From 922394c68fd872d87edf79a420fbf41ce509c72e Mon Sep 17 00:00:00 2001
From: vfbd <ve.forbryderne@gmail.com>
Date: Wed, 22 Jun 2022 11:23:03 -0400
Subject: [PATCH 1/2] Don't blacklist </s> token in "s" newline mode

---
 aiserver.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/aiserver.py b/aiserver.py
index 4bbd89b0..71cd67dc 100644
--- a/aiserver.py
+++ b/aiserver.py
@@ -1995,7 +1995,7 @@ def load_model(use_gpu=True, gpu_layers=None, disk_layers=None, initial_load=Fal
                             shutil.rmtree("cache/")
 
                 if(vars.badwordsids is vars.badwordsids_default and vars.model_type not in ("gpt2", "gpt_neo", "gptj")):
-                    vars.badwordsids = [[v] for k, v in tokenizer.get_vocab().items() if any(c in str(k) for c in "<>[]")]
+                    vars.badwordsids = [[v] for k, v in tokenizer.get_vocab().items() if any(c in str(k) for c in "<>[]") if vars.newlinemode != "s" or str(k) != "</s>"]
 
                 patch_causallm(model)
 
@@ -2162,7 +2162,7 @@ def load_model(use_gpu=True, gpu_layers=None, disk_layers=None, initial_load=Fal
             vars.modeldim = int(tpu_mtj_backend.params.get("d_embed", tpu_mtj_backend.params["d_model"]))
             tokenizer = tpu_mtj_backend.tokenizer
             if(vars.badwordsids is vars.badwordsids_default and vars.model_type not in ("gpt2", "gpt_neo", "gptj")):
-                vars.badwordsids = [[v] for k, v in tokenizer.get_vocab().items() if any(c in str(k) for c in "<>[]")]
+                vars.badwordsids = [[v] for k, v in tokenizer.get_vocab().items() if any(c in str(k) for c in "<>[]") if vars.newlinemode != "s" or str(k) != "</s>"]
         else:
             loadsettings()
     

From 53034ee533227e6b650aabbd3f6e599829a5ee8d Mon Sep 17 00:00:00 2001
From: vfbd <ve.forbryderne@gmail.com>
Date: Wed, 22 Jun 2022 12:07:36 -0400
Subject: [PATCH 2/2] Delete all torch tensors before loading model

---
 aiserver.py | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/aiserver.py b/aiserver.py
index 71cd67dc..6afee9a6 100644
--- a/aiserver.py
+++ b/aiserver.py
@@ -1528,6 +1528,14 @@ def load_model(use_gpu=True, gpu_layers=None, disk_layers=None, initial_load=Fal
     model = None
     generator = None
     model_config = None
+    for tensor in gc.get_objects():
+        try:
+            if torch.is_tensor(tensor):
+                with torch.no_grad():
+                    tensor.set_(torch.tensor((), device=tensor.device, dtype=tensor.dtype))
+        except:
+            pass
+    gc.collect()
     try:
         torch.cuda.empty_cache()
     except: