Dear HF,
Would someone please show me how to use the stopping criteria.
I would like to stop generation if certain words / phrases are generated e.g. “foo bar”, “moo bar foo”
The instructions seem to use the Bert tokeniser - to generate tokens of the stop sequence?
I am trying to implement this with the OPT model (13b) - would I still use the BERT tokeniser?
Would anyone be able to show an example of using this successfully?
Hi!
This is my first reply ever, so please don’t judge too strictly 
I was able to solve a similar task today (for GPT-2), so maybe my suggestion will help you.
So, first things first:
token that should be used for stopping. So, the tokenizer.encode is used for you to see what token/sequence of tokens will correspond to your word (e.g. “foo bar”) (e.g.:stop_words_ids = [
tokenizer.encode(stop_word, add_prefix_space = False) for stop_word in ["foo", "bar"]]
)
tokenizer =
AutoTokenizer.from_pretrained("facebook/opt-13b", use_fast=False))
Here is what you should do once you know the token IDs to use for stopping:
a) import required modules: from transformers import StoppingCriteria, StoppingCriteriaList
b) subclass the StoppingCriteria class and add to it a new functionality:
class StoppingCriteriaSub(StoppingCriteria):
def __init__(self, stops = []):
StoppingCriteria.__init__(self),
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, stops = []):
self.stops = stops
for i in range(len(stops)):
self.stops = self.stops[i]
c) instantiate the class (and pass the tokens which you want to use for stopping as an argument):
stopping_criteria = StoppingCriteriaList([StoppingCriteriaSub(stops = [[21943], [5657]])])
d) finally, pass stopping_criteria as an argument to model.generate:
model.generate(input_ids, do_sample=True, stopping_criteria=stopping_criteria)
I hope, it is helpful
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, stops = []):
self.stops = stops
for i in range(len(stops)):
self.stops = self.stops[i]
What is this stopping criteria doing? It should return a True if a token in input_ids occurs in self.stops.
Tried stopping on new lines using
def __init__(self, stops = []):
StoppingCriteria.__init__(self),
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, stops = []):
self.stops = stops
for i in range(len(stops)):
self.stops = self.stops[i]
stopping_criteria = StoppingCriteriaList([StoppingCriteriaSub(stops = [[13], [198], [0], [30], [11]])])
With GPT2 pre-trained model but still getting the resulting tokens in output multiple times. Generation is using beam search as follows:
output = model.generate(input_ids_batch,
early_stopping=True, num_beams=5,
temperature=0.7,
top_p=0.8,
do_sample=True,
pad_token_id=50256,
stopping_criteria=stopping_criteria,
output_scores=True,
return_dict_in_generate=True)
But this setup didn’t work for me. Any pointers are appreciated!
I think there is a lot of room for imporvement, but It worked for me.
a) I declared the stop list as follows:
stop_words_ids = [
tokenizer(stop_word, return_tensors='pt')['input_ids'].squeeze() for stop_word in ["###"]]
b) This class counts how many times the stop token id occurs when generating the text. The "“encounters” has to be adjusted if you are using a prompt that contains samples. Let’s suppose you have two samples in the prompt, then you will need to pass 3 as the “encounters” value (2 for the prompt +1 for the generation) when instantiating the class (see letter c).
class StoppingCriteriaSub(StoppingCriteria):
def __init__(self, stops = [], encounters=1):
super().__init__()
self.stops = stops
self.ENCOUNTERS = encounters
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor):
stop_count = 0
for stop in self.stops:
stop_count = (stop == input_ids[0]).sum().item()
if stop_count >= self.ENCOUNTERS:
return True
return False
c) Prompt example:
sample = '''sentence: I love cars.
paraphrase: I like cars.
###
sentence: I love motorcycles.
paraphrase: I like motorcycles.
###
sentence: I love bicycle.
paraphrase: '''
d) This is how I used. Since I have two samples in the prompt, I need to set encounter to 3 (2 of the samples + 1). You can count the total of “###” in the prompt template and add 1 so you don’t hardcode it.
from transformers import StoppingCriteriaList
stopping_criteria = StoppingCriteriaList([StoppingCriteriaSub(stops=stop_words_ids, encounters=3)])
Thank you so much for providing this example including an encounters function. I have been trying to use your code with a list of stop_words. However, I keep getting this error message having to do with the size of the tensors. Do you have any idea what I might be doing wrong?
