Use pre-compiled python module for ExLlama (#2770)

modules/LoRA.py

@@ -20,10 +20,13 @@ def add_lora_to_model(lora_names):
 def add_lora_exllama(lora_names):
 
     try:
-        from repositories.exllama.lora import ExLlamaLora
+        from exllama.lora import ExLlamaLora
     except:
-        logger.error("Could not find the file repositories/exllama/lora.py. Make sure that exllama is cloned inside repositories/ and is up to date.")
+        try:
-        return
+            from repositories.exllama.lora import ExLlamaLora
+        except:
+            logger.error("Could not find the file repositories/exllama/lora.py. Make sure that exllama is cloned inside repositories/ and is up to date.")
+            return
 
     if len(lora_names) == 0:
         shared.model.generator.lora = None

modules/exllama.py

@@ -3,12 +3,23 @@ from pathlib import Path
 
 from modules import shared
 from modules.logging_colors import logger
-from modules.relative_imports import RelativeImport
 
-with RelativeImport("repositories/exllama"):
+try:
-    from generator import ExLlamaGenerator
+    from exllama.generator import ExLlamaGenerator
-    from model import ExLlama, ExLlamaCache, ExLlamaConfig
+    from exllama.model import ExLlama, ExLlamaCache, ExLlamaConfig
-    from tokenizer import ExLlamaTokenizer
+    from exllama.tokenizer import ExLlamaTokenizer
+except:
+    logger.warning('Exllama module failed to load. Will attempt to load from repositories.')
+    try:
+        from modules.relative_imports import RelativeImport
+
+        with RelativeImport("repositories/exllama"):
+            from generator import ExLlamaGenerator
+            from model import ExLlama, ExLlamaCache, ExLlamaConfig
+            from tokenizer import ExLlamaTokenizer
+    except:
+        logger.error("Could not find repositories/exllama/. Make sure that exllama is cloned inside repositories/ and is up to date.")
+        raise
 
 
 class ExllamaModel:

modules/exllama_hf.py

@@ -9,10 +9,19 @@ from transformers.modeling_outputs import CausalLMOutputWithPast
 
 from modules import shared
 from modules.logging_colors import logger
-from modules.relative_imports import RelativeImport
 
-with RelativeImport("repositories/exllama"):
+try:
-    from model import ExLlama, ExLlamaCache, ExLlamaConfig
+    from exllama.model import ExLlama, ExLlamaCache, ExLlamaConfig
+except:
+    logger.warning('Exllama module failed to load. Will attempt to load from repositories.')
+    try:
+        from modules.relative_imports import RelativeImport
+
+        with RelativeImport("repositories/exllama"):
+            from model import ExLlama, ExLlamaCache, ExLlamaConfig
+    except:
+        logger.error("Could not find repositories/exllama/. Make sure that exllama is cloned inside repositories/ and is up to date.")
+        raise
 
 
 class ExllamaHF(PreTrainedModel):
@@ -68,7 +77,7 @@ class ExllamaHF(PreTrainedModel):
         assert len(model_args) == 0 and len(kwargs) == 0, "extra args is currently not supported"
         if isinstance(pretrained_model_name_or_path, str):
             pretrained_model_name_or_path = Path(pretrained_model_name_or_path)
-            
+
         pretrained_model_name_or_path = Path(f'{shared.args.model_dir}') / Path(pretrained_model_name_or_path)
         config = ExLlamaConfig(pretrained_model_name_or_path / 'config.json')
 
@@ -86,7 +95,7 @@ class ExllamaHF(PreTrainedModel):
         if shared.args.gpu_split:
             config.set_auto_map(shared.args.gpu_split)
             config.gpu_peer_fix = True
-        
+
         # This slowes down a bit but align better with autogptq generation.
         # TODO: Should give user choice to tune the exllama config
         # config.fused_attn = False

requirements.txt

@@ -22,4 +22,6 @@ https://github.com/jllllll/bitsandbytes-windows-webui/releases/download/wheels/b
 llama-cpp-python==0.1.64; platform_system != "Windows"
 https://github.com/abetlen/llama-cpp-python/releases/download/v0.1.64/llama_cpp_python-0.1.64-cp310-cp310-win_amd64.whl; platform_system == "Windows"
 https://github.com/PanQiWei/AutoGPTQ/releases/download/v0.2.2/auto_gptq-0.2.2+cu117-cp310-cp310-win_amd64.whl; platform_system == "Windows"
-https://github.com/PanQiWei/AutoGPTQ/releases/download/v0.2.2/auto_gptq-0.2.2+cu117-cp310-cp310-linux_x86_64.whl; platform_system == "Linux"
+https://github.com/PanQiWei/AutoGPTQ/releases/download/v0.2.2/auto_gptq-0.2.2+cu117-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"
+https://github.com/jllllll/exllama/releases/download/0.0.3/exllama-0.0.3+cu117-cp310-cp310-win_amd64.whl; platform_system == "Windows"
+https://github.com/jllllll/exllama/releases/download/0.0.3/exllama-0.0.3+cu117-cp310-cp310-linux_x86_64.whl; platform_system == "Linux" and platform_machine == "x86_64"

server.py

@@ -237,8 +237,8 @@ def create_model_menus():
                         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='Make sure to inspect the .py files inside the model folder before loading it with this option enabled.')
                         shared.gradio['gptq_for_llama_info'] = gr.Markdown('GPTQ-for-LLaMa is currently 2x faster than AutoGPTQ on some systems. It is installed by default with the one-click installers. Otherwise, it has to be installed manually following the instructions here: [instructions](https://github.com/oobabooga/text-generation-webui/blob/main/docs/GPTQ-models-(4-bit-mode).md#installation-1).')
-                        shared.gradio['exllama_info'] = gr.Markdown('ExLlama has to be installed manually. See the instructions here: [instructions](https://github.com/oobabooga/text-generation-webui/blob/main/docs/ExLlama.md).')
+                        shared.gradio['exllama_info'] = gr.Markdown('For more information, consult the [docs](https://github.com/oobabooga/text-generation-webui/blob/main/docs/ExLlama.md).')
-                        shared.gradio['exllama_HF_info'] = gr.Markdown('ExLlama_HF is a wrapper that lets you use ExLlama like a Transformers model, which means it can use the Transformers samplers. It\'s still a bit buggy, so feel free to help out by fixing issues.\n\nCheck out PR [#2777](https://github.com/oobabooga/text-generation-webui/pull/2777) for more details.')
+                        shared.gradio['exllama_HF_info'] = gr.Markdown('ExLlama_HF is a wrapper that lets you use ExLlama like a Transformers model, which means it can use the Transformers samplers. It\'s a bit slower than the regular ExLlama and doesn\'t support LoRA.')
 
         with gr.Column():
             with gr.Row():