[New Features] support dynamic src_length (#7740)

* support dynamic src_length

* revert max_position_embedding

* update doc

* update flask_server

* update max_length control

* update request flask_server

* fix max-position-embeddings

* update error message

* update predictor length init

llm/README.md

@@ -378,8 +378,11 @@ python -m paddle.distributed.launch --gpus "0,1,2,3,4,5,6,7" flask_server.py \
 - `flask_port`: Flask服务端口号，默认8010。
 - 其他参数请参见动态图推理中参数。
 
+此外，如果想通过API脚本的方式跑推理，可参考：`./request_flask_server.py` 文件。
+
 </div></details>
 
+<<<<<<< HEAD
 ## 6. 量化
 
 量化算法可以将模型权重和激活转为更低比特数值类型表示，能够有效减少显存占用和计算开销。下面我们提供GPTQ和PaddleSlim自研的PTQ策略，分别实现WINT4和W8A8量化。更多技术细节详见[量化策略详细教程](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/docs/zh_cn/tutorials/quant/advanced_quantization.md)
@@ -465,6 +468,12 @@ python  finetune_generation.py ./llama/gptq_argument.json
 ### 7.2 转化 Pytorch 权重
 
 PaddleNLP 提供了可自动将 Pytorch 相关的权重转化为 Paddle 权重的接口，代码如下：
+=======
+
+
+### 6. PyTorch模型权重转换
+PaddleNLP 提供了可自动将 PyTorch 相关的权重转化为 Paddle 权重的接口，代码如下：
+>>>>>>> 04dc6251f ([New Features] support dynamic src_length (#7740))
 
 ```python
 from paddlenlp.transformers import AutoModelForCausalLM

llm/flask_server.py

@@ -17,7 +17,7 @@
 import os
 import socket
 from contextlib import closing
-from dataclasses import dataclass, field
+from dataclasses import asdict, dataclass, field
 from time import sleep
 
 import requests
@@ -66,6 +66,7 @@ def __init__(self, args: ServerArgument, predictor: BasePredictor):
             self.args.base_port + port_interval * predictor.tensor_parallel_rank,
             self.args.base_port + port_interval * (predictor.tensor_parallel_rank + 1),
         )
+        self.total_max_length = predictor.config.src_length + predictor.config.max_length
 
         if self.predictor.tensor_parallel_rank == 0:
             # fetch port info
@@ -121,16 +122,44 @@ def streaming(data):
 
                 # build chat template
                 if self.predictor.tokenizer.chat_template is not None:
-                    history = json.loads(history)
+                    if not history:
+                        history = []
+                    # also support history data
+                    elif isinstance(history, str):
+                        history = json.loads(history)
+
                     assert len(history) % 2 == 0
                     chat_query = []
                     for idx in range(0, len(history), 2):
-                        chat_query.append(["", ""])
-                        chat_query[-1][0], chat_query[-1][1] = history[idx]["utterance"], history[idx + 1]["utterance"]
-                    query = [chat_query]
+                        if isinstance(history[idx], str):
+                            chat_query.append([history[idx], history[idx + 1]])
+                        elif isinstance(history[idx], dict):
+                            chat_query.append([history[idx]["utterance"], history[idx + 1]["utterance"]])
+                        else:
+                            raise ValueError(
+                                "history data should be list[str] or list[dict], eg: ['sentence-1', 'sentece-2', ...], or "
+                                "[{'utterance': 'sentence-1'}, {'utterance': 'sentence-2'}, ...]"
+                            )
+
+                    # the input of predictor should be batched.
+                    # batched query: [ [[user, bot], [user, bot], ..., [user]]  ]
+                    query = [chat_query + [[query]]]
 
                 generation_args = data
                 self.predictor.config.max_length = generation_args["max_length"]
+                if "src_length" in generation_args:
+                    self.predictor.config.src_length = generation_args["src_length"]
+
+                if self.predictor.config.src_length + self.predictor.config.max_length > self.total_max_length:
+                    output = {
+                        "error_code": 1,
+                        "error_msg": f"The sum of src_length<{self.predictor.config.src_length}> and "
+                        f"max_length<{self.predictor.config.max_length}> should be smaller than or equal to "
+                        f"the max-total-length<{self.total_max_length}>",
