Json Decode && Mutl-Turns (#4)

benchmark/multi_turns/README.md

@@ -0,0 +1,66 @@
+### Benchmark sglang
+
+Run llama-7b
+
+```
+python3 -m sglang.launch_server --model-path meta-llama/Llama-2-7b-chat-hf --port 30000
+```
+
+Run mixtral-8x7b
+(When there is a CUDA out-of-memory error, try to reduce the `--mem-fraction-static`)
+
+```
+python3 -m sglang.launch_server --model-path mistralai/Mixtral-8x7B-Instruct-v0.1 --port 30000 --tp-size 8
+```
+
+Benchmark(short output)
+
+```
+python3 bench_sglang.py --tokenizer meta-llama/Llama-2-7b-chat-hf
+```
+
+Benchmark(long output)
+
+```
+python3 bench_sglang.py --tokenizer meta-llama/Llama-2-7b-chat-hf --long
+```
+
+### Benchmark vLLM
+
+Run llama-7b
+
+```
+python3 -m vllm.entrypoints.api_server --tokenizer-mode auto --model meta-llama/Llama-2-7b-chat-hf  --disable-log-requests --port 21000
+```
+
+Run mixtral-8x7b
+
+```
+python3 -m vllm.entrypoints.api_server --tokenizer-mode auto --model mistralai/Mixtral-8x7B-Instruct-v0.1 --disable-log-requests --port 21000 --tensor-parallel-size 8
+```
+
+Benchmark(short output)
+
+```
+python3 bench_other.py --tokenizer meta-llama/Llama-2-7b-chat-hf --backend vllm
+```
+
+Benchmark(long output)
+
+```
+python3 bench_other.py --tokenizer meta-llama/Llama-2-7b-chat-hf --backend vllm --long
+```
+
+### Benchmark guidance
+
+Benchmark llama-7b(short output)
+
+```
+python3 bench_other.py --tokenizer meta-llama/Llama-2-7b-chat-hf --backend guidance --parallel 1
+```
+
+Benchmark llama-7b(long output)
+
+```
+python3 bench_other.py --tokenizer meta-llama/Llama-2-7b-chat-hf --backend guidance --parallel 1 --long
+```

benchmark/multi_turns/bench_other.py

@@ -0,0 +1,133 @@
+import json
+import time
+from argparse import ArgumentParser
+from concurrent.futures import ThreadPoolExecutor
+
+import requests
+from sglang.test.test_utils import add_common_other_args_and_parse
+from sglang.utils import dump_state_text
+from tqdm import tqdm
+from vllm.transformers_utils.tokenizer import get_tokenizer
+
+from data_gen import gen_arguments
+
+
+def get_generate(args):
+    # Select backend
+    if args.backend == "vllm":
+        url = f"{args.host}:{args.port}/generate"
+
+        def generate(prompt, max_tokens, stop=None, temperature=0, url=url, n=1):
+            data = {
+                "prompt": prompt,
+                "temperature": temperature,
+                "max_tokens": max_tokens,
+                "ignore_eos": True,
+                "stop": stop,
+                "stream": False,
+                "n": n,
+            }
+            res = requests.post(url, json=data)
+            assert res.status_code == 200
+            return res.json()["text"][0][len(prompt) :]
+
+    elif args.backend == "guidance":
+        from guidance import gen, models
+
+        model = models.LlamaCpp(
+            "/home/ubuntu/model_weights/Llama-2-7b-chat-hf/ggml-model-f16.gguf",
+            n_gpu_layers=-1,
+            n_ctx=4096,
+        )
+
+        def generate(prompt, max_tokens, stop=None):
+            out = (
+                model
+                + prompt
+                + gen(name="answer", max_tokens=max_tokens, temperature=0, stop=stop)
+            )
+            return out["answer"]
+
+        # warmup
+        for _ in range(3):
+            generate("Hello!" * 10, max_tokens=64, stop=None)
+    else:
+        raise ValueError(f"Invalid backend: {args.backend}")
+
+    return generate
+
+
+def multi_turns(generate, qas):
+    s = ""
+    for qa in qas:
+        s += qa["prompt"]
+        s += generate(s, max_tokens=qa["new_tokens"]) 
+
+    return s
+
+
+def main(args):
+    print(args)
+
+    tokenizer = get_tokenizer(args.tokenizer, trust_remote_code=args.trust_remote_code)
+
+    multi_qas = gen_arguments(args, tokenizer)
+
+    states = [None] * args.num_qa
+
+    generate = get_generate(args)
+
+    def get_one_answer(i):
+        states[i] = multi_turns(generate=generate, **multi_qas[i])
+
+    tic = time.time()
+    if args.parallel == 1:
+        for i in tqdm(range(len(multi_qas))):
+            get_one_answer(i)
+    else:
+        with ThreadPoolExecutor(args.parallel) as executor:
+            rets = executor.map(get_one_answer, list(range(len(multi_qas))))
+            for _ in rets:
+                pass
+
+    latency = time.time() - tic
+
+    # Compute accuracy
+    print(f"Latency: {latency:.3f}")
+
+    dump_state_text(f"tmp_output_{args.backend}.txt", states)
+
+    with open(args.result_file, "a") as fout:
+        value = {
+            "task": "multi_turns",
+            "backend": args.backend,
+            "num_gpus": 1,
+            "latency": round(latency, 3),
+            "num_requests": args.num_qa,
+            "num_turns": args.turns,
+            "other": {
+                "parallel": args.parallel,
+                "output_mode": "long" if args.long else "short",
+            },
+        }
+        fout.write(json.dumps(value) + "\n")
+
+
+if __name__ == "__main__":
+    parser = ArgumentParser()
+    parser.add_argument("--turns", type=int, default=4)
+    parser.add_argument("--num-qa", type=int, default=20)
+    parser.add_argument("--min-len-q", type=int, default=256)
+    parser.add_argument("--max-len-q", type=int, default=512)
+    parser.add_argument("--min-len-a", type=int, default=4)
+    parser.add_argument("--max-len-a", type=int, default=8)
+    parser.add_argument("--tokenizer", type=str, required=True)
+    parser.add_argument("--trust-remote-code", action="store_true")
+    parser.add_argument("--long", action="store_true")
+    args = add_common_other_args_and_parse(parser)
+
+    if args.long:
+        args.min_len_a = 256
+        args.max_len_a = 512
+        args.num_qa = 20
+    main(args)

