aiasistentworld/
Kimi-VL-A3B-Thinking-2506-LLM:latest

330 3 weeks ago

This is an improved version of Kimi-VL-A3B-Thinking. Please consider using this updated model instead of the previous version.

... /
model
72253d82d21c · 11GB
    Metadata
  • general.architecture
    deepseek2
  • general.file_type
    Q4_K_M
  • deepseek2.attention.head_count
    16
  • deepseek2.attention.head_count_kv
    1
  • deepseek2.attention.key_length
    576
  • deepseek2.attention.key_length_mla
    192
  • deepseek2.attention.kv_lora_rank
    512
  • deepseek2.attention.layer_norm_rms_epsilon
    1e-05
  • deepseek2.attention.value_length
    512
  • deepseek2.attention.value_length_mla
    128
  • deepseek2.block_count
    27
  • deepseek2.context_length
    131072
  • deepseek2.embedding_length
    2048
  • deepseek2.expert_count
    64
  • deepseek2.expert_feed_forward_length
    1408
  • deepseek2.expert_gating_func
    2
  • deepseek2.expert_shared_count
    2
  • deepseek2.expert_used_count
    6
  • deepseek2.expert_weights_norm
    true
  • deepseek2.expert_weights_scale
    2.446
  • deepseek2.feed_forward_length
    11264
  • deepseek2.leading_dense_block_count
    1
  • deepseek2.rope.dimension_count
    64
  • deepseek2.rope.freq_base
    800000
  • deepseek2.vocab_size
    163840
  • tokenizer.ggml.bos_token_id
    163584
  • tokenizer.ggml.eos_token_id
    163585
  • tokenizer.ggml.merges
    [Ġ Ġ, ĠĠ ĠĠ, Ġ t, i n, ä ¸, ...]
  • tokenizer.ggml.model
    gpt2
  • tokenizer.ggml.padding_token_id
    163839
  • tokenizer.ggml.pre
    kimi-k2
  • tokenizer.ggml.token_type
    [1, 1, 1, 1, 1, ...]
  • tokenizer.ggml.tokens
    [!, ", #, $, %, ...]
    • Tensor
  • token_embd.weight
    Q4_K
    [2048, 163840]
    • blk.0
  • blk.0.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.0.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.0.attn_kv_a_norm.weight
    F32
    [512]
  • blk.0.attn_norm.weight
    F32
    [2048]
  • blk.0.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.0.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.0.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.0.ffn_down.weight
    Q6_K
    [11264, 2048]
  • blk.0.ffn_gate.weight
    Q4_K
    [2048, 11264]
  • blk.0.ffn_norm.weight
    F32
    [2048]
  • blk.0.ffn_up.weight
    Q4_K
    [2048, 11264]
    • blk.1
  • blk.1.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.1.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.1.attn_kv_a_norm.weight
    F32
    [512]
  • blk.1.attn_norm.weight
    F32
    [2048]
  • blk.1.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.1.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.1.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.1.exp_probs_b.bias
    F32
    [64]
  • blk.1.ffn_down_exps.weight
    Q8_0
    [1408, 2048, 64]
  • blk.1.ffn_down_shexp.weight
    Q6_K
    [2816, 2048]
  • blk.1.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.1.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.1.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.1.ffn_norm.weight
    F32
    [2048]
  • blk.1.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.1.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.2
  • blk.2.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.2.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.2.attn_kv_a_norm.weight
    F32
    [512]
  • blk.2.attn_norm.weight
    F32
    [2048]
  • blk.2.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.2.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.2.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.2.exp_probs_b.bias
    F32
    [64]
  • blk.2.ffn_down_exps.weight
    Q8_0
    [1408, 2048, 64]
  • blk.2.ffn_down_shexp.weight
    Q6_K
    [2816, 2048]
  • blk.2.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.2.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.2.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.2.ffn_norm.weight
    F32
    [2048]
  • blk.2.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.2.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.3
  • blk.3.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.3.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.3.attn_kv_a_norm.weight
    F32
    [512]
  • blk.3.attn_norm.weight
    F32
    [2048]
  • blk.3.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.3.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.3.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.3.exp_probs_b.bias
    F32
    [64]
  • blk.3.ffn_down_exps.weight
    Q5_0
    [1408, 2048, 64]
  • blk.3.ffn_down_shexp.weight
    Q4_K
    [2816, 2048]
  • blk.3.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.3.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.3.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.3.ffn_norm.weight
    F32
    [2048]
  • blk.3.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.3.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.4
  • blk.4.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.4.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.4.attn_kv_a_norm.weight
    F32
    [512]
  • blk.4.attn_norm.weight
    F32
    [2048]
  • blk.4.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.4.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.4.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.4.exp_probs_b.bias
    F32
    [64]
  • blk.4.ffn_down_exps.weight
    Q5_0
    [1408, 2048, 64]
  • blk.4.ffn_down_shexp.weight
    Q4_K
    [2816, 2048]
  • blk.4.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.4.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.4.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.4.ffn_norm.weight
    F32
    [2048]
  • blk.4.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.4.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.5
  • blk.5.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.5.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.5.attn_kv_a_norm.weight
    F32
    [512]
  • blk.5.attn_norm.weight
    F32
    [2048]
  • blk.5.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.5.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.5.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.5.exp_probs_b.bias
    F32
    [64]
  • blk.5.ffn_down_exps.weight
    Q8_0
    [1408, 2048, 64]
  • blk.5.ffn_down_shexp.weight
    Q6_K
