### UD-Q8_K_XL_APPROX (Q8_0) ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices ggml_metal_library_init: using embedded metal library ggml_metal_library_init: loaded in 0.008 sec ggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s) ggml_metal_device_init: GPU name: MTL0 (Apple M4 Max) ggml_metal_device_init: GPU family: MTLGPUFamilyApple9 (1009) ggml_metal_device_init: GPU family: MTLGPUFamilyCommon3 (3003) ggml_metal_device_init: GPU family: MTLGPUFamilyMetal4 (5002) ggml_metal_device_init: simdgroup reduction = true ggml_metal_device_init: simdgroup matrix mul. = true ggml_metal_device_init: has unified memory = true ggml_metal_device_init: has bfloat = true ggml_metal_device_init: has tensor = false ggml_metal_device_init: use residency sets = true ggml_metal_device_init: use shared buffers = true ggml_metal_device_init: recommendedMaxWorkingSetSize = 55662.79 MB llama_print_build_info: build = 1 (399739d) llama_print_build_info: built with AppleClang 17.0.0.17000013 for Darwin arm64 llama_quantize: calculating quantization size for 'release_work/lfm25-moe8-mac-trainer/artifacts/gguf/bf16/LFM-2.5-8B-1B-hermes-ft-BF16.gguf' as Q8_0 llama_model_loader: loaded meta data with 32 key-value pairs and 256 tensors from release_work/lfm25-moe8-mac-trainer/artifacts/gguf/bf16/LFM-2.5-8B-1B-hermes-ft-BF16.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 = lfm2moe llama_model_loader: - kv 1: general.type str = model llama_model_loader: - kv 2: general.name str = Model_Upload_Iter10_Fused_Dequantized llama_model_loader: - kv 3: general.size_label str = 32x959M llama_model_loader: - kv 4: general.tags arr[str,2] = ["mlx", "text-generation"] llama_model_loader: - kv 5: general.languages arr[str,1] = ["en"] llama_model_loader: - kv 6: lfm2moe.block_count u32 = 24 llama_model_loader: - kv 7: lfm2moe.context_length u32 = 128000 llama_model_loader: - kv 8: lfm2moe.embedding_length u32 = 2048 llama_model_loader: - kv 9: lfm2moe.feed_forward_length u32 = 7168 llama_model_loader: - kv 10: lfm2moe.attention.head_count u32 = 32 llama_model_loader: - kv 11: lfm2moe.attention.head_count_kv arr[i32,24] = [0, 0, 8, 0, 0, 0, 8, 0, 0, 0, 8, 0, ... llama_model_loader: - kv 12: lfm2moe.rope.freq_base f32 = 5000000.000000 llama_model_loader: - kv 13: lfm2moe.attention.layer_norm_rms_epsilon f32 = 0.000010 llama_model_loader: - kv 14: lfm2moe.expert_count u32 = 32 llama_model_loader: - kv 15: lfm2moe.expert_used_count u32 = 4 llama_model_loader: - kv 16: general.file_type u32 = 32 llama_model_loader: - kv 17: lfm2moe.expert_feed_forward_length u32 = 1792 llama_model_loader: - kv 18: lfm2moe.leading_dense_block_count u32 = 2 llama_model_loader: - kv 19: lfm2moe.expert_gating_func u32 = 2 llama_model_loader: - kv 20: lfm2moe.vocab_size u32 = 128000 llama_model_loader: - kv 21: lfm2moe.shortconv.l_cache u32 = 3 llama_model_loader: - kv 22: general.quantization_version u32 = 2 llama_model_loader: - kv 23: tokenizer.ggml.model str = gpt2 llama_model_loader: - kv 24: tokenizer.ggml.pre str = lfm2 llama_model_loader: - kv 25: tokenizer.ggml.tokens arr[str,128000] = ["ÄŠÄŠÄŠÄŠÄŠÄŠÄŠÄŠÄŠÄŠ", "ÄŠÄŠÄŠÄŠÄŠÄ... llama_model_loader: - kv 26: tokenizer.ggml.token_type arr[i32,128000] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ... llama_model_loader: - kv 27: tokenizer.ggml.merges arr[str,293320] = ["ÄŠÄŠÄŠÄŠÄŠÄŠÄŠÄŠÄŠ ÄŠ", "ÄŠÄŠÄŠÄŠÄŠ... llama_model_loader: - kv 28: tokenizer.ggml.bos_token_id u32 = 124894 llama_model_loader: - kv 29: tokenizer.ggml.eos_token_id u32 = 124900 llama_model_loader: - kv 30: tokenizer.ggml.padding_token_id u32 = 124893 llama_model_loader: - kv 31: tokenizer.chat_template str = {{- bos_token -}}\n{%- set preserve_th... llama_model_loader: - type f32: 123 tensors llama_model_loader: - type bf16: 133 tensors llama_model_quantize_impl: have importance matrix data with 154 entries llama_tensor_get_type: token_embd.weight - applying manual override: q8_0 -> f16 llama_tensor_get_type: blk.2.attn_output.weight - applying manual override: q8_0 -> f16 llama_tensor_get_type: blk.6.attn_output.weight - applying manual override: q8_0 -> f16 llama_tensor_get_type: blk.10.attn_output.weight - applying manual override: q8_0 -> f16 llama_tensor_get_type: blk.14.attn_output.weight - applying manual override: q8_0 -> f16 llama_tensor_get_type: blk.18.attn_output.weight - applying manual override: q8_0 -> f16 llama_tensor_get_type: blk.21.attn_output.weight - applying manual override: q8_0 -> f16 [ 1/ 256] token_embd.weight - [ 2048, 128000, 1, 1], type = bf16, size = 500.00 MiB -> 500.00 MiB (f16) [ 2/ 256] token_embd_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 3/ 256] blk.0.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 4/ 256] blk.0.ffn_down.weight - [ 7168, 2048, 1, 1], type = bf16, size = 28.00 MiB -> 14.88 MiB (q8_0) [ 5/ 256] blk.0.ffn_gate.weight - [ 2048, 7168, 1, 1], type = bf16, size = 28.00 MiB -> 14.88 MiB (q8_0) [ 6/ 256] blk.0.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 7/ 256] blk.0.ffn_up.weight - [ 2048, 7168, 1, 1], type = bf16, size = 28.00 MiB -> 14.88 MiB (q8_0) [ 8/ 256] blk.0.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 9/ 256] blk.0.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 12.75 MiB (q8_0) [ 10/ 256] blk.0.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 11/ 256] blk.1.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 12/ 256] blk.1.ffn_down.weight - [ 7168, 2048, 1, 1], type = bf16, size = 28.00 MiB -> 14.88 MiB (q8_0) [ 13/ 256] blk.1.ffn_gate.weight - [ 2048, 7168, 1, 1], type = bf16, size = 28.00 MiB -> 14.88 MiB (q8_0) [ 14/ 256] blk.1.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 15/ 256] blk.1.ffn_up.weight - [ 2048, 7168, 1, 1], type = bf16, size = 28.00 MiB -> 14.88 MiB (q8_0) [ 16/ 256] blk.1.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 17/ 256] blk.1.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 12.75 MiB (q8_0) [ 18/ 256] blk.1.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 19/ 256] blk.2.attn_k.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 1.06 MiB (q8_0) [ 20/ 256] blk.2.attn_k_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 21/ 256] blk.2.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 22/ 256] blk.2.attn_output.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 8.00 MiB (f16) [ 23/ 256] blk.2.attn_q.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 24/ 256] blk.2.attn_q_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 25/ 256] blk.2.attn_v.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 1.06 MiB (q8_0) [ 26/ 256] blk.2.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 27/ 256] blk.2.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 28/ 256] blk.2.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 29/ 256] blk.2.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 30/ 256] blk.2.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 31/ 256] blk.2.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 32/ 256] blk.3.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 33/ 256] blk.3.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 34/ 256] blk.3.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 35/ 256] blk.3.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 36/ 256] blk.3.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 37/ 256] blk.3.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 38/ 256] blk.3.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 39/ 256] blk.3.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 40/ 256] blk.3.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 12.75 MiB (q8_0) [ 41/ 256] blk.3.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 42/ 256] blk.4.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 43/ 256] blk.4.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 44/ 256] blk.4.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 45/ 256] blk.4.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 46/ 256] blk.4.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 47/ 256] blk.4.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 48/ 256] blk.4.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 49/ 256] blk.4.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 50/ 256] blk.4.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 12.75 MiB (q8_0) [ 51/ 256] blk.4.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 52/ 256] blk.5.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 53/ 256] blk.5.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 54/ 256] blk.5.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 55/ 256] blk.5.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 56/ 256] blk.5.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 57/ 256] blk.5.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 58/ 256] blk.5.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 59/ 256] blk.5.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 60/ 256] blk.5.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 12.75 MiB (q8_0) [ 61/ 256] blk.5.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 62/ 256] blk.6.attn_k.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 1.06 MiB (q8_0) [ 63/ 256] blk.6.attn_k_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 64/ 256] blk.6.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 65/ 256] blk.6.attn_output.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 8.00 MiB (f16) [ 66/ 256] blk.6.attn_q.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 67/ 256] blk.6.attn_q_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 68/ 256] blk.6.attn_v.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 1.06 MiB (q8_0) [ 69/ 256] blk.6.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 70/ 256] blk.6.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 71/ 256] blk.6.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 72/ 256] blk.6.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 73/ 256] blk.6.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 74/ 256] blk.6.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 75/ 256] blk.7.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 76/ 256] blk.7.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 77/ 256] blk.7.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 78/ 256] blk.7.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 79/ 256] blk.7.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 80/ 256] blk.7.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 81/ 256] blk.7.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 82/ 256] blk.7.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 83/ 256] blk.7.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 12.75 MiB (q8_0) [ 84/ 256] blk.7.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 85/ 256] blk.8.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 86/ 256] blk.8.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 87/ 256] blk.8.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 88/ 256] blk.8.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 89/ 256] blk.8.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 90/ 256] blk.8.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 91/ 256] blk.8.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 92/ 256] blk.8.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 93/ 256] blk.8.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 12.75 MiB (q8_0) [ 94/ 256] blk.8.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 95/ 256] blk.9.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 96/ 256] blk.9.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 97/ 256] blk.9.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 98/ 256] blk.9.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 99/ 256] blk.9.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 100/ 256] blk.9.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 101/ 256] blk.9.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 102/ 256] blk.9.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 103/ 256] blk.9.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 12.75 MiB (q8_0) [ 104/ 256] blk.9.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 105/ 256] blk.10.attn_k.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 1.06 MiB (q8_0) [ 106/ 256] blk.10.attn_k_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 107/ 256] blk.10.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 108/ 256] blk.10.attn_output.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 8.00 MiB (f16) [ 109/ 256] blk.10.attn_q.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 110/ 256] blk.10.attn_q_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 111/ 256] blk.10.attn_v.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 1.06 MiB (q8_0) [ 112/ 256] blk.10.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 113/ 256] blk.10.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 114/ 256] blk.10.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 115/ 256] blk.10.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 116/ 256] blk.10.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 117/ 256] blk.10.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 118/ 256] blk.11.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 119/ 256] blk.11.