release initial code

Co-authored-by: Ying Sheng <sqy1415@gmail.com>
Co-authored-by: Liangsheng Yin <hnyls2002@gmail.com>
Co-authored-by: Zhiqiang Xie <xiezhq@stanford.edu>
Co-authored-by: parasol-aser <3848358+parasol-aser@users.noreply.github.com>
Co-authored-by: LiviaSun <33578456+ChuyueSun@users.noreply.github.com>
Co-authored-by: Cody Yu <hao.yu.cody@gmail.com>

benchmark/multi_document_qa/README.md

@@ -0,0 +1,47 @@
+## Run benchmark
+
+### Benchmark sglang
+```
+python3 -m sglang.launch_server --model-path codellama/CodeLlama-7b-instruct-hf --port 30000
+```
+
+```
+python3 bench_sglang.py --num-questions 10 --parallel 1
+```
+
+
+### Benchmark vllm
+```
+python3 -m vllm.entrypoints.api_server --tokenizer-mode auto --model codellama/CodeLlama-7b-instruct-hf  --disable-log-requests --port 21000 --gpu 0.97
+```
+
+```
+python3 bench_other.py --backend vllm --num-questions 64
+```
+
+
+### Benchmark guidance
+```
+python3 bench_other.py --backend guidance --num-questions 32 --parallel 1
+```
+
+
+
+### Build dataset
+
+```
+pip install PyPDF2
+python3 build_dataset.py
+```
+
+```python
+import PyPDF2
+
+with open('llama2.pdf', 'rb') as file:
+    reader = PyPDF2.PdfReader(file)
+    text = ''
+    for page_num in range(len(reader.pages)):
+        text += reader.pages[page_num].extract_text()
+    with open('output.txt', 'w') as text_file:
+        text_file.write(text)
+```

benchmark/multi_document_qa/bench_other.py

@@ -0,0 +1,126 @@
+import argparse
+import asyncio
+from concurrent.futures import ThreadPoolExecutor
+from functools import partial
+import json
+import time
+
+from tqdm import tqdm
+import numpy as np
+from sglang.test.test_utils import add_common_other_args_and_parse, call_generate_lightllm, call_generate_vllm, call_generate_srt_raw
+from sglang.utils import read_jsonl, dump_state_text
+
+
+USER_PREFIX = "[INST] "
+USER_SUFFIX = " [/INST]"
+ASSISTANT_PREFIX = ""
+ASSISTANT_SUFFIX = " </s><s>"
+
+
+def multi_document_qa(docs, question, generate):
+    s = USER_PREFIX
+    s += "Pleaes answer a question according to given documents.\n"
+    s += "Question:" + question + "Documents begin.\n"
+
+    s += "".join(docs)
+
+    s += "\nDocuments end."
+    s += ("\n\nBased on the above documents, please answer this question:\n" + question + "\nAnswer in three words or fewer.")
+    s += USER_SUFFIX
+    s += ASSISTANT_PREFIX
+    answer = generate(s, max_tokens=16, stop=None)
+    return answer
+
+
+def main(args):
+    lines = read_jsonl(args.data_path)
+    l = lines[0]
+    arguments = []
+    labels = []
+
+    num_docs = 10
+    if args.backend == "guidance":
+        num_docs = 7  # due to OOM
+
+    for i in range(len(l["questions"][:args.num_questions])):
+        arguments.append({
+            "docs": l["documents"][:num_docs],
+            "question": l["questions"][i],
+        })
+        labels.append(l["answers"][i])
+    states = [None] * len(arguments)
+
+    # Select backend
+    if args.backend == "lightllm":
+        url = f"{args.host}:{args.port}/generate"
+        generate = partial(call_generate_lightllm, url=url, temperature=0)
+    elif args.backend == "vllm":
+        url = f"{args.host}:{args.port}/generate"
+        generate = partial(call_generate_vllm, url=url, temperature=0)
+    elif args.backend == "srt-raw":
+        url = f"{args.host}:{args.port}/generate"
+        generate = partial(call_generate_srt_raw, url=url, temperature=0)
+    elif args.backend == "guidance":
+        from guidance import models, gen
+
+        model = models.LlamaCpp("/home/ubuntu/model_weights/CodeLlama-7b-instruct-hf.gguf", n_gpu_layers=-1, n_ctx=11000)
+
+        def generate(prompt, max_tokens, stop):
+            out = model + prompt + gen(name="answer",
+                max_tokens=max_tokens, temperature=0, stop=stop)
+            return out["answer"]
+
+        # warmup
+        generate("Hello!", max_tokens=8, stop=None)
+    else:
+        raise ValueError(f"Invalid backend: {args.backend}")
+
+    # Run requests
+    def get_one_answer(i):
+        states[i] = multi_document_qa(generate=generate, **arguments[i])
+
+    tic = time.time()
+    if args.parallel == 1:
+        for i in tqdm(range(len(labels))):
+            get_one_answer(i)
+    else:
+        with ThreadPoolExecutor(args.parallel) as executor:
+            executor.map(get_one_answer, list(range(len(labels))))
+    latency = time.time() - tic
+
+    # Compute accuracy
+    print(states)
+    correct = 0
+    for s, label in zip(states, labels):
+        answer = s.lower()
+        if all(x in answer for x in label.lower().split(" ")):
+            correct += 1
+    accuracy = correct / len(labels)
+    print(f"Accuracy: {accuracy:.3f}")
+    print(f"Latency: {latency:.3f}")
+
+    # Write results
+    dump_state_text(f"tmp_output_{args.backend}.txt", states)
+
+    with open(args.result_file, "a") as fout:
+        value = {
+            "task": "multi_document_qa",
+            "backend": args.backend,
+            "num_gpus": 1,
+            "latency": round(latency, 3),
+            "num_requests": args.num_questions,
+            "accuracy": accuracy,
+            "other": {
+                "num_questions": args.num_questions,
+                "parallel": args.parallel,
+            }
+        }
+        fout.write(json.dumps(value) + "\n")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--data-path", type=str, default="questions.jsonl")
+    parser.add_argument("--num-questions", type=int, default=100)
+    args = add_common_other_args_and_parse(parser)
+    main(args)

