231 19 hours ago

Laguna XS.2 is a 33B total parameter Mixture-of-Experts model with 3B activated parameters per token designed for agentic coding and long-horizon work on a local machine.

tools thinking
f453c452e1be · 37GB
    Metadata
  • general.architecture
    laguna
  • general.file_type
    Q8_0
  • laguna.attention.gating_type
    1
  • laguna.attention.head_count
    [48, 64, 64, 64, 48, ...]
  • laguna.attention.head_count_kv
    8
  • laguna.attention.key_length
    128
  • laguna.attention.layer_norm_rms_epsilon
    1e-06
  • laguna.attention.layer_types
    [0, 1, 1, 1, 0, ...]
  • laguna.attention.qk_norm
    true
  • laguna.attention.sink_enabled
    false
  • laguna.attention.sliding_window
    512
  • laguna.attention.sliding_window_pattern
    [false, true, true, true, false, ...]
  • laguna.attention.value_length
    128
  • laguna.block_count
    40
  • laguna.context_length
    131072
  • laguna.decoder_sparse_step
    1
  • laguna.dense_layers
    [0]
  • laguna.embedding_length
    2048
  • laguna.expert_count
    256
  • laguna.expert_feed_forward_length
    512
  • laguna.expert_gating_func
    2
  • laguna.expert_shared_count
    1
  • laguna.expert_shared_feed_forward_length
    512
  • laguna.expert_used_count
    8
  • laguna.expert_weights_norm
    true
  • laguna.expert_weights_scale
    2.5
  • laguna.feed_forward_length
    8192
  • laguna.leading_dense_block_count
    1
  • laguna.rope.dimension_count
    64
  • laguna.rope.freq_base
    500000
  • laguna.rope.partial_rotary_factor
    0.5
  • laguna.rope.scaling.attn_factor
    1
  • laguna.rope.scaling.beta_fast
    64
  • laguna.rope.scaling.beta_slow
    1
  • laguna.rope.scaling.factor
    32
  • laguna.rope.scaling.original_context_length
    4096
  • laguna.rope.scaling.type
    yarn
  • laguna.rope.swa.dimension_count
    128
  • laguna.rope.swa.freq_base
    10000
  • laguna.rope.swa.partial_rotary_factor
    1
  • laguna.rope.swa.scaling.factor
    1
  • laguna.rope.swa.scaling.type
    default
  • tokenizer.ggml.add_bos_token
    false
  • tokenizer.ggml.add_eos_token
    false
  • tokenizer.ggml.add_padding_token
    false
  • tokenizer.ggml.add_unknown_token
    false
  • tokenizer.ggml.bos_token_id
    2
  • tokenizer.ggml.eos_token_id
    2
  • tokenizer.ggml.eos_token_ids
    [2, 24]
  • tokenizer.ggml.merges
    [i n, Ġ t, Ġ Ġ, e r, Ġ a, ...]
  • tokenizer.ggml.model
    gpt2
  • tokenizer.ggml.padding_token_id
    9
  • tokenizer.ggml.pre
    laguna
  • tokenizer.ggml.scores
    [0, 1, 2, 3, 4, ...]
  • tokenizer.ggml.token_type
    [3, 3, 3, 3, 3, ...]
  • tokenizer.ggml.tokens
    [〈|UNK|〉, 〈|CODE_START|〉, 〈|EOS|〉, 〈|CODE_END|〉, 〈|META_START|〉, ...]
