Thank you so much for the work you are doing, this is my little
contribution to this great thing you have built. I hope it is useful and
helpful, please don't hesitate to discuss any matters that are not
clear!
I am basing my implementation of frequency penalty on OpenAI's
implementation:
https://platform.openai.com/docs/guides/text-generation/parameter-details
The problem I see with TGI's current implementation is that is not
taking into account the frequency of tokens which have already been
sampled in the current generation stream. Also, the scaling is of the
adjusted token logits is done differently for positive and negative
logits. While in OpenAI's implementation token frequency is taking into
account and the scaling is always done with a subtraction (if penalty is
positive) or add operation (if penalty is negative).
This leads to corrupt generations as I mentioned in issue #1810 .
Moreover, after my tests, other issues are also gone like the one about
some request's with ``penalty_frequency = 1.0`` overruling other
requests (with ``frequency_penalty = 0.0``) in the same batch and
therefore corrupting all generations in the batch. Basically, padding
does not affect this implementation so I believe this ``score *=
input_ids.ne(0)`` is not needed anymore.
Frequency penalty | -1.0 | 0.0 | 1.0
-- | -- | -- | --
Before my change | https://paste.mozilla.org/JxqGJkWY |
https://paste.mozilla.org/hrztJ56h | https://paste.mozilla.org/pBSEH2zw
After my change | https://paste.mozilla.org/7gXCi7zo |
https://paste.mozilla.org/ZR9rJ92g | https://paste.mozilla.org/gHaD2YnC
---------
Co-authored-by: martini <martin.iglesiasgoyanes@adyen.com>
This PR resolves an issue with the penalty processors during batched
generation where extra padding tokens incorrectly impact the penalty
scores.
generation is impacted in the case where at least one item in the batch
includes a `frequency_penalty`
reproduction script below
```python
import requests
from concurrent import futures
import time
headers = {
"Content-Type": "application/json",
}
json_data = {
"inputs": "[INST] Whats the capitol of France? [/INST]",
"parameters": {
"max_new_tokens": 100,
"seed": 20,
"do_sample": False,
},
}
json_data2 = {
"inputs": "<s>[INST]Write a mind bending story: I saw a puppy a cat a rat and a raccoon during my bike ride in the park[/INST]",
"parameters": {
"max_new_tokens": 100,
"seed": 2,
"do_sample": False,
# OFFENDING LINE
"frequency_penalty": 1.05,
},
}
base_url = "http://localhost:3000/generate"
def req():
response = requests.post(base_url, headers=headers, json=json_data)
print("[req ]", response.json())
def req2():
response = requests.post(base_url, headers=headers, json=json_data2)
print("[req2]", response.json())
n = 1
for i in range(0, 3):
print(f"- {n} threads -")
with futures.ThreadPoolExecutor(max_workers=n) as executor:
executor.submit(req)
for i in range(3):
executor.submit(req2)
n += 1
# - 1 threads -
# [req ] {'generated_text': ' The capital of France is Paris.'}
# [req2] {'generated_text': " As you were riding your bicycle through Central Park, enjoying some fresh air on an otherwise gloomy day. You couldn't help but notice that it was eerily quiet for this time of year - usually there would be hordes"}
# [req2] {'generated_text': " As you were riding your bicycle through Central Park, enjoying some fresh air on an otherwise gloomy day. You couldn't help but notice that it was eerily quiet for this time of year - usually there would be hordes"}
# [req2] {'generated_text': " As you were riding your bicycle through Central Park, enjoying some fresh air on an otherwise gloomy day. You couldn't help but notice that it was eerily quiet for this time of year - usually there would be hordes"}
# - 2 threads -
# [req ] {'generated_text': ' The capital city'}
# [req2] {'generated_text': ' As""%\n================'}
# [req2] {'generated_text': ' As""%%$\n================'}
# [req2] {'generated_text': " As you were riding your bicycle through Central Park, enjoying some fresh air on an otherwise gloomy day. You couldn't help but notice that it was eerily quiet for this time of year - usually there would be hordes"}
# output with this PR's changes:
# - 1 threads -
# [req ] {'generated_text': ' The capital of France is Paris.'}
# [req2] {'generated_text': " As you were riding your bicycle through Central Park, enjoying some fresh air on an otherwise gloomy day. You couldn't help but notice that it was eerily quiet for this time of year - usually there would be hordes"}
# [req2] {'generated_text': " As you were riding your bicycle through Central Park, enjoying some fresh air on an otherwise gloomy day. You couldn't help but notice that it was eerily quiet for this time of year - usually there would be hordes"}
# [req2] {'generated_text': " As you were riding your bicycle through Central Park, enjoying some fresh air on an otherwise gloomy day. You couldn't help but notice that it was eerily quiet for this time of year - usually there would be hordes"}
# - 2 threads -
# [req ] {'generated_text': ' The capital city'}
# [req2] {'generated_text': " As you were riding your bicycle through Central Park, enjoying some fresh air on an otherwise gloomy day. You couldn't help but notice that it was eerily quiet for this time of year - usually there would be hordes"}
# [req2] {'generated_text': " As you were riding your bicycle through Central Park, enjoying some fresh air on an otherwise gloomy day. You couldn't help but notice that it was eerily quiet for this time of year - usually there would be hordes"}
# [req2] {'generated_text': " As you were riding your bicycle through Central Park, enjoying some fresh air on an otherwise gloomy day. You couldn't help but notice that it was eerily quiet for this time of year - usually there would be hordes"}
```
**divergence from expected generation is easier to reproduce with
batched grammar requests as they are more sensitive to unexpected
outputs.
this PR resolves the issue by setting the penalty score to 0 where input
ids are padding tokens (0).
---------
Co-authored-by: OlivierDehaene <olivier@huggingface.co>
This PR correctly handles batches with a mixture of constrained and non
constrained generations.
Currently if batch contains mixed generations the generation will throw
an error because it will incorrectly attempt to constrain a request with
an empty grammar.
We now handled `None` grammars and only apply the mask if needed
Fixes:
https://github.com/huggingface/text-generation-inference/issues/1643
This PR resolves a couple
- [X] adjusts the tool response to align with openai's tools response
type
- [X] bumps pydantic to `2.6.4` in all apps (resolves dependency issue
when running tests)
- [X] bump `outlines` version and fix import for new name
Martin Iglesias Goyanes