Upload folder using huggingface_hub

.gitattributes

@@ -67,3 +67,8 @@ sim_traverse_bs/traverse_bs_util_std.nsys-rep filter=lfs diff=lfs merge=lfs -tex
 traverse_bs_util_std.nsys-rep filter=lfs diff=lfs merge=lfs -text
 traverse_bs_util_sim_decoding.nsys-rep filter=lfs diff=lfs merge=lfs -text
 traverse_bs_util_sim_prefill.nsys-rep filter=lfs diff=lfs merge=lfs -text
+sim_traverse_bs/traverse_bs_util_sim_decoding.nsys-rep filter=lfs diff=lfs merge=lfs -text
+sim_traverse_bs/traverse_bs_util_sim_decoding_1024.nsys-rep filter=lfs diff=lfs merge=lfs -text
+sim_traverse_bs/traverse_bs_util_sim_prefill.nsys-rep filter=lfs diff=lfs merge=lfs -text
+sim_traverse_bs/traverse_bs_util_sim_prefill_1152.nsys-rep filter=lfs diff=lfs merge=lfs -text
+std_traverse_bs/traverse_bs_util_std.nsys-rep filter=lfs diff=lfs merge=lfs -text

bench_vllm_v1.py

@@ -0,0 +1,242 @@
+import os
+import time
+import statistics
+from typing import List, Tuple
+
+from vllm import LLM, SamplingParams
+from transformers import AutoTokenizer
+
+# ==== 关键：V1 开关（也可在外部 export）====
+os.environ.setdefault("VLLM_USE_V1", "1")
+
+# 可选：多进程安全（官方建议的 spawn）
+os.environ.setdefault("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
+
+# V1 metrics 类型（Counter/Gauge/Histogram/Vector）
+try:
+    from vllm.v1.metrics.reader import Counter, Gauge, Histogram, Vector  # type: ignore
+except Exception:
+    # 老版本兜底：若导入失败，后面代码会用 hasattr/反射方式尽力兼容
+    Counter = Gauge = Histogram = Vector = type("X", (), {})  # dummy
+
+# =============== 可调参数 ===============
+TP = 1
+DTYPE = "bfloat16"
+GPU_MEM_UTIL = 0.9
+WARMUP = 2
+RUNS = 5
+INPUT_TOKENS = 1024
+OUTPUT_TOKENS = 128
+SEED = 1234
+TEMPERATURE = 0.0
+MODELS = [
+    # "meta-llama/Meta-Llama-3-70B",
+    # "meta-llama/Meta-Llama-3-8B",
+    "Qwen/Qwen2-1.5B",
+]
+# =====================================
+
+def build_exact_token_prompt(tokenizer, target_len: int) -> str:
+    base_text = "You are a helpful assistant. " \
+                "Please analyze the following input and respond succinctly. "
+    chunk = " ".join(["data"] * 100) + ". "
+    text = base_text + chunk * 200
+
+    lo, hi = 0, len(text)
+    target_ids = None
+    while lo <= hi:
+        mid = (lo + hi) // 2
+        ids = tokenizer(text[:mid], add_special_tokens=False)["input_ids"]
+        if len(ids) == target_len:
+            target_ids = ids
+            break
+        if len(ids) < target_len:
+            lo = mid + 1
+        else:
+            hi = mid - 1
+
+    if target_ids is None:
+        ids = tokenizer(text[:lo], add_special_tokens=False)["input_ids"]
+        if len(ids) > target_len:
+            target_ids = ids[:target_len]
+        else:
+            filler = " data"
+            while len(ids) < target_len:
+                ids = tokenizer(tokenizer.decode(ids) + filler, add_special_tokens=False)["input_ids"]
+            target_ids = ids[:target_len]
+
+    prompt = tokenizer.decode(target_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)
+    assert len(tokenizer(prompt, add_special_tokens=False)["input_ids"]) == target_len
+    return prompt
+
+# --------- V1 metrics 抽取工具 ---------
+TTFT_METRIC_NAME = "vllm:time_to_first_token_seconds"
+TPOT_METRIC_NAME = "vllm:time_per_output_token_seconds"  # per-output-token latency
+
+def _iter_children_of_vector(vec_obj):
+    """
+    兼容不同版本 Vector 的内部字段名（children/metrics/series/...）。
+    """
+    for attr in ("children", "metrics", "series", "values", "samples", "items"):
+        if hasattr(vec_obj, attr):
+            val = getattr(vec_obj, attr)
+            if isinstance(val, dict):
+                for v in val.values():
+                    yield v
+            else:
+                try:
+                    for v in val:
+                        yield v
+                except TypeError:
+                    pass
+
+def _collect_hist_sum_count(metrics, metric_name: str):
+    """
+    聚合（可能带 labels 的）Histogram 的 sum 与 count。
+    既兼容直接 Histogram，也兼容 Vector[Histogram] 形式。
+    """
+    total_sum = 0.0
+    total_count = 0.0
+    for m in metrics:
+        mname = getattr(m, "name", None)
+        if mname != metric_name:
+            continue
+
+        # 直接 Histogram
+        if isinstance(m, Histogram) or m.__class__.__name__ == "Histogram":
+            total_sum += float(getattr(m, "sum", 0.0))
+            total_count += float(getattr(m, "count", 0.0))
+            continue
+
+        # Vector[Histogram]
+        if isinstance(m, Vector) or m.__class__.__name__ == "Vector":
+            for child in _iter_children_of_vector(m):
+                if isinstance(child, Histogram) or child.__class__.__name__ == "Histogram":
+                    total_sum += float(getattr(child, "sum", 0.0))
+                    total_count += float(getattr(child, "count", 0.0))
+    return total_sum, total_count
+
+def _metrics_snapshot(llm):
+    """拿一份快照，并解析出 TTFT/TPOT 的 (sum, count)。"""
+    mets = llm.get_metrics()  # V1: 返回 Metric 列表（包含 Histogram/Vector 等）
+    ttft_sum, ttft_cnt = _collect_hist_sum_count(mets, TTFT_METRIC_NAME)
+    tpot_sum, tpot_cnt = _collect_hist_sum_count(mets, TPOT_METRIC_NAME)
+    return {
+        "ttft_sum": ttft_sum, "ttft_cnt": ttft_cnt,
+        "tpot_sum": tpot_sum, "tpot_cnt": tpot_cnt,
+    }
+
+def _metrics_delta(before: dict, after: dict):
+    """计算 after - before 的增量。"""
+    return {
+        "ttft_sum": after["ttft_sum"] - before["ttft_sum"],
+        "ttft_cnt": after["ttft_cnt"] - before["ttft_cnt"],
+        "tpot_sum": after["tpot_sum"] - before["tpot_sum"],
+        "tpot_cnt": after["tpot_cnt"] - before["tpot_cnt"],
+    }
+
+def measure_one_model(model_name: str) -> None:
+    print(f"\n===== Testing model: {model_name} | TP={TP} | dtype={DTYPE} (V1 metrics) =====")
+    tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=True, trust_remote_code=True)
+    prompt = build_exact_token_prompt(tokenizer, INPUT_TOKENS)
+
+    llm = LLM(
+        model=model_name,
+        tensor_parallel_size=TP,
+        dtype=DTYPE,
+        trust_remote_code=True,
+        gpu_memory_utilization=GPU_MEM_UTIL,
+        max_num_seqs=1,
+        max_model_len=8192,
+        disable_log_stats=False,               # <<< 加这一行
+    )
+
+    sampling_params = SamplingParams(
+        max_tokens=OUTPUT_TOKENS,
+        temperature=TEMPERATURE,
+        top_p=1.0,
+        seed=SEED,
+        n=1,
+        stop=None
+    )
+
+    # 预热
+    for _ in range(WARMUP):
+        _ = llm.generate([prompt], sampling_params)
+
+    # 统计容器
+    latencies: List[float] = []
+    out_token_speeds: List[float] = []
+    total_token_speeds: List[float] = []
+    ttft_list: List[float] = []            # 从 V1 metrics 读出的 TTFT（秒）
+    decode_tps_list: List[float] = []      # 从 V1 metrics 读出的 decoding tokens/s
+
+    for i in range(RUNS):
+        # 取前快照
+        snap_before = _metrics_snapshot(llm)
+
+        # 发请求并计时（方便对照）
+        t0 = time.perf_counter()
+        outputs = llm.generate([prompt], sampling_params)
+        t1 = time.perf_counter()
+        dur = t1 - t0
+
+        # 取后快照
+        snap_after = _metrics_snapshot(llm)
+        delta = _metrics_delta(snap_before, snap_after)
+
+        # 解析本轮的 TTFT 与 decoding tokens/s
+        # avg_ttft = sum_delta / count_delta
+        if delta["ttft_cnt"] > 0:
+            ttft = delta["ttft_sum"] / delta["ttft_cnt"]
+            ttft_list.append(ttft)
+        else:
+            ttft_list.append(float("nan"))
+
+        # avg_time_per_output_token = sum_delta / count_delta
+        if delta["tpot_cnt"] > 0:
+            avg_tpot = delta["tpot_sum"] / delta["tpot_cnt"]  # seconds / token
+            decode_tokens_per_s = 1.0 / avg_tpot
+            decode_tps_list.append(decode_tokens_per_s)
+        else:
+            decode_tps_list.append(float("nan"))
+
+        # 与原脚本同口径的吞吐统计（供对照）
+        text_out = outputs[0].outputs[0].text
+        out_token_cnt = len(tokenizer(text_out, add_special_tokens=False)["input_ids"])
+        latencies.append(dur)
+        out_token_speeds.append(out_token_cnt / dur if dur > 0 else float("inf"))
+        total_token_speeds.append((INPUT_TOKENS + out_token_cnt) / dur if dur > 0 else float("inf"))
+
+        print(
+            f"Run {i+1}/{RUNS}: latency={dur:.3f}s, "
+            f"gen_tokens={out_token_cnt}, "
+            f"gen_toks/s(raw)={out_token_speeds[-1]:.2f}, "
+            f"TTFT(V1)={ttft_list[-1]:.3f}s, "
+            f"decode_toks/s(V1)={decode_tps_list[-1]:.2f}"
+        )
+
+    def fmt_stats(x: List[float]) -> Tuple[float, float, float]:
+        xs = [v for v in x if (v == v)]  # 过滤 NaN
+        if not xs:
+            return (float("nan"), float("nan"), float("nan"))
+        return (statistics.mean(xs), statistics.median(xs), statistics.quantiles(xs, n=10)[-1])  # p90
+
+    lat_mean, lat_med, lat_p90 = fmt_stats(latencies)
+    gen_mean, gen_med, gen_p90 = fmt_stats(out_token_speeds)
+    tot_mean, tot_med, tot_p90 = fmt_stats(total_token_speeds)
+    ttft_mean, ttft_med, ttft_p90 = fmt_stats(ttft_list)
+    dtps_mean, dtps_med, dtps_p90 = fmt_stats(decode_tps_list)
+
+    print("\n--- Summary ---")
+    print(f"Latency (s)         : mean={lat_mean:.3f}, median={lat_med:.3f}, p90={lat_p90:.3f}")
+    print(f"Gen tok/s (raw)     : mean={gen_mean:.2f}, median={gen_med:.2f}, p90={gen_p90:.2f}")
+    print(f"Total tok/s (raw)   : mean={tot_mean:.2f}, median={tot_med:.2f}, p90={tot_p90:.2f}")
+    print(f"TTFT (V1 metrics, s): mean={ttft_mean:.3f}, median={ttft_med:.3f}, p90={ttft_p90:.3f}")
+    print(f"Decode tok/s (V1)   : mean={dtps_mean:.2f}, median={dtps_med:.2f}, p90={dtps_p90:.2f}")
+    print("(V1 的 TTFT/Decode tok/s 来自 Prometheus 直方图指标；raw 为你原先用时长推算的数值)")
+
+if __name__ == "__main__":
+    print(f"CUDA_VISIBLE_DEVICES = {os.getenv('CUDA_VISIBLE_DEVICES')}")
+    for m in MODELS:
+        measure_one_model(m)