  • tokenizer.ggml.unknown_token_id
    0
  • Tensor
  • token_embd.weight
    BF16
    [2048, 100352]
  • blk.0
  • blk.0.attn_g.weight
    BF16
    [2048, 48]
  • blk.0.attn_k.weight
    BF16
    [2048, 1024]
  • blk.0.attn_k_norm.weight
    F32
    [128]
  • blk.0.attn_norm.weight
    F32
    [2048]
  • blk.0.attn_output.weight
    BF16
    [6144, 2048]
  • blk.0.attn_q.weight
    BF16
    [2048, 6144]
  • blk.0.attn_q_norm.weight
    F32
    [128]
  • blk.0.attn_v.weight
    BF16
    [2048, 1024]
  • blk.0.ffn_down.weight
    BF16
    [8192, 2048]
  • blk.0.ffn_gate.weight
    BF16
    [2048, 8192]
  • blk.0.ffn_norm.weight
    F32
    [2048]
  • blk.0.ffn_up.weight
    BF16
    [2048, 8192]
  • blk.1
  • blk.1.attn_g.weight
    BF16
    [2048, 64]
  • blk.1.attn_k.weight
    BF16
    [2048, 1024]
  • blk.1.attn_k_norm.weight
    F32
    [128]
  • blk.1.attn_norm.weight
    F32
    [2048]
  • blk.1.attn_output.weight
    BF16
    [8192, 2048]
  • blk.1.attn_q.weight
    BF16
    [2048, 8192]
  • blk.1.attn_q_norm.weight
    F32
    [128]
  • blk.1.attn_v.weight
    BF16
    [2048, 1024]
  • blk.1.exp_probs_b.bias
    F32
    [256]
  • blk.1.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.1.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.1.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.1.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.1.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.1.ffn_norm.weight
    F32
    [2048]
  • blk.1.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.1.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.2
  • blk.2.attn_g.weight
    BF16
    [2048, 64]
  • blk.2.attn_k.weight
    BF16
    [2048, 1024]
  • blk.2.attn_k_norm.weight
    F32
    [128]
  • blk.2.attn_norm.weight
    F32
    [2048]
  • blk.2.attn_output.weight
    BF16
    [8192, 2048]
  • blk.2.attn_q.weight
    BF16
    [2048, 8192]
  • blk.2.attn_q_norm.weight
    F32
    [128]
  • blk.2.attn_v.weight
    BF16
    [2048, 1024]
  • blk.2.exp_probs_b.bias
    F32
    [256]
  • blk.2.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.2.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.2.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.2.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.2.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.2.ffn_norm.weight
    F32
    [2048]
  • blk.2.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.2.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.3
  • blk.3.attn_g.weight
    BF16
    [2048, 64]
  • blk.3.attn_k.weight
    BF16
    [2048, 1024]
  • blk.3.attn_k_norm.weight
    F32
    [128]
  • blk.3.attn_norm.weight
    F32
    [2048]
  • blk.3.attn_output.weight
    BF16
    [8192, 2048]
  • blk.3.attn_q.weight
    BF16
    [2048, 8192]
  • blk.3.attn_q_norm.weight
    F32
    [128]
  • blk.3.attn_v.weight
    BF16
    [2048, 1024]
  • blk.3.exp_probs_b.bias
    F32
    [256]
  • blk.3.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.3.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.3.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.3.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.3.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.3.ffn_norm.weight
    F32
    [2048]
  • blk.3.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.3.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.4
  • blk.4.attn_g.weight
    BF16
    [2048, 48]
  • blk.4.attn_k.weight
    BF16
    [2048, 1024]
  • blk.4.attn_k_norm.weight
    F32
    [128]
  • blk.4.attn_norm.weight
    F32
    [2048]
  • blk.4.attn_output.weight
    BF16
    [6144, 2048]
  • blk.4.attn_q.weight
    BF16
    [2048, 6144]
  • blk.4.attn_q_norm.weight
    F32
    [128]
  • blk.4.attn_v.weight
    BF16
    [2048, 1024]
  • blk.4.exp_probs_b.bias
    F32
    [256]
  • blk.4.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.4.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.4.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.4.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.4.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.4.ffn_norm.weight
    F32
    [2048]
  • blk.4.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.4.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.5
  • blk.5.attn_g.weight
    BF16
    [2048, 64]
  • blk.5.attn_k.weight
    BF16
    [2048, 1024]
  • blk.5.attn_k_norm.weight
    F32
    [128]
  • blk.5.attn_norm.weight
    F32
    [2048]
  • blk.5.attn_output.weight
    BF16
    [8192, 2048]
  • blk.5.attn_q.weight
    BF16
    [2048, 8192]
  • blk.5.attn_q_norm.weight
    F32
    [128]
  • blk.5.attn_v.weight
    BF16
    [2048, 1024]
  • blk.5.exp_probs_b.bias
    F32
    [256]
  • blk.5.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.5.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.5.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.5.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.5.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.5.ffn_norm.weight
    F32
    [2048]
  • blk.5.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.5.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.6
