93 layers › 68 KDA layers
Store and restart
×no parametersif self.layer_idx % self.attn_res_block_size == 0 · 995–998
In layers 13, 25, …, 85: append the running sum to the stored sums and start a new sum. In the other 61 KDA layers and the running sum goes on unchanged.
Why
Blocks. Attending over every earlier layer’s output would keep all of them alive. Summing the layers of a block into one representation and attending over the blocks cuts that memory and communication from one entry per layer to one per block, and bounds the state at inference. Kimi K3 report §2.2
The authors report that about eight blocks recover most of the benefit across model scales; Kimi K3 uses blocks of 12 layers. Kimi K3 report §2.2
The first stored entry is the token embedding, so that it is always among the sources. Kimi K3 report §2.2
- in
- [B batch, 1 positions, m stored sums, 7168 features] [B batch, 1 positions, 7168 d]
- out
- [B batch, 1 positions, m' stored sums, 7168 features]
Explanation
7 of these 68 layers store; the plate draws the step dashed because the other 61 skip it.
It runs after the attention mix has read the sums and before attention. In a layer that stores, has one entry more than and the attention’s update starts the new sum; in every other layer nothing is stored and . So in the layers that store and elsewhere.
The stored sums grow by one entry at each of these layers: is 1 after layer 1, which stores the embedding, and 8 after layer 85.
Codemodeling_kimi_linear.py 973–1046 · config.json 18–254
def _forward_attn_residual(
self,
hidden_states: torch.Tensor,
attention_mask: torch.Tensor | None = None,
position_ids: torch.LongTensor | None = None,
past_key_values: tuple[torch.Tensor] | None = None,
output_attentions: bool | None = False,
use_cache: bool | None = False,
block_residual: torch.Tensor | None = None,
**kwargs: Unpack[FlashAttentionKwargs],
):
batch_size, seq_len, hidden_size = hidden_states.shape
prefix_sum = hidden_states
if block_residual is not None and block_residual.shape[1] > 0:
hidden_states = _apply_attn_res(
prefix_sum.view(-1, hidden_size),
block_residual,
self.self_attention_res_proj,
self.self_attention_res_norm,
).view(batch_size, seq_len, hidden_size)
if self.layer_idx % self.attn_res_block_size == 0:
block_residual = torch.cat(
[block_residual, prefix_sum.view(-1, hidden_size).unsqueeze(1)], dim=1)
prefix_sum = None
hidden_states = self.input_layernorm(hidden_states)
# Self Attention
if self.is_linear_attn is False:
hidden_states = self.self_attn(
hidden_states=hidden_states,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_values=past_key_values,
output_attentions=output_attentions,
use_cache=use_cache,
**kwargs,
)
else:
hidden_states = self.self_attn(
hidden_states=hidden_states,
attention_mask=attention_mask,
cache_params=past_key_values,
output_attentions=output_attentions,
use_cache=use_cache,
**kwargs,
)
if prefix_sum is not None:
prefix_sum = prefix_sum + hidden_states
else:
prefix_sum = hidden_states
hidden_states = _apply_attn_res(
prefix_sum.view(-1, hidden_size),
block_residual,
self.mlp_res_proj,
self.mlp_res_norm,
).view(batch_size, seq_len, hidden_size)
hidden_states = self.post_attention_layernorm(hidden_states)
if hasattr(self, "block_sparse_moe"):
hidden_states = self.block_sparse_moe(hidden_states)
else:
hidden_states = self.mlp(hidden_states)
if prefix_sum is None:
prefix_sum = hidden_states
else:
prefix_sum = prefix_sum + hidden_states
return prefix_sum, block_residual "text_config": {
"_name_or_path": "",
"activation_situ_beta": 4.0,
"activation_situ_linear_beta": 25.0,
"add_cross_attention": false,
"architectures": [
"KimiLinearForCausalLM"
],
"attn_res_block_size": 12, "first_k_dense_replace": 1,
"forced_bos_token_id": null,
"forced_eos_token_id": null,
"hidden_act": "situ", "linear_attn_config": {
"full_attn_layers": [
4,
8,
12,
16,
20,
24,
28,
32,
36,
40,
44,
48,
52,
56,
60,
64,
68,
72,
76,
80,
84,
88,
92,
93
],
"gate_lower_bound": -5.0,
"head_dim": 128,
"kda_layers": [
1,
2,
3,
5,
6,
7,
9,
10,
11,
13,
14,
15,
17,
18,
19,
21,
22,
23,
25,
26,
