Notify all remote blocks of session end (clears their KV caches).

Returns the GPTQ group size if this is GPTQ quantization, None otherwise.

Main loop handling communication with the master.

atexit handler: terminates process before HIP's broken atexit cleanup runs. Registered after HIP init, so runs before HIP's handler (LIFO order).

SIGABRT handler: HIP runtime crashes in Monitor::tryLock during process exit because its atexit cleanup runs after threads are torn down. We catch the

Return the unique identity or local.

Serve the single-page UI.

The input modality this model accepts (always Text for now).

Check if a flat layer index is the first layer of its stack.

Check if a flat layer index is the last layer of its stack.

Extract the flat layer index from a name like "fm_decoder.layers.3".

Extract the layer index from a block name like "model.language_model.layers.3".

Layer normalization: `(x - mean) / sqrt(var + eps) * weight + bias`. Matches candle_nn::LayerNorm::forward() — uses fused kernel when contiguous + has

Linear layer forward: `x @ weight^T + bias`. Matches candle_nn::Linear::forward() semantics exactly: - For contiguous 3D/4D inputs: reshape to 2D → m