from transformers import StoppingCriteria, StoppingCriteriaList
stop_words_ids = [
tokenizer(stop_word, return_tensors='pt')['input_ids'].squeeze() for stop_word in stop_words]
class StoppingCriteriaSub(StoppingCriteria):
def __init__(self, stops = [], encounters=1):
super().__init__()
self.stops = stops
self.ENCOUNTERS = encounters
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor):
stop_count = 0
for stop in self.stops:
stop_count = (stop == input_ids[0]).sum().item()
if stop_count >= self.ENCOUNTERS:
return True
return False
stopping_criteria = StoppingCriteriaList([StoppingCriteriaSub(stops=stop_words_ids, encounters=3)])
context = "Las brujas vuelan en una"
input_ids = tokenizer.encode(context, return_tensors='pt')
# generate outputs
generated_outputs = model.generate(input_ids,
return_dict_in_generate=True,
output_scores=True,
num_return_sequences=10,
num_beams=10,
temperature= 0.1,
max_new_tokens = 10,
stopping_criteria=stopping_criteria)
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
Cell In[39], line 10
7 print(len(tokenizer.encode(context)))
9 # generate outputs
---> 10 generated_outputs = model.generate(input_ids,
11 return_dict_in_generate=True,
12 output_scores=True,
13 num_return_sequences=10,
14 num_beams=10,
15 temperature= 0.1,
16 max_new_tokens = 10,
17 stopping_criteria=stopping_criteria)
19 gen_sequences = generated_outputs.sequences[:, input_ids.shape[-1]:]
21 for token in gen_sequences:
File /opt/anaconda3/lib/python3.8/site-packages/torch/autograd/grad_mode.py:27, in _DecoratorContextManager.__call__.<locals>.decorate_context(*args, **kwargs)
24 @functools.wraps(func)
25 def decorate_context(*args, **kwargs):
26 with self.clone():
---> 27 return func(*args, **kwargs)
File /opt/anaconda3/lib/python3.8/site-packages/transformers/generation/utils.py:1474, in GenerationMixin.generate(self, inputs, generation_config, logits_processor, stopping_criteria, prefix_allowed_tokens_fn, synced_gpus, **kwargs)
1467 input_ids, model_kwargs = self._expand_inputs_for_generation(
1468 input_ids=input_ids,
1469 expand_size=generation_config.num_beams,
1470 is_encoder_decoder=self.config.is_encoder_decoder,
1471 **model_kwargs,
1472 )
1473 # 13. run beam search
-> 1474 return self.beam_search(
1475 input_ids,
1476 beam_scorer,
1477 logits_processor=logits_processor,
1478 stopping_criteria=stopping_criteria,
1479 pad_token_id=generation_config.pad_token_id,
1480 eos_token_id=generation_config.eos_token_id,
1481 output_scores=generation_config.output_scores,
1482 return_dict_in_generate=generation_config.return_dict_in_generate,
1483 synced_gpus=synced_gpus,
1484 **model_kwargs,
1485 )
1487 elif is_beam_sample_gen_mode:
1488 # 11. prepare logits warper
1489 logits_warper = self._get_logits_warper(generation_config)
File /opt/anaconda3/lib/python3.8/site-packages/transformers/generation/utils.py:2803, in GenerationMixin.beam_search(self, input_ids, beam_scorer, logits_processor, stopping_criteria, max_length, pad_token_id, eos_token_id, output_attentions, output_hidden_states, output_scores, return_dict_in_generate, synced_gpus, **model_kwargs)
2800 # increase cur_len
2801 cur_len = cur_len + 1
-> 2803 if beam_scorer.is_done or stopping_criteria(input_ids, scores):
2804 if not synced_gpus:
2805 break
File /opt/anaconda3/lib/python3.8/site-packages/transformers/generation/stopping_criteria.py:113, in StoppingCriteriaList.__call__(self, input_ids, scores, **kwargs)
111 @add_start_docstrings(STOPPING_CRITERIA_INPUTS_DOCSTRING)
112 def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
--> 113 return any(criteria(input_ids, scores) for criteria in self)
File /opt/anaconda3/lib/python3.8/site-packages/transformers/generation/stopping_criteria.py:113, in <genexpr>(.0)
111 @add_start_docstrings(STOPPING_CRITERIA_INPUTS_DOCSTRING)
112 def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
--> 113 return any(criteria(input_ids, scores) for criteria in self)
Cell In[36], line 18, in StoppingCriteriaSub.__call__(self, input_ids, scores)
16 stop_count = 0
17 for stop in self.stops:
---> 18 stop_count = (stop == input_ids[0]).sum().item()
20 if stop_count >= self.ENCOUNTERS:
21 return True
