From a2e6d0012855101bb474fca7f0709c5d7fc0f996 Mon Sep 17 00:00:00 2001
From: oobabooga <112222186+oobabooga@users.noreply.github.com>
Date: Sun, 19 Nov 2023 09:22:08 -0800
Subject: [PATCH] Use convert_ids_to_tokens instead of decode in logits
 endpoint

This preserves the llama tokenizer spaces.
---
 modules/logits.py | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/modules/logits.py b/modules/logits.py
index 383659e0..5d0d3210 100644
--- a/modules/logits.py
+++ b/modules/logits.py
@@ -55,7 +55,10 @@ def get_next_logits(prompt, state, use_samplers, previous, return_dict=False):
     if is_non_hf_exllamav1 or is_non_hf_llamacpp:
         topk_indices = [i.expand((1, 1)) for i in topk_indices]
 
-    tokens = [shared.tokenizer.decode(i) for i in topk_indices]
+    if hasattr(shared.tokenizer, 'convert_ids_to_tokens'):
+        tokens = [shared.tokenizer.convert_ids_to_tokens(int(i)) for i in topk_indices]
+    else:
+        tokens = [shared.tokenizer.decode(i) for i in topk_indices]
 
     if return_dict:
         output = {}