    [2816, 2048]
  • blk.5.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.5.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.5.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.5.ffn_norm.weight
    F32
    [2048]
  • blk.5.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.5.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.6
  • blk.6.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.6.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.6.attn_kv_a_norm.weight
    F32
    [512]
  • blk.6.attn_norm.weight
    F32
    [2048]
  • blk.6.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.6.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.6.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.6.exp_probs_b.bias
    F32
    [64]
  • blk.6.ffn_down_exps.weight
    Q5_0
    [1408, 2048, 64]
  • blk.6.ffn_down_shexp.weight
    Q4_K
    [2816, 2048]
  • blk.6.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.6.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.6.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.6.ffn_norm.weight
    F32
    [2048]
  • blk.6.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.6.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.7
  • blk.7.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.7.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.7.attn_kv_a_norm.weight
    F32
    [512]
  • blk.7.attn_norm.weight
    F32
    [2048]
  • blk.7.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.7.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.7.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.7.exp_probs_b.bias
    F32
    [64]
  • blk.7.ffn_down_exps.weight
    Q5_0
    [1408, 2048, 64]
  • blk.7.ffn_down_shexp.weight
    Q4_K
    [2816, 2048]
  • blk.7.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.7.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.7.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.7.ffn_norm.weight
    F32
    [2048]
  • blk.7.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.7.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.8
  • blk.8.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.8.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.8.attn_kv_a_norm.weight
    F32
    [512]
  • blk.8.attn_norm.weight
    F32
    [2048]
  • blk.8.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.8.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.8.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.8.exp_probs_b.bias
    F32
    [64]
  • blk.8.ffn_down_exps.weight
    Q8_0
    [1408, 2048, 64]
  • blk.8.ffn_down_shexp.weight
    Q6_K
    [2816, 2048]
  • blk.8.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.8.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.8.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.8.ffn_norm.weight
    F32
    [2048]
  • blk.8.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.8.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.9
  • blk.9.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.9.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.9.attn_kv_a_norm.weight
    F32
    [512]
  • blk.9.attn_norm.weight
    F32
    [2048]
  • blk.9.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.9.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.9.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.9.exp_probs_b.bias
    F32
    [64]
  • blk.9.ffn_down_exps.weight
    Q5_0
    [1408, 2048, 64]
  • blk.9.ffn_down_shexp.weight
    Q4_K
    [2816, 2048]
  • blk.9.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.9.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.9.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.9.ffn_norm.weight
    F32
    [2048]
  • blk.9.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.9.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.10
  • blk.10.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.10.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.10.attn_kv_a_norm.weight
    F32
    [512]
  • blk.10.attn_norm.weight
    F32
    [2048]
  • blk.10.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.10.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.10.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.10.exp_probs_b.bias
    F32
    [64]
  • blk.10.ffn_down_exps.weight
    Q5_0
    [1408, 2048, 64]
  • blk.10.ffn_down_shexp.weight
    Q4_K
    [2816, 2048]
  • blk.10.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.10.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.10.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.10.ffn_norm.weight
    F32
    [2048]
  • blk.10.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.10.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.11
  • blk.11.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.11.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.11.attn_kv_a_norm.weight
    F32
    [512]
  • blk.11.attn_norm.weight
    F32
    [2048]
  • blk.11.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.11.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.11.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.11.exp_probs_b.bias
    F32
    [64]
  • blk.11.ffn_down_exps.weight
    Q8_0
    [1408, 2048, 64]
  • blk.11.ffn_down_shexp.weight
    Q6_K
    [2816, 2048]
  • blk.11.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.11.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.11.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.11.ffn_norm.weight
    F32
    [2048]
  • blk.11.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.11.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.12
  • blk.12.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.12.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.12.attn_kv_a_norm.weight
    F32
    [512]
  • blk.12.attn_norm.weight
    F32
    [2048]
  • blk.12.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.12.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.12.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.12.exp_probs_b.bias
    F32
    [64]
  • blk.12.ffn_down_exps.weight
    Q5_0
    [1408, 2048, 64]
  • blk.12.ffn_down_shexp.weight
    Q4_K
    [2816, 2048]
  • blk.12.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.12.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.12.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.12.ffn_norm.weight