  • blk.6.attn_g.weight
    BF16
    [2048, 64]
  • blk.6.attn_k.weight
    BF16
    [2048, 1024]
  • blk.6.attn_k_norm.weight
    F32
    [128]
  • blk.6.attn_norm.weight
    F32
    [2048]
  • blk.6.attn_output.weight
    BF16
    [8192, 2048]
  • blk.6.attn_q.weight
    BF16
    [2048, 8192]
  • blk.6.attn_q_norm.weight
    F32
    [128]
  • blk.6.attn_v.weight
    BF16
    [2048, 1024]
  • blk.6.exp_probs_b.bias
    F32
    [256]
  • blk.6.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.6.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.6.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.6.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.6.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.6.ffn_norm.weight
    F32
    [2048]
  • blk.6.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.6.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.7
  • blk.7.attn_g.weight
    BF16
    [2048, 64]
  • blk.7.attn_k.weight
    BF16
    [2048, 1024]
  • blk.7.attn_k_norm.weight
    F32
    [128]
  • blk.7.attn_norm.weight
    F32
    [2048]
  • blk.7.attn_output.weight
    BF16
    [8192, 2048]
  • blk.7.attn_q.weight
    BF16
    [2048, 8192]
  • blk.7.attn_q_norm.weight
    F32
    [128]
  • blk.7.attn_v.weight
    BF16
    [2048, 1024]
  • blk.7.exp_probs_b.bias
    F32
    [256]
  • blk.7.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.7.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.7.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.7.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.7.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.7.ffn_norm.weight
    F32
    [2048]
  • blk.7.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.7.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.8
  • blk.8.attn_g.weight
    BF16
    [2048, 48]
  • blk.8.attn_k.weight
    BF16
    [2048, 1024]
  • blk.8.attn_k_norm.weight
    F32
    [128]
  • blk.8.attn_norm.weight
    F32
    [2048]
  • blk.8.attn_output.weight
    BF16
    [6144, 2048]
  • blk.8.attn_q.weight
    BF16
    [2048, 6144]
  • blk.8.attn_q_norm.weight
    F32
    [128]
  • blk.8.attn_v.weight
    BF16
    [2048, 1024]
  • blk.8.exp_probs_b.bias
    F32
    [256]
  • blk.8.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.8.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.8.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.8.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.8.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.8.ffn_norm.weight
    F32
    [2048]
  • blk.8.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.8.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.9
  • blk.9.attn_g.weight
    BF16
    [2048, 64]
  • blk.9.attn_k.weight
    BF16
    [2048, 1024]
  • blk.9.attn_k_norm.weight
    F32
    [128]
  • blk.9.attn_norm.weight
    F32
    [2048]
  • blk.9.attn_output.weight
    BF16
    [8192, 2048]
  • blk.9.attn_q.weight
    BF16
    [2048, 8192]
  • blk.9.attn_q_norm.weight
    F32
    [128]
  • blk.9.attn_v.weight
    BF16
    [2048, 1024]
  • blk.9.exp_probs_b.bias
    F32
    [256]
  • blk.9.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.9.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.9.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.9.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.9.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.9.ffn_norm.weight
    F32
    [2048]
  • blk.9.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.9.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.10
  • blk.10.attn_g.weight
    BF16
    [2048, 64]
  • blk.10.attn_k.weight
    BF16
    [2048, 1024]
  • blk.10.attn_k_norm.weight
    F32
    [128]
  • blk.10.attn_norm.weight
    F32
    [2048]
  • blk.10.attn_output.weight
    BF16
    [8192, 2048]
  • blk.10.attn_q.weight
    BF16
    [2048, 8192]
  • blk.10.attn_q_norm.weight
    F32
    [128]
  • blk.10.attn_v.weight
    BF16
    [2048, 1024]
  • blk.10.exp_probs_b.bias
    F32
    [256]
  • blk.10.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.10.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.10.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.10.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.10.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.10.ffn_norm.weight
    F32
    [2048]
  • blk.10.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.10.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.11
  • blk.11.attn_g.weight
    BF16
    [2048, 64]
  • blk.11.attn_k.weight
    BF16
    [2048, 1024]
  • blk.11.attn_k_norm.weight
    F32
    [128]
  • blk.11.attn_norm.weight
    F32
    [2048]
  • blk.11.attn_output.weight
    BF16
    [8192, 2048]
  • blk.11.attn_q.weight
    BF16
    [2048, 8192]
  • blk.11.attn_q_norm.weight
    F32
    [128]
  • blk.11.attn_v.weight
    BF16
    [2048, 1024]
  • blk.11.exp_probs_b.bias
    F32
    [256]
  • blk.11.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.11.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.11.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.11.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.11.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.11.ffn_norm.weight
