nnterp · encyclopedia ernie4_5_moe · nnterp.families.ernie4_5_moe

family ernie4_5_moe

ERNIE 4.5 MoE

Llama's pre-norm block with rotary on adjacent pairs of dimensions, a dense MLP on the first blocks and a mixture after them, whose experts are chosen by a softmax plus a correction bias that the weights leave out, renormalized to sum to one.

Interleaved rotaryMixture of expertsDense first blocksGrouped-query attention

The block

01 · the residual stream

One block of the stack. Hover any part for its nnterp name; click to pin the card. The slider picks a block, each tick's color gives its block's shape, and the diagram redraws for the block picked.

embed_tokenstoken_embeddings
→
layers× 28
→
normfinal_norm_output
→
lm_headlm_head_output
→
logitsthe output
i = 0

identity layers[i].input + self_attn.attention_output + mlp.mlp_output == layer_output

The API

02 · support()

Values, by host

⚠︎ marks a value this checkpoint has only under some load: hover it for the condition.

model Ernie4_5_MoeForCausalLMlayoutwhat it isread at
logits Logits [batch seq vocab_size] The logits the model returns, after anything it does past lm_head (a softcap, a scale)

the module's own output
token_embeddings Residual [batch seq hidden_size] The embedding module's output; layers[0].input is what enters the first block

the child module embed_tokens's output
final_norm_output Residual [batch seq hidden_size] The residual stream after the final norm, what lm_head reads

the child module norm's output
lm_head_output Logits [batch seq vocab_size] The head's logits before anything the model does past lm_head (a softcap, a scale)

the child module lm_head's output
next_token_probs NextTokenProbs [batch vocab_size] The next-token distribution at the last position; derived, read-only

the module's own output
input_ids Tokens [batch seq] The token ids the model was called with

the module's own input
attention_mask Tokens [batch seq] The attention mask the model was called with; zeros are padding

the module's own input
input_size scalar [batch, seq] of the current call; read-only

the module's own input
model.layers[i] Ernie4_5_MoeDecoderLayerlayoutwhat it isread at
layer_input Residual [batch seq hidden_size] The residual stream entering the block

the module's own input
attention_norm_output Residual [batch seq hidden_size] The normed stream entering the attention

the child module input_layernorm's output
layer_mid Residual [batch seq hidden_size] The residual stream after the attention add

the child module post_attention_layernorm's input
mlp_norm_output Residual [batch seq hidden_size] The normed stream entering the MLP

the child module post_attention_layernorm's output
mlp_input Residual [batch seq hidden_size] What the MLP is called with

the child module mlp's input
layer_output Residual [batch seq hidden_size] The residual stream leaving the block, a tensor even when the block returns a tuple

the module's own output
model.layers[i].self_attn Ernie4_5_MoeAttentionlayoutwhat it isread at
⚠︎attention_queries Queries [batch num_heads seq qk_head_dim] The queries entering attention

inside the forward: the arguments of attention_interface_1, element 1
⚠︎attention_keys Keys [batch num_kv_heads seq qk_head_dim] The keys entering attention

inside the forward: the arguments of attention_interface_1, element 2
⚠︎attention_values Values [batch num_kv_heads seq head_dim] The values entering attention

inside the forward: the arguments of attention_interface_1, element 3
⚠︎attention_scores Pattern [batch num_heads query key] The attention scores entering the softmax, masked

inside the forward: the input of nn_functional_softmax_0 inside attention_interface_1
attention_premix HeadOutputs [batch seq num_heads head_dim] Every head's output side by side, as the output projection takes it

the child module o_proj's input
attention_output Residual [batch seq hidden_size] What the attention adds to the residual stream

the module's own output
⚠︎attention_probabilities Pattern [batch num_heads query key] The attention pattern the values are mixed with

inside the forward: the output of nn_functional_dropout_0 inside attention_interface_1
⚠︎attention_head_outputs HeadOutputs [batch seq num_heads head_dim] The per-head outputs before the output projection

inside the forward: the output of attention_interface_1, element 0
model.layers[i].mlp Ernie4_5_MoeMLP / Ernie4_5_MoeSparseMoeBlocklayoutwhat it isread at
⚠︎mlp_activation Neurons [batch seq intermediate_size] The activation function's output

the child module act_fn's output
⚠︎mlp_neurons Neurons [batch seq intermediate_size] What the down projection receives: act(gate)·up on a gated MLP, the activation itself otherwise

the child module down_proj's input
mlp_output Residual [batch seq hidden_size] What the MLP adds to the residual stream

