jetelain/ Gemma-4-31B:latest

172 3 months ago

Gemma-4-31B UD-Q4_K_XL MTP from Unsloth

vision tools thinking
... /
draft
5ae8b0117bed · 515MB
    Metadata
  • general.architecture
    gemma4-assistant
  • general.file_type
    Q8_0
  • gemma4-assistant.attention.head_count
    32
  • gemma4-assistant.attention.head_count_kv
    [16, 16, 16, 4]
  • gemma4-assistant.attention.key_length
    512
  • gemma4-assistant.attention.key_length_swa
    256
  • gemma4-assistant.attention.layer_norm_rms_epsilon
    1e-06
  • gemma4-assistant.attention.shared_kv_layers
    4
  • gemma4-assistant.attention.sliding_window
    1024
  • gemma4-assistant.attention.sliding_window_pattern
    [true, true, true, false]
  • gemma4-assistant.attention.value_length
    512
  • gemma4-assistant.attention.value_length_swa
    256
  • gemma4-assistant.block_count
    4
  • gemma4-assistant.context_length
    262144
  • gemma4-assistant.embedding_length
    1024
  • gemma4-assistant.embedding_length_out
    5376
  • gemma4-assistant.embedding_length_per_layer_input
    0
  • gemma4-assistant.feed_forward_length
    8192
  • gemma4-assistant.nextn_predict_layers
    4
  • gemma4-assistant.rope.dimension_count
    512
  • gemma4-assistant.rope.dimension_count_swa
    256
  • gemma4-assistant.rope.freq_base
    1e+06
  • gemma4-assistant.rope.freq_base_swa
    10000
  • tokenizer.ggml.add_bos_token
    true
  • tokenizer.ggml.add_space_prefix
    false
  • tokenizer.ggml.bos_token_id
    2
  • tokenizer.ggml.eos_token_id
    1
  • tokenizer.ggml.mask_token_id
    4
  • tokenizer.ggml.merges
    [ , ▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁ ▁, , , ▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁ ▁▁, ...]
  • tokenizer.ggml.model
    gemma4
  • tokenizer.ggml.padding_token_id
    0
  • tokenizer.ggml.scores
    [-1000, -1000, -1000, -1000, -1000, ...]
  • tokenizer.ggml.token_type
    [3, 3, 3, 3, 3, ...]
  • tokenizer.ggml.tokens
    [<pad>, <eos>, <bos>, <unk>, <mask>, ...]
  • tokenizer.ggml.unknown_token_id
    3
    • Tensor
  • token_embd.weight
    Q8_0
    [1024, 262144]
    • blk.0
  • blk.0.attn_norm.weight
    F32
    [1024]
  • blk.0.attn_output.weight
    Q8_0
    [8192, 1024]
  • blk.0.attn_q.weight
    Q8_0
    [1024, 8192]
  • blk.0.attn_q_norm.weight
    F32
    [256]
  • blk.0.ffn_down.weight
    Q8_0
    [8192, 1024]
  • blk.0.ffn_gate.weight
    Q8_0
    [1024, 8192]
  • blk.0.ffn_norm.weight
    F32
    [1024]
  • blk.0.ffn_up.weight
    Q8_0
    [1024, 8192]
  • blk.0.layer_output_scale.weight
    F32
    [1]
  • blk.0.post_attention_norm.weight
    F32
    [1024]
  • blk.0.post_ffw_norm.weight
    F32
    [1024]
    • blk.1
  • blk.1.attn_norm.weight
    F32
    [1024]
  • blk.1.attn_output.weight
    Q8_0
    [8192, 1024]
  • blk.1.attn_q.weight
    Q8_0
    [1024, 8192]
  • blk.1.attn_q_norm.weight
    F32
    [256]
  • blk.1.ffn_down.weight
    Q8_0
    [8192, 1024]
  • blk.1.ffn_gate.weight
    Q8_0
    [1024, 8192]
  • blk.1.ffn_norm.weight
    F32
    [1024]
  • blk.1.ffn_up.weight
    Q8_0
    [1024, 8192]
  • blk.1.layer_output_scale.weight
    F32
    [1]
  • blk.1.post_attention_norm.weight
    F32
    [1024]
  • blk.1.post_ffw_norm.weight
    F32
    [1024]
    • blk.2
  • blk.2.attn_norm.weight
    F32
    [1024]
  • blk.2.attn_output.weight
    Q8_0
    [8192, 1024]
  • blk.2.attn_q.weight
    Q8_0
    [1024, 8192]
  • blk.2.attn_q_norm.weight
    F32
    [256]
  • blk.2.ffn_down.weight
    Q8_0
    [8192, 1024]
  • blk.2.ffn_gate.weight
    Q8_0
    [1024, 8192]
  • blk.2.ffn_norm.weight
    F32
    [1024]
  • blk.2.ffn_up.weight
    Q8_0
    [1024, 8192]
  • blk.2.layer_output_scale.weight
    F32
    [1]
  • blk.2.post_attention_norm.weight
    F32
    [1024]
  • blk.2.post_ffw_norm.weight
    F32
    [1024]
    • blk.3
  • blk.3.attn_norm.weight
    F32
    [1024]
  • blk.3.attn_output.weight
    Q8_0
    [16384, 1024]
  • blk.3.attn_q.weight
    Q8_0
    [1024, 16384]
  • blk.3.attn_q_norm.weight
    F32
    [512]
  • blk.3.ffn_down.weight
    Q8_0
    [8192, 1024]
  • blk.3.ffn_gate.weight
    Q8_0
    [1024, 8192]
  • blk.3.ffn_norm.weight
    F32
    [1024]
  • blk.3.ffn_up.weight
    Q8_0
    [1024, 8192]
  • blk.3.layer_output_scale.weight
    F32
    [1]
  • blk.3.post_attention_norm.weight
    F32
    [1024]
  • blk.3.post_ffw_norm.weight
    F32
    [1024]
  • nextn.post_projection.weight
    Q8_0
    [1024, 5376]
  • nextn.pre_projection.weight
    Q8_0
    [10752, 1024]
  • rope_freqs.weight
    F32
    [256]
  • output_norm.weight
    F32
    [1024]