Add ChatGLM support (#1256)

---------

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

README.md

@@ -219,6 +219,7 @@ Optionally, you can use the following command-line flags:
 | `--no-cache`                                | Set `use_cache` to False while generating text. This reduces the VRAM usage a bit with a performance cost. |
 | `--xformers`                                | Use xformer's memory efficient attention. This should increase your tokens/s. |
 | `--sdp-attention`                           | Use torch 2.0's sdp attention. |
+| `--trust-remote-code`                       | Set trust_remote_code=True while loading a model. Necessary for ChatGLM. |
 
 #### llama.cpp
 

characters/instruction-following/ChatGLM.yaml

@@ -0,0 +1,3 @@
+name: "答:"
+your_name: "[Round <|round|>]\n问:"
+context: ""

download-model.py

@@ -108,7 +108,7 @@ def get_download_links_from_huggingface(model, branch, text_only=False):
             is_safetensors = re.match(".*\.safetensors", fname)
             is_pt = re.match(".*\.pt", fname)
             is_ggml = re.match("ggml.*\.bin", fname)
-            is_tokenizer = re.match("tokenizer.*\.model", fname)
+            is_tokenizer = re.match("(tokenizer|ice).*\.model", fname)
             is_text = re.match(".*\.(txt|json|py|md)", fname) or is_tokenizer
 
             if any((is_pytorch, is_safetensors, is_pt, is_ggml, is_tokenizer, is_text)):

models/config.yaml

@@ -45,3 +45,6 @@ llama-[0-9]*b-4bit$:
 .*koala:
   mode: 'instruct'
   instruction_template: 'Koala'
+.*chatglm:
+  mode: 'instruct'
+  instruction_template: 'ChatGLM'

modules/chat.py

@@ -49,7 +49,8 @@ def generate_chat_prompt(user_input, state, **kwargs):
 
         string = shared.history['internal'][i][0]
         if string not in ['', '<|BEGIN-VISIBLE-CHAT|>']:
-            rows.insert(1, f"{prefix1}{string.strip()}{state['end_of_turn']}\n")
+            this_prefix1 = prefix1.replace('<|round|>', f'{i}')  # for ChatGLM
+            rows.insert(1, f"{this_prefix1}{string.strip()}{state['end_of_turn']}\n")
 
         i -= 1
 
@@ -60,7 +61,8 @@ def generate_chat_prompt(user_input, state, **kwargs):
 
         # Adding the user message
         if len(user_input) > 0:
-            rows.append(f"{prefix1}{user_input}{state['end_of_turn']}\n")
+            this_prefix1 = prefix1.replace('<|round|>', f'{len(shared.history["internal"])}')  # for ChatGLM
+            rows.append(f"{this_prefix1}{user_input}{state['end_of_turn']}\n")
 
         # Adding the Character prefix
         rows.append(apply_extensions(f"{prefix2.strip() if not is_instruct else prefix2}", "bot_prefix"))

modules/models.py

@@ -10,8 +10,8 @@ import numpy as np
 import torch
 import transformers
 from accelerate import infer_auto_device_map, init_empty_weights
-from transformers import (AutoConfig, AutoModelForCausalLM, AutoTokenizer,
-                          BitsAndBytesConfig, LlamaTokenizer)
+from transformers import (AutoConfig, AutoModel, AutoModelForCausalLM,
+                          AutoTokenizer, BitsAndBytesConfig, LlamaTokenizer)
 
 import modules.shared as shared
 from modules import llama_attn_hijack
@@ -44,10 +44,16 @@ def load_model(model_name):
 
     shared.is_RWKV = 'rwkv-' in model_name.lower()
     shared.is_llamacpp = len(list(Path(f'{shared.args.model_dir}/{model_name}').glob('ggml*.bin'))) > 0
+    if 'chatglm' in model_name.lower():
+        LoaderClass = AutoModel
+        trust_remote_code = shared.args.trust_remote_code
+    else:
+        LoaderClass = AutoModelForCausalLM
+        trust_remote_code = False
 
