[fix] fix bench*.py.

bench.py

@@ -4,18 +4,18 @@ from random import randint, seed
 from nanovllm import LLM, SamplingParams
 
 
-def bench_decode(llm, num_seqs, max_input_len, max_output_len):
+def bench_decode(llm, num_seqs, input_len, 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)]
+    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=output_len)
 
     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)
+    total_output_tokens = num_seqs * output_len
     throughput = total_output_tokens / t
-    print(f"[Decode] Output: {total_output_tokens}tok, Time: {t:.2f}s, Throughput: {throughput:.2f}tok/s")
+    print(f"[Decode] Input: {num_seqs}x{input_len}tok, Output: {total_output_tokens}tok, Time: {t:.2f}s, Throughput: {throughput:.2f}tok/s")
 
 
 def bench_prefill(llm, num_seqs, input_len):
@@ -34,8 +34,10 @@ def bench_prefill(llm, num_seqs, input_len):
 
 
 def main():
-    path = os.path.expanduser("~/models/Qwen3-4B-Instruct-2507/")
-    llm = LLM(path, enforce_eager=False, max_model_len=4096)
+    path = os.path.expanduser("~/models/Qwen3-0.6B/")
+    # Note: Qwen3-0.6B max_position_embeddings = 40960, cannot exceed this
+    max_len = 40960
+    llm = LLM(path, enforce_eager=False, max_model_len=max_len, max_num_batched_tokens=max_len)
 
     # Warmup
     llm.generate(["Benchmark: "], SamplingParams())
@@ -45,15 +47,15 @@ def main():
     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=1, input_len=max_len - 1)
     # 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=1, max_input_len=4072, max_output_len=16)
+    # bench_decode(llm, num_seqs=1, input_len=1024, output_len=1024)
+    bench_decode(llm, num_seqs=1, input_len=max_len - 128, output_len=128)  # input + output <= max_len
 
 
 if __name__ == "__main__":