GPT-2

All models, with previews →

OpenAI · Feb–Nov 2019Dense decoder

Masked multi-head self-attention

Reference code openai/gpt-2 @ 9b63575, huggingface/transformers @ 71c6f69, configuration openai-community/gpt2 @ 607a30d · checkpoint tensors not independently counted · How to read

Full sequenceTrainingPrefillDecode

Transformer block › Attention

Retrieved values by head

×

tensora · 110

Each head contains the attention-weighted combination of its value vectors, before merging.

shape
HiH_{i} head i [B batch, T positions, 64 d_h]HH all heads [B batch, 12 heads, T positions, 64 d_h]
Sources

Contextual evidence from Weighted value retrieval

The maintained eager attention path normalizes scores, applies training-dependent dropout, and multiplies the resulting weights by values.

transformers · huggingface/transformers / src/transformers/models/gpt2/modeling_gpt2.py
Pinned revision 71c6f699ac9b ↗ · lines 63–69

Read code excerpt
    attn_weights = nn.functional.softmax(attn_weights, dim=-1)

    # Downcast (if necessary) back to V's dtype (if in mixed-precision) -- No-Op otherwise
    attn_weights = attn_weights.type(value.dtype)
    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)

    attn_output = torch.matmul(attn_weights, value)

Contextual evidence from Weighted value retrieval

The normalized score matrix multiplies the value field inside each head.

openai · openai/gpt-2 / src/model.py
Pinned revision 9b63575ef427 ↗ · lines 91–99

Read code excerpt
        w = mask_attn_weights(w)
        w = softmax(w)
        a = tf.matmul(w, v)
        return a

Contextual evidence from Merge attention heads

The released merge_heads function reverses the head permutation and merges the final two axes.

openai · openai/gpt-2 / src/model.py
Pinned revision 9b63575ef427 ↗ · lines 79–81

Read code excerpt
    def merge_heads(x):

… lines 80–80 omitted …

        return merge_states(tf.transpose(x, [0, 2, 1, 3]))