sim_traverse_bs/plot.py

@@ -0,0 +1,96 @@
+import numpy as np
+import matplotlib.pyplot as plt
+
+def kv_cache_size_bytes(bs, seq):
+    """计算 Qwen2ForCausalLM 模型的 KV cache 大小（字节）"""
+    n_layers = 28
+    n_kv_heads = 2
+    hidden_size = 1536
+    num_attention_heads = 12
+    head_dim = hidden_size // num_attention_heads
+    bytes_per_elem = 2  # bfloat16
+    return bs * seq * n_kv_heads * head_dim * 2 * n_layers * bytes_per_elem
+
+# ===== Prefill 数据 =====
+# seq=640
+avg_ns_prefill_640 = np.array([
+    1418534991, 1033939782, 494241378, 252646178, 135127951,
+    69450854, 42089250, 35031755, 28423503, 22718985, 15558545
+], dtype=np.float64)
+bs_prefill_640 = np.array([1024, 512, 256, 128, 64, 32, 16, 8, 4, 2, 1])
+seq_prefill_640 = 640
+throughput_prefill_640 = bs_prefill_640 * seq_prefill_640 / (avg_ns_prefill_640 * 1e-9)
+norm_prefill_640 = throughput_prefill_640 / throughput_prefill_640.max()
+kv_prefill_640 = kv_cache_size_bytes(bs_prefill_640, seq_prefill_640) / 1024**3  # 改成 GB
+
+# seq=1152
+# avg_ns_prefill_1152 = np.array([
+#     2423366261, 1417657581, 1040325509, 389808438, 200051514,
+#     110723882, 61663721, 39934013, 30382401, 21185162, 15818426
+# ], dtype=np.float64)
+# bs_prefill_1152 = np.array([1024, 512, 256, 128, 64, 32, 16, 8, 4, 2, 1])
+avg_ns_prefill_1152 = np.array([
+    1417657581, 1040325509, 389808438, 200051514,
+    110723882, 61663721, 39934013, 30382401, 21185162, 15818426
+], dtype=np.float64)
+bs_prefill_1152 = np.array([ 512, 256, 128, 64, 32, 16, 8, 4, 2, 1])
+seq_prefill_1152 = 1152
+throughput_prefill_1152 = bs_prefill_1152 * seq_prefill_1152 / (avg_ns_prefill_1152 * 1e-9)
+norm_prefill_1152 = throughput_prefill_1152 / throughput_prefill_1152.max()
+kv_prefill_1152 = kv_cache_size_bytes(bs_prefill_1152, seq_prefill_1152) / 1024**3  # 改成 GB
+
+# ===== Decoding 数据 =====
+# seq=512
+avg_ns_decoding_512 = np.array([
+    25285906551, 12679311252, 7440608085, 4841914697, 4272889441,
+    3997075015, 3825710172, 3726603655, 3648294896, 3635960724, 3319210677
+], dtype=np.float64)
+bs_decoding_512 = np.array([1024, 512, 256, 128, 64, 32, 16, 8, 2, 4, 1])
+seq_decoding_512 = 512
+throughput_decoding_512 = bs_decoding_512 * seq_decoding_512 / (avg_ns_decoding_512 * 1e-9)
+norm_decoding_512 = throughput_decoding_512 / throughput_decoding_512.max()
+kv_decoding_512 = kv_cache_size_bytes(bs_decoding_512, seq_decoding_512) / 1024**3  # 改成 GB
+
+# seq=1024
+# avg_ns_decoding_1024 = np.array([
+#     65019709544, 30598009745, 16552100314, 11165129518, 8835508288,
+#     8037503827, 8020861613, 7502439278, 7275415153, 7204870191, 6423331403
+# ], dtype=np.float64)
+# bs_decoding_1024 = np.array([1024, 512, 256, 128, 64, 32, 16, 8, 4, 2, 1])
+avg_ns_decoding_1024 = np.array([
+     30598009745, 16552100314, 11165129518, 8835508288,
+    8037503827, 8020861613, 7502439278, 7275415153, 7204870191, 6423331403
+], dtype=np.float64)
+bs_decoding_1024 = np.array([ 512, 256, 128, 64, 32, 16, 8, 4, 2, 1])
+seq_decoding_1024 = 1024
+throughput_decoding_1024 = bs_decoding_1024 * seq_decoding_1024 / (avg_ns_decoding_1024 * 1e-9)
+norm_decoding_1024 = throughput_decoding_1024 / throughput_decoding_1024.max()
+kv_decoding_1024 = kv_cache_size_bytes(bs_decoding_1024, seq_decoding_1024) / 1024**3  # 改成 GB
+
+# ===== 绘图 =====
+fig, axes = plt.subplots(1, 2, figsize=(14, 5))
+
+# Prefill 子图
+axes[0].plot(kv_prefill_640, norm_prefill_640, marker='o', label="seq=640")
+axes[0].plot(kv_prefill_1152, norm_prefill_1152, marker='s', label="seq=1152")
+axes[0].set_xscale('log')
+axes[0].set_xlabel("KV Cache Size (GB, log scale)")
+axes[0].set_ylabel("Normalized GPU Utilization")
+axes[0].set_title("Prefill")
+axes[0].grid(True, which="both", ls="--", alpha=0.5)
+axes[0].legend()
+
+# Decoding 子图
+axes[1].plot(kv_decoding_512, norm_decoding_512, marker='o', label="seq=512")
+axes[1].plot(kv_decoding_1024, norm_decoding_1024, marker='s', label="seq=1024")
+axes[1].set_xscale('log')
+axes[1].set_xlabel("KV Cache Size (GB, log scale)")
+axes[1].set_ylabel("Normalized GPU Utilization")
+axes[1].set_title("Decoding")
+axes[1].grid(True, which="both", ls="--", alpha=0.5)
+axes[1].legend()
+
+plt.suptitle("Normalized GPU Utilization vs KV Cache Size")
+plt.savefig("kv_cache_vs_util_gb.png", dpi=300, bbox_inches='tight')
+plt.savefig("kv_cache_vs_util_gb.pdf", dpi=300, bbox_inches='tight')
+plt.show()

