lfm3:1.7b-qad-q4_0

7,771 36 minutes ago

LFM3-1.7B is a tri-hybrid model for local agents and specialized applications.

tools thinking 1.7b
43a7d09a0e9c · 1.0GB
    Metadata
  • general.architecture
    lfm3
  • general.file_type
    Q4_0
  • lfm3.attention.head_count
    16
  • lfm3.attention.head_count_kv
    [0, 0, 4, 0, 4, ...]
  • lfm3.attention.key_length
    128
  • lfm3.attention.layer_norm_rms_epsilon
    1e-05
  • lfm3.attention.sliding_window
    256
  • lfm3.attention.sliding_window_pattern
    [false, false, false, false, true, ...]
  • lfm3.attention.value_length
    128
  • lfm3.block_count
    24
  • lfm3.context_length
    262144
  • lfm3.embedding_length
    2048
  • lfm3.feed_forward_length
    7680
  • lfm3.rope.dimension_count
    128
  • lfm3.rope.freq_base
    1e+06
  • lfm3.rope.freq_base_swa
    10000
  • lfm3.shortconv.l_cache
    7
  • tokenizer.ggml.add_bos_token
    false
  • tokenizer.ggml.add_eos_token
    false
  • tokenizer.ggml.bos_token_id
    124894
  • tokenizer.ggml.eos_token_id
    124900
  • tokenizer.ggml.merges
    [ĊĊĊĊĊĊĊĊĊ Ċ, ĊĊĊĊĊĊĊĊ ĊĊ, ĊĊĊĊĊĊĊ ĊĊĊ, ĊĊĊĊĊĊ ĊĊĊĊ, ĊĊĊĊĊ ĊĊĊĊĊ, ...]
  • tokenizer.ggml.model
    gpt2
  • tokenizer.ggml.padding_token_id
    124893
  • tokenizer.ggml.pre
    lfm2
  • tokenizer.ggml.token_type
    [1, 1, 1, 1, 1, ...]
  • tokenizer.ggml.tokens
    [ĊĊĊĊĊĊĊĊĊĊ, ĊĊĊĊĊĊĊĊĊ, ĊĊĊĊĊĊĊĊ, ĊĊĊĊĊĊĊ, ĊĊĊĊĊĊ, ...]
  • Tensor
  • token_embd.weight
    Q6_K
    [2048, 128000]
  • blk.0
  • blk.0.attn_norm.weight
    F32
    [2048]
  • blk.0.ffn_down.weight
    Q4_0
    [7680, 2048]
  • blk.0.ffn_gate.weight
    Q4_0
    [2048, 7680]
  • blk.0.ffn_norm.weight
    F32
    [2048]
  • blk.0.ffn_up.weight
    Q4_0
    [2048, 7680]
  • blk.0.post_attention_norm.weight
    F32
    [2048]
  • blk.0.post_ffw_norm.weight
    F32
    [2048]
  • blk.0.shortconv.conv.weight
    F32
    [7, 2048]
  • blk.0.shortconv.in_proj.weight
    Q4_0
    [2048, 6144]
  • blk.0.shortconv.out_proj.weight
    Q4_0
    [2048, 2048]
  • blk.1
  • blk.1.attn_norm.weight
    F32
    [2048]
  • blk.1.ffn_down.weight
    Q4_0
    [7680, 2048]
  • blk.1.ffn_gate.weight
    Q4_0
    [2048, 7680]
  • blk.1.ffn_norm.weight
    F32
    [2048]
  • blk.1.ffn_up.weight
    Q4_0
    [2048, 7680]
  • blk.1.post_attention_norm.weight
    F32
    [2048]
  • blk.1.post_ffw_norm.weight
    F32
    [2048]
  • blk.1.shortconv.conv.weight
    F32
    [7, 2048]
  • blk.1.shortconv.in_proj.weight
    Q4_0
    [2048, 6144]
  • blk.1.shortconv.out_proj.weight
    Q4_0
    [2048, 2048]
  • blk.2
  • blk.2.attn_gate.weight
    Q4_0
    [2048, 16]
  • blk.2.attn_k.weight
    Q4_0
    [2048, 512]
  • blk.2.attn_k_norm.weight
    F32
    [128]
  • blk.2.attn_norm.weight
    F32
    [2048]