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 120/ 256] blk.11.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 121/ 256] blk.11.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 122/ 256] blk.11.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 123/ 256] blk.11.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 124/ 256] blk.11.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 125/ 256] blk.11.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 126/ 256] blk.11.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 12.75 MiB (q8_0) [ 127/ 256] blk.11.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 128/ 256] blk.12.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 129/ 256] blk.12.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 130/ 256] blk.12.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 131/ 256] blk.12.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 132/ 256] blk.12.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 133/ 256] blk.12.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 134/ 256] blk.12.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 135/ 256] blk.12.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 136/ 256] blk.12.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 12.75 MiB (q8_0) [ 137/ 256] blk.12.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 138/ 256] blk.13.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 139/ 256] blk.13.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 140/ 256] blk.13.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 141/ 256] blk.13.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 142/ 256] blk.13.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 143/ 256] blk.13.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 144/ 256] blk.13.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 145/ 256] blk.13.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 146/ 256] blk.13.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 12.75 MiB (q8_0) [ 147/ 256] blk.13.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 148/ 256] blk.14.attn_k.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 1.06 MiB (q8_0) [ 149/ 256] blk.14.attn_k_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 150/ 256] blk.14.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 151/ 256] blk.14.attn_output.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 8.00 MiB (f16) [ 152/ 256] blk.14.attn_q.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 153/ 256] blk.14.attn_q_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 154/ 256] blk.14.attn_v.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 1.06 MiB (q8_0) [ 155/ 256] blk.14.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 156/ 256] blk.14.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 157/ 256] blk.14.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 158/ 256] blk.14.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 159/ 256] blk.14.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 160/ 256] blk.14.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 161/ 256] blk.15.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 162/ 256] blk.15.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 163/ 256] blk.15.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 164/ 256] blk.15.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 165/ 256] blk.15.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 166/ 256] blk.15.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 167/ 256] blk.15.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 168/ 256] blk.15.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 169/ 256] blk.15.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 12.75 MiB (q8_0) [ 170/ 256] blk.15.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 171/ 256] blk.16.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 172/ 256] blk.16.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 173/ 256] blk.16.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 174/ 256] blk.16.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 175/ 256] blk.16.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 176/ 256] blk.16.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 177/ 256] blk.16.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 178/ 256] blk.16.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 179/ 256] blk.16.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 12.75 MiB (q8_0) [ 180/ 256] blk.16.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 181/ 256] blk.17.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 182/ 256] blk.17.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 183/ 256] blk.17.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 184/ 256] blk.17.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 185/ 256] blk.17.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 186/ 256] blk.17.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 187/ 256] blk.17.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 188/ 256] blk.17.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 189/ 256] blk.17.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 12.75 MiB (q8_0) [ 190/ 256] blk.17.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 191/ 256] blk.18.attn_k.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 1.06 MiB (q8_0) [ 192/ 256] blk.18.attn_k_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 193/ 256] blk.18.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 194/ 256] blk.18.attn_output.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 8.00 MiB (f16) [ 195/ 256] blk.18.attn_q.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 196/ 256] blk.18.attn_q_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 197/ 256] blk.18.attn_v.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 1.06 MiB (q8_0) [ 198/ 256] blk.18.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 199/ 256] blk.18.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 200/ 256] blk.18.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 201/ 256] blk.18.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 202/ 256] blk.18.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 203/ 256] blk.18.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 204/ 256] blk.19.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 205/ 256] blk.19.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 206/ 256] blk.19.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 207/ 256] blk.19.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 208/ 256] blk.19.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 209/ 256] blk.19.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 210/ 256] blk.19.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 211/ 256] blk.19.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 212/ 256] blk.19.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 12.75 MiB (q8_0) [ 213/ 256] blk.19.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 214/ 256] blk.20.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 215/ 256] blk.20.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 216/ 256] blk.20.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 217/ 256] blk.20.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 218/ 256] blk.20.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 219/ 256] blk.20.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 220/ 256] blk.20.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 221/ 256] blk.20.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 222/ 256] blk.20.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 12.75 MiB (q8_0) [ 223/ 256] blk.20.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 224/ 256] blk.21.attn_k.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 1.06 MiB (q8_0) [ 225/ 256] blk.21.attn_k_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 226/ 256] blk.21.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 227/ 256] blk.21.attn_output.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 8.00 MiB (f16) [ 228/ 256] blk.21.attn_q.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 229/ 256] blk.21.attn_q_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 230/ 256] blk.21.attn_v.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 1.06 MiB (q8_0) [ 231/ 256] blk.21.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 232/ 256] blk.21.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 233/ 256] blk.21.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 234/ 256] blk.21.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 235/ 256] blk.21.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 236/ 256] blk.21.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 237/ 256] blk.22.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 238/ 256] blk.22.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 239/ 256] blk.22.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 240/ 256] blk.22.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 241/ 256] blk.22.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 242/ 256] blk.22.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 243/ 256] blk.22.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 244/ 256] blk.22.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 245/ 256] blk.22.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 12.75 MiB (q8_0) [ 246/ 256] blk.22.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 247/ 256] blk.23.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 248/ 256] blk.23.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 249/ 256] blk.23.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 250/ 256] blk.23.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 251/ 256] blk.23.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 252/ 256] blk.23.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 253/ 256] blk.23.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 254/ 256] blk.23.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 255/ 256] blk.23.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 12.75 MiB (q8_0) [ 256/ 256] blk.23.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) llama_model_quantize_impl: model size = 16154.31 MiB (16.00 BPW) llama_model_quantize_impl: quant size = 8841.81 MiB (8.76 BPW) load_imatrix: imatrix datasets=['release_work/lfm25-moe8-mac-trainer/artifacts/gguf/calibration/hermes_tool_router_calibration.txt'] load_imatrix: loaded 154 importance matrix entries from release_work/lfm25-moe8-mac-trainer/artifacts/gguf/calibration/hermes_tool_router_imatrix.gguf computed on 32 chunks prepare_imatrix: have 154 importance matrix entries llama_quantize: quantize time = 68.83 ms llama_quantize: total time = 68.83 ms ### UD-Q6_K_XL_APPROX (Q6_K) ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices ggml_metal_library_init: using embedded metal library ggml_metal_library_init: loaded in 0.009 sec ggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s) ggml_metal_device_init: GPU name: MTL0 (Apple M4 Max) ggml_metal_device_init: GPU family: MTLGPUFamilyApple9 (1009) ggml_metal_device_init: GPU family: MTLGPUFamilyCommon3 (3003) ggml_metal_device_init: GPU family: MTLGPUFamilyMetal4 (5002) ggml_metal_device_init: simdgroup reduction = true ggml_metal_device_init: simdgroup matrix mul. = true ggml_metal_device_init: has unified memory = true ggml_metal_device_init: has bfloat = true ggml_metal_device_init: has tensor = false ggml_metal_device_init: use residency sets = true ggml_metal_device_init: use shared buffers = true ggml_metal_device_init: recommendedMaxWorkingSetSize = 55662.79 MB llama_print_build_info: build = 1 (399739d) llama_print_build_info: built with AppleClang 17.0.0.17000013 for Darwin arm64 llama_quantize: calculating quantization size for 'release_work/lfm25-moe8-mac-trainer/artifacts/gguf/bf16/LFM-2.5-8B-1B-hermes-ft-BF16.gguf' as Q6_K llama_model_loader: loaded meta data with 32 key-value pairs and 256 tensors from release_work/lfm25-moe8-mac-trainer/artifacts/gguf/bf16/LFM-2.5-8B-1B-hermes-ft-BF16.