- leading_dense_block_count => hparams.n_leading_dense_layer,
- expert_feed_forward_length => hparams.n_expert_ff,
- expert_shared_count => hparams.n_expert_shared,
- attention.q_lora_rank => hparams.n_lora_q,
- attention.kv_lora_rank => hparams.n_lora_kv

Added missing scaling of kq_scale parameter.

…ts in the final mscale value equal to 1.0.

…omments.

… multiplier of the ln(s) from the sqrt(1/t) = 0.1 ln(s) + 1 equation.

…k-V2 models.

…_head_qk_rope, n_embd_head_qk_nope

…) implementation

…states, v_states

…ace with single ggml_concat in build_deepseek2()