frob/ deepseek-v4.1-flash:552b

205 5 days ago

1017

vision 552b
... /
draft
e62f1887cd82 · 8.5GB
    Metadata
  • general.architecture
    dflash
  • general.file_type
    (!unknown_file_type 38!)
  • dflash.attention.compress_ratios
    [0, 0, 0]
  • dflash.attention.head_count
    64
  • dflash.attention.head_count_kv
    1
  • dflash.attention.key_length
    512
  • dflash.attention.layer_norm_rms_epsilon
    1e-20
  • dflash.attention.output_group_count
    8
  • dflash.attention.output_lora_rank
    1024
  • dflash.attention.q_lora_rank
    1280
  • dflash.attention.sliding_window
    128
  • dflash.attention.value_length
    512
  • dflash.block_count
    3
  • dflash.block_size
    5
  • dflash.context_length
    1048576
  • dflash.embedding_length
    5120
  • dflash.expert_count
    128
  • dflash.expert_feed_forward_length
    2304
  • dflash.expert_gating_func
    4
  • dflash.expert_shared_count
    1
  • dflash.expert_used_count
    3
  • dflash.expert_weights_norm
    true
  • dflash.expert_weights_scale
    1.5
  • dflash.feed_forward_length
    2304
  • dflash.has_confidence_head
    true
  • dflash.hyper_connection.count
    4
  • dflash.hyper_connection.epsilon
    1e-06
  • dflash.hyper_connection.single_pass
    true
  • dflash.hyper_connection.sinkhorn_iterations
    20
  • dflash.rope.dimension_count
    64
  • dflash.rope.freq_base
    10000
  • dflash.sample_from_anchor
    true
  • dflash.swiglu_clamp_exp
    [10, 10, 10]
  • dflash.swiglu_clamp_shexp
    [10, 10, 10]
  • dflash.target_layers
    [37, 38, 39]
  • tokenizer.ggml.add_bos_token
    false
  • tokenizer.ggml.add_eos_token
    false
  • tokenizer.ggml.bos_token_id
    0
  • tokenizer.ggml.clean_spaces
    false
  • tokenizer.ggml.eos_token_id
    1
  • tokenizer.ggml.mask_token_id
    128799
  • tokenizer.ggml.merges
    [Ġ t, Ġ a, i n, Ġ Ġ, h e, ...]
  • tokenizer.ggml.model
    gpt2
  • tokenizer.ggml.padding_token_id
    1
  • tokenizer.ggml.pre
    joyai-llm
  • tokenizer.ggml.token_type
    [3, 3, 3, 1, 1, ...]
  • tokenizer.ggml.tokens
    [<｜begin▁of▁sentence｜>, <｜end▁of▁sentence｜>, <｜▁pad▁｜>, !, ", ...]
    • Tensor
    blk.0
  • blk.0.attn_kv.weight
    BF16
    [5120, 512]
  • blk.0.attn_kv_a_norm.weight
    F32
    [512]
  • blk.0.attn_norm.weight
    F32
    [5120]
  • blk.0.attn_output_a.weight
    BF16
    [4096, 8192]
  • blk.0.attn_output_b.weight
    BF16
    [8192, 5120]
  • blk.0.attn_q_a.weight
    BF16
    [5120, 1280]
  • blk.0.attn_q_a_norm.weight
    F32
    [1280]
  • blk.0.attn_q_b.weight
    BF16
    [1280, 32768]
  • blk.0.attn_sinks.weight
    F32
    [64]
  • blk.0.exp_probs_b.bias
    F32
    [128]
  • blk.0.ffn_down_exps.weight
    MXFP4
    [2304, 5120, 128]
  • blk.0.ffn_down_shexp.weight
    BF16
    [2304, 5120]
  • blk.0.ffn_gate_exps.weight
    MXFP4
    [5120, 2304, 128]
  • blk.0.ffn_gate_inp.weight
    F32
    [5120, 128]
  • blk.0.ffn_gate_shexp.weight
    BF16
    [5120, 2304]
  • blk.0.ffn_norm.weight
    F32
    [5120]
  • blk.0.ffn_up_exps.weight
    MXFP4
    [5120, 2304, 128]
  • blk.0.ffn_up_shexp.weight
    BF16