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 = lfm2moe llama_model_loader: - kv 1: general.type str = model llama_model_loader: - kv 2: general.name str = Model_Upload_Iter10_Fused_Dequantized llama_model_loader: - kv 3: general.size_label str = 32x959M llama_model_loader: - kv 4: general.tags arr[str,2] = ["mlx", "text-generation"] llama_model_loader: - kv 5: general.languages arr[str,1] = ["en"] llama_model_loader: - kv 6: lfm2moe.block_count u32 = 24 llama_model_loader: - kv 7: lfm2moe.context_length u32 = 128000 llama_model_loader: - kv 8: lfm2moe.embedding_length u32 = 2048 llama_model_loader: - kv 9: lfm2moe.feed_forward_length u32 = 7168 llama_model_loader: - kv 10: lfm2moe.attention.head_count u32 = 32 llama_model_loader: - kv 11: lfm2moe.attention.head_count_kv arr[i32,24] = [0, 0, 8, 0, 0, 0, 8, 0, 0, 0, 8, 0, ... llama_model_loader: - kv 12: lfm2moe.rope.freq_base f32 = 5000000.000000 llama_model_loader: - kv 13: lfm2moe.attention.layer_norm_rms_epsilon f32 = 0.000010 llama_model_loader: - kv 14: lfm2moe.expert_count u32 = 32 llama_model_loader: - kv 15: lfm2moe.expert_used_count u32 = 4 llama_model_loader: - kv 16: general.file_type u32 = 32 llama_model_loader: - kv 17: lfm2moe.expert_feed_forward_length u32 = 1792 llama_model_loader: - kv 18: lfm2moe.leading_dense_block_count u32 = 2 llama_model_loader: - kv 19: lfm2moe.expert_gating_func u32 = 2 llama_model_loader: - kv 20: lfm2moe.vocab_size u32 = 128000 llama_model_loader: - kv 21: lfm2moe.shortconv.l_cache u32 = 3 llama_model_loader: - kv 22: general.quantization_version u32 = 2 llama_model_loader: - kv 23: tokenizer.ggml.model str = gpt2 llama_model_loader: - kv 24: tokenizer.ggml.pre str = lfm2 llama_model_loader: - kv 25: tokenizer.ggml.tokens arr[str,128000] = ["ÄŠÄŠÄŠÄŠÄŠÄŠÄŠÄŠÄŠÄŠ", "ÄŠÄŠÄŠÄŠÄŠÄ... llama_model_loader: - kv 26: tokenizer.ggml.token_type arr[i32,128000] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ... llama_model_loader: - kv 27: tokenizer.ggml.merges arr[str,293320] = ["ÄŠÄŠÄŠÄŠÄŠÄŠÄŠÄŠÄŠ ÄŠ", "ÄŠÄŠÄŠÄŠÄŠ... llama_model_loader: - kv 28: tokenizer.ggml.bos_token_id u32 = 124894 llama_model_loader: - kv 29: tokenizer.ggml.eos_token_id u32 = 124900 llama_model_loader: - kv 30: tokenizer.ggml.padding_token_id u32 = 124893 llama_model_loader: - kv 31: tokenizer.chat_template str = {{- bos_token -}}\n{%- set preserve_th... llama_model_loader: - type f32: 123 tensors llama_model_loader: - type bf16: 133 tensors llama_model_quantize_impl: have importance matrix data with 154 entries llama_tensor_get_type: token_embd.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.0.ffn_gate.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.1.ffn_gate.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.2.attn_k.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.2.attn_output.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.2.attn_q.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.2.attn_v.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.2.ffn_gate_exps.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.3.ffn_gate_exps.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.4.ffn_gate_exps.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.5.ffn_gate_exps.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.6.attn_k.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.6.attn_output.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.6.attn_q.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.6.attn_v.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.6.ffn_gate_exps.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.7.ffn_gate_exps.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.8.ffn_gate_exps.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.9.ffn_gate_exps.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.10.attn_k.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.10.attn_output.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.10.attn_q.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.10.attn_v.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.10.ffn_gate_exps.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.11.ffn_gate_exps.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.12.ffn_gate_exps.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.13.ffn_gate_exps.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.14.attn_k.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.14.attn_output.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.14.attn_q.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.14.attn_v.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.14.ffn_gate_exps.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.15.ffn_gate_exps.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.16.ffn_gate_exps.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.17.ffn_gate_exps.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.18.attn_k.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.18.attn_output.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.18.attn_q.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.18.attn_v.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.18.ffn_gate_exps.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.19.ffn_gate_exps.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.20.ffn_gate_exps.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.21.attn_k.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.21.attn_output.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.21.attn_q.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.21.attn_v.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.21.ffn_gate_exps.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.22.ffn_gate_exps.weight - applying manual override: q6_K -> q8_0 llama_tensor_get_type: blk.23.ffn_gate_exps.weight - applying manual override: q6_K -> q8_0 [ 1/ 256] token_embd.weight - [ 2048, 128000, 1, 1], type = bf16, size = 500.00 MiB -> 265.62 MiB (q8_0) [ 2/ 256] token_embd_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 3/ 256] blk.0.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 4/ 256] blk.0.ffn_down.weight - [ 7168, 2048, 1, 1], type = bf16, size = 28.00 MiB -> 11.48 MiB (q6_K) [ 5/ 256] blk.0.ffn_gate.weight - [ 2048, 7168, 1, 1], type = bf16, size = 28.00 MiB -> 14.88 MiB (q8_0) [ 6/ 256] blk.0.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 7/ 256] blk.0.ffn_up.weight - [ 2048, 7168, 1, 1], type = bf16, size = 28.00 MiB -> 11.48 MiB (q6_K) [ 8/ 256] blk.0.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 9/ 256] blk.0.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 9.84 MiB (q6_K) [ 10/ 256] blk.0.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 11/ 256] blk.1.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 12/ 256] blk.1.ffn_down.weight - [ 7168, 2048, 1, 1], type = bf16, size = 28.00 MiB -> 11.48 MiB (q6_K) [ 13/ 256] blk.1.ffn_gate.weight - [ 2048, 7168, 1, 1], type = bf16, size = 28.00 MiB -> 14.88 MiB (q8_0) [ 14/ 256] blk.1.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 15/ 256] blk.1.ffn_up.weight - [ 2048, 7168, 1, 1], type = bf16, size = 28.00 MiB -> 11.48 MiB (q6_K) [ 16/ 256] blk.1.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 17/ 256] blk.1.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 9.84 MiB (q6_K) [ 18/ 256] blk.1.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 19/ 256] blk.2.attn_k.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 1.06 MiB (q8_0) [ 20/ 256] blk.2.attn_k_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 21/ 256] blk.2.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 22/ 256] blk.2.attn_output.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 23/ 256] blk.2.attn_q.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 24/ 256] blk.2.attn_q_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 25/ 256] blk.2.attn_v.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 1.06 MiB (q8_0) [ 26/ 256] blk.2.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 27/ 256] blk.2.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 28/ 256] blk.2.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 29/ 256] blk.2.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 30/ 256] blk.2.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 31/ 256] blk.2.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 32/ 256] blk.3.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 33/ 256] blk.3.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 34/ 256] blk.3.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 35/ 256] blk.3.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 36/ 256] blk.3.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 37/ 256] blk.3.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 38/ 256] blk.3.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 39/ 256] blk.3.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 40/ 256] blk.3.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 9.84 MiB (q6_K) [ 41/ 256] blk.3.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 42/ 256] blk.4.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 43/ 256] blk.4.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 44/ 256] blk.4.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 45/ 256] blk.4.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 46/ 256] blk.4.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 47/ 256] blk.4.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 48/ 256] blk.4.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 49/ 256] blk.4.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 50/ 256] blk.4.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 9.84 MiB (q6_K) [ 51/ 256] blk.4.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 52/ 256] blk.5.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 53/ 256] blk.5.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 54/ 256] blk.5.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 55/ 256] blk.5.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 56/ 256] blk.5.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 57/ 256] blk.5.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 58/ 256] blk.5.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 59/ 256] blk.5.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 60/ 256] blk.5.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 9.84 MiB (q6_K) [ 61/ 256] blk.5.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 62/ 256] blk.6.attn_k.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 1.06 MiB (q8_0) [ 63/ 256] blk.6.attn_k_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 64/ 256] blk.6.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 65/ 256] blk.6.attn_output.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 66/ 256] blk.6.attn_q.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 67/ 256] blk.6.attn_q_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 68/ 256] blk.6.attn_v.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 1.06 MiB (q8_0) [ 69/ 256] blk.6.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 70/ 256] blk.6.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 71/ 256] blk.6.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 72/ 256] blk.6.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 73/ 256] blk.6.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 74/ 256] blk.6.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 75/ 256] blk.7.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 76/ 256] blk.7.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 77/ 256] blk.7.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 78/ 256] blk.7.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 79/ 256] blk.7.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 80/ 256] blk.7.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 81/ 256] blk.7.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 82/ 256] blk.7.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 83/ 256] blk.7.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 9.84 MiB (q6_K) [ 84/ 256] blk.7.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 85/ 256] blk.8.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 86/ 256] blk.8.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 87/ 256] blk.8.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 88/ 256] blk.8.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 89/ 256] blk.8.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 90/ 256] blk.8.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 91/ 256] blk.8.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 92/ 256] blk.8.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 93/ 256] blk.8.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 9.84 MiB (q6_K) [ 94/ 256] blk.8.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 95/ 256] blk.9.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 96/ 256] blk.9.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 97/ 256] blk.9.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 98/ 256] blk.9.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 99/ 256] blk.9.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 100/ 256] blk.9.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 101/ 256] blk.9.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 102/ 256] blk.9.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 103/ 256] blk.9.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 9.84 MiB (q6_K) [ 104/ 256] blk.9.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 105/ 256] blk.10.attn_k.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 1.06 MiB (q8_0) [ 106/ 256] blk.10.attn_k_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 107/ 256] blk.10.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 108/ 256] blk.10.attn_output.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 109/ 256] blk.10.attn_q.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 110/ 256] blk.10.attn_q_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 111/ 256] blk.10.attn_v.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 1.06 MiB (q8_0) [ 112/ 256] blk.10.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 113/ 256] blk.10.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 114/ 256] blk.10.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 115/ 256] blk.10.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 116/ 256] blk.10.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 117/ 256] blk.10.