  • blk.2.attn_output.weight
    Q4_0
    [2048, 2048]
  • blk.2.attn_q.weight
    Q4_0
    [2048, 2048]
  • blk.2.attn_q_norm.weight
    F32
    [128]
  • blk.2.attn_v.weight
    Q4_0
    [2048, 512]
  • blk.2.ffn_down.weight
    Q4_0
    [7680, 2048]
  • blk.2.ffn_gate.weight
    Q4_0
    [2048, 7680]
  • blk.2.ffn_norm.weight
    F32
    [2048]
  • blk.2.ffn_up.weight
    Q4_0
    [2048, 7680]
  • blk.2.post_attention_norm.weight
    F32
    [2048]
  • blk.2.post_ffw_norm.weight
    F32
    [2048]
  • blk.3
  • blk.3.attn_norm.weight
    F32
    [2048]
  • blk.3.ffn_down.weight
    Q4_0
    [7680, 2048]
  • blk.3.ffn_gate.weight
    Q4_0
    [2048, 7680]
  • blk.3.ffn_norm.weight
    F32
    [2048]
  • blk.3.ffn_up.weight
    Q4_0
    [2048, 7680]
  • blk.3.post_attention_norm.weight
    F32
    [2048]
  • blk.3.post_ffw_norm.weight
    F32
    [2048]
  • blk.3.shortconv.conv.weight
    F32
    [7, 2048]
  • blk.3.shortconv.in_proj.weight
    Q4_0
    [2048, 6144]
  • blk.3.shortconv.out_proj.weight
    Q4_0
    [2048, 2048]
  • blk.4
  • blk.4.attn_gate.weight
    Q4_0
    [2048, 16]
  • blk.4.attn_k.weight
    Q4_0
    [2048, 512]
  • blk.4.attn_k_norm.weight
    F32
    [128]
  • blk.4.attn_norm.weight
    F32
    [2048]
  • blk.4.attn_output.weight
    Q4_0
    [2048, 2048]
  • blk.4.attn_q.weight
    Q4_0
    [2048, 2048]
  • blk.4.attn_q_norm.weight
    F32
    [128]
  • blk.4.attn_v.weight
    Q4_0
    [2048, 512]
  • blk.4.ffn_down.weight
    Q4_0
    [7680, 2048]
  • blk.4.ffn_gate.weight
    Q4_0
    [2048, 7680]
  • blk.4.ffn_norm.weight
    F32
    [2048]
  • blk.4.ffn_up.weight
    Q4_0
    [2048, 7680]
  • blk.4.post_attention_norm.weight
    F32
    [2048]
  • blk.4.post_ffw_norm.weight
    F32
    [2048]
  • blk.5
  • blk.5.attn_norm.weight
    F32
    [2048]
  • blk.5.ffn_down.weight
    Q4_0
    [7680, 2048]
  • blk.5.ffn_gate.weight
    Q4_0
    [2048, 7680]
  • blk.5.ffn_norm.weight
    F32
    [2048]
  • blk.5.ffn_up.weight
    Q4_0
    [2048, 7680]
  • blk.5.post_attention_norm.weight
    F32
    [2048]
  • blk.5.post_ffw_norm.weight
    F32
    [2048]
  • blk.5.shortconv.conv.weight
    F32
    [7, 2048]
  • blk.5.shortconv.in_proj.weight
    Q4_0
    [2048, 6144]
  • blk.5.shortconv.out_proj.weight
    Q4_0
    [2048, 2048]
  • blk.6
  • blk.6.attn_gate.weight
    Q4_0
    [2048, 16]
  • blk.6.attn_k.weight
    Q4_0
    [2048, 512]
  • blk.6.attn_k_norm.weight
    F32
    [128]
  • blk.6.attn_norm.weight
    F32
    [2048]
  • blk.6.attn_output.weight
    Q4_0
    [2048, 2048]
  • blk.6.attn_q.weight
    Q4_0
    [2048, 2048]
  • blk.6.attn_q_norm.weight
    F32
    [128]
  • blk.6.attn_v.weight
    Q4_0
    [2048, 512]
  • blk.6.ffn_down.weight
    Q4_0
    [7680, 2048]
  • blk.6.ffn_gate.weight
