search

openai / transformer-debugger

MIT License
4.01k stars 231 forks source link

error with /attention_head_record endpoint #25

Open Ruibn opened 4 months ago

Ruibn commented 4 months ago

Hi Team,

I hit error when I tried following request body to /attention_head_record { "dst": "logits", "layerIndex": 9, "activationIndex": 8, "datasets": [ "https://openaipublic.blob.core.windows.net/neuron-explainer/gpt2_small_data/collated-activations/" ] }

Error as below: File "/home/ruibn/debug/transformer-debugger/neuron_explainer/activation_server/read_routes.py", line 265, in attention_head_record ) = convert_activation_records_to_token_and_attention_activations_lists( ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/ruibn/debug/transformer-debugger/neuron_explainer/activation_server/read_routes.py", line 191, in convert_activation_records_to_token_and_attention_activations_lists return normalize_attention_token_scalars(zipped_tokens_and_raw_attention_activations) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/ruibn/debug/transformer-debugger/neuron_explainer/activation_server/read_routes.py", line 455, in normalize_attention_token_scalars ) = compute_scalar_summary_and_scale(compute_max_scalar_out, list_of_sequence_lists) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/ruibn/debug/transformer-debugger/neuron_explainer/activation_server/read_routes.py", line 427, in compute_scalar_summary_and_scale scalar_indexed_by_token_sequence_list: list[list[list[float]]] = [ ^ File "/home/ruibn/debug/transformer-debugger/neuron_explainer/activation_server/read_routes.py", line 428, in [compute_scalar_summary_function(sequence) for sequence in sequence_list] File "/home/ruibn/debug/transformer-debugger/neuron_explainer/activation_server/read_routes.py", line 428, in [compute_scalar_summary_function(sequence) for sequence in sequence_list] ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/ruibn/debug/transformer-debugger/neuron_explainer/activation_server/read_routes.py", line 421, in compute_max_scalar_out return compute_summary_of_scalar_out(attn_token_sequence_list, _max_non_none) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/ruibn/debug/transformer-debugger/neuron_explainer/activation_server/read_routes.py", line 392, in compute_summary_of_scalar_out scalar_out = [ ^ File "/home/ruibn/debug/transformer-debugger/neuron_explainer/activation_server/read_routes.py", line 393, in operation( File "/home/ruibn/debug/transformer-debugger/neuron_explainer/activation_server/read_routes.py", line 380, in _max_non_none return max([a for a in scalars if a is not None])

Any idea? Same request was working fine on /neuron_record. Will it due to wrong dataset I used? @WuTheFWasThat