lfm2.5-thinking:latest

98 12 hours ago

LFM2.5 is a new family of hybrid models designed for on-device deployment.

1.2b
... /
model
a7b19185f316 · 731MB
    Metadata
  • general.architecture
    lfm2
  • general.file_type
    Q4_K_M
  • lfm2.attention.head_count
    32
  • lfm2.attention.head_count_kv
    [0, 0, 8, 0, 0, ...]
  • lfm2.attention.key_length
    64
  • lfm2.attention.layer_norm_rms_epsilon
    1e-05
  • lfm2.attention.value_length
    64
  • lfm2.block_count
    16
  • lfm2.context_length
    128000
  • lfm2.embedding_length
    2048
  • lfm2.feed_forward_length
    12288
  • lfm2.rope.freq_base
    1e+06
  • lfm2.shortconv.l_cache
    3
  • lfm2.vocab_size
    65536
  • tokenizer.ggml.add_bos_token
    true
  • tokenizer.ggml.add_eos_token
    false
  • tokenizer.ggml.add_padding_token
    false
  • tokenizer.ggml.bos_token_id
    1
  • tokenizer.ggml.eos_token_id
    7
  • tokenizer.ggml.merges
    [Ċ Ċ, Ċ ĊĊ, ĊĊ Ċ, Ċ ĊĊĊ, ĊĊ ĊĊ, ...]
  • tokenizer.ggml.model
    gpt2
  • tokenizer.ggml.padding_token_id
    0
  • tokenizer.ggml.pre
    llama-bpe
  • tokenizer.ggml.scores
    [0, 1, 2, 3, 4, ...]
  • tokenizer.ggml.token_type
    [3, 3, 3, 3, 3, ...]
  • tokenizer.ggml.tokens
    [<|pad|>, <|startoftext|>, <|endoftext|>, <|fim_pre|>, <|fim_mid|>, ...]
    • Tensor
  • token_embd.weight
    Q6_K
    [2048, 65536]
    • blk.0
  • blk.0.attn_norm.weight
    F32
    [2048]
  • blk.0.ffn_down.weight
    Q6_K
    [8192, 2048]
  • blk.0.ffn_gate.weight
    Q4_K
    [2048, 8192]
  • blk.0.ffn_norm.weight
    F32
    [2048]
  • blk.0.ffn_up.weight
    Q4_K
    [2048, 8192]
  • blk.0.shortconv.conv.weight
    F32
    [3, 2048]
  • blk.0.shortconv.in_proj.weight
    Q4_K
    [2048, 6144]
  • blk.0.shortconv.out_proj.weight
    Q4_K
    [2048, 2048]
    • blk.1
  • blk.1.attn_norm.weight
    F32
    [2048]
  • blk.1.ffn_down.weight
    Q6_K
    [8192, 2048]
  • blk.1.ffn_gate.weight
    Q4_K
    [2048, 8192]
  • blk.1.ffn_norm.weight
    F32
    [2048]
  • blk.1.ffn_up.weight
    Q4_K
    [2048, 8192]
  • blk.1.shortconv.conv.weight
    F32
    [3, 2048]
  • blk.1.shortconv.in_proj.weight
    Q4_K
    [2048, 6144]
  • blk.1.shortconv.out_proj.weight
    Q4_K
    [2048, 2048]
    • blk.2
  • blk.2.attn_k.weight
    Q4_K
    [2048, 512]
  • blk.2.attn_k_norm.weight
    F32
    [64]
  • blk.2.attn_norm.weight
    F32
    [2048]
  • blk.2.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.2.attn_q.weight
    Q4_K
    [2048, 2048]
  • blk.2.attn_q_norm.weight
    F32
    [64]
  • blk.2.attn_v.weight
    Q4_K
    [2048, 512]
  • blk.2.ffn_down.weight
    Q4_K
    [8192, 2048]
  • blk.2.ffn_gate.weight
    Q4_K
    [2048, 8192]
  • blk.2.ffn_norm.weight
    F32
    [2048]
  • blk.2.ffn_up.weight
    Q4_K
    [2048, 8192]
    • blk.3
  • blk.3.attn_norm.weight
    F32
    [2048]
  • blk.3.ffn_down.weight
    Q4_K
    [8192, 2048]
  • blk.3.ffn_gate.weight
    Q4_K
    [2048, 8192]
  • blk.3.ffn_norm.weight
    F32
    [2048]
  • blk.3.ffn_up.weight
    Q4_K
    [2048, 8192]
