Eval bug: mtmd: "flash attention is disabled / please report this on github as an issue" #16950
Description
Name and 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 4060 Ti, compute capability 8.9, VMM: yes
version: 6926 (2f966b8ed)
built with cc (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0 for x86_64-linux-gnuOperating systems
Linux
GGML backends
CUDA
Hardware
4060 Ti 16GB
Ryzen 7 7700X
128GB DDR5-6400 @ 4800 MT/s
Models
Qwen_Qwen3-VL-32B-Instruct-Q5_K_M.gguf from 🤗 bartowski/Qwen_Qwen3-VL-32B-Instruct-GGUF
Same error occurs on other Qwen3-VL models when using vision
Problem description & steps to reproduce
The following warning messages appear when loading Qwen3-VL-32B model with its bf16 mmproj when flash_attn is on or auto. The warning does not appear when flash_attn is off.
warmup: *****************************************************************
warmup: WARNING: flash attention not supported by CUDA0, memory usage will increase
warmup: op params:
warmup: dst: type = f32, ne = [72 16 1024 1], nb = [4 288 4608 4718592]
warmup: src0: type = f32, ne = [72 1024 16 1], nb = [4 4608 288 4718592]
warmup: src1: type = f16, ne = [72 1024 16 1], nb = [2 144 147456 2359296]
warmup: src2: type = f16, ne = [72 1024 16 1], nb = [2 144 147456 2359296]
warmup: please report this on github as an issue
warmup: *****************************************************************
alloc_compute_meta: warmup with image size = 512 x 512
alloc_compute_meta: CUDA0 compute buffer size = 112.02 MiB
alloc_compute_meta: CPU compute buffer size = 13.50 MiB
alloc_compute_meta: graph splits = 1, 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=CUDA0):
warmup: SOFT_MAX: type = f32, ne = [1024 1024 16 1]
warmup: CONT: type = f32, ne = [1024 72 16 1]
warmup: PERMUTE: type = f32, ne = [72 1024 16 1]
warmup: ROPE: type = f32, ne = [72 16 1024 1]
warmup: VIEW: type = f32, ne = [72 16 1024 1]
warmup: VIEW: type = f32, ne = [72 16 1024 1]
warmup: MUL_MAT: type = f32, ne = [3456 1024 1 1]
warmup: MUL: type = f32, ne = [1152 1024 1 1]
warmup: ADD: type = f32, ne = [1152 1024 1 1]
warmup: UNARY: type = f32, ne = [4608 256 1 1]
warmup: ADD: type = f32, ne = [4304 1024 1 1]
warmup: ADD: type = f32, ne = [1152 1024 1 1]
warmup: NORM: type = f32, ne = [1152 1024 1 1]
warmup: ADD: type = f32, ne = [1152 1024 1 1]
warmup: CONT: type = f32, ne = [1152 1024 1 1]
warmup: MUL_MAT: type = f32, ne = [72 1024 16 1]
warmup: MUL_MAT: type = f32, ne = [1024 1024 16 1]
warmup: NORM: type = f32, ne = [1152 1024 1 1]
warmup: PERMUTE: type = f32, ne = [72 1024 16 1]
warmup: ROPE: type = f32, ne = [72 16 1024 1]