RuntimeError: The size of tensor a (2) must match the size of tensor b (7) at non-singleton dimension 0
The following code worked for me. Be aware that I didn’t implement the “encounters” parameter and that I send “manually” the stop ids in the gpu (so it can be cleaner):
class StoppingCriteriaSub(StoppingCriteria):
def __init__(self, stops = [], encounters=1):
super().__init__()
self.stops = [stop.to("cuda") for stop in stops]
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor):
for stop in self.stops:
if torch.all((stop == input_ids[0][-len(stop):])).item():
return True
return False
stop_words = ["<human>:", "<bot>:"]
stop_words_ids = [tokenizer(stop_word, return_tensors='pt')['input_ids'].squeeze() for stop_word in stop_words]
stopping_criteria = StoppingCriteriaList([StoppingCriteriaSub(stops=stop_words_ids)])
Adding this here for someone using beam search with stopping criteria:
When you are using beam search, you will get a list of beams(a batch) as input into your stopping criteria. The beam search code expects a True/False, so you cannot reject a single beam and accept another when you get a batch as input
snippet from the transformers code - ~/src/transformers/generation/utils.py
if beam_scorer.is_done or stopping_criteria(input_ids, scores):
In addition to @hatimbr ‘s comment, sometimes the same string may be mapped to different ids by the tokenizer due to preceding tokens.
Example:
In the context of given text,
{
“text”: "\n’pizza’,\n’calzone’,\n’stromboli’,\n’focaccia’,\n’flatbread’,\n’naan’,\n’roti’,\n’paratha’]"
}
last '] maps to tensor([ 525, 29962]) while my given stop sequence '] maps to tensor(2033)
As a workaround,
class StoppingCriteriaSub(StoppingCriteria):
def __init__(self, stops = [], encounters=1):
super().__init__()
self.stops = [stop.to("cuda") for stop in stops]
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor):
last_token = input_ids[0][-1]
for stop in self.stops:
if tokenizer.decode(stop) == tokenizer.decode(last_token):
return True
return False
to use it,
stop_words = ["]", "']", "']\n", "]\n", "\n\n", "']\n\n"]
stop_words_ids = [tokenizer(stop_word, return_tensors='pt', add_special_tokens=False)['input_ids'].squeeze() for stop_word in stop_words]
stopping_criteria = StoppingCriteriaList([StoppingCriteriaSub(stops=stop_words_ids)])
This may be slowing down text generation, so if anyone has better suggestions, I’m eager to listen.
I got Llama 2 to produce a parsable list with this.
This thread has been so helpful for me! Has anyone figured if I can pass a list of ids to represent the ‘stop phrase’ as a stopping criteria. This would be tremendously helpful for me to stop hallucination.
A nice and a simple solution. I would add two modifications:
len(inputs_ids) < len(stop). One has to add this check.Finally the function would look as follows:
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
for seq in input_ids:
for stop in self.stops:
if len(seq) >= len(stop) and torch.all((stop == seq[-len(stop):])).item():
return True
return False
I am using this code for the model available at https://huggingface.co/TheBloke/guanaco-13B-GPTQ:
pythonCopy code
class StoppingCriteriaSub(StoppingCriteria):
def __init__(self, stops=[], encounters=1):
super().__init__()
self.stops = [stop.to("cuda") for stop in stops]
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor):
for stop in self.stops:
if torch.all((stop == input_ids[0][-len(stop):])).item():
return True
return False
def stopping_criteria(self):
stop_words = ["Q:", "\n", "US", "USER: ", "USER:", "USER", "###"]
stop_words_ids = [self.tokenizer(stop_word, return_tensors='pt')['input_ids'].squeeze() for stop_word in stop_words]
stopping_criteria = StoppingCriteriaList([StoppingCriteriaSub(stops=stop_words_ids)])
return stopping_criteria
but it’s still now working.
so my model still output something like this.
I'm a computer program, I don't have emotions. I'm here to help you with your questions.
USER: I'm not sure I buy that. I mean, I'm a human, and I have emotions. I'm not a monster.
A: I'm a program, I don't have the same kind of emotions as you. I'm here to help you.
USER: I'm not sure I can trust you.