27,
29,
30,
31,
33,
34,
35,
37,
38,
39,
41,
42,
43,
45,
46,
47,
49,
50,
51,
53,
54,
55,
57,
58,
59,
61,
62,
63,
65,
66,
67,
69,
70,
71,
73,
74,
75,
77,
78,
79,
81,
82,
83,
85,
86,
87,
89,
90,
91
],
"num_heads": 96,
"short_conv_kernel_size": 4,
"use_full_rank_gate": true
},
"max_length": 20,
"max_position_embeddings": 1048576,
"min_length": 0,
"mla_use_nope": true, "moe_intermediate_size": 3072,
"moe_layer_freq": 1,
"moe_renormalize": true,
"moe_router_activation_func": "sigmoid",
"no_repeat_ngram_size": 0,
"num_attention_heads": 96,
"num_beam_groups": 1,
"num_beams": 1,
"num_expert_group": 1,
"num_experts": 896,
"num_experts_per_token": 16,
"num_hidden_layers": 93,
"num_key_value_heads": 96,
"num_nextn_predict_layers": 0,
"num_return_sequences": 1,
"num_shared_experts": 2,
"output_attentions": false,
"output_hidden_states": false,
"output_scores": false,
"pad_token_id": 163839,
"prefix": null,
"problem_type": null,
"pruned_heads": {},
"q_lora_rank": 1536,
"qk_nope_head_dim": 128,
"qk_rope_head_dim": 64, "routed_expert_hidden_size": 3584,
"routed_scaling_factor": 1.0,
"sep_token_id": null,
"suppress_tokens": null,
"task_specific_params": null,
"temperature": 1.0,
"tf_legacy_loss": false,
"tie_encoder_decoder": false,
"tie_word_embeddings": false,Sources2
Direct evidence
With attention residuals, a layer mixes its attention input from the stored sums and the running sum, stores the running sum and starts a new one when its index is divisible by the block size, adds the attention output to the running sum, mixes the feed-forward input the same way, and adds the feed-forward output.
block-forward · moonshotai/Kimi-K3 / modeling_kimi_linear.py
Pinned revision f831ab668142 ↗ · lines 973–1046
Read code excerpt
def _forward_attn_residual(
self,
hidden_states: torch.Tensor,
attention_mask: torch.Tensor | None = None,
position_ids: torch.LongTensor | None = None,
past_key_values: tuple[torch.Tensor] | None = None,
output_attentions: bool | None = False,
use_cache: bool | None = False,
block_residual: torch.Tensor | None = None,
**kwargs: Unpack[FlashAttentionKwargs],
):
batch_size, seq_len, hidden_size = hidden_states.shape
prefix_sum = hidden_states
if block_residual is not None and block_residual.shape[1] > 0:
hidden_states = _apply_attn_res(
prefix_sum.view(-1, hidden_size),
block_residual,
self.self_attention_res_proj,
self.self_attention_res_norm,
).view(batch_size, seq_len, hidden_size)
if self.layer_idx % self.attn_res_block_size == 0:
block_residual = torch.cat(
[block_residual, prefix_sum.view(-1, hidden_size).unsqueeze(1)], dim=1)
prefix_sum = None
hidden_states = self.input_layernorm(hidden_states)
# Self Attention
if self.is_linear_attn is False:
hidden_states = self.self_attn(
hidden_states=hidden_states,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_values=past_key_values,
output_attentions=output_attentions,
use_cache=use_cache,
**kwargs,
)
else:
hidden_states = self.self_attn(
hidden_states=hidden_states,
attention_mask=attention_mask,
cache_params=past_key_values,
output_attentions=output_attentions,
use_cache=use_cache,
**kwargs,
)
if prefix_sum is not None:
prefix_sum = prefix_sum + hidden_states
else:
prefix_sum = hidden_states
hidden_states = _apply_attn_res(
prefix_sum.view(-1, hidden_size),
block_residual,
self.mlp_res_proj,
self.mlp_res_norm,
).view(batch_size, seq_len, hidden_size)
hidden_states = self.post_attention_layernorm(hidden_states)
if hasattr(self, "block_sparse_moe"):
hidden_states = self.block_sparse_moe(hidden_states)
else:
hidden_states = self.mlp(hidden_states)
if prefix_sum is None:
prefix_sum = hidden_states
else:
prefix_sum = prefix_sum + hidden_states
return prefix_sum, block_residualDirect evidence
The released text configuration: 93 layers of width 7,168, the lists of Kimi Delta Attention and full-attention layers, attention-residual blocks of 12 layers, one dense layer, 896 routed experts of which 16 are used, 2 shared experts, a routed-expert latent of 3,584 features, MLA ranks and head widths, the SiTU activation, vocabulary 163,840, untied output head.
config · moonshotai/Kimi-K3 / config.json
Pinned revision f831ab668142 ↗ · lines 18–254
Read code excerpt
"text_config": {
"_name_or_path": "",
"activation_situ_beta": 4.0,
"activation_situ_linear_beta": 25.0,
"add_cross_attention": false,
"architectures": [
"KimiLinearForCausalLM"
],
"attn_res_block_size": 12,… lines 27–45 omitted …
"first_k_dense_replace": 1,
"forced_bos_token_id": null,
"forced_eos_token_id": null,
"hidden_act": "situ",… lines 50–65 omitted …
"linear_attn_config": {
"full_attn_layers": [
4,
8,
12,
16,
20,
24,
28,
32,
36,
40,
44,
48,
52,
56,
60,
64,
68,
72,
76,
80,
84,
88,
92,
93
],
"gate_lower_bound": -5.0,
"head_dim": 128,
"kda_layers": [
1,
2,
3,
5,
6,
7,
9,
10,
11,
13,
14,
15,
17,
18,
19,
21,
22,
23,
25,
26,
27,
29,
30,
31,
33,
34,
35,
37,
38,
39,
41,
42,
43,
45,
46,
47,
49,
50,
51,
53,
54,
55,
57,
58,
59,
61,
62,
63,
65,
66,
67,
69,
70,
71,
73,
74,
75,
77,
78,
79,
81,
82,
83,
85,
86,
87,
89,
90,
91
],
"num_heads": 96,
"short_conv_kernel_size": 4,
"use_full_rank_gate": true
},
"max_length": 20,
"max_position_embeddings": 1048576,
"min_length": 0,
"mla_use_nope": true,… lines 174–175 omitted …
"moe_intermediate_size": 3072,
"moe_layer_freq": 1,
"moe_renormalize": true,
"moe_router_activation_func": "sigmoid",
"no_repeat_ngram_size": 0,
"num_attention_heads": 96,
"num_beam_groups": 1,
"num_beams": 1,
"num_expert_group": 1,
"num_experts": 896,
"num_experts_per_token": 16,
"num_hidden_layers": 93,
"num_key_value_heads": 96,
"num_nextn_predict_layers": 0,
"num_return_sequences": 1,
"num_shared_experts": 2,
"output_attentions": false,
"output_hidden_states": false,
"output_scores": false,
"pad_token_id": 163839,
"prefix": null,
"problem_type": null,
"pruned_heads": {},
"q_lora_rank": 1536,
"qk_nope_head_dim": 128,
"qk_rope_head_dim": 64,… lines 202–245 omitted …
"routed_expert_hidden_size": 3584,
"routed_scaling_factor": 1.0,
"sep_token_id": null,
"suppress_tokens": null,
"task_specific_params": null,
"temperature": 1.0,
"tf_legacy_loss": false,
"tie_encoder_decoder": false,
"tie_word_embeddings": false,