fzkun/ minimind3-ascend:64m

19 1 month ago

基于昇腾910B纯国产环境，从零训练。

tools thinking 64m
... /
model
c30e26133edf · 128MB
    Metadata
  • general.architecture
    qwen3
  • general.file_type
    F16
  • qwen3.attention.head_count
    8
  • qwen3.attention.head_count_kv
    4
  • qwen3.attention.key_length
    96
  • qwen3.attention.layer_norm_rms_epsilon
    1e-06
  • qwen3.attention.value_length
    96
  • qwen3.block_count
    8
  • qwen3.context_length
    8192
  • qwen3.embedding_length
    768
  • qwen3.feed_forward_length
    2432
  • qwen3.rope.freq_base
    1e+06
  • tokenizer.ggml.add_bos_token
    false
  • tokenizer.ggml.add_eos_token
    false
  • tokenizer.ggml.bos_token_id
    1
  • tokenizer.ggml.eos_token_id
    2
  • tokenizer.ggml.merges
    [ï ¼, ä ¸, ï¼ Į, ç ļ, çļ Ħ, ...]
  • tokenizer.ggml.model
    gpt2
  • tokenizer.ggml.padding_token_id
    0
  • tokenizer.ggml.pre
    qwen2
  • tokenizer.ggml.token_type
    [3, 3, 3, 3, 3, ...]
  • tokenizer.ggml.tokens
    [<|endoftext|>, <|im_start|>, <|im_end|>, <|object_ref_start|>, <|object_ref_end|>, ...]
  • tokenizer.ggml.unknown_token_id
    0
    • Tensor
  • token_embd.weight
    F16
    [768, 6400]
    • blk.0
  • blk.0.attn_k.weight
    F16
    [768, 384]
  • blk.0.attn_k_norm.weight
    F32
    [96]
  • blk.0.attn_norm.weight
    F32
    [768]
  • blk.0.attn_output.weight
    F16
    [768, 768]
  • blk.0.attn_q.weight
    F16
    [768, 768]
  • blk.0.attn_q_norm.weight
    F32
    [96]
  • blk.0.attn_v.weight
    F16
    [768, 384]
  • blk.0.ffn_down.weight
    F16
    [2432, 768]
  • blk.0.ffn_gate.weight
    F16
    [768, 2432]
  • blk.0.ffn_norm.weight
    F32
    [768]
  • blk.0.ffn_up.weight
    F16
    [768, 2432]
    • blk.1
  • blk.1.attn_k.weight
    F16
    [768, 384]
  • blk.1.attn_k_norm.weight
    F32
    [96]
  • blk.1.attn_norm.weight
    F32
    [768]
  • blk.1.attn_output.weight
    F16
    [768, 768]
  • blk.1.attn_q.weight
    F16
    [768, 768]
  • blk.1.attn_q_norm.weight
    F32
    [96]
  • blk.1.attn_v.weight
    F16
    [768, 384]
  • blk.1.ffn_down.weight
    F16
    [2432, 768]
  • blk.1.ffn_gate.weight
    F16
    [768, 2432]
  • blk.1.ffn_norm.weight
    F32
    [768]
  • blk.1.ffn_up.weight
    F16
    [768, 2432]
    • blk.2
  • blk.2.attn_k.weight
    F16
    [768, 384]
  • blk.2.attn_k_norm.weight
    F32
    [96]
  • blk.2.attn_norm.weight
    F32
    [768]
  • blk.2.attn_output.weight
    F16
    [768, 768]
  • blk.2.attn_q.weight
    F16
    [768, 768]
  • blk.2.attn_q_norm.weight
    F32
    [96]
  • blk.2.attn_v.weight
    F16
    [768, 384]
  • blk.2.ffn_down.weight
    F16
    [2432, 768]
  • blk.2.ffn_gate.weight
    F16
    [768, 2432]
  • blk.2.ffn_norm.weight
    F32
    [768]
