[inference] Async dynamic batching (#4894)

* finish input and output logic

* add generate

* test forward

* 1

colossalai/inference/dynamic_batching/io_struct.py

@@ -102,17 +102,21 @@ def mark_finished_req(self, eos_id):
                 has_new_finish = True
         return has_new_finish
 
-    def filter_finished(self):
+    def filter_finished(self)->List[Req]:
         """
         Filter finished requests from the batch, the finished ones will be removed from 'reqs'.
         """
         # TODO: the logic of return should be defined here.
         unfinished_req = []
+        finished_req = []
         for req in self.reqs:
             if not req.has_generate_finished:
-                unfinished_req.append(req)
+                unfinished_req.append(req)   
+            else:
+                finished_req.append(req)             
         self.reqs = unfinished_req
         self.id_to_reqs = {req.request_id: req for req in self.reqs}
+        return finished_req
 
     def is_clear(self):
         return len(self.reqs) == 0

colossalai/inference/manager.py

@@ -1,5 +1,6 @@
 import time
 from typing import List
+import asyncio
 
 from .dynamic_batching.infer_batch import InferBatch
 from .dynamic_batching.io_struct import Batch, Req
@@ -8,6 +9,8 @@
 from .dynamic_batching.stats import Stats
 from .tensor_parallel import TPInferEngine
 
+from transformers import AutoTokenizer
+_FAST_LLAMA_TOKENIZER = "hf-internal-testing/llama-tokenizer"
 
 class DynamicBatchManager:
     def __init__(
@@ -54,6 +57,20 @@ def add_req(self, prompt_ids: List[int], sampling_params: SamplingParams, reques
         self.req_queue.append(req)
         return
 
+    def add_input(self, request_id, sampling_params, input_ids):
+        """
+        Encode and Add new input to req queue. support one sequence input for now.
+        """
+        prompt_ids = self.tokenizer.encode(input_ids)
+        prompt_len = len(prompt_ids)
+        if prompt_len > self.engine.max_input_len:
+            raise ValueError(
+                f"the input prompt token len {prompt_len} is too long > {self.engine.max_input_len}"
+            )
+        sampling_params.stop_sentences_to_token_ids(self.tokenizer)
+        self.add_req(prompt_ids, sampling_params, request_id)
+        return
+
     def abort(self, request_id):
         if self.running_batch is not None:
             for req in self.running_batch.reqs:
@@ -66,13 +83,15 @@ def abort(self, request_id):
                 req.aborted = True
         return
 
-    def loop_for_fwd(self):
+    async def loop_for_fwd(self):
         """
         The main loop for a dynamic batching process.
         """
         counter_count = 0
-        while self.running_batch is not None or self.req_queue.waiting_req_list:
-            self._step()
+        #self.running_batch is not None or self.req_queue.waiting_req_list
+        while True:
+            async for item in self._step():
+                yield item
             counter_count += 1
             if self.running_batch is not None:
                 if counter_count % self.mem_usage_interval == 0:
@@ -87,6 +106,26 @@ def loop_for_fwd(self):
             if self.running_batch is None:
                 time.sleep(0.1)  # 10ms
 
+    def _set_tokenizer(self, tokenizer=None, tokenizer_name: str = "", trust_remote_code: bool = False, use_fast:bool = True,):
+        if tokenizer is not None:
+            self.tokenizer = tokenizer 
+        else:
+            if "llama" in tokenizer_name.lower() and use_fast == True:
+                print(
+                "For some LLaMA-based models, initializing the fast tokenizer may "
+                "take a long time. To eliminate the initialization time, consider "
+                f"using '{_FAST_LLAMA_TOKENIZER}' instead of the original "
+                "tokenizer. This is done automatically in Colossalai.")
+
+                tokenizer_name = _FAST_LLAMA_TOKENIZER  
+
+            try: 
+                self.tokenizer = AutoTokenizer.from_pretrained(tokenizer_name, use_fast=use_fast,trust_remote_code=trust_remote_code)
+            except TypeError as e:
+                use_fast = False
+                self.tokenizer = AutoTokenizer.from_pretrained(tokenizer_name, use_fast=use_fast,trust_remote_code=trust_remote_code)
+
+
     def _step(self):
         """
         Logic for handling requests
@@ -97,32 +136,33 @@ def _step(self):
             if new_batch is not None:
                 self.stats_tool.count_prompt_tokens(new_batch)
                 self.running_batch = new_batch
-                self._prefill_batch(self.running_batch)
+                yield from self._prefill_batch(self.running_batch)
                 self._filter_runing_batch()
                 self.has_wait_tokens = 0
             return
 