sim_traverse_bs/traverse_bs_util_sim_decoding.log

@@ -0,0 +1,390 @@
+WARNING: CPU IP/backtrace sampling not supported, disabling.
+Try the 'nsys status --environment' command to learn more.
+
+WARNING: CPU context switch tracing not supported, disabling.
+Try the 'nsys status --environment' command to learn more.
+
+INFO 08-13 19:21:37 [__init__.py:235] Automatically detected platform cuda.
+CUDA_VISIBLE_DEVICES = 3
+--- vLLM V1 基准测试（含 NVTX 标记）---
+模型: Qwen/Qwen2-1.5B
+批量大小: [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024]
+场景: ['prefill1_decode512']
+------------------------------------------------------------
+加载分词器/模型中...
+INFO 08-13 19:21:46 [config.py:1604] Using max model len 4096
+INFO 08-13 19:21:47 [config.py:2434] Chunked prefill is enabled with max_num_batched_tokens=8192.
+INFO 08-13 19:21:52 [__init__.py:235] Automatically detected platform cuda.
+INFO 08-13 19:21:54 [core.py:572] Waiting for init message from front-end.
+INFO 08-13 19:21:54 [core.py:71] Initializing a V1 LLM engine (v0.10.0) with config: model='Qwen/Qwen2-1.5B', speculative_config=None, tokenizer='Qwen/Qwen2-1.5B', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config={}, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=4096, download_dir=None, load_format=LoadFormat.AUTO, tensor_parallel_size=1, pipeline_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=False, kv_cache_dtype=auto,  device_config=cuda, decoding_config=DecodingConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_backend=''), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None), seed=0, served_model_name=Qwen/Qwen2-1.5B, num_scheduler_steps=1, multi_step_stream_outputs=True, enable_prefix_caching=True, chunked_prefill_enabled=True, use_async_output_proc=True, pooler_config=None, compilation_config={"level":3,"debug_dump_path":"","cache_dir":"","backend":"","custom_ops":[],"splitting_ops":["vllm.unified_attention","vllm.unified_attention_with_output","vllm.mamba_mixer2"],"use_inductor":true,"compile_sizes":[],"inductor_compile_config":{"enable_auto_functionalized_v2":false},"inductor_passes":{},"use_cudagraph":true,"cudagraph_num_of_warmups":1,"cudagraph_capture_sizes":[512,504,496,488,480,472,464,456,448,440,432,424,416,408,400,392,384,376,368,360,352,344,336,328,320,312,304,296,288,280,272,264,256,248,240,232,224,216,208,200,192,184,176,168,160,152,144,136,128,120,112,104,96,88,80,72,64,56,48,40,32,24,16,8,4,2,1],"cudagraph_copy_inputs":false,"full_cuda_graph":false,"max_capture_size":512,"local_cache_dir":null}
+INFO 08-13 19:21:56 [parallel_state.py:1102] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, TP rank 0, EP rank 0
+WARNING 08-13 19:21:56 [topk_topp_sampler.py:59] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
+INFO 08-13 19:21:56 [gpu_model_runner.py:1843] Starting to load model Qwen/Qwen2-1.5B...
+INFO 08-13 19:21:56 [gpu_model_runner.py:1875] Loading model from scratch...
+INFO 08-13 19:21:56 [cuda.py:290] Using Flash Attention backend on V1 engine.
+INFO 08-13 19:21:57 [weight_utils.py:296] Using model weights format ['*.safetensors']
+INFO 08-13 19:21:57 [weight_utils.py:349] No model.safetensors.index.json found in remote.
+
+
+
+
+INFO 08-13 19:21:58 [default_loader.py:262] Loading weights took 0.63 seconds
+INFO 08-13 19:21:58 [gpu_model_runner.py:1892] Model loading took 2.9105 GiB and 1.878581 seconds
+INFO 08-13 19:22:04 [backends.py:530] Using cache directory: /home/cy/.cache/vllm/torch_compile_cache/40b61c71e9/rank_0_0/backbone for vLLM's torch.compile
+INFO 08-13 19:22:04 [backends.py:541] Dynamo bytecode transform time: 5.72 s
+INFO 08-13 19:22:09 [backends.py:161] Directly load the compiled graph(s) for dynamic shape from the cache, took 4.036 s
+INFO 08-13 19:22:10 [monitor.py:34] torch.compile takes 5.72 s in total
+INFO 08-13 19:22:11 [gpu_worker.py:255] Available KV cache memory: 12.81 GiB
+INFO 08-13 19:22:11 [kv_cache_utils.py:833] GPU KV cache size: 479,536 tokens
+INFO 08-13 19:22:11 [kv_cache_utils.py:837] Maximum concurrency for 4,096 tokens per request: 117.07x
+
+INFO 08-13 19:22:13 [gpu_model_runner.py:2485] Graph capturing finished in 2 secs, took 0.49 GiB
+INFO 08-13 19:22:13 [core.py:193] init engine (profile, create kv cache, warmup model) took 14.62 seconds
+模型加载完成。
+
+===== 场景：prefill1_decode512 | prefill=1, decode=512 =====
+
+--- 批量大小 bs=1 ---
+
+
+执行时间: 3.3194 s
+实际平均输入 tokens: 1.00（目标 1）
+生成总 tokens: 512
+吞吐(生成tokens/秒): 154.24
+TTFT (V1 metrics): 0.0190 s
+解码吞吐 (V1 metrics): 155.07 tok/s
+
+--- 批量大小 bs=2 ---
+
+
+执行时间: 3.6484 s
+实际平均输入 tokens: 1.00（目标 1）
+生成总 tokens: 1024
+吞吐(生成tokens/秒): 280.67
+TTFT (V1 metrics): 0.0120 s
+解码吞吐 (V1 metrics): 140.73 tok/s
+
+--- 批量大小 bs=4 ---
+
+
+执行时间: 3.6361 s
+实际平均输入 tokens: 1.00（目标 1）
+生成总 tokens: 2048
+吞吐(生成tokens/秒): 563.24
+TTFT (V1 metrics): 0.0140 s
+解码吞吐 (V1 metrics): 141.25 tok/s
+
+--- 批量大小 bs=8 ---
+
+
+执行时间: 3.7267 s
+实际平均输入 tokens: 1.00（目标 1）
+生成总 tokens: 4096
+吞吐(生成tokens/秒): 1099.08
+TTFT (V1 metrics): 0.0149 s
+解码吞吐 (V1 metrics): 137.87 tok/s
+
+--- 批量大小 bs=16 ---
+
+
+执行时间: 3.8260 s
+实际平均输入 tokens: 1.00（目标 1）
+生成总 tokens: 8192
+吞吐(生成tokens/秒): 2141.13
+TTFT (V1 metrics): 0.0136 s
+解码吞吐 (V1 metrics): 134.33 tok/s
+
+--- 批量大小 bs=32 ---
+
+
+执行时间: 3.9972 s
+实际平均输入 tokens: 1.00（目标 1）
+生成总 tokens: 16384
+吞吐(生成tokens/秒): 4098.85
+TTFT (V1 metrics): 0.0164 s
+解码吞吐 (V1 metrics): 128.92 tok/s
+
+--- 批量大小 bs=64 ---
+
+
+执行时间: 4.2731 s
+实际平均输入 tokens: 1.00（目标 1）
+生成总 tokens: 32768
+吞吐(生成tokens/秒): 7668.51
+TTFT (V1 metrics): 0.0198 s
+解码吞吐 (V1 metrics): 120.83 tok/s
+
+--- 批量大小 bs=128 ---
+
+
+执行时间: 4.8421 s
+实际平均输入 tokens: 1.00（目标 1）
+生成总 tokens: 65536
+吞吐(生成tokens/秒): 13534.75
+TTFT (V1 metrics): 0.0316 s
+解码吞吐 (V1 metrics): 107.35 tok/s
+
+--- 批量大小 bs=256 ---
+
+
+执行时间: 7.4408 s
+实际平均输入 tokens: 1.00（目标 1）
+生成总 tokens: 131072
+吞吐(生成tokens/秒): 17615.41
+TTFT (V1 metrics): 0.0433 s
+解码吞吐 (V1 metrics): 69.70 tok/s
+
+--- 批量大小 bs=512 ---
+
+
+执行时间: 12.6794 s
+实际平均输入 tokens: 1.00（目标 1）
+生成总 tokens: 262144
+吞吐(生成tokens/秒): 20674.72
+TTFT (V1 metrics): 0.1809 s
+解码吞吐 (V1 metrics): 42.38 tok/s
+
+--- 批量大小 bs=1024 ---
+
+
+[rank0]:[W813 19:23:32.135663883 ProcessGroupNCCL.cpp:1479] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
+执行时间: 25.2865 s
+实际平均输入 tokens: 1.00（目标 1）
+生成总 tokens: 524288
+吞吐(生成tokens/秒): 20733.89
+TTFT (V1 metrics): 0.1191 s
+解码吞吐 (V1 metrics): 20.77 tok/s
+
+完成。提示：在 Nsight Systems 中可通过 NVTX 区间快速定位各场景/批量的调用。
+GPU 3: General Metrics for NVIDIA AD10x (any frequency)
+Generating '/tmp/nsys-report-bd68.qdstrm'
+
+
+[3/8] Executing 'nvtx_sum' stats report
+
+ Time (%)  Total Time (ns)  Instances      Avg (ns)          Med (ns)         Min (ns)        Max (ns)     StdDev (ns)   Style                   Range                 
+ --------  ---------------  ---------  ----------------  ----------------  --------------  --------------  -----------  -------  --------------------------------------
+     31.5   35,202,819,763          1  35,202,819,763.0  35,202,819,763.0  35,202,819,763  35,202,819,763          0.0  PushPop  :LLM_init                             
+     22.6   25,285,906,551          1  25,285,906,551.0  25,285,906,551.0  25,285,906,551  25,285,906,551          0.0  PushPop  :generate [prefill1_decode512] bs=1024
+     11.3   12,679,311,252          1  12,679,311,252.0  12,679,311,252.0  12,679,311,252  12,679,311,252          0.0  PushPop  :generate [prefill1_decode512] bs=512 
+      6.7    7,440,608,085          1   7,440,608,085.0   7,440,608,085.0   7,440,608,085   7,440,608,085          0.0  PushPop  :generate [prefill1_decode512] bs=256 
+      4.3    4,841,914,697          1   4,841,914,697.0   4,841,914,697.0   4,841,914,697   4,841,914,697          0.0  PushPop  :generate [prefill1_decode512] bs=128 
+      3.8    4,272,889,441          1   4,272,889,441.0   4,272,889,441.0   4,272,889,441   4,272,889,441          0.0  PushPop  :generate [prefill1_decode512] bs=64  
+      3.6    3,997,075,015          1   3,997,075,015.0   3,997,075,015.0   3,997,075,015   3,997,075,015          0.0  PushPop  :generate [prefill1_decode512] bs=32  
+      3.4    3,825,710,172          1   3,825,710,172.0   3,825,710,172.0   3,825,710,172   3,825,710,172          0.0  PushPop  :generate [prefill1_decode512] bs=16  
+      3.3    3,726,603,655          1   3,726,603,655.0   3,726,603,655.0   3,726,603,655   3,726,603,655          0.0  PushPop  :generate [prefill1_decode512] bs=8   
+      3.3    3,648,294,896          1   3,648,294,896.0   3,648,294,896.0   3,648,294,896   3,648,294,896          0.0  PushPop  :generate [prefill1_decode512] bs=2   
+      3.2    3,635,960,724          1   3,635,960,724.0   3,635,960,724.0   3,635,960,724   3,635,960,724          0.0  PushPop  :generate [prefill1_decode512] bs=4   
+      3.0    3,319,210,677          1   3,319,210,677.0   3,319,210,677.0   3,319,210,677   3,319,210,677          0.0  PushPop  :generate [prefill1_decode512] bs=1   
+      0.0           90,630          2          45,315.0          45,315.0          41,468          49,162      5,440.5  PushPop  CCCL:cub::DeviceSegmentedRadixSort    
+
+[4/8] Executing 'osrt_sum' stats report
+
+ Time (%)   Total Time (ns)   Num Calls     Avg (ns)          Med (ns)      Min (ns)      Max (ns)       StdDev (ns)              Name         
+ --------  -----------------  ---------  ---------------  ----------------  ---------  --------------  ----------------  ----------------------
+     29.7  1,284,023,267,442     49,709     25,830,800.6          28,820.0      1,000  96,534,138,823     969,651,617.8  pthread_cond_timedwait
+     24.2  1,045,773,118,476     73,135     14,299,215.4      10,062,843.0      1,010  81,708,518,315     463,992,514.8  epoll_wait            
+     23.9  1,031,963,983,489        549  1,879,715,816.9          15,827.0      1,644  96,535,575,076  13,212,126,179.1  pthread_cond_wait     
+      8.3    357,289,068,553         57  6,268,229,272.9  10,000,073,611.0     10,419  10,000,146,360   4,730,020,985.8  sem_timedwait         
+      8.2    355,562,427,280     39,343      9,037,501.6           1,512.0      1,000  12,219,368,064     127,255,975.8  poll                  
+      3.0    131,715,516,186     11,467     11,486,484.4       7,170,376.0     28,456     585,263,318      14,022,810.8  sem_wait              
+      2.6    112,179,081,690     41,286      2,717,121.6           2,213.0      1,000  94,667,308,194     468,888,772.5  read                  
+      0.0        793,423,700        330      2,404,314.2       1,354,549.5      1,900      18,406,079       2,639,127.9  pthread_rwlock_wrlock 
+      0.0        494,035,258    199,029          2,482.2           1,380.0      1,000      72,123,224         161,811.1  munmap                
+      0.0        298,099,876      8,608         34,630.6          10,089.5      1,002      29,694,619         390,579.6  ioctl                 
+      0.0        220,903,685        369        598,655.0           2,510.0      1,159      22,536,588       3,264,985.3  fopen                 
+      0.0        121,576,605         24      5,065,691.9       5,064,718.5  5,053,737       5,087,479           7,589.0  nanosleep             
+      0.0        110,429,690     30,645          3,603.5           2,536.0      1,000      19,587,041         111,890.3  open64                
+      0.0         88,325,670         79      1,118,046.5           3,103.0      1,011      81,521,546       9,166,201.0  waitpid               
+      0.0         76,471,088     18,154          4,212.4           3,660.0      1,000       1,659,563          15,417.3  mmap64                
+      0.0         74,586,704         96        776,944.8           3,874.0      1,020      19,635,272       3,727,062.1  open                  
+      0.0         72,521,171      8,897          8,151.2           4,707.0      1,022       2,826,985          34,660.7  recv                  
+      0.0         71,841,066      8,895          8,076.6           5,429.0      1,571          84,986           7,239.5  send                  
+      0.0         69,801,955     41,067          1,699.7           1,627.0      1,000          32,060             794.7  pthread_cond_signal   
+      0.0         67,085,379         39      1,720,137.9         470,979.0      3,544      10,373,042       3,340,509.2  pthread_join          
+      0.0         56,617,564         10      5,661,756.4          18,705.5      8,315      56,388,994      17,823,747.2  connect               
+      0.0         51,207,160     14,809          3,457.8           2,380.0      1,013         139,725           5,733.7  write                 
+      0.0         40,211,592      4,773          8,424.8           6,319.0      1,000         661,710          13,119.1  pthread_mutex_lock    
+      0.0         16,225,859     10,123          1,602.9           1,387.0      1,000          17,416             730.6  epoll_ctl             
+      0.0          9,852,733        147         67,025.4          68,737.0     55,805          95,256           5,155.2  sleep                 
+      0.0          7,858,705         22        357,213.9         474,706.5      8,796         678,261         278,233.7  pthread_rwlock_rdlock 
+      0.0          7,721,440        131         58,942.3          56,260.0     21,296         195,560          26,034.2  pthread_create        
+      0.0          7,224,126        929          7,776.2           3,096.0      1,000          86,864          11,142.9  fgets                 
+      0.0          1,723,609        344          5,010.5           4,755.0      1,827          40,649           2,516.1  fopen64               
+      0.0          1,708,972         62         27,564.1           2,983.5      1,002         230,421          59,773.0  futex                 
+      0.0          1,347,355      1,069          1,260.4           1,023.0      1,000          12,904             880.3  fclose                
+      0.0          1,149,466        196          5,864.6           3,579.5      1,105         168,420          13,582.2  mmap                  
+      0.0            878,967          1        878,967.0         878,967.0    878,967         878,967               0.0  fork                  
+      0.0            364,215         65          5,603.3           5,028.0      1,909          15,104           3,123.9  pipe2                 
+      0.0            247,833         41          6,044.7           4,941.0      1,709          17,457           4,172.6  socket                
+      0.0            188,362         19          9,913.8           3,097.0      1,045          62,742          16,639.4  bind                  
+      0.0            128,433         34          3,777.4           3,261.0      1,187          14,840           2,461.4  pthread_cond_broadcast
+      0.0             76,747          7         10,963.9           9,959.0      3,576          31,262           9,493.1  fread                 
+      0.0             65,399         41          1,595.1           1,200.0      1,012           5,988           1,063.6  fcntl                 
+      0.0             49,079          5          9,815.8           9,542.0      4,750          17,158           4,761.0  accept4               
+      0.0             42,725         25          1,709.0           1,806.0      1,011           2,296             397.2  sigaction             
+      0.0             40,441         20          2,022.1           2,166.5      1,063           3,618             818.8  dup2                  
+      0.0             39,878         15          2,658.5           2,065.0      1,267           7,040           1,459.6  stat                  
+      0.0             31,245         12          2,603.8           1,918.0      1,006           5,220           1,771.4  fflush                
+      0.0             27,179          5          5,435.8           5,277.0      1,662           9,374           3,035.3  fwrite                
+      0.0             21,540          4          5,385.0           5,545.5      4,572           5,877             575.6  lstat                 
+      0.0             17,255          4          4,313.8           4,516.5      2,856           5,366           1,051.1  flock                 
+      0.0             16,827          9          1,869.7           1,599.0      1,008           3,313             844.3  pread                 
+      0.0             15,569         10          1,556.9           1,444.0      1,184           2,260             325.7  listen                
+      0.0             13,074          3          4,358.0           4,294.0      4,285           4,495             118.7  fputs_unlocked        
+      0.0             12,439          5          2,487.8           2,713.0      1,831           3,023             566.6  mprotect              
+      0.0              7,489          4          1,872.3           1,856.5      1,636           2,140             206.8  flockfile             
+      0.0              6,919          1          6,919.0           6,919.0      6,919           6,919               0.0  kill                  
+      0.0              5,460          2          2,730.0           2,730.0      2,008           3,452           1,021.1  openat64              
+      0.0              5,297          3          1,765.7           1,842.0      1,157           2,298             574.3  fstat                 
+      0.0              3,627          1          3,627.0           3,627.0      3,627           3,627               0.0  fputs                 
+
+[5/8] Executing 'cuda_api_sum' stats report
+
+ Time (%)  Total Time (ns)  Num Calls   Avg (ns)     Med (ns)    Min (ns)   Max (ns)    StdDev (ns)                     Name                   
+ --------  ---------------  ---------  -----------  -----------  --------  -----------  -----------  ------------------------------------------
+     65.2   20,318,348,211     12,196  1,665,984.6      3,806.0     1,713  143,905,820  4,877,424.6  cudaStreamSynchronize                     
+     19.6    6,114,479,110    979,090      6,245.1      4,893.0       826   61,317,707    103,504.7  cudaLaunchKernel                          
+      5.9    1,844,232,020    151,960     12,136.3      9,991.0     7,437    6,397,280     55,443.5  cudaGraphLaunch_v10000                    
+      4.4    1,366,270,456     61,593     22,182.2      8,605.0     2,898   97,438,741    427,651.1  cudaMemcpyAsync                           
+      2.1      657,913,070    123,014      5,348.3      4,791.0       646   11,218,430     74,172.4  cuLaunchKernel                            
+      0.7      225,191,175      1,943    115,898.7     75,223.0    40,921    1,507,774    191,247.4  cudaGraphInstantiateWithFlags_v11040      
+      0.6      190,028,321      2,135     89,006.2     32,930.0     5,778  121,430,749  2,627,383.4  cudaDeviceSynchronize                     
+      0.4      131,733,072     24,728      5,327.3      5,346.0       183    7,263,692     48,804.8  cudaMemsetAsync                           
+      0.4      117,166,497    154,261        759.5        737.0       297        9,955        164.6  cudaStreamIsCapturing_v10000              
+      0.2       54,817,946        222    246,927.7    125,544.5    64,964    2,389,846    359,442.8  cudaFree                                  
+      0.1       41,574,143        348    119,465.9    111,957.5     6,496    1,314,648     70,124.0  cudaMalloc                                
+      0.1       25,470,407         10  2,547,040.7  2,568,202.0    60,182    4,674,895  1,473,736.1  cuLibraryLoadData                         
+      0.0       14,126,639     13,502      1,046.3        512.0       267    4,070,645     36,594.4  cuKernelGetFunction                       
+      0.0       11,511,739        169     68,116.8     73,800.0    26,538      398,968     40,288.3  cuModuleLoadData                          
+      0.0        9,477,345     18,895        501.6        477.0       305        7,151        120.4  cudaStreamGetCaptureInfo_v2_v11030        
+      0.0        8,547,159      1,943      4,398.9      4,349.0     3,274       12,306        644.1  cudaStreamBeginCapture_v10000             
+      0.0        7,583,507      1,943      3,903.0      3,886.0     2,371       10,115        530.0  cudaGraphDestroy_v10000                   
+      0.0        2,953,354        128     23,073.1      2,127.0     1,339      976,651    118,496.6  cudaStreamCreateWithPriority              
+      0.0        2,583,759      1,943      1,329.8      1,322.0       973        2,362        129.3  cudaStreamEndCapture_v10000               
