Eval bug: Response always starts with </think> tag when running Qwen3.5 9B #20516
Description
Name and Version
llama-cli.exe --version
load_backend: loaded RPC backend from P:\Llama\llama.cpp_vulkan\ggml-rpc.dll
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon(TM) Graphics (AMD proprietary driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 32 | shared memory: 32768 | int dot: 1 | matrix cores: none
load_backend: loaded Vulkan backend from P:\Llama\llama.cpp_vulkan\ggml-vulkan.dll
load_backend: loaded CPU backend from P:\Llama\llama.cpp_vulkan\ggml-cpu-haswell.dll
version: 8322 (557fe2d91)
built with Clang 19.1.5 for Windows x86_64
Operating systems
Windows
GGML backends
Vulkan
Hardware
AMD 6800U (680M)
Models
Qwen3.5-9B-UD-Q4_K_XL__unsloth_6f5d30.gguf
Problem description & steps to reproduce
When I run Qwen3.5 9B with the following parameters, the response always starts with a </think> tag.
.\llama.cpp_vulkan\llama-server.exe --webui-config-file config_webui.json --host 0.0.0.0 --port 8090 --props --slots --metrics --gpu-layers all --cache-type-k q8_0 --cache-type-v q8_0 --cache-reuse 256 --model .\_model\Qwen3.5-9B\Qwen3.5-9B-UD-Q4_K_XL__unsloth_6f5d30.gguf --mmproj .\_model\Qwen3.5-9B\Qwen3.5-9B-mmproj-F16__unsloth_f70dc3.gguf --temp 0.7 --top-k 20 --top-p 0.8 --min-p 0.0 --repeat-penalty 1.0 --presence-penalty 1.5 --reasoning-budget 0 --chat-template-kwargs "{\"enable_thinking\":false}" --alias Qwen3.5-9B-Instruct --parallel -1 --no-mmap
First Bad Commit
No response
Relevant log output
Logs
load_backend: loaded RPC backend from P:\Llama\llama.cpp_vulkan\ggml-rpc.dll
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon(TM) Graphics (AMD proprietary driver) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 32 | shared memory: 32768 | int dot: 1 | matrix cores: none
load_backend: loaded Vulkan backend from P:\Llama\llama.cpp_vulkan\ggml-vulkan.dll
load_backend: loaded CPU backend from P:\Llama\llama.cpp_vulkan\ggml-cpu-haswell.dll
Setting 'enable_thinking' via --chat-template-kwargs is deprecated. Use --reasoning on / --reasoning off instead.
main: n_parallel is set to auto, using n_parallel = 4 and kv_unified = true
build: 8322 (557fe2d91) with Clang 19.1.5 for Windows x86_64
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 |
init: using 15 threads for HTTP server
start: binding port with default address family
main: loading model
srv load_model: loading model '.\_model\Qwen3.5-9B\Qwen3.5-9B-UD-Q4_K_XL__unsloth_6f5d30.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
llama_params_fit_impl: projected to use 10486 MiB of device memory vs. 17182 MiB of free device memory
llama_params_fit_impl: will leave 6696 >= 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 0.56 seconds
llama_model_load_from_file_impl: using device Vulkan0 (AMD Radeon(TM) Graphics) (unknown id) - 17387 MiB free
llama_model_loader: loaded meta data with 46 key-value pairs and 427 tensors from .\_model\Qwen3.5-9B\Qwen3.5-9B-UD-Q4_K_XL__unsloth_6f5d30.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 = qwen35
llama_model_loader: - kv 1: general.type str = model
llama_model_loader: - kv 2: general.name str = Qwen3.5-9B
llama_model_loader: - kv 3: general.basename str = Qwen3.5-9B
llama_model_loader: - kv 4: general.quantized_by str = Unsloth
llama_model_loader: - kv 5: general.size_label str = 9B
llama_model_loader: - kv 6: general.license str = apache-2.0
llama_model_loader: - kv 7: general.license.link str = https://huggingface.co/Qwen/Qwen3.5-9...