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 118/ 256] blk.11.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 119/ 256] blk.11.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 120/ 256] blk.11.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 121/ 256] blk.11.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 122/ 256] blk.11.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 123/ 256] blk.11.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 124/ 256] blk.11.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 125/ 256] blk.11.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 126/ 256] blk.11.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 9.84 MiB (q6_K) [ 127/ 256] blk.11.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 128/ 256] blk.12.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 129/ 256] blk.12.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 130/ 256] blk.12.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 131/ 256] blk.12.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 132/ 256] blk.12.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 133/ 256] blk.12.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 134/ 256] blk.12.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 135/ 256] blk.12.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 136/ 256] blk.12.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 9.84 MiB (q6_K) [ 137/ 256] blk.12.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 138/ 256] blk.13.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 139/ 256] blk.13.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 140/ 256] blk.13.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 141/ 256] blk.13.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 142/ 256] blk.13.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 143/ 256] blk.13.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 144/ 256] blk.13.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 145/ 256] blk.13.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 146/ 256] blk.13.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 9.84 MiB (q6_K) [ 147/ 256] blk.13.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 148/ 256] blk.14.attn_k.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 1.06 MiB (q8_0) [ 149/ 256] blk.14.attn_k_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 150/ 256] blk.14.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 151/ 256] blk.14.attn_output.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 152/ 256] blk.14.attn_q.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 153/ 256] blk.14.attn_q_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 154/ 256] blk.14.attn_v.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 1.06 MiB (q8_0) [ 155/ 256] blk.14.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 156/ 256] blk.14.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 157/ 256] blk.14.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 158/ 256] blk.14.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 159/ 256] blk.14.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 160/ 256] blk.14.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 161/ 256] blk.15.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 162/ 256] blk.15.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 163/ 256] blk.15.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 164/ 256] blk.15.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 165/ 256] blk.15.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 166/ 256] blk.15.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 167/ 256] blk.15.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 168/ 256] blk.15.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 169/ 256] blk.15.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 9.84 MiB (q6_K) [ 170/ 256] blk.15.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 171/ 256] blk.16.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 172/ 256] blk.16.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 173/ 256] blk.16.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 174/ 256] blk.16.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 175/ 256] blk.16.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 176/ 256] blk.16.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 177/ 256] blk.16.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 178/ 256] blk.16.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 179/ 256] blk.16.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 9.84 MiB (q6_K) [ 180/ 256] blk.16.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 181/ 256] blk.17.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 182/ 256] blk.17.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 183/ 256] blk.17.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 184/ 256] blk.17.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 185/ 256] blk.17.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 186/ 256] blk.17.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 187/ 256] blk.17.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 188/ 256] blk.17.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 189/ 256] blk.17.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 9.84 MiB (q6_K) [ 190/ 256] blk.17.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 191/ 256] blk.18.attn_k.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 1.06 MiB (q8_0) [ 192/ 256] blk.18.attn_k_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 193/ 256] blk.18.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 194/ 256] blk.18.attn_output.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 195/ 256] blk.18.attn_q.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 196/ 256] blk.18.attn_q_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 197/ 256] blk.18.attn_v.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 1.06 MiB (q8_0) [ 198/ 256] blk.18.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 199/ 256] blk.18.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 200/ 256] blk.18.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 201/ 256] blk.18.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 202/ 256] blk.18.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 203/ 256] blk.18.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 204/ 256] blk.19.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 205/ 256] blk.19.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 206/ 256] blk.19.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 207/ 256] blk.19.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 208/ 256] blk.19.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 209/ 256] blk.19.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 210/ 256] blk.19.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 211/ 256] blk.19.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 212/ 256] blk.19.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 9.84 MiB (q6_K) [ 213/ 256] blk.19.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 214/ 256] blk.20.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 215/ 256] blk.20.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 216/ 256] blk.20.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 217/ 256] blk.20.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 218/ 256] blk.20.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 219/ 256] blk.20.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 220/ 256] blk.20.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 221/ 256] blk.20.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 222/ 256] blk.20.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 9.84 MiB (q6_K) [ 223/ 256] blk.20.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 224/ 256] blk.21.attn_k.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 1.06 MiB (q8_0) [ 225/ 256] blk.21.attn_k_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 226/ 256] blk.21.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 227/ 256] blk.21.attn_output.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 228/ 256] blk.21.attn_q.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 229/ 256] blk.21.attn_q_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 230/ 256] blk.21.attn_v.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 1.06 MiB (q8_0) [ 231/ 256] blk.21.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 232/ 256] blk.21.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 233/ 256] blk.21.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 234/ 256] blk.21.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 235/ 256] blk.21.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 236/ 256] blk.21.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 237/ 256] blk.22.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 238/ 256] blk.22.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 239/ 256] blk.22.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 240/ 256] blk.22.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 241/ 256] blk.22.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 242/ 256] blk.22.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 243/ 256] blk.22.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 244/ 256] blk.22.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 245/ 256] blk.22.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 9.84 MiB (q6_K) [ 246/ 256] blk.22.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 247/ 256] blk.23.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 248/ 256] blk.23.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 249/ 256] blk.23.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 250/ 256] blk.23.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 251/ 256] blk.23.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 252/ 256] blk.23.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 253/ 256] blk.23.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 91.88 MiB (q6_K) [ 254/ 256] blk.23.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 255/ 256] blk.23.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 9.84 MiB (q6_K) [ 256/ 256] blk.23.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) llama_model_quantize_impl: model size = 16154.31 MiB (16.00 BPW) llama_model_quantize_impl: quant size = 7308.12 MiB (7.24 BPW) load_imatrix: imatrix datasets=['release_work/lfm25-moe8-mac-trainer/artifacts/gguf/calibration/hermes_tool_router_calibration.txt'] load_imatrix: loaded 154 importance matrix entries from release_work/lfm25-moe8-mac-trainer/artifacts/gguf/calibration/hermes_tool_router_imatrix.gguf computed on 32 chunks prepare_imatrix: have 154 importance matrix entries llama_quantize: quantize time = 70.15 ms llama_quantize: total time = 70.15 ms ### UD-Q5_K_XL_APPROX (Q5_K_M) ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices ggml_metal_library_init: using embedded metal library ggml_metal_library_init: loaded in 0.010 sec ggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s) ggml_metal_device_init: GPU name: MTL0 (Apple M4 Max) ggml_metal_device_init: GPU family: MTLGPUFamilyApple9 (1009) ggml_metal_device_init: GPU family: MTLGPUFamilyCommon3 (3003) ggml_metal_device_init: GPU family: MTLGPUFamilyMetal4 (5002) ggml_metal_device_init: simdgroup reduction = true ggml_metal_device_init: simdgroup matrix mul. = true ggml_metal_device_init: has unified memory = true ggml_metal_device_init: has bfloat = true ggml_metal_device_init: has tensor = false ggml_metal_device_init: use residency sets = true ggml_metal_device_init: use shared buffers = true ggml_metal_device_init: recommendedMaxWorkingSetSize = 55662.79 MB llama_print_build_info: build = 1 (399739d) llama_print_build_info: built with AppleClang 17.0.0.17000013 for Darwin arm64 llama_quantize: calculating quantization size for 'release_work/lfm25-moe8-mac-trainer/artifacts/gguf/bf16/LFM-2.5-8B-1B-hermes-ft-BF16.gguf' as Q5_K_M llama_model_loader: loaded meta data with 32 key-value pairs and 256 tensors from release_work/lfm25-moe8-mac-trainer/artifacts/gguf/bf16/LFM-2.5-8B-1B-hermes-ft-BF16.