warmup: VIEW: type = f32, ne = [72 16 1024 1]
warmup: ADD: type = f32, ne = [3456 1024 1 1]
warmup: ADD: type = f32, ne = [1152 1024 1 1]
warmup: NORM: type = f32, ne = [1152 1024 1 1]
warmup: ADD: type = f32, ne = [1152 1024 1 1]
warmup: VIEW: type = f32, ne = [72 16 1024 1]
warmup: MUL_MAT: type = f32, ne = [4608 256 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: *****************************************************************
First Bad Commit
2f966b8ed87514e74bb96592217226cb6a6974dd
Relevant log output
Command
exportv LLAMA_ARG_THREADS 8
exportv LLAMA_LOG_COLORS on
exportv LLAMA_LOG_PREFIX 1
exportv LLAMA_LOG_TIMESTAMPS 1
exportv LLAMA_ARG_CTX_CHECKPOINTS 32
exportv LLAMA_ARG_CONTEXT_SHIFT 0
exportv LLAMA_ARG_CACHE_REUSE 0
exportv LLAMA_ARG_CACHE_RAM 51200
exportv LLAMA_ARG_THREADS_HTTP $(nproc)
exportv LLAMA_ARG_ALIAS llama
exportv LLAMA_ARG_SSL_KEY_FILE ~/keys/server.key
exportv LLAMA_ARG_SSL_CERT_FILE ~/keys/server.crt
exportv LLAMA_API_KEY GERGANOV
exportv LLAMA_ARG_HOST 0.0.0.0
exportv LLAMA_ARG_PORT 12800
exportv LLAMA_ARG_NO_MMAP 1
exportv LLAMA_ARG_MLOCK 0
unset LLAMA_CHAT_TEMPLATE_KWARGS
#exportv LLAMA_CHAT_TEMPLATE_KWARGS "{\"enable_thinking\": false}"
unset LLAMA_ARG_CHAT_TEMPLATE_FILE
exportv LLAMA_ARG_JINJA 1
exportv LLAMA_ARG_MODEL ~/gguf/Qwen3-VL-32B-Instruct-Q5_K_M.gguf
exportv LLAMA_ARG_FLASH_ATTN auto
exportv LLAMA_ARG_CTX_SIZE 8192
exportv LLAMA_ARG_N_PARALLEL 1
exportv LLAMA_ARG_BATCH 1024
exportv LLAMA_ARG_UBATCH 1024
exportv LLAMA_ARG_N_GPU_LAYERS 999
exportv OT "ffn_(?:up|gate|down)=CPU"
exportv LLAMA_ARG_MMPROJ ~/gguf/mmproj-Qwen3-VL-32B-Instruct-bf16.gguf
~/llama.cpp/build/bin/llama-server -ot $OT --samplers "top_k;top_p;temperature" --top-k 64 --top-p 0.9 --temp 1.25Console 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 4060 Ti, compute capability 8.9, VMM: yes
0.00.000.478 W main: setting n_parallel = 4 and kv_unified = true
0.00.000.488 I build: 6926 (2f966b8ed) with cc (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0 for x86_64-linux-gnu
0.00.000.522 I system info: n_threads = 8, n_threads_batch = 8, total_threads = 16
0.00.000.523 I
0.00.000.581 I system_info: n_threads = 8 (n_threads_batch = 8) / 16 | CUDA : ARCHS = 890 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | FA_ALL_QUANTS = 1 | 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 |
0.00.000.582 I
0.00.000.582 I Running with SSL: key = /home/dylan/keys/server.key, cert = /home/dylan/keys/server.crt
0.00.002.629 I main: binding port with default address family
0.00.003.750 I main: HTTP server is listening, hostname: 0.0.0.0, port: 12800, http threads: 16
0.00.003.752 I main: loading model
