[bench] Added vllm vs nano-vllm bench.

bench.py

@@ -2,30 +2,58 @@ import os
 import time
 from random import randint, seed
 from nanovllm import LLM, SamplingParams
-# from vllm import LLM, SamplingParams
+
+
+def bench_decode(llm, num_seqs, max_input_len, max_output_len):
+    """Benchmark decode performance (original test)"""
+    seed(0)
+    prompt_token_ids = [[randint(0, 10000) for _ in range(randint(100, max_input_len))] for _ in range(num_seqs)]
+    sampling_params = [SamplingParams(temperature=0.6, ignore_eos=True, max_tokens=randint(100, max_output_len)) for _ in range(num_seqs)]
+
+    t = time.time()
+    llm.generate(prompt_token_ids, sampling_params, use_tqdm=False)
+    t = time.time() - t
+    total_output_tokens = sum(sp.max_tokens for sp in sampling_params)
+    throughput = total_output_tokens / t
+    print(f"[Decode] Output: {total_output_tokens}tok, Time: {t:.2f}s, Throughput: {throughput:.2f}tok/s")
+
+
+def bench_prefill(llm, num_seqs, input_len):
+    """Benchmark prefill performance"""
+    seed(0)
+    # Fixed length input, minimal output to focus on prefill
+    prompt_token_ids = [[randint(0, 10000) for _ in range(input_len)] for _ in range(num_seqs)]
+    sampling_params = SamplingParams(temperature=0.6, ignore_eos=True, max_tokens=1)
+
+    t = time.time()
+    llm.generate(prompt_token_ids, sampling_params, use_tqdm=False)
+    t = time.time() - t
+    total_input_tokens = num_seqs * input_len
+    throughput = total_input_tokens / t
+    print(f"[Prefill] Input: {total_input_tokens}tok ({num_seqs}x{input_len}), Time: {t:.2f}s, Throughput: {throughput:.2f}tok/s")
 
 
 def main():
-    seed(0)
-    num_seqs = 256
-    max_input_len = 1024
-    max_ouput_len = 1024
-
-    path = os.path.expanduser("~/huggingface/Qwen3-0.6B/")
+    path = os.path.expanduser("~/models/Qwen3-4B-Instruct-2507/")
     llm = LLM(path, enforce_eager=False, max_model_len=4096)
 
-    prompt_token_ids = [[randint(0, 10000) for _ in range(randint(100, max_input_len))] for _ in range(num_seqs)]
-    sampling_params = [SamplingParams(temperature=0.6, ignore_eos=True, max_tokens=randint(100, max_ouput_len)) for _ in range(num_seqs)]
-    # uncomment the following line for vllm
-    # prompt_token_ids = [dict(prompt_token_ids=p) for p in prompt_token_ids]
-
+    # Warmup
     llm.generate(["Benchmark: "], SamplingParams())
-    t = time.time()
-    llm.generate(prompt_token_ids, sampling_params, use_tqdm=False)
-    t = (time.time() - t)
-    total_tokens = sum(sp.max_tokens for sp in sampling_params)
-    throughput = total_tokens / t
-    print(f"Total: {total_tokens}tok, Time: {t:.2f}s, Throughput: {throughput:.2f}tok/s")
+
+    print("=" * 60)
+    print("Prefill Benchmark")
+    print("=" * 60)
+    bench_prefill(llm, num_seqs=1, input_len=1024)
+    bench_prefill(llm, num_seqs=1, input_len=2048)
+    bench_prefill(llm, num_seqs=1, input_len=4095)
+    bench_prefill(llm, num_seqs=16, input_len=1024)
+    bench_prefill(llm, num_seqs=64, input_len=1024)
+
+    print("=" * 60)
+    print("Decode Benchmark")
+    print("=" * 60)
+    bench_decode(llm, num_seqs=1, max_input_len=1024, max_output_len=1024)
+    bench_decode(llm, num_seqs=256, max_input_len=1024, max_output_len=1024)
 
 
 if __name__ == "__main__":