61 transformer blocks › 60 MoE blocks › Latent attention
KV cache
×state · kept between callspast_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
The rebuilt keys and the values of every head and position so far, kept between forward calls.
- shape
- k [B batch, 64 heads, 1001 cached positions, 192 content + rotary] v [B batch, 64 heads, 1001 cached positions, 128 value]
This call rebuilds the new token’s keys and values and hands them to the cache, which returns all 1,001 positions for the scores. It holds 64 heads of 192 key and 128 value features per position, not the 512-feature latent they are rebuilt from: 20,480 values per position against 576. It has no fixed size; the slider stops at the configuration’s 131,072 positions (max_position_embeddings), which the code does not enforce.
Codemodeling_deepseek.py 776–799, 810–824 · config.json 1–68
kv_seq_len = value_states.shape[-2]
if past_key_value is not None:
if self.layer_idx is None:
raise ValueError(
f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
"for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
"with a layer index."
)
kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len) if past_key_value is not None:
cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
key_states, value_states = past_key_value.update(
key_states, value_states, self.layer_idx, cache_kwargs
)
attn_weights = (
torch.matmul(query_states, key_states.transpose(2, 3)) * self.softmax_scale
)
if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
raise ValueError(
f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
f" {attn_weights.size()}"
) compressed_kv = self.kv_a_proj_with_mqa(hidden_states)
compressed_kv, k_pe = torch.split(
compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1
)
k_pe = k_pe.view(bsz, q_len, 1, self.qk_rope_head_dim).transpose(1, 2)
kv = (
self.kv_b_proj(self.kv_a_layernorm(compressed_kv))
.view(bsz, q_len, self.num_heads, self.qk_nope_head_dim + self.v_head_dim)
.transpose(1, 2)
)
k_nope, value_states = torch.split(
kv, [self.qk_nope_head_dim, self.v_head_dim], dim=-1
){
"architectures": [
"DeepseekV3ForCausalLM"
],
"attention_bias": false,
"attention_dropout": 0.0,
"auto_map": {
"AutoConfig": "configuration_deepseek.DeepseekV3Config",
"AutoModel": "modeling_deepseek.DeepseekV3Model",
"AutoModelForCausalLM": "modeling_deepseek.DeepseekV3ForCausalLM"
},
"aux_loss_alpha": 0.001,
"bos_token_id": 163584,
"eos_token_id": 163585,
"first_k_dense_replace": 1,
"hidden_act": "silu",
"hidden_size": 7168,
"initializer_range": 0.02,
"intermediate_size": 18432,
"kv_lora_rank": 512,
"max_position_embeddings": 131072,
"model_type": "kimi_k2",
"moe_intermediate_size": 2048,
"moe_layer_freq": 1,
"n_group": 1,
"n_routed_experts": 384,
"n_shared_experts": 1,
"norm_topk_prob": true,
"num_attention_heads": 64,
"num_experts_per_tok": 8,
"num_hidden_layers": 61,
"num_key_value_heads": 64,
"num_nextn_predict_layers": 0,
"pretraining_tp": 1,
"q_lora_rank": 1536,
"qk_nope_head_dim": 128,
"qk_rope_head_dim": 64,
"quantization_config": {
"activation_scheme": "dynamic",
"fmt": "e4m3",
"quant_method": "fp8",
"weight_block_size": [
128,
128
]
},
"rms_norm_eps": 1e-06,
"rope_theta": 50000.0,
"routed_scaling_factor": 2.827,
"rope_scaling": {
"beta_fast": 1.0,
"beta_slow": 1.0,
"factor": 32.0,
"mscale": 1.0,
"mscale_all_dim": 1.0,
"original_max_position_embeddings": 4096,
"type": "yarn"
},
"scoring_func": "sigmoid",
"seq_aux": true,
"tie_word_embeddings": false,
"topk_group": 1,
"topk_method": "noaux_tc",
"torch_dtype": "bfloat16",
"transformers_version": "4.48.3",
"use_cache": true,
"v_head_dim": 128,
"vocab_size": 163840Sources3
Contextual evidence
With a cache, the cached length is added to this call’s, the rebuilt and rotated keys and the values are handed to the cache object, which returns every cached key and value so far, and the scores cover all of them.
kv-cache · moonshotai/Kimi-K2-Instruct / modeling_deepseek.py
Pinned revision fd1984e2b7a3 ↗ · lines 790–824
Read code excerpt
kv_seq_len = value_states.shape[-2]
if past_key_value is not None:
if self.layer_idx is None:
raise ValueError(
f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
"for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
"with a layer index."
