ggml_cuda_init: found 1 CUDA devices (Total VRAM: 124610 MiB):
Device 0: NVIDIA GB10, compute capability 12.1, VMM: yes, VRAM: 124610 MiB
system info: n_threads = 20, n_threads_batch = 20, total_threads = 20
system_info: n_threads = 20 (n_threads_batch = 20) / 20 | CUDA : ARCHS = 1210 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | FA_ALL_QUANTS = 1 | BLACKWELL_NATIVE_FP4 = 1 | CPU : NEON = 1 | ARM_FMA = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |
Running without SSL
init: using 19 threads for HTTP server
start: binding port with default address family
main: loading model
srv load_model: loading model '/home/coder/models/gemma-4-26B-A4B-it-f16.gguf'
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GB10) (000f:01:00.0) - 116297 MiB free
llama_model_loader: loaded meta data with 47 key-value pairs and 658 tensors from /home/coder/models/gemma-4-26B-A4B-it-f16.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 = gemma4
llama_model_loader: - kv 1: general.type str = model
llama_model_loader: - kv 2: general.sampling.top_k i32 = 64
llama_model_loader: - kv 3: general.sampling.top_p f32 = 0.950000
llama_model_loader: - kv 4: general.sampling.temp f32 = 1.000000
llama_model_loader: - kv 5: general.size_label str = 128x2.6B
llama_model_loader: - kv 6: general.license str = apache-2.0
llama_model_loader: - kv 7: general.license.link str = https://ai.google.dev/gemma/docs/gemm...
llama_model_loader: - kv 8: general.tags arr[str,1] = ["image-text-to-text"]
llama_model_loader: - kv 9: gemma4.block_count u32 = 30
llama_model_loader: - kv 10: gemma4.context_length u32 = 262144
llama_model_loader: - kv 11: gemma4.embedding_length u32 = 2816
llama_model_loader: - kv 12: gemma4.feed_forward_length u32 = 2112
llama_model_loader: - kv 13: gemma4.attention.head_count u32 = 16
llama_model_loader: - kv 14: gemma4.attention.head_count_kv arr[i32,30] = [8, 8, 8, 8, 8, 2, 8, 8, 8, 8, 8, 2, ...
llama_model_loader: - kv 15: gemma4.rope.freq_base f32 = 1000000.000000
llama_model_loader: - kv 16: gemma4.rope.freq_base_swa f32 = 10000.000000
llama_model_loader: - kv 17: gemma4.attention.layer_norm_rms_epsilon f32 = 0.000001
llama_model_loader: - kv 18: gemma4.expert_count u32 = 128
llama_model_loader: - kv 19: gemma4.expert_used_count u32 = 8
llama_model_loader: - kv 20: gemma4.attention.key_length u32 = 512
llama_model_loader: - kv 21: gemma4.attention.value_length u32 = 512
llama_model_loader: - kv 22: general.file_type u32 = 1
llama_model_loader: - kv 23: gemma4.final_logit_softcapping f32 = 30.000000
llama_model_loader: - kv 24: gemma4.attention.sliding_window u32 = 1024
llama_model_loader: - kv 25: gemma4.attention.shared_kv_layers u32 = 0
llama_model_loader: - kv 26: gemma4.embedding_length_per_layer_input u32 = 0
llama_model_loader: - kv 27: gemma4.attention.sliding_window_pattern arr[bool,30] = [true, true, true, true, true, false,...
llama_model_loader: - kv 28: gemma4.attention.key_length_swa u32 = 256
llama_model_loader: - kv 29: gemma4.attention.value_length_swa u32 = 256
llama_model_loader: - kv 30: gemma4.expert_feed_forward_length u32 = 704
llama_model_loader: - kv 31: gemma4.rope.dimension_count u32 = 512
llama_model_loader: - kv 32: gemma4.rope.dimension_count_swa u32 = 256
llama_model_loader: - kv 33: general.quantization_version u32 = 2
llama_model_loader: - kv 34: tokenizer.ggml.model str = gemma4
llama_model_loader: - kv 35: tokenizer.ggml.tokens arr[str,262144] = ["<pad>", "<eos>", "<bos>", "<unk>", ...
llama_model_loader: - kv 36: tokenizer.ggml.scores arr[f32,262144] = [-1000.000000, -1000.000000, -1000.00...
llama_model_loader: - kv 37: tokenizer.ggml.token_type arr[i32,262144] = [3, 3, 3, 3, 3, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv 38: tokenizer.ggml.merges arr[str,514906] = ["\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n \n", ...