  • blk.2.ffn_up.weight
    F16
    [768, 2432]
    • blk.3
  • blk.3.attn_k.weight
    F16
    [768, 384]
  • blk.3.attn_k_norm.weight
    F32
    [96]
  • blk.3.attn_norm.weight
    F32
    [768]
  • blk.3.attn_output.weight
    F16
    [768, 768]
  • blk.3.attn_q.weight
    F16
    [768, 768]
  • blk.3.attn_q_norm.weight
    F32
    [96]
  • blk.3.attn_v.weight
    F16
    [768, 384]
  • blk.3.ffn_down.weight
    F16
    [2432, 768]
  • blk.3.ffn_gate.weight
    F16
    [768, 2432]
  • blk.3.ffn_norm.weight
    F32
    [768]
  • blk.3.ffn_up.weight
    F16
    [768, 2432]
    • blk.4
  • blk.4.attn_k.weight
    F16
    [768, 384]
  • blk.4.attn_k_norm.weight
    F32
    [96]
  • blk.4.attn_norm.weight
    F32
    [768]
  • blk.4.attn_output.weight
    F16
    [768, 768]
  • blk.4.attn_q.weight
    F16
    [768, 768]
  • blk.4.attn_q_norm.weight
    F32
    [96]
  • blk.4.attn_v.weight
    F16
    [768, 384]
  • blk.4.ffn_down.weight
    F16
    [2432, 768]
  • blk.4.ffn_gate.weight
    F16
    [768, 2432]
  • blk.4.ffn_norm.weight
    F32
    [768]
  • blk.4.ffn_up.weight
    F16
    [768, 2432]
    • blk.5
  • blk.5.attn_k.weight
    F16
    [768, 384]
  • blk.5.attn_k_norm.weight
    F32
    [96]
  • blk.5.attn_norm.weight
    F32
    [768]
  • blk.5.attn_output.weight
    F16
    [768, 768]
  • blk.5.attn_q.weight
    F16
    [768, 768]
  • blk.5.attn_q_norm.weight
    F32
    [96]
  • blk.5.attn_v.weight
    F16
    [768, 384]
  • blk.5.ffn_down.weight
    F16
    [2432, 768]
  • blk.5.ffn_gate.weight
    F16
    [768, 2432]
  • blk.5.ffn_norm.weight
    F32
    [768]
  • blk.5.ffn_up.weight
    F16
    [768, 2432]
    • blk.6
  • blk.6.attn_k.weight
    F16
    [768, 384]
  • blk.6.attn_k_norm.weight
    F32
    [96]
  • blk.6.attn_norm.weight
    F32
    [768]
  • blk.6.attn_output.weight
    F16
    [768, 768]
  • blk.6.attn_q.weight
    F16
    [768, 768]
  • blk.6.attn_q_norm.weight
    F32
    [96]
  • blk.6.attn_v.weight
    F16
    [768, 384]
  • blk.6.ffn_down.weight
    F16
    [2432, 768]
  • blk.6.ffn_gate.weight
    F16
    [768, 2432]
  • blk.6.ffn_norm.weight
    F32
    [768]
  • blk.6.ffn_up.weight
    F16
    [768, 2432]
    • blk.7
  • blk.7.attn_k.weight
    F16
    [768, 384]
  • blk.7.attn_k_norm.weight
    F32
    [96]
  • blk.7.attn_norm.weight
    F32
    [768]
  • blk.7.attn_output.weight
    F16
    [768, 768]
  • blk.7.attn_q.weight
    F16
    [768, 768]
  • blk.7.attn_q_norm.weight
    F32
    [96]
  • blk.7.attn_v.weight
    F16
    [768, 384]
  • blk.7.ffn_down.weight
    F16
    [2432, 768]
  • blk.7.ffn_gate.weight
    F16
    [768, 2432]
  • blk.7.ffn_norm.weight
    F32
    [768]
  • blk.7.ffn_up.weight
    F16
    [768, 2432]
  • output_norm.weight
    F32
    [768]