adapt to sglang v0.5.2rc1 on dcu

benchmark/multi_turn_chat/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 --n-ctx 4096 --model-path path/to/gguf
+```
+
+Benchmark Llama-7B (long output)
+
+```
+python3 bench_other.py --tokenizer meta-llama/Llama-2-7b-chat-hf --backend guidance --parallel 1 --n-ctx 4096 --model-path path/to/gguf --long
+```

benchmark/multi_turn_chat/bench_other.py

@@ -0,0 +1,93 @@
+import json
+import time
+from argparse import ArgumentParser
+from concurrent.futures import ThreadPoolExecutor
+from functools import partial
+
+from data_gen import gen_arguments
+from tqdm import tqdm
+from vllm.transformers_utils.tokenizer import get_tokenizer
+
+from sglang.test.test_utils import add_common_other_args_and_parse, get_call_generate
+from sglang.utils import dump_state_text
+
+
+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
+
+    call_generate = partial(get_call_generate(args), temperature=0)
+
+    def get_one_answer(i):
+        states[i] = multi_turns(generate=call_generate, **multi_qas[i])
+
+    tic = time.perf_counter()
+    if args.parallel == 1:
+        for i in tqdm(range(len(multi_qas))):
+            get_one_answer(i)
+    else:
+        with ThreadPoolExecutor(args.parallel) as executor:
+            rets = list(
+                tqdm(
+                    executor.map(get_one_answer, list(range(len(multi_qas)))),
+                    total=len(multi_qas),
+                )
+            )
+            for _ in rets:
+                pass
+
+    latency = time.perf_counter() - 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_turn_chat",
+            "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_turn_chat/bench_sglang.py

@@ -0,0 +1,79 @@
+import json
+import time
+from argparse import ArgumentParser
+
+from data_gen import gen_arguments
+from vllm.transformers_utils.tokenizer import get_tokenizer
+
+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
+
+
+@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):
+    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.perf_counter()
+    states = multi_turns.run_batch(
+        multi_qas,
+        temperature=0,
+        backend=backend,
+        num_threads=args.parallel,
+        progress_bar=True,
+    )
+    latency = time.perf_counter() - 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_turn_chat",
+            "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
+
+    print(args)
+    main(args)

benchmark/multi_turn_chat/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

benchmark/multi_turn_chat/long_prompt_multi_turn.py

@@ -0,0 +1,129 @@
+import json
+import random
+import time
+from argparse import ArgumentParser
+from pathlib import Path
+
+from tqdm import tqdm
+
+import sglang as sgl
+from sglang.srt.hf_transformers_utils import get_tokenizer
+from sglang.test.test_utils import (
+    add_common_sglang_args_and_parse,
+    select_sglang_backend,
+)
+from sglang.utils import dump_state_text
+
+
+def gen_prompt(tokenizer, token_num):
+    all_available_tokens = list(tokenizer.get_vocab().values())
+    selected_tokens = random.choices(all_available_tokens, k=token_num)
+    ret = tokenizer.decode(selected_tokens)
+    return ret
+
+
+def get_cache_path(args):
+    # Create cache directory under ~/.cache/sglang
+    cache_dir = Path.home() / ".cache" / "sglang"
+
+    # Create a unique cache filename based on the arguments that affect generation
+    cache_key = f"qa_{args.num_qa}_{args.turns}_{args.system_prompt_len}_{args.len_q}_{args.len_a}_{args.tokenizer.replace('/', '_')}.json"
+    return cache_dir / cache_key
+
+
+def gen_arguments(args, tokenizer):
+    cache_path = get_cache_path(args)
+
+    # Try to load from cache first
+    if cache_path.exists():
+        print(f"Loading cached arguments from {cache_path}")
+        with open(cache_path, "r") as f:
+            return json.load(f)
+
+    print("Generating new arguments...")
+    # First progress bar for system prompts
+    multi_qas = []
+    for _ in tqdm(range(args.num_qa), desc="Generating system prompts"):
+        multi_qas.append(
+            {"system_prompt": gen_prompt(tokenizer, args.system_prompt_len), "qas": []}
+        )
+
+    # Nested progress bars for QA pairs
+    for i in tqdm(range(args.num_qa), desc="Generating QA pairs"):
+        qas = multi_qas[i]["qas"]
+        for j in range(args.turns):
+            qas.append(
+                {
+                    "prompt": gen_prompt(tokenizer, args.len_q),
+                    "new_tokens": args.len_a,
+                }
+            )
+
+    # Save to cache
+    cache_path.parent.mkdir(parents=True, exist_ok=True)
+    with open(cache_path, "w") as f:
+        json.dump(multi_qas, f)
+    print(f"Cached arguments saved to {cache_path}")
+
+    return multi_qas
+
+
+@sgl.function
+def multi_turns(s, system_prompt, qas):
+    s += system_prompt
+
+    for i, qa in enumerate(qas):
+        s += qa["prompt"]
+        s += sgl.gen(max_tokens=qa["new_tokens"], ignore_eos=True)
+
+
+def main(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.perf_counter()
+    states = multi_turns.run_batch(
+        multi_qas,
+        temperature=0,
+        backend=backend,
+        num_threads="auto",
+        progress_bar=True,
+    )
+    latency = time.perf_counter() - 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_turn_system_prompt_chat",
+            "backend": args.backend,
+            "latency": round(latency, 3),
+            "num_requests": args.num_qa,
+            "num_turns": args.turns,
+            "other": {
+                "parallel": args.parallel,
+            },
+        }
+        fout.write(json.dumps(value) + "\n")
+
+
+if __name__ == "__main__":
+    parser = ArgumentParser()
+    parser.add_argument("--turns", type=int, default=8)
+    parser.add_argument("--num-qa", type=int, default=128)
+    parser.add_argument("--system-prompt-len", type=int, default=2048)
+    parser.add_argument("--len-q", type=int, default=32)
+    parser.add_argument("--len-a", type=int, default=128)
+    parser.add_argument(
+        "--tokenizer", type=str, default="meta-llama/Meta-Llama-3-8B-Instruct"
+    )
+    parser.add_argument("--trust-remote-code", action="store_true")
+    args = add_common_sglang_args_and_parse(parser)
+
+    print(args)
+    main(args)