LLM: Enable batch generate (world_size>1) in Deepspeed-AutoTP-FastAPI example (#10876)

Enable batch generate (world_size>1) in Deepspeed-AutoTP-FastAPI example.

python/llm/example/GPU/Deepspeed-AutoTP-FastAPI/run_llama2_7b_chat_hf_arc_2_card.sh

@@ -30,6 +30,9 @@ export USE_XETLA=OFF
 export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=2
 export TORCH_LLM_ALLREDUCE=0
 
+export WORLD_SIZE=2
+export MAX_NUM_SEQS=16
+
 mpirun -np $NUM_GPUS --prepend-rank \
         python serving.py --repo-id-or-model-path YOUR_REPO_ID_OR_MODEL_PATH --low-bit 'sym_int4' --port 8000
 

python/llm/example/GPU/Deepspeed-AutoTP-FastAPI/serving.py

@@ -25,6 +25,12 @@ from fastapi import FastAPI, HTTPException
 from pydantic import BaseModel
 import uvicorn
 
+import asyncio, uuid
+from typing import Dict, List, Optional
+
+from transformers.utils import logging
+logger = logging.get_logger(__name__)
+
 def get_int_from_env(env_keys, default):
     """Returns the first positive env value found in the `env_keys` list or the default."""
     for e in env_keys:
@@ -35,6 +41,7 @@ def get_int_from_env(env_keys, default):
 
 local_rank = get_int_from_env(["LOCAL_RANK","PMI_RANK"], "0")
 world_size = get_int_from_env(["WORLD_SIZE","PMI_SIZE"], "1")
+max_num_seqs = get_int_from_env(["MAX_NUM_SEQS"], "16")
 os.environ["RANK"] = str(local_rank)
 os.environ["WORLD_SIZE"] = str(world_size)
 os.environ["MASTER_PORT"] = os.environ.get("MASTER_PORT", "29500")
@@ -70,7 +77,7 @@ def load_model(model_path, low_bit):
 
     model = deepspeed.init_inference(
         model,
-        mp_size=world_size,
+        tensor_parallel={"tp_size": world_size},
         dtype=torch.bfloat16,
         replace_method="auto",
     )
@@ -96,11 +103,22 @@ def load_model(model_path, low_bit):
     init_distributed()
 
     # Load tokenizer
-    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True, padding_side='left')
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = tokenizer.eos_token
 
-def generate_text(prompt: str, n_predict: int = 32):
-    input_ids = tokenizer.encode(prompt, return_tensors="pt").to(f'xpu:{local_rank}')
+def generate_text(prompt: List[str], n_predict = 32):
+    while prompt[-1] == "":
+        prompt = prompt[:-1]
+    if isinstance(n_predict, list):
+        n_predict = max(n_predict)
+        
+    inputs = tokenizer(prompt, return_tensors="pt", padding=True)
+    input_ids = inputs.input_ids.to(f'xpu:{local_rank}')
+    # print(input_ids)
+    attention_mask = inputs.attention_mask.to(f'xpu:{local_rank}')
     output = model.generate(input_ids,
+                            attention_mask=attention_mask,
                             max_new_tokens=n_predict,
                             use_cache=True)
     torch.xpu.synchronize()
@@ -111,19 +129,59 @@ class PromptRequest(BaseModel):
     prompt: str
     n_predict: int = 32  
 
+empty_req = PromptRequest(prompt="", n_predict=0)
+
 app = FastAPI()
 
+request_queue: asyncio.Queue = asyncio.Queue()
+result_dict: Dict[str, str] = {}
+
 @app.post("/generate/")
 async def generate(prompt_request: PromptRequest):
-    if local_rank == 0:
-        object_list = [prompt_request]
+    request_id = str(uuid.uuid4())
+    await request_queue.put((request_id, prompt_request))
+    while True:
+        await asyncio.sleep(0.1)
+        if request_id in result_dict:
+            output_str = result_dict.pop(request_id)
+            return {"generated_text": output_str}
+
+
+async def process_requests():
+    while True:
+        request_ids, prompt_requests = [], []
+        for _ in range(max_num_seqs):
+            if request_queue.empty():
+                break
+            request_id, prompt_request = await request_queue.get()
+            request_ids.append(request_id)
+            prompt_requests.append(prompt_request)
+
+        if local_rank == 0 and prompt_requests:
+            # import pdb
+            # pdb.set_trace()
+            object_list = prompt_requests
+            if len(object_list) < max_num_seqs:
+                object_list = object_list + [empty_req] * (max_num_seqs - len(object_list))
+            logger.info(f"Running: {len(prompt_requests)}, Pending: {request_queue.qsize()}")
             dist.broadcast_object_list(object_list, src=0)
             start_time = time.time()
-        output = generate_text(object_list[0].prompt, object_list[0].n_predict)
+            outputs = generate_text([req.prompt for req in object_list], [req.n_predict for req in object_list])
             generate_time = time.time() - start_time
-        output = output.cpu()
-        output_str = tokenizer.decode(output[0], skip_special_tokens=True)
-        return {"generated_text": output_str, "generate_time": f'{generate_time:.3f}s'}
+            outputs = outputs.cpu()
+            output_strs = tokenizer.batch_decode(outputs, skip_special_tokens=True)
+            output_strs = output_strs[:len(prompt_requests)]
+
+            for request_id, output_str in zip(request_ids, output_strs):
+                result_dict[request_id] = output_str
+
+        await asyncio.sleep(0.1)
+
+
+@app.on_event("startup")
+async def startup_event():
+    asyncio.create_task(process_requests())
+
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(description='Predict Tokens using fastapi by leveraging DeepSpeed-AutoTP')
@@ -143,7 +201,6 @@ if __name__ == "__main__":
         uvicorn.run(app, host="0.0.0.0", port=args.port)
     else:
         while True:
-            object_list = [None]
+            object_list = [None] * max_num_seqs
             dist.broadcast_object_list(object_list, src=0)
-            output = generate_text(object_list[0].prompt, object_list[0].n_predict)
-
+            output = generate_text([req.prompt for req in object_list], [req.n_predict for req in object_list])