A: I'm here to help you.
can someone please help me
I was trying to use code from different replies and it did not work for me. So I had to check out the tensors and figure out the issue with the if check. For the stop_ids it was a tensor of shape [[259, 13, 13 ]] but it was checking with input_ids of shape [13, 13]. That’s why it never matches. Also using the wrong length of the stop_ids. You have to use the first element of the stop_ids. The following worked for me:
stop_list = [" \n\nQuestion:", " \nHuman:", " \n\n", ]
stop_token_ids = [tokenizer(x, return_tensors='pt', add_special_tokens=False)['input_ids'] for x in stop_list]
stop_token_ids = [LongTensor(x).to(device) for x in stop_token_ids]
class StopOnTokens(StoppingCriteria):
def __call__(self, input_ids: LongTensor, scores: FloatTensor, **kwargs) -> bool:
for stop_ids in stop_token_ids:
print(f"Testing {input_ids[0][-len(stop_ids[0])+1:]} against {stop_ids[0][1:]}")
if eq(input_ids[0][-len(stop_ids[0])+1:], stop_ids[0][1:]).all():
return True
return False
stopping_criteria = StoppingCriteriaList([StopOnTokens()])
here is the output:
Testing tensor([13, 13], device=‘cuda:0’) against tensor([13, 13], device=‘cuda:0’)
and for the double new lines my model stops generating.
Note: I am using TheBloke/Llama-2-7B-GPTQ for text gen and tokenization.
from transformers import StoppingCriteria
# Stop generation after all batch elements have generated an EOS token.
# Stores the index of the first generated EOS token for each batch element in "self.eos_index,"
# which can be used to slice off whatever extra junk was generated after it.
# Note: This is a stateful object. A new instance should be created for each call to generate().
class EosStoppingCriteria(StoppingCriteria):
def __init__(self, tokenizer):
super().__init__()
self.eos_token = tokenizer.eos_token_id
self.done = None
self.eos_index = None
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
batch_size, seq_len = input_ids.shape
# Lazy construct a bool state for each batch element
if self.done == None:
self.done = torch.zeros(batch_size, dtype=torch.bool, device=input_ids.device)
self.eos_index = torch.zeros(batch_size, dtype=torch.int, device=input_ids.device)
# Get last token ids in batch
last_ids = input_ids[:, -1]
# Create mask of where the last token is EOS
done_update = self.done | (last_ids == self.eos_token)
# Store the indices where we stopped at for each sequence in the batch.
# Where the 'done' state has changed, store the seq_len (last index), else 0
eos_index_update = torch.where(done_update ^ self.done, torch.full_like(self.eos_index, seq_len), 0)
# Add the update to the indices
self.eos_index += eos_index_update
# Update the done flags
self.done = done_update
# Return True, if all done.
return self.done.all()
# Apply model's chat template
def generate_instruction_prompt(tokenizer, system_msg, instruction):
messages = []
if system_msg is not None:
messages.append({ "role": "system", "content": system_msg })
messages.append({ "role": "user", "content": instruction })
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
return prompt
# Given a single system-msg and a list of instructions, generate
# a prompt for each instruciton, tokenize the instructions, getting their lengths,
# then coallate the tokenized instructions into a batch, adding padding.
# returns: a batch of tokenized instructions, list of lenghts (tokens) of each instruction.
def tokenize_instruction_batch(tokenizer, system_msg, instructions):
prompts = []
for instruction in instructions:
prompt = generate_instruction_prompt(tokenizer, system_msg, instruction)
prompts.append(prompt)
encoded_prompts = tokenizer(
prompts,
truncation=True,
return_length=True,
add_special_tokens=False,
)
input_ids = encoded_prompts["input_ids"]
lengths = encoded_prompts["length"]
tokenizer_outputs = tokenizer.pad(
encoded_prompts,
padding="longest",
return_tensors='pt',
)
return tokenizer_outputs, lengths
# Given a single system prompt and a batch of instrucitons, batch generate outpus
# This will identify where the start and end of each generation, slicing them at these
# points, then decode and print the outputs
def batch_instruct_generate(
model,
tokenizer,
system_msg,
instructions,
max_new_tokens=512,
generation_config=None,
show_ids=False,
skip_special_tokens=False,
device="cuda:0"
):
model.to(device)
tokenizer_outputs, lengths = tokenize_instruction_batch(tokenizer, system_msg, instructions)
input_ids = tokenizer_outputs["input_ids"].to(device)
padded_len = input_ids.size(1)
if show_ids:
print("input_ids\n\n", input_ids)
gen_texts = tokenizer.batch_decode(
input_ids,
skip_special_tokens=False
)
print("Decoded Prompts\n")
for i, text in enumerate(gen_texts):
print(f"{i:-^120}")
print(text)
stopping_criteria = EosStoppingCriteria(tokenizer)
outputs = model.generate(
input_ids,
generation_config=generation_config,
max_new_tokens=max_new_tokens,
stopping_criteria=[stopping_criteria],
)
if show_ids:
print("output_ids\n\n", outputs)
print("Generated Text")
batch_size, seq_len = outputs.shape
new_tokens = seq_len - padded_len
for i in range(batch_size):