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 = Qwen3.5 9B
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/Qwen3.5-9B
llama_model_loader: - kv 13: general.tags arr[str,2] = ["unsloth", "image-text-to-text"]
llama_model_loader: - kv 14: qwen35.block_count u32 = 32
llama_model_loader: - kv 15: qwen35.context_length u32 = 262144
llama_model_loader: - kv 16: qwen35.embedding_length u32 = 4096
llama_model_loader: - kv 17: qwen35.feed_forward_length u32 = 12288
llama_model_loader: - kv 18: qwen35.attention.head_count u32 = 16
llama_model_loader: - kv 19: qwen35.attention.head_count_kv u32 = 4
llama_model_loader: - kv 20: qwen35.rope.dimension_sections arr[i32,4] = [11, 11, 10, 0]
llama_model_loader: - kv 21: qwen35.rope.freq_base f32 = 10000000.000000
llama_model_loader: - kv 22: qwen35.attention.layer_norm_rms_epsilon f32 = 0.000001
llama_model_loader: - kv 23: qwen35.attention.key_length u32 = 256
llama_model_loader: - kv 24: qwen35.attention.value_length u32 = 256
llama_model_loader: - kv 25: qwen35.ssm.conv_kernel u32 = 4
llama_model_loader: - kv 26: qwen35.ssm.state_size u32 = 128
llama_model_loader: - kv 27: qwen35.ssm.group_count u32 = 16
llama_model_loader: - kv 28: qwen35.ssm.time_step_rank u32 = 32
llama_model_loader: - kv 29: qwen35.ssm.inner_size u32 = 4096
llama_model_loader: - kv 30: qwen35.full_attention_interval u32 = 4
llama_model_loader: - kv 31: qwen35.rope.dimension_count u32 = 64
llama_model_loader: - kv 32: tokenizer.ggml.model str = gpt2
llama_model_loader: - kv 33: tokenizer.ggml.pre str = qwen35
llama_model_loader: - kv 34: tokenizer.ggml.tokens arr[str,248320] = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv 35: tokenizer.ggml.token_type arr[i32,248320] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv 36: tokenizer.ggml.merges arr[str,247587] = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv 37: tokenizer.ggml.eos_token_id u32 = 248046
llama_model_loader: - kv 38: tokenizer.ggml.padding_token_id u32 = 248055
llama_model_loader: - kv 39: tokenizer.chat_template str = {%- set image_count = namespace(value...
llama_model_loader: - kv 40: general.quantization_version u32 = 2
llama_model_loader: - kv 41: general.file_type u32 = 15
llama_model_loader: - kv 42: quantize.imatrix.file str = Qwen3.5-9B-GGUF/imatrix_unsloth.gguf
llama_model_loader: - kv 43: quantize.imatrix.dataset str = unsloth_calibration_Qwen3.5-9B.txt
llama_model_loader: - kv 44: quantize.imatrix.entries_count u32 = 248
llama_model_loader: - kv 45: quantize.imatrix.chunks_count u32 = 80
llama_model_loader: - type f32: 177 tensors
llama_model_loader: - type f16: 48 tensors
llama_model_loader: - type q8_0: 24 tensors
llama_model_loader: - type q4_K: 77 tensors
llama_model_loader: - type q5_K: 67 tensors
llama_model_loader: - type q6_K: 24 tensors
llama_model_loader: - type iq4_xs: 10 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type = Q4_K - Medium
print_info: file size = 5.55 GiB (5.32 BPW)
load: 0 unused tokens
load: printing all EOG tokens:
load: - 248044 ('<|endoftext|>')
load: - 248046 ('<|im_end|>')
load: - 248063 ('<|fim_pad|>')
load: - 248064 ('<|repo_name|>')
load: - 248065 ('<|file_sep|>')
load: special tokens cache size = 33
load: token to piece cache size = 1.7581 MB
print_info: arch = qwen35
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 = 4096
print_info: n_layer = 32
print_info: n_head = 16
print_info: n_head_kv = 4
print_info: n_rot = 64
print_info: n_swa = 0
print_info: is_swa_any = 0
print_info: n_embd_head_k = 256
print_info: n_embd_head_v = 256
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 = 10000000.0
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 = [11, 11, 10, 0]
print_info: ssm_d_conv = 4
print_info: ssm_d_inner = 4096
print_info: ssm_d_state = 128
print_info: ssm_dt_rank = 32
print_info: ssm_n_group = 16
print_info: ssm_dt_b_c_rms = 0