    F32
    [2048]
  • blk.11.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.11.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.12
  • blk.12.attn_g.weight
    BF16
    [2048, 48]
  • blk.12.attn_k.weight
    BF16
    [2048, 1024]
  • blk.12.attn_k_norm.weight
    F32
    [128]
  • blk.12.attn_norm.weight
    F32
    [2048]
  • blk.12.attn_output.weight
    BF16
    [6144, 2048]
  • blk.12.attn_q.weight
    BF16
    [2048, 6144]
  • blk.12.attn_q_norm.weight
    F32
    [128]
  • blk.12.attn_v.weight
    BF16
    [2048, 1024]
  • blk.12.exp_probs_b.bias
    F32
    [256]
  • blk.12.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.12.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.12.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.12.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.12.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.12.ffn_norm.weight
    F32
    [2048]
  • blk.12.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.12.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.13
  • blk.13.attn_g.weight
    BF16
    [2048, 64]
  • blk.13.attn_k.weight
    BF16
    [2048, 1024]
  • blk.13.attn_k_norm.weight
    F32
    [128]
  • blk.13.attn_norm.weight
    F32
    [2048]
  • blk.13.attn_output.weight
    BF16
    [8192, 2048]
  • blk.13.attn_q.weight
    BF16
    [2048, 8192]
  • blk.13.attn_q_norm.weight
    F32
    [128]
  • blk.13.attn_v.weight
    BF16
    [2048, 1024]
  • blk.13.exp_probs_b.bias
    F32
    [256]
  • blk.13.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.13.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.13.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.13.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.13.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.13.ffn_norm.weight
    F32
    [2048]
  • blk.13.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.13.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.14
  • blk.14.attn_g.weight
    BF16
    [2048, 64]
  • blk.14.attn_k.weight
    BF16
    [2048, 1024]
  • blk.14.attn_k_norm.weight
    F32
    [128]
  • blk.14.attn_norm.weight
    F32
    [2048]
  • blk.14.attn_output.weight
    BF16
    [8192, 2048]
  • blk.14.attn_q.weight
    BF16
    [2048, 8192]
  • blk.14.attn_q_norm.weight
    F32
    [128]
  • blk.14.attn_v.weight
    BF16
    [2048, 1024]
  • blk.14.exp_probs_b.bias
    F32
    [256]
  • blk.14.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.14.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.14.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.14.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.14.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.14.ffn_norm.weight
    F32
    [2048]
  • blk.14.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.14.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.15
  • blk.15.attn_g.weight
    BF16
    [2048, 64]
  • blk.15.attn_k.weight
    BF16
    [2048, 1024]
  • blk.15.attn_k_norm.weight
    F32
    [128]
  • blk.15.attn_norm.weight
    F32
    [2048]
  • blk.15.attn_output.weight
    BF16
    [8192, 2048]
  • blk.15.attn_q.weight
    BF16
    [2048, 8192]
  • blk.15.attn_q_norm.weight
    F32
    [128]
  • blk.15.attn_v.weight
    BF16
    [2048, 1024]
  • blk.15.exp_probs_b.bias
    F32
    [256]
  • blk.15.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.15.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.15.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.15.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.15.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.15.ffn_norm.weight
    F32
    [2048]
  • blk.15.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.15.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.16
  • blk.16.attn_g.weight
    BF16
    [2048, 48]
  • blk.16.attn_k.weight
    BF16
    [2048, 1024]
  • blk.16.attn_k_norm.weight
    F32
    [128]
  • blk.16.attn_norm.weight
    F32
    [2048]
  • blk.16.attn_output.weight
    BF16
    [6144, 2048]
  • blk.16.attn_q.weight
    BF16
    [2048, 6144]
  • blk.16.attn_q_norm.weight
    F32
    [128]
  • blk.16.attn_v.weight
    BF16
    [2048, 1024]
  • blk.16.exp_probs_b.bias
    F32
    [256]
  • blk.16.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.16.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.16.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.16.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.16.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.16.ffn_norm.weight
    F32
    [2048]
  • blk.16.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.16.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.17
  • blk.17.attn_g.weight
    BF16
    [2048, 64]
  • blk.17.attn_k.weight
    BF16
    [2048, 1024]
  • blk.17.attn_k_norm.weight
    F32
    [128]
  • blk.17.attn_norm.weight
    F32
    [2048]
  • blk.17.attn_output.weight
    BF16
    [8192, 2048]
  • blk.17.attn_q.weight
    BF16
    [2048, 8192]
  • blk.17.attn_q_norm.weight
    F32
    [128]
  • blk.17.attn_v.weight