+                    }
+                    yield json.dumps(output, ensure_ascii=False) + "\n"
+                    return
+
                 self.predictor.config.top_p = generation_args["top_p"]
                 self.predictor.config.temperature = generation_args["temperature"]
                 self.predictor.config.top_k = generation_args["top_k"]
@@ -158,13 +187,13 @@ def streaming(data):
         # refer to: https://github.com/pallets/flask/blob/main/src/flask/app.py#L605
         app.run(host="0.0.0.0", port=self.port, threaded=False)
 
-    def start_ui_service(self, args):
+    def start_ui_service(self, args, predictor_args):
         # do not support start ui service in one command
         from multiprocessing import Process
 
         from gradio_ui import main
 
-        p = Process(target=main, args=(args,))
+        p = Process(target=main, args=(args, predictor_args))
         p.daemon = True
         p.start()
 
@@ -184,6 +213,6 @@ def start_ui_service(self, args):
     server = PredictorServer(server_args, predictor)
 
     if server.predictor.tensor_parallel_rank == 0:
-        server.start_ui_service(server_args)
+        server.start_ui_service(server_args, asdict(predictor.config))
 
     server.start_flask_server()

llm/gradio_ui.py

@@ -30,7 +30,29 @@ def setup_args():
     return args
 
 
-def launch(args):
+def create_src_slider(value, maximum):
+    return gr.Slider(
+        minimum=1,
+        maximum=maximum,
+        value=value,
+        step=1,
+        label="Max Src Length",
+        info="最大输入长度。",
+    )
+
+
+def create_max_slider(value, maximum):
+    return gr.Slider(
+        minimum=1,
+        maximum=maximum,
+        value=value,
+        step=1,
+        label="Max Decoding Length",
+        info="生成结果的最大长度。",
+    )
+
+
+def launch(args, default_params: dict = {}):
     """Launch characters dialogue demo."""
 
     def rollback(state):
@@ -42,7 +64,7 @@ def rollback(state):
         shown_context = get_shown_context(context)
         return utterance, shown_context, context, state
 
-    def regen(state, top_k, top_p, temperature, repetition_penalty, max_length):
+    def regen(state, top_k, top_p, temperature, repetition_penalty, max_length, src_length):
         """Regenerate response."""
         context = state.setdefault("context", [])
         if len(context) < 2:
@@ -74,13 +96,13 @@ def begin(utterance, state):
         shown_context = get_shown_context(context)
         return utterance, shown_context, context, state
 
-    def infer(utterance, state, top_k, top_p, temperature, repetition_penalty, max_length):
+    def infer(utterance, state, top_k, top_p, temperature, repetition_penalty, max_length, src_length):
         """Model inference."""
         utterance = utterance.strip().replace("<br>", "\n")
         context = state.setdefault("context", [])
 
         if not utterance:
-            gr.Warning("invalid inputs111")
+            gr.Warning("invalid inputs")
             # gr.Warning("请输入有效问题")
             shown_context = get_shown_context(context)
             return None, shown_context, context, state
@@ -93,11 +115,17 @@ def infer(utterance, state, top_k, top_p, temperature, repetition_penalty, max_l
             "temperature": temperature,
             "repetition_penalty": repetition_penalty,
             "max_length": max_length,
+            "src_length": src_length,
             "min_length": 1,
         }
         res = requests.post(f"http://0.0.0.0:{args.base_port}/api/chat", json=data, stream=True)
         for line in res.iter_lines():
             result = json.loads(line)
+            if result["error_code"] != 0:
+                gr.Warning(result["error_msg"])
+                shown_context = get_shown_context(context)
+                return None, shown_context, context, state
+
             bot_response = result["result"]["response"]
 
             # replace \n with br: https://github.com/gradio-app/gradio/issues/4344
@@ -156,30 +184,57 @@ def get_shown_context(context):
         with gr.Row():
             with gr.Column(scale=1):
                 top_k = gr.Slider(
-                    minimum=1, maximum=100, value=50, step=1, label="Top-k", info="该参数越大，模型生成结果更加随机，反之生成结果更加确定。"
