cookieeatersixtynine/ gemma4-31b-dflash:latest

26 3 days ago

31b using dflash for gemma4 made to be faster and keep the full quality of normal models

vision tools thinking
... /
draft
ec9df1ab74e0 · 912MB
    Metadata
  • general.architecture
    dflash
  • general.file_type
    Q4_K_M
  • dflash.attention.head_count
    64
  • dflash.attention.head_count_kv
    8
  • dflash.attention.key_length
    128
  • dflash.attention.layer_norm_rms_epsilon
    1e-06
  • dflash.attention.sliding_window
    2048
  • dflash.attention.sliding_window_pattern
    [true, true, true, true, false]
  • dflash.attention.value_length
    128
  • dflash.block_count
    5
  • dflash.block_size
    16
  • dflash.context_length
    262144
  • dflash.embedding_length
    5376
  • dflash.feed_forward_length
    10752
  • dflash.rope.freq_base
    1e+06
  • dflash.target_layers
    [2, 13, 24, 36, 47, ...]
  • tokenizer.ggml.bos_token_id
    2
  • tokenizer.ggml.eos_token_id
    1
  • tokenizer.ggml.mask_token_id
    4
  • tokenizer.ggml.merges
    [ , ▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁ ▁, , , ▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁ ▁▁, ...]
  • tokenizer.ggml.model
    gpt2
  • tokenizer.ggml.padding_token_id
    0
  • tokenizer.ggml.pre
    granite-embed-multi-311m
  • tokenizer.ggml.token_type
    [3, 3, 3, 3, 3, ...]
  • tokenizer.ggml.tokens
    [<pad>, <eos>, <bos>, <unk>, <mask>, ...]
  • tokenizer.ggml.unknown_token_id
    3
    • Tensor
    blk.0
  • blk.0.attn_k.weight
    Q4_K
    [5376, 1024]
  • blk.0.attn_k_norm.weight
    F32
    [128]
  • blk.0.attn_norm.weight
    F32
    [5376]
  • blk.0.attn_output.weight
    Q4_K
    [8192, 5376]
  • blk.0.attn_q.weight
    Q4_K
    [5376, 8192]
  • blk.0.attn_q_norm.weight
    F32
    [128]
  • blk.0.attn_v.weight
    Q4_K
    [5376, 1024]
  • blk.0.ffn_down.weight
    Q4_K
    [10752, 5376]
  • blk.0.ffn_gate.weight
    Q4_K
    [5376, 10752]
  • blk.0.ffn_norm.weight
    F32
    [5376]
  • blk.0.ffn_up.weight
    Q4_K
    [5376, 10752]
    • blk.1
  • blk.1.attn_k.weight
    Q4_K
    [5376, 1024]
  • blk.1.attn_k_norm.weight
    F32
    [128]
  • blk.1.attn_norm.weight
    F32
    [5376]
  • blk.1.attn_output.weight
    Q4_K
    [8192, 5376]
  • blk.1.attn_q.weight
    Q4_K
    [5376, 8192]
  • blk.1.attn_q_norm.weight
    F32
    [128]
  • blk.1.attn_v.weight
    Q4_K
    [5376, 1024]
  • blk.1.ffn_down.weight
    Q4_K
    [10752, 5376]
  • blk.1.ffn_gate.weight
    Q4_K
    [5376, 10752]
  • blk.1.ffn_norm.weight
    F32
    [5376]
  • blk.1.ffn_up.weight
    Q4_K
    [5376, 10752]
    • blk.2
  • blk.2.attn_k.weight
    Q4_K
    [5376, 1024]
  • blk.2.attn_k_norm.weight
    F32
    [128]
  • blk.2.attn_norm.weight
    F32
    [5376]
  • blk.2.attn_output.weight
    Q4_K
    [8192, 5376]
  • blk.2.attn_q.weight
    Q4_K
    [5376, 8192]
  • blk.2.attn_q_norm.weight
    F32
    [128]
  • blk.2.attn_v.weight
    Q6_K
    [5376, 1024]
  • blk.2.ffn_down.weight
    Q6_K
    [10752, 5376]
  • blk.2.ffn_gate.weight
    Q4_K
    [5376, 10752]
  • blk.2.ffn_norm.weight
    F32
    [5376]
  • blk.2.ffn_up.weight
    Q4_K
    [5376, 10752]
    • blk.3
  • blk.3.attn_k.weight
    Q4_K
    [5376, 1024]
  • blk.3.attn_k_norm.weight
    F32
    [128]
  • blk.3.attn_norm.weight
    F32
    [5376]
  • blk.3.attn_output.weight
    Q4_K
    [8192, 5376]
  • blk.3.attn_q.weight
    Q4_K
    [5376, 8192]
  • blk.3.attn_q_norm.weight
    F32
    [128]
  • blk.3.attn_v.weight
    Q4_K
    [5376, 1024]
  • blk.3.ffn_down.weight
    Q4_K
    [10752, 5376]
  • blk.3.ffn_gate.weight
    Q4_K
    [5376, 10752]
  • blk.3.ffn_norm.weight
    F32
    [5376]
  • blk.3.ffn_up.weight
    Q4_K
    [5376, 10752]
    • blk.4
  • blk.4.attn_k.weight
    Q4_K
    [5376, 1024]
  • blk.4.attn_k_norm.weight
    F32
    [128]
  • blk.4.attn_norm.weight
    F32
    [5376]
  • blk.4.attn_output.weight
    Q4_K
    [8192, 5376]
  • blk.4.attn_q.weight
    Q4_K
    [5376, 8192]
  • blk.4.attn_q_norm.weight
    F32
    [128]
  • blk.4.attn_v.weight
    Q6_K
    [5376, 1024]
  • blk.4.ffn_down.weight
    Q6_K
    [10752, 5376]
  • blk.4.ffn_gate.weight
    Q4_K
    [5376, 10752]
  • blk.4.ffn_norm.weight
    F32
    [5376]
  • blk.4.ffn_up.weight
    Q4_K
    [5376, 10752]
  • enc.output_norm.weight
    F32
    [5376]
  • fc.weight
    Q4_K
    [32256, 5376]
  • output_norm.weight
    F32
    [5376]