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 = lfm2moe llama_model_loader: - kv 1: general.type str = model llama_model_loader: - kv 2: general.name str = Model_Upload_Iter10_Fused_Dequantized llama_model_loader: - kv 3: general.size_label str = 32x959M llama_model_loader: - kv 4: general.tags arr[str,2] = ["mlx", "text-generation"] llama_model_loader: - kv 5: general.languages arr[str,1] = ["en"] llama_model_loader: - kv 6: lfm2moe.block_count u32 = 24 llama_model_loader: - kv 7: lfm2moe.context_length u32 = 128000 llama_model_loader: - kv 8: lfm2moe.embedding_length u32 = 2048 llama_model_loader: - kv 9: lfm2moe.feed_forward_length u32 = 7168 llama_model_loader: - kv 10: lfm2moe.attention.head_count u32 = 32 llama_model_loader: - kv 11: lfm2moe.attention.head_count_kv arr[i32,24] = [0, 0, 8, 0, 0, 0, 8, 0, 0, 0, 8, 0, ... llama_model_loader: - kv 12: lfm2moe.rope.freq_base f32 = 5000000.000000 llama_model_loader: - kv 13: lfm2moe.attention.layer_norm_rms_epsilon f32 = 0.000010 llama_model_loader: - kv 14: lfm2moe.expert_count u32 = 32 llama_model_loader: - kv 15: lfm2moe.expert_used_count u32 = 4 llama_model_loader: - kv 16: general.file_type u32 = 32 llama_model_loader: - kv 17: lfm2moe.expert_feed_forward_length u32 = 1792 llama_model_loader: - kv 18: lfm2moe.leading_dense_block_count u32 = 2 llama_model_loader: - kv 19: lfm2moe.expert_gating_func u32 = 2 llama_model_loader: - kv 20: lfm2moe.vocab_size u32 = 128000 llama_model_loader: - kv 21: lfm2moe.shortconv.l_cache u32 = 3 llama_model_loader: - kv 22: general.quantization_version u32 = 2 llama_model_loader: - kv 23: tokenizer.ggml.model str = gpt2 llama_model_loader: - kv 24: tokenizer.ggml.pre str = lfm2 llama_model_loader: - kv 25: tokenizer.ggml.tokens arr[str,128000] = ["ÄŠÄŠÄŠÄŠÄŠÄŠÄŠÄŠÄŠÄŠ", "ÄŠÄŠÄŠÄŠÄŠÄ... llama_model_loader: - kv 26: tokenizer.ggml.token_type arr[i32,128000] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ... llama_model_loader: - kv 27: tokenizer.ggml.merges arr[str,293320] = ["ÄŠÄŠÄŠÄŠÄŠÄŠÄŠÄŠÄŠ ÄŠ", "ÄŠÄŠÄŠÄŠÄŠ... llama_model_loader: - kv 28: tokenizer.ggml.bos_token_id u32 = 124894 llama_model_loader: - kv 29: tokenizer.ggml.eos_token_id u32 = 124900 llama_model_loader: - kv 30: tokenizer.ggml.padding_token_id u32 = 124893 llama_model_loader: - kv 31: tokenizer.chat_template str = {{- bos_token -}}\n{%- set preserve_th... llama_model_loader: - type f32: 123 tensors llama_model_loader: - type bf16: 133 tensors llama_model_quantize_impl: have importance matrix data with 154 entries llama_tensor_get_type: token_embd.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.0.ffn_gate.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.1.ffn_gate.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.2.attn_k.weight - applying manual override: q5_K -> q6_K llama_tensor_get_type: blk.2.attn_output.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.2.attn_q.weight - applying manual override: q5_K -> q6_K llama_tensor_get_type: blk.2.attn_v.weight - applying manual override: q5_K -> q6_K llama_tensor_get_type: blk.2.ffn_gate_exps.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.3.ffn_gate_exps.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.4.ffn_gate_exps.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.5.ffn_gate_exps.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.6.attn_k.weight - applying manual override: q5_K -> q6_K llama_tensor_get_type: blk.6.attn_output.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.6.attn_q.weight - applying manual override: q5_K -> q6_K llama_tensor_get_type: blk.6.attn_v.weight - applying manual override: q5_K -> q6_K llama_tensor_get_type: blk.6.ffn_gate_exps.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.7.ffn_gate_exps.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.8.ffn_gate_exps.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.9.ffn_gate_exps.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.10.attn_k.weight - applying manual override: q5_K -> q6_K llama_tensor_get_type: blk.10.attn_output.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.10.attn_q.weight - applying manual override: q5_K -> q6_K llama_tensor_get_type: blk.10.attn_v.weight - applying manual override: q5_K -> q6_K llama_tensor_get_type: blk.10.ffn_gate_exps.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.11.ffn_gate_exps.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.12.ffn_gate_exps.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.13.ffn_gate_exps.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.14.attn_k.weight - applying manual override: q5_K -> q6_K llama_tensor_get_type: blk.14.attn_output.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.14.attn_q.weight - applying manual override: q5_K -> q6_K llama_tensor_get_type: blk.14.attn_v.weight - applying manual override: q5_K -> q6_K llama_tensor_get_type: blk.14.ffn_gate_exps.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.15.ffn_gate_exps.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.16.ffn_gate_exps.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.17.ffn_gate_exps.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.18.attn_k.weight - applying manual override: q5_K -> q6_K llama_tensor_get_type: blk.18.attn_output.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.18.attn_q.weight - applying manual override: q5_K -> q6_K llama_tensor_get_type: blk.18.attn_v.weight - applying manual override: q5_K -> q6_K llama_tensor_get_type: blk.18.ffn_gate_exps.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.19.ffn_gate_exps.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.20.ffn_gate_exps.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.21.attn_k.weight - applying manual override: q5_K -> q6_K llama_tensor_get_type: blk.21.attn_output.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.21.attn_q.weight - applying manual override: q5_K -> q6_K llama_tensor_get_type: blk.21.attn_v.weight - applying manual override: q5_K -> q6_K llama_tensor_get_type: blk.21.ffn_gate_exps.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.22.ffn_gate_exps.weight - applying manual override: q5_K -> q8_0 llama_tensor_get_type: blk.23.ffn_gate_exps.weight - applying manual override: q5_K -> q8_0 [ 1/ 256] token_embd.weight - [ 2048, 128000, 1, 1], type = bf16, size = 500.00 MiB -> 265.62 MiB (q8_0) [ 2/ 256] token_embd_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 3/ 256] blk.0.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 4/ 256] blk.0.ffn_down.weight - [ 7168, 2048, 1, 1], type = bf16, size = 28.00 MiB -> 9.62 MiB (q5_K) [ 5/ 256] blk.0.ffn_gate.weight - [ 2048, 7168, 1, 1], type = bf16, size = 28.00 MiB -> 14.88 MiB (q8_0) [ 6/ 256] blk.0.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 7/ 256] blk.0.ffn_up.weight - [ 2048, 7168, 1, 1], type = bf16, size = 28.00 MiB -> 9.62 MiB (q5_K) [ 8/ 256] blk.0.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 9/ 256] blk.0.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 8.25 MiB (q5_K) [ 10/ 256] blk.0.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.75 MiB (q5_K) [ 11/ 256] blk.1.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 12/ 256] blk.1.ffn_down.weight - [ 7168, 2048, 1, 1], type = bf16, size = 28.00 MiB -> 9.62 MiB (q5_K) [ 13/ 256] blk.1.ffn_gate.weight - [ 2048, 7168, 1, 1], type = bf16, size = 28.00 MiB -> 14.88 MiB (q8_0) [ 14/ 256] blk.1.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 15/ 256] blk.1.ffn_up.weight - [ 2048, 7168, 1, 1], type = bf16, size = 28.00 MiB -> 9.62 MiB (q5_K) [ 16/ 256] blk.1.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 17/ 256] blk.1.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 8.25 MiB (q5_K) [ 18/ 256] blk.1.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.75 MiB (q5_K) [ 19/ 256] blk.2.attn_k.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 0.82 MiB (q6_K) [ 20/ 256] blk.2.attn_k_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 21/ 256] blk.2.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 22/ 256] blk.2.attn_output.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 23/ 256] blk.2.attn_q.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 24/ 256] blk.2.attn_q_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 25/ 256] blk.2.attn_v.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 0.82 MiB (q6_K) [ 26/ 256] blk.2.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 27/ 256] blk.2.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 28/ 256] blk.2.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 29/ 256] blk.2.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 30/ 256] blk.2.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 31/ 256] blk.2.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 32/ 256] blk.3.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 33/ 256] blk.3.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 34/ 256] blk.3.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 35/ 256] blk.3.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 36/ 256] blk.3.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 37/ 256] blk.3.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 38/ 256] blk.3.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 39/ 256] blk.3.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 40/ 256] blk.3.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 8.25 MiB (q5_K) [ 41/ 256] blk.3.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.75 MiB (q5_K) [ 42/ 256] blk.4.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 43/ 256] blk.4.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 44/ 256] blk.4.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 45/ 256] blk.4.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 46/ 256] blk.4.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 47/ 256] blk.4.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 48/ 256] blk.4.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 49/ 256] blk.4.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 50/ 256] blk.4.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 8.25 MiB (q5_K) [ 51/ 256] blk.4.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.75 MiB (q5_K) [ 52/ 256] blk.5.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 53/ 256] blk.5.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 54/ 256] blk.5.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 55/ 256] blk.5.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 56/ 256] blk.5.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 57/ 256] blk.5.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 58/ 256] blk.5.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 59/ 256] blk.5.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 60/ 256] blk.5.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 8.25 MiB (q5_K) [ 61/ 256] blk.5.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.75 MiB (q5_K) [ 62/ 256] blk.6.attn_k.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 0.82 MiB (q6_K) [ 63/ 256] blk.6.attn_k_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 64/ 256] blk.6.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 65/ 256] blk.6.attn_output.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 66/ 256] blk.6.attn_q.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 67/ 256] blk.6.attn_q_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 68/ 256] blk.6.attn_v.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 0.82 MiB (q6_K) [ 69/ 256] blk.6.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 70/ 256] blk.6.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 71/ 256] blk.6.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 72/ 256] blk.6.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 73/ 256] blk.6.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 74/ 256] blk.6.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 75/ 256] blk.7.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 76/ 256] blk.7.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 77/ 256] blk.7.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 78/ 256] blk.7.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 79/ 256] blk.7.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 80/ 256] blk.7.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 81/ 256] blk.7.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 82/ 256] blk.7.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 83/ 256] blk.7.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 8.25 MiB (q5_K) [ 84/ 256] blk.7.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.75 MiB (q5_K) [ 85/ 256] blk.8.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 86/ 256] blk.8.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 87/ 256] blk.8.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 88/ 256] blk.8.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 89/ 256] blk.8.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 90/ 256] blk.8.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 91/ 256] blk.8.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 92/ 256] blk.8.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 93/ 256] blk.8.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 8.25 MiB (q5_K) [ 94/ 256] blk.8.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.75 MiB (q5_K) [ 95/ 256] blk.9.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 96/ 256] blk.9.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 97/ 256] blk.9.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 98/ 256] blk.9.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 99/ 256] blk.9.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 100/ 256] blk.9.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 101/ 256] blk.9.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 102/ 256] blk.9.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 103/ 256] blk.9.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 8.25 MiB (q5_K) [ 104/ 256] blk.9.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.75 MiB (q5_K) [ 105/ 256] blk.10.attn_k.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 0.82 MiB (q6_K) [ 106/ 256] blk.10.attn_k_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 107/ 256] blk.10.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 108/ 256] blk.10.attn_output.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 109/ 256] blk.10.attn_q.