+      0.0        1,910,887         26     73,495.7     12,773.5     3,625    1,207,162    232,915.8  cudaHostAlloc                             
+      0.0        1,625,828      1,943        836.8        771.0       625        3,016        254.6  cudaGraphGetNodes_v10000                  
+      0.0          943,862        310      3,044.7      2,639.0       879       11,991      1,944.4  cudaEventQuery                            
+      0.0          731,374        311      2,351.7      2,439.0       991        7,657      1,133.3  cudaEventRecord                           
+      0.0          219,541          8     27,442.6     26,305.5     8,804       64,233     18,995.2  cudaMemGetInfo                            
+      0.0          140,500        810        173.5        143.0        85        1,704        110.6  cuGetProcAddress_v2                       
+      0.0           21,914         21      1,043.5        438.0       339        4,729      1,202.9  cudaEventCreateWithFlags                  
+      0.0           16,258         16      1,016.1        849.5       502        2,663        551.3  cuLibraryGetKernel                        
+      0.0            8,991         14        642.2        586.0       346        1,420        261.0  cudaThreadExchangeStreamCaptureMode_v10010
+      0.0            4,849          3      1,616.3      1,664.0     1,386        1,799        210.6  cuInit                                    
+      0.0            3,460          4        865.0        749.0       110        1,852        882.7  cuModuleGetLoadingMode                    
+      0.0            3,416          1      3,416.0      3,416.0     3,416        3,416          0.0  cudaStreamWaitEvent                       
+      0.0            1,901          1      1,901.0      1,901.0     1,901        1,901          0.0  cudaEventDestroy                          
+      0.0            1,166          2        583.0        583.0       248          918        473.8  cudaGetDriverEntryPoint_v11030            
+
+[6/8] Executing 'cuda_gpu_kern_sum' stats report
+
+ Time (%)  Total Time (ns)  Instances   Avg (ns)     Med (ns)    Min (ns)   Max (ns)   StdDev (ns)                                                  Name                                                
+ --------  ---------------  ---------  -----------  -----------  ---------  ---------  -----------  ----------------------------------------------------------------------------------------------------
+     33.5    9,507,682,829     84,588    112,399.9     58,880.0      5,728    569,477    137,306.6  void flash::flash_fwd_splitkv_kernel<Flash_fwd_kernel_traits<(int)128, (int)64, (int)128, (int)4, (…
+     27.2    7,733,720,607     29,164    265,180.4    333,123.0     33,344    763,622    112,810.5  ampere_bf16_s1688gemm_bf16_64x128_sliced1x2_ldg8_f2f_tn                                             
+      7.4    2,089,759,636      1,164  1,795,326.1  1,390,859.0     40,065  4,518,698  1,024,759.1  ampere_bf16_s1688gemm_bf16_128x128_ldg8_f2f_stages_32x1_tn                                          
+      3.3      942,739,585      5,754    163,840.7     13,376.0      1,951  1,008,234    293,922.8  void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::TensorIterator…
+      3.3      942,069,183     76,664     12,288.3      8,032.0      6,240     73,248      8,482.6  void flash::flash_fwd_splitkv_kernel<Flash_fwd_kernel_traits<(int)128, (int)64, (int)128, (int)4, (…
+      3.3      926,915,977      5,958    155,575.0     73,120.5      7,649    549,540    194,687.6  void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_32x6_tn_align8>(T1::Param…
+      2.8      781,701,830      1,991    392,617.7    496,547.0     10,528    506,724    194,713.8  void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_16x16_128x2_tn_align8>(T1::Par…
+      2.5      718,309,736      5,756    124,793.2      9,920.0      5,151    716,420    213,447.3  void at::native::reduce_kernel<(int)512, (int)1, at::native::ReduceOp<float, at::native::ArgMaxOps<…
+      2.4      679,170,209    292,768      2,319.8      1,889.0      1,631      6,304        962.1  void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_nocast<at::n…
+      2.1      605,586,081     14,252     42,491.3     42,529.0     26,240     62,817      1,642.0  ampere_bf16_s1688gemm_bf16_128x64_sliced1x2_ldg8_relu_f2f_tn                                        
+      1.8      516,743,427     13,776     37,510.4     37,472.0     36,608     42,560        332.8  ampere_bf16_s1688gemm_bf16_64x64_sliced1x4_ldg8_f2f_tn                                              
+      1.3      366,625,152     16,268     22,536.6     23,936.0      1,055    462,659     19,521.4  triton_poi_fused_mul_silu_1                                                                         
+      1.2      345,171,402        112  3,081,887.5  3,078,316.5  3,036,028  3,128,477     29,283.9  void flash::flash_fwd_splitkv_kernel<Flash_fwd_kernel_traits<(int)128, (int)64, (int)128, (int)4, (…
+      0.9      260,532,523        513    507,860.7    507,843.0    506,403    509,475        427.2  void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_16x16_128x1_tn_align8>(T1::Par…
+      0.9      255,622,286        604    423,215.7    487,970.0      7,008    488,866    160,519.3  std::enable_if<!T7, void>::type internal::gemvx::kernel<int, int, __nv_bfloat16, __nv_bfloat16, __n…
+      0.9      242,493,430    161,056      1,505.6      1,280.0      1,023      3,488        458.1  void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0…
+      0.7      206,425,050    146,384      1,410.2      1,344.0      1,183      2,208        221.6  void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_nocast<at::n…
+      0.7      203,499,133        184  1,105,973.5    579,541.0    369,795  2,808,909    981,858.0  ampere_bf16_s16816gemm_bf16_128x64_ldg8_f2f_tn                                                      
+      0.7      187,786,954      1,120    167,666.9    158,929.0     40,416  1,415,463    206,084.0  ampere_bf16_s1688gemm_bf16_128x64_sliced1x2_ldg8_f2f_tn                                             
+      0.6      180,432,004     16,268     11,091.2     11,936.0      1,505    111,617      4,905.8  triton_red_fused__to_copy_add_mean_mul_pow_rsqrt_2                                                  
+      0.5      135,678,243     43,792      3,098.2      3,104.0      2,943      3,616         73.3  void flash::flash_fwd_splitkv_combine_kernel<Flash_fwd_kernel_traits<(int)128, (int)64, (int)128, (…
+      0.3       98,055,325     32,872      2,982.9      2,945.0      2,847      3,233         98.3  void flash::flash_fwd_splitkv_combine_kernel<Flash_fwd_kernel_traits<(int)128, (int)64, (int)128, (…
+      0.3       83,656,872     16,268      5,142.4      5,472.0      1,536     79,136      3,233.8  triton_red_fused__to_copy_add_mean_mul_pow_rsqrt_0                                                  
+      0.2       52,052,585     15,687      3,318.2      3,457.0      1,344     22,048        905.0  triton_poi_fused_cat_3                                                                              
+      0.1       39,252,777          8  4,906,597.1  4,863,769.0  4,802,745  5,085,370    117,043.6  void at_cuda_detail::cub::DeviceSegmentedRadixSortKernel<at_cuda_detail::cub::DeviceRadixSortPolicy…
+      0.1       35,101,287     15,687      2,237.6      2,336.0        863     16,672        687.2  triton_poi_fused_view_5                                                                             
+      0.1       26,987,031     17,256      1,563.9      1,408.0      1,023      2,784        465.1  void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::TensorIterator…
+      0.1       22,634,043        784     28,870.0     12,543.5     11,616     62,720     20,595.1  ampere_bf16_s16816gemm_bf16_64x64_ldg8_f2f_stages_64x5_tn                                           
+      0.1       22,584,985     15,687      1,439.7      1,440.0      1,215      6,720        194.4  triton_poi_fused_cat_4                                                                              
+      0.1       20,606,081      5,888      3,499.7      3,136.0      2,687      7,488        970.6  void at::native::index_elementwise_kernel<(int)128, (int)4, void at::native::gpu_index_kernel<void …
+      0.1       20,451,721          4  5,112,930.3  5,112,153.5  4,937,754  5,289,660    201,308.2  void at_cuda_detail::cub::DeviceSegmentedRadixSortKernel<at_cuda_detail::cub::DeviceRadixSortPolicy…
+      0.1       15,066,944      5,752      2,619.4      2,368.0      1,952      3,969        554.4  void at::native::index_elementwise_kernel<(int)128, (int)4, void at::native::gpu_index_kernel<void …
+      0.1       14,293,381         28    510,477.9    512,002.5    468,706    513,474      8,208.5  void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<signed char>, std::a…
+      0.0        9,733,874          4  2,433,468.5  2,435,244.5  2,367,692  2,495,693     60,628.3  void at::native::<unnamed>::cunn_SoftMaxForward<(int)4, float, float, float, at::native::<unnamed>:…
+      0.0        9,136,049         28    326,287.5    326,210.0    324,514    329,538        982.3  ampere_bf16_s1688gemm_bf16_128x128_ldg8_relu_f2f_stages_32x1_tn                                     
+      0.0        8,425,678        224     37,614.6     37,568.5     36,608     38,880        409.5  void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x128_32x6_tn_align8>(T1::Para…
+      0.0        7,777,256          2  3,888,628.0  3,888,628.0  3,705,715  4,071,541    258,678.0  void at::native::_scatter_gather_elementwise_kernel<(int)128, (int)8, void at::native::_cuda_scatte…
+      0.0        7,758,429      8,970        864.9        864.0        767      1,280         77.7  void at::native::vectorized_elementwise_kernel<(int)2, at::native::FillFunctor<long>, std::array<ch…
+      0.0        7,519,688      5,754      1,306.9      1,152.0      1,023      2,048        279.6  void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::TensorIterator…
+      0.0        7,231,396        336     21,522.0     21,440.0     21,056     22,592        362.7  ampere_bf16_s16816gemm_bf16_128x64_ldg8_relu_f2f_stages_64x3_tn                                     
+      0.0        6,130,365        476     12,878.9     12,800.5     11,744     14,432        645.4  ampere_bf16_s16816gemm_bf16_64x64_ldg8_relu_f2f_stages_64x5_tn                                      
+      0.0        5,896,061          4  1,474,015.3  1,473,463.0  1,473,191  1,475,944      1,292.3  void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::masked_fill_kernel(at…
+      0.0        5,380,367      5,752        935.4        896.0        863      1,344         76.3  void at::native::unrolled_elementwise_kernel<at::native::CUDAFunctorOnSelf_add<int>, std::array<cha…
+      0.0        4,603,692      5,292        869.9        864.0        767      1,185         33.2  void at::native::unrolled_elementwise_kernel<at::native::FillFunctor<int>, std::array<char *, (unsi…
+      0.0        3,996,949          2  1,998,474.5  1,998,474.5  1,996,490  2,000,459      2,806.5  void at::native::vectorized_elementwise_kernel<(int)4, at::native::BinaryFunctor<float, float, floa…
+      0.0        3,842,156      4,143        927.4        896.0        800      1,856         93.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<int>, std::array<cha…
+      0.0        3,593,747         56     64,174.1     64,144.5     63,105     65,728        478.5  void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_64x1_tn_align8>(T1::Para…
+      0.0        3,433,971          4    858,492.8    858,901.0    855,877    860,292      2,168.9  void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_nocast<at::n…
+      0.0        3,191,215          2  1,595,607.5  1,595,607.5  1,560,359  1,630,856     49,848.9  void at::native::tensor_kernel_scan_innermost_dim<float, std::plus<float>>(T1 *, const T1 *, unsign…
+      0.0        2,871,002      1,512      1,898.8      1,760.0      1,312      2,912        445.3  void cublasLt::splitKreduce_kernel<(int)32, (int)16, int, __nv_bfloat16, __nv_bfloat16, float, (boo…
+      0.0        2,643,071        581      4,549.2      4,671.0      1,984     36,256      1,427.2  triton_red_fused__to_copy_add_embedding_mean_mul_pow_rsqrt_0                                        
+      0.0        2,581,742          2  1,290,871.0  1,290,871.0  1,290,663  1,291,079        294.2  at::native::<unnamed>::fill_reverse_indices_kernel(long *, int, at::cuda::detail::IntDivider<unsign…
+      0.0        2,581,389          2  1,290,694.5  1,290,694.5  1,290,406  1,290,983        408.0  void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_nocast<at::n…
+      0.0        2,421,998        112     21,625.0     21,552.0      9,408     34,465     12,020.4  void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align8>(T1::Par…
+      0.0        1,835,794        581      3,159.7      3,200.0      1,632     39,200      1,543.9  triton_poi_fused_cat_1                                                                              
+      0.0        1,365,128          2    682,564.0    682,564.0    677,764    687,364      6,788.2  void at::native::<unnamed>::distribution_elementwise_grid_stride_kernel<float, (int)4, void at::nat…
+      0.0        1,304,994        581      2,246.1      2,368.0        863     14,272        629.4  triton_poi_fused_view_3                                                                             
+      0.0        1,202,982         56     21,481.8     21,456.0     21,152     21,888        275.5  ampere_bf16_s16816gemm_bf16_128x64_ldg8_f2f_stages_32x6_tn                                          
+      0.0        1,027,854      1,153        891.5        896.0        800      1,216         36.7  void at::native::vectorized_elementwise_kernel<(int)2, at::native::FillFunctor<int>, std::array<cha…
+      0.0          956,098         28     34,146.4     34,736.5     17,920     35,200      3,188.0  std::enable_if<!T7, void>::type internal::gemvx::kernel<int, int, __nv_bfloat16, float, float, floa…
+      0.0          847,695        581      1,459.0      1,440.0      1,216      9,408        336.4  triton_poi_fused_cat_2                                                                              
+      0.0          611,794        673        909.1        896.0        864      1,025         28.1  void at::native::unrolled_elementwise_kernel<at::native::FillFunctor<long>, std::array<char *, (uns…
+      0.0          417,574        308      1,355.8      1,344.0      1,311      1,504         21.9  void vllm::merge_attn_states_kernel<__nv_bfloat16, (unsigned int)128>(T1 *, float *, const T1 *, co…
+      0.0          295,335        168      1,757.9      1,760.0      1,535      2,080        119.1  void cublasLt::splitKreduce_kernel<(int)32, (int)16, int, __nv_bfloat16, __nv_bfloat16, float, (boo…
+      0.0          155,841          1    155,841.0    155,841.0    155,841    155,841          0.0  void at::native::<unnamed>::CatArrayBatchedCopy_aligned16_contig<at::native::<unnamed>::OpaqueType<…
+      0.0           78,880          1     78,880.0     78,880.0     78,880     78,880          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::bfloat16_copy_kernel_cuda(at::Te…
+      0.0           63,740         58      1,099.0        896.0        864     11,360      1,372.6  void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<c10::BFloat16>, std:…
+      0.0           43,936          1     43,936.0     43,936.0     43,936     43,936          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::sin_kernel_cuda(at::TensorIterat…
+      0.0           36,570         28      1,306.1      1,312.0      1,280      1,376         19.5  void cublasLt::splitKreduce_kernel<(int)32, (int)16, int, float, __nv_bfloat16, float, (bool)0, __n…
+      0.0           26,816          1     26,816.0     26,816.0     26,816     26,816          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::cos_kernel_cuda(at::TensorIterat…
+      0.0           19,520          1     19,520.0     19,520.0     19,520     19,520          0.0  void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_nocast<at::n…
+      0.0           11,936         11      1,085.1        864.0        864      1,568        286.4  void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<float>, std::array<c…
+      0.0           10,752          2      5,376.0      5,376.0      5,120      5,632        362.0  void at::native::_scatter_gather_elementwise_kernel<(int)128, (int)8, void at::native::_cuda_scatte…
+      0.0            9,152          2      4,576.0      4,576.0      4,480      4,672        135.8  void at::native::<unnamed>::distribution_elementwise_grid_stride_kernel<float, (int)4, void at::nat…
+      0.0            3,616          2      1,808.0      1,808.0      1,600      2,016        294.2  void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_nocast<at::n…
+      0.0            3,424          2      1,712.0      1,712.0      1,664      1,760         67.9  void at::native::vectorized_elementwise_kernel<(int)2, at::native::CUDAFunctorOnOther_add<long>, st…
+      0.0            3,136          2      1,568.0      1,568.0      1,504      1,632         90.5  void at::native::vectorized_elementwise_kernel<(int)2, at::native::<unnamed>::where_kernel_impl(at:…
+      0.0            3,104          2      1,552.0      1,552.0      1,344      1,760        294.2  void at::native::vectorized_elementwise_kernel<(int)4, void at::native::compare_scalar_kernel<float…
+      0.0            2,975          2      1,487.5      1,487.5        992      1,983        700.7  void <unnamed>::elementwise_kernel_with_index<int, at::native::arange_cuda_out(const c10::Scalar &,…
+      0.0            2,912          2      1,456.0      1,456.0      1,344      1,568        158.4  void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnOther_add<float>, s…
+      0.0            2,336          1      2,336.0      2,336.0      2,336      2,336          0.0  void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl<at::native::…
+      0.0            1,184          1      1,184.0      1,184.0      1,184      1,184          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::reciprocal_kernel_cuda(at::Tenso…
+      0.0            1,024          1      1,024.0      1,024.0      1,024      1,024          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, float, floa…
+      0.0            1,024          1      1,024.0      1,024.0      1,024      1,024          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::BUnaryFunctor<float, float, floa…
+      0.0              896          1        896.0        896.0        896        896          0.0  void at::native::vectorized_elementwise_kernel<(int)2, at::native::FillFunctor<double>, std::array<…
+
+[7/8] Executing 'cuda_gpu_mem_time_sum' stats report
+
+ Time (%)  Total Time (ns)  Count   Avg (ns)  Med (ns)  Min (ns)   Max (ns)   StdDev (ns)            Operation           
+ --------  ---------------  ------  --------  --------  --------  ----------  -----------  ------------------------------
+     93.2      540,571,743  41,277  13,096.2     352.0       287  97,068,545    513,408.1  [CUDA memcpy Host-to-Device]  
+      3.2       18,710,334  14,564   1,284.7     896.0       864   1,362,855     22,521.7  [CUDA memcpy Device-to-Device]
+      2.5       14,536,294  21,760     668.0     768.0       287       7,744        311.5  [CUDA memset]                 
+      1.1        6,503,130   5,752   1,130.6   1,120.0       863       1,760         95.6  [CUDA memcpy Device-to-Host]  
+
+[8/8] Executing 'cuda_gpu_mem_size_sum' stats report
+
+ Total (MB)  Count   Avg (MB)  Med (MB)  Min (MB)  Max (MB)  StdDev (MB)            Operation           
+ ----------  ------  --------  --------  --------  --------  -----------  ------------------------------
+  4,190.741  41,277     0.102     0.000     0.000   466.747        2.619  [CUDA memcpy Host-to-Device]  
+  2,534.048  14,564     0.174     0.003     0.003   622.330       10.312  [CUDA memcpy Device-to-Device]
+     14.589  21,760     0.001     0.001     0.000     0.006        0.000  [CUDA memset]                 
+      4.192   5,752     0.001     0.000     0.000     0.004        0.001  [CUDA memcpy Device-to-Host]  
+
+Generated:
+    /data/cy/kv_cache_vs_util/sim_traverse_bs/traverse_bs_util_sim_decoding.nsys-rep
+    /data/cy/kv_cache_vs_util/sim_traverse_bs/traverse_bs_util_sim_decoding.sqlite

