def test_top_k_dist_warper(self):
input_ids = None
vocab_size = 10
batch_size = 2
# create ramp distribution
ramp_logits = np.broadcast_to(np.arange(vocab_size)[None, :], (batch_size, vocab_size)).copy()
ramp_logits[1:, : vocab_size // 2] = ramp_logits[1:, : vocab_size // 2] + vocab_size
top_k_warp = FlaxTopKLogitsWarper(3)
scores = top_k_warp(input_ids, ramp_logits, cur_len=None)
# check that correct tokens are filtered
self.assertListEqual(jnp.isinf(scores[0]).tolist(), 7 * [True] + 3 * [False])
self.assertListEqual(jnp.isinf(scores[1]).tolist(), 2 * [True] + 3 * [False] + 5 * [True])
# check special case
length = 5
top_k_warp_safety_check = FlaxTopKLogitsWarper(top_k=1, filter_value=0.0, min_tokens_to_keep=3)
ramp_logits = np.broadcast_to(np.arange(length)[None, :], (batch_size, length)).copy()
scores = top_k_warp_safety_check(input_ids, ramp_logits, cur_len=None)
# min_tokens overwrites k: 3 tokens are kept => 2 tokens are nullified
self.assertListEqual((scores == 0.0).sum(axis=-1).tolist(), [2, 2])
def test_processor_list(self):
batch_size = 4
sequence_length = 10
vocab_size = 15
# dummy input_ids and scores
input_ids = ids_tensor((batch_size, sequence_length), vocab_size)
input_ids_comp = input_ids.copy()
scores = self._get_uniform_logits(batch_size, vocab_size)
scores_comp = scores.copy()
# instantiate all dist processors
temp_dist_warp = FlaxTemperatureLogitsWarper(temperature=0.5)
top_k_warp = FlaxTopKLogitsWarper(3)
top_p_warp = FlaxTopPLogitsWarper(0.8)
# no processor list
scores = temp_dist_warp(input_ids, scores)
scores = top_k_warp(input_ids, scores)
scores = top_p_warp(input_ids, scores)
# with processor list
processor = FlaxLogitsProcessorList(
[temp_dist_warp, top_k_warp, top_p_warp])
scores_comp = processor(input_ids, scores_comp)
# scores should be equal
self.assertTrue(jnp.allclose(scores, scores_comp, atol=1e-3))
# input_ids should never be changed
self.assertListEqual(input_ids.tolist(), input_ids_comp.tolist())
def test_processor_list_jitted(self):
batch_size = 4
sequence_length = 10
vocab_size = 15
eos_token_id = 2
bos_token_id = 1
max_length = 15
# dummy input_ids and scores
input_ids = ids_tensor((batch_size, sequence_length), vocab_size)
input_ids_comp = input_ids.copy()
scores = self._get_uniform_logits(batch_size, vocab_size)
scores_comp = scores.copy()
# instantiate all dist processors
temp_dist_warp = FlaxTemperatureLogitsWarper(temperature=0.5)
top_k_warp = FlaxTopKLogitsWarper(3)
top_p_warp = FlaxTopPLogitsWarper(0.8)
# instantiate all logits processors
min_dist_proc = FlaxMinLengthLogitsProcessor(min_length=10, eos_token_id=eos_token_id)
bos_dist_proc = FlaxForcedBOSTokenLogitsProcessor(bos_token_id=bos_token_id)
eos_dist_proc = FlaxForcedEOSTokenLogitsProcessor(max_length=max_length, eos_token_id=eos_token_id)
cur_len = 10
# no processor list
def run_no_processor_list(input_ids, scores, cur_len):
scores = temp_dist_warp(input_ids, scores, cur_len=cur_len)
scores = top_k_warp(input_ids, scores, cur_len=cur_len)
scores = top_p_warp(input_ids, scores, cur_len=cur_len)
scores = min_dist_proc(input_ids, scores, cur_len=cur_len)
scores = bos_dist_proc(input_ids, scores, cur_len=cur_len)
scores = eos_dist_proc(input_ids, scores, cur_len=cur_len)
return scores
# with processor list
def run_processor_list(input_ids, scores, cur_len):
processor = FlaxLogitsProcessorList(
[temp_dist_warp, top_k_warp, top_p_warp, min_dist_proc, bos_dist_proc, eos_dist_proc]
)
scores = processor(input_ids, scores, cur_len=cur_len)
return scores
jitted_run_no_processor_list = jax.jit(run_no_processor_list)
jitted_run_processor_list = jax.jit(run_processor_list)
scores = jitted_run_no_processor_list(input_ids, scores, cur_len)
scores_comp = jitted_run_processor_list(input_ids, scores_comp, cur_len)
# scores should be equal
self.assertTrue(jnp.allclose(scores, scores_comp, atol=1e-3))
# input_ids should never be changed
self.assertListEqual(input_ids.tolist(), input_ids_comp.tolist())