aiconjured/ qwen3.8-flash-next-gsq-rco-gguf-mtp:latest

36 4 days ago

Text + MTP IQ3_S qwen3.8-flash-next-gsq-rco-gguf-mtp

... /
draft
8087dbb39fc7 · 2.8GB
    Metadata
  • general.architecture
    qwen4exp
  • general.file_type
    Q4_K_M
  • qwen4exp.attention.compress_ratios
    [0, 0, 0, 4, 0, ...]
  • qwen4exp.attention.head_count
    24
  • qwen4exp.attention.head_count_kv
    2
  • qwen4exp.attention.indexer.head_count
    4
  • qwen4exp.attention.indexer.key_length
    128
  • qwen4exp.attention.indexer.top_k
    2048
  • qwen4exp.attention.key_length
    256
  • qwen4exp.attention.layer_norm_rms_epsilon
    1e-06
  • qwen4exp.attention.value_length
    256
  • qwen4exp.block_count
    49
  • qwen4exp.context_length
    262144
  • qwen4exp.embedding_length
    2560
  • qwen4exp.expert_count
    512
  • qwen4exp.expert_feed_forward_length
    640
  • qwen4exp.expert_shared_feed_forward_length
    640
  • qwen4exp.expert_used_count
    10
  • qwen4exp.full_attention_interval
    4
  • qwen4exp.hyper_connection.count
    4
  • qwen4exp.hyper_connection.low_rank
    320
  • qwen4exp.nextn_predict_layers
    1
  • qwen4exp.rope.dimension_count
    64
  • qwen4exp.rope.dimension_sections
    [11, 11, 10, 0]
  • qwen4exp.rope.freq_base
    1e+07
  • qwen4exp.ssm.conv_kernel
    4
  • qwen4exp.ssm.group_count
    16
  • qwen4exp.ssm.inner_size
    6144
  • qwen4exp.ssm.state_size
    128
  • qwen4exp.ssm.time_step_rank
    48
  • tokenizer.ggml.add_bos_token
    false
  • tokenizer.ggml.bos_token_id
    248044
  • tokenizer.ggml.eos_token_id
    248046
  • tokenizer.ggml.merges
    [Ġ Ġ, ĠĠ ĠĠ, i n, Ġ t, ĠĠĠĠ ĠĠĠĠ, ...]
  • tokenizer.ggml.model
    gpt2
  • tokenizer.ggml.padding_token_id
    248044
  • tokenizer.ggml.pre
    qwen35
  • tokenizer.ggml.token_type
    [1, 1, 1, 1, 1, ...]
  • tokenizer.ggml.tokens
    [!, ", #, $, %, ...]
    • Tensor
  • token_embd.weight
    Q4_K
    [2560, 248320]
    • blk.48
  • blk.48.attn_k.weight
    Q4_K
    [2560, 512]
  • blk.48.attn_k_norm.weight
    F32
    [256]
  • blk.48.attn_output.weight
    Q4_K
    [6144, 2560]
  • blk.48.attn_q.weight
    Q4_K
    [2560, 12288]
  • blk.48.attn_q_norm.weight
    F32
    [256]
  • blk.48.attn_v.weight
    Q6_K
    [2560, 512]
  • blk.48.ffn_down_exps.weight
    Q8_0
    [640, 2560, 512]
  • blk.48.ffn_down_shexp.weight
    Q8_0
    [640, 2560]
  • blk.48.ffn_gate_exps.weight
    Q4_K
    [2560, 640, 512]
  • blk.48.ffn_gate_inp.weight
    F32
    [2560, 512]
  • blk.48.ffn_gate_inp_shexp.weight
    F32
    [2560]
  • blk.48.ffn_gate_shexp.weight
    Q4_K
    [2560, 640]
  • blk.48.ffn_up_exps.weight
    Q4_K
    [2560, 640, 512]
  • blk.48.ffn_up_shexp.weight
    Q4_K
    [2560, 640]
  • blk.48.hc_attn_down.weight
    Q4_K
    [10240, 320]
  • blk.48.hc_attn_inject.weight
    Q4_K
    [10240, 4]
  • blk.48.hc_attn_norm.weight
    F32
    [10240]
  • blk.48.hc_attn_up.weight
    Q5_0
    [320, 10240]
  • blk.48.hc_ffn_down.weight
    Q6_K
    [10240, 320]
  • blk.48.hc_ffn_inject.weight
    Q4_K
    [10240, 4]
  • blk.48.hc_ffn_norm.weight
    F32
    [10240]
  • blk.48.hc_ffn_up.weight
    Q5_0
    [320, 10240]
  • blk.48.indexer.k_norm.weight
    F32
    [128]
  • blk.48.indexer.k_proj.weight
    BF16
    [2560, 128]
  • blk.48.indexer.q_norm.weight
    F32
    [128]
  • blk.48.indexer.q_proj.weight
    BF16
    [2560, 512]
  • blk.48.nextn.eh_proj.weight
    Q4_K
    [5120, 2560]
  • blk.48.nextn.enorm.weight
    F32
    [2560]
  • blk.48.nextn.hc_head_down.weight
    Q4_K
    [10240, 320]
  • blk.48.nextn.hc_head_norm.weight
    F32
    [10240]
  • blk.48.nextn.hc_head_up.weight
    Q5_0
    [320, 10240]
  • blk.48.nextn.hnorm.weight
    F32
    [10240]
  • output.weight
    Q6_K
    [2560, 248320]