323f593a96
Custom unpickler to avoid pickle's arbitrary code execution vulnerability
2022-10-06 20:08:08 -04:00
bdfa6d86b7
Seed has to be a 64-bit unsigned int or PyTorch will throw an error
...
tpu_mtj_backend's seed can be an integer of arbitrary size but we will
limit it to a 64-bit unsigned integer anyways for consistency.
2022-10-02 17:50:32 -04:00
1a59a4acea
Allow changing sampler seed and sampler order from API
2022-10-02 16:25:51 -04:00
6c32bc18d7
GPT2Tokenizer for TPU
2022-09-27 18:33:31 +02:00
11455697ef
Tokenizer Fixes (Slow first to keep coherency)
2022-09-27 17:57:18 +02:00
07896867b2
Revert Tokenizer Change
2022-09-27 15:36:08 +02:00
82a250aa1b
Revert "Fix tokenizer selection code"
...
This reverts commit 7fba1fd28a
2022-09-27 15:33:08 +02:00
79ae0f17ec
Merge branch 'main' into merge
2022-09-26 16:10:10 -04:00
7fba1fd28a
Fix tokenizer selection code
2022-09-26 14:37:25 -04:00
943614b5e6
Merge branch 'main' into dependency-fix
2022-09-15 17:33:48 -04:00
7bf6c9a23f
Remove TPU Colab's dependency on optax and chex
2022-09-15 13:47:48 -04:00
cbacfbdfac
Fix error that occurs when using dynamic TPU backend
2022-08-27 17:42:49 -04:00
938e1eddf3
Fix `jax.lax.cond` call
2022-08-23 18:13:46 -04:00
6ffaf43548
Repetition penalty is now sampler #6 in the sampler order
2022-08-23 15:10:21 -04:00
74922966bd
Merge branch 'avril' into rep-pen-order
2022-08-23 14:47:29 -04:00
9e140e3ba9
Merge branch 'KoboldAI:main' into united
2022-07-05 21:35:53 +02:00
2a78b66932
Fix base OPT-125M and finetuned OPT models in Colab TPU instances
2022-07-05 15:28:58 -04:00
048bd0ff3b
Add support for setting the RNG seed and full determinism
2022-06-28 13:21:05 -04:00
6a89ad5b94
Merge branch 'main' into united
2022-06-23 21:07:56 +02:00
d94f29a68a
Merge pull request #127 from VE-FORBRYDERNE/tracer
...
Fix JAX UnexpectedTracerError
2022-06-23 19:29:51 +02:00
33a2a318db
Fix 20B TPU model
2022-06-21 17:16:01 -04:00
0ea4fa9c87
Automatically calculate badwords and pad_token_id
2022-06-21 14:35:52 -04:00
ea7d278ff4
Fix 20B TPU model
2022-06-21 13:16:45 -04:00
5e71f7fe97
Use slow tokenizer if fast tokenizer is not available
2022-06-17 21:08:37 -04:00
2d3db7b4ba
Implement support for sampler order in the backend code
2022-06-13 19:12:23 -04:00
fdb2a7fa4c
Top-A sampling
2022-06-10 22:28:20 -04:00
f2558e39d9
Fix JAX UnexpectedTracerError
2022-05-31 13:25:41 -04:00
707316de31
Kaggle TPU support
2022-05-31 12:20:16 -04:00
d4e8f56789
Remove debugging code from tpu_mtj_backend.py
2022-05-14 12:00:44 -04:00
0c5ca5261e
Loading a sharded model will now display only one progress bar
2022-05-13 23:32:16 -04:00
b1d8797a54
Allow TPU Colab to load sharded HF models
2022-05-12 23:51:40 -04:00
4fa5f1cd6a
Add TPU support for OPT-350M
...
The 350M model seems to have a different structure than the other ones ???
2022-05-12 22:21:15 -04:00
f5e689a725
Upload maps/opt.json and update requirements
2022-05-12 19:09:31 -04:00
b97b2a02d6
Add `--revision` command line flag
2022-05-10 22:14:56 -04:00
c117bfd0ad
Fix lazy loader
2022-04-08 19:38:15 -04:00
fabbdf2bb1
Lazy loader Python 3.6 compatibility
...
The current lazy loader relies on a feature of the Python zipfile module
that was added in Python 3.7.0:
https://bugs.python.org/issue22908
This commit adds compatibility for Python 3.6.
2022-04-02 15:02:54 -04:00
67e28d2b5c
Typical sampling needs to use nansum instead of sum
...
If `probs` is zero then `log_probs` will be negative infinity, and the
calculation of `neg_entropy` would then give NaN because zero times
infinity is a mathematically indeterminate value.
We need to use nansum so that those NaN values are treated as zeros to
ignore them in the entropy calculation.
2022-03-28 00:02:31 -04:00
d5989d4c62
Hide division by zero warning in JAX typical filter
...
This warning happens when `np.log` gets an input containing zeros.
In that case, NumPy will throw a warning and output negative infinity.
Negative infinity is the correct behaviour here, so we can safely ignore
the warning.
2022-03-27 16:57:12 -04:00
20e48b11d7
Typical sampling
2022-03-27 16:25:50 -04:00
73aecc0510
Divide NeoX replicated bias layers by 4 again instead of by 8
2022-03-20 01:04:55 -04:00
05fc46b253
Changing this again to divide by 8
2022-03-19 02:09:41 -04:00
6c20d0d657
Nevermind, dividing by 4 is actually correct...
2022-03-19 00:55:04 -04:00
f16b61ec77
Should divide NeoX replicated parameters by 8 (not by 4)
...
Also, suppresses the PyTorch 1.11 warning about transposing tensors with
ndim != 2 in the new code
2022-03-19 00:48:33 -04:00
c2c139e940
Change default PE type for NeoX to `neox_rotary`
2022-03-19 00:26:04 -04:00
85a4959efa
Merge branch 'united' into neox
2022-03-18 11:19:03 -04:00
c444260eac
Silence PyTorch warning about transposing tensors with dimension != 2
2022-03-17 15:16:56 -04:00
eaf190469d
Add PyTorch 1.11 support for lazy loader
2022-03-17 12:51:41 -04:00
95c4251db9
Print two newlines before loading HF models
2022-03-15 13:58:53 -04:00
9e2848e48f
Show parameter count when loading GPT-NeoX in Colab TPU instance
2022-03-15 13:55:27 -04:00
88f247d535
GPT-NeoX-20B support in Colab TPU instances
2022-03-14 23:14:20 -04:00
vfbd
Henk
Gnome Ann