| 1 | ```Qwen3-4B-Thinking-2507-Q4_K_M.gguf (version GGUF V3 (latest))
|
| 2 | llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
|
| 3 | llama_model_loader: - kv 0: general.architecture str = qwen3
|
| 4 | llama_model_loader: - kv 1: general.type str = model
|
| 5 | llama_model_loader: - kv 2: general.name str = Qwen3 4B Thinking 2507
|
| 6 | llama_model_loader: - kv 3: general.version str = 2507
|
| 7 | llama_model_loader: - kv 4: general.finetune str = Thinking
|
| 8 | llama_model_loader: - kv 5: general.basename str = Qwen3
|
| 9 | llama_model_loader: - kv 6: general.size_label str = 4B
|
| 10 | llama_model_loader: - kv 7: general.license str = apache-2.0
|
| 11 | llama_model_loader: - kv 8: general.license.link str = https://huggingface.co/Qwen/Qwen3-4B-...
|
| 12 | llama_model_loader: - kv 9: general.tags arr[str,1] = ["text-generation"]
|
| 13 | llama_model_loader: - kv 10: qwen3.block_count u32 = 36
|
| 14 | llama_model_loader: - kv 11: qwen3.context_length u32 = 262144
|
| 15 | llama_model_loader: - kv 12: qwen3.embedding_length u32 = 2560
|
| 16 | llama_model_loader: - kv 13: qwen3.feed_forward_length u32 = 9728
|
| 17 | llama_model_loader: - kv 14: qwen3.attention.head_count u32 = 32
|
| 18 | llama_model_loader: - kv 15: qwen3.attention.head_count_kv u32 = 8
|
| 19 | llama_model_loader: - kv 16: qwen3.rope.freq_base f32 = 5000000.000000
|
| 20 | llama_model_loader: - kv 17: qwen3.attention.layer_norm_rms_epsilon f32 = 0.000001
|
| 21 | llama_model_loader: - kv 18: qwen3.attention.key_length u32 = 128
|
| 22 | llama_model_loader: - kv 19: qwen3.attention.value_length u32 = 128
|
| 23 | llama_model_loader: - kv 20: tokenizer.ggml.model str = gpt2
|
| 24 | llama_model_loader: - kv 21: tokenizer.ggml.pre str = qwen2
|
| 25 | llama_model_loader: - kv 22: tokenizer.ggml.tokens arr[str,151936] = ["!", "\"", "#", "$", "%", "&", "'", ...
|
| 26 | llama_model_loader: - kv 23: tokenizer.ggml.token_type arr[i32,151936] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
|
| 27 | llama_model_loader: - kv 24: tokenizer.ggml.merges arr[str,151387] = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
|
| 28 | llama_model_loader: - kv 25: tokenizer.ggml.eos_token_id u32 = 151645
|
| 29 | llama_model_loader: - kv 26: tokenizer.ggml.padding_token_id u32 = 151643
|
| 30 | llama_model_loader: - kv 27: tokenizer.ggml.bos_token_id u32 = 151643
|
| 31 | llama_model_loader: - kv 28: tokenizer.ggml.add_bos_token bool = false
|
| 32 | llama_model_loader: - kv 29: tokenizer.chat_template str = {%- if tools %}\n {{- '<|im_start|>...
|
| 33 | llama_model_loader: - kv 30: general.quantization_version u32 = 2
|
| 34 | llama_model_loader: - kv 31: general.file_type u32 = 15
|
| 35 | llama_model_loader: - type f32: 145 tensors
|
| 36 | llama_model_loader: - type q4_K: 216 tensors
|
| 37 | llama_model_loader: - type q6_K: 37 tensors
|
| 38 | print_info: file format = GGUF V3 (latest)
|
| 39 | print_info: file type = Q4_K - Medium
|
| 40 | print_info: file size = 2.32 GiB (4.95 BPW)
|
| 41 | load: printing all EOG tokens:
|
| 42 | load: - 151643 ('<|endoftext|>')
|
| 43 | load: - 151645 ('<|im_end|>')
|
| 44 | load: - 151662 ('<|fim_pad|>')
|
| 45 | load: - 151663 ('<|repo_name|>')
|
| 46 | load: - 151664 ('<|file_sep|>')
|
| 47 | load: special tokens cache size = 26
|
| 48 | load: token to piece cache size = 0.9311 MB
|
| 49 | print_info: arch = qwen3
|
| 50 | print_info: vocab_only = 0
|
| 51 | print_info: n_ctx_train = 262144
|
| 52 | print_info: n_embd = 2560
|
| 53 | print_info: n_embd_inp = 2560
|
| 54 | print_info: n_layer = 36
|
| 55 | print_info: n_head = 32
|
| 56 | print_info: n_head_kv = 8
|
| 57 | print_info: n_rot = 128
|
| 58 | print_info: n_swa = 0
|
| 59 | print_info: is_swa_any = 0
|
| 60 | print_info: n_embd_head_k = 128
|
| 61 | print_info: n_embd_head_v = 128
|
| 62 | print_info: n_gqa = 4