)
kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)… lines 800–809 omitted …
if past_key_value is not None:
cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
key_states, value_states = past_key_value.update(
key_states, value_states, self.layer_idx, cache_kwargs
)
attn_weights = (
torch.matmul(query_states, key_states.transpose(2, 3)) * self.softmax_scale
)
if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
raise ValueError(
f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
f" {attn_weights.size()}"
)Contextual evidence
The joint down-projection yields the key–value latent and one rotary key shared by all heads; the latent is normalized and up-projected into per-head content keys and values.
mla-kv · moonshotai/Kimi-K2-Instruct / modeling_deepseek.py
Pinned revision fd1984e2b7a3 ↗ · lines 776–789
Read code excerpt
compressed_kv = self.kv_a_proj_with_mqa(hidden_states)
compressed_kv, k_pe = torch.split(
compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1
)
k_pe = k_pe.view(bsz, q_len, 1, self.qk_rope_head_dim).transpose(1, 2)
kv = (
self.kv_b_proj(self.kv_a_layernorm(compressed_kv))
.view(bsz, q_len, self.num_heads, self.qk_nope_head_dim + self.v_head_dim)
.transpose(1, 2)
)
k_nope, value_states = torch.split(
kv, [self.qk_nope_head_dim, self.v_head_dim], dim=-1
)Contextual evidence
The released configuration: width, layers, dense layers, heads, routed and shared experts, routing (sigmoid scores, one expert group, top-8 of 384, route scale 2.827), MLA ranks and head widths, vocabulary, YaRN RoPE.
config · moonshotai/Kimi-K2-Instruct / config.json
Pinned revision fd1984e2b7a3 ↗ · lines 1–68
Read code excerpt
{
"architectures": [
"DeepseekV3ForCausalLM"
],
"attention_bias": false,
"attention_dropout": 0.0,
"auto_map": {
"AutoConfig": "configuration_deepseek.DeepseekV3Config",
"AutoModel": "modeling_deepseek.DeepseekV3Model",
"AutoModelForCausalLM": "modeling_deepseek.DeepseekV3ForCausalLM"
},
"aux_loss_alpha": 0.001,
"bos_token_id": 163584,
"eos_token_id": 163585,
"first_k_dense_replace": 1,
"hidden_act": "silu",
"hidden_size": 7168,
"initializer_range": 0.02,
"intermediate_size": 18432,
"kv_lora_rank": 512,
"max_position_embeddings": 131072,
"model_type": "kimi_k2",
"moe_intermediate_size": 2048,
"moe_layer_freq": 1,
"n_group": 1,
"n_routed_experts": 384,
"n_shared_experts": 1,
"norm_topk_prob": true,
"num_attention_heads": 64,
"num_experts_per_tok": 8,
"num_hidden_layers": 61,
"num_key_value_heads": 64,
"num_nextn_predict_layers": 0,
"pretraining_tp": 1,
"q_lora_rank": 1536,
"qk_nope_head_dim": 128,
"qk_rope_head_dim": 64,
"quantization_config": {
"activation_scheme": "dynamic",
"fmt": "e4m3",
"quant_method": "fp8",
"weight_block_size": [
128,
128
]
},
"rms_norm_eps": 1e-06,
"rope_theta": 50000.0,
"routed_scaling_factor": 2.827,
"rope_scaling": {
"beta_fast": 1.0,
"beta_slow": 1.0,
"factor": 32.0,
"mscale": 1.0,
"mscale_all_dim": 1.0,
"original_max_position_embeddings": 4096,
"type": "yarn"
},
"scoring_func": "sigmoid",
"seq_aux": true,
"tie_word_embeddings": false,
"topk_group": 1,
"topk_method": "noaux_tc",
"torch_dtype": "bfloat16",
"transformers_version": "4.48.3",
"use_cache": true,
"v_head_dim": 128,
"vocab_size": 163840