    Q4_0
    [2048, 7680]
  • blk.6.ffn_norm.weight
    F32
    [2048]
  • blk.6.ffn_up.weight
    Q4_0
    [2048, 7680]
  • blk.6.post_attention_norm.weight
    F32
    [2048]
  • blk.6.post_ffw_norm.weight
    F32
    [2048]
  • blk.7
  • blk.7.attn_norm.weight
    F32
    [2048]
  • blk.7.ffn_down.weight
    Q4_0
    [7680, 2048]
  • blk.7.ffn_gate.weight
    Q4_0
    [2048, 7680]
  • blk.7.ffn_norm.weight
    F32
    [2048]
  • blk.7.ffn_up.weight
    Q4_0
    [2048, 7680]
  • blk.7.post_attention_norm.weight
    F32
    [2048]
  • blk.7.post_ffw_norm.weight
    F32
    [2048]
  • blk.7.shortconv.conv.weight
    F32
    [7, 2048]
  • blk.7.shortconv.in_proj.weight
    Q4_0
    [2048, 6144]
  • blk.7.shortconv.out_proj.weight
    Q4_0
    [2048, 2048]
  • blk.8
  • blk.8.attn_gate.weight
    Q4_0
    [2048, 16]
  • blk.8.attn_k.weight
    Q4_0
    [2048, 512]
  • blk.8.attn_k_norm.weight
    F32
    [128]
  • blk.8.attn_norm.weight
    F32
    [2048]
  • blk.8.attn_output.weight
    Q4_0
    [2048, 2048]
  • blk.8.attn_q.weight
    Q4_0
    [2048, 2048]
  • blk.8.attn_q_norm.weight
    F32
    [128]
  • blk.8.attn_v.weight
    Q4_0
    [2048, 512]
  • blk.8.ffn_down.weight
    Q4_0
    [7680, 2048]
  • blk.8.ffn_gate.weight
    Q4_0
    [2048, 7680]
  • blk.8.ffn_norm.weight
    F32
    [2048]
  • blk.8.ffn_up.weight
    Q4_0
    [2048, 7680]
  • blk.8.post_attention_norm.weight
    F32
    [2048]
  • blk.8.post_ffw_norm.weight
    F32
    [2048]
  • blk.9
  • blk.9.attn_norm.weight
    F32
    [2048]
  • blk.9.ffn_down.weight
    Q4_0
    [7680, 2048]
  • blk.9.ffn_gate.weight
    Q4_0
    [2048, 7680]
  • blk.9.ffn_norm.weight
    F32
    [2048]
  • blk.9.ffn_up.weight
    Q4_0
    [2048, 7680]
  • blk.9.post_attention_norm.weight
    F32
    [2048]
  • blk.9.post_ffw_norm.weight
    F32
    [2048]
  • blk.9.shortconv.conv.weight
    F32
    [7, 2048]
  • blk.9.shortconv.in_proj.weight
    Q4_0
    [2048, 6144]
  • blk.9.shortconv.out_proj.weight
    Q4_0
    [2048, 2048]
  • blk.10
  • blk.10.attn_gate.weight
    Q4_0
    [2048, 16]
  • blk.10.attn_k.weight
    Q4_0
    [2048, 512]
  • blk.10.attn_k_norm.weight
    F32
    [128]
  • blk.10.attn_norm.weight
    F32
    [2048]
  • blk.10.attn_output.weight
    Q4_0
    [2048, 2048]
  • blk.10.attn_q.weight
    Q4_0
    [2048, 2048]
  • blk.10.attn_q_norm.weight
    F32
    [128]
  • blk.10.attn_v.weight
    Q4_0
    [2048, 512]
  • blk.10.ffn_down.weight
    Q4_0
    [7680, 2048]
  • blk.10.ffn_gate.weight
    Q4_0
    [2048, 7680]
  • blk.10.ffn_norm.weight
    F32
    [2048]
  • blk.10.ffn_up.weight
    Q4_0
    [2048, 7680]
  • blk.10.post_attention_norm.weight
    F32
    [2048]
  • blk.10.post_ffw_norm.weight
    F32
    [2048]
  • blk.11