  • blk.3.shortconv.conv.weight
    F32
    [3, 2048]
  • blk.3.shortconv.in_proj.weight
    Q4_K
    [2048, 6144]
  • blk.3.shortconv.out_proj.weight
    Q4_K
    [2048, 2048]
    • blk.4
  • blk.4.attn_norm.weight
    F32
    [2048]
  • blk.4.ffn_down.weight
    Q6_K
    [8192, 2048]
  • blk.4.ffn_gate.weight
    Q4_K
    [2048, 8192]
  • blk.4.ffn_norm.weight
    F32
    [2048]
  • blk.4.ffn_up.weight
    Q4_K
    [2048, 8192]
  • blk.4.shortconv.conv.weight
    F32
    [3, 2048]
  • blk.4.shortconv.in_proj.weight
    Q4_K
    [2048, 6144]
  • blk.4.shortconv.out_proj.weight
    Q4_K
    [2048, 2048]
    • blk.5
  • blk.5.attn_k.weight
    Q4_K
    [2048, 512]
  • blk.5.attn_k_norm.weight
    F32
    [64]
  • blk.5.attn_norm.weight
    F32
    [2048]
  • blk.5.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.5.attn_q.weight
    Q4_K
    [2048, 2048]
  • blk.5.attn_q_norm.weight
    F32
    [64]
  • blk.5.attn_v.weight
    Q4_K
    [2048, 512]
  • blk.5.ffn_down.weight
    Q4_K
    [8192, 2048]
  • blk.5.ffn_gate.weight
    Q4_K
    [2048, 8192]
  • blk.5.ffn_norm.weight
    F32
    [2048]
  • blk.5.ffn_up.weight
    Q4_K
    [2048, 8192]
    • blk.6
  • blk.6.attn_norm.weight
    F32
    [2048]
  • blk.6.ffn_down.weight
    Q4_K
    [8192, 2048]
  • blk.6.ffn_gate.weight
    Q4_K
    [2048, 8192]
  • blk.6.ffn_norm.weight
    F32
    [2048]
  • blk.6.ffn_up.weight
    Q4_K
    [2048, 8192]
  • blk.6.shortconv.conv.weight
    F32
    [3, 2048]
  • blk.6.shortconv.in_proj.weight
    Q4_K
    [2048, 6144]
  • blk.6.shortconv.out_proj.weight
    Q4_K
    [2048, 2048]
    • blk.7
  • blk.7.attn_norm.weight
    F32
    [2048]
  • blk.7.ffn_down.weight
    Q6_K
    [8192, 2048]
  • blk.7.ffn_gate.weight
    Q4_K
    [2048, 8192]
  • blk.7.ffn_norm.weight
    F32
    [2048]
  • blk.7.ffn_up.weight
    Q4_K
    [2048, 8192]
  • blk.7.shortconv.conv.weight
    F32
    [3, 2048]
  • blk.7.shortconv.in_proj.weight
    Q4_K
    [2048, 6144]
  • blk.7.shortconv.out_proj.weight
    Q4_K
    [2048, 2048]
    • blk.8
  • blk.8.attn_k.weight
    Q4_K
    [2048, 512]
  • blk.8.attn_k_norm.weight
    F32
    [64]
  • blk.8.attn_norm.weight
    F32
    [2048]
  • blk.8.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.8.attn_q.weight
    Q4_K
    [2048, 2048]
  • blk.8.attn_q_norm.weight
    F32
    [64]
  • blk.8.attn_v.weight
    Q6_K
    [2048, 512]
  • blk.8.ffn_down.weight
    Q4_K
    [8192, 2048]
  • blk.8.ffn_gate.weight
    Q4_K
    [2048, 8192]
  • blk.8.ffn_norm.weight
    F32
    [2048]
  • blk.8.ffn_up.weight
    Q4_K
    [2048, 8192]
    • blk.9
  • blk.9.attn_norm.weight
    F32
    [2048]
  • blk.9.ffn_down.weight
    Q4_K
    [8192, 2048]
  • blk.9.ffn_gate.weight
    Q4_K
    [2048, 8192]
  • blk.9.ffn_norm.weight
    F32
    [2048]
  • blk.9.ffn_up.weight
    Q4_K
    [2048, 8192]
  • blk.9.shortconv.conv.weight
    F32
    [3, 2048]
  • blk.9.shortconv.in_proj.weight
    Q4_K
    [2048, 6144]
  • blk.9.shortconv.out_proj.weight
    Q4_K
    [2048, 2048]
    • blk.10
  • blk.10.attn_k.weight