print_info: model type = 9B
print_info: model params = 8.95 B
print_info: general.name = Qwen3.5-9B
print_info: vocab type = BPE
print_info: n_vocab = 248320
print_info: n_merges = 247587
print_info: BOS token = 11 ','
print_info: EOS token = 248046 '<|im_end|>'
print_info: EOT token = 248046 '<|im_end|>'
print_info: PAD token = 248055 '<|vision_pad|>'
print_info: LF token = 198 'Ċ'
print_info: FIM PRE token = 248060 '<|fim_prefix|>'
print_info: FIM SUF token = 248062 '<|fim_suffix|>'
print_info: FIM MID token = 248061 '<|fim_middle|>'
print_info: FIM PAD token = 248063 '<|fim_pad|>'
print_info: FIM REP token = 248064 '<|repo_name|>'
print_info: FIM SEP token = 248065 '<|file_sep|>'
print_info: EOG token = 248044 '<|endoftext|>'
print_info: EOG token = 248046 '<|im_end|>'
print_info: EOG token = 248063 '<|fim_pad|>'
print_info: EOG token = 248064 '<|repo_name|>'
print_info: EOG token = 248065 '<|file_sep|>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = false, direct_io = false)
load_tensors: offloading output layer to GPU
load_tensors: offloading 31 repeating layers to GPU
load_tensors: offloaded 33/33 layers to GPU
load_tensors: Vulkan0 model buffer size = 5133.63 MiB
load_tensors: Vulkan_Host model buffer size = 545.63 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 = 4
llama_context: n_ctx = 262144
llama_context: n_ctx_seq = 262144
llama_context: n_batch = 2048
llama_context: n_ubatch = 512
llama_context: causal_attn = 1
llama_context: flash_attn = auto
llama_context: kv_unified = true
llama_context: freq_base = 10000000.0
llama_context: freq_scale = 1
llama_context: Vulkan_Host output buffer size = 3.79 MiB
llama_kv_cache: Vulkan0 KV buffer size = 4352.00 MiB
llama_kv_cache: size = 4352.00 MiB (262144 cells, 8 layers, 4/1 seqs), K (q8_0): 2176.00 MiB, V (q8_0): 2176.00 MiB
llama_memory_recurrent: Vulkan0 RS buffer size = 201.00 MiB
llama_memory_recurrent: size = 201.00 MiB ( 4 cells, 32 layers, 4 seqs), R (f32): 9.00 MiB, S (f32): 192.00 MiB
sched_reserve: reserving ...
sched_reserve: Flash Attention was auto, set to enabled
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: Vulkan0 compute buffer size = 800.02 MiB
sched_reserve: Vulkan_Host compute buffer size = 528.02 MiB
sched_reserve: graph nodes = 1833
sched_reserve: graph splits = 2
sched_reserve: reserve took 138.98 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: Qwen3.5-9B
clip_model_loader: description:
clip_model_loader: GGUF version: 3
clip_model_loader: alignment: 32
clip_model_loader: n_tensors: 334
clip_model_loader: n_kv: 32
clip_model_loader: has vision encoder
clip_ctx: CLIP using Vulkan0 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: 4194304
load_hparams: model size: 875.61 MiB
load_hparams: metadata size: 0.12 MiB
warmup: warmup with image size = 1472 x 1472
alloc_compute_meta: Vulkan0 compute buffer size = 248.10 MiB
alloc_compute_meta: CPU compute buffer size = 24.93 MiB
alloc_compute_meta: graph splits = 1, nodes = 823
warmup: flash attention is enabled
srv load_model: loaded multimodal model, '.\_model\Qwen3.5-9B\Qwen3.5-9B-mmproj-F16__unsloth_f70dc3.gguf'
srv load_model: cache_reuse is not supported by multimodal, it will be disabled
srv load_model: initializing slots, n_slots = 4
common_speculative_is_compat: the target context does not support partial sequence removal
srv load_model: speculative decoding not supported by this context
slot load_model: id 0 | task -1 | new slot, n_ctx = 262144
slot load_model: id 1 | task -1 | new slot, n_ctx = 262144
slot load_model: id 2 | task -1 | new slot, n_ctx = 262144
slot load_model: id 3 | task -1 | new slot, n_ctx = 262144
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
<think>
</think>
'
srv init: init: chat template, thinking = 1
main: model loaded
main: server is listening on http://0.0.0.0:8090
main: starting the main loop...
srv update_slots: all slots are idle