0.00.003.752 I srv load_model: loading model '/home/dylan/gguf/Qwen3-VL-32B-Instruct-Q5_K_M.gguf'
0.00.100.528 I llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 4060 Ti) (0000:01:00.0) - 15805 MiB free
0.00.124.519 I llama_model_loader: loaded meta data with 36 key-value pairs and 707 tensors from /home/dylan/gguf/Qwen3-VL-32B-Instruct-Q5_K_M.gguf (version GGUF V3 (latest))
0.00.124.535 I llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
0.00.124.542 I llama_model_loader: - kv 0: general.architecture str = qwen3vl
0.00.124.543 I llama_model_loader: - kv 1: general.type str = model
0.00.124.543 I llama_model_loader: - kv 2: general.name str = Qwen3 VL 32B Instruct
0.00.124.544 I llama_model_loader: - kv 3: general.finetune str = Instruct
0.00.124.544 I llama_model_loader: - kv 4: general.basename str = Qwen3-VL
0.00.124.544 I llama_model_loader: - kv 5: general.size_label str = 32B
0.00.124.544 I llama_model_loader: - kv 6: general.license str = apache-2.0
0.00.124.559 I llama_model_loader: - kv 7: general.tags arr[str,1] = ["image-text-to-text"]
0.00.124.561 I llama_model_loader: - kv 8: qwen3vl.block_count u32 = 64
0.00.124.562 I llama_model_loader: - kv 9: qwen3vl.context_length u32 = 262144
0.00.124.563 I llama_model_loader: - kv 10: qwen3vl.embedding_length u32 = 5120
0.00.124.563 I llama_model_loader: - kv 11: qwen3vl.feed_forward_length u32 = 25600
0.00.124.563 I llama_model_loader: - kv 12: qwen3vl.attention.head_count u32 = 64
0.00.124.564 I llama_model_loader: - kv 13: qwen3vl.attention.head_count_kv u32 = 8
0.00.124.569 I llama_model_loader: - kv 14: qwen3vl.rope.freq_base f32 = 5000000.000000
0.00.124.569 I llama_model_loader: - kv 15: qwen3vl.attention.layer_norm_rms_epsilon f32 = 0.000001
0.00.124.570 I llama_model_loader: - kv 16: qwen3vl.attention.key_length u32 = 128
0.00.124.570 I llama_model_loader: - kv 17: qwen3vl.attention.value_length u32 = 128
0.00.124.572 I llama_model_loader: - kv 18: qwen3vl.rope.dimension_sections arr[i32,4] = [24, 20, 20, 0]
0.00.124.572 I llama_model_loader: - kv 19: qwen3vl.n_deepstack_layers u32 = 3
0.00.124.573 I llama_model_loader: - kv 20: tokenizer.ggml.model str = gpt2
0.00.124.573 I llama_model_loader: - kv 21: tokenizer.ggml.pre str = qwen2
0.00.135.533 I llama_model_loader: - kv 22: tokenizer.ggml.tokens arr[str,151936] = ["!", "\"", "#", "$", "%", "&", "'", ...
0.00.138.776 I llama_model_loader: - kv 23: tokenizer.ggml.token_type arr[i32,151936] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
0.00.150.481 I llama_model_loader: - kv 24: tokenizer.ggml.merges arr[str,151387] = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
