llm-architectures-explained

/models

Qwen1.5-MoE-A2.7B

Alibaba Cloud (Qwen) · Qwen MoE · open weights

Facts and where they come from

Released2024-02config.jsonconfig.jsonHugging Face repository creation date (api.createdAt)
Licenceotherconfig.jsonconfig.jsonREADME metadata: license
Total parameters14.3Blabmodel cardREADME: 14.3B parameters in total and 2.7B activated
Active parameters2.7Blabmodel cardREADME: 2.7B activated
Context length8K tokensconfig.jsonconfig.jsonmax_position_embeddings
Norm placementprecodemodelling codetransformers 5.18.0 / repo modelling code for qwen2_moe: input_layernorm before attention, post_attention_layernorm before the MLP
Norm typeRMSNormcodemodelling codetransformers 5.18.0 / repo modelling code for qwen2_moe: input_layernorm before attention, post_attention_layernorm before the MLP
QK-normnocodemodelling codeno q/k normalisation in the attention block
Positional encodingRoPEcodemodelling coderotary on the full head (default)
Parallel attention and MLPnocodemodelling codetransformers 5.18.0 / repo modelling code for qwen2_moe: input_layernorm before attention, post_attention_layernorm before the MLP

Architecture, drawn from the data

MHA 16q/16kv. Each column is one layer: its token mixer above, its feed-forward block below. Paler columns reuse another layer’s keys and values.

Qwen1.5-MoE-A2.7B: layer stack and blockslayers (24)mixer / FFNlayer 0: MHA: 16 query / 16 KV heads · head 128layer 0: MoE: 60 experts, 4 active · expert 1408 · 1 sharedlayer 1: MHA: 16 query / 16 KV heads · head 128layer 1: MoE: 60 experts, 4 active · expert 1408 · 1 sharedlayer 2: MHA: 16 query / 16 KV heads · head 128layer 2: MoE: 60 experts, 4 active · expert 1408 · 1 sharedlayer 3: MHA: 16 query / 16 KV heads · head 128layer 3: MoE: 60 experts, 4 active · expert 1408 · 1 sharedlayer 4: MHA: 16 query / 16 KV heads · head 128layer 4: MoE: 60 experts, 4 active · expert 1408 · 1 sharedlayer 5: MHA: 16 query / 16 KV heads · head 128layer 5: MoE: 60 experts, 4 active · expert 1408 · 1 sharedlayer 6: MHA: 16 query / 16 KV heads · head 128layer 6: MoE: 60 experts, 4 active · expert 1408 · 1 sharedlayer 7: MHA: 16 query / 16 KV heads · head 128layer 7: MoE: 60 experts, 4 active · expert 1408 · 1 sharedlayer 8: MHA: 16 query / 16 KV heads · head 128layer 8: MoE: 60 experts, 4 active · expert 1408 · 1 sharedlayer 9: MHA: 16 query / 16 KV heads · head 128layer 9: MoE: 60 experts, 4 active · expert 1408 · 1 sharedlayer 10: MHA: 16 query / 16 KV heads · head 128layer 10: MoE: 60 experts, 4 active · expert 1408 · 1 sharedlayer 11: MHA: 16 query / 16 KV heads · head 128layer 11: MoE: 60 experts, 4 active · expert 1408 · 1 sharedlayer 12: MHA: 16 query / 16 KV heads · head 128layer 12: MoE: 60 experts, 4 active · expert 1408 · 1 sharedlayer 13: MHA: 16 query / 16 KV heads · head 128layer 13: MoE: 60 experts, 4 active · expert 1408 · 1 sharedlayer 14: MHA: 16 query / 16 KV heads · head 128layer 14: MoE: 60 experts, 4 active · expert 1408 · 1 sharedlayer 15: MHA: 16 query / 16 KV heads · head 128layer 15: MoE: 60 experts, 4 active · expert 1408 · 1 sharedlayer 16: MHA: 16 query / 16 KV heads · head 128layer 16: MoE: 60 experts, 4 active · expert 1408 · 1 sharedlayer 17: MHA: 16 query / 16 KV heads · head 128layer 17: MoE: 60 experts, 4 active · expert 1408 · 1 sharedlayer 18: MHA: 16 query / 16 KV heads · head 128layer 18: MoE: 60 experts, 4 active · expert 1408 · 1 sharedlayer 19: MHA: 16 query / 16 KV heads · head 128layer 19: MoE: 60 experts, 4 active · expert 1408 · 1 sharedlayer 20: MHA: 16 query / 16 KV heads · head 128layer 20: MoE: 60 experts, 4 active · expert 1408 · 1 sharedlayer 21: MHA: 16 query / 16 KV heads · head 128layer 21: MoE: 60 experts, 4 active · expert 1408 · 1 sharedlayer 22: MHA: 16 query / 16 KV heads · head 128layer 22: MoE: 60 experts, 4 active · expert 1408 · 1 sharedlayer 23: MHA: 16 query / 16 KV heads · head 128layer 23: MoE: 60 experts, 4 active · expert 1408 · 1 shared01223× 24normMHA: 16 query / 16 KV heads · head 128+normMoE: 60 experts, 4 active · expert 1408 · 1 shared+full attentionMoE FFN

Modelled costs

From the cost model, batch size 1. Totals the lab states are in the table above; differences come from rounding, from what a lab counts, or from parts the model does not describe (listed on the about page).

Parameters (modelled)14.3B
Active per token (modelled)2.69B
Without embeddings and output head13.7B total, 2.07B active
Published weights (Hugging Face count)14.3B
KV cache per token, BF16 (layers that grow with context)192 KiB
KV cache + state at 8K tokens, BF161.5 GiB
Decode FLOPs per token at 4K context5.56 GFLOP
Prefill FLOPs for a 4K prompt18.6 TFLOP

KV cache against context

Qwen1.5-MoE-A2.7B: KV cache bytes against context length101001,000980 KiB9.5 MiB95 MiB950 MiBcontext (tokens)KV cache + state (BF16)Qwen1.5-MoE-A2.7B

Compare with other models →

Every architecture field

FieldValueSource
d_model2,048config.jsonconfig.jsonhidden_size
vocab151,936config.jsonconfig.jsonvocab_size
tied_embeddingsfalseconfig.jsonconfig.jsontie_word_embeddings
mixers.full.typeattncodemodelling codeattention block
mixers.full.heads16config.jsonconfig.jsonnum_attention_heads
mixers.full.kv_heads16config.jsonconfig.jsonnum_key_value_heads
mixers.full.head_dim128codemodelling codetransformers 5.18.0: head_dim = hidden_size / num_attention_heads
mixers.full.biastruecodemodelling codeqwen2_moe: qkv biases
ffns.moe.typemoecodemodelling codeMoE block
ffns.moe.experts60config.jsonconfig.jsonnum_experts
ffns.moe.active4config.jsonconfig.jsonnum_experts_per_tok
ffns.moe.d_expert1,408config.jsonconfig.jsonmoe_intermediate_size
ffns.moe.gatedtruecodemodelling codeexperts are gated MLPs
ffns.moe.shared1codemodelling codeone shared expert of shared_expert_intermediate_size
ffns.moe.d_shared5,632config.jsonconfig.jsonshared_expert_intermediate_size
layout24× full/moeconfig.jsonconfig.jsonnum_hidden_layers

Sources