     # Load the model in simple 16-bit mode by default
     if not any([shared.args.cpu, shared.args.load_in_8bit, shared.args.wbits, shared.args.auto_devices, shared.args.disk, shared.args.gpu_memory is not None, shared.args.cpu_memory is not None, shared.args.deepspeed, shared.args.flexgen, shared.is_RWKV, shared.is_llamacpp]):
-        model = AutoModelForCausalLM.from_pretrained(Path(f"{shared.args.model_dir}/{shared.model_name}"), low_cpu_mem_usage=True, torch_dtype=torch.bfloat16 if shared.args.bf16 else torch.float16)
+        model = LoaderClass.from_pretrained(Path(f"{shared.args.model_dir}/{shared.model_name}"), low_cpu_mem_usage=True, torch_dtype=torch.bfloat16 if shared.args.bf16 else torch.float16, trust_remote_code=trust_remote_code)
         if torch.has_mps:
             device = torch.device('mps')
             model = model.to(device)
@@ -79,7 +85,7 @@ def load_model(model_name):
 
     # DeepSpeed ZeRO-3
     elif shared.args.deepspeed:
-        model = AutoModelForCausalLM.from_pretrained(Path(f"{shared.args.model_dir}/{shared.model_name}"), torch_dtype=torch.bfloat16 if shared.args.bf16 else torch.float16)
+        model = LoaderClass.from_pretrained(Path(f"{shared.args.model_dir}/{shared.model_name}"), torch_dtype=torch.bfloat16 if shared.args.bf16 else torch.float16)
         model = deepspeed.initialize(model=model, config_params=ds_config, model_parameters=None, optimizer=None, lr_scheduler=None)[0]
         model.module.eval()  # Inference
         print(f"DeepSpeed ZeRO-3 is enabled: {is_deepspeed_zero3_enabled()}")
@@ -120,6 +126,7 @@ def load_model(model_name):
             params["torch_dtype"] = torch.float32
         else:
             params["device_map"] = 'auto'
+            params["trust_remote_code"] = trust_remote_code
             if shared.args.load_in_8bit and any((shared.args.auto_devices, shared.args.gpu_memory)):
                 params['quantization_config'] = BitsAndBytesConfig(load_in_8bit=True, llm_int8_enable_fp32_cpu_offload=True)
             elif shared.args.load_in_8bit:
@@ -156,7 +163,7 @@ def load_model(model_name):
         if shared.args.load_in_8bit and params.get('max_memory', None) is not None and params['device_map'] == 'auto':
             config = AutoConfig.from_pretrained(checkpoint)
             with init_empty_weights():
-                model = AutoModelForCausalLM.from_config(config)
+                model = LoaderClass.from_config(config)
             model.tie_weights()
             params['device_map'] = infer_auto_device_map(
                 model,
@@ -165,7 +172,7 @@ def load_model(model_name):
                 no_split_module_classes=model._no_split_modules
             )
 
-        model = AutoModelForCausalLM.from_pretrained(checkpoint, **params)
+        model = LoaderClass.from_pretrained(checkpoint, **params)
 
     # Hijack attention with xformers
     if any((shared.args.xformers, shared.args.sdp_attention)):
@@ -185,7 +192,7 @@ def load_model(model_name):
         except:
             pass
     else:
-        tokenizer = AutoTokenizer.from_pretrained(Path(f"{shared.args.model_dir}/{shared.model_name}/"))
+        tokenizer = AutoTokenizer.from_pretrained(Path(f"{shared.args.model_dir}/{shared.model_name}/"), trust_remote_code=trust_remote_code)
 
     print(f"Loaded the model in {(time.time()-t0):.2f} seconds.")
     return model, tokenizer

modules/shared.py

@@ -113,6 +113,7 @@ parser.add_argument('--bf16', action='store_true', help='Load the model with bfl
 parser.add_argument('--no-cache', action='store_true', help='Set use_cache to False while generating text. This reduces the VRAM usage a bit at a performance cost.')
 parser.add_argument('--xformers', action='store_true', help="Use xformer's memory efficient attention. This should increase your tokens/s.")
 parser.add_argument('--sdp-attention', action='store_true', help="Use torch 2.0's sdp attention.")
+parser.add_argument('--trust-remote-code', action='store_true', help="Set trust_remote_code=True while loading a model. Necessary for ChatGLM.")
 
 # llama.cpp
 parser.add_argument('--threads', type=int, default=0, help='Number of threads to use in llama.cpp.')
@@ -162,6 +163,10 @@ if args.cai_chat:
     print("Warning: --cai-chat is deprecated. Use --chat instead.")
     args.chat = True
 
+# Security warnings
+if args.trust_remote_code:
+    print("Warning: trust_remote_code is enabled. This is dangerous.")
+
 
 def is_chat():
     return args.chat