7 1 week ago

Low latency instruct LLM by JetBrains

tools
d34481305eaa · 7.0GB
    Metadata
  • general.architecture
    mellum
  • general.file_type
    (!unknown_file_type 38!)
  • mellum.attention.head_count
    32
  • mellum.attention.head_count_kv
    4
  • mellum.attention.key_length
    128
  • mellum.attention.layer_norm_rms_epsilon
    1e-06
  • mellum.attention.sliding_window
    1024
  • mellum.attention.sliding_window_pattern
    [true, true, true, false, true, ...]
  • mellum.attention.value_length
    128
  • mellum.block_count
    28
  • mellum.context_length
    131072
  • mellum.embedding_length
    2304
  • mellum.expert_count
    64
  • mellum.expert_feed_forward_length
    896
  • mellum.expert_used_count
    8
  • mellum.feed_forward_length
    7168
  • mellum.rope.freq_base
    500000
  • mellum.rope.freq_base_swa
    500000
  • mellum.rope.scaling.factor
    16
  • mellum.rope.scaling.original_context_length
    8192
  • mellum.rope.scaling.type
    yarn
  • mellum.rope.scaling.yarn_attn_factor
    1.2772589
  • mellum.rope.scaling.yarn_beta_fast
    32
  • mellum.rope.scaling.yarn_beta_slow
    1
  • tokenizer.ggml.bos_token_id
    0
  • tokenizer.ggml.eos_token_id
    28
  • tokenizer.ggml.merges
    [e r, o n, h e, r e, a n, ...]
  • tokenizer.ggml.model
    gpt2
  • tokenizer.ggml.padding_token_id
    0
  • tokenizer.ggml.pre
    mellum2
  • tokenizer.ggml.token_type
    [3, 3, 3, 3, 3, ...]
  • tokenizer.ggml.tokens
    [<|endoftext|>, <assistant>, </assistant>, <commit_before>, <commit_after>, ...]
  • tokenizer.ggml.unknown_token_id
    0
  • Tensor
  • token_embd.weight
    Q8_0
    [2304, 98304]
  • blk.0
  • blk.0.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.0.attn_k_norm.weight
    F32
    [128]
  • blk.0.attn_norm.weight
    F32
    [2304]
  • blk.0.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.0.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.0.attn_q_norm.weight
    F32
    [128]
  • blk.0.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.0.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.0.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.0.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.0.ffn_norm.weight
    F32
    [2304]
  • blk.0.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.1
  • blk.1.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.1.attn_k_norm.weight
    F32
    [128]
  • blk.1.attn_norm.weight
    F32
    [2304]
  • blk.1.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.1.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.1.attn_q_norm.weight
    F32
    [128]
  • blk.1.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.1.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.1.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.1.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.1.ffn_norm.weight
    F32
    [2304]
  • blk.1.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.2
  • blk.2.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.2.attn_k_norm.weight
    F32
    [128]
  • blk.2.attn_norm.weight
    F32
    [2304]
  • blk.2.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.2.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.2.attn_q_norm.weight
    F32
    [128]
  • blk.2.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.2.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.2.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.2.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.2.ffn_norm.weight
    F32
    [2304]
  • blk.2.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.3
  • blk.3.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.3.attn_k_norm.weight
    F32
    [128]
  • blk.3.attn_norm.weight
    F32
    [2304]
  • blk.3.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.3.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.3.attn_q_norm.weight
    F32
    [128]
  • blk.3.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.3.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.3.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.3.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.3.ffn_norm.weight
    F32
    [2304]
  • blk.3.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.4
  • blk.4.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.4.attn_k_norm.weight
    F32
    [128]
  • blk.4.attn_norm.weight
    F32
    [2304]
  • blk.4.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.4.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.4.attn_q_norm.weight
    F32
    [128]
  • blk.4.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.4.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.4.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.4.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.4.ffn_norm.weight
    F32
    [2304]
  • blk.4.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.5
  • blk.5.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.5.attn_k_norm.weight
    F32
    [128]
  • blk.5.attn_norm.weight
    F32
    [2304]
  • blk.5.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.5.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.5.attn_q_norm.weight
    F32
    [128]
  • blk.5.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.5.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.5.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.5.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.5.ffn_norm.weight
    F32
    [2304]
  • blk.5.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.6
  • blk.6.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.6.attn_k_norm.weight
    F32
    [128]
  • blk.6.attn_norm.weight
    F32
    [2304]
  • blk.6.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.6.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.6.attn_q_norm.weight
    F32
    [128]
  • blk.6.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.6.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.6.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.6.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.6.ffn_norm.weight
    F32
    [2304]
  • blk.6.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.7
  • blk.7.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.7.attn_k_norm.weight
    F32
    [128]
  • blk.7.attn_norm.weight
    F32
    [2304]
  • blk.7.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.7.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.7.attn_q_norm.weight
    F32
    [128]
  • blk.7.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.7.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.7.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.7.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.7.ffn_norm.weight