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 110/ 256] blk.10.attn_q_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 111/ 256] blk.10.attn_v.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 0.82 MiB (q6_K) [ 112/ 256] blk.10.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 113/ 256] blk.10.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 114/ 256] blk.10.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 115/ 256] blk.10.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 116/ 256] blk.10.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 117/ 256] blk.10.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 118/ 256] blk.11.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 119/ 256] blk.11.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 120/ 256] blk.11.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 121/ 256] blk.11.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 122/ 256] blk.11.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 123/ 256] blk.11.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 124/ 256] blk.11.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 125/ 256] blk.11.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 126/ 256] blk.11.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 8.25 MiB (q5_K) [ 127/ 256] blk.11.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.75 MiB (q5_K) [ 128/ 256] blk.12.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 129/ 256] blk.12.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 130/ 256] blk.12.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 131/ 256] blk.12.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 132/ 256] blk.12.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 133/ 256] blk.12.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 134/ 256] blk.12.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 135/ 256] blk.12.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 136/ 256] blk.12.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 8.25 MiB (q5_K) [ 137/ 256] blk.12.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.75 MiB (q5_K) [ 138/ 256] blk.13.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 139/ 256] blk.13.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 140/ 256] blk.13.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 141/ 256] blk.13.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 142/ 256] blk.13.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 143/ 256] blk.13.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 144/ 256] blk.13.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 145/ 256] blk.13.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 146/ 256] blk.13.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 8.25 MiB (q5_K) [ 147/ 256] blk.13.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.75 MiB (q5_K) [ 148/ 256] blk.14.attn_k.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 0.82 MiB (q6_K) [ 149/ 256] blk.14.attn_k_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 150/ 256] blk.14.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 151/ 256] blk.14.attn_output.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 152/ 256] blk.14.attn_q.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 153/ 256] blk.14.attn_q_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 154/ 256] blk.14.attn_v.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 0.82 MiB (q6_K) [ 155/ 256] blk.14.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 156/ 256] blk.14.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 157/ 256] blk.14.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 158/ 256] blk.14.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 159/ 256] blk.14.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 160/ 256] blk.14.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 161/ 256] blk.15.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 162/ 256] blk.15.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 163/ 256] blk.15.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 164/ 256] blk.15.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 165/ 256] blk.15.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 166/ 256] blk.15.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 167/ 256] blk.15.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 168/ 256] blk.15.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 169/ 256] blk.15.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 8.25 MiB (q5_K) [ 170/ 256] blk.15.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.75 MiB (q5_K) [ 171/ 256] blk.16.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 172/ 256] blk.16.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 173/ 256] blk.16.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 174/ 256] blk.16.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 175/ 256] blk.16.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 176/ 256] blk.16.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 177/ 256] blk.16.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 178/ 256] blk.16.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 179/ 256] blk.16.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 8.25 MiB (q5_K) [ 180/ 256] blk.16.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.75 MiB (q5_K) [ 181/ 256] blk.17.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 182/ 256] blk.17.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 183/ 256] blk.17.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 184/ 256] blk.17.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 185/ 256] blk.17.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 186/ 256] blk.17.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 187/ 256] blk.17.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 188/ 256] blk.17.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 189/ 256] blk.17.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 8.25 MiB (q5_K) [ 190/ 256] blk.17.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.75 MiB (q5_K) [ 191/ 256] blk.18.attn_k.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 0.82 MiB (q6_K) [ 192/ 256] blk.18.attn_k_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 193/ 256] blk.18.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 194/ 256] blk.18.attn_output.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 195/ 256] blk.18.attn_q.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 196/ 256] blk.18.attn_q_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 197/ 256] blk.18.attn_v.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 0.82 MiB (q6_K) [ 198/ 256] blk.18.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 199/ 256] blk.18.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 200/ 256] blk.18.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 201/ 256] blk.18.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 202/ 256] blk.18.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 203/ 256] blk.18.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 204/ 256] blk.19.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 205/ 256] blk.19.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 206/ 256] blk.19.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 207/ 256] blk.19.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 208/ 256] blk.19.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 209/ 256] blk.19.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 210/ 256] blk.19.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 211/ 256] blk.19.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 212/ 256] blk.19.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 8.25 MiB (q5_K) [ 213/ 256] blk.19.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.75 MiB (q5_K) [ 214/ 256] blk.20.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 215/ 256] blk.20.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 216/ 256] blk.20.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 217/ 256] blk.20.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 218/ 256] blk.20.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 219/ 256] blk.20.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 220/ 256] blk.20.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 221/ 256] blk.20.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 222/ 256] blk.20.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 8.25 MiB (q5_K) [ 223/ 256] blk.20.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.75 MiB (q5_K) [ 224/ 256] blk.21.attn_k.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 0.82 MiB (q6_K) [ 225/ 256] blk.21.attn_k_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 226/ 256] blk.21.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 227/ 256] blk.21.attn_output.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 4.25 MiB (q8_0) [ 228/ 256] blk.21.attn_q.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 229/ 256] blk.21.attn_q_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 230/ 256] blk.21.attn_v.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 0.82 MiB (q6_K) [ 231/ 256] blk.21.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 232/ 256] blk.21.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 233/ 256] blk.21.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 234/ 256] blk.21.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 235/ 256] blk.21.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 236/ 256] blk.21.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 237/ 256] blk.22.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 238/ 256] blk.22.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 239/ 256] blk.22.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 240/ 256] blk.22.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 241/ 256] blk.22.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 242/ 256] blk.22.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 243/ 256] blk.22.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 244/ 256] blk.22.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 245/ 256] blk.22.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 8.25 MiB (q5_K) [ 246/ 256] blk.22.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.75 MiB (q5_K) [ 247/ 256] blk.23.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 248/ 256] blk.23.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 249/ 256] blk.23.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 250/ 256] blk.23.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 251/ 256] blk.23.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 252/ 256] blk.23.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 253/ 256] blk.23.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 77.00 MiB (q5_K) [ 254/ 256] blk.23.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 255/ 256] blk.23.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 8.25 MiB (q5_K) [ 256/ 256] blk.23.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.75 MiB (q5_K) llama_model_quantize_impl: model size = 16154.31 MiB (16.00 BPW) llama_model_quantize_impl: quant size = 6599.22 MiB (6.54 BPW) load_imatrix: imatrix datasets=['release_work/lfm25-moe8-mac-trainer/artifacts/gguf/calibration/hermes_tool_router_calibration.txt'] load_imatrix: loaded 154 importance matrix entries from release_work/lfm25-moe8-mac-trainer/artifacts/gguf/calibration/hermes_tool_router_imatrix.gguf computed on 32 chunks prepare_imatrix: have 154 importance matrix entries llama_quantize: quantize time = 74.64 ms llama_quantize: total time = 74.64 ms ### UD-Q4_K_XL_APPROX (Q4_K_M) ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices ggml_metal_library_init: using embedded metal library ggml_metal_library_init: loaded in 0.009 sec ggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s) ggml_metal_device_init: GPU name: MTL0 (Apple M4 Max) ggml_metal_device_init: GPU family: MTLGPUFamilyApple9 (1009) ggml_metal_device_init: GPU family: MTLGPUFamilyCommon3 (3003) ggml_metal_device_init: GPU family: MTLGPUFamilyMetal4 (5002) ggml_metal_device_init: simdgroup reduction = true ggml_metal_device_init: simdgroup matrix mul. = true ggml_metal_device_init: has unified memory = true ggml_metal_device_init: has bfloat = true ggml_metal_device_init: has tensor = false ggml_metal_device_init: use residency sets = true ggml_metal_device_init: use shared buffers = true ggml_metal_device_init: recommendedMaxWorkingSetSize = 55662.79 MB llama_print_build_info: build = 1 (399739d) llama_print_build_info: built with AppleClang 17.0.0.17000013 for Darwin arm64 llama_quantize: calculating quantization size for 'release_work/lfm25-moe8-mac-trainer/artifacts/gguf/bf16/LFM-2.5-8B-1B-hermes-ft-BF16.gguf' as Q4_K_M llama_model_loader: loaded meta data with 32 key-value pairs and 256 tensors from release_work/lfm25-moe8-mac-trainer/artifacts/gguf/bf16/LFM-2.5-8B-1B-hermes-ft-BF16.