nnterp · encyclopedia solar_open · nnterp.families.solar_open

family solar_open

Solar Open

Llama's pre-norm block with a mixture of experts on every block: a sigmoid router whose correction bias affects only which experts are chosen, and a shared expert that runs on every token.

Mixture of expertsGrouped-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× 48
→
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 SolarOpenForCausalLMlayoutwhat 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] SolarOpenDecoderLayerlayoutwhat 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 SolarOpenAttentionlayoutwhat 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 SolarOpenMoElayoutwhat 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 upstage/Solar-Open-100B: the native modules, standard names and every standard value with its layout.

SolarOpenForCausalLM(
  (lm_head): Linear(in_features=4096, out_features=196608, bias=False)
  (generator): Generator(
    (streamer): Streamer()
  )
  (embed_tokens): Embedding(196608, 4096, padding_idx=2)
  (layers): ModuleList(
    (0-47): 48 x SolarOpenDecoderLayer(
      (self_attn): SolarOpenAttention(
        (q_proj): Linear(in_features=4096, out_features=8192, bias=False)
        (k_proj): Linear(in_features=4096, out_features=1024, bias=False)
        (v_proj): Linear(in_features=4096, out_features=1024, bias=False)
        (o_proj): Linear(in_features=8192, out_features=4096, 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): SolarOpenMoE(
        (experts): SolarOpenExperts(
          (act_fn): SiLUActivation()
        )
        (router/gate): SolarOpenTopkRouter()
        (shared_experts): SolarOpenMLP(
          (gate_proj): Linear(in_features=4096, out_features=1280, bias=False)
          (up_proj): Linear(in_features=4096, out_features=1280, bias=False)
          (down_proj): Linear(in_features=1280, out_features=4096, 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): SolarOpenRMSNorm((4096,), eps=1e-05)
      (post_attention_layernorm): SolarOpenRMSNorm((4096,), 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): SolarOpenRMSNorm((4096,), 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_layers48
model.hidden_size4096
model.num_heads64
model.num_kv_heads8
model.head_dim128
model.qk_head_dim128
model.vocab_size196608
model.intermediate_size10240
architecturesSolarOpenForCausalLM
num_hidden_layers48
hidden_size4096
intermediate_size10240
num_attention_heads64
num_key_value_heads8
head_dim128
vocab_size196608
max_position_embeddings131072
hidden_actsilu
rms_norm_eps1e-05
tie_word_embeddingsFalse
dtypebfloat16
first_k_dense_replace0
moe_intermediate_size1280
n_routed_experts128
num_experts_per_tok8
n_shared_experts1
routed_scaling_factor1.0
norm_topk_probTrue
attention_biasFalse
n_group1
topk_group1

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))    # 8 of 128 experts, plus the shared expert

Llama's pre-norm block and Llama's names. Nothing norms or scales a sublayer's output, so attention_output is o_proj's output and mlp_output the mixture's, shared expert included, and the identity is the plain sum, exact in float32 on the pinned tiny checkpoint:

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)

Every block is a mixture

Transformers builds SolarOpenMoE on every block; the config's first_k_dense_replace is 0, and SolarOpenMLP is only the shared expert's class. The experts are moe_intermediate_size (1280) wide, and so is the one shared expert (n_shared_experts × 1280). model.intermediate_size reads the config's intermediate_size, 10240, which no module of Solar-Open-100B has.

The router: sigmoid scores, a bias for the choice, weights renormalized

router (native gate) scores the 128 experts in float32 and takes their sigmoid. It adds e_score_correction_bias (a buffer) to choose the 8 experts and then reads the weights from the unbiased sigmoid, renormalizes them to sum to one and multiplies them by routed_scaling_factor, 1.0 on Solar-Open-100B. n_group and topk_group are 1, so no group limits the choice. router_logits is before the sigmoid:

moe = model.layers[1].mlp
with model.trace(prompt):
    logits = moe.router_logits.save()       # [batch, seq, num_experts]
    w = moe.expert_weights.save()           # [batch, seq, top_k]
    idx = moe.expert_indices.save()

bias = moe._module.gate.e_score_correction_bias
scores = logits.float().sigmoid()
chosen = (scores + bias).topk(moe.top_k, dim=-1).indices
assert torch.equal(chosen.sort(-1).values, idx.sort(-1).values)   # biased choice
weights = scores.gather(-1, idx)
torch.testing.assert_close((weights / weights.sum(-1, keepdim=True)).to(w.dtype), w)

The slots are not sorted by weight. A logit written at router_logits changes both the choice and the weights.

The shared expert runs on every token

mlp_output is routed_output + shared_expert_output. The shared expert (mlp.shared_experts) reads the same normed input as the router, and its output is added after the routed experts', so read shared_expert_output after routed_output in one trace. An expert ablation leaves it alone: zeroing expert e's weight where expert_indices == e changes mlp_output on exactly the tokens that chose e. The other 7 weights are not renormalized, so those tokens lose expert e's term and keep the others' as they were.

moe, e = model.layers[1].mlp, 3
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 == e, 0)
    ablated = moe.mlp_output.save()

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

Attention: YaRN rotary and grouped heads

Rotary covers the whole head (partial_rotary_factor 1.0) with YaRN, factor 2.0 over original_max_position_embeddings 65536. YaRN multiplies the rotary cosines and sines by 0.1 * ln(2) + 1 = 1.0693, so attention_queries and attention_keys are 1.0693 times the rotated projections, and the scores carry that factor twice (1.143), on top of 1 / sqrt(head_dim).

64 query heads of 128 share 8 key/value heads, so q_proj maps the 4096-wide stream to 8192, and an edit to one key or value head reaches the 8 query heads that read it. The attention interior needs attn_implementation="eager"; the default load runs sdpa.

The readout and the tokenizer

logits is lm_head.output, with no cap or scale; lm_head and embed_tokens are separate weights. The tokenizer prepends nothing, so model.input_ids is the prompt's tokens alone; <|endoftext|> (id 2) is the end of text.

From the family module

04 · nnterp/families/solar_open.py

Solar Open (SolarOpenForCausalLM).

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. The rotary covers partial_rotary_factor of each head (1.0, the whole head, on Solar-Open-100B). Every block's MLP is a mixture of experts: transformers builds SolarOpenMoE on every block whatever first_k_dense_replace says. It adds a shared expert (mlp.shared_experts, a SolarOpenMLP) before it returns; the experts are moe_intermediate_size wide. SolarOpenMLP is keyed to Mlp, so the shared expert carries its own mlp_output.