    F32
    [2304]
  • blk.7.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.8
  • blk.8.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.8.attn_k_norm.weight
    F32
    [128]
  • blk.8.attn_norm.weight
    F32
    [2304]
  • blk.8.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.8.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.8.attn_q_norm.weight
    F32
    [128]
  • blk.8.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.8.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.8.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.8.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.8.ffn_norm.weight
    F32
    [2304]
  • blk.8.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.9
  • blk.9.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.9.attn_k_norm.weight
    F32
    [128]
  • blk.9.attn_norm.weight
    F32
    [2304]
  • blk.9.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.9.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.9.attn_q_norm.weight
    F32
    [128]
  • blk.9.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.9.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.9.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.9.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.9.ffn_norm.weight
    F32
    [2304]
  • blk.9.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.10
  • blk.10.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.10.attn_k_norm.weight
    F32
    [128]
  • blk.10.attn_norm.weight
    F32
    [2304]
  • blk.10.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.10.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.10.attn_q_norm.weight
    F32
    [128]
  • blk.10.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.10.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.10.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.10.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.10.ffn_norm.weight
    F32
    [2304]
  • blk.10.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.11
  • blk.11.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.11.attn_k_norm.weight
    F32
    [128]
  • blk.11.attn_norm.weight
    F32
    [2304]
  • blk.11.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.11.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.11.attn_q_norm.weight
    F32
    [128]
  • blk.11.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.11.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.11.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.11.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.11.ffn_norm.weight
    F32
    [2304]
  • blk.11.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.12
  • blk.12.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.12.attn_k_norm.weight
    F32
    [128]
  • blk.12.attn_norm.weight
    F32
    [2304]
  • blk.12.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.12.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.12.attn_q_norm.weight
    F32
    [128]
  • blk.12.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.12.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.12.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.12.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.12.ffn_norm.weight
    F32
    [2304]
  • blk.12.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.13
  • blk.13.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.13.attn_k_norm.weight
    F32
    [128]
  • blk.13.attn_norm.weight
    F32
    [2304]
  • blk.13.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.13.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.13.attn_q_norm.weight
    F32
    [128]
  • blk.13.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.13.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.13.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.13.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.13.ffn_norm.weight
    F32
    [2304]
  • blk.13.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.14
  • blk.14.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.14.attn_k_norm.weight
    F32
    [128]
  • blk.14.attn_norm.weight
    F32
    [2304]
  • blk.14.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.14.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.14.attn_q_norm.weight
    F32
    [128]
  • blk.14.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.14.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.14.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.14.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.14.ffn_norm.weight
    F32
    [2304]
  • blk.14.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.15
  • blk.15.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.15.attn_k_norm.weight
    F32
    [128]
  • blk.15.attn_norm.weight
    F32
    [2304]
  • blk.15.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.15.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.15.attn_q_norm.weight
    F32
    [128]
  • blk.15.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.15.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.15.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.15.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.15.ffn_norm.weight
    F32
    [2304]
  • blk.15.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.16
  • blk.16.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.16.attn_k_norm.weight
    F32
    [128]
  • blk.16.attn_norm.weight
    F32
    [2304]
  • blk.16.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.16.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.16.attn_q_norm.weight
    F32
    [128]
  • blk.16.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.16.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.16.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.16.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.16.ffn_norm.weight
    F32
    [2304]
  • blk.16.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.17
  • blk.17.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.17.attn_k_norm.weight
    F32
    [128]
  • blk.17.attn_norm.weight
    F32
    [2304]
  • blk.17.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.17.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.17.attn_q_norm.weight
    F32
    [128]
  • blk.17.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.17.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.17.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.17.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.17.ffn_norm.weight
    F32
    [2304]
  • blk.17.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.18
