kokoro-sima / runtime /albert_inputs.py
florianvoss's picture
Add minimal Kokoro Modalix runtime, CLI, webapp and speech API
dc5b70c verified
Raw History Blame Contribute Delete
1.3 kB
"""CPU embedding lookup and static-mask contract for Kokoro ALBERT."""
import numpy as np
def prepare_inputs(token_ids, embeddings, capacity=480):
ids = np.asarray(token_ids)
if ids.ndim != 1 or not np.issubdtype(ids.dtype, np.integer):
raise ValueError("Expected a one-dimensional integer token sequence")
if not 1 <= len(ids) <= capacity:
raise ValueError(f"Expected 1..{capacity} tokens, including boundary tokens")
if np.any(ids < 0) or np.any(ids >= len(embeddings['word'])):
raise ValueError("Token ID outside vocabulary")
padded = np.zeros(capacity, dtype=np.int64)
padded[:len(ids)] = ids
summed = embeddings['word'][padded] + embeddings['type'][0]
summed = summed + embeddings['position'][:capacity]
mask = np.full((1, capacity, 1, 1), -10000, dtype=np.float32)
mask[:, :len(ids)] = 0
return {'embedding_sum': np.ascontiguousarray(summed.T[None, :, None, :]),
'attention_mask': mask}
def unpack_output(hidden_states, valid_tokens):
"""Remove padding BEFORE the downstream CPU projection/LSTMs."""
if not 1 <= valid_tokens <= hidden_states.shape[-1]:
raise ValueError("Invalid valid-token count")
return np.ascontiguousarray(hidden_states[:, :, 0, :valid_tokens].transpose(0, 2, 1))