Faster sampling with mx.compile (#937)

* faster sampling with compile

* fix test
This commit is contained in:
Awni Hannun
2024-08-15 11:29:09 -07:00
committed by GitHub
parent 95840f32e2
commit 9b83004631
3 changed files with 32 additions and 28 deletions

View File

@@ -20,7 +20,7 @@ from transformers import PreTrainedTokenizer
# Local imports
from .models.base import KVCache
from .sample_utils import top_p_sampling
from .sample_utils import categorical_sampling, top_p_sampling
from .tokenizer_utils import TokenizerWrapper, load_tokenizer
from .tuner.utils import apply_lora_layers
from .tuner.utils import dequantize as dequantize_model
@@ -169,7 +169,7 @@ def generate_step(
if top_p > 0 and top_p < 1.0:
token = top_p_sampling(logits, top_p, temp)
else:
token = mx.random.categorical(logits * (1 / temp))
token = categorical_sampling(logits, temp)
return token, logprobs