benchmark/multi_document_qa/bench_sglang.py

@@ -0,0 +1,84 @@
+import argparse
+import json
+import time
+
+import numpy as np
+import sglang as sgl
+from sglang.test.test_utils import add_common_sglang_args_and_parse, select_sglang_backend
+from sglang.utils import read_jsonl, dump_state_text
+
+
+@sgl.function
+def multi_document_qa(s, docs, question):
+    s += sgl.user_begin()
+    s += "Pleaes answer a question according to given documents.\n"
+    s += "Question:" + question + "Documents begin.\n"
+
+    forks = s.fork(len(docs))
+    forks += lambda i: docs[i]
+    forks.join("concate_and_append")
+
+    s += "\nDocuments end."
+    s += ("\n\nBased on the above documents, please answer this question:\n" + question + "\nAnswer in three words or fewer.")
+    s += sgl.user_end()
+    s += sgl.assistant(sgl.gen("answer", max_tokens=16))
+
+
+def main(args):
+    lines = read_jsonl(args.data_path)
+    l = lines[0]
+    arguments = []
+    labels = []
+    for i in range(len(l["questions"][:args.num_questions])):
+        arguments.append({
+            "docs": l["documents"][:10],
+            "question": l["questions"][i],
+        })
+        labels.append(l["answers"][i])
+
+    # Select backend
+    backend = select_sglang_backend(args)
+    sgl.set_default_backend(backend)
+
+    # Run requests
+    tic = time.time()
+    states = multi_document_qa.run_batch(
+        arguments, temperature=0, num_threads=args.parallel)
+    latency = time.time() - tic
+
+    # Compute accuracy 
+    print([s["answer"] for s in states])
+    correct = 0
+    for s, label in zip(states, labels):
+        answer = s["answer"].lower()
+        if all(x in answer for x in label.lower().split(" ")):
+            correct += 1
+    accuracy = correct / len(labels)
+    print(f"Accuracy: {accuracy:.3f}")
+    print(f"Latency: {latency:.3f}")
+
+    # Write results
+    dump_state_text(f"tmp_output_{args.backend}.txt", states)
+
+    with open(args.result_file, "a") as fout:
+        value = {
+            "task": "multi_document_qa",
+            "backend": args.backend,
+            "num_gpus": 1,
+            "latency": round(latency, 3),
+            "num_requests": args.num_questions,
+            "accuracy": accuracy,
+            "other": {
+                "num_questions": args.num_questions,
+                "parallel": args.parallel,
+            }
+        }
+        fout.write(json.dumps(value) + "\n")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--data-path", type=str, default="questions.jsonl")
+    parser.add_argument("--num-questions", type=int, default=100)
+    args = add_common_sglang_args_and_parse(parser)
+    main(args)

benchmark/multi_document_qa/build_dataset.py

@@ -0,0 +1,64 @@
+import json
+
+import transformers
+
+content = "\n".join(
+    open("llama2.txt", 'r', encoding='utf-8', errors='ignore').readlines())
+content = content.replace("\n\n", "\n")
+
+# Count token
+name = "meta-llama/Llama-2-7b-chat-hf"
+t = transformers.AutoTokenizer.from_pretrained(name)
+print(f"num tokens: {len(t.encode(content))}")
+
+# Segment
+SEP = "\n\n"
+parts = content.split(SEP)
+print(f"num segments: {len(parts)}")
+
+segment_len = 1100
+
+segments = []
+tmp = []
+tmp_len = 0
+for i in range(len(parts)):
+    tmp.append(parts[i])
+    tmp_len += len(t.encode(parts[i]))
+
+    if tmp_len > segment_len:
+        segments.append(SEP.join(tmp))
+        tmp = []
+        tmp_len = 0
+
+for i, s in enumerate(segments):
+    print(i, len(t.encode(segments[i])))
+
+# Dump
+with open("questions.jsonl", "w") as fout:
+    fout.write(json.dumps({
+        "documents": segments[:30],
+        "questions": [
+            "What is the name of the fine-tuned LLMs?",
+            "Which figure shows the helpfulness human evaluation results for Llama 2-Chat?",
+            "What is the number of parameters in the largest Llama 2 model?",
+            "What is the batch size of fine-tuning?",
+            "Where can we find the details of potential data contamination?",
+            "What is the full name of MPT?",
+            "What is the power consumption of RSC in Watt?",
+            "How many tokens of data do they train on?",
+            "Which model's release is delayed due to a lack of time to sufficiently red team?",
+            "Which activation function is used in Llama?"
+        ],
+        "answers": [
+            "Llama 2 Chat",
+            "1",
+            "70 B",
+            "64",
+            "A 6",
+            "MosaicML",
+            "400",
+            "2 trillion",
+            "34 B",
+            "SwiGLU",
+        ],
+    }) + "\n")