llm-architectures-explained

/models

Qwen2.5 72B

Alibaba Cloud (Qwen) · Qwen2.5 · open weights

Facts and where they come from

Released2024-09config.jsonconfig.jsonHugging Face repository creation date (api.createdAt)
Licenceotherconfig.jsonconfig.jsonREADME metadata: license
Total parameters72.7Blabmodel cardREADME: Number of Parameters: 72.7B
Active parametersnot disclosednot disclosed
Context length128K tokenslabmodel cardREADME: Context Length: 131,072 tokens
Norm placementprecodemodelling codetransformers 5.18.0 / repo modelling code for qwen2: input_layernorm before attention, post_attention_layernorm before the MLP
Norm typeRMSNormcodemodelling codetransformers 5.18.0 / repo modelling code for qwen2: 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: input_layernorm before attention, post_attention_layernorm before the MLP

Architecture, drawn from the data

GQA 64q/8kv. Each column is one layer: its token mixer above, its feed-forward block below. Paler columns reuse another layer’s keys and values.

Qwen2.5 72B: layer stack and blockslayers (80)mixer / FFNlayer 0: GQA: 64 query / 8 KV heads · head 128layer 0: Gated MLP: 29568layer 1: GQA: 64 query / 8 KV heads · head 128layer 1: Gated MLP: 29568layer 2: GQA: 64 query / 8 KV heads · head 128layer 2: Gated MLP: 29568layer 3: GQA: 64 query / 8 KV heads · head 128layer 3: Gated MLP: 29568layer 4: GQA: 64 query / 8 KV heads · head 128layer 4: Gated MLP: 29568layer 5: GQA: 64 query / 8 KV heads · head 128layer 5: Gated MLP: 29568layer 6: GQA: 64 query / 8 KV heads · head 128layer 6: Gated MLP: 29568layer 7: GQA: 64 query / 8 KV heads · head 128layer 7: Gated MLP: 29568layer 8: GQA: 64 query / 8 KV heads · head 128layer 8: Gated MLP: 29568layer 9: GQA: 64 query / 8 KV heads · head 128layer 9: Gated MLP: 29568layer 10: GQA: 64 query / 8 KV heads · head 128layer 10: Gated MLP: 29568layer 11: GQA: 64 query / 8 KV heads · head 128layer 11: Gated MLP: 29568layer 12: GQA: 64 query / 8 KV heads · head 128layer 12: Gated MLP: 29568layer 13: GQA: 64 query / 8 KV heads · head 128layer 13: Gated MLP: 29568layer 14: GQA: 64 query / 8 KV heads · head 128layer 14: Gated MLP: 29568layer 15: GQA: 64 query / 8 KV heads · head 128layer 15: Gated MLP: 29568layer 16: GQA: 64 query / 8 KV heads · head 128layer 16: Gated MLP: 29568layer 17: GQA: 64 query / 8 KV heads · head 128layer 17: Gated MLP: 29568layer 18: GQA: 64 query / 8 KV heads · head 128layer 18: Gated MLP: 29568layer 19: GQA: 64 query / 8 KV heads · head 128layer 19: Gated MLP: 29568layer 20: GQA: 64 query / 8 KV heads · head 128layer 20: Gated MLP: 29568layer 21: GQA: 64 query / 8 KV heads · head 128layer 21: Gated MLP: 29568layer 22: GQA: 64 query / 8 KV heads · head 128layer 22: Gated MLP: 29568layer 23: GQA: 64 query / 8 KV heads · head 128layer 23: Gated MLP: 29568layer 24: GQA: 64 query / 8 KV heads · head 128layer 24: Gated MLP: 29568layer 25: GQA: 64 query / 8 KV heads · head 128layer 25: Gated MLP: 29568layer 26: GQA: 64 query / 8 KV heads · head 128layer 26: Gated MLP: 29568layer 27: GQA: 64 query / 8 KV heads · head 128layer 27: Gated MLP: 29568layer 28: GQA: 64 query / 8 KV heads · head 128layer 28: Gated MLP: 29568layer 29: GQA: 64 query / 8 KV heads · head 128layer 29: Gated MLP: 29568layer 30: GQA: 64 query / 8 KV heads · head 128layer 30: Gated MLP: 29568layer 31: GQA: 64 query / 8 KV heads · head 128layer 31: Gated MLP: 29568layer 32: GQA: 64 query / 8 KV heads · head 128layer 32: Gated MLP: 29568layer 33: GQA: 64 query / 8 KV heads · head 128layer 33: Gated MLP: 29568layer 34: GQA: 64 query / 8 KV heads · head 128layer 34: Gated MLP: 29568layer 35: GQA: 64 query / 8 KV heads · head 128layer 35: Gated MLP: 29568layer 36: GQA: 64 query / 8 KV heads · head 128layer 36: Gated MLP: 29568layer 37: GQA: 64 query / 8 KV heads · head 128layer 37: Gated MLP: 29568layer 38: GQA: 64 query / 8 KV heads · head 128layer 38: Gated MLP: 29568layer 39: GQA: 64 query / 8 KV heads · head 128layer 39: Gated MLP: 29568layer 40: GQA: 64 query / 8 KV heads · head 128layer 40: Gated MLP: 29568layer 41: GQA: 64 query / 8 KV heads · head 128layer 41: Gated MLP: 29568layer 42: GQA: 64 query / 8 KV heads · head 128layer 42: Gated MLP: 29568layer 43: GQA: 64 query / 8 KV heads · head 128layer 43: Gated MLP: 29568layer 44: GQA: 64 query / 8 KV heads · head 128layer 44: Gated MLP: 29568layer 45: GQA: 64 query / 8 KV heads · head 128layer 45: Gated MLP: 29568layer 46: GQA: 64 query / 8 KV heads · head 128layer 46: Gated MLP: 29568layer 47: GQA: 64 query / 8 KV heads · head 128layer 47: Gated MLP: 29568layer 48: GQA: 64 query / 8 KV heads · head 128layer 48: Gated MLP: 29568layer 49: GQA: 64 query / 8 KV heads · head 128layer 49: Gated MLP: 29568layer 50: GQA: 64 query / 8 KV heads · head 128layer 50: Gated MLP: 29568layer 51: GQA: 64 query / 8 KV heads · head 128layer 51: Gated MLP: 29568layer 52: GQA: 64 query / 8 KV heads · head 128layer 52: Gated MLP: 29568layer 53: GQA: 64 query / 8 KV heads · head 128layer 53: Gated MLP: 29568layer 54: GQA: 64 query / 8 KV heads · head 128layer 54: Gated MLP: 29568layer 55: GQA: 64 query / 8 KV heads · head 128layer 55: Gated MLP: 29568layer 56: GQA: 64 query / 8 KV heads · head 128layer 56: Gated MLP: 29568layer 57: GQA: 64 query / 8 KV heads · head 128layer 57: Gated MLP: 29568layer 58: GQA: 64 query / 8 KV heads · head 128layer 58: Gated MLP: 29568layer 59: GQA: 64 query / 8 KV heads · head 128layer 59: Gated MLP: 29568layer 60: GQA: 64 query / 8 KV heads · head 128layer 60: Gated MLP: 29568layer 61: GQA: 64 query / 8 KV heads · head 128layer 61: Gated MLP: 29568layer 62: GQA: 64 query / 8 KV heads · head 128layer 62: Gated MLP: 29568layer 63: GQA: 64 query / 8 KV heads · head 128layer 63: Gated MLP: 29568layer 64: GQA: 64 query / 8 KV heads · head 128layer 64: Gated MLP: 29568layer 65: GQA: 64 query / 8 KV heads · head 128layer 65: Gated MLP: 29568layer 66: GQA: 64 query / 8 KV heads · head 128layer 66: Gated MLP: 29568layer 67: GQA: 64 query / 8 KV heads · head 128layer 67: Gated MLP: 29568layer 68: GQA: 64 query / 8 KV heads · head 128layer 68: Gated MLP: 29568layer 69: GQA: 64 query / 8 KV heads · head 128layer 69: Gated MLP: 29568layer 70: GQA: 64 query / 8 KV heads · head 128layer 70: Gated MLP: 29568layer 71: GQA: 64 query / 8 KV heads · head 128layer 71: Gated MLP: 29568layer 72: GQA: 64 query / 8 KV heads · head 128layer 72: Gated MLP: 29568layer 73: GQA: 64 query / 8 KV heads · head 128layer 73: Gated MLP: 29568layer 74: GQA: 64 query / 8 KV heads · head 128layer 74: Gated MLP: 29568layer 75: GQA: 64 query / 8 KV heads · head 128layer 75: Gated MLP: 29568layer 76: GQA: 64 query / 8 KV heads · head 128layer 76: Gated MLP: 29568layer 77: GQA: 64 query / 8 KV heads · head 128layer 77: Gated MLP: 29568layer 78: GQA: 64 query / 8 KV heads · head 128layer 78: Gated MLP: 29568layer 79: GQA: 64 query / 8 KV heads · head 128layer 79: Gated MLP: 2956804079× 80normGQA: 64 query / 8 KV heads · head 128+normGated MLP: 29568+full attentiondense 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)72.7B
Active per token (modelled)72.7B
Without embeddings and output head70.2B total, 70.2B active
Published weights (Hugging Face count)72.7B
KV cache per token, BF16 (layers that grow with context)320 KiB
KV cache + state at 128K tokens, BF1640 GiB
Decode FLOPs per token at 4K context154 GFLOP
Prefill FLOPs for a 4K prompt597 TFLOP

KV cache against context

Qwen2.5 72B: KV cache bytes against context length101001,00010,000100,000980 KiB9.5 MiB95 MiB950 MiB9.3 GiBcontext (tokens)KV cache + state (BF16)Qwen2.5 72B

Compare with other models →

Every architecture field

FieldValueSource
d_model8,192config.jsonconfig.jsonhidden_size
vocab152,064config.jsonconfig.jsonvocab_size
tied_embeddingsfalseconfig.jsonconfig.jsontie_word_embeddings
mixers.full.typeattncodemodelling codeattention block
mixers.full.heads64config.jsonconfig.jsonnum_attention_heads
mixers.full.kv_heads8config.jsonconfig.jsonnum_key_value_heads
mixers.full.head_dim128codemodelling codetransformers 5.18.0: head_dim = hidden_size / num_attention_heads
mixers.full.biastruecodemodelling codetransformers 5.18.0 qwen2: q/k/v projections have biases
ffns.dense.typedensecodemodelling codeMLP block
ffns.dense.d_ff29,568config.jsonconfig.jsonintermediate_size
ffns.dense.gatedtruecodemodelling codeMLP: gated (SwiGLU/GeGLU)
layout80× full/denseconfig.jsonconfig.jsonnum_hidden_layers

Sources