Update cache_4bit documentation (#5649)

---------

Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>

README.md

@@ -279,7 +279,7 @@ List of command-line flags
 |`--cfg-cache`                         | ExLlamav2_HF: Create an additional cache for CFG negative prompts. Necessary to use CFG with that loader. |
 |`--no_flash_attn`                     | Force flash-attention to not be used. |
 |`--cache_8bit`                        | Use 8-bit cache to save VRAM. |
-|`--cache_4bit`                        | Use 4-bit cache to save VRAM. |
+|`--cache_4bit`                        | Use Q4 cache to save VRAM. |
 |`--num_experts_per_token NUM_EXPERTS_PER_TOKEN` |  Number of experts to use for generation. Applies to MoE models like Mixtral. |
 
 #### AutoGPTQ

docs/04 - Model Tab.md

@@ -46,6 +46,7 @@ Examples:
 * **cfg-cache**: Creates a second cache to hold the CFG negative prompts. You need to set this if and only if you intend to use CFG in the "Parameters" > "Generation" tab. Checking this parameter doubles the cache VRAM usage.
 * **no_flash_attn**: Disables flash attention. Otherwise, it is automatically used as long as the library is installed.
 * **cache_8bit**: Create a 8-bit precision cache instead of a 16-bit one. This saves VRAM but increases perplexity (I don't know by how much).
+* **cache_4bit**: Creates a Q4 cache using grouped quantization.
 
 ### ExLlamav2
 

modules/exllamav2_hf.py

@@ -63,6 +63,8 @@ class Exllamav2HF(PreTrainedModel):
         if shared.args.cfg_cache:
             if shared.args.cache_8bit:
                 self.ex_cache_negative = ExLlamaV2Cache_8bit(self.ex_model)
+            elif shared.args.cache_4bit:
+                self.ex_cache_negative = ExLlamaV2Cache_Q4(self.ex_model)
             else:
                 self.ex_cache_negative = ExLlamaV2Cache(self.ex_model)
 

modules/shared.py

@@ -139,7 +139,7 @@ group.add_argument('--max_seq_len', type=int, default=2048, help='Maximum sequen
 group.add_argument('--cfg-cache', action='store_true', help='ExLlamav2_HF: Create an additional cache for CFG negative prompts. Necessary to use CFG with that loader.')
 group.add_argument('--no_flash_attn', action='store_true', help='Force flash-attention to not be used.')
 group.add_argument('--cache_8bit', action='store_true', help='Use 8-bit cache to save VRAM.')
-group.add_argument('--cache_4bit', action='store_true', help='Use 4-bit cache to save VRAM.')
+group.add_argument('--cache_4bit', action='store_true', help='Use Q4 cache to save VRAM.')
 group.add_argument('--num_experts_per_token', type=int, default=2, help='Number of experts to use for generation. Applies to MoE models like Mixtral.')
 
 # AutoGPTQ

modules/ui_model_menu.py

@@ -132,7 +132,7 @@ def create_ui():
                             shared.gradio['disk'] = gr.Checkbox(label="disk", value=shared.args.disk)
                             shared.gradio['bf16'] = gr.Checkbox(label="bf16", value=shared.args.bf16)
                             shared.gradio['cache_8bit'] = gr.Checkbox(label="cache_8bit", value=shared.args.cache_8bit, info='Use 8-bit cache to save VRAM.')
-                            shared.gradio['cache_4bit'] = gr.Checkbox(label="cache_4bit", value=shared.args.cache_4bit, info='Use 4-bit cache to save VRAM.')
+                            shared.gradio['cache_4bit'] = gr.Checkbox(label="cache_4bit", value=shared.args.cache_4bit, info='Use Q4 cache to save VRAM.')
                             shared.gradio['autosplit'] = gr.Checkbox(label="autosplit", value=shared.args.autosplit, info='Automatically split the model tensors across the available GPUs.')
                             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.')
                             shared.gradio['cfg_cache'] = gr.Checkbox(label="cfg-cache", value=shared.args.cfg_cache, info='Necessary to use CFG with this loader.')