    BF16
    [2048, 1024]
  • blk.17.exp_probs_b.bias
    F32
    [256]
  • blk.17.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.17.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.17.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.17.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.17.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.17.ffn_norm.weight
    F32
    [2048]
  • blk.17.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.17.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.18
  • blk.18.attn_g.weight
    BF16
    [2048, 64]
  • blk.18.attn_k.weight
    BF16
    [2048, 1024]
  • blk.18.attn_k_norm.weight
    F32
    [128]
  • blk.18.attn_norm.weight
    F32
    [2048]
  • blk.18.attn_output.weight
    BF16
    [8192, 2048]
  • blk.18.attn_q.weight
    BF16
    [2048, 8192]
  • blk.18.attn_q_norm.weight
    F32
    [128]
  • blk.18.attn_v.weight
    BF16
    [2048, 1024]
  • blk.18.exp_probs_b.bias
    F32
    [256]
  • blk.18.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.18.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.18.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.18.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.18.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.18.ffn_norm.weight
    F32
    [2048]
  • blk.18.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.18.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.19
  • blk.19.attn_g.weight
    BF16
    [2048, 64]
  • blk.19.attn_k.weight
    BF16
    [2048, 1024]
  • blk.19.attn_k_norm.weight
    F32
    [128]
  • blk.19.attn_norm.weight
    F32
    [2048]
  • blk.19.attn_output.weight
    BF16
    [8192, 2048]
  • blk.19.attn_q.weight
    BF16
    [2048, 8192]
  • blk.19.attn_q_norm.weight
    F32
    [128]
  • blk.19.attn_v.weight
    BF16
    [2048, 1024]
  • blk.19.exp_probs_b.bias
    F32
    [256]
  • blk.19.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.19.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.19.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.19.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.19.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.19.ffn_norm.weight
    F32
    [2048]
  • blk.19.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.19.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.20
  • blk.20.attn_g.weight
    BF16
    [2048, 48]
  • blk.20.attn_k.weight
    BF16
    [2048, 1024]
  • blk.20.attn_k_norm.weight
    F32
    [128]
  • blk.20.attn_norm.weight
    F32
    [2048]
  • blk.20.attn_output.weight
    BF16
    [6144, 2048]
  • blk.20.attn_q.weight
    BF16
    [2048, 6144]
  • blk.20.attn_q_norm.weight
    F32
    [128]
  • blk.20.attn_v.weight
    BF16
    [2048, 1024]
  • blk.20.exp_probs_b.bias
    F32
    [256]
  • blk.20.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.20.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.20.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.20.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.20.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.20.ffn_norm.weight
    F32
    [2048]
  • blk.20.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.20.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.21
  • blk.21.attn_g.weight
    BF16
    [2048, 64]
  • blk.21.attn_k.weight
    BF16
    [2048, 1024]
  • blk.21.attn_k_norm.weight
    F32
    [128]
  • blk.21.attn_norm.weight
    F32
    [2048]
  • blk.21.attn_output.weight
    BF16
    [8192, 2048]
  • blk.21.attn_q.weight
    BF16
    [2048, 8192]
  • blk.21.attn_q_norm.weight
    F32
    [128]
  • blk.21.attn_v.weight
    BF16
    [2048, 1024]
  • blk.21.exp_probs_b.bias
    F32
    [256]
  • blk.21.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.21.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.21.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.21.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.21.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.21.ffn_norm.weight
    F32
    [2048]
  • blk.21.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.21.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.22
  • blk.22.attn_g.weight
    BF16
    [2048, 64]
  • blk.22.attn_k.weight
    BF16
    [2048, 1024]
  • blk.22.attn_k_norm.weight
    F32
    [128]
  • blk.22.attn_norm.weight
    F32
    [2048]
  • blk.22.attn_output.weight
    BF16
    [8192, 2048]
  • blk.22.attn_q.weight
    BF16
    [2048, 8192]
  • blk.22.attn_q_norm.weight
    F32
    [128]
  • blk.22.attn_v.weight
    BF16
    [2048, 1024]
  • blk.22.exp_probs_b.bias
    F32
    [256]
  • blk.22.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.22.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.22.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.22.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.22.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.22.ffn_norm.weight
    F32
    [2048]
  • blk.22.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.22.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.23
  • blk.23.attn_g.weight
    BF16
    [2048, 64]
  • blk.23.attn_k.weight