|
| 63 | print_info: n_embd_k_gqa = 1024
|
| 64 | print_info: n_embd_v_gqa = 1024
|
| 65 | print_info: f_norm_eps = 0.0e+00
|
| 66 | print_info: f_norm_rms_eps = 1.0e-06
|
| 67 | print_info: f_clamp_kqv = 0.0e+00
|
| 68 | print_info: f_max_alibi_bias = 0.0e+00
|
| 69 | print_info: f_logit_scale = 0.0e+00
|
| 70 | print_info: f_attn_scale = 0.0e+00
|
| 71 | print_info: n_ff = 9728
|
| 72 | print_info: n_expert = 0
|
| 73 | print_info: n_expert_used = 0
|
| 74 | print_info: n_expert_groups = 0
|
| 75 | print_info: n_group_used = 0
|
| 76 | print_info: causal attn = 1
|
| 77 | print_info: pooling type = -1
|
| 78 | print_info: rope type = 2
|
| 79 | print_info: rope scaling = linear
|
| 80 | print_info: freq_base_train = 5000000.0
|
| 81 | print_info: freq_scale_train = 1
|
| 82 | print_info: n_ctx_orig_yarn = 262144
|
| 83 | print_info: rope_finetuned = unknown
|
| 84 | print_info: model type = 4B
|
| 85 | print_info: model params = 4.02 B
|
| 86 | print_info: general.name = Qwen3 4B Thinking 2507
|
| 87 | print_info: vocab type = BPE
|
| 88 | print_info: n_vocab = 151936
|
| 89 | print_info: n_merges = 151387
|
| 90 | print_info: BOS token = 151643 '<|endoftext|>'
|
| 91 | print_info: EOS token = 151645 '<|im_end|>'
|
| 92 | print_info: EOT token = 151645 '<|im_end|>'
|
| 93 | print_info: PAD token = 151643 '<|endoftext|>'
|
| 94 | print_info: LF token = 198 'Ċ'
|
| 95 | print_info: FIM PRE token = 151659 '<|fim_prefix|>'
|
| 96 | print_info: FIM SUF token = 151661 '<|fim_suffix|>'
|
| 97 | print_info: FIM MID token = 151660 '<|fim_middle|>'
|
| 98 | print_info: FIM PAD token = 151662 '<|fim_pad|>'
|
| 99 | print_info: FIM REP token = 151663 '<|repo_name|>'
|
| 100 | print_info: FIM SEP token = 151664 '<|file_sep|>'
|
| 101 | print_info: EOG token = 151643 '<|endoftext|>'
|
| 102 | print_info: EOG token = 151645 '<|im_end|>'
|
| 103 | print_info: EOG token = 151662 '<|fim_pad|>'
|
| 104 | print_info: EOG token = 151663 '<|repo_name|>'
|
| 105 | print_info: EOG token = 151664 '<|file_sep|>'
|
| 106 | print_info: max token length = 256
|
| 107 | load_tensors: loading model tensors, this can take a while... (mmap = true)
|
| 108 | load_tensors: offloading 36 repeating layers to GPU
|
| 109 | load_tensors: offloading output layer to GPU
|
| 110 | load_tensors: offloaded 37/37 layers to GPU
|
| 111 | load_tensors: CPU_Mapped model buffer size = 304.28 MiB
|
| 112 | load_tensors: Metal_Mapped model buffer size = 2375.91 MiB
|
| 113 | ................................................................................
|
| 114 | llama_context: constructing llama_context
|
| 115 | llama_context: n_seq_max = 4
|
| 116 | llama_context: n_ctx = 10240
|
| 117 | llama_context: n_ctx_seq = 10240
|
| 118 | llama_context: n_batch = 2048
|
| 119 | llama_context: n_ubatch = 512
|
| 120 | llama_context: causal_attn = 1
|
| 121 | llama_context: flash_attn = auto
|
| 122 | llama_context: kv_unified = true
|
| 123 | llama_context: freq_base = 5000000.0
|
| 124 | llama_context: freq_scale = 1
|
| 125 | llama_context: n_ctx_seq (10240) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
|
| 126 | ggml_metal_init: allocating
|
| 127 | ggml_metal_init: found device: Apple M3
|
| 128 | ggml_metal_init: picking default device: Apple M3
|
| 129 | ggml_metal_init: use fusion = true
|
| 130 | ggml_metal_init: use concurrency = true
|
| 131 | ggml_metal_init: use graph optimize = true
|
| 132 | llama_context: CPU output buffer size = 2.32 MiB
|
| 133 | llama_kv_cache: Metal KV buffer size = 1440.00 MiB
|
| 134 | llama_kv_cache: size = 1440.00 MiB ( 10240 cells, 36 layers, 4/1 seqs), K (f16): 720.00 MiB, V (f16): 720.00 MiB
|
| 135 | llama_context: Flash Attention was auto, set to enabled
|
| 136 | llama_context: Metal compute buffer size = 301.75 MiB
|
| 137 | llama_context: CPU compute buffer size = 25.01 MiB```
|