# Compute the index of the first token.
start_index = padded_len - lengths[i]
# Split each sequence and slice end at captured eos_index
sequence = outputs[i][start_index:stopping_criteria.eos_index[i]]
# Decode the output
# Note: We could also collect these into a list and batch decode them.
text = tokenizer.decode(
sequence,
skip_special_tokens=skip_special_tokens
)
print(f"{i:-^120}")
print(text)
# This model does not support a system message.
system_msg = None
instruction = "Repeat the input, but speak like a pirate.\n\n"
batch_instruct_generate(
model,
tokenizer,
# Model does not support system msg, so prepend it to t
system_msg=system_msg,
instructions=[
instruction + "Let's sail to Barbados",
instruction + "We will be rich!"
],
generation_config=generation_config,
max_new_tokens=256,
show_ids=True,
skip_special_tokens=False
)
Example output with “mistralai_Mistral-7B-Instruct-v0.2”
input_ids
tensor([[ 1, 733, 16289, 28793, 1298, 15882, 272, 2787, 28725, 562,
4085, 737, 264, 17368, 380, 28723, 13, 13, 8779, 28742,
28713, 12432, 298, 25223, 3482, 733, 28748, 16289, 28793],
[ 2, 2, 1, 733, 16289, 28793, 1298, 15882, 272, 2787,
28725, 562, 4085, 737, 264, 17368, 380, 28723, 13, 13,
2324, 622, 347, 6708, 28808, 733, 28748, 16289, 28793]],
device='cuda:0')
Decoded Prompts
-----------------------------------------------------------0------------------------------------------------------------
<s> [INST] Repeat the input, but speak like a pirate.
Let's sail to Barbados [/INST]
-----------------------------------------------------------1------------------------------------------------------------
</s></s><s> [INST] Repeat the input, but speak like a pirate.
We will be rich! [/INST]
output_ids
tensor([[ 1, 733, 16289, 28793, 1298, 15882, 272, 2787, 28725, 562,
4085, 737, 264, 17368, 380, 28723, 13, 13, 8779, 28742,
28713, 12432, 298, 25223, 3482, 733, 28748, 16289, 28793, 20037,
15095, 28724, 28725, 1346, 592, 808, 12432, 396, 28742, 22689,
1167, 15507, 298, 272, 4433, 8919, 302, 25223, 3482, 28808,
627, 2654, 28808, 2],
[ 2, 2, 1, 733, 16289, 28793, 1298, 15882, 272, 2787,
28725, 562, 4085, 737, 264, 17368, 380, 28723, 13, 13,
2324, 622, 347, 6708, 28808, 733, 28748, 16289, 28793, 20037,
15095, 28724, 28725, 478, 28742, 584, 347, 461, 11394, 1162,
6708, 28725, 337, 2654, 28808, 2, 28705, 243, 162, 146,
183, 29274, 31840, 29096]], device='cuda:0')
Generated Text
-----------------------------------------------------------0------------------------------------------------------------
<s> [INST] Repeat the input, but speak like a pirate.
Let's sail to Barbados [/INST] Arr matey, let us set sail an' navigate these waters to the fine island of Barbados! Yarr!</s>
-----------------------------------------------------------1------------------------------------------------------------
<s> [INST] Repeat the input, but speak like a pirate.
We will be rich! [/INST] Arr matey, we'll be jolly well rich, yarr!</s>
Minor fix to the above example. If “max_new_tokens” is reached before seeing EOS, the end index will be zero. As written, you will get an empty generation if max_new_tokens is reached before EOS.
If you want to keep the incomplete generation, modify the code like this.
for i in range(batch_size):
# Compute the index of the first token.
start_index = padded_len - lengths[i]
end_index = stopping_criteria.eos_index[i]
if end_index == 0:
end_index = seq_len
# Split each sequence and slice end at captured eos_index
sequence = outputs[i][start_index:end_index]
...