+                    minimum=0,
+                    maximum=default_params.get("top_k", 20),
+                    value=0,
+                    step=1,
+                    label="Top-k",
+                    info="该参数越大，模型生成结果更加随机，反之生成结果更加确定。",
                 )
                 top_p = gr.Slider(
-                    minimum=0, maximum=1, value=0.7, step=0.05, label="Top-p", info="该参数越大，模型生成结果更加随机，反之生成结果更加确定。"
+                    minimum=0,
+                    maximum=1,
+                    value=default_params.get("top_p", 0.7),
+                    step=0.05,
+                    label="Top-p",
+                    info="该参数越大，模型生成结果更加随机，反之生成结果更加确定。",
                 )
                 temperature = gr.Slider(
                     minimum=0.05,
                     maximum=1.5,
-                    value=0.95,
+                    value=default_params.get("temperature", 0.95),
                     step=0.05,
                     label="Temperature",
                     info="该参数越小，模型生成结果更加随机，反之生成结果更加确定。",
                 )
                 repetition_penalty = gr.Slider(
                     minimum=0.1,
                     maximum=10,
-                    value=1.0,
+                    value=default_params.get("repetition_penalty", 1.2),
                     step=0.05,
                     label="Repetition Penalty",
                     info="该参数越大，生成结果重复的概率越低。设置 1 则不开启。",
                 )
-                max_length = gr.Slider(
-                    minimum=1, maximum=1024, value=50, step=1, label="Max Length", info="生成结果的最大长度。"
-                )
+                default_src_length = default_params["src_length"]
+                total_length = default_params["src_length"] + default_params["max_length"]
+                src_length = create_src_slider(default_src_length, total_length)
+                max_length = create_max_slider(50, total_length)
+
+                def src_length_change_event(src_length_value, max_length_value):
+                    return create_max_slider(
+                        min(total_length - src_length_value, max_length_value),
+                        total_length - src_length_value,
+                    )
+
+                def max_length_change_event(src_length_value, max_length_value):
+                    return create_src_slider(
+                        min(total_length - max_length_value, src_length_value),
+                        total_length - max_length_value,
+                    )
+
+                src_length.change(src_length_change_event, inputs=[src_length, max_length], outputs=max_length)
+                max_length.change(max_length_change_event, inputs=[src_length, max_length], outputs=src_length)
+
             with gr.Column(scale=4):
                 state = gr.State({})
                 context_chatbot = gr.Chatbot(label="Context")
@@ -200,7 +255,7 @@ def get_shown_context(context):
                 api_name="chat",
             ).then(
                 infer,
-                inputs=[utt_text, state, top_k, top_p, temperature, repetition_penalty, max_length],
+                inputs=[utt_text, state, top_k, top_p, temperature, repetition_penalty, max_length, src_length],
                 outputs=[utt_text, context_chatbot, raw_context_json, state],
             )
 
@@ -219,13 +274,13 @@ def get_shown_context(context):
             )
             regen_btn.click(
                 regen,
-                inputs=[state, top_k, top_p, temperature, repetition_penalty, max_length],
+                inputs=[state, top_k, top_p, temperature, repetition_penalty, max_length, src_length],
                 outputs=[utt_text, context_chatbot, raw_context_json, state],
                 queue=False,
                 api_name="chat",
             ).then(
                 infer,
-                inputs=[utt_text, state, top_k, top_p, temperature, repetition_penalty, max_length],
+                inputs=[utt_text, state, top_k, top_p, temperature, repetition_penalty, max_length, src_length],
                 outputs=[utt_text, context_chatbot, raw_context_json, state],
             )
 
@@ -237,15 +292,15 @@ def get_shown_context(context):
                 api_name="chat",
             ).then(