  • blk.11.attn_norm.weight
    F32
    [2048]
  • blk.11.ffn_down.weight
    Q4_0
    [7680, 2048]
  • blk.11.ffn_gate.weight
    Q4_0
    [2048, 7680]
  • blk.11.ffn_norm.weight
    F32
    [2048]
  • blk.11.ffn_up.weight
    Q4_0
    [2048, 7680]
  • blk.11.post_attention_norm.weight
    F32
    [2048]
  • blk.11.post_ffw_norm.weight
    F32
    [2048]
  • blk.11.shortconv.conv.weight
    F32
    [7, 2048]
  • blk.11.shortconv.in_proj.weight
    Q4_0
    [2048, 6144]
  • blk.11.shortconv.out_proj.weight
    Q4_0
    [2048, 2048]
  • blk.12
  • blk.12.attn_gate.weight
    Q4_0
    [2048, 16]
  • blk.12.attn_k.weight
    Q4_0
    [2048, 512]
  • blk.12.attn_k_norm.weight
    F32
    [128]
  • blk.12.attn_norm.weight
    F32
    [2048]
  • blk.12.attn_output.weight
    Q4_0
    [2048, 2048]
  • blk.12.attn_q.weight
    Q4_0
    [2048, 2048]
  • blk.12.attn_q_norm.weight
    F32
    [128]
  • blk.12.attn_v.weight
    Q4_0
    [2048, 512]
  • blk.12.ffn_down.weight
    Q4_0
    [7680, 2048]
  • blk.12.ffn_gate.weight
    Q4_0
    [2048, 7680]
  • blk.12.ffn_norm.weight
    F32
    [2048]
  • blk.12.ffn_up.weight
    Q4_0
    [2048, 7680]
  • blk.12.post_attention_norm.weight
    F32
    [2048]
  • blk.12.post_ffw_norm.weight
    F32
    [2048]
  • blk.13
  • blk.13.attn_norm.weight
    F32
    [2048]
  • blk.13.ffn_down.weight
    Q4_0
    [7680, 2048]
  • blk.13.ffn_gate.weight
    Q4_0
    [2048, 7680]
  • blk.13.ffn_norm.weight
    F32
    [2048]
  • blk.13.ffn_up.weight
    Q4_0
    [2048, 7680]
  • blk.13.post_attention_norm.weight
    F32
    [2048]
  • blk.13.post_ffw_norm.weight
    F32
    [2048]
  • blk.13.shortconv.conv.weight
    F32
    [7, 2048]
  • blk.13.shortconv.in_proj.weight
    Q4_0
    [2048, 6144]
  • blk.13.shortconv.out_proj.weight
    Q4_0
    [2048, 2048]
  • blk.14
  • blk.14.attn_gate.weight
    Q4_0
    [2048, 16]
  • blk.14.attn_k.weight
    Q4_0
    [2048, 512]
  • blk.14.attn_k_norm.weight
    F32
    [128]
  • blk.14.attn_norm.weight
    F32
    [2048]
  • blk.14.attn_output.weight
    Q4_0
    [2048, 2048]
  • blk.14.attn_q.weight
    Q4_0
    [2048, 2048]
  • blk.14.attn_q_norm.weight
    F32
    [128]
  • blk.14.attn_v.weight
    Q4_0
    [2048, 512]
  • blk.14.ffn_down.weight
    Q4_0
    [7680, 2048]
  • blk.14.ffn_gate.weight
    Q4_0
    [2048, 7680]
  • blk.14.ffn_norm.weight
    F32
    [2048]
  • blk.14.ffn_up.weight
    Q4_0
    [2048, 7680]
  • blk.14.post_attention_norm.weight
    F32
    [2048]
  • blk.14.post_ffw_norm.weight
    F32
    [2048]
  • blk.15
  • blk.15.attn_norm.weight
    F32
    [2048]
  • blk.15.ffn_down.weight
    Q4_0
    [7680, 2048]
  • blk.15.ffn_gate.weight
    Q4_0
    [2048, 7680]
  • blk.15.ffn_norm.weight
    F32
    [2048]