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 = lfm2moe llama_model_loader: - kv 1: general.type str = model llama_model_loader: - kv 2: general.name str = Model_Upload_Iter10_Fused_Dequantized llama_model_loader: - kv 3: general.size_label str = 32x959M llama_model_loader: - kv 4: general.tags arr[str,2] = ["mlx", "text-generation"] llama_model_loader: - kv 5: general.languages arr[str,1] = ["en"] llama_model_loader: - kv 6: lfm2moe.block_count u32 = 24 llama_model_loader: - kv 7: lfm2moe.context_length u32 = 128000 llama_model_loader: - kv 8: lfm2moe.embedding_length u32 = 2048 llama_model_loader: - kv 9: lfm2moe.feed_forward_length u32 = 7168 llama_model_loader: - kv 10: lfm2moe.attention.head_count u32 = 32 llama_model_loader: - kv 11: lfm2moe.attention.head_count_kv arr[i32,24] = [0, 0, 8, 0, 0, 0, 8, 0, 0, 0, 8, 0, ... llama_model_loader: - kv 12: lfm2moe.rope.freq_base f32 = 5000000.000000 llama_model_loader: - kv 13: lfm2moe.attention.layer_norm_rms_epsilon f32 = 0.000010 llama_model_loader: - kv 14: lfm2moe.expert_count u32 = 32 llama_model_loader: - kv 15: lfm2moe.expert_used_count u32 = 4 llama_model_loader: - kv 16: general.file_type u32 = 32 llama_model_loader: - kv 17: lfm2moe.expert_feed_forward_length u32 = 1792 llama_model_loader: - kv 18: lfm2moe.leading_dense_block_count u32 = 2 llama_model_loader: - kv 19: lfm2moe.expert_gating_func u32 = 2 llama_model_loader: - kv 20: lfm2moe.vocab_size u32 = 128000 llama_model_loader: - kv 21: lfm2moe.shortconv.l_cache u32 = 3 llama_model_loader: - kv 22: general.quantization_version u32 = 2 llama_model_loader: - kv 23: tokenizer.ggml.model str = gpt2 llama_model_loader: - kv 24: tokenizer.ggml.pre str = lfm2 llama_model_loader: - kv 25: tokenizer.ggml.tokens arr[str,128000] = ["ÄŠÄŠÄŠÄŠÄŠÄŠÄŠÄŠÄŠÄŠ", "ÄŠÄŠÄŠÄŠÄŠÄ... llama_model_loader: - kv 26: tokenizer.ggml.token_type arr[i32,128000] = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ... llama_model_loader: - kv 27: tokenizer.ggml.merges arr[str,293320] = ["ÄŠÄŠÄŠÄŠÄŠÄŠÄŠÄŠÄŠ ÄŠ", "ÄŠÄŠÄŠÄŠÄŠ... llama_model_loader: - kv 28: tokenizer.ggml.bos_token_id u32 = 124894 llama_model_loader: - kv 29: tokenizer.ggml.eos_token_id u32 = 124900 llama_model_loader: - kv 30: tokenizer.ggml.padding_token_id u32 = 124893 llama_model_loader: - kv 31: tokenizer.chat_template str = {{- bos_token -}}\n{%- set preserve_th... llama_model_loader: - type f32: 123 tensors llama_model_loader: - type bf16: 133 tensors llama_model_quantize_impl: have importance matrix data with 154 entries llama_tensor_get_type: token_embd.weight - applying manual override: q4_K -> q6_K llama_tensor_get_type: blk.0.ffn_gate.weight - applying manual override: q4_K -> q8_0 llama_tensor_get_type: blk.1.ffn_gate.weight - applying manual override: q4_K -> q8_0 llama_tensor_get_type: blk.2.attn_k.weight - applying manual override: q4_K -> q6_K llama_tensor_get_type: blk.2.attn_output.weight - applying manual override: q4_K -> q6_K llama_tensor_get_type: blk.2.attn_q.weight - applying manual override: q4_K -> q6_K llama_tensor_get_type: blk.2.attn_v.weight - applying manual override: q4_K -> q5_K llama_tensor_get_type: blk.2.ffn_gate_exps.weight - applying manual override: q4_K -> q8_0 llama_tensor_get_type: blk.3.ffn_gate_exps.weight - applying manual override: q4_K -> q8_0 llama_tensor_get_type: blk.4.ffn_gate_exps.weight - applying manual override: q4_K -> q8_0 llama_tensor_get_type: blk.5.ffn_gate_exps.weight - applying manual override: q4_K -> q8_0 llama_tensor_get_type: blk.6.attn_k.weight - applying manual override: q4_K -> q6_K llama_tensor_get_type: blk.6.attn_output.weight - applying manual override: q4_K -> q6_K llama_tensor_get_type: blk.6.attn_q.weight - applying manual override: q4_K -> q6_K llama_tensor_get_type: blk.6.attn_v.weight - applying manual override: q4_K -> q5_K llama_tensor_get_type: blk.6.ffn_gate_exps.weight - applying manual override: q4_K -> q8_0 llama_tensor_get_type: blk.7.ffn_gate_exps.weight - applying manual override: q4_K -> q8_0 llama_tensor_get_type: blk.8.ffn_gate_exps.weight - applying manual override: q4_K -> q8_0 llama_tensor_get_type: blk.9.ffn_gate_exps.weight - applying manual override: q4_K -> q8_0 llama_tensor_get_type: blk.10.attn_k.weight - applying manual override: q4_K -> q6_K llama_tensor_get_type: blk.10.attn_output.weight - applying manual override: q4_K -> q6_K llama_tensor_get_type: blk.10.attn_q.weight - applying manual override: q4_K -> q6_K llama_tensor_get_type: blk.10.attn_v.weight - applying manual override: q4_K -> q5_K llama_tensor_get_type: blk.10.ffn_gate_exps.weight - applying manual override: q4_K -> q8_0 llama_tensor_get_type: blk.11.ffn_gate_exps.weight - applying manual override: q4_K -> q8_0 llama_tensor_get_type: blk.12.ffn_gate_exps.weight - applying manual override: q4_K -> q8_0 llama_tensor_get_type: blk.13.ffn_gate_exps.weight - applying manual override: q4_K -> q8_0 llama_tensor_get_type: blk.14.attn_k.weight - applying manual override: q4_K -> q6_K llama_tensor_get_type: blk.14.attn_output.weight - applying manual override: q4_K -> q6_K llama_tensor_get_type: blk.14.attn_q.weight - applying manual override: q4_K -> q6_K llama_tensor_get_type: blk.14.attn_v.weight - applying manual override: q4_K -> q5_K llama_tensor_get_type: blk.14.ffn_gate_exps.weight - applying manual override: q4_K -> q8_0 llama_tensor_get_type: blk.15.ffn_gate_exps.weight - applying manual override: q4_K -> q8_0 llama_tensor_get_type: blk.16.ffn_gate_exps.weight - applying manual override: q4_K -> q8_0 llama_tensor_get_type: blk.17.ffn_gate_exps.weight - applying manual override: q4_K -> q8_0 llama_tensor_get_type: blk.18.attn_k.weight - applying manual override: q4_K -> q6_K llama_tensor_get_type: blk.18.attn_output.weight - applying manual override: q4_K -> q6_K llama_tensor_get_type: blk.18.attn_q.weight - applying manual override: q4_K -> q6_K llama_tensor_get_type: blk.18.attn_v.weight - applying manual override: q4_K -> q5_K llama_tensor_get_type: blk.18.ffn_gate_exps.weight - applying manual override: q4_K -> q8_0 llama_tensor_get_type: blk.19.ffn_gate_exps.weight - applying manual override: q4_K -> q8_0 llama_tensor_get_type: blk.20.ffn_gate_exps.weight - applying manual override: q4_K -> q8_0 llama_tensor_get_type: blk.21.attn_k.weight - applying manual override: q4_K -> q6_K llama_tensor_get_type: blk.21.attn_output.weight - applying manual override: q4_K -> q6_K llama_tensor_get_type: blk.21.attn_q.weight - applying manual override: q4_K -> q6_K llama_tensor_get_type: blk.21.attn_v.weight - applying manual override: q4_K -> q5_K llama_tensor_get_type: blk.21.ffn_gate_exps.weight - applying manual override: q4_K -> q8_0 llama_tensor_get_type: blk.22.ffn_gate_exps.weight - applying manual override: q4_K -> q8_0 llama_tensor_get_type: blk.23.ffn_gate_exps.weight - applying manual override: q4_K -> q8_0 [ 1/ 256] token_embd.weight - [ 2048, 128000, 1, 1], type = bf16, size = 500.00 MiB -> 205.08 MiB (q6_K) [ 2/ 256] token_embd_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 3/ 256] blk.0.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 4/ 256] blk.0.ffn_down.weight - [ 7168, 2048, 1, 1], type = bf16, size = 28.00 MiB -> 7.88 MiB (q4_K) [ 5/ 256] blk.0.ffn_gate.weight - [ 2048, 7168, 1, 1], type = bf16, size = 28.00 MiB -> 14.88 MiB (q8_0) [ 6/ 256] blk.0.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 7/ 256] blk.0.ffn_up.weight - [ 2048, 7168, 1, 1], type = bf16, size = 28.00 MiB -> 7.88 MiB (q4_K) [ 8/ 256] blk.0.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 9/ 256] blk.0.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 6.75 MiB (q4_K) [ 10/ 256] blk.0.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.25 MiB (q4_K) [ 11/ 256] blk.1.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 12/ 256] blk.1.ffn_down.weight - [ 7168, 2048, 1, 1], type = bf16, size = 28.00 MiB -> 7.88 MiB (q4_K) [ 13/ 256] blk.1.ffn_gate.weight - [ 2048, 7168, 1, 1], type = bf16, size = 28.00 MiB -> 14.88 MiB (q8_0) [ 14/ 256] blk.1.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 15/ 256] blk.1.ffn_up.weight - [ 2048, 7168, 1, 1], type = bf16, size = 28.00 MiB -> 7.88 MiB (q4_K) [ 16/ 256] blk.1.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 17/ 256] blk.1.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 6.75 MiB (q4_K) [ 18/ 256] blk.1.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.25 MiB (q4_K) [ 19/ 256] blk.2.attn_k.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 0.82 MiB (q6_K) [ 20/ 256] blk.2.attn_k_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 21/ 256] blk.2.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 22/ 256] blk.2.attn_output.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 23/ 256] blk.2.attn_q.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 24/ 256] blk.2.attn_q_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 25/ 256] blk.2.attn_v.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 0.69 MiB (q5_K) [ 26/ 256] blk.2.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 27/ 256] blk.2.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 28/ 256] blk.2.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 29/ 256] blk.2.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 30/ 256] blk.2.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 31/ 256] blk.2.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 32/ 256] blk.3.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 33/ 256] blk.3.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 34/ 256] blk.3.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 35/ 256] blk.3.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 36/ 256] blk.3.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 37/ 256] blk.3.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 38/ 256] blk.3.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 39/ 256] blk.3.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 40/ 256] blk.3.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 6.75 MiB (q4_K) [ 41/ 256] blk.3.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.25 MiB (q4_K) [ 42/ 256] blk.4.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 43/ 256] blk.4.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 44/ 256] blk.4.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 45/ 256] blk.4.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 46/ 256] blk.4.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 47/ 256] blk.4.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 48/ 256] blk.4.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 49/ 256] blk.4.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 50/ 256] blk.4.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 6.75 MiB (q4_K) [ 51/ 256] blk.4.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.25 MiB (q4_K) [ 52/ 256] blk.5.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 53/ 256] blk.5.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 54/ 256] blk.5.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 55/ 256] blk.5.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 56/ 256] blk.5.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 57/ 256] blk.5.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 58/ 256] blk.5.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 59/ 256] blk.5.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 60/ 256] blk.5.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 6.75 MiB (q4_K) [ 61/ 256] blk.5.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.25 MiB (q4_K) [ 62/ 256] blk.6.attn_k.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 0.82 MiB (q6_K) [ 63/ 256] blk.6.attn_k_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 64/ 256] blk.6.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 65/ 256] blk.6.attn_output.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 66/ 256] blk.6.attn_q.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 67/ 256] blk.6.attn_q_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 68/ 256] blk.6.attn_v.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 0.69 MiB (q5_K) [ 69/ 256] blk.6.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 70/ 256] blk.6.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 71/ 256] blk.6.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 72/ 256] blk.6.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 73/ 256] blk.6.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 74/ 256] blk.6.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 75/ 256] blk.7.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 76/ 256] blk.7.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 77/ 256] blk.7.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 78/ 256] blk.7.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 79/ 256] blk.7.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 80/ 256] blk.7.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 81/ 256] blk.7.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 82/ 256] blk.7.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 83/ 256] blk.7.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 6.75 MiB (q4_K) [ 84/ 256] blk.7.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.25 MiB (q4_K) [ 85/ 256] blk.8.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 86/ 256] blk.8.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 87/ 256] blk.8.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 88/ 256] blk.8.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 89/ 256] blk.8.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 90/ 256] blk.8.