diff --git a/README.md b/README.md
index 3c58d0ec..bd356a18 100644
--- a/README.md
+++ b/README.md
@@ -280,6 +280,7 @@ Optionally, you can use the following command-line flags:
 | `--n_gqa N_GQA`         | grouped-query attention. Must be 8 for llama-2 70b. |
 | `--rms_norm_eps RMS_NORM_EPS`  | 5e-6 is a good value for llama-2 models. |
 | `--cpu`                        | Use the CPU version of llama-cpp-python instead of the GPU-accelerated version. |
+|`--cfg-cache`                   | llamacpp_HF: Create an additional cache for CFG negative prompts. |
 
 #### ctransformers
 
diff --git a/modules/llamacpp_hf.py b/modules/llamacpp_hf.py
index 3c7314d1..f37d710a 100644
--- a/modules/llamacpp_hf.py
+++ b/modules/llamacpp_hf.py
@@ -38,16 +38,17 @@ class LlamacppHF(PreTrainedModel):
         self.llamacpp_cache = {
             'n_tokens': self.model.n_tokens,
             'input_ids': self.model.input_ids,
-            'scores': self.model.scores
+            'scores': self.model.scores,
+            'ctx': self.model.ctx
         }
 
         if shared.args.cfg_cache:
-            logger.warning('CFG is currently bugged and not functional for llamacpp_HF. Contributions are welcome.')
             self.past_seq_negative = None
             self.llamacpp_cache_negative = {
                 'n_tokens': self.model.n_tokens,
                 'input_ids': self.model.input_ids.copy(),
-                'scores': self.model.scores.copy()
+                'scores': self.model.scores.copy(),
+                'ctx': llama_cpp_lib().llama_new_context_with_model(model.model, model.params)
             }
 
     def _validate_model_class(self):
@@ -63,25 +64,29 @@ class LlamacppHF(PreTrainedModel):
         self.llamacpp_cache.update({
             'n_tokens': self.model.n_tokens,
             'input_ids': self.model.input_ids,
-            'scores': self.model.scores
+            'scores': self.model.scores,
+            'ctx': self.model.ctx
         })
 
     def save_negative_cache(self):
         self.llamacpp_cache_negative.update({
             'n_tokens': self.model.n_tokens,
             'input_ids': self.model.input_ids,
-            'scores': self.model.scores
+            'scores': self.model.scores,
+            'ctx': self.model.ctx
         })
 
     def load_cache(self):
         self.model.n_tokens = self.llamacpp_cache['n_tokens']
         self.model.input_ids = self.llamacpp_cache['input_ids']
         self.model.scores = self.llamacpp_cache['scores']
+        self.model.ctx = self.llamacpp_cache['ctx']
 
     def load_negative_cache(self):
         self.model.n_tokens = self.llamacpp_cache_negative['n_tokens']
         self.model.input_ids = self.llamacpp_cache_negative['input_ids']
         self.model.scores = self.llamacpp_cache_negative['scores']
+        self.model.ctx = self.llamacpp_cache_negative['ctx']
 
     @property
     def device(self) -> torch.device:
@@ -95,7 +100,6 @@ class LlamacppHF(PreTrainedModel):
         if len(args) > 0:
             if not shared.args.cfg_cache:
                 logger.error("Please enable the cfg-cache option to use CFG with llamacpp_HF.")
-                logger.warning('CFG is currently bugged and not functional for llamacpp_HF. Contributions are welcome.')
                 return
 
             input_ids = args[0]
diff --git a/modules/loaders.py b/modules/loaders.py
index b8660a46..dde40c9b 100644
--- a/modules/loaders.py
+++ b/modules/loaders.py
@@ -95,6 +95,7 @@ loaders_and_params = OrderedDict({
         'alpha_value',
         'compress_pos_emb',
         'cpu',
+        'cfg_cache',
         'llamacpp_HF_info',
     ],
     'ctransformers': [
@@ -268,6 +269,8 @@ loaders_samplers = {
         'mirostat_mode',
         'mirostat_tau',
         'mirostat_eta',
+        'guidance_scale',
+        'negative_prompt',
         'ban_eos_token',
         'add_bos_token',
         'skip_special_tokens',