Merge branch 'llm_fxp' into falcon

pilot/configs/model_config.py

@@ -36,6 +36,7 @@ LLM_MODEL_CONFIG = {
     "text2vec-base": os.path.join(MODEL_PATH, "text2vec-base-chinese"),
     "guanaco-33b-merged": os.path.join(MODEL_PATH, "guanaco-33b-merged"),
     "falcon-40b": os.path.join(MODEL_PATH, "falcon-40b"),
+    "gorilla-7b": os.path.join(MODEL_PATH, "gorilla-7b"),
     "proxyllm": "proxyllm",
 }
 

pilot/model/adapter.py

@@ -123,6 +123,20 @@ class FalconAdapater(BaseLLMAdaper):
                 **from_pretrained_kwagrs
             )
         return model, tokenizer
+    
+    
+class GorillaAdapter(BaseLLMAdaper):
+    """TODO Support guanaco"""
+
+    def match(self, model_path: str):
+        return "gorilla" in model_path
+
+    def loader(self, model_path: str, from_pretrained_kwargs: dict):
+        tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False)
+        model = AutoModelForCausalLM.from_pretrained(
+            model_path, low_cpu_mem_usage=True, **from_pretrained_kwargs
+        )
+        return model, tokenizer
 
 
 class CodeGenAdapter(BaseLLMAdaper):
@@ -181,6 +195,7 @@ register_llm_model_adapters(VicunaLLMAdapater)
 register_llm_model_adapters(ChatGLMAdapater)
 register_llm_model_adapters(GuanacoAdapter)
 register_llm_model_adapters(FalconAdapater)
+register_llm_model_adapters(GorillaAdapter)
 # TODO Default support vicuna, other model need to tests and Evaluate
 
 # just for test, remove this later

pilot/model/llm_out/gorilla_llm.py

@@ -0,0 +1,58 @@
+import torch
+
+@torch.inference_mode()
+def generate_stream(
+    model, tokenizer, params, device, context_len=42048, stream_interval=2
+):
+    """Fork from https://github.com/ShishirPatil/gorilla/blob/main/inference/serve/gorilla_cli.py"""
+    prompt = params["prompt"]
+    l_prompt = len(prompt)
+    max_new_tokens = int(params.get("max_new_tokens", 1024))
+    stop_str = params.get("stop", None)
+
+    input_ids = tokenizer(prompt).input_ids
+    output_ids = list(input_ids)
+    input_echo_len = len(input_ids)
+    max_src_len = context_len - max_new_tokens - 8
+    input_ids = input_ids[-max_src_len:]
+    past_key_values = out = None
+
+    for i in range(max_new_tokens):
+        if i == 0:
+            out = model(torch.as_tensor([input_ids], device=device), use_cache=True)
+            logits = out.logits
+            past_key_values = out.past_key_values
+        else:     
+            out = model(
+                input_ids=torch.as_tensor([[token]], device=device),
+                use_cache=True,
+                past_key_values=past_key_values,
+            )
+            logits = out.logits
+            past_key_values = out.past_key_values
+
+        last_token_logits = logits[0][-1]
+
+        probs = torch.softmax(last_token_logits, dim=-1)
+        token = int(torch.multinomial(probs, num_samples=1))
+        output_ids.append(token)
+
+
+        if token == tokenizer.eos_token_id:
+            stopped = True
+        else:
+            stopped = False
+
+        if i % stream_interval == 0 or i == max_new_tokens - 1 or stopped:
+            tmp_output_ids = output_ids[input_echo_len:]
+            output = tokenizer.decode(tmp_output_ids, skip_special_tokens=True, spaces_between_special_tokens=False,)
+            pos = output.rfind(stop_str, l_prompt)
+            if pos != -1:
+                output = output[:pos]
+                stopped = True
+            yield output
+
+        if stopped:
+            break
+
+    del past_key_values

pilot/server/chat_adapter.py

@@ -94,6 +94,7 @@ class GuanacoChatAdapter(BaseChatAdpter):
 
         return guanaco_generate_stream
 
+   
 class FalconChatAdapter(BaseChatAdpter):
     """Model chat adapter for Guanaco"""
 
@@ -119,6 +120,7 @@ register_llm_model_chat_adapter(VicunaChatAdapter)
 register_llm_model_chat_adapter(ChatGLMChatAdapter)
 register_llm_model_chat_adapter(GuanacoChatAdapter)
 register_llm_model_adapters(FalconChatAdapter)
+register_llm_model_chat_adapter(GorillaChatAdapter)
 
 # Proxy model for test and develop, it's cheap for us now.
 register_llm_model_chat_adapter(ProxyllmChatAdapter)