llama_model_loader: - kv 39: tokenizer.ggml.bos_token_id u32 = 2
llama_model_loader: - kv 40: tokenizer.ggml.eos_token_id u32 = 1
llama_model_loader: - kv 41: tokenizer.ggml.unknown_token_id u32 = 3
llama_model_loader: - kv 42: tokenizer.ggml.padding_token_id u32 = 0
llama_model_loader: - kv 43: tokenizer.ggml.mask_token_id u32 = 4
llama_model_loader: - kv 44: tokenizer.chat_template str = {%- macro format_parameters(propertie...
llama_model_loader: - kv 45: tokenizer.ggml.add_space_prefix bool = false
llama_model_loader: - kv 46: tokenizer.ggml.add_bos_token bool = false
llama_model_loader: - type f32: 392 tensors
llama_model_loader: - type f16: 266 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type = F16
print_info: file size = 47.02 GiB (16.01 BPW)
load: 0 unused tokens
load: control-looking token: 212 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
load: printing all EOG tokens:
load: - 1 ('<eos>')
load: - 106 ('<turn|>')
load: - 212 ('</s>')
load: special tokens cache size = 25
load: token to piece cache size = 1.9445 MB
print_info: arch = gemma4
print_info: vocab_only = 0
print_info: no_alloc = 0
print_info: n_ctx_train = 262144
print_info: n_embd = 2816
print_info: n_embd_inp = 2816
print_info: n_layer = 30
print_info: n_head = 16
print_info: n_head_kv = [8, 8, 8, 8, 8, 2, 8, 8, 8, 8, 8, 2, 8, 8, 8, 8, 8, 2, 8, 8, 8, 8, 8, 2, 8, 8, 8, 8, 8, 2]
print_info: n_rot = 512
print_info: n_swa = 1024
print_info: is_swa_any = 1
print_info: n_embd_head_k = 512
print_info: n_embd_head_v = 512
print_info: n_gqa = [2, 2, 2, 2, 2, 8, 2, 2, 2, 2, 2, 8, 2, 2, 2, 2, 2, 8, 2, 2, 2, 2, 2, 8, 2, 2, 2, 2, 2, 8]
print_info: n_embd_k_gqa = [2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024]
print_info: n_embd_v_gqa = [2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 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 = 1.0e+00
print_info: n_ff = 2112
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 = -1
print_info: rope type = 2
print_info: rope scaling = linear
print_info: freq_base_train = 1000000.0
print_info: freq_scale_train = 1
print_info: freq_base_swa = 10000.0
print_info: freq_scale_swa = 1
print_info: n_embd_head_k_swa = 256
print_info: n_embd_head_v_swa = 256
print_info: n_rot_swa = 256
print_info: n_ctx_orig_yarn = 262144
print_info: rope_yarn_log_mul = 0.0000
print_info: rope_finetuned = unknown
print_info: model type = ?B
print_info: model params = 25.23 B
print_info: general.name = n/a
print_info: vocab type = SPM
print_info: n_vocab = 262144
print_info: n_merges = 0
print_info: BOS token = 2 '<bos>'
print_info: EOS token = 1 '<eos>'
print_info: UNK token = 3 '<unk>'
print_info: PAD token = 0 '<pad>'
print_info: MASK token = 4 '<mask>'
print_info: LF token = 248 '<0x0A>'
print_info: EOG token = 1 '<eos>'
print_info: EOG token = 106 '<turn|>'
print_info: EOG token = 212 '</s>'
print_info: max token length = 93
load_tensors: loading model tensors, this can take a while... (mmap = false, direct_io = false)
str: cannot properly format tensor name output with suffix=weight bid=-1 xid=-1
load_tensors: offloading output layer to GPU
load_tensors: offloading 29 repeating layers to GPU
load_tensors: offloaded 31/31 layers to GPU
load_tensors: CUDA0 model buffer size = 48150.36 MiB
load_tensors: CUDA_Host model buffer size = 1408.00 MiB
........................................................................