                 infer,
-                inputs=[utt_text, state, top_k, top_p, temperature, repetition_penalty, max_length],
+                inputs=[utt_text, state, top_k, top_p, temperature, repetition_penalty, max_length, src_length],
                 outputs=[utt_text, context_chatbot, raw_context_json, state],
             )
 
     block.queue().launch(server_name="0.0.0.0", server_port=args.port, debug=True)
 
 
-def main(args):
-    launch(args)
+def main(args, default_params: dict = {}):
+    launch(args, default_params)
 
 
 if __name__ == "__main__":

llm/predictor.py

@@ -29,7 +29,10 @@
 from utils import (
     dybatch_preprocess,
     get_alibi_slopes,
+    get_default_max_decoding_length,
+    get_default_max_encoding_length,
     get_infer_model_path,
+    get_model_max_position_embeddings,
     get_prefix_tuning_params,
     init_chat_template,
     load_real_time_tokens,
@@ -56,8 +59,8 @@
 class PredictorArgument:
     model_name_or_path: str = field(default=None, metadata={"help": "The directory of model."})
     model_prefix: str = field(default="model", metadata={"help": "the prefix name of static model"})
-    src_length: int = field(default=1024, metadata={"help": "The max length of source text."})
-    max_length: int = field(default=2048, metadata={"help": "the max length for decoding."})
+    src_length: int = field(default=None, metadata={"help": "The max length of source text."})
+    max_length: int = field(default=None, metadata={"help": "the max length for decoding."})
     top_k: int = field(default=0, metadata={"help": "top_k parameter for generation"})
     top_p: float = field(default=0.7, metadata={"help": "top_p parameter for generation"})
     temperature: float = field(default=0.95, metadata={"help": "top_p parameter for generation"})
@@ -732,6 +735,40 @@ def create_predictor(
     if isinstance(tokenizer, LlamaTokenizer) and not tokenizer.pad_token:
         tokenizer.pad_token = tokenizer.unk_token
 
+    config = AutoConfig.from_pretrained(predictor_args.model_name_or_path)
+
+    max_position_embeddings = get_model_max_position_embeddings(config)
+    if max_position_embeddings is None:
+        max_position_embeddings = 2048
+        logger.warning("Can not retrieval `max_position_embeddings` from config.json, use default value 2048")
+
+    if predictor_args.src_length is None:
+        if predictor_args.max_length is None:
+            predictor_args.src_length = get_default_max_encoding_length(config)
+            predictor_args.max_length = get_default_max_decoding_length(config)
+        else:
+            predictor_args.src_length = max_position_embeddings - predictor_args.max_length
+            if predictor_args.src_length <= 0:
+                raise ValueError(
+                    f"--max_length<{predictor_args.max_length}> param should be smaller "
+                    f"than max_position_embeddings<{max_position_embeddings}>"
+                )
+    else:
+        if predictor_args.max_length is None:
+            predictor_args.max_length = max_position_embeddings - predictor_args.src_length
+            if predictor_args.max_length <= 0:
+                raise ValueError(
+                    f"--src_length<{predictor_args.src_length}> param should be smaller "
+                    f"than max_position_embeddings<{max_position_embeddings}>"
+                )
+        else:
+            if predictor_args.src_length + predictor_args.max_length > max_position_embeddings:
+                raise ValueError(
+                    f"The sum of src_length<{predictor_args.src_length}> and "
+                    f"max_length<{predictor_args.max_length}> should be smaller than or equal to "
+                    f"the maximum position embedding size<{max_position_embeddings}>"
+                )
+
     # update config parameter for inference predictor
     if predictor_args.decode_strategy == "greedy_search":
         predictor_args.top_p = 0.0
@@ -925,6 +962,7 @@ def create_predictor(
             predictor = StaticInferencePredictor(predictor_args, cache_kvs_shape, tokenizer=tokenizer)
         else:
             raise ValueError("the `mode` should be one of [dynamic, static]")
+
     return predictor