    Q4_K
    [2048, 512]
  • blk.10.attn_k_norm.weight
    F32
    [64]
  • blk.10.attn_norm.weight
    F32
    [2048]
  • blk.10.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.10.attn_q.weight
    Q4_K
    [2048, 2048]
  • blk.10.attn_q_norm.weight
    F32
    [64]
  • blk.10.attn_v.weight
    Q4_K
    [2048, 512]
  • blk.10.ffn_down.weight
    Q6_K
    [8192, 2048]
  • blk.10.ffn_gate.weight
    Q4_K
    [2048, 8192]
  • blk.10.ffn_norm.weight
    F32
    [2048]
  • blk.10.ffn_up.weight
    Q4_K
    [2048, 8192]
    • blk.11
  • blk.11.attn_norm.weight
    F32
    [2048]
  • blk.11.ffn_down.weight
    Q4_K
    [8192, 2048]
  • blk.11.ffn_gate.weight
    Q4_K
    [2048, 8192]
  • blk.11.ffn_norm.weight
    F32
    [2048]
  • blk.11.ffn_up.weight
    Q4_K
    [2048, 8192]
  • blk.11.shortconv.conv.weight
    F32
    [3, 2048]
  • blk.11.shortconv.in_proj.weight
    Q4_K
    [2048, 6144]
  • blk.11.shortconv.out_proj.weight
    Q4_K
    [2048, 2048]
    • blk.12
  • blk.12.attn_k.weight
    Q4_K
    [2048, 512]
  • blk.12.attn_k_norm.weight
    F32
    [64]
  • blk.12.attn_norm.weight
    F32
    [2048]
  • blk.12.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.12.attn_q.weight
    Q4_K
    [2048, 2048]
  • blk.12.attn_q_norm.weight
    F32
    [64]
  • blk.12.attn_v.weight
    Q4_K
    [2048, 512]
  • blk.12.ffn_down.weight
    Q4_K
    [8192, 2048]
  • blk.12.ffn_gate.weight
    Q4_K
    [2048, 8192]
  • blk.12.ffn_norm.weight
    F32
    [2048]
  • blk.12.ffn_up.weight
    Q4_K
    [2048, 8192]
    • blk.13
  • blk.13.attn_norm.weight
    F32
    [2048]
  • blk.13.ffn_down.weight
    Q6_K
    [8192, 2048]
  • blk.13.ffn_gate.weight
    Q4_K
    [2048, 8192]
  • blk.13.ffn_norm.weight
    F32
    [2048]
  • blk.13.ffn_up.weight
    Q4_K
    [2048, 8192]
  • blk.13.shortconv.conv.weight
    F32
    [3, 2048]
  • blk.13.shortconv.in_proj.weight
    Q4_K
    [2048, 6144]
  • blk.13.shortconv.out_proj.weight
    Q4_K
    [2048, 2048]
    • blk.14
  • blk.14.attn_k.weight
    Q4_K
    [2048, 512]
  • blk.14.attn_k_norm.weight
    F32
    [64]
  • blk.14.attn_norm.weight
    F32
    [2048]
  • blk.14.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.14.attn_q.weight
    Q4_K
    [2048, 2048]
  • blk.14.attn_q_norm.weight
    F32
    [64]
  • blk.14.attn_v.weight
    Q6_K
    [2048, 512]
  • blk.14.ffn_down.weight
    Q6_K
    [8192, 2048]
  • blk.14.ffn_gate.weight
    Q4_K
    [2048, 8192]
  • blk.14.ffn_norm.weight
    F32
    [2048]
  • blk.14.ffn_up.weight
    Q4_K
    [2048, 8192]
    • blk.15
  • blk.15.attn_norm.weight
    F32
    [2048]
  • blk.15.ffn_down.weight
    Q6_K
    [8192, 2048]
  • blk.15.ffn_gate.weight
    Q4_K
    [2048, 8192]
  • blk.15.ffn_norm.weight
    F32
    [2048]
  • blk.15.ffn_up.weight
    Q4_K
    [2048, 8192]
  • blk.15.shortconv.conv.weight
    F32
    [3, 2048]
  • blk.15.shortconv.in_proj.weight
    Q4_K
    [2048, 6144]
  • blk.15.shortconv.out_proj.weight
    Q4_K
    [2048, 2048]
  • output_norm.weight
    F32
    [2048]