    BF16
    [2048, 1024]
  • blk.23.attn_k_norm.weight
    F32
    [128]
  • blk.23.attn_norm.weight
    F32
    [2048]
  • blk.23.attn_output.weight
    BF16
    [8192, 2048]
  • blk.23.attn_q.weight
    BF16
    [2048, 8192]
  • blk.23.attn_q_norm.weight
    F32
    [128]
  • blk.23.attn_v.weight
    BF16
    [2048, 1024]
  • blk.23.exp_probs_b.bias
    F32
    [256]
  • blk.23.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.23.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.23.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.23.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.23.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.23.ffn_norm.weight
    F32
    [2048]
  • blk.23.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.23.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.24
  • blk.24.attn_g.weight
    BF16
    [2048, 48]
  • blk.24.attn_k.weight
    BF16
    [2048, 1024]
  • blk.24.attn_k_norm.weight
    F32
    [128]
  • blk.24.attn_norm.weight
    F32
    [2048]
  • blk.24.attn_output.weight
    BF16
    [6144, 2048]
  • blk.24.attn_q.weight
    BF16
    [2048, 6144]
  • blk.24.attn_q_norm.weight
    F32
    [128]
  • blk.24.attn_v.weight
    BF16
    [2048, 1024]
  • blk.24.exp_probs_b.bias
    F32
    [256]
  • blk.24.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.24.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.24.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.24.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.24.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.24.ffn_norm.weight
    F32
    [2048]
  • blk.24.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.24.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.25
  • blk.25.attn_g.weight
    BF16
    [2048, 64]
  • blk.25.attn_k.weight
    BF16
    [2048, 1024]
  • blk.25.attn_k_norm.weight
    F32
    [128]
  • blk.25.attn_norm.weight
    F32
    [2048]
  • blk.25.attn_output.weight
    BF16
    [8192, 2048]
  • blk.25.attn_q.weight
    BF16
    [2048, 8192]
  • blk.25.attn_q_norm.weight
    F32
    [128]
  • blk.25.attn_v.weight
    BF16
    [2048, 1024]
  • blk.25.exp_probs_b.bias
    F32
    [256]
  • blk.25.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.25.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.25.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.25.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.25.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.25.ffn_norm.weight
    F32
    [2048]
  • blk.25.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.25.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.26
  • blk.26.attn_g.weight
    BF16
    [2048, 64]
  • blk.26.attn_k.weight
    BF16
    [2048, 1024]
  • blk.26.attn_k_norm.weight
    F32
    [128]
  • blk.26.attn_norm.weight
    F32
    [2048]
  • blk.26.attn_output.weight
    BF16
    [8192, 2048]
  • blk.26.attn_q.weight
    BF16
    [2048, 8192]
  • blk.26.attn_q_norm.weight
    F32
    [128]
  • blk.26.attn_v.weight
    BF16
    [2048, 1024]
  • blk.26.exp_probs_b.bias
    F32
    [256]
  • blk.26.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.26.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.26.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.26.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.26.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.26.ffn_norm.weight
    F32
    [2048]
  • blk.26.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.26.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.27
  • blk.27.attn_g.weight
    BF16
    [2048, 64]
  • blk.27.attn_k.weight
    BF16
    [2048, 1024]
  • blk.27.attn_k_norm.weight
    F32
    [128]
  • blk.27.attn_norm.weight
    F32
    [2048]
  • blk.27.attn_output.weight
    BF16
    [8192, 2048]
  • blk.27.attn_q.weight
    BF16
    [2048, 8192]
  • blk.27.attn_q_norm.weight
    F32
    [128]
  • blk.27.attn_v.weight
    BF16
    [2048, 1024]
  • blk.27.exp_probs_b.bias
    F32
    [256]
  • blk.27.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.27.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.27.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.27.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.27.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.27.ffn_norm.weight
    F32
    [2048]
  • blk.27.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.27.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.28
  • blk.28.attn_g.weight
    BF16
    [2048, 48]
  • blk.28.attn_k.weight
    BF16
    [2048, 1024]
  • blk.28.attn_k_norm.weight
    F32
    [128]
  • blk.28.attn_norm.weight
    F32
    [2048]
  • blk.28.attn_output.weight
    BF16
    [6144, 2048]
  • blk.28.attn_q.weight
    BF16
    [2048, 6144]
  • blk.28.attn_q_norm.weight
    F32
    [128]
  • blk.28.attn_v.weight
    BF16
    [2048, 1024]
  • blk.28.exp_probs_b.bias
    F32
    [256]
  • blk.28.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.28.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.28.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.28.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.28.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.28.ffn_norm.weight
