fix(server): Use cleanup_tokenization_spaces=False for lossless decoding (#13)

Fixes #12 in the easiest way I could think of.
2023-01-03 11:07:05 +01:00 · 2023-01-03 11:07:05 +01:00 · b94f30215f
parent 60472f9d2b
commit b94f30215f
1 changed files with 2 additions and 1 deletions
--- a/server/text_generation/models/causal_lm.py
+++ b/server/text_generation/models/causal_lm.py
@ -354,7 +354,8 @@ class CausalLM(Model):
            if stop:
                # Decode all tokens
                output_text = self.tokenizer.decode(
-                    all_input_ids.squeeze(-1), skip_special_tokens=True
+                    all_input_ids.squeeze(-1), skip_special_tokens=True,
+                    cleanup_tokenization_spaces=False
                )
                # Slice with input_length to remove padding
                token_ids = all_input_ids[-new_input_length:]