCPZUDQ3MUJWM27HIESUL4SPLL7AVCNFSM6AAAAACU3R2HA2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DCMRWGUYDIOJRHA> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[Hisir0909]

@Hisir0909 I can run the models by the same cmd. The time is about 14000 ms on Arc 770.

I run the model on first time. I feel the response speed is OK since the image is 1129*1127.

Is this actually a normal speed?
It seems that the Intel Arc A770 is really no longer good I can run much faster with the 3060 than it。
Thank you for clarifying your doubts

./llama-server -m ~/models/qwen3.5-9b/Qwen3.5-9B-UD-Q6_K_XL.gguf  --mmproj ~/models/qwe
n3.5-9b/mmproj-BF16.gguf -c 7000 -n 512 --host 0.0.0.0 --port 8097  -b 4096


encoding image slice...                                                                                                         
image slice encoded in 4736 ms                                                                                                  
decoding image batch 1/1, n_tokens_batch = 2809                                                                                 
find_slot: non-consecutive token position 6 after 5 for sequence 3 with 512 new tokens                                          
find_slot: non-consecutive token position 6 after 6 for sequence 3 with 512 new tokens                                          
find_slot: non-consecutive token position 6 after 6 for sequence 3 with 512 new tokens                                          
find_slot: non-consecutive token position 6 after 6 for sequence 3 with 512 new tokens                                          
find_slot: non-consecutive token position 6 after 6 for sequence 3 with 512 new tokens                                          
find_slot: non-consecutive token position 6 after 6 for sequence 3 with 249 new tokens                                          
find_slot: non-consecutive token position 6 after 5 for sequence 3 with 512 new tokens                                          
find_slot: non-consecutive token position 6 after 6 for sequence 3 with 512 new tokens                                          
find_slot: non-consecutive token position 6 after 6 for sequence 3 with 512 new tokens                                          
find_slot: non-consecutive token position 6 after 6 for sequence 3 with 512 new tokens                                          
find_slot: non-consecutive token position 6 after 6 for sequence 3 with 512 new tokens                                          
find_slot: non-consecutive token position 6 after 6 for sequence 3 with 249 new tokens                                          
image decoded (batch 1/1) in 1693 ms                                                                                            
srv  process_chun: image processed in 6429 ms

[NeoZhangJianyu]

@Hisir0909
Yes, there is the performance gap.
There will be optimization of mul_mat in next.
Hope it help this case!

Thank you!