    F32
    [2048]
  • blk.28.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.28.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.29
  • blk.29.attn_g.weight
    BF16
    [2048, 64]
  • blk.29.attn_k.weight
    BF16
    [2048, 1024]
  • blk.29.attn_k_norm.weight
    F32
    [128]
  • blk.29.attn_norm.weight
    F32
    [2048]
  • blk.29.attn_output.weight
    BF16
    [8192, 2048]
  • blk.29.attn_q.weight
    BF16
    [2048, 8192]
  • blk.29.attn_q_norm.weight
    F32
    [128]
  • blk.29.attn_v.weight
    BF16
    [2048, 1024]
  • blk.29.exp_probs_b.bias
    F32
    [256]
  • blk.29.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.29.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.29.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.29.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.29.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.29.ffn_norm.weight
    F32
    [2048]
  • blk.29.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.29.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.30
  • blk.30.attn_g.weight
    BF16
    [2048, 64]
  • blk.30.attn_k.weight
    BF16
    [2048, 1024]
  • blk.30.attn_k_norm.weight
    F32
    [128]
  • blk.30.attn_norm.weight
    F32
    [2048]
  • blk.30.attn_output.weight
    BF16
    [8192, 2048]
  • blk.30.attn_q.weight
    BF16
    [2048, 8192]
  • blk.30.attn_q_norm.weight
    F32
    [128]
  • blk.30.attn_v.weight
    BF16
    [2048, 1024]
  • blk.30.exp_probs_b.bias
    F32
    [256]
  • blk.30.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.30.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.30.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.30.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.30.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.30.ffn_norm.weight
    F32
    [2048]
  • blk.30.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.30.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.31
  • blk.31.attn_g.weight
    BF16
    [2048, 64]
  • blk.31.attn_k.weight
    BF16
    [2048, 1024]
  • blk.31.attn_k_norm.weight
    F32
    [128]
  • blk.31.attn_norm.weight
    F32
    [2048]
  • blk.31.attn_output.weight
    BF16
    [8192, 2048]
  • blk.31.attn_q.weight
    BF16
    [2048, 8192]
  • blk.31.attn_q_norm.weight
    F32
    [128]
  • blk.31.attn_v.weight
    BF16
    [2048, 1024]
  • blk.31.exp_probs_b.bias
    F32
    [256]
  • blk.31.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.31.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.31.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.31.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.31.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.31.ffn_norm.weight
    F32
    [2048]
  • blk.31.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.31.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.32
  • blk.32.attn_g.weight
    BF16
    [2048, 48]
  • blk.32.attn_k.weight
    BF16
    [2048, 1024]
  • blk.32.attn_k_norm.weight
    F32
    [128]
  • blk.32.attn_norm.weight
    F32
    [2048]
  • blk.32.attn_output.weight
    BF16
    [6144, 2048]
  • blk.32.attn_q.weight
    BF16
    [2048, 6144]
  • blk.32.attn_q_norm.weight
    F32
    [128]
  • blk.32.attn_v.weight
    BF16
    [2048, 1024]
  • blk.32.exp_probs_b.bias
    F32
    [256]
  • blk.32.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.32.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.32.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.32.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.32.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.32.ffn_norm.weight
    F32
    [2048]
  • blk.32.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.32.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.33
  • blk.33.attn_g.weight
    BF16
    [2048, 64]
  • blk.33.attn_k.weight
    BF16
    [2048, 1024]
  • blk.33.attn_k_norm.weight
    F32
    [128]
  • blk.33.attn_norm.weight
    F32
    [2048]
  • blk.33.attn_output.weight
    BF16
    [8192, 2048]
  • blk.33.attn_q.weight
    BF16
    [2048, 8192]
  • blk.33.attn_q_norm.weight
    F32
    [128]
  • blk.33.attn_v.weight
    BF16
    [2048, 1024]
  • blk.33.exp_probs_b.bias
    F32
    [256]
  • blk.33.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.33.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.33.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.33.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.33.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.33.ffn_norm.weight
    F32
    [2048]
  • blk.33.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.33.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.34
  • blk.34.attn_g.weight
    BF16
    [2048, 64]
  • blk.34.attn_k.weight
    BF16
    [2048, 1024]
  • blk.34.attn_k_norm.weight
    F32
    [128]
  • blk.34.attn_norm.weight
    F32
    [2048]
  • blk.34.attn_output.weight
    BF16
    [8192, 2048]
  • blk.34.attn_q.weight
    BF16
    [2048, 8192]
  • blk.34.attn_q_norm.weight
    F32
    [128]
  • blk.34.attn_v.weight
    BF16