benchmark/multi_turns/bench_sglang.py

@@ -0,0 +1,77 @@
+import json
+import time
+from argparse import ArgumentParser
+
+import sglang as sgl
+from sglang.test.test_utils import (
+    add_common_sglang_args_and_parse,
+    select_sglang_backend,
+)
+from sglang.utils import dump_state_text
+from vllm.transformers_utils.tokenizer import get_tokenizer
+
+from data_gen import gen_arguments
+
+
+@sgl.function
+def multi_turns(s, qas):
+    for qa in qas:
+        s += qa["prompt"]
+        s += sgl.gen(max_tokens=qa["new_tokens"], ignore_eos=True) 
+
+
+def main(args):
+    print(args)
+
+    tokenizer = get_tokenizer(args.tokenizer, trust_remote_code=args.trust_remote_code)
+
+    multi_qas = gen_arguments(args, tokenizer)
+
+    backend = select_sglang_backend(args)
+
+    tic = time.time()
+    states = multi_turns.run_batch(
+        multi_qas, temperature=0, backend=backend, num_threads=args.parallel
+    )
+    for state in states:
+        state.sync()
+    latency = time.time() - tic
+
+    print(f"Latency: {latency:.3f}")
+
+    dump_state_text(f"tmp_output_{args.backend}.txt", states)
+
+    with open(args.result_file, "a") as fout:
+        value = {
+            "task": "multi_turns",
+            "backend": args.backend,
+            "num_gpus": 1,
+            "latency": round(latency, 3),
+            "num_requests": args.num_qa,
+            "num_turns": args.turns,
+            "other": {
+                "parallel": args.parallel,
+                "output_mode": "long" if args.long else "short",
+            },
+        }
+        fout.write(json.dumps(value) + "\n")
+
+
+if __name__ == "__main__":
+    parser = ArgumentParser()
+    parser.add_argument("--turns", type=int, default=4)
+    parser.add_argument("--num-qa", type=int, default=20)
+    parser.add_argument("--min-len-q", type=int, default=256)
+    parser.add_argument("--max-len-q", type=int, default=512)
+    parser.add_argument("--min-len-a", type=int, default=4)
+    parser.add_argument("--max-len-a", type=int, default=8)
+    parser.add_argument("--tokenizer", type=str, required=True)
+    parser.add_argument("--trust-remote-code", action="store_true")
+    parser.add_argument("--long", action="store_true")
+    args = add_common_sglang_args_and_parse(parser)
+
+    if args.long:
+        args.min_len_a = 256
+        args.max_len_a = 512
+        args.num_qa = 20
+    main(args)

benchmark/multi_turns/data_gen.py

@@ -0,0 +1,29 @@
+import random
+import string
+
+random.seed(42)
+
+
+def gen_prompt(tokenizer, token_num):
+    cha_set = string.ascii_letters + string.digits
+    ret = "".join(random.choices(cha_set, k=token_num))
+    while len(tokenizer(ret).input_ids) < token_num:
+        ret += random.choice(cha_set)
+    return ret
+
+
+def gen_arguments(args, tokenizer):
+    multi_qas = [{"qas": []} for _ in range(args.num_qa)]
+    for i in range(args.num_qa):
+        qas = multi_qas[i]["qas"]
+        for _ in range(args.turns):
+            prompt_len = random.randint(args.min_len_q, args.max_len_q)
+            new_tokens = random.randint(args.min_len_a, args.max_len_a)
+            qas.append(
+                {
+                    "prompt": gen_prompt(tokenizer, prompt_len),
+                    "new_tokens": new_tokens,
+                }
+            )
+
+    return multi_qas