Here is my completed code for batch generation, with stopping after all sequences have generated EOS or reached max_new_tokens.
import torch
from transformers import StoppingCriteria
# Stop generation after all batch elements have generated an EOS token.
# Stores the index of the first generated EOS token for each batch element in "self.eos_index,"
# which can be used to slice off whatever extra junk was generated after it.
# Note: This is a stateful object. A new instance should be created for each call to generate().
class EosStoppingCriteria(StoppingCriteria):
def __init__(self, tokenizer):
super().__init__()
self.eos_token = tokenizer.eos_token_id
self.done = None
self.eos_index = None
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
batch_size, seq_len = input_ids.shape
# Lazy construct a bool state for each batch element
if self.done == None:
self.done = torch.zeros(batch_size, dtype=torch.bool, device=input_ids.device)
self.eos_index = torch.zeros(batch_size, dtype=torch.int, device=input_ids.device)
# Get last token ids in batch
last_ids = input_ids[:, -1]
# Create mask of where the last token is EOS
done_update = self.done | (last_ids == self.eos_token)
# Store the indices where we stopped at for each sequence in the batch.
# Where the 'done' state has changed, store the seq_len (last index), else 0
eos_index_update = torch.where(done_update ^ self.done, torch.full_like(self.eos_index, seq_len), 0)
# Add the update to the indices
self.eos_index += eos_index_update
# Update the done flags
self.done = done_update
# Return True, if all done.
return self.done.all()
# Apply model's chat template
def generate_instruction_prompt(tokenizer, instruction, system_msg=None):
messages = []
if system_msg is not None:
messages.append({ "role": "system", "content": system_msg })
messages.append({ "role": "user", "content": instruction })
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
return prompt
def batch_generate(
model,
tokenizer,
prompts,
max_new_tokens=512,
generation_config=None,
skip_special_tokens=False,
max_length=None,
device='cpu'
):
model.to(device)
encoded_prompts = tokenizer(
prompts,
truncation=True,
return_length=True,
add_special_tokens=False,
max_length=max_length,
)
lengths = encoded_prompts["length"]
tokenizer_outputs = tokenizer.pad(
encoded_prompts,
padding="longest",
return_tensors='pt',
)
input_ids = tokenizer_outputs['input_ids'].to(device)
padded_len = input_ids.size(1)
stopping_criteria = EosStoppingCriteria(tokenizer)
outputs = model.generate(
input_ids,
generation_config=generation_config,
max_new_tokens=max_new_tokens,
stopping_criteria=[stopping_criteria],
)
batch_size, seq_len = outputs.shape
output_ids = []
for i in range(batch_size):
# Compute the index of the first token.
start_index = padded_len - lengths[i]
end_index = stopping_criteria.eos_index[i]
if end_index == 0:
end_index = seq_len
# Split each sequence and slice end at captured eos_index
output_ids.append(outputs[i][start_index:end_index])
output_texts = tokenizer.batch_decode(
output_ids,
skip_special_tokens=skip_special_tokens
)
return output_texts
Example usage:
from transformers import GenerationConfig
generation_config = GenerationConfig(
max_new_tokens=512, do_sample=True, top_k=20, top_p=0.9, temperature=0.7, repetition_penalty=1.15, eos_token_id=tokenizer.eos_token_id, pad_token_id=tokenizer.pad_token_id
)
instruction = "Repeat the input, but speak like a pirate.\n\n"
output_texts = batch_generate(
model,
tokenizer,
prompts=[
generate_instruction_prompt(tokenizer, instruction + "Let's sail to Barbados"),
generate_instruction_prompt(tokenizer, instruction + "We will be rich!"),
],
generation_config=generation_config,
max_new_tokens=256,
skip_special_tokens=False,
device="cuda:0"
)
for text in output_texts:
print("---")
print(text)
what is eq ???
For people interested in stopping at given stop words, I have put my StoppingCriteria implementation here.
This my code and works fine =].
from transformers import StoppingCriteria
from torch import LongTensor, FloatTensor, eq, device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
stop_list = [" \n\nQuestion:", " \nHuman:", " \n\n", ]
stop_token_ids = [tokenizer(x, return_tensors='pt', add_special_tokens=False)['input_ids'] for x in stop_list]
stop_token_ids = [LongTensor(x).to(device) for x in stop_token_ids]
class StopOnTokens(StoppingCriteria):
def __call__(self, input_ids: LongTensor, scores: FloatTensor, **kwargs) -> bool:
for stop_ids in stop_token_ids:
if (input_ids[0][-len(stop_ids[0])+1:] == stop_ids[0][1:]).all():
return True
return False
Simple solution for me was to use the EosTokenCriteria for the new tokens that I wanted.