Signal Reference

PyTorch signals exported by OneXMonitor.watch(model) (non-exhaustive). Use assessments=[...] to filter fields; when unset, full payloads are sent.

Per-forward signals

`signal_type`	When	Purpose
`representation.<policy>`	After each `forward` (causal / GPT-2)	OOD-style `embedding_vector` (`last_token`, `prompt_last`, or `mean_pool`)
`output_distribution`	After each `forward`	Entropy, top-k confidence, logit stats, optional `logits_extras`
`input_statistics`	After each `forward`	Input shape / norm / token stats
`llm.final_norm`	Final norm layer (causal LMs)	Pre-`lm_head` hidden state summaries
`llm.lm_head`	`lm_head` forward	Bounded logits summary

Per-layer signals (architecture-dependent)

Pattern	Description
`hidden_states.causal_decoder.layer_*`	Decoder block output (Qwen, Llama, etc.)
`hidden_states.bert_encoder.layer_*`	BERT encoder layers
`hidden_states.gpt_block.layer_*`	GPT-2 blocks
`attention.layer_*`	Attention weights (when returned by the model)
`embeddings.*`	Embedding layer outputs
`generic_activation.*`	Linear / LayerNorm statistics
`pre_classification.*`	Classifier / head inputs (BERT-style OOD)

`llm_extension` (when enabled)

Optional nested JSON on exported signals with generation context, output summaries, and model metadata. See Configuration Reference for related flags.

Qwen / PyTorch tips

Call monitor.watch() on AutoModelForCausalLM (not only model.model) so logits and representation use the full forward output.
For long generations, raise signal_queue_maxsize if the exporter drops signals.

Start Here

Models

AI Agents

Reference

Per-forward signals

Per-layer signals (architecture-dependent)

`llm_extension` (when enabled)

Qwen / PyTorch tips

​Per-forward signals

​Per-layer signals (architecture-dependent)

​llm_extension (when enabled)

​Qwen / PyTorch tips

Per-forward signals

Per-layer signals (architecture-dependent)

`llm_extension` (when enabled)

Qwen / PyTorch tips