         if self.has_wait_tokens < self.max_wait_tokens:
             self.stats_tool.count_output_tokens(self.running_batch)
-            self._decode_batch(self.running_batch)
+            yield from self._decode_batch(self.running_batch)
             self._filter_runing_batch()
             self.has_wait_tokens += 1
             return
         else:
             new_mini_batch = self.req_queue.generate_new_batch(self.running_batch)
             if new_mini_batch is not None:
                 self.stats_tool.count_prompt_tokens(new_mini_batch)
-                self._prefill_batch(new_mini_batch)
+                yield from self._prefill_batch(new_mini_batch)
                 if not new_mini_batch.is_clear():
                     self._merge_batch(self.running_batch, new_mini_batch)
                     self.running_batch.merge(new_mini_batch)
                 self.has_wait_tokens = 0
+
             else:
                 self.stats_tool.count_output_tokens(self.running_batch)
-                self._decode_batch(self.running_batch)
+                yield from self._decode_batch(self.running_batch)
                 self._filter_runing_batch()
                 self.has_wait_tokens += 1
-
+         
         return
 
     def _init_batch(self, batch: Batch, dtype="fp16"):
@@ -158,7 +198,8 @@ def _prefill_batch(self, batch):
         req_to_out_token_id = ans
         self._add_token_id_to_req(batch, req_to_out_token_id)
         has_new_finished_req = batch.mark_finished_req(self.eos_id)
-        self._handle_finish_req(batch, has_new_finished_req)
+        yield from self._handle_finish_req(batch, has_new_finished_req)
+
         # delete finished reqs
 
     def _decode_batch(self, batch: Batch):
@@ -169,7 +210,7 @@ def _decode_batch(self, batch: Batch):
         req_to_out_token_id = ans
         self._add_token_id_to_req(batch, req_to_out_token_id)
         has_new_finished_req = batch.mark_finished_req(self.eos_id)
-        self._handle_finish_req(batch, has_new_finished_req)
+        yield from self._handle_finish_req(batch, has_new_finished_req)
 
     def _filter_batch(self, batch: Batch):
         batch_id = batch.batch_id
@@ -201,11 +242,13 @@ def _remove_batch(self, batch):
 
     def _handle_finish_req(self, batch: Batch, has_new_finished_req):
         if has_new_finished_req:
-            batch.filter_finished()
+            finished_reqs=batch.filter_finished()
             if batch.is_clear():
                 self._remove_batch(batch)
             else:
                 self._filter_batch(batch)
+            yield from self._output_process(finished_reqs)
+
 
     def _filter_runing_batch(self):
         if self.running_batch is not None and self.running_batch.is_clear():
@@ -218,26 +261,47 @@ def _add_token_id_to_req(self, batch: Batch, req_ans):
             req.output_metadata_list.append(new_gen_metadata)
         return
 
+    async def _output_process(self, finished_reqs: List[Req]):
+        """
+        Process the output of a batch.
+        """
+        for req in finished_reqs:
+            output = self.tokenizer.decode(req.output_ids)
+            yield output, req.request_id, req.output_metadata_list 
+
     def clean_up(self):
         # this logic should be implemented in the future.
         pass
 