common_init_result: added <eos> logit bias = -inf
common_init_result: added <turn|> logit bias = -inf
common_init_result: added </s> logit bias = -inf
llama_context: constructing llama_context
llama_context: n_seq_max = 1
llama_context: n_ctx = 262144
llama_context: n_ctx_seq = 262144
llama_context: n_batch = 2048
llama_context: n_ubatch = 2048
llama_context: causal_attn = 1
llama_context: flash_attn = enabled
llama_context: kv_unified = false
llama_context: freq_base = 1000000.0
llama_context: freq_scale = 1
llama_context: CUDA_Host output buffer size = 1.00 MiB
llama_kv_cache_iswa: creating non-SWA KV cache, size = 262144 cells
llama_kv_cache: CUDA0 KV buffer size = 5120.00 MiB
llama_kv_cache: size = 5120.00 MiB (262144 cells, 5 layers, 1/1 seqs), K (f16): 2560.00 MiB, V (f16): 2560.00 MiB
llama_kv_cache: attn_rot_k = 0
llama_kv_cache: attn_rot_v = 0
llama_kv_cache_iswa: creating SWA KV cache, size = 3072 cells
llama_kv_cache: CUDA0 KV buffer size = 600.00 MiB
llama_kv_cache: size = 600.00 MiB ( 3072 cells, 25 layers, 1/1 seqs), K (f16): 300.00 MiB, V (f16): 300.00 MiB
llama_kv_cache: attn_rot_k = 0
llama_kv_cache: attn_rot_v = 0
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: CUDA0 compute buffer size = 3280.08 MiB
sched_reserve: CUDA_Host compute buffer size = 2116.09 MiB
sched_reserve: graph nodes = 2647
sched_reserve: graph splits = 2
sched_reserve: reserve took 857.64 ms, sched copies = 1
clip_model_loader: model name:
clip_model_loader: description:
clip_model_loader: GGUF version: 3
clip_model_loader: alignment: 32
clip_model_loader: n_tensors: 356
clip_model_loader: n_kv: 23
clip_model_loader: has vision encoder
clip_ctx: CLIP using CUDA0 backend
load_hparams: projector: gemma4v
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_quick
load_hparams: projection_dim: 2816
--- vision hparams ---
load_hparams: image_size: 224
load_hparams: patch_size: 16
load_hparams: has_llava_proj: 0
load_hparams: minicpmv_version: 0
load_hparams: n_merge: 3
load_hparams: n_wa_pattern: 0
load_hparams: image_min_pixels: 580608
load_hparams: image_max_pixels: 645120
load_hparams: model size: 1137.77 MiB
load_hparams: metadata size: 0.12 MiB
srv load_model: loaded multimodal model, '/home/coder/models/gemma-4-26B-A4B-it-mmproj-f16.gguf'
srv load_model: speculative decoding is not supported by multimodal, it will be disabled
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 = 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: '<bos><|turn>system
<|think|>You are a helpful assistant<turn|>
<|turn>user
Hello<turn|>
<|turn>model
Hi there<turn|>
<|turn>user
How are you?<turn|>
<|turn>model
'
srv init: init: chat template, thinking = 1
main: model loaded
main: server is listening on http://0.0.0.0:5808
main: starting the main loop...
srv update_slots: all slots are idle
srv log_server_r: done request: GET / 127.0.0.1 200
srv log_server_r: done request: GET /bundle.css 127.0.0.1 200
srv log_server_r: done request: GET /bundle.js 127.0.0.1 200
srv log_server_r: done request: HEAD /cors-proxy 127.0.0.1 404
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 = 262144, n_keep = 0, task.n_tokens = 20
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 = 16, batch.n_tokens = 16, progress = 0.800000
srv log_server_r: done request: POST /v1/chat/completions 127.0.0.1 200
slot update_slots: id 0 | task 0 | n_tokens = 16, memory_seq_rm [16, end)
slot init_sampler: id 0 | task 0 | init sampler, took 0.01 ms, tokens: text = 20, total = 20
slot update_slots: id 0 | task 0 | prompt processing done, n_tokens = 20, batch.n_tokens = 4
reasoning-budget: activated, budget=2147483647 tokens
reasoning-budget: deactivated (natural end)
srv stop: cancel task, id_task = 0
slot release: id 0 | task 0 | stop processing: n_tokens = 1299, truncated = 0
srv update_slots: all slots are idle
srv operator(): operator(): cleaning up before exit...
llama_memory_breakdown_print: | memory breakdown [MiB] | total free self model context compute unaccounted |
llama_memory_breakdown_print: | - CUDA0 (GB10) | 124610 = 56906 + (57150 = 48150 + 5720 + 3280) + 10554 |
llama_memory_breakdown_print: | - Host | 3524 = 1408 + 0 + 2116 |
Name and Version
version: 8638 (5803c8d)
built with GNU 13.3.0 for Linux aarch64
Operating systems
Linux
GGML backends
CUDA
Hardware
DGX Spark
Models
ggml-org/gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-f16.gguf
Problem description & steps to reproduce
I asked a general knowledge question ("What is the LHC?), and partway through the response, the model began spamming an unbounded stream of
<unused24>tokens.I stopped the generation manually at this point.
Perhaps we should just not sample unused tokens?
First Bad Commit
No response
Relevant log output
Logs