    [2048, 1024]
  • blk.34.exp_probs_b.bias
    F32
    [256]
  • blk.34.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.34.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.34.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.34.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.34.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.34.ffn_norm.weight
    F32
    [2048]
  • blk.34.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.34.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.35
  • blk.35.attn_g.weight
    BF16
    [2048, 64]
  • blk.35.attn_k.weight
    BF16
    [2048, 1024]
  • blk.35.attn_k_norm.weight
    F32
    [128]
  • blk.35.attn_norm.weight
    F32
    [2048]
  • blk.35.attn_output.weight
    BF16
    [8192, 2048]
  • blk.35.attn_q.weight
    BF16
    [2048, 8192]
  • blk.35.attn_q_norm.weight
    F32
    [128]
  • blk.35.attn_v.weight
    BF16
    [2048, 1024]
  • blk.35.exp_probs_b.bias
    F32
    [256]
  • blk.35.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.35.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.35.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.35.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.35.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.35.ffn_norm.weight
    F32
    [2048]
  • blk.35.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.35.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.36
  • blk.36.attn_g.weight
    BF16
    [2048, 48]
  • blk.36.attn_k.weight
    BF16
    [2048, 1024]
  • blk.36.attn_k_norm.weight
    F32
    [128]
  • blk.36.attn_norm.weight
    F32
    [2048]
  • blk.36.attn_output.weight
    BF16
    [6144, 2048]
  • blk.36.attn_q.weight
    BF16
    [2048, 6144]
  • blk.36.attn_q_norm.weight
    F32
    [128]
  • blk.36.attn_v.weight
    BF16
    [2048, 1024]
  • blk.36.exp_probs_b.bias
    F32
    [256]
  • blk.36.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.36.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.36.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.36.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.36.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.36.ffn_norm.weight
    F32
    [2048]
  • blk.36.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.36.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.37
  • blk.37.attn_g.weight
    BF16
    [2048, 64]
  • blk.37.attn_k.weight
    BF16
    [2048, 1024]
  • blk.37.attn_k_norm.weight
    F32
    [128]
  • blk.37.attn_norm.weight
    F32
    [2048]
  • blk.37.attn_output.weight
    BF16
    [8192, 2048]
  • blk.37.attn_q.weight
    BF16
    [2048, 8192]
  • blk.37.attn_q_norm.weight
    F32
    [128]
  • blk.37.attn_v.weight
    BF16
    [2048, 1024]
  • blk.37.exp_probs_b.bias
    F32
    [256]
  • blk.37.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.37.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.37.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.37.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.37.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.37.ffn_norm.weight
    F32
    [2048]
  • blk.37.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.37.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.38
  • blk.38.attn_g.weight
    BF16
    [2048, 64]
  • blk.38.attn_k.weight
    BF16
    [2048, 1024]
  • blk.38.attn_k_norm.weight
    F32
    [128]
  • blk.38.attn_norm.weight
    F32
    [2048]
  • blk.38.attn_output.weight
    BF16
    [8192, 2048]
  • blk.38.attn_q.weight
    BF16
    [2048, 8192]
  • blk.38.attn_q_norm.weight
    F32
    [128]
  • blk.38.attn_v.weight
    BF16
    [2048, 1024]
  • blk.38.exp_probs_b.bias
    F32
    [256]
  • blk.38.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.38.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.38.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.38.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.38.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.38.ffn_norm.weight
    F32
    [2048]
  • blk.38.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.38.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • blk.39
  • blk.39.attn_g.weight
    BF16
    [2048, 64]
  • blk.39.attn_k.weight
    BF16
    [2048, 1024]
  • blk.39.attn_k_norm.weight
    F32
    [128]
  • blk.39.attn_norm.weight
    F32
    [2048]
  • blk.39.attn_output.weight
    BF16
    [8192, 2048]
  • blk.39.attn_q.weight
    BF16
    [2048, 8192]
  • blk.39.attn_q_norm.weight
    F32
    [128]
  • blk.39.attn_v.weight
    BF16
    [2048, 1024]
  • blk.39.exp_probs_b.bias
    F32
    [256]
  • blk.39.ffn_down_exps.weight
    Q8_0
    [512, 2048, 256]
  • blk.39.ffn_down_shexp.weight
    BF16
    [512, 2048]
  • blk.39.ffn_gate_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.39.ffn_gate_inp.weight
    F32
    [2048, 256]
  • blk.39.ffn_gate_shexp.weight
    BF16
    [2048, 512]
  • blk.39.ffn_norm.weight
    F32
    [2048]
  • blk.39.ffn_up_exps.weight
    Q8_0
    [2048, 512, 256]
  • blk.39.ffn_up_shexp.weight
    BF16
    [2048, 512]
  • output.weight
    BF16
    [2048, 100352]
  • output_norm.weight
    F32
    [2048]