    F32
    [2048]
  • blk.12.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.12.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.13
  • blk.13.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.13.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.13.attn_kv_a_norm.weight
    F32
    [512]
  • blk.13.attn_norm.weight
    F32
    [2048]
  • blk.13.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.13.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.13.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.13.exp_probs_b.bias
    F32
    [64]
  • blk.13.ffn_down_exps.weight
    Q5_0
    [1408, 2048, 64]
  • blk.13.ffn_down_shexp.weight
    Q4_K
    [2816, 2048]
  • blk.13.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.13.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.13.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.13.ffn_norm.weight
    F32
    [2048]
  • blk.13.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.13.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.14
  • blk.14.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.14.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.14.attn_kv_a_norm.weight
    F32
    [512]
  • blk.14.attn_norm.weight
    F32
    [2048]
  • blk.14.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.14.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.14.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.14.exp_probs_b.bias
    F32
    [64]
  • blk.14.ffn_down_exps.weight
    Q8_0
    [1408, 2048, 64]
  • blk.14.ffn_down_shexp.weight
    Q6_K
    [2816, 2048]
  • blk.14.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.14.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.14.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.14.ffn_norm.weight
    F32
    [2048]
  • blk.14.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.14.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.15
  • blk.15.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.15.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.15.attn_kv_a_norm.weight
    F32
    [512]
  • blk.15.attn_norm.weight
    F32
    [2048]
  • blk.15.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.15.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.15.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.15.exp_probs_b.bias
    F32
    [64]
  • blk.15.ffn_down_exps.weight
    Q5_0
    [1408, 2048, 64]
  • blk.15.ffn_down_shexp.weight
    Q4_K
    [2816, 2048]
  • blk.15.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.15.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.15.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.15.ffn_norm.weight
    F32
    [2048]
  • blk.15.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.15.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.16
  • blk.16.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.16.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.16.attn_kv_a_norm.weight
    F32
    [512]
  • blk.16.attn_norm.weight
    F32
    [2048]
  • blk.16.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.16.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.16.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.16.exp_probs_b.bias
    F32
    [64]
  • blk.16.ffn_down_exps.weight
    Q5_0
    [1408, 2048, 64]
  • blk.16.ffn_down_shexp.weight
    Q4_K
    [2816, 2048]
  • blk.16.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.16.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.16.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.16.ffn_norm.weight
    F32
    [2048]
  • blk.16.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.16.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.17
  • blk.17.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.17.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.17.attn_kv_a_norm.weight
    F32
    [512]
  • blk.17.attn_norm.weight
    F32
    [2048]
  • blk.17.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.17.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.17.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.17.exp_probs_b.bias
    F32
    [64]
  • blk.17.ffn_down_exps.weight
    Q8_0
    [1408, 2048, 64]
  • blk.17.ffn_down_shexp.weight
    Q6_K
    [2816, 2048]
  • blk.17.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.17.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.17.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.17.ffn_norm.weight
    F32
    [2048]
  • blk.17.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.17.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.18
  • blk.18.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.18.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.18.attn_kv_a_norm.weight
    F32
    [512]
  • blk.18.attn_norm.weight
    F32
    [2048]
  • blk.18.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.18.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.18.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.18.exp_probs_b.bias
    F32
    [64]
  • blk.18.ffn_down_exps.weight
    Q5_0
    [1408, 2048, 64]
  • blk.18.ffn_down_shexp.weight
    Q4_K
    [2816, 2048]
  • blk.18.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.18.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.18.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.18.ffn_norm.weight
    F32
    [2048]
  • blk.18.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.18.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.19
  • blk.19.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.19.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.19.attn_kv_a_norm.weight
    F32
    [512]
  • blk.19.attn_norm.weight
    F32
    [2048]
  • blk.19.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.19.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.19.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.19.exp_probs_b.bias
    F32
    [64]
  • blk.19.ffn_down_exps.weight
    Q5_0
    [1408, 2048, 64]
  • blk.19.ffn_down_shexp.weight
    Q4_K
    [2816, 2048]
  • blk.19.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.19.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.19.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.19.ffn_norm.weight
    F32
    [2048]
  • blk.19.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.19.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.20
