nnterp · encyclopedia ernie4_5 · nnterp.families.ernie4_5

family ernie4_5

ERNIE 4.5

Llama's pre-norm block whose rotary turns adjacent pairs of query and key dimensions, (2i, 2i + 1), and whose grouped query heads are together wider than the residual stream.

Interleaved rotaryGrouped-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.

embed_tokenstoken_embeddings
→
layers× 18
→
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_5ForCausalLMlayoutwhat 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_5DecoderLayerlayoutwhat 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_5Attentionlayoutwhat 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_5MLPlayoutwhat 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

Printout

What print(model) shows on baidu/ERNIE-4.5-0.3B-Base-PT: the native modules, standard names and every standard value with its layout.

Ernie4_5ForCausalLM(
  (lm_head): Linear(in_features=1024, out_features=103424, bias=False)
  (generator): Generator(
    (streamer): Streamer()
  )
  (embed_tokens): Embedding(103424, 1024, padding_idx=0)
  (layers): ModuleList(
    (0-17): 18 x Ernie4_5DecoderLayer(
      (self_attn): Ernie4_5Attention(
        (q_proj): Linear(in_features=1024, out_features=2048, bias=False)
        (k_proj): Linear(in_features=1024, out_features=256, bias=False)
        (v_proj): Linear(in_features=1024, out_features=256, bias=False)
        (o_proj): Linear(in_features=2048, out_features=1024, 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_5MLP(
        (gate_proj): Linear(in_features=1024, out_features=3072, bias=False)
        (up_proj): Linear(in_features=1024, out_features=3072, bias=False)
        (down_proj): Linear(in_features=3072, out_features=1024, 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_5RMSNorm((1024,), eps=1e-05)
      (post_attention_layernorm): Ernie4_5RMSNorm((1024,), 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_5RMSNorm((1024,), 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_layers18
model.hidden_size1024
model.num_heads16
model.num_kv_heads2
model.head_dim128
model.qk_head_dim128
model.vocab_size103424
model.intermediate_size3072
architecturesErnie4_5ForCausalLM
num_hidden_layers18
hidden_size1024
intermediate_size3072
num_attention_heads16
num_key_value_heads2
head_dim128
vocab_size103424
max_position_embeddings131072
hidden_actsilu
rms_norm_eps1e-05
tie_word_embeddingsTrue
dtypebfloat16

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))

Llama's pre-norm block and Llama's names. Nothing norms a sublayer's output, so attention_output is o_proj's output and mlp_output the MLP's, and the identity is the plain sum:

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)

No projection has a bias (use_bias is false).

The heads are wider than the stream

0.3B has hidden_size 1024 and 16 query heads of head_dim 128, so q_proj maps 1024 to 2048 and o_proj 2048 back to 1024. The 16 query heads share 2 key/value heads, 8 to each: an edit to key head 0 at attention_keys reaches query heads 0 to 7. The attention interior needs attn_implementation="eager".

Rotary turns adjacent pairs

The rotary embedding rotates dimensions (2i, 2i + 1) of each query and key head together, at frequency rope_theta ** (-2i / head_dim) (rope_theta 500000), where Llama's rotates i with i + 64. The served queries and keys keep the projection's order: at position 0 attention_queries is q_proj's output split into heads, and at position 1 dimensions 0 and 1 are that pair turned by one radian:

attn = model.layers[1].self_attn
with model.trace(prompt):
    proj = attn.q_proj.output.save()
with model.trace(prompt):
    q = attn.attention_queries.save()          # [batch, num_heads, seq, head_dim]

p = proj.view(1, -1, model.num_heads, model.head_dim)
torch.testing.assert_close(q[0, :, 0], p[0, 0])
c, s = math.cos(1.0), math.sin(1.0)
a, b = p[0, 1, 0, 0], p[0, 1, 0, 1]
torch.testing.assert_close(q[0, 0, 1, :2], torch.stack([a * c - b * s, b * c + a * s]))

A probe or a patch that splits a head into a rotating half and its partner, as a rotate-half layout would, pairs the wrong dimensions here.

The readout

logits is lm_head.output, with no cap or scale, and project_on_vocab on the last block's layer_output equals logits. On 0.3B lm_head is embed_tokens's matrix (tie_word_embeddings).

From the family module

04 · nnterp/families/ernie4_5.py

ERNIE 4.5 dense (Ernie4_5ForCausalLM).

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.