0.00.150.483 I llama_model_loader: - kv 25: tokenizer.ggml.eos_token_id u32 = 151645
0.00.150.483 I llama_model_loader: - kv 26: tokenizer.ggml.padding_token_id u32 = 151643
0.00.150.483 I llama_model_loader: - kv 27: tokenizer.ggml.bos_token_id u32 = 151643
0.00.150.484 I llama_model_loader: - kv 28: tokenizer.ggml.add_bos_token bool = false
0.00.150.485 I llama_model_loader: - kv 29: tokenizer.chat_template str = {%- if tools %}\n {{- '<|im_start|>...
0.00.150.485 I llama_model_loader: - kv 30: general.quantization_version u32 = 2
0.00.150.485 I llama_model_loader: - kv 31: general.file_type u32 = 17
0.00.150.486 I llama_model_loader: - kv 32: quantize.imatrix.file str = /models_out/Qwen3-VL-32B-Instruct-GGU...
0.00.150.486 I llama_model_loader: - kv 33: quantize.imatrix.dataset str = /training_dir/calibration_datav5.txt
0.00.150.486 I llama_model_loader: - kv 34: quantize.imatrix.entries_count u32 = 448
0.00.150.487 I llama_model_loader: - kv 35: quantize.imatrix.chunks_count u32 = 818
0.00.150.488 I llama_model_loader: - type f32: 257 tensors
0.00.150.489 I llama_model_loader: - type q5_K: 385 tensors
0.00.150.489 I llama_model_loader: - type q6_K: 65 tensors
0.00.150.491 I print_info: file format = GGUF V3 (latest)
0.00.150.491 I print_info: file type = Q5_K - Medium
0.00.150.492 I print_info: file size = 21.61 GiB (5.67 BPW)
0.00.207.400 I load: printing all EOG tokens:
0.00.207.401 I load: - 151643 ('<|endoftext|>')
0.00.207.402 I load: - 151645 ('<|im_end|>')
0.00.207.402 I load: - 151662 ('<|fim_pad|>')
0.00.207.403 I load: - 151663 ('<|repo_name|>')
0.00.207.403 I load: - 151664 ('<|file_sep|>')
0.00.207.470 I load: special tokens cache size = 26
0.00.241.920 I load: token to piece cache size = 0.9311 MB
0.00.241.931 I print_info: arch = qwen3vl
0.00.241.931 I print_info: vocab_only = 0
0.00.241.931 I print_info: n_ctx_train = 262144
0.00.241.931 I print_info: n_embd = 20480
0.00.241.931 I print_info: n_layer = 64
0.00.241.939 I print_info: n_head = 64
0.00.241.941 I print_info: n_head_kv = 8
0.00.241.941 I print_info: n_rot = 128
0.00.241.941 I print_info: n_swa = 0
0.00.241.943 I print_info: is_swa_any = 0
0.00.241.943 I print_info: n_embd_head_k = 128
0.00.241.943 I print_info: n_embd_head_v = 128
0.00.241.945 I print_info: n_gqa = 8
0.00.241.947 I print_info: n_embd_k_gqa = 1024
0.00.241.948 I print_info: n_embd_v_gqa = 1024
0.00.241.949 I print_info: f_norm_eps = 0.0e+00
0.00.241.950 I print_info: f_norm_rms_eps = 1.0e-06
0.00.241.950 I print_info: f_clamp_kqv = 0.0e+00
0.00.241.950 I print_info: f_max_alibi_bias = 0.0e+00
0.00.241.950 I print_info: f_logit_scale = 0.0e+00
0.00.241.951 I print_info: f_attn_scale = 0.0e+00
0.00.241.952 I print_info: n_ff = 25600
0.00.241.952 I print_info: n_expert = 0
0.00.241.952 I print_info: n_expert_used = 0
0.00.241.953 I print_info: n_expert_groups = 0
0.00.241.953 I print_info: n_group_used = 0
0.00.241.953 I print_info: causal attn = 1
0.00.241.953 I print_info: pooling type = 0
0.00.241.953 I print_info: rope type = 40
0.00.241.954 I print_info: rope scaling = linear
0.00.241.955 I print_info: freq_base_train = 5000000.0
0.00.241.955 I print_info: freq_scale_train = 1
0.00.241.955 I print_info: n_ctx_orig_yarn = 262144
0.00.241.955 I print_info: rope_finetuned = unknown
0.00.241.956 I print_info: mrope sections = [24, 20, 20, 0]
0.00.241.956 I print_info: model type = 32B
0.00.241.957 I print_info: model params = 32.76 B
0.00.241.957 I print_info: general.name = Qwen3 VL 32B Instruct
0.00.241.960 I print_info: vocab type = BPE