  • blk.18.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.18.attn_k_norm.weight
    F32
    [128]
  • blk.18.attn_norm.weight
    F32
    [2304]
  • blk.18.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.18.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.18.attn_q_norm.weight
    F32
    [128]
  • blk.18.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.18.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.18.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.18.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.18.ffn_norm.weight
    F32
    [2304]
  • blk.18.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.19
  • blk.19.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.19.attn_k_norm.weight
    F32
    [128]
  • blk.19.attn_norm.weight
    F32
    [2304]
  • blk.19.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.19.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.19.attn_q_norm.weight
    F32
    [128]
  • blk.19.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.19.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.19.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.19.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.19.ffn_norm.weight
    F32
    [2304]
  • blk.19.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.20
  • blk.20.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.20.attn_k_norm.weight
    F32
    [128]
  • blk.20.attn_norm.weight
    F32
    [2304]
  • blk.20.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.20.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.20.attn_q_norm.weight
    F32
    [128]
  • blk.20.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.20.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.20.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.20.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.20.ffn_norm.weight
    F32
    [2304]
  • blk.20.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.21
  • blk.21.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.21.attn_k_norm.weight
    F32
    [128]
  • blk.21.attn_norm.weight
    F32
    [2304]
  • blk.21.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.21.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.21.attn_q_norm.weight
    F32
    [128]
  • blk.21.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.21.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.21.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.21.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.21.ffn_norm.weight
    F32
    [2304]
  • blk.21.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.22
  • blk.22.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.22.attn_k_norm.weight
    F32
    [128]
  • blk.22.attn_norm.weight
    F32
    [2304]
  • blk.22.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.22.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.22.attn_q_norm.weight
    F32
    [128]
  • blk.22.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.22.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.22.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.22.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.22.ffn_norm.weight
    F32
    [2304]
  • blk.22.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.23
  • blk.23.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.23.attn_k_norm.weight
    F32
    [128]
  • blk.23.attn_norm.weight
    F32
    [2304]
  • blk.23.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.23.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.23.attn_q_norm.weight
    F32
    [128]
  • blk.23.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.23.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.23.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.23.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.23.ffn_norm.weight
    F32
    [2304]
  • blk.23.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.24
  • blk.24.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.24.attn_k_norm.weight
    F32
    [128]
  • blk.24.attn_norm.weight
    F32
    [2304]
  • blk.24.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.24.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.24.attn_q_norm.weight
    F32
    [128]
  • blk.24.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.24.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.24.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.24.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.24.ffn_norm.weight
    F32
    [2304]
  • blk.24.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.25
  • blk.25.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.25.attn_k_norm.weight
    F32
    [128]
  • blk.25.attn_norm.weight
    F32
    [2304]
  • blk.25.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.25.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.25.attn_q_norm.weight
    F32
    [128]
  • blk.25.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.25.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.25.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.25.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.25.ffn_norm.weight
    F32
    [2304]
  • blk.25.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.26
  • blk.26.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.26.attn_k_norm.weight
    F32
    [128]
  • blk.26.attn_norm.weight
    F32
    [2304]
  • blk.26.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.26.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.26.attn_q_norm.weight
    F32
    [128]
  • blk.26.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.26.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.26.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.26.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.26.ffn_norm.weight
    F32
    [2304]
  • blk.26.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.27
  • blk.27.attn_k.weight
    Q8_0
    [2304, 512]
  • blk.27.attn_k_norm.weight
    F32
    [128]
  • blk.27.attn_norm.weight
    F32
    [2304]
  • blk.27.attn_output.weight
    Q8_0
    [4096, 2304]
  • blk.27.attn_q.weight
    Q8_0
    [2304, 4096]
  • blk.27.attn_q_norm.weight
    F32
    [128]
  • blk.27.attn_v.weight
    Q8_0
    [2304, 512]
  • blk.27.ffn_down_exps.weight
    MXFP4
    [896, 2304, 64]
  • blk.27.ffn_gate_exps.weight
    MXFP4
    [2304, 896, 64]
  • blk.27.ffn_gate_inp.weight
    F32
    [2304, 64]
  • blk.27.ffn_norm.weight
    F32
    [2304]
  • blk.27.ffn_up_exps.weight
    MXFP4
    [2304, 896, 64]
  • output.weight
    Q8_0
    [2304, 98304]
  • output_norm.weight
    F32
    [2304]