sim_traverse_bs/traverse_bs_util_sim_decoding.nsys-rep

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9cb1e2cf07fcb0e9c21d44c66cdf7762bc995625a1a55e835b8cffc5a6104b8
+size 109416019

sim_traverse_bs/traverse_bs_util_sim_decoding.py

@@ -0,0 +1,274 @@
+import os
+import time
+import statistics
+from typing import List, Tuple, Dict
+
+import torch
+import torch.cuda.nvtx as nvtx
+
+from vllm import LLM, SamplingParams
+from transformers import AutoTokenizer
+
+# ========= 强制使用 vLLM V1 =========
+os.environ.setdefault("VLLM_USE_V1", "1")
+os.environ.setdefault("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
+
+# 可选：打开 V1 metrics 统计
+os.environ.setdefault("VLLM_LOGGING_LEVEL", "INFO")
+
+# ========= 试图导入 V1 metrics 类型（兼容不同版本）=========
+try:
+    from vllm.v1.metrics.reader import Counter, Gauge, Histogram, Vector  # type: ignore
+except Exception:
+    Counter = Gauge = Histogram = Vector = type("X", (), {})  # dummy
+
+# ========= 配置 =========
+MODEL_NAME = "Qwen/Qwen2-1.5B"
+DTYPE = "bfloat16"
+TP = 1
+GPU_MEM_UTIL = 0.90
+TRUST_REMOTE_CODE = True
+
+# 场景：prefill=输入tokens，decode=输出tokens
+SCENARIOS = [
+    # {"name": "prefill640_decode1", "prompt_tokens": 640, "max_new_tokens": 1},
+    {"name": "prefill1_decode512", "prompt_tokens": 1,   "max_new_tokens": 1024},
+    # {"name": "prefill640_decode512", "prompt_tokens": 640, "max_new_tokens": 512},
+]
+
+BATCH_SIZES = [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024]
+
+SEED = 1234
+TEMPERATURE = 0.0
+TOP_P = 1.0
+WARMUP_PER_BS = 1  # 每个批次做一次预热
+
+# ========= 构造“精确 token 数量”的 prompt =========
+def build_exact_token_prompt(tokenizer, target_len: int) -> str:
+    if target_len <= 1:
+        # 最小化 prompt：用一个简单 token（避免空串导致0 token）
+        ids = tokenizer("A", add_special_tokens=False)["input_ids"]
+        if len(ids) >= 1:
+            return tokenizer.decode(ids[:1], skip_special_tokens=True, clean_up_tokenization_spaces=False)
+
+    base_text = (
+        "You are a helpful assistant. "
+        "Please analyze the following input and respond succinctly. "
+    )
+    chunk = " ".join(["data"] * 100) + ". "
+    text = base_text + chunk * 200  # 足够长的文本
+
+    lo, hi = 0, len(text)
+    target_ids = None
+    while lo <= hi:
+        mid = (lo + hi) // 2
+        ids = tokenizer(text[:mid], add_special_tokens=False)["input_ids"]
+        if len(ids) == target_len:
+            target_ids = ids
+            break
+        if len(ids) < target_len:
+            lo = mid + 1
+        else:
+            hi = mid - 1
+
+    if target_ids is None:
+        ids = tokenizer(text[:lo], add_special_tokens=False)["input_ids"]
+        if len(ids) > target_len:
+            target_ids = ids[:target_len]
+        else:
+            filler = " data"
+            while len(ids) < target_len:
+                ids = tokenizer(tokenizer.decode(ids) + filler, add_special_tokens=False)["input_ids"]
+            target_ids = ids[:target_len]
+
+    prompt = tokenizer.decode(target_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)
+    # 断言精确长度
+    assert len(tokenizer(prompt, add_special_tokens=False)["input_ids"]) == target_len
+    return prompt
+
+# ========= V1 metrics 抽取工具 =========
+TTFT_METRIC_NAME = "vllm:time_to_first_token_seconds"
+TPOT_METRIC_NAME = "vllm:time_per_output_token_seconds"  # per-output-token latency
+
+def _iter_children_of_vector(vec_obj):
+    for attr in ("children", "metrics", "series", "values", "samples", "items"):
+        if hasattr(vec_obj, attr):
+            val = getattr(vec_obj, attr)
+            if isinstance(val, dict):
+                for v in val.values():
+                    yield v
+            else:
+                try:
+                    for v in val:
+                        yield v
+                except TypeError:
+                    pass
+
+def _collect_hist_sum_count(metrics, metric_name: str):
+    total_sum = 0.0
+    total_count = 0.0
+    for m in metrics:
+        mname = getattr(m, "name", None)
+        if mname != metric_name:
+            continue
+        # 直接 Histogram
+        if isinstance(m, Histogram) or m.__class__.__name__ == "Histogram":
+            total_sum += float(getattr(m, "sum", 0.0))
+            total_count += float(getattr(m, "count", 0.0))
+            continue
+        # Vector[Histogram]
+        if isinstance(m, Vector) or m.__class__.__name__ == "Vector":
+            for child in _iter_children_of_vector(m):
+                if isinstance(child, Histogram) or child.__class__.__name__ == "Histogram":
+                    total_sum += float(getattr(child, "sum", 0.0))
+                    total_count += float(getattr(child, "count", 0.0))
+    return total_sum, total_count
+
+def _metrics_snapshot(llm) -> Dict[str, float]:
+    try:
+        mets = llm.get_metrics()  # V1: 返回 Metric 列表（包含 Histogram/Vector 等）
+    except Exception:
+        return {"ttft_sum": 0.0, "ttft_cnt": 0.0, "tpot_sum": 0.0, "tpot_cnt": 0.0}
+    ttft_sum, ttft_cnt = _collect_hist_sum_count(mets, TTFT_METRIC_NAME)
+    tpot_sum, tpot_cnt = _collect_hist_sum_count(mets, TPOT_METRIC_NAME)
+    return {"ttft_sum": ttft_sum, "ttft_cnt": ttft_cnt, "tpot_sum": tpot_sum, "tpot_cnt": tpot_cnt}
+
+def _metrics_delta(before: dict, after: dict):
+    return {
+        "ttft_sum": after["ttft_sum"] - before["ttft_sum"],
+        "ttft_cnt": after["ttft_cnt"] - before["ttft_cnt"],
+        "tpot_sum": after["tpot_sum"] - before["tpot_sum"],
+        "tpot_cnt": after["tpot_cnt"] - before["tpot_cnt"],
+    }
+
+# ========= 带 NVTX 的 generate 包装 =========
+def decorated_generate(llm: LLM, prompts: List[str], params: SamplingParams):
+    return llm.generate(prompts, params)
+
+# ========= 统计格式化 =========
+def fmt_stats(x: List[float]) -> Tuple[float, float, float]:
+    xs = [v for v in x if (v == v)]  # 过滤 NaN
+    if not xs:
+        return (float("nan"), float("nan"), float("nan"))
+    return (statistics.mean(xs), statistics.median(xs), statistics.quantiles(xs, n=10)[-1])  # p90
+
+def main():
+    print("--- vLLM V1 基准测试（含 NVTX 标记）---")
+    print(f"模型: {MODEL_NAME}")
+    print(f"批量大小: {BATCH_SIZES}")
+    print(f"场景: {[s['name'] for s in SCENARIOS]}")
+    print("-" * 60)
+
+    if not torch.cuda.is_available():
+        print("错误：需要 CUDA GPU。")
+        return
+
+    print("加载分词器/模型中...")
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, use_fast=True, trust_remote_code=TRUST_REMOTE_CODE)
+
+    # 用 NVTX 标记模型加载阶段
+    nvtx.range_push("LLM_init")
+    llm = LLM(
+        model=MODEL_NAME,
+        tensor_parallel_size=TP,
+        dtype=DTYPE,
+        trust_remote_code=TRUST_REMOTE_CODE,
+        gpu_memory_utilization=GPU_MEM_UTIL,
+        max_num_seqs=1024,         # 足够覆盖本次扫描
+        max_model_len=4096,
+        disable_log_stats=False,  # 开启 V1 metrics 收集
+    )
+    nvtx.range_pop()
+    print("模型加载完成。")
+
+    for sc in SCENARIOS:
+        name = sc["name"]
+        prompt_tokens = sc["prompt_tokens"]
+        max_new_tokens = sc["max_new_tokens"]
+
+        print(f"\n===== 场景：{name} | prefill={prompt_tokens}, decode={max_new_tokens} =====")
+
+        # 准备精确长度 prompt
+        prompt_text = build_exact_token_prompt(tokenizer, prompt_tokens)
+
+        # 采样参数（贪心）
+        # sampling_params = SamplingParams(
+        #     max_tokens=max_new_tokens,
+        #     temperature=TEMPERATURE,
+        #     top_p=TOP_P,
+        #     seed=SEED,
+        #     n=1,
+        # )
+
+        sampling_params = SamplingParams(
+            max_tokens=max_new_tokens,      # 比如 512
+            # 关键点：
+            ignore_eos=True,          # 忽略 EOS，继续生成
+            stop=None,                # 不设 stop 字符串
+            stop_token_ids=[],        # 不设 stop token
+            # 选配：如果你的 vLLM 版本支持
+            min_tokens=max_new_tokens,      # 至少生成 N 个 token（≤ max_tokens）
+            temperature=0.0,
+            top_p=1.0,
+        )
+
+        # 记录每个 bs 的结果（便于后续统计或外部解析）
+        for bs in BATCH_SIZES:
+            print(f"\n--- 批量大小 bs={bs} ---")
+
+            prompts = [prompt_text] * bs
+
+            # 预热
+            # print("预热中...")
+            # nvtx.range_push(f"WARMUP [{name}] bs={bs}")
+            # _ = decorated_generate(llm, [prompts[0]], sampling_params)
+            # torch.cuda.synchronize()
+            # nvtx.range_pop()
+
+            # 正式计时与 V1 metrics
+            # nvtx.range_push(f"RUN [{name}] bs={bs}")
+            torch.cuda.synchronize()
+            snap_before = _metrics_snapshot(llm)
+            t0 = time.perf_counter()
+
+            nvtx.range_push(f"generate [{name}] bs={bs}")
+            outputs = decorated_generate(llm, prompts, sampling_params)
+            nvtx.range_pop()  # generate
+
+            torch.cuda.synchronize()
+            t1 = time.perf_counter()
+            snap_after = _metrics_snapshot(llm)
+            # nvtx.range_pop()  # RUN
+
+            duration = t1 - t0
+
+            # 统计 token 与吞吐
+            total_output_tokens = sum(len(o.outputs[0].token_ids) for o in outputs)
+            avg_prompt_tokens = sum(len(o.prompt_token_ids) for o in outputs) / bs
+            throughput = total_output_tokens / duration if duration > 0 else float("inf")
+
+            # 解析 V1 TTFT / 解码吞吐
+            delta = _metrics_delta(snap_before, snap_after)
+            if delta["ttft_cnt"] > 0:
+                ttft = delta["ttft_sum"] / delta["ttft_cnt"]
+            else:
+                ttft = float("nan")
+
+            if delta["tpot_cnt"] > 0:
+                avg_tpot = delta["tpot_sum"] / delta["tpot_cnt"]  # seconds/token
+                decode_tps = 1.0 / avg_tpot
+            else:
+                decode_tps = float("nan")
+
+            print(f"执行时间: {duration:.4f} s")
+            print(f"实际平均输入 tokens: {avg_prompt_tokens:.2f}（目标 {prompt_tokens}）")
+            print(f"生成总 tokens: {total_output_tokens}")
+            print(f"吞吐(生成tokens/秒): {throughput:.2f}")
+            print(f"TTFT (V1 metrics): {ttft:.4f} s")
+            print(f"解码吞吐 (V1 metrics): {decode_tps:.2f} tok/s")
+
+    print("\n完成。提示：在 Nsight Systems 中可通过 NVTX 区间快速定位各场景/批量的调用。")
+
+if __name__ == "__main__":
+    print(f"CUDA_VISIBLE_DEVICES = {os.getenv('CUDA_VISIBLE_DEVICES')}")
+    main()