0.00.241.961 I print_info: n_vocab = 151936
0.00.241.961 I print_info: n_merges = 151387
0.00.241.962 I print_info: BOS token = 151643 '<|endoftext|>'
0.00.241.962 I print_info: EOS token = 151645 '<|im_end|>'
0.00.241.962 I print_info: EOT token = 151645 '<|im_end|>'
0.00.241.962 I print_info: PAD token = 151643 '<|endoftext|>'
0.00.241.963 I print_info: LF token = 198 'Ċ'
0.00.241.963 I print_info: FIM PRE token = 151659 '<|fim_prefix|>'
0.00.241.963 I print_info: FIM SUF token = 151661 '<|fim_suffix|>'
0.00.241.963 I print_info: FIM MID token = 151660 '<|fim_middle|>'
0.00.241.963 I print_info: FIM PAD token = 151662 '<|fim_pad|>'
0.00.241.963 I print_info: FIM REP token = 151663 '<|repo_name|>'
0.00.241.964 I print_info: FIM SEP token = 151664 '<|file_sep|>'
0.00.241.964 I print_info: EOG token = 151643 '<|endoftext|>'
0.00.241.964 I print_info: EOG token = 151645 '<|im_end|>'
0.00.241.964 I print_info: EOG token = 151662 '<|fim_pad|>'
0.00.241.964 I print_info: EOG token = 151663 '<|repo_name|>'
0.00.241.964 I print_info: EOG token = 151664 '<|file_sep|>'
0.00.241.964 I print_info: max token length = 256
0.00.241.965 I load_tensors: loading model tensors, this can take a while... (mmap = false)
0.06.347.134 I load_tensors: offloading 64 repeating layers to GPU
0.06.347.136 I load_tensors: offloading output layer to GPU
0.06.347.137 I load_tensors: offloaded 65/65 layers to GPU
0.06.347.143 I load_tensors: CPU model buffer size = 510.04 MiB
0.06.347.144 I load_tensors: CUDA0 model buffer size = 4592.40 MiB
0.06.347.145 I load_tensors: CUDA_Host model buffer size = 17031.25 MiB
.................................................................................................
0.08.133.282 I llama_context: constructing llama_context
0.08.133.284 I llama_context: n_seq_max = 4
0.08.133.285 I llama_context: n_ctx = 8192
0.08.133.285 I llama_context: n_ctx_seq = 8192
0.08.133.285 I llama_context: n_batch = 1024
0.08.133.285 I llama_context: n_ubatch = 1024
0.08.133.286 I llama_context: causal_attn = 1
0.08.133.287 I llama_context: flash_attn = auto
0.08.133.287 I llama_context: kv_unified = true
0.08.133.291 I llama_context: freq_base = 5000000.0
0.08.133.291 I llama_context: freq_scale = 1
0.08.133.291 W llama_context: n_ctx_seq (8192) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
0.08.134.335 I llama_context: CUDA_Host output buffer size = 2.32 MiB
0.08.134.593 I llama_kv_cache: CUDA0 KV buffer size = 2048.00 MiB
0.08.142.396 I llama_kv_cache: size = 2048.00 MiB ( 8192 cells, 64 layers, 4/1 seqs), K (f16): 1024.00 MiB, V (f16): 1024.00 MiB
0.08.143.107 I llama_context: Flash Attention was auto, set to enabled
0.08.163.141 I llama_context: CUDA0 compute buffer size = 613.50 MiB
0.08.163.142 I llama_context: CUDA_Host compute buffer size = 52.04 MiB
0.08.163.143 I llama_context: graph nodes = 2247
0.08.163.143 I llama_context: graph splits = 194 (with bs=1024), 130 (with bs=1)
0.08.164.054 I common_init_from_params: added <|endoftext|> logit bias = -inf
0.08.164.055 I common_init_from_params: added <|im_end|> logit bias = -inf
0.08.164.056 I common_init_from_params: added <|fim_pad|> logit bias = -inf
0.08.164.056 I common_init_from_params: added <|repo_name|> logit bias = -inf
0.08.164.056 I common_init_from_params: added <|file_sep|> logit bias = -inf
0.08.164.059 I common_init_from_params: setting dry_penalty_last_n to ctx_size = 8192
0.08.164.059 W 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 32B 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: 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: 5120