    [5120, 2304]
  • blk.0.hc_attn_base.weight
    F32
    [24]
  • blk.0.hc_attn_fn.weight
    F32
    [20480, 24]
  • blk.0.hc_attn_scale.weight
    F32
    [3]
  • blk.0.hc_ffn_base.weight
    F32
    [24]
  • blk.0.hc_ffn_fn.weight
    F32
    [20480, 24]
  • blk.0.hc_ffn_scale.weight
    F32
    [3]
    • blk.1
  • blk.1.attn_kv.weight
    BF16
    [5120, 512]
  • blk.1.attn_kv_a_norm.weight
    F32
    [512]
  • blk.1.attn_norm.weight
    F32
    [5120]
  • blk.1.attn_output_a.weight
    BF16
    [4096, 8192]
  • blk.1.attn_output_b.weight
    BF16
    [8192, 5120]
  • blk.1.attn_q_a.weight
    BF16
    [5120, 1280]
  • blk.1.attn_q_a_norm.weight
    F32
    [1280]
  • blk.1.attn_q_b.weight
    BF16
    [1280, 32768]
  • blk.1.attn_sinks.weight
    F32
    [64]
  • blk.1.exp_probs_b.bias
    F32
    [128]
  • blk.1.ffn_down_exps.weight
    MXFP4
    [2304, 5120, 128]
  • blk.1.ffn_down_shexp.weight
    BF16
    [2304, 5120]
  • blk.1.ffn_gate_exps.weight
    MXFP4
    [5120, 2304, 128]
  • blk.1.ffn_gate_inp.weight
    F32
    [5120, 128]
  • blk.1.ffn_gate_shexp.weight
    BF16
    [5120, 2304]
  • blk.1.ffn_norm.weight
    F32
    [5120]
  • blk.1.ffn_up_exps.weight
    MXFP4
    [5120, 2304, 128]
  • blk.1.ffn_up_shexp.weight
    BF16
    [5120, 2304]
  • blk.1.hc_attn_base.weight
    F32
    [24]
  • blk.1.hc_attn_fn.weight
    F32
    [20480, 24]
  • blk.1.hc_attn_scale.weight
    F32
    [3]
  • blk.1.hc_ffn_base.weight
    F32
    [24]
  • blk.1.hc_ffn_fn.weight
    F32
    [20480, 24]
  • blk.1.hc_ffn_scale.weight
    F32
    [3]
    • blk.2
  • blk.2.attn_kv.weight
    BF16
    [5120, 512]
  • blk.2.attn_kv_a_norm.weight
    F32
    [512]
  • blk.2.attn_norm.weight
    F32
    [5120]
  • blk.2.attn_output_a.weight
    BF16
    [4096, 8192]
  • blk.2.attn_output_b.weight
    BF16
    [8192, 5120]
  • blk.2.attn_q_a.weight
    BF16
    [5120, 1280]
  • blk.2.attn_q_a_norm.weight
    F32
    [1280]
  • blk.2.attn_q_b.weight
    BF16
    [1280, 32768]
  • blk.2.attn_sinks.weight
    F32
    [64]
  • blk.2.exp_probs_b.bias
    F32
    [128]
  • blk.2.ffn_down_exps.weight
    MXFP4
    [2304, 5120, 128]
  • blk.2.ffn_down_shexp.weight
    BF16
    [2304, 5120]
  • blk.2.ffn_gate_exps.weight
    MXFP4
    [5120, 2304, 128]
  • blk.2.ffn_gate_inp.weight
    F32
    [5120, 128]
  • blk.2.ffn_gate_shexp.weight
    BF16
    [5120, 2304]
  • blk.2.ffn_norm.weight
    F32
    [5120]
  • blk.2.ffn_up_exps.weight
    MXFP4
    [5120, 2304, 128]
  • blk.2.ffn_up_shexp.weight
    BF16
    [5120, 2304]
  • blk.2.hc_attn_base.weight
    F32
    [24]
  • blk.2.hc_attn_fn.weight
    F32
    [20480, 24]
  • blk.2.hc_attn_scale.weight
    F32
    [3]
  • blk.2.hc_ffn_base.weight
    F32
    [24]
  • blk.2.hc_ffn_fn.weight
    F32
    [20480, 24]
  • blk.2.hc_ffn_scale.weight
    F32
    [3]
  • conf_proj.weight
    F32
    [5376, 1]
  • enc.output_norm.weight
    F32
    [5120]
  • fc.weight
    BF16
    [15360, 5120]
  • markov_w1.weight
    BF16
    [256, 129280]
  • markov_w2.weight
    BF16
    [256, 129280]
  • output_norm.weight
    F32
    [5120]