Organize Benchmark (#381)

benchmark/tip_suggestion/README.md

@@ -23,5 +23,11 @@ python3 bench_other.py --backend vllm --num-questions 64
 
 ### Benchmark guidance
 ```
-python3 bench_other.py --backend guidance --num-questions 32 --parallel 1
+python3 bench_other.py --backend guidance --num-questions 32 --parallel 1 --n-ctx 4096 --model-path path/to/gguf
 ```
+
+### Benchmark lmql
+
+```
+python3 bench_other.py --backend lmql --num-questions 32 --parallel 1
+```

benchmark/tip_suggestion/bench_other.py

@@ -6,12 +6,7 @@ from functools import partial
 
 from tqdm import tqdm
 
-from sglang.test.test_utils import (
-    add_common_other_args_and_parse,
-    call_generate_lightllm,
-    call_generate_srt_raw,
-    call_generate_vllm,
-)
+from sglang.test.test_utils import add_common_other_args_and_parse, get_call_generate
 from sglang.utils import dump_state_text, read_jsonl
 
 number = 5
@@ -70,48 +65,43 @@ def main(args):
     states = [None] * len(lines)
 
     # 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 gen, models
-
-        model = models.LlamaCpp(
-            "/home/ubuntu/model_weights/Llama-2-7b-chat.gguf",
-            n_gpu_layers=-1,
-            n_ctx=4096,
-        )
-
-        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}")
+    call_generate = partial(get_call_generate(args), temperature=0)
 
     # Run requests
-    def get_one_answer(i):
-        states[i] = suggest_tips(lines[i]["topic"], generate)
-
     tic = time.time()
-    if args.parallel == 1:
-        for i in tqdm(range(len(lines))):
-            get_one_answer(i)
+    if args.backend != "lmql":
+
+        def get_one_answer(i):
+            states[i] = suggest_tips(lines[i]["topic"], call_generate)
+
+        if args.parallel == 1:
+            for i in tqdm(range(len(lines))):
+                get_one_answer(i)
+        else:
+            with ThreadPoolExecutor(args.parallel) as executor:
+                list(
+                    tqdm(
+                        executor.map(get_one_answer, list(range(len(lines)))),
+                        total=len(lines),
+                    )
+                )
+
     else:
-        with ThreadPoolExecutor(args.parallel) as executor:
-            executor.map(get_one_answer, list(range(len(lines))))
+        import asyncio
+
+        from lmql_funcs import suggest_tips_async
+
+        async def get_one_answer_async(i):
+            states[i] = await suggest_tips_async(lines[i]["topic"], call_generate)
+
+        batches = []
+        for i in range(0, len(lines), args.parallel):
+            batches.append(list(range(i, min(i + args.parallel, len(lines)))))
+        loop = asyncio.get_event_loop()
+        for batch in tqdm(batches):
+            loop.run_until_complete(
+                asyncio.gather(*[get_one_answer_async(i) for i in batch])
+            )
     latency = time.time() - tic
 
     # Compute accuracy

benchmark/tip_suggestion/lmql_funcs.py

@@ -0,0 +1,50 @@
+number = 5
+
+
+async def expand_tip_async(topic, tip, generate):
+    s = (
+        """Please expand a tip for a topic into a detailed paragraph.
+
+Topic: staying healthy
+Tip: Regular Exercise
+Paragraph: Incorporate physical activity into your daily routine. This doesn't necessarily mean intense gym workouts; it can be as simple as walking, cycling, or yoga. Regular exercise helps in maintaining a healthy weight, improves cardiovascular health, boosts mental health, and can enhance cognitive function, which is crucial for fields that require intense intellectual engagement.
+
+Topic: building a campfire
+Tip: Choose the Right Location
+Paragraph: Always build your campfire in a safe spot. This means selecting a location that's away from trees, bushes, and other flammable materials. Ideally, use a fire ring if available. If you're building a fire pit, it should be on bare soil or on a bed of stones, not on grass or near roots which can catch fire underground. Make sure the area above is clear of low-hanging branches.
+
+Topic: writing a blog post
+Tip: structure your content effectively
+Paragraph: A well-structured post is easier to read and more enjoyable. Start with an engaging introduction that hooks the reader and clearly states the purpose of your post. Use headings and subheadings to break up the text and guide readers through your content. Bullet points and numbered lists can make information more digestible. Ensure each paragraph flows logically into the next, and conclude with a summary or call-to-action that encourages reader engagement.
+
+Topic: """
+        + topic
+        + "\nTip: "
+        + tip
+        + "\nParagraph:"
+    )
+    return await generate(s, max_tokens=128, stop="\n\n")
+
+
+async def suggest_tips_async(topic, generate):
+    s = "Please act as a helpful assistant. Your job is to provide users with useful tips on a specific topic.\n"
+    s += "USER: Give some tips for " + topic + ".\n"
+    s += (
+        "ASSISTANT: Okay. Here are "
+        + str(number)
+        + " concise tips, each under 8 words:\n"
+    )
+
+    tips = []
+    for i in range(1, 1 + number):
+        s += f"{i}."
+        # NOTE: stop is different due to lmql does not support a list of stop tokens
+        tip = await generate(s, max_tokens=24, stop=".\n")
+        s += tip + ".\n"
+        tips.append(tip)
+
+    paragraphs = [await expand_tip_async(topic, tip, generate=generate) for tip in tips]
+
+    for i in range(1, 1 + number):
+        s += f"Tip {i}:" + paragraphs[i - 1] + "\n"
+    return s