  • blk.15.ffn_up.weight
    Q4_0
    [2048, 7680]
  • blk.15.post_attention_norm.weight
    F32
    [2048]
  • blk.15.post_ffw_norm.weight
    F32
    [2048]
  • blk.15.shortconv.conv.weight
    F32
    [7, 2048]
  • blk.15.shortconv.in_proj.weight
    Q4_0
    [2048, 6144]
  • blk.15.shortconv.out_proj.weight
    Q4_0
    [2048, 2048]
  • blk.16
  • blk.16.attn_gate.weight
    Q4_0
    [2048, 16]
  • blk.16.attn_k.weight
    Q4_0
    [2048, 512]
  • blk.16.attn_k_norm.weight
    F32
    [128]
  • blk.16.attn_norm.weight
    F32
    [2048]
  • blk.16.attn_output.weight
    Q4_0
    [2048, 2048]
  • blk.16.attn_q.weight
    Q4_0
    [2048, 2048]
  • blk.16.attn_q_norm.weight
    F32
    [128]
  • blk.16.attn_v.weight
    Q4_0
    [2048, 512]
  • blk.16.ffn_down.weight
    Q4_0
    [7680, 2048]
  • blk.16.ffn_gate.weight
    Q4_0
    [2048, 7680]
  • blk.16.ffn_norm.weight
    F32
    [2048]
  • blk.16.ffn_up.weight
    Q4_0
    [2048, 7680]
  • blk.16.post_attention_norm.weight
    F32
    [2048]
  • blk.16.post_ffw_norm.weight
    F32
    [2048]
  • blk.17
  • blk.17.attn_norm.weight
    F32
    [2048]
  • blk.17.ffn_down.weight
    Q4_0
    [7680, 2048]
  • blk.17.ffn_gate.weight
    Q4_0
    [2048, 7680]
  • blk.17.ffn_norm.weight
    F32
    [2048]
  • blk.17.ffn_up.weight
    Q4_0
    [2048, 7680]
  • blk.17.post_attention_norm.weight
    F32
    [2048]
  • blk.17.post_ffw_norm.weight
    F32
    [2048]
  • blk.17.shortconv.conv.weight
    F32
    [7, 2048]
  • blk.17.shortconv.in_proj.weight
    Q4_0
    [2048, 6144]
  • blk.17.shortconv.out_proj.weight
    Q4_0
    [2048, 2048]
  • blk.18
  • blk.18.attn_gate.weight
    Q4_0
    [2048, 16]
  • blk.18.attn_k.weight
    Q4_0
    [2048, 512]
  • blk.18.attn_k_norm.weight
    F32
    [128]
  • blk.18.attn_norm.weight
    F32
    [2048]
  • blk.18.attn_output.weight
    Q4_0
    [2048, 2048]
  • blk.18.attn_q.weight
    Q4_0
    [2048, 2048]
  • blk.18.attn_q_norm.weight
    F32
    [128]
  • blk.18.attn_v.weight
    Q4_0
    [2048, 512]
  • blk.18.ffn_down.weight
    Q4_0
    [7680, 2048]
  • blk.18.ffn_gate.weight
    Q4_0
    [2048, 7680]
  • blk.18.ffn_norm.weight
    F32
    [2048]
  • blk.18.ffn_up.weight
    Q4_0
    [2048, 7680]
  • blk.18.post_attention_norm.weight
    F32
    [2048]
  • blk.18.post_ffw_norm.weight
    F32
    [2048]
  • blk.19
  • blk.19.attn_norm.weight
    F32
    [2048]
  • blk.19.ffn_down.weight
    Q4_0
    [7680, 2048]
  • blk.19.ffn_gate.weight
    Q4_0
    [2048, 7680]
  • blk.19.ffn_norm.weight
    F32
    [2048]
  • blk.19.ffn_up.weight
    Q4_0
    [2048, 7680]
  • blk.19.post_attention_norm.weight
    F32
    [2048]
  • blk.19.post_ffw_norm.weight
    F32
    [2048]
  • blk.19.shortconv.conv.weight
    F32
    [7, 2048]