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 91/ 256] blk.8.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 92/ 256] blk.8.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 93/ 256] blk.8.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 6.75 MiB (q4_K) [ 94/ 256] blk.8.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.25 MiB (q4_K) [ 95/ 256] blk.9.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 96/ 256] blk.9.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 97/ 256] blk.9.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 98/ 256] blk.9.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 99/ 256] blk.9.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 100/ 256] blk.9.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 101/ 256] blk.9.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 102/ 256] blk.9.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 103/ 256] blk.9.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 6.75 MiB (q4_K) [ 104/ 256] blk.9.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.25 MiB (q4_K) [ 105/ 256] blk.10.attn_k.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 0.82 MiB (q6_K) [ 106/ 256] blk.10.attn_k_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 107/ 256] blk.10.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 108/ 256] blk.10.attn_output.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 109/ 256] blk.10.attn_q.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 110/ 256] blk.10.attn_q_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 111/ 256] blk.10.attn_v.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 0.69 MiB (q5_K) [ 112/ 256] blk.10.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 113/ 256] blk.10.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 114/ 256] blk.10.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 115/ 256] blk.10.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 116/ 256] blk.10.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 117/ 256] blk.10.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 118/ 256] blk.11.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 119/ 256] blk.11.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 120/ 256] blk.11.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 121/ 256] blk.11.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 122/ 256] blk.11.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 123/ 256] blk.11.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 124/ 256] blk.11.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 125/ 256] blk.11.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 126/ 256] blk.11.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 6.75 MiB (q4_K) [ 127/ 256] blk.11.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.25 MiB (q4_K) [ 128/ 256] blk.12.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 129/ 256] blk.12.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 130/ 256] blk.12.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 131/ 256] blk.12.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 132/ 256] blk.12.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 133/ 256] blk.12.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 134/ 256] blk.12.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 135/ 256] blk.12.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 136/ 256] blk.12.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 6.75 MiB (q4_K) [ 137/ 256] blk.12.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.25 MiB (q4_K) [ 138/ 256] blk.13.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 139/ 256] blk.13.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 140/ 256] blk.13.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 141/ 256] blk.13.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 142/ 256] blk.13.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 143/ 256] blk.13.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 144/ 256] blk.13.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 145/ 256] blk.13.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 146/ 256] blk.13.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 6.75 MiB (q4_K) [ 147/ 256] blk.13.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.25 MiB (q4_K) [ 148/ 256] blk.14.attn_k.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 0.82 MiB (q6_K) [ 149/ 256] blk.14.attn_k_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 150/ 256] blk.14.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 151/ 256] blk.14.attn_output.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 152/ 256] blk.14.attn_q.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 153/ 256] blk.14.attn_q_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 154/ 256] blk.14.attn_v.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 0.69 MiB (q5_K) [ 155/ 256] blk.14.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 156/ 256] blk.14.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 157/ 256] blk.14.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 158/ 256] blk.14.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 159/ 256] blk.14.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 160/ 256] blk.14.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 161/ 256] blk.15.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 162/ 256] blk.15.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 163/ 256] blk.15.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 164/ 256] blk.15.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 165/ 256] blk.15.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 166/ 256] blk.15.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 167/ 256] blk.15.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 168/ 256] blk.15.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 169/ 256] blk.15.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 6.75 MiB (q4_K) [ 170/ 256] blk.15.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.25 MiB (q4_K) [ 171/ 256] blk.16.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 172/ 256] blk.16.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 173/ 256] blk.16.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 174/ 256] blk.16.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 175/ 256] blk.16.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 176/ 256] blk.16.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 177/ 256] blk.16.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 178/ 256] blk.16.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 179/ 256] blk.16.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 6.75 MiB (q4_K) [ 180/ 256] blk.16.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.25 MiB (q4_K) [ 181/ 256] blk.17.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 182/ 256] blk.17.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 183/ 256] blk.17.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 184/ 256] blk.17.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 185/ 256] blk.17.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 186/ 256] blk.17.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 187/ 256] blk.17.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 188/ 256] blk.17.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 189/ 256] blk.17.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 6.75 MiB (q4_K) [ 190/ 256] blk.17.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.25 MiB (q4_K) [ 191/ 256] blk.18.attn_k.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 0.82 MiB (q6_K) [ 192/ 256] blk.18.attn_k_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 193/ 256] blk.18.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 194/ 256] blk.18.attn_output.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 195/ 256] blk.18.attn_q.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 196/ 256] blk.18.attn_q_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 197/ 256] blk.18.attn_v.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 0.69 MiB (q5_K) [ 198/ 256] blk.18.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 199/ 256] blk.18.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 200/ 256] blk.18.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 201/ 256] blk.18.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 202/ 256] blk.18.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 203/ 256] blk.18.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 204/ 256] blk.19.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 205/ 256] blk.19.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 206/ 256] blk.19.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 207/ 256] blk.19.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 208/ 256] blk.19.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 209/ 256] blk.19.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 210/ 256] blk.19.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 211/ 256] blk.19.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 212/ 256] blk.19.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 6.75 MiB (q4_K) [ 213/ 256] blk.19.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.25 MiB (q4_K) [ 214/ 256] blk.20.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 215/ 256] blk.20.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 216/ 256] blk.20.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 217/ 256] blk.20.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 218/ 256] blk.20.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 219/ 256] blk.20.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 220/ 256] blk.20.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 221/ 256] blk.20.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 222/ 256] blk.20.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 6.75 MiB (q4_K) [ 223/ 256] blk.20.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.25 MiB (q4_K) [ 224/ 256] blk.21.attn_k.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 0.82 MiB (q6_K) [ 225/ 256] blk.21.attn_k_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 226/ 256] blk.21.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 227/ 256] blk.21.attn_output.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 228/ 256] blk.21.attn_q.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 3.28 MiB (q6_K) [ 229/ 256] blk.21.attn_q_norm.weight - [ 64, 1, 1, 1], type = f32, size = 0.000 MiB [ 230/ 256] blk.21.attn_v.weight - [ 2048, 512, 1, 1], type = bf16, size = 2.00 MiB -> 0.69 MiB (q5_K) [ 231/ 256] blk.21.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 232/ 256] blk.21.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 233/ 256] blk.21.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 234/ 256] blk.21.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 235/ 256] blk.21.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 236/ 256] blk.21.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 237/ 256] blk.22.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 238/ 256] blk.22.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 239/ 256] blk.22.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 240/ 256] blk.22.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 241/ 256] blk.22.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 242/ 256] blk.22.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 243/ 256] blk.22.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 244/ 256] blk.22.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 245/ 256] blk.22.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 6.75 MiB (q4_K) [ 246/ 256] blk.22.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.25 MiB (q4_K) [ 247/ 256] blk.23.attn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 248/ 256] blk.23.exp_probs_b.bias - [ 32, 1, 1, 1], type = f32, size = 0.000 MiB [ 249/ 256] blk.23.ffn_down_exps.weight - [ 1792, 2048, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 250/ 256] blk.23.ffn_gate_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 119.00 MiB (q8_0) [ 251/ 256] blk.23.ffn_gate_inp.weight - [ 2048, 32, 1, 1], type = f32, size = 0.250 MiB [ 252/ 256] blk.23.ffn_norm.weight - [ 2048, 1, 1, 1], type = f32, size = 0.008 MiB [ 253/ 256] blk.23.ffn_up_exps.weight - [ 2048, 1792, 32, 1], type = bf16, size = 224.00 MiB -> 63.00 MiB (q4_K) [ 254/ 256] blk.23.shortconv.conv.weight - [ 3, 2048, 1, 1], type = f32, size = 0.023 MiB [ 255/ 256] blk.23.shortconv.in_proj.weight - [ 2048, 6144, 1, 1], type = bf16, size = 24.00 MiB -> 6.75 MiB (q4_K) [ 256/ 256] blk.23.shortconv.out_proj.weight - [ 2048, 2048, 1, 1], type = bf16, size = 8.00 MiB -> 2.25 MiB (q4_K) llama_model_quantize_impl: model size = 16154.31 MiB (16.00 BPW) llama_model_quantize_impl: quant size = 5873.06 MiB (5.82 BPW) load_imatrix: imatrix datasets=['release_work/lfm25-moe8-mac-trainer/artifacts/gguf/calibration/hermes_tool_router_calibration.txt'] load_imatrix: loaded 154 importance matrix entries from release_work/lfm25-moe8-mac-trainer/artifacts/gguf/calibration/hermes_tool_router_imatrix.gguf computed on 32 chunks prepare_imatrix: have 154 importance matrix entries llama_quantize: quantize time = 75.32 ms llama_quantize: total time = 75.32 ms