Add --autosplit flag for ExLlamaV2 (oobabooga#5524)

Author: oobabooga

modules/exllamav2.py

@@ -51,18 +51,21 @@ def from_pretrained(self, path_to_model):
 
         model = ExLlamaV2(config)
 
-        split = None
-        if shared.args.gpu_split:
-            split = [float(alloc) for alloc in shared.args.gpu_split.split(",")]
-
-        model.load(split)
-
-        tokenizer = ExLlamaV2Tokenizer(config)
         if shared.args.cache_8bit:
-            cache = ExLlamaV2Cache_8bit(model)
+            cache = ExLlamaV2Cache_8bit(model, lazy=True)
         else:
-            cache = ExLlamaV2Cache(model)
+            cache = ExLlamaV2Cache(model, lazy=True)
 
+        if shared.args.autosplit:
+            model.load_autosplit(cache)
+        else:
+            split = None
+            if shared.args.gpu_split:
+                split = [float(alloc) for alloc in shared.args.gpu_split.split(",")]
+
+            model.load(split)
+
+        tokenizer = ExLlamaV2Tokenizer(config)
         generator = ExLlamaV2StreamingGenerator(model, cache, tokenizer)
 
         result = self()

modules/exllamav2_hf.py

@@ -37,18 +37,22 @@ def __init__(self, config: ExLlamaV2Config):
         super().__init__(PretrainedConfig())
         self.ex_config = config
         self.ex_model = ExLlamaV2(config)
-        split = None
-        if shared.args.gpu_split:
-            split = [float(alloc) for alloc in shared.args.gpu_split.split(",")]
-
-        self.ex_model.load(split)
-        self.generation_config = GenerationConfig()
         self.loras = None
+        self.generation_config = GenerationConfig()
 
         if shared.args.cache_8bit:
-            self.ex_cache = ExLlamaV2Cache_8bit(self.ex_model)
+            self.ex_cache = ExLlamaV2Cache_8bit(self.ex_model, lazy=True)
         else:
-            self.ex_cache = ExLlamaV2Cache(self.ex_model)
+            self.ex_cache = ExLlamaV2Cache(self.ex_model, lazy=True)
+
+        if shared.args.autosplit:
+            self.ex_model.load_autosplit(self.ex_cache)
+        else:
+            split = None
+            if shared.args.gpu_split:
+                split = [float(alloc) for alloc in shared.args.gpu_split.split(",")]
+
+            self.ex_model.load(split)
 
         self.past_seq = None
         if shared.args.cfg_cache:

modules/loaders.py

@@ -78,6 +78,7 @@
         'no_flash_attn',
         'num_experts_per_token',
         'cache_8bit',
+        'autosplit',
         'alpha_value',
         'compress_pos_emb',
         'trust_remote_code',
@@ -89,6 +90,7 @@
         'no_flash_attn',
         'num_experts_per_token',
         'cache_8bit',
+        'autosplit',
         'alpha_value',
         'compress_pos_emb',
         'exllamav2_info',

modules/shared.py

@@ -134,6 +134,7 @@
 # ExLlamaV2
 group = parser.add_argument_group('ExLlamaV2')
 group.add_argument('--gpu-split', type=str, help='Comma-separated list of VRAM (in GB) to use per GPU device for model layers. Example: 20,7,7.')
+group.add_argument('--autosplit', action='store_true', help='Autosplit the model tensors across the available GPUs. This causes --gpu-split to be ignored.')
 group.add_argument('--max_seq_len', type=int, default=2048, help='Maximum sequence length.')
 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.')

modules/ui.py

@@ -76,6 +76,7 @@ def list_model_elements():
         'no_flash_attn',
         'num_experts_per_token',
         'cache_8bit',
+        'autosplit',
         'threads',
         'threads_batch',
         'n_batch',

modules/ui_model_menu.py

@@ -132,6 +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['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.')
                             shared.gradio['num_experts_per_token'] = gr.Number(label="Number of experts per token", value=shared.args.num_experts_per_token, info='Only applies to MoE models like Mixtral.')