sim_traverse_bs/traverse_bs_util_sim_decoding_1024.log

@@ -0,0 +1,388 @@
+WARNING: CPU IP/backtrace sampling not supported, disabling.
+Try the 'nsys status --environment' command to learn more.
+
+WARNING: CPU context switch tracing not supported, disabling.
+Try the 'nsys status --environment' command to learn more.
+
+INFO 08-13 21:32:20 [__init__.py:235] Automatically detected platform cuda.
+CUDA_VISIBLE_DEVICES = 3
+--- vLLM V1 基准测试（含 NVTX 标记）---
+模型: Qwen/Qwen2-1.5B
+批量大小: [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024]
+场景: ['prefill1_decode512']
+------------------------------------------------------------
+加载分词器/模型中...
+INFO 08-13 21:32:29 [config.py:1604] Using max model len 4096
+INFO 08-13 21:32:30 [config.py:2434] Chunked prefill is enabled with max_num_batched_tokens=8192.
+INFO 08-13 21:32:35 [__init__.py:235] Automatically detected platform cuda.
+INFO 08-13 21:32:36 [core.py:572] Waiting for init message from front-end.
+INFO 08-13 21:32:36 [core.py:71] Initializing a V1 LLM engine (v0.10.0) with config: model='Qwen/Qwen2-1.5B', speculative_config=None, tokenizer='Qwen/Qwen2-1.5B', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config={}, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=4096, download_dir=None, load_format=LoadFormat.AUTO, tensor_parallel_size=1, pipeline_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=False, kv_cache_dtype=auto,  device_config=cuda, decoding_config=DecodingConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_backend=''), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None), seed=0, served_model_name=Qwen/Qwen2-1.5B, num_scheduler_steps=1, multi_step_stream_outputs=True, enable_prefix_caching=True, chunked_prefill_enabled=True, use_async_output_proc=True, pooler_config=None, compilation_config={"level":3,"debug_dump_path":"","cache_dir":"","backend":"","custom_ops":[],"splitting_ops":["vllm.unified_attention","vllm.unified_attention_with_output","vllm.mamba_mixer2"],"use_inductor":true,"compile_sizes":[],"inductor_compile_config":{"enable_auto_functionalized_v2":false},"inductor_passes":{},"use_cudagraph":true,"cudagraph_num_of_warmups":1,"cudagraph_capture_sizes":[512,504,496,488,480,472,464,456,448,440,432,424,416,408,400,392,384,376,368,360,352,344,336,328,320,312,304,296,288,280,272,264,256,248,240,232,224,216,208,200,192,184,176,168,160,152,144,136,128,120,112,104,96,88,80,72,64,56,48,40,32,24,16,8,4,2,1],"cudagraph_copy_inputs":false,"full_cuda_graph":false,"max_capture_size":512,"local_cache_dir":null}
+INFO 08-13 21:32:38 [parallel_state.py:1102] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, TP rank 0, EP rank 0
+WARNING 08-13 21:32:38 [topk_topp_sampler.py:59] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
+INFO 08-13 21:32:38 [gpu_model_runner.py:1843] Starting to load model Qwen/Qwen2-1.5B...
+INFO 08-13 21:32:38 [gpu_model_runner.py:1875] Loading model from scratch...
+INFO 08-13 21:32:38 [cuda.py:290] Using Flash Attention backend on V1 engine.
+INFO 08-13 21:32:39 [weight_utils.py:296] Using model weights format ['*.safetensors']
+INFO 08-13 21:32:39 [weight_utils.py:349] No model.safetensors.index.json found in remote.
+
+
+
+
+INFO 08-13 21:32:40 [default_loader.py:262] Loading weights took 0.63 seconds
+INFO 08-13 21:32:40 [gpu_model_runner.py:1892] Model loading took 2.9105 GiB and 1.831000 seconds
+INFO 08-13 21:32:46 [backends.py:530] Using cache directory: /home/cy/.cache/vllm/torch_compile_cache/40b61c71e9/rank_0_0/backbone for vLLM's torch.compile
+INFO 08-13 21:32:46 [backends.py:541] Dynamo bytecode transform time: 5.56 s
+INFO 08-13 21:32:51 [backends.py:161] Directly load the compiled graph(s) for dynamic shape from the cache, took 4.562 s
+INFO 08-13 21:32:52 [monitor.py:34] torch.compile takes 5.56 s in total
+INFO 08-13 21:32:53 [gpu_worker.py:255] Available KV cache memory: 12.81 GiB
+INFO 08-13 21:32:53 [kv_cache_utils.py:833] GPU KV cache size: 479,536 tokens
+INFO 08-13 21:32:53 [kv_cache_utils.py:837] Maximum concurrency for 4,096 tokens per request: 117.07x
+
+INFO 08-13 21:32:55 [gpu_model_runner.py:2485] Graph capturing finished in 2 secs, took 0.49 GiB
+INFO 08-13 21:32:55 [core.py:193] init engine (profile, create kv cache, warmup model) took 14.97 seconds
+模型加载完成。
+
+===== 场景：prefill1_decode512 | prefill=1, decode=1024 =====
+
+--- 批量大小 bs=1 ---
+
+
+执行时间: 6.4234 s
+实际平均输入 tokens: 1.00（目标 1）
+生成总 tokens: 1024
+吞吐(生成tokens/秒): 159.42
+TTFT (V1 metrics): 0.0138 s
+解码吞吐 (V1 metrics): 159.67 tok/s
+
+--- 批量大小 bs=2 ---
+
+
+执行时间: 7.2050 s
+实际平均输入 tokens: 1.00（目标 1）
+生成总 tokens: 2048
+吞吐(生成tokens/秒): 284.25
+TTFT (V1 metrics): 0.0127 s
+解码吞吐 (V1 metrics): 142.32 tok/s
+
+--- 批量大小 bs=4 ---
+
+
+执行时间: 7.2755 s
+实际平均输入 tokens: 1.00（目标 1）
+生成总 tokens: 4096
+吞吐(生成tokens/秒): 562.99
+TTFT (V1 metrics): 0.0141 s
+解码吞吐 (V1 metrics): 140.94 tok/s
+
+--- 批量大小 bs=8 ---
+
+
+执行时间: 7.5025 s
+实际平均输入 tokens: 1.00（目标 1）
+生成总 tokens: 8192
+吞吐(生成tokens/秒): 1091.90
+TTFT (V1 metrics): 0.0187 s
+解码吞吐 (V1 metrics): 136.74 tok/s
+
+--- 批量大小 bs=16 ---
+
+
+执行时间: 8.0210 s
+实际平均输入 tokens: 1.00（目标 1）
+生成总 tokens: 16384
+吞吐(生成tokens/秒): 2042.63
+TTFT (V1 metrics): 0.0178 s
+解码吞吐 (V1 metrics): 127.93 tok/s
+
+--- 批量大小 bs=32 ---
+
+
+执行时间: 8.0376 s
+实际平均输入 tokens: 1.00（目标 1）
+生成总 tokens: 32768
+吞吐(生成tokens/秒): 4076.81
+TTFT (V1 metrics): 0.0169 s
+解码吞吐 (V1 metrics): 127.81 tok/s
+
+--- 批量大小 bs=64 ---
+
+
+执行时间: 8.8357 s
+实际平均输入 tokens: 1.00（目标 1）
+生成总 tokens: 65536
+吞吐(生成tokens/秒): 7417.22
+TTFT (V1 metrics): 0.0206 s
+解码吞吐 (V1 metrics): 116.44 tok/s
+
+--- 批量大小 bs=128 ---
+
+
+执行时间: 11.1653 s
+实际平均输入 tokens: 1.00（目标 1）
+生成总 tokens: 131072
+吞吐(生成tokens/秒): 11739.23
+TTFT (V1 metrics): 0.0323 s
+解码吞吐 (V1 metrics): 92.27 tok/s
+
+--- 批量大小 bs=256 ---
+
+
+执行时间: 16.5525 s
+实际平均输入 tokens: 1.00（目标 1）
+生成总 tokens: 262144
+吞吐(生成tokens/秒): 15837.09
+TTFT (V1 metrics): 0.0482 s
+解码吞吐 (V1 metrics): 62.35 tok/s
+
+--- 批量大小 bs=512 ---
+
+
+执行时间: 30.5987 s
+实际平均输入 tokens: 1.00（目标 1）
+生成总 tokens: 524288
+吞吐(生成tokens/秒): 17134.34
+TTFT (V1 metrics): 0.1151 s
+解码吞吐 (V1 metrics): 34.41 tok/s
+
+--- 批量大小 bs=1024 ---
+
+
+[rank0]:[W813 21:35:54.931843453 ProcessGroupNCCL.cpp:1479] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
+执行时间: 65.0204 s
+实际平均输入 tokens: 1.00（目标 1）
+生成总 tokens: 1048576
+吞吐(生成tokens/秒): 16126.88
+TTFT (V1 metrics): 0.1185 s
+解码吞吐 (V1 metrics): 17.89 tok/s
+
+完成。提示：在 Nsight Systems 中可通过 NVTX 区间快速定位各场景/批量的调用。
+GPU 3: General Metrics for NVIDIA AD10x (any frequency)
+Generating '/tmp/nsys-report-5e6c.qdstrm'
+
+
+[3/8] Executing 'nvtx_sum' stats report
+
+ Time (%)  Total Time (ns)  Instances      Avg (ns)          Med (ns)         Min (ns)        Max (ns)     StdDev (ns)   Style                   Range                 
+ --------  ---------------  ---------  ----------------  ----------------  --------------  --------------  -----------  -------  --------------------------------------
+     30.8   65,019,709,544          1  65,019,709,544.0  65,019,709,544.0  65,019,709,544  65,019,709,544          0.0  PushPop  :generate [prefill1_decode512] bs=1024
+     16.4   34,528,225,294          1  34,528,225,294.0  34,528,225,294.0  34,528,225,294  34,528,225,294          0.0  PushPop  :LLM_init                             
+     14.5   30,598,009,745          1  30,598,009,745.0  30,598,009,745.0  30,598,009,745  30,598,009,745          0.0  PushPop  :generate [prefill1_decode512] bs=512 
+      7.8   16,552,100,314          1  16,552,100,314.0  16,552,100,314.0  16,552,100,314  16,552,100,314          0.0  PushPop  :generate [prefill1_decode512] bs=256 
+      5.3   11,165,129,518          1  11,165,129,518.0  11,165,129,518.0  11,165,129,518  11,165,129,518          0.0  PushPop  :generate [prefill1_decode512] bs=128 
+      4.2    8,835,508,288          1   8,835,508,288.0   8,835,508,288.0   8,835,508,288   8,835,508,288          0.0  PushPop  :generate [prefill1_decode512] bs=64  
+      3.8    8,037,503,827          1   8,037,503,827.0   8,037,503,827.0   8,037,503,827   8,037,503,827          0.0  PushPop  :generate [prefill1_decode512] bs=32  
+      3.8    8,020,861,613          1   8,020,861,613.0   8,020,861,613.0   8,020,861,613   8,020,861,613          0.0  PushPop  :generate [prefill1_decode512] bs=16  
+      3.6    7,502,439,278          1   7,502,439,278.0   7,502,439,278.0   7,502,439,278   7,502,439,278          0.0  PushPop  :generate [prefill1_decode512] bs=8   
+      3.4    7,275,415,153          1   7,275,415,153.0   7,275,415,153.0   7,275,415,153   7,275,415,153          0.0  PushPop  :generate [prefill1_decode512] bs=4   
+      3.4    7,204,870,191          1   7,204,870,191.0   7,204,870,191.0   7,204,870,191   7,204,870,191          0.0  PushPop  :generate [prefill1_decode512] bs=2   
+      3.0    6,423,331,403          1   6,423,331,403.0   6,423,331,403.0   6,423,331,403   6,423,331,403          0.0  PushPop  :generate [prefill1_decode512] bs=1   
+      0.0           97,220          2          48,610.0          48,610.0          47,061          50,159      2,190.6  PushPop  CCCL:cub::DeviceSegmentedRadixSort    
+
+[4/8] Executing 'osrt_sum' stats report
+
+ Time (%)   Total Time (ns)   Num Calls     Avg (ns)          Med (ns)      Min (ns)      Max (ns)        StdDev (ns)              Name         
+ --------  -----------------  ---------  ---------------  ----------------  ---------  ---------------  ----------------  ----------------------
+     29.4  2,592,016,067,300     89,110     29,087,824.8          37,176.5      1,000  196,381,349,289   1,472,090,154.0  pthread_cond_timedwait
+     24.1  2,131,420,272,351        832  2,561,803,212.0          17,028.0      6,962  196,382,583,332  22,164,783,605.9  pthread_cond_wait     
+     23.8  2,103,264,859,788    139,802     15,044,597.8      10,063,165.0      1,000  182,210,183,998     708,819,423.9  epoll_wait            
+      8.6    757,597,580,457         97  7,810,284,334.6  10,000,087,607.0      9,123   10,000,141,025   4,049,277,230.6  sem_timedwait         
+      8.1    715,974,502,372     73,851      9,694,851.8           1,424.0      1,000   30,237,953,598     166,855,063.6  poll                  
+      3.5    307,209,613,206     23,827     12,893,340.0       7,249,780.0     28,136      424,017,307      14,586,617.3  sem_wait              
+      2.4    212,178,624,327     52,115      4,071,354.2           2,015.0      1,000  194,509,488,551     853,307,229.4  read                  
+      0.0      4,204,021,023        771      5,452,686.2       3,184,955.0      1,046       47,259,678       6,190,683.1  pthread_rwlock_wrlock 
+      0.0        556,982,469    205,066          2,716.1           1,395.0      1,015       72,058,054         159,155.6  munmap                
+      0.0        508,440,444        693        733,680.3           1,074.0      1,000      501,674,334      19,056,845.3  waitpid               
+      0.0        282,788,889      8,972         31,519.0           8,207.5      1,000       19,677,480         317,215.8  ioctl                 
+      0.0        207,036,840        369        561,075.4           2,063.0      1,059       22,500,907       3,196,025.5  fopen                 
+      0.0        137,675,005     16,197          8,500.0           5,777.0      1,292           74,868           7,416.4  send                  
+      0.0        134,110,298     26,256          5,107.8           5,241.0      1,000        4,804,910          33,899.5  mmap64                
+      0.0        125,003,708     16,189          7,721.5           5,348.0      1,028        2,620,379          27,095.4  recv                  
+      0.0        124,828,097     71,937          1,735.2           1,593.0      1,000           57,072             823.9  pthread_cond_signal   
+      0.0        121,590,749         24      5,066,281.2       5,064,224.5  5,053,813        5,077,752           6,272.7  nanosleep             
+      0.0        109,066,343     11,638          9,371.6           6,763.0      1,012          350,375           9,043.1  pthread_mutex_lock    
+      0.0         93,279,596     27,220          3,426.9           2,358.5      1,024          153,763           5,757.8  write                 
+      0.0         88,935,356     31,463          2,826.7           2,452.0      1,000          113,592           1,855.8  open64                
+      0.0         69,874,922         94        743,350.2           3,294.5      1,031       18,719,114       3,530,346.8  open                  
+      0.0         68,894,671         38      1,813,017.7         580,319.0      4,100       10,403,699       3,383,579.7  pthread_join          
+      0.0         58,572,963         10      5,857,296.3          19,236.5      8,226       58,388,275      18,457,510.0  connect               
+      0.0         32,577,628         39        835,323.8       1,010,499.0     19,103        1,444,190         573,956.4  pthread_rwlock_rdlock 
+      0.0         29,134,753     18,187          1,602.0           1,332.0      1,000           44,360             802.2  epoll_ctl             
+      0.0         13,985,406        147         95,138.8          68,564.0     54,005        4,263,847         346,238.1  sleep                 
+      0.0          7,106,157        131         54,245.5          47,729.0     19,806          155,514          26,100.4  pthread_create        
+      0.0          5,907,972        868          6,806.4           4,068.0      1,008           97,687           8,868.4  fgets                 
+      0.0          1,715,987         65         26,399.8           3,330.0      1,055          473,306          73,248.1  futex                 
+      0.0          1,658,264        344          4,820.5           4,549.0      1,638           51,357           2,832.0  fopen64               
+      0.0          1,333,956      1,140          1,170.1           1,036.0      1,000            9,415             572.9  fclose                
+      0.0          1,043,640          1      1,043,640.0       1,043,640.0  1,043,640        1,043,640               0.0  fork                  
+      0.0            913,975        195          4,687.1           3,377.0      1,101           52,405           6,163.2  mmap                  
+      0.0            836,573          2        418,286.5         418,286.5      6,782          829,791         581,955.2  kill                  
+      0.0            358,003         65          5,507.7           4,718.0      2,188           14,024           2,925.3  pipe2                 
+      0.0            215,020         41          5,244.4           3,970.0      1,620           17,421           3,543.4  socket                
+      0.0            161,868         37          4,374.8           2,946.0      1,083           14,013           3,353.3  pthread_cond_broadcast
+      0.0            121,455         17          7,144.4           2,996.0      1,029           34,938          10,194.7  bind                  
+      0.0             77,221          7         11,031.6           7,581.0      4,133           36,236          11,339.1  fread                 
+      0.0             39,450         15          2,630.0           1,819.0      1,186            7,083           1,552.8  stat                  
+      0.0             37,451         23          1,628.3           1,663.0      1,048            2,012             323.2  sigaction             
+      0.0             35,497          5          7,099.4           8,915.0      3,047           10,850           3,562.4  accept4               
+      0.0             33,739         18          1,874.4           2,021.0      1,184            2,697             563.0  dup2                  
+      0.0             24,284         12          2,023.7           1,304.0      1,000            5,336           1,446.7  fcntl                 
+      0.0             23,323          5          4,664.6           3,577.0      2,140            8,317           2,528.9  fwrite                
+      0.0             22,194          7          3,170.6           3,746.0      1,157            5,067           1,793.0  fflush                
+      0.0             18,242          4          4,560.5           4,522.5      4,277            4,920             305.0  lstat                 
+      0.0             16,543          4          4,135.8           3,653.0      2,777            6,460           1,657.0  flock                 
+      0.0             15,834          8          1,979.3           1,650.0      1,263            3,278             825.8  pread                 
+      0.0             12,843          5          2,568.6           2,703.0      1,847            3,016             448.7  mprotect              
+      0.0             12,668          3          4,222.7           4,276.0      4,068            4,324             136.1  fputs_unlocked        
+      0.0              9,955          7          1,422.1           1,315.0      1,127            2,334             418.4  listen                
+      0.0              6,128          3          2,042.7           2,107.0      1,885            2,136             137.3  flockfile             
+      0.0              4,323          3          1,441.0           1,234.0      1,233            1,856             359.4  fstat                 
+      0.0              4,204          2          2,102.0           2,102.0      1,413            2,791             974.4  openat64              
+      0.0              3,018          1          3,018.0           3,018.0      3,018            3,018               0.0  fputs                 
+
+[5/8] Executing 'cuda_api_sum' stats report
+
+ Time (%)  Total Time (ns)  Num Calls   Avg (ns)     Med (ns)    Min (ns)   Max (ns)    StdDev (ns)                     Name                   
+ --------  ---------------  ---------  -----------  -----------  --------  -----------  -----------  ------------------------------------------
+     45.1   12,813,768,740     19,112    670,456.7      3,838.0     1,387  143,887,295  1,931,593.0  cudaStreamSynchronize                     
+     34.5    9,797,481,802  1,507,570      6,498.9      4,801.0       787   60,705,277    111,719.0  cudaLaunchKernel                          
+     11.7    3,315,922,316    267,434     12,399.0      9,714.0     7,444   13,005,521    119,290.9  cudaGraphLaunch_v10000                    
+      5.4    1,523,067,920    103,470     14,719.9      8,395.0     2,778  106,794,386    373,560.2  cudaMemcpyAsync                           
+      0.8      220,113,686      1,943    113,285.5     73,444.0    40,977    2,057,968    189,715.3  cudaGraphInstantiateWithFlags_v11040      
+      0.8      215,892,156    269,735        800.4        768.0       324       15,838        178.2  cudaStreamIsCapturing_v10000              
+      0.5      141,444,195      2,135     66,250.2     31,760.0     6,027   72,254,094  1,563,196.6  cudaDeviceSynchronize                     
+      0.5      137,833,300     38,438      3,585.9      4,058.0       652      155,081      2,294.0  cuLaunchKernel                            
+      0.2       67,108,465     14,485      4,633.0      5,021.0       183    3,200,351     28,944.6  cudaMemsetAsync                           
+      0.2       53,001,649        222    238,746.2    117,647.0    63,745    2,285,490    340,113.2  cudaFree                                  
+      0.1       40,868,863        348    117,439.3    109,378.5     5,752    1,125,389     63,264.5  cudaMalloc                                
+      0.1       24,861,223         10  2,486,122.3  2,553,390.5    56,705    4,320,787  1,404,300.5  cuLibraryLoadData                         
+      0.1       16,696,302     15,454      1,080.4        655.0       263    4,184,955     35,857.1  cuKernelGetFunction                       
+      0.0       13,006,871        169     76,963.7     68,902.0    27,343      356,192     46,120.4  cuModuleLoadData                          
+      0.0        9,079,336     18,895        480.5        453.0       297        8,904        131.4  cudaStreamGetCaptureInfo_v2_v11030        
+      0.0        8,354,276      1,943      4,299.7      4,189.0     3,373       20,932        803.4  cudaStreamBeginCapture_v10000             
+      0.0        7,675,564      1,943      3,950.4      3,913.0     2,470       13,545        599.9  cudaGraphDestroy_v10000                   
