predictivemanish/ sarvam-30b:latest

164 4 months ago

Sarvam-30B - Multilingual Indian LLM. Converted and packaged for Ollama.

... /
model
94d8a9374055 · 20GB
    Metadata
  • general.architecture
    bailingmoe2
  • general.file_type
    Q4_K_M
  • bailingmoe2.attention.head_count
    64
  • bailingmoe2.attention.head_count_kv
    4
  • bailingmoe2.attention.key_length
    64
  • bailingmoe2.attention.layer_norm_rms_epsilon
    1e-06
  • bailingmoe2.attention.value_length
    64
  • bailingmoe2.block_count
    19
  • bailingmoe2.context_length
    131072
  • bailingmoe2.embedding_length
    4096
  • bailingmoe2.expert_count
    128
  • bailingmoe2.expert_feed_forward_length
    1024
  • bailingmoe2.expert_gating_func
    2
  • bailingmoe2.expert_group_count
    1
  • bailingmoe2.expert_group_used_count
    1
  • bailingmoe2.expert_shared_count
    1
  • bailingmoe2.expert_shared_feed_forward_length
    1024
  • bailingmoe2.expert_used_count
    6
  • bailingmoe2.expert_weights_norm
    true
  • bailingmoe2.expert_weights_scale
    2.5
  • bailingmoe2.feed_forward_length
    8192
  • bailingmoe2.leading_dense_block_count
    1
  • bailingmoe2.rope.dimension_count
    64
  • bailingmoe2.rope.freq_base
    8e+06
  • bailingmoe2.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
    6
  • split.no
    0
  • split.tensors.count
    264
    • Tensor
  • token_embd.weight
    Q4_K
    [4096, 262144]
    • blk.0
  • blk.0.attn_k_norm.weight
    F32
    [64]
  • blk.0.attn_norm.weight
    F32
    [4096]
  • blk.0.attn_output.weight
    Q4_K
    [4096, 4096]
  • blk.0.attn_q_norm.weight
    F32
    [64]
  • blk.0.attn_qkv.weight
    Q6_K
    [4096, 4608]
  • blk.0.ffn_down.weight
    Q6_K
    [8192, 4096]
  • blk.0.ffn_gate.weight
    Q4_K
    [4096, 8192]
  • blk.0.ffn_norm.weight
    F32
    [4096]
  • blk.0.ffn_up.weight
    Q4_K
    [4096, 8192]
    • blk.1
  • blk.1.attn_k_norm.weight
    F32
    [64]
  • blk.1.attn_norm.weight
    F32
    [4096]
  • blk.1.attn_output.weight
    Q4_K
    [4096, 4096]
  • blk.1.attn_q_norm.weight
    F32
    [64]
  • blk.1.attn_qkv.weight
    Q6_K
    [4096, 4608]
  • blk.1.exp_probs_b.bias
    F32
    [128]
  • blk.1.ffn_down_exps.weight
    Q6_K
    [1024, 4096, 128]
  • blk.1.ffn_down_shexp.weight
    Q6_K
    [1024, 4096]
  • blk.1.ffn_gate_exps.weight
    Q4_K
    [4096, 1024, 128]
  • blk.1.ffn_gate_inp.weight
    F32
    [4096, 128]
  • blk.1.ffn_gate_shexp.weight
    Q4_K
    [4096, 1024]
  • blk.1.ffn_norm.weight
    F32
    [4096]
  • blk.1.ffn_up_exps.weight
    Q4_K
    [4096, 1024, 128]
  • blk.1.ffn_up_shexp.weight
    Q4_K
    [4096, 1024]
    • blk.2
  • blk.2.attn_k_norm.weight
    F32
    [64]
  • blk.2.attn_norm.weight
    F32
    [4096]
  • blk.2.attn_output.weight
    Q4_K
    [4096, 4096]
  • blk.2.attn_q_norm.weight
    F32
    [64]
  • blk.2.attn_qkv.weight
    Q4_K
    [4096, 4608]
  • blk.2.exp_probs_b.bias
    F32
    [128]
  • blk.2.ffn_down_exps.weight
    Q4_K
    [1024, 4096, 128]
  • blk.2.ffn_down_shexp.weight
    Q4_K
    [1024, 4096]
  • blk.2.ffn_gate_exps.weight
    Q4_K
    [4096, 1024, 128]
  • blk.2.ffn_gate_inp.weight
    F32
    [4096, 128]
  • blk.2.ffn_gate_shexp.weight
    Q4_K
    [4096, 1024]
  • blk.2.ffn_norm.weight
    F32
    [4096]
  • blk.2.ffn_up_exps.weight
    Q4_K
    [4096, 1024, 128]
  • blk.2.ffn_up_shexp.weight
    Q4_K
    [4096, 1024]
    • blk.3
  • blk.3.attn_k_norm.weight
    F32
    [64]
  • blk.3.attn_norm.weight
    F32
    [4096]
  • blk.3.attn_output.weight
    Q4_K
    [4096, 4096]
  • blk.3.attn_q_norm.weight
    F32
    [64]
  • blk.3.attn_qkv.weight
    Q4_K
    [4096, 4608]
  • blk.3.exp_probs_b.bias
    F32
    [128]
  • blk.3.ffn_down_exps.weight
    Q4_K
    [1024, 4096, 128]
  • blk.3.ffn_down_shexp.weight
    Q4_K
    [1024, 4096]
  • output.weight
    Q6_K
    [4096, 262144]
  • output_norm.weight
    F32
    [4096]