+    async def generate(self,request_id,prompt_id,sampling_params):
+        """
+        Generate the output of a request.
+        """
+        self.add_input(request_id,prompt_id,sampling_params)
+
 
 def start_dynamic_batching(args, tp_engine, waiting_req_list):
-    # try:
-    batch_manager = DynamicBatchManager(
-        tp_engine=tp_engine,
-        max_total_token_num=args.max_total_token_num,
-        batch_max_tokens=args.batch_max_tokens,
-        eos_id=args.eos_id,
-        log_stats=not args.disable_log_stats,
-        log_stats_interval=args.log_stats_interval,
-        waiting_req_list=waiting_req_list,
-    )
-
-    # except Exception:
-    #     batch_manager.clean_up()
-    #     raise
-
-    batch_manager.loop_for_fwd()
-    return
+    try:
+        batch_manager = DynamicBatchManager(
+            tp_engine=tp_engine,
+            max_total_token_num=args.max_total_token_num,
+            batch_max_tokens=args.batch_max_tokens,
+            eos_id=args.eos_id,
+            log_stats=not args.disable_log_stats,
+            log_stats_interval=args.log_stats_interval,
+            waiting_req_list=waiting_req_list,
+        )
+
+    except Exception:
+        batch_manager.clean_up()
+        raise
+
+    batch_manager._set_tokenizer(tokenizer_name = tp_engine.model.__class__.__name__)
+    prod_task = asyncio.create_task(batch_manager.add_input(4,sampling_params=SamplingParams(),input_ids="hello world"))
+
+    asyncio.run(prod_task)
+
+    for item in batch_manager.loop_for_fwd():
+        print(item)
+
+    return batch_manager

colossalai/inference/test_async.py

@@ -0,0 +1,33 @@
+import asyncio
+
+shared_list = []
+
+async def producer():
+    for i in range(5):
+        await asyncio.sleep(1)  # 模拟异步获取数据的操作
+        shared_list.append(i)
+        print(f"Produced {i}")
+
+async def consumer():
+    last_index = 0
+    while True:
+        await asyncio.sleep(0.5)  # 为了不使循环过于紧凑，增加了小的延迟
+        if last_index < len(shared_list):
+            item = shared_list[last_index]
+            print(f"Consumed {item}")
+            yield item
+            last_index += 1
+
+async def main():
+    # 创建生产者和消费者任务
+    prod_task = asyncio.create_task(producer())
+
+    # 等待生产者任务完成
+    await prod_task
+
+    async for  data in consumer():
+        print(data)
+    # 为了示例的目的，我们只等待一段时间，然后停止消费者
+    await asyncio.sleep(5)
+
+asyncio.run(main())

tests/test_infer/test_dynamic_batching/test_forward.py

@@ -42,15 +42,21 @@ def run():
     waiting_list.append(req2)
     waiting_list.append(req3)
     waiting_list.append(req4)
-
+    
     llama_config = LlamaConfig(num_hidden_layers=2, bos_token_id=0, eos_token_id=1, vocab_size=1200, hidden_size=1024)
     model = LlamaForCausalLM(llama_config)
     model = model.half()
 
     shard_config = ShardConfig(enable_tensor_parallelism=True if TP_SIZE > 1 else False, inference_only=True)
 
     infer_engine = TPInferEngine(model, shard_config, MAX_BATCH_SIZE, MAX_INPUT_LEN, MAX_OUTPUT_LEN)
-    start_dynamic_batching(arg, tp_engine=infer_engine, waiting_req_list=waiting_list)
+    manager = start_dynamic_batching(arg, tp_engine=infer_engine, waiting_req_list=waiting_list)
+    manager._set_tokenizer(tokenizer_name = model.__class__.__name__)
+    result_generator = manager.loop_for_fwd()
+    for result in result_generator:
+        print(result)
+
+
 
 
 def check_dynamic_forward(rank, world_size, port):