the module's own output
router_logits RouterLogits [batch seq num_experts] The router's logits, one per expert, before the scoring

inside router's forward: the output of F_linear_0
expert_weights ExpertWeights [batch seq top_k] The weight each selected expert's output is scaled by

the child module experts's inputs, element 2
expert_indices ExpertIndices [batch seq top_k] The expert each routing slot sends the token to

the child module experts's inputs, element 1
expert_outputs ExpertOutputs [batch seq top_k hidden_size] Each routing slot's weighted expert output

inside experts's forward: the output of weighted_out_view_0 inside experts_forward_1
routed_output Residual [batch seq hidden_size] The routed experts' combined output

the child module experts's output
shared_expert_output Residual [batch seq hidden_size] The shared expert's output

the child module shared_experts's output

Printout

What print(model) shows on baidu/ERNIE-4.5-21B-A3B-Base-PT: the native modules, standard names and every standard value with its layout.

Ernie4_5_MoeForCausalLM(
  (lm_head): Linear(in_features=2560, out_features=103424, bias=False)
  (generator): Generator(
    (streamer): Streamer()
  )
  (embed_tokens): Embedding(103424, 2560, padding_idx=0)
  (layers): ModuleList(
    (0): Ernie4_5_MoeDecoderLayer(
      (self_attn): Ernie4_5_MoeAttention(
        (q_proj): Linear(in_features=2560, out_features=2560, bias=False)
        (k_proj): Linear(in_features=2560, out_features=512, bias=False)
        (v_proj): Linear(in_features=2560, out_features=512, bias=False)
        (o_proj): Linear(in_features=2560, out_features=2560, bias=False)
        (attention_queries) -> Queries [batch num_heads seq qk_head_dim]: The queries entering attention
        (attention_keys) -> Keys [batch num_kv_heads seq qk_head_dim]: The keys entering attention
        (attention_values) -> Values [batch num_kv_heads seq head_dim]: The values entering attention
        (attention_scores) -> Pattern [batch num_heads query key]: The attention scores entering the softmax, masked
        (attention_premix) -> HeadOutputs [batch seq num_heads head_dim]: Every head's output side by side, as the output projection takes it
        (attention_output) -> Residual [batch seq hidden_size]: What the attention adds to the residual stream
        (attention_probabilities) -> Pattern [batch num_heads query key]: The attention pattern the values are mixed with
        (attention_head_outputs) -> HeadOutputs [batch seq num_heads head_dim]: The per-head outputs before the output projection
      )
      (mlp): Ernie4_5_MoeMLP(
        (gate_proj): Linear(in_features=2560, out_features=12288, bias=False)
        (up_proj): Linear(in_features=2560, out_features=12288, bias=False)
        (down_proj): Linear(in_features=12288, out_features=2560, bias=False)
        (act_fn): SiLUActivation()
        (mlp_activation) -> Neurons [batch seq intermediate_size]: The activation function's output
        (mlp_neurons) -> Neurons [batch seq intermediate_size]: What the down projection receives: act(gate)·up on a gated MLP, the activation itself otherwise
        (mlp_output) -> Residual [batch seq hidden_size]: What the MLP adds to the residual stream
      )
      (input_layernorm): Ernie4_5_MoeRMSNorm((2560,), eps=1e-05)
      (post_attention_layernorm): Ernie4_5_MoeRMSNorm((2560,), eps=1e-05)
      (layer_input) -> Residual [batch seq hidden_size]: The residual stream entering the block
      (attention_norm_output) -> Residual [batch seq hidden_size]: The normed stream entering the attention
      (layer_mid) -> Residual [batch seq hidden_size]: The residual stream after the attention add
      (mlp_norm_output) -> Residual [batch seq hidden_size]: The normed stream entering the MLP
      (mlp_input) -> Residual [batch seq hidden_size]: What the MLP is called with
      (layer_output) -> Residual [batch seq hidden_size]: The residual stream leaving the block, a tensor even when the block returns a tuple
    )
    (1-27): 27 x Ernie4_5_MoeDecoderLayer(
      (self_attn): Ernie4_5_MoeAttention(
        (q_proj): Linear(in_features=2560, out_features=2560, bias=False)
        (k_proj): Linear(in_features=2560, out_features=512, bias=False)
        (v_proj): Linear(in_features=2560, out_features=512, bias=False)
        (o_proj): Linear(in_features=2560, out_features=2560, bias=False)
        (attention_queries) -> Queries [batch num_heads seq qk_head_dim]: The queries entering attention
