predictivemanish/ sarvam-105b:latest

67 4 months ago

Sarvam-105B (Q4_K_M GGUF) packaged for Ollama. Multilingual Indian MoE LLM made available for local experimentation.

... /
model
d42bfb5fff07 · 64GB
    Metadata
  • general.architecture
    deepseek2
  • general.file_type
    Q4_K_M
  • deepseek2.attention.head_count
    64
  • 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-06
  • deepseek2.attention.q_lora_rank
    0
  • deepseek2.attention.value_length
    512
  • deepseek2.attention.value_length_mla
    128
  • deepseek2.block_count
    32
  • deepseek2.context_length
    131072
  • deepseek2.embedding_length
    4096
  • deepseek2.expert_count
    128
  • deepseek2.expert_feed_forward_length
    2048
  • deepseek2.expert_shared_count
    1
  • deepseek2.expert_used_count
    8
  • deepseek2.expert_weights_norm
    true
  • deepseek2.expert_weights_scale
    2.5
  • deepseek2.feed_forward_length
    16384
  • deepseek2.leading_dense_block_count
    1
  • deepseek2.rope.dimension_count
    64
  • deepseek2.rope.freq_base
    10000
  • deepseek2.rope.scaling.yarn_log_multiplier
    0.1
  • deepseek2.vocab_size
    262144
  • tokenizer.ggml.add_bos_token
    false
  • tokenizer.ggml.add_eos_token
    false
  • tokenizer.ggml.add_sep_token
    false
  • tokenizer.ggml.add_space_prefix
    false
  • tokenizer.ggml.bos_token_id
    6
  • tokenizer.ggml.eos_token_id
    26
  • tokenizer.ggml.model
    llama
  • tokenizer.ggml.padding_token_id
    0
  • tokenizer.ggml.pre
    default
  • tokenizer.ggml.scores
    [-513739, -513739, -513739, -513739, -513739, ...]
  • tokenizer.ggml.token_type
    [3, 3, 3, 3, 3, ...]
  • tokenizer.ggml.tokens
    [<pad>, <eos>, <bos>, <unk>, <mask>, ...]
  • tokenizer.ggml.unknown_token_id
    3
  • split.count
    9
  • split.no
    0
  • split.tensors.count
    510
    • Tensor
  • token_embd.weight
    Q4_K
    [4096, 262144]
    • blk.0
  • blk.0.attn_k_b.weight
    Q5_0
    [128, 512, 64]
  • blk.0.attn_kv_a_mqa.weight
    Q4_K
    [4096, 576]
  • blk.0.attn_kv_a_norm.weight
    F32
    [512]
  • blk.0.attn_norm.weight
    F32
    [4096]
  • blk.0.attn_output.weight
    Q4_K
    [8192, 4096]
  • blk.0.attn_q.weight
    Q4_K
    [4096, 12288]
  • blk.0.attn_v_b.weight
    Q4_K
    [512, 128, 64]
  • blk.0.ffn_down.weight
    Q6_K
    [16384, 4096]
  • blk.0.ffn_gate.weight
    Q4_K
    [4096, 16384]
  • blk.0.ffn_norm.weight
    F32
    [4096]
  • blk.0.ffn_up.weight
    Q4_K
    [4096, 16384]
    • blk.1
  • blk.1.attn_k_b.weight
    Q5_0
    [128, 512, 64]
  • blk.1.attn_kv_a_mqa.weight
    Q4_K
    [4096, 576]
  • blk.1.attn_kv_a_norm.weight
    F32
    [512]
  • blk.1.attn_norm.weight
    F32
    [4096]
  • blk.1.attn_output.weight
    Q4_K
    [8192, 4096]
  • blk.1.attn_q.weight
    Q4_K
    [4096, 12288]
  • blk.1.attn_v_b.weight
    Q4_K
    [512, 128, 64]
  • blk.1.exp_probs_b.bias
    F32
    [128]
  • blk.1.ffn_down_exps.weight
    Q6_K
    [2048, 4096, 128]
  • blk.1.ffn_down_shexp.weight
    Q6_K
    [2048, 4096]
  • blk.1.ffn_gate_exps.weight
    Q4_K
    [4096, 2048, 128]
  • blk.1.ffn_gate_inp.weight
    F32
    [4096, 128]
  • blk.1.ffn_gate_shexp.weight
    Q4_K
    [4096, 2048]
  • blk.1.ffn_norm.weight
    F32
    [4096]
  • blk.1.ffn_up_exps.weight
    Q4_K
    [4096, 2048, 128]
  • blk.1.ffn_up_shexp.weight
    Q4_K
    [4096, 2048]
    • blk.2
  • blk.2.attn_k_b.weight
    Q5_0
    [128, 512, 64]
  • blk.2.attn_kv_a_mqa.weight
    Q4_K
    [4096, 576]
  • blk.2.attn_kv_a_norm.weight
    F32
    [512]
  • blk.2.attn_norm.weight
    F32
    [4096]
  • blk.2.attn_output.weight
    Q4_K
    [8192, 4096]
  • blk.2.attn_q.weight
    Q4_K
    [4096, 12288]
  • blk.2.attn_v_b.weight
    Q4_K
    [512, 128, 64]
  • blk.2.exp_probs_b.bias
    F32
    [128]
  • blk.2.ffn_down_exps.weight
    Q6_K
    [2048, 4096, 128]
  • blk.2.ffn_down_shexp.weight
    Q6_K
    [2048, 4096]
  • blk.2.ffn_gate_exps.weight
    Q4_K
    [4096, 2048, 128]
  • blk.2.ffn_gate_inp.weight
    F32
    [4096, 128]
  • blk.2.ffn_gate_shexp.weight
    Q4_K
    [4096, 2048]
  • blk.2.ffn_norm.weight
    F32
    [4096]
  • blk.2.ffn_up_exps.weight
    Q4_K
    [4096, 2048, 128]
  • blk.2.ffn_up_shexp.weight
    Q4_K
    [4096, 2048]
    • blk.3
  • blk.3.attn_k_b.weight
    Q5_0
    [128, 512, 64]
  • blk.3.attn_kv_a_mqa.weight
    Q4_K
    [4096, 576]
  • blk.3.attn_kv_a_norm.weight
    F32
    [512]
  • blk.3.attn_norm.weight
    F32
    [4096]
  • blk.3.attn_output.weight
    Q4_K
    [8192, 4096]
  • blk.3.attn_q.weight
    Q4_K
    [4096, 12288]
  • blk.3.attn_v_b.weight
    Q4_K
    [512, 128, 64]
  • blk.3.exp_probs_b.bias
    F32
    [128]
  • blk.3.ffn_down_exps.weight
    Q6_K
    [2048, 4096, 128]
  • blk.3.ffn_down_shexp.weight
    Q6_K
    [2048, 4096]
  • blk.3.ffn_gate_exps.weight
    Q4_K
    [4096, 2048, 128]
  • blk.3.ffn_gate_inp.weight
    F32
    [4096, 128]
  • blk.3.ffn_gate_shexp.weight
    Q4_K
    [4096, 2048]
  • blk.3.ffn_norm.weight
    F32
    [4096]
  • output.weight
    Q6_K
    [4096, 262144]
  • output_norm.weight
    F32
    [4096]