  • blk.20.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.20.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.20.attn_kv_a_norm.weight
    F32
    [512]
  • blk.20.attn_norm.weight
    F32
    [2048]
  • blk.20.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.20.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.20.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.20.exp_probs_b.bias
    F32
    [64]
  • blk.20.ffn_down_exps.weight
    Q8_0
    [1408, 2048, 64]
  • blk.20.ffn_down_shexp.weight
    Q6_K
    [2816, 2048]
  • blk.20.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.20.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.20.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.20.ffn_norm.weight
    F32
    [2048]
  • blk.20.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.20.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.21
  • blk.21.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.21.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.21.attn_kv_a_norm.weight
    F32
    [512]
  • blk.21.attn_norm.weight
    F32
    [2048]
  • blk.21.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.21.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.21.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.21.exp_probs_b.bias
    F32
    [64]
  • blk.21.ffn_down_exps.weight
    Q5_0
    [1408, 2048, 64]
  • blk.21.ffn_down_shexp.weight
    Q4_K
    [2816, 2048]
  • blk.21.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.21.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.21.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.21.ffn_norm.weight
    F32
    [2048]
  • blk.21.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.21.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.22
  • blk.22.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.22.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.22.attn_kv_a_norm.weight
    F32
    [512]
  • blk.22.attn_norm.weight
    F32
    [2048]
  • blk.22.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.22.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.22.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.22.exp_probs_b.bias
    F32
    [64]
  • blk.22.ffn_down_exps.weight
    Q5_0
    [1408, 2048, 64]
  • blk.22.ffn_down_shexp.weight
    Q4_K
    [2816, 2048]
  • blk.22.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.22.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.22.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.22.ffn_norm.weight
    F32
    [2048]
  • blk.22.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.22.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.23
  • blk.23.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.23.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.23.attn_kv_a_norm.weight
    F32
    [512]
  • blk.23.attn_norm.weight
    F32
    [2048]
  • blk.23.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.23.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.23.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.23.exp_probs_b.bias
    F32
    [64]
  • blk.23.ffn_down_exps.weight
    Q8_0
    [1408, 2048, 64]
  • blk.23.ffn_down_shexp.weight
    Q6_K
    [2816, 2048]
  • blk.23.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.23.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.23.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.23.ffn_norm.weight
    F32
    [2048]
  • blk.23.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.23.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.24
  • blk.24.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.24.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.24.attn_kv_a_norm.weight
    F32
    [512]
  • blk.24.attn_norm.weight
    F32
    [2048]
  • blk.24.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.24.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.24.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.24.exp_probs_b.bias
    F32
    [64]
  • blk.24.ffn_down_exps.weight
    Q8_0
    [1408, 2048, 64]
  • blk.24.ffn_down_shexp.weight
    Q6_K
    [2816, 2048]
  • blk.24.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.24.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.24.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.24.ffn_norm.weight
    F32
    [2048]
  • blk.24.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.24.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.25
  • blk.25.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.25.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.25.attn_kv_a_norm.weight
    F32
    [512]
  • blk.25.attn_norm.weight
    F32
    [2048]
  • blk.25.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.25.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.25.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.25.exp_probs_b.bias
    F32
    [64]
  • blk.25.ffn_down_exps.weight
    Q8_0
    [1408, 2048, 64]
  • blk.25.ffn_down_shexp.weight
    Q6_K
    [2816, 2048]
  • blk.25.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.25.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.25.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.25.ffn_norm.weight
    F32
    [2048]
  • blk.25.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.25.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
    • blk.26
  • blk.26.attn_k_b.weight
    Q5_0
    [128, 512, 16]
  • blk.26.attn_kv_a_mqa.weight
    Q4_K
    [2048, 576]
  • blk.26.attn_kv_a_norm.weight
    F32
    [512]
  • blk.26.attn_norm.weight
    F32
    [2048]
  • blk.26.attn_output.weight
    Q4_K
    [2048, 2048]
  • blk.26.attn_q.weight
    Q4_K
    [2048, 3072]
  • blk.26.attn_v_b.weight
    Q4_K
    [512, 128, 16]
  • blk.26.exp_probs_b.bias
    F32
    [64]
  • blk.26.ffn_down_exps.weight
    Q8_0
    [1408, 2048, 64]
  • blk.26.ffn_down_shexp.weight
    Q6_K
    [2816, 2048]
  • blk.26.ffn_gate_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.26.ffn_gate_inp.weight
    F32
    [2048, 64]
  • blk.26.ffn_gate_shexp.weight
    Q4_K
    [2048, 2816]
  • blk.26.ffn_norm.weight
    F32
    [2048]
  • blk.26.ffn_up_exps.weight
    Q4_K
    [2048, 1408, 64]
  • blk.26.ffn_up_shexp.weight
    Q4_K
    [2048, 2816]
  • output.weight
    Q6_K
    [2048, 163840]
  • output_norm.weight
    F32
    [2048]