From 20bd645f6a355c39b1907fe213fa44e38a83dc8c Mon Sep 17 00:00:00 2001
From: oobabooga <112222186+oobabooga@users.noreply.github.com>
Date: Mon, 6 Mar 2023 15:58:18 -0300
Subject: [PATCH] Fix bug in multigpu setups (attempt 3)

---
 modules/text_generation.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/modules/text_generation.py b/modules/text_generation.py
index f585c013..0e8aec51 100644
--- a/modules/text_generation.py
+++ b/modules/text_generation.py
@@ -113,7 +113,7 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
 
     input_ids = encode(question, max_new_tokens)
     cuda = "" if (shared.args.cpu or shared.args.deepspeed or shared.args.flexgen) else ".cuda()"
-    n = torch.tensor(shared.tokenizer.eos_token_id) if eos_token is None else encode(eos_token)[0][-1]
+    n = shared.tokenizer.eos_token_id if eos_token is None else int(encode(eos_token)[0][-1])
     if stopping_string is not None:
         # The stopping_criteria code below was copied from
         # https://github.com/PygmalionAI/gradio-ui/blob/master/src/model.py
@@ -195,7 +195,7 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
             yield formatted_outputs(reply, shared.model_name)
 
             if not shared.args.flexgen:
-                if output[-1].to("cpu") == n.to("cpu"):
+                if output[-1] == n:
                     break
                 input_ids = torch.reshape(output, (1, output.shape[0]))
             else: