MCPcopy Create free account
hub / github.com/NVIDIA/FasterTransformer / compute

Method compute

examples/tensorflow/t5/utils/ft_decoding.py:252–281  ·  view source on GitHub ↗
(self, input_tokens, beam_width, max_seq_len, top_k=1, top_p = 0.0, beam_search_diversity_rate = 0.0,
                temperature = 1.0, len_penalty = 0.0, repetition_penalty = 1.0, random_seed=0)

Source from the content-addressed store, hash-verified

250 self.decoding_params = decoding_params
251
252 def compute(self, input_tokens, beam_width, max_seq_len, top_k=1, top_p = 0.0, beam_search_diversity_rate = 0.0,
253 temperature = 1.0, len_penalty = 0.0, repetition_penalty = 1.0, random_seed=0):
254
255 input_ids = tf.cast(input_tokens.input_ids, tf.int32) # maybe convert to int32
256
257 mem_seq_len = 0
258 if hasattr(input_tokens, "attention_mask"):
259 mem_seq_len = np.sum(input_tokens.attention_mask, axis=1)
260 else:
261 mem_seq_len = input_tokens.seq_len
262 mem_seq_len = tf.cast(mem_seq_len, tf.int32)
263
264 encoder_outputs = ftt5_encoder(input_ids, mem_seq_len, self.encoder_params)
265 ft_decoding_output_ids, ft_decoding_seq_lens, ft_output_log_probs, ft_cum_log_probs = ftt5_decoding(encoder_outputs,
266 mem_seq_len,
267 self.decoding_params,
268 max_seq_len,
269 beam_width,
270 top_k,
271 top_p,
272 beam_search_diversity_rate,
273 temperature,
274 len_penalty,
275 repetition_penalty,
276 random_seed=random_seed)
277
278 ft_decoding_output_ids = tf.reshape(ft_decoding_output_ids, [-1, beam_width, max_seq_len])
279 ft_decoding_seq_lens = tf.reshape(ft_decoding_seq_lens, [-1, beam_width])
280
281 return ft_decoding_output_ids.numpy(), ft_decoding_seq_lens.numpy()

Callers 5

translateFunction · 0.95
mainFunction · 0.80
mainFunction · 0.80
mainFunction · 0.80
xnli_taskFunction · 0.80

Calls 2

ftt5_encoderFunction · 0.90
ftt5_decodingFunction · 0.85

Tested by

no test coverage detected