--- 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
load_hparams: model size: 1144.71 MiB
load_hparams: metadata size: 0.12 MiB
alloc_compute_meta: warmup with image size = 512 x 512
alloc_compute_meta: CUDA0 compute buffer size = 55.52 MiB
alloc_compute_meta: CPU compute buffer size = 13.50 MiB
alloc_compute_meta: graph splits = 55, nodes = 853
warmup: *****************************************************************
warmup: WARNING: flash attention not supported by CUDA0, memory usage will increase
warmup: op params:
warmup: dst: type = f32, ne = [72 16 1024 1], nb = [4 288 4608 4718592]
warmup: src0: type = f32, ne = [72 1024 16 1], nb = [4 4608 288 4718592]
warmup: src1: type = f16, ne = [72 1024 16 1], nb = [2 144 147456 2359296]
warmup: src2: type = f16, ne = [72 1024 16 1], nb = [2 144 147456 2359296]
warmup: please report this on github as an issue
warmup: *****************************************************************
alloc_compute_meta: warmup with image size = 512 x 512
alloc_compute_meta: CUDA0 compute buffer size = 112.02 MiB
alloc_compute_meta: CPU compute buffer size = 13.50 MiB
alloc_compute_meta: graph splits = 1, 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=CUDA0):
warmup: SOFT_MAX: type = f32, ne = [1024 1024 16 1]
warmup: CONT: type = f32, ne = [1024 72 16 1]
warmup: PERMUTE: type = f32, ne = [72 1024 16 1]
warmup: ROPE: type = f32, ne = [72 16 1024 1]
warmup: VIEW: type = f32, ne = [72 16 1024 1]
warmup: VIEW: type = f32, ne = [72 16 1024 1]
warmup: MUL_MAT: type = f32, ne = [3456 1024 1 1]
warmup: MUL: type = f32, ne = [1152 1024 1 1]
warmup: ADD: type = f32, ne = [1152 1024 1 1]
warmup: UNARY: type = f32, ne = [4608 256 1 1]
warmup: ADD: type = f32, ne = [4304 1024 1 1]
warmup: ADD: type = f32, ne = [1152 1024 1 1]
warmup: NORM: type = f32, ne = [1152 1024 1 1]
warmup: ADD: type = f32, ne = [1152 1024 1 1]
warmup: CONT: type = f32, ne = [1152 1024 1 1]
warmup: MUL_MAT: type = f32, ne = [72 1024 16 1]
warmup: MUL_MAT: type = f32, ne = [1024 1024 16 1]
warmup: NORM: type = f32, ne = [1152 1024 1 1]
warmup: PERMUTE: type = f32, ne = [72 1024 16 1]
warmup: ROPE: type = f32, ne = [72 16 1024 1]
warmup: VIEW: type = f32, ne = [72 16 1024 1]
warmup: ADD: type = f32, ne = [3456 1024 1 1]
warmup: ADD: type = f32, ne = [1152 1024 1 1]
warmup: NORM: type = f32, ne = [1152 1024 1 1]
warmup: ADD: type = f32, ne = [1152 1024 1 1]
warmup: VIEW: type = f32, ne = [72 16 1024 1]
warmup: MUL_MAT: type = f32, ne = [4608 256 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: *****************************************************************
0.08.779.898 I srv load_model: loaded multimodal model, '/home/dylan/gguf/mmproj-Qwen3-VL-32B-Instruct-bf16.gguf'
0.08.779.902 I srv init: initializing slots, n_slots = 4
0.08.779.905 I slot init: id 0 | task -1 | new slot, n_ctx = 8192
0.08.779.908 I slot init: id 1 | task -1 | new slot, n_ctx = 8192
0.08.779.910 I slot init: id 2 | task -1 | new slot, n_ctx = 8192
0.08.779.910 I slot init: id 3 | task -1 | new slot, n_ctx = 8192
0.08.779.931 W srv init: prompt cache is enabled, size limit: 51200 MiB
0.08.779.932 W srv init: use `--cache-ram 0` to disable the prompt cache
0.08.779.932 W srv init: for more info see https://github.com/ggml-org/llama.cpp/pull/16391
0.08.780.025 I srv init: thinking = 0
0.08.780.027 I main: model loaded
0.08.780.072 I 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
'
0.08.780.074 I main: server is listening on http://0.0.0.0:12800 - starting the main loop
0.08.780.075 I srv update_slots: all slots are idle