Disable logits_all in llamacpp_HF (makes processing 3x faster)

modules/llamacpp_hf.py

@@ -204,7 +204,7 @@ class LlamacppHF(PreTrainedModel):
             'rope_freq_base': RoPE.get_rope_freq_base(shared.args.alpha_value, shared.args.rope_freq_base),
             'tensor_split': tensor_split_list,
             'rope_freq_scale': 1.0 / shared.args.compress_pos_emb,
-            'logits_all': True,
+            'logits_all': shared.args.logits_all,
         }
 
         Llama = llama_cpp_lib().Llama

modules/loaders.py

@@ -123,6 +123,7 @@ loaders_and_params = OrderedDict({
         'numa',
         'cfg_cache',
         'use_fast',
+        'logits_all',
         'llamacpp_HF_info',
     ],
     'ctransformers': [

modules/shared.py

@@ -113,6 +113,7 @@ parser.add_argument('--n-gpu-layers', type=int, default=0, help='Number of layer
 parser.add_argument('--tensor_split', type=str, default=None, help='Split the model across multiple GPUs. Comma-separated list of proportions. Example: 18,17.')
 parser.add_argument('--llama_cpp_seed', type=int, default=0, help='Seed for llama-cpp models. Default is 0 (random).')
 parser.add_argument('--numa', action='store_true', help='Activate NUMA task allocation for llama.cpp.')
+parser.add_argument('--logits_all', action='store_true', help='Needs to be set for perplexity evaluation to work. Otherwise, ignore it, as it makes prompt processing slower.')
 parser.add_argument('--cache-capacity', type=str, help='Maximum cache capacity (llama-cpp-python). Examples: 2000MiB, 2GiB. When provided without units, bytes will be assumed.')
 
 # ExLlama

modules/ui.py

@@ -87,6 +87,7 @@ def list_model_elements():
         'alpha_value',
         'rope_freq_base',
         'numa',
+        'logits_all',
     ]
     if is_torch_xpu_available():
         for i in range(torch.xpu.device_count()):

modules/ui_model_menu.py

@@ -124,6 +124,7 @@ def create_ui():
                             shared.gradio['llama_cpp_seed'] = gr.Number(label='Seed (0 for random)', value=shared.args.llama_cpp_seed)
                             shared.gradio['trust_remote_code'] = gr.Checkbox(label="trust-remote-code", value=shared.args.trust_remote_code, info='To enable this option, start the web UI with the --trust-remote-code flag. It is necessary for some models.', interactive=shared.args.trust_remote_code)
                             shared.gradio['use_fast'] = gr.Checkbox(label="use_fast", value=shared.args.use_fast, info='Set use_fast=True while loading the tokenizer. May trigger a conversion that takes several minutes.')
+                            shared.gradio['logits_all'] = gr.Checkbox(label="logits_all", value=shared.args.logits_all, info='Needs to be set for perplexity evaluation to work. Otherwise, ignore it, as it makes prompt processing slower.')
                             shared.gradio['use_flash_attention_2'] = gr.Checkbox(label="use_flash_attention_2", value=shared.args.use_flash_attention_2, info='Set use_flash_attention_2=True while loading the model.')
                             shared.gradio['disable_exllama'] = gr.Checkbox(label="disable_exllama", value=shared.args.disable_exllama, info='Disable ExLlama kernel.')
                             shared.gradio['no_flash_attn'] = gr.Checkbox(label="no_flash_attn", value=shared.args.no_flash_attn, info='Force flash-attention to not be used.')