        (attention_keys) -> Keys [batch num_kv_heads seq qk_head_dim]: The keys entering attention
        (attention_values) -> Values [batch num_kv_heads seq head_dim]: The values entering attention
        (attention_scores) -> Pattern [batch num_heads query key]: The attention scores entering the softmax, masked
        (attention_premix) -> HeadOutputs [batch seq num_heads head_dim]: Every head's output side by side, as the output projection takes it
        (attention_output) -> Residual [batch seq hidden_size]: What the attention adds to the residual stream
        (attention_probabilities) -> Pattern [batch num_heads query key]: The attention pattern the values are mixed with
        (attention_head_outputs) -> HeadOutputs [batch seq num_heads head_dim]: The per-head outputs before the output projection
      )
      (mlp): Ernie4_5_MoeSparseMoeBlock(
        (router/gate): Ernie4_5_MoeTopKRouter(
          (moe_statics): Ernie4_5_MoeStatics()
        )
        (experts): Ernie4_5_MoeExperts(
          (act_fn): SiLUActivation()
        )
        (shared_experts): Ernie4_5_MoeMLP(
          (gate_proj): Linear(in_features=2560, out_features=3072, bias=False)
          (up_proj): Linear(in_features=2560, out_features=3072, bias=False)
          (down_proj): Linear(in_features=3072, out_features=2560, bias=False)
          (act_fn): SiLUActivation()
          (mlp_activation) -> Neurons [batch seq intermediate_size]: The activation function's output
          (mlp_neurons) -> Neurons [batch seq intermediate_size]: What the down projection receives: act(gate)·up on a gated MLP, the activation itself otherwise
          (mlp_output) -> Residual [batch seq hidden_size]: What the MLP adds to the residual stream
        )
        (mlp_activation) -> Neurons [batch seq intermediate_size]: The activation function's output
        (mlp_neurons) -> Neurons [batch seq intermediate_size]: What the down projection receives: act(gate)·up on a gated MLP, the activation itself otherwise
        (router_logits) -> RouterLogits [batch seq num_experts]: The router's logits, one per expert, before the scoring
        (expert_weights) -> ExpertWeights [batch seq top_k]: The weight each selected expert's output is scaled by
        (expert_indices) -> ExpertIndices [batch seq top_k]: The expert each routing slot sends the token to
        (expert_outputs) -> ExpertOutputs [batch seq top_k hidden_size]: Each routing slot's weighted expert output
        (routed_output) -> Residual [batch seq hidden_size]: The routed experts' combined output
        (shared_expert_output) -> Residual [batch seq hidden_size]: The shared expert's output
        (mlp_output) -> Residual [batch seq hidden_size]: What the MLP adds to the residual stream
      )
      (input_layernorm): Ernie4_5_MoeRMSNorm((2560,), eps=1e-05)
      (post_attention_layernorm): Ernie4_5_MoeRMSNorm((2560,), eps=1e-05)
      (layer_input) -> Residual [batch seq hidden_size]: The residual stream entering the block
      (attention_norm_output) -> Residual [batch seq hidden_size]: The normed stream entering the attention
      (layer_mid) -> Residual [batch seq hidden_size]: The residual stream after the attention add
      (mlp_norm_output) -> Residual [batch seq hidden_size]: The normed stream entering the MLP
      (mlp_input) -> Residual [batch seq hidden_size]: What the MLP is called with
      (layer_output) -> Residual [batch seq hidden_size]: The residual stream leaving the block, a tensor even when the block returns a tuple
    )
  )
  (norm): Ernie4_5_MoeRMSNorm((2560,), eps=1e-05)
  (logits) -> Logits [batch seq vocab_size]: The logits the model returns, after anything it does past lm_head (a softcap, a scale)
  (token_embeddings) -> Residual [batch seq hidden_size]: The embedding module's output; layers[0].input is what enters the first block
  (final_norm_output) -> Residual [batch seq hidden_size]: The residual stream after the final norm, what lm_head reads
  (lm_head_output) -> Logits [batch seq vocab_size]: The head's logits before anything the model does past lm_head (a softcap, a scale)
  (next_token_probs) -> NextTokenProbs [batch vocab_size]: The next-token distribution at the last position; derived, read-only
  (input_ids) -> Tokens [batch seq]: The token ids the model was called with
  (attention_mask) -> Tokens [batch seq]: The attention mask the model was called with; zeros are padding
  (input_size): [batch, seq] of the current call; read-only
)

Config

model.num_layers28
model.hidden_size2560
model.num_heads20
model.num_kv_heads4
model.head_dim128
model.qk_head_dim128
model.vocab_size103424
model.intermediate_size12288
architecturesErnie4_5_MoeForCausalLM
num_hidden_layers28
hidden_size2560
intermediate_size12288
num_attention_heads20
num_key_value_heads4
vocab_size103424
max_position_embeddings131072
hidden_actsilu
rms_norm_eps1e-05
tie_word_embeddingsTrue
dtypebfloat16
moe_intermediate_size1536
moe_layer_start_index1
moe_k6
moe_num_shared_experts2

The Notes

03 · what to know before you run

The block, in order

h   = x + self_attn(input_layernorm(x))
out = h + mlp(post_attention_layernorm(h))    # dense first, a mixture after

Llama's pre-norm block and Llama's names. Block i (counted from 0) has a mixture when moe_layer_start_index <= i <= moe_layer_end_index and i + 1 is a multiple of moe_layer_interval (1 on every checkpoint here). Every other block has a dense MLP intermediate_size wide. On 21B-A3B block 0 is dense and blocks 1 to 27 are mixtures; on 300B-A47B blocks 0 to 2 are dense and 3 to 53 mixtures. Nothing norms a sublayer's output, so the identity is the plain sum on both kinds:

with model.trace(prompt):
    x = model.layers[1].input.save()
    attn = model.layers[1].self_attn.attention_output.save()
    mlp = model.layers[1].mlp.mlp_output.save()
    out = model.layers[1].layer_output.save()

torch.testing.assert_close(x + attn + mlp, out)

Attention

20 query heads over 4 key/value heads on 21B-A3B, 64 over 8 on 300B-A47B, head_dim 128 on both. The rotary turns adjacent pairs of dimensions (2i, 2i + 1) of each query and key head, as on the dense ERNIE 4.5, so attention_queries keeps q_proj's dimension order: at position 0 it is q_proj's output split into heads. No projection has a bias. The attention interior needs attn_implementation="eager".

The router: a softmax, a correction bias, weights summing to one

64 routed experts on every checkpoint, 6 per token on 21B-A3B (moe_k) and 8 on 300B-A47B. router_logits are float32 logits over all 64. The router takes their softmax, adds the correction bias (router.moe_statics.e_score_correction_bias) to the probabilities to choose the experts, and returns the chosen experts' probabilities without the bias, renormalized to sum to one (the sum is clamped at moe_norm_min, 1e-12):

moe = model.layers[1].mlp
with model.trace(prompt):
    logits = moe.router_logits.save()
    w = moe.expert_weights.save()
    idx = moe.expert_indices.save()

chosen = logits.float().softmax(-1).gather(-1, idx)
torch.testing.assert_close(w, (chosen / chosen.sum(-1, keepdim=True)).to(w.dtype))

A written router_logits moves the choice and the weights; a written bias moves only the choice. The bias is added to probabilities, so a bias of 0.1 outweighs any gap between two probabilities smaller than 0.1.

The shared expert runs before the router

21B-A3B has moe_num_shared_experts 2, merged into one Ernie4_5_MoeMLP (mlp.shared_experts) of width 2 × 1536. It runs on every token before the router, so in one trace read shared_expert_output before router_logits. routed_output + shared_expert_output is mlp_output:

with model.trace(prompt):
    shared = moe.shared_expert_output.save()
    routed = moe.routed_output.save()
    out = moe.mlp_output.save()

torch.testing.assert_close(routed + shared, out)

300B-A47B has no shared expert (moe_num_shared_experts 0): shared_expert_output is unavailable there, and routed_output is mlp_output.

Ablating an expert

Zero the weight of every slot that chose it. The token's other weights keep their values, so they then sum to less than one, and mlp_output changes on exactly the tokens that chose it:

with model.trace(prompt):
    idx = moe.expert_indices.save()
    clean = moe.mlp_output.save()
with model.trace(prompt):
    moe.expert_weights = moe.expert_weights.masked_fill(moe.expert_indices == 3, 0)
    ablated = moe.mlp_output.save()

assert torch.equal((ablated != clean).any(-1), (idx == 3).any(-1))

The dense blocks have no router: support() reports every mixture value missing there.

From the family module

04 · nnterp/families/ernie4_5_moe.py

ERNIE 4.5 MoE (Ernie4_5_MoeForCausalLM).

Llama's tree and Llama's block: model.{embed_tokens, layers[i].{input_layernorm, self_attn, post_attention_layernorm, mlp}, norm} and lm_head, the residual added in the block, attention through the shared eager forward. Blocks from moe_layer_start_index to moe_layer_end_index, every moe_layer_interval-th, have a mixture of experts (with optional shared experts, added before it returns); the others a dense MLP. The experts are moe_intermediate_size wide.