+      0.0        3,107,178        128     24,274.8      2,264.5     1,433    1,101,304    127,352.9  cudaStreamCreateWithPriority              
+      0.0        2,506,159      1,943      1,289.8      1,277.0       976        1,937        112.9  cudaStreamEndCapture_v10000               
+      0.0        1,952,904         24     81,371.0      8,307.5     3,591    1,241,446    249,577.9  cudaHostAlloc                             
+      0.0        1,595,388      1,943        821.1        749.0       636        2,762        252.4  cudaGraphGetNodes_v10000                  
+      0.0          293,162         98      2,991.4      3,357.0       887        6,560      1,554.2  cudaEventQuery                            
+      0.0          268,050         99      2,707.6      2,796.0     1,146        7,573      1,146.9  cudaEventRecord                           
+      0.0          255,527          8     31,940.9     29,099.5     9,678       69,087     22,331.0  cudaMemGetInfo                            
+      0.0          132,043        810        163.0        131.5        69        1,615        113.9  cuGetProcAddress_v2                       
+      0.0           22,323         16      1,395.2        921.0       431        4,897      1,232.4  cuLibraryGetKernel                        
+      0.0           20,937         21        997.0        467.0       277        4,218      1,099.3  cudaEventCreateWithFlags                  
+      0.0            8,033         14        573.8        570.5       312          918        173.3  cudaThreadExchangeStreamCaptureMode_v10010
+      0.0            4,402          3      1,467.3      1,282.0     1,262        1,858        338.5  cuInit                                    
+      0.0            3,502          1      3,502.0      3,502.0     3,502        3,502          0.0  cudaStreamWaitEvent                       
+      0.0            2,574          4        643.5        523.0        87        1,441        663.1  cuModuleGetLoadingMode                    
+      0.0            1,563          1      1,563.0      1,563.0     1,563        1,563          0.0  cudaEventDestroy                          
+      0.0              918          2        459.0        459.0       270          648        267.3  cudaGetDriverEntryPoint_v11030            
+
+[6/8] Executing 'cuda_gpu_kern_sum' stats report
+
+ Time (%)  Total Time (ns)  Instances   Avg (ns)     Med (ns)    Min (ns)   Max (ns)   StdDev (ns)                                                  Name                                                
+ --------  ---------------  ---------  -----------  -----------  ---------  ---------  -----------  ----------------------------------------------------------------------------------------------------
+     38.9    7,460,281,214     82,205     90,752.2     69,057.0      5,760    316,868     87,744.0  void flash::flash_fwd_splitkv_kernel<Flash_fwd_kernel_traits<(int)128, (int)64, (int)128, (int)4, (…
+     16.0    3,075,389,176    176,008     17,473.0      9,088.0      6,304    136,609     16,733.1  void flash::flash_fwd_splitkv_kernel<Flash_fwd_kernel_traits<(int)128, (int)64, (int)128, (int)4, (…
+      8.0    1,542,686,749      3,521    438,138.8    496,805.0     10,624    504,710    155,358.7  void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_16x16_128x2_tn_align8>(T1::Par…
+      7.1    1,364,722,257      2,113    645,869.5    573,543.0    186,818    763,722    119,005.8  ampere_bf16_s1688gemm_bf16_64x128_sliced1x2_ldg8_f2f_tn                                             
+      6.7    1,288,109,699      5,868    219,514.3     86,640.5      7,584    548,071    233,154.4  void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_32x6_tn_align8>(T1::Param…
+      4.7      899,528,528    458,577      1,961.6      1,888.0      1,632      3,583        325.8  void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_nocast<at::n…
+      2.7      520,608,894      1,025    507,911.1    507,878.0    506,278    509,990        414.6  void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_16x16_128x1_tn_align8>(T1::Par…
+      2.6      505,946,595      1,117    452,951.3    488,004.0      7,071    488,933    122,330.3  std::enable_if<!T7, void>::type internal::gemvx::kernel<int, int, __nv_bfloat16, __nv_bfloat16, __n…
+      1.9      368,335,550      9,223     39,936.6      7,808.0      1,952  1,005,962     67,492.9  void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::TensorIterator…
+      1.8      345,618,817    258,213      1,338.5      1,248.0        992      2,047        191.4  void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0…
+      1.6      316,168,888      9,225     34,273.1      8,224.0      5,152    712,902     62,162.9  void at::native::reduce_kernel<(int)512, (int)1, at::native::ReduceOp<float, at::native::ArgMaxOps<…
+      1.6      309,202,466    229,288      1,348.5      1,312.0      1,183      2,080        171.7  void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_nocast<at::n…
+      1.3      250,234,957     71,932      3,478.8      3,488.0      3,231      4,128        128.3  void flash::flash_fwd_splitkv_combine_kernel<Flash_fwd_kernel_traits<(int)128, (int)64, (int)128, (…
+      0.9      179,336,549     57,456      3,121.3      3,135.0      2,943      4,992         78.7  void flash::flash_fwd_splitkv_combine_kernel<Flash_fwd_kernel_traits<(int)128, (int)64, (int)128, (…
+      0.8      157,635,915        924    170,601.6    162,545.5     40,256  1,415,021    225,677.4  ampere_bf16_s1688gemm_bf16_128x64_sliced1x2_ldg8_f2f_tn                                             
+      0.7      140,934,798     46,620      3,023.1      3,040.0      2,847      3,232         98.2  void flash::flash_fwd_splitkv_combine_kernel<Flash_fwd_kernel_traits<(int)128, (int)64, (int)128, (…
+      0.4       79,079,082         29  2,726,864.9  2,804,313.0    584,168  2,808,793    412,123.5  ampere_bf16_s16816gemm_bf16_128x64_ldg8_f2f_tn                                                      
+      0.2       39,299,362          8  4,912,420.3  4,870,556.0  4,820,171  5,103,438    113,542.1  void at_cuda_detail::cub::DeviceSegmentedRadixSortKernel<at_cuda_detail::cub::DeviceRadixSortPolicy…
+      0.2       37,974,768     27,663      1,372.8      1,376.0      1,055      2,145        222.9  void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::TensorIterator…
+      0.2       29,065,343      9,357      3,106.3      3,104.0      2,687      7,488        192.9  void at::native::index_elementwise_kernel<(int)128, (int)4, void at::native::gpu_index_kernel<void …
+      0.1       22,668,389        784     28,913.8     12,512.0     11,584     62,529     20,679.1  ampere_bf16_s16816gemm_bf16_64x64_ldg8_f2f_stages_64x5_tn                                           
+      0.1       21,873,260      9,221      2,372.1      2,336.0      1,952      3,008        199.7  void at::native::index_elementwise_kernel<(int)128, (int)4, void at::native::gpu_index_kernel<void …
+      0.1       21,556,364      1,932     11,157.5      4,384.0      1,055    461,252     54,521.7  triton_poi_fused_mul_silu_1                                                                         
+      0.1       20,487,032          4  5,121,758.0  5,121,598.5  4,943,820  5,300,015    200,903.2  void at_cuda_detail::cub::DeviceSegmentedRadixSortKernel<at_cuda_detail::cub::DeviceRadixSortPolicy…
+      0.1       14,291,746         28    510,419.5    511,797.0    469,540    512,996      8,022.2  void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<signed char>, std::a…
+      0.1       13,279,798        142     93,519.7     51,680.0     50,400  3,040,379    351,298.3  ampere_bf16_s1688gemm_bf16_128x128_ldg8_f2f_stages_32x1_tn                                          
+      0.1       10,969,719      9,223      1,189.4      1,152.0      1,023      1,760        134.6  void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::TensorIterator…
+      0.1       10,416,865     12,059        863.8        895.0        767      1,280         54.5  void at::native::vectorized_elementwise_kernel<(int)2, at::native::FillFunctor<long>, std::array<ch…
+      0.1       10,035,051      1,932      5,194.1      3,455.5      1,568    112,066     12,675.8  triton_red_fused__to_copy_add_mean_mul_pow_rsqrt_2                                                  
+      0.1        9,775,446          4  2,443,861.5  2,443,525.5  2,389,237  2,499,158     60,370.1  void at::native::<unnamed>::cunn_SoftMaxForward<(int)4, float, float, float, at::native::<unnamed>:…
+      0.0        9,151,987         28    326,856.7    327,058.5    324,132    331,139      1,248.5  ampere_bf16_s1688gemm_bf16_128x128_ldg8_relu_f2f_stages_32x1_tn                                     
+      0.0        8,406,218        224     37,527.8     37,536.0     36,608     38,752        441.5  void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x128_32x6_tn_align8>(T1::Para…
+      0.0        8,365,002      9,222        907.1        896.0        895        993         15.6  void at::native::unrolled_elementwise_kernel<at::native::CUDAFunctorOnSelf_add<int>, std::array<cha…
+      0.0        8,036,497      9,222        871.4        864.0        831        960         17.1  void at::native::unrolled_elementwise_kernel<at::native::FillFunctor<int>, std::array<char *, (unsi…
+      0.0        7,799,398          2  3,899,699.0  3,899,699.0  3,708,001  4,091,397    271,101.9  void at::native::_scatter_gather_elementwise_kernel<(int)128, (int)8, void at::native::_cuda_scatte…
+      0.0        7,231,222        336     21,521.5     21,440.0     21,088     22,752        357.7  ampere_bf16_s16816gemm_bf16_128x64_ldg8_relu_f2f_stages_64x3_tn                                     
+      0.0        6,510,065      7,157        909.6        895.0        832      2,080         79.6  void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<int>, std::array<cha…
+      0.0        6,285,367      1,932      3,253.3      2,017.0      1,535     78,849      9,097.9  triton_red_fused__to_copy_add_mean_mul_pow_rsqrt_0                                                  
+      0.0        6,109,827        476     12,835.8     12,832.0     11,584     14,368        733.3  ampere_bf16_s16816gemm_bf16_64x64_ldg8_relu_f2f_stages_64x5_tn                                      
+      0.0        5,896,822          4  1,474,205.5  1,473,277.0  1,473,006  1,477,262      2,045.1  void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::masked_fill_kernel(at…
+      0.0        4,385,449        140     31,324.6     27,232.0     26,080     50,113      8,339.5  ampere_bf16_s1688gemm_bf16_128x64_sliced1x2_ldg8_relu_f2f_tn                                        
+      0.0        4,015,431      1,863      2,155.4      1,823.0      1,375     22,753      2,300.3  triton_poi_fused_cat_3                                                                              
+      0.0        3,999,332          2  1,999,666.0  1,999,666.0  1,997,394  2,001,938      3,213.1  void at::native::vectorized_elementwise_kernel<(int)4, at::native::BinaryFunctor<float, float, floa…
+      0.0        3,594,888         56     64,194.4     64,161.0     63,041     65,313        437.2  void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_64x1_tn_align8>(T1::Para…
+      0.0        3,434,878          4    858,719.5    857,975.5    856,167    862,760      3,016.9  void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_nocast<at::n…
+      0.0        3,192,956          2  1,596,478.0  1,596,478.0  1,565,902  1,627,054     43,241.0  void at::native::tensor_kernel_scan_innermost_dim<float, std::plus<float>>(T1 *, const T1 *, unsign…
+      0.0        2,856,853      1,512      1,889.5      1,712.0      1,311      2,944        453.7  void cublasLt::splitKreduce_kernel<(int)32, (int)16, int, __nv_bfloat16, __nv_bfloat16, float, (boo…
+      0.0        2,644,615      1,863      1,419.5      1,151.0        863     16,640      1,781.0  triton_poi_fused_view_5                                                                             
+      0.0        2,613,678      1,863      1,402.9      1,344.0      1,215      6,624        562.0  triton_poi_fused_cat_4                                                                              
+      0.0        2,581,335          2  1,290,667.5  1,290,667.5  1,290,379  1,290,956        408.0  at::native::<unnamed>::fill_reverse_indices_kernel(long *, int, at::cuda::detail::IntDivider<unsign…
+      0.0        2,579,032          2  1,289,516.0  1,289,516.0  1,287,756  1,291,276      2,489.0  void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_nocast<at::n…
+      0.0        2,426,037        112     21,661.0     21,568.0      9,473     34,721     12,050.1  void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align8>(T1::Par…
+      0.0        1,840,620      2,069        889.6        896.0        863        960         16.1  void at::native::vectorized_elementwise_kernel<(int)2, at::native::FillFunctor<int>, std::array<cha…
+      0.0        1,371,084          2    685,542.0    685,542.0    680,966    690,118      6,471.4  void at::native::<unnamed>::distribution_elementwise_grid_stride_kernel<float, (int)4, void at::nat…
+      0.0        1,149,932         28     41,069.0     41,072.0     40,449     43,297        529.2  ampere_bf16_s1688gemm_bf16_64x64_sliced1x4_ldg8_f2f_tn                                              
+      0.0          962,005      1,054        912.7        897.0        895        992         18.8  void at::native::unrolled_elementwise_kernel<at::native::FillFunctor<long>, std::array<char *, (uns…
+      0.0          957,484         28     34,195.9     34,848.5     18,177     35,297      3,147.6  std::enable_if<!T7, void>::type internal::gemvx::kernel<int, int, __nv_bfloat16, float, float, floa…
+      0.0          295,678        168      1,760.0      1,760.0      1,535      1,984        116.8  void cublasLt::splitKreduce_kernel<(int)32, (int)16, int, __nv_bfloat16, __nv_bfloat16, float, (boo…
+      0.0          254,179         69      3,683.8      3,040.0      2,016     36,896      4,118.3  triton_red_fused__to_copy_add_embedding_mean_mul_pow_rsqrt_0                                        
+      0.0          186,527         69      2,703.3      2,080.0      1,600     39,296      4,485.1  triton_poi_fused_cat_1                                                                              
+      0.0          155,234          1    155,234.0    155,234.0    155,234    155,234          0.0  void at::native::<unnamed>::CatArrayBatchedCopy_aligned16_contig<at::native::<unnamed>::OpaqueType<…
+      0.0           99,648         69      1,444.2      1,344.0      1,216      8,704        890.0  triton_poi_fused_cat_2                                                                              
+      0.0           97,824         69      1,417.7      1,184.0        864     14,337      1,605.4  triton_poi_fused_view_3                                                                             
+      0.0           79,521          1     79,521.0     79,521.0     79,521     79,521          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::bfloat16_copy_kernel_cuda(at::Te…
+      0.0           64,605         58      1,113.9        896.0        895     11,328      1,368.2  void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<c10::BFloat16>, std:…
+      0.0           44,704          1     44,704.0     44,704.0     44,704     44,704          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::sin_kernel_cuda(at::TensorIterat…
+      0.0           36,892         28      1,317.6      1,312.0      1,280      1,375         23.0  void cublasLt::splitKreduce_kernel<(int)32, (int)16, int, float, __nv_bfloat16, float, (bool)0, __n…
+      0.0           26,657          1     26,657.0     26,657.0     26,657     26,657          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::cos_kernel_cuda(at::TensorIterat…
+      0.0           19,520          1     19,520.0     19,520.0     19,520     19,520          0.0  void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_nocast<at::n…
+      0.0           11,903         11      1,082.1        865.0        863      1,567        288.8  void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<float>, std::array<c…
+      0.0           11,009          2      5,504.5      5,504.5      5,472      5,537         46.0  void at::native::_scatter_gather_elementwise_kernel<(int)128, (int)8, void at::native::_cuda_scatte…
+      0.0            9,217          2      4,608.5      4,608.5      4,545      4,672         89.8  void at::native::<unnamed>::distribution_elementwise_grid_stride_kernel<float, (int)4, void at::nat…
+      0.0            3,809          2      1,904.5      1,904.5      1,665      2,144        338.7  void at::native::vectorized_elementwise_kernel<(int)2, at::native::CUDAFunctorOnOther_add<long>, st…
+      0.0            3,616          2      1,808.0      1,808.0      1,600      2,016        294.2  void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_nocast<at::n…
+      0.0            3,232          2      1,616.0      1,616.0      1,504      1,728        158.4  void at::native::vectorized_elementwise_kernel<(int)2, at::native::<unnamed>::where_kernel_impl(at:…
+      0.0            3,103          2      1,551.5      1,551.5      1,120      1,983        610.2  void <unnamed>::elementwise_kernel_with_index<int, at::native::arange_cuda_out(const c10::Scalar &,…
+      0.0            2,913          2      1,456.5      1,456.5      1,345      1,568        157.7  void at::native::vectorized_elementwise_kernel<(int)4, void at::native::compare_scalar_kernel<float…
+      0.0            2,880          2      1,440.0      1,440.0      1,312      1,568        181.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnOther_add<float>, s…
+      0.0            2,304          1      2,304.0      2,304.0      2,304      2,304          0.0  void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl<at::native::…
+      0.0            1,216          1      1,216.0      1,216.0      1,216      1,216          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::reciprocal_kernel_cuda(at::Tenso…
+      0.0            1,025          1      1,025.0      1,025.0      1,025      1,025          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, float, floa…
+      0.0              992          1        992.0        992.0        992        992          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::BUnaryFunctor<float, float, floa…
+      0.0              895          1        895.0        895.0        895        895          0.0  void at::native::vectorized_elementwise_kernel<(int)2, at::native::FillFunctor<double>, std::array<…
+
+[7/8] Executing 'cuda_gpu_mem_time_sum' stats report
+
+ Time (%)  Total Time (ns)  Count   Avg (ns)  Med (ns)  Min (ns)   Max (ns)    StdDev (ns)            Operation           
+ --------  ---------------  ------  --------  --------  --------  -----------  -----------  ------------------------------
+     91.9      532,693,904  65,321   8,155.0     352.0       287  106,336,118    441,449.3  [CUDA memcpy Host-to-Device]  
+      5.5       31,855,936  28,928   1,101.2     897.0       864    1,362,988     15,981.0  [CUDA memcpy Device-to-Device]
+      1.8       10,196,329   9,221   1,105.8   1,120.0       863        1,696         90.0  [CUDA memcpy Device-to-Host]  
+      0.8        4,628,573  11,573     399.9     352.0       288        2,240        166.8  [CUDA memset]                 
+
+[8/8] Executing 'cuda_gpu_mem_size_sum' stats report
+
+ Total (MB)  Count   Avg (MB)  Med (MB)  Min (MB)  Max (MB)  StdDev (MB)            Operation           
+ ----------  ------  --------  --------  --------  --------  -----------  ------------------------------
+  3,634.332  65,321     0.056     0.000     0.000   466.747        2.080  [CUDA memcpy Host-to-Device]  
+  2,578.174  28,928     0.089     0.003     0.003   622.330        7.318  [CUDA memcpy Device-to-Device]
+      4.048  11,573     0.000     0.000     0.000     0.006        0.001  [CUDA memset]                 
+      2.077   9,221     0.000     0.000     0.000     0.001        0.000  [CUDA memcpy Device-to-Host]  
+
+Generated:
+    /data/cy/kv_cache_vs_util/sim_traverse_bs/traverse_bs_util_sim_decoding_1024.nsys-rep
+    /data/cy/kv_cache_vs_util/sim_traverse_bs/traverse_bs_util_sim_decoding_1024.sqlite