  • blk.19.shortconv.in_proj.weight
    Q4_0
    [2048, 6144]
  • blk.19.shortconv.out_proj.weight
    Q4_0
    [2048, 2048]
  • blk.20
  • blk.20.attn_gate.weight
    Q4_0
    [2048, 16]
  • blk.20.attn_k.weight
    Q4_0
    [2048, 512]
  • blk.20.attn_k_norm.weight
    F32
    [128]
  • blk.20.attn_norm.weight
    F32
    [2048]
  • blk.20.attn_output.weight
    Q4_0
    [2048, 2048]
  • blk.20.attn_q.weight
    Q4_0
    [2048, 2048]
  • blk.20.attn_q_norm.weight
    F32
    [128]
  • blk.20.attn_v.weight
    Q4_0
    [2048, 512]
  • blk.20.ffn_down.weight
    Q4_0
    [7680, 2048]
  • blk.20.ffn_gate.weight
    Q4_0
    [2048, 7680]
  • blk.20.ffn_norm.weight
    F32
    [2048]
  • blk.20.ffn_up.weight
    Q4_0
    [2048, 7680]
  • blk.20.post_attention_norm.weight
    F32
    [2048]
  • blk.20.post_ffw_norm.weight
    F32
    [2048]
  • blk.21
  • blk.21.attn_norm.weight
    F32
    [2048]
  • blk.21.ffn_down.weight
    Q4_0
    [7680, 2048]
  • blk.21.ffn_gate.weight
    Q4_0
    [2048, 7680]
  • blk.21.ffn_norm.weight
    F32
    [2048]
  • blk.21.ffn_up.weight
    Q4_0
    [2048, 7680]
  • blk.21.post_attention_norm.weight
    F32
    [2048]
  • blk.21.post_ffw_norm.weight
    F32
    [2048]
  • blk.21.shortconv.conv.weight
    F32
    [7, 2048]
  • blk.21.shortconv.in_proj.weight
    Q4_0
    [2048, 6144]
  • blk.21.shortconv.out_proj.weight
    Q4_0
    [2048, 2048]
  • blk.22
  • blk.22.attn_gate.weight
    Q4_0
    [2048, 16]
  • blk.22.attn_k.weight
    Q4_0
    [2048, 512]
  • blk.22.attn_k_norm.weight
    F32
    [128]
  • blk.22.attn_norm.weight
    F32
    [2048]
  • blk.22.attn_output.weight
    Q4_0
    [2048, 2048]
  • blk.22.attn_q.weight
    Q4_0
    [2048, 2048]
  • blk.22.attn_q_norm.weight
    F32
    [128]
  • blk.22.attn_v.weight
    Q4_0
    [2048, 512]
  • blk.22.ffn_down.weight
    Q4_0
    [7680, 2048]
  • blk.22.ffn_gate.weight
    Q4_0
    [2048, 7680]
  • blk.22.ffn_norm.weight
    F32
    [2048]
  • blk.22.ffn_up.weight
    Q4_0
    [2048, 7680]
  • blk.22.post_attention_norm.weight
    F32
    [2048]
  • blk.22.post_ffw_norm.weight
    F32
    [2048]
  • blk.23
  • blk.23.attn_norm.weight
    F32
    [2048]
  • blk.23.ffn_down.weight
    Q4_0
    [7680, 2048]
  • blk.23.ffn_gate.weight
    Q4_0
    [2048, 7680]
  • blk.23.ffn_norm.weight
    F32
    [2048]
  • blk.23.ffn_up.weight
    Q4_0
    [2048, 7680]
  • blk.23.post_attention_norm.weight
    F32
    [2048]
  • blk.23.post_ffw_norm.weight
    F32
    [2048]
  • blk.23.shortconv.conv.weight
    F32
    [7, 2048]
  • blk.23.shortconv.in_proj.weight
    Q4_0
    [2048, 6144]
  • blk.23.shortconv.out_proj.weight
    Q4_0
    [2048, 2048]
  • rope_freqs.weight
    F32
    [64]
  • output_norm.weight
    F32
    [2048]