sim_traverse_bs/traverse_bs_util_sim_decoding_1024.nsys-rep

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:076aaf8297b058dc3a50d9858116d1dbaa2dbe038c6ca5881b0edb7f2fb3d314
+size 177488267

sim_traverse_bs/traverse_bs_util_sim_prefill.log

@@ -0,0 +1,386 @@
+WARNING: CPU IP/backtrace sampling not supported, disabling.
+Try the 'nsys status --environment' command to learn more.
+
+WARNING: CPU context switch tracing not supported, disabling.
+Try the 'nsys status --environment' command to learn more.
+
+INFO 08-13 20:36:20 [__init__.py:235] Automatically detected platform cuda.
+CUDA_VISIBLE_DEVICES = 3
+--- vLLM V1 基准测试（含 NVTX 标记）---
+模型: Qwen/Qwen2-1.5B
+批量大小: [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024]
+场景: ['prefill640_decode1']
+------------------------------------------------------------
+加载分词器/模型中...
+INFO 08-13 20:36:30 [config.py:1604] Using max model len 4096
+INFO 08-13 20:36:30 [config.py:2434] Chunked prefill is enabled with max_num_batched_tokens=8192.
+INFO 08-13 20:36:35 [__init__.py:235] Automatically detected platform cuda.
+INFO 08-13 20:36:37 [core.py:572] Waiting for init message from front-end.
+INFO 08-13 20:36:37 [core.py:71] Initializing a V1 LLM engine (v0.10.0) with config: model='Qwen/Qwen2-1.5B', speculative_config=None, tokenizer='Qwen/Qwen2-1.5B', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config={}, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=4096, download_dir=None, load_format=LoadFormat.AUTO, tensor_parallel_size=1, pipeline_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=False, kv_cache_dtype=auto,  device_config=cuda, decoding_config=DecodingConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_backend=''), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None), seed=0, served_model_name=Qwen/Qwen2-1.5B, num_scheduler_steps=1, multi_step_stream_outputs=True, enable_prefix_caching=True, chunked_prefill_enabled=True, use_async_output_proc=True, pooler_config=None, compilation_config={"level":3,"debug_dump_path":"","cache_dir":"","backend":"","custom_ops":[],"splitting_ops":["vllm.unified_attention","vllm.unified_attention_with_output","vllm.mamba_mixer2"],"use_inductor":true,"compile_sizes":[],"inductor_compile_config":{"enable_auto_functionalized_v2":false},"inductor_passes":{},"use_cudagraph":true,"cudagraph_num_of_warmups":1,"cudagraph_capture_sizes":[512,504,496,488,480,472,464,456,448,440,432,424,416,408,400,392,384,376,368,360,352,344,336,328,320,312,304,296,288,280,272,264,256,248,240,232,224,216,208,200,192,184,176,168,160,152,144,136,128,120,112,104,96,88,80,72,64,56,48,40,32,24,16,8,4,2,1],"cudagraph_copy_inputs":false,"full_cuda_graph":false,"max_capture_size":512,"local_cache_dir":null}
+INFO 08-13 20:36:39 [parallel_state.py:1102] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, TP rank 0, EP rank 0
+WARNING 08-13 20:36:39 [topk_topp_sampler.py:59] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
+INFO 08-13 20:36:39 [gpu_model_runner.py:1843] Starting to load model Qwen/Qwen2-1.5B...
+INFO 08-13 20:36:39 [gpu_model_runner.py:1875] Loading model from scratch...
+INFO 08-13 20:36:39 [cuda.py:290] Using Flash Attention backend on V1 engine.
+INFO 08-13 20:36:40 [weight_utils.py:296] Using model weights format ['*.safetensors']
+INFO 08-13 20:36:40 [weight_utils.py:349] No model.safetensors.index.json found in remote.
+
+
+
+
+INFO 08-13 20:36:41 [default_loader.py:262] Loading weights took 0.74 seconds
+INFO 08-13 20:36:42 [gpu_model_runner.py:1892] Model loading took 2.9105 GiB and 2.308288 seconds
+INFO 08-13 20:36:48 [backends.py:530] Using cache directory: /home/cy/.cache/vllm/torch_compile_cache/40b61c71e9/rank_0_0/backbone for vLLM's torch.compile
+INFO 08-13 20:36:48 [backends.py:541] Dynamo bytecode transform time: 6.35 s
+INFO 08-13 20:36:53 [backends.py:161] Directly load the compiled graph(s) for dynamic shape from the cache, took 3.954 s
+INFO 08-13 20:36:54 [monitor.py:34] torch.compile takes 6.35 s in total
+INFO 08-13 20:36:54 [gpu_worker.py:255] Available KV cache memory: 12.81 GiB
+INFO 08-13 20:36:55 [kv_cache_utils.py:833] GPU KV cache size: 479,536 tokens
+INFO 08-13 20:36:55 [kv_cache_utils.py:837] Maximum concurrency for 4,096 tokens per request: 117.07x
+
+INFO 08-13 20:36:57 [gpu_model_runner.py:2485] Graph capturing finished in 2 secs, took 0.49 GiB
+INFO 08-13 20:36:57 [core.py:193] init engine (profile, create kv cache, warmup model) took 15.10 seconds
+模型加载完成。
+
+===== 场景：prefill640_decode1 | prefill=640, decode=1 =====
+
+--- 批量大小 bs=1 ---
+预热中...
+
+
+
+
+执行时间: 0.0106 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 1
+吞吐(生成tokens/秒): 94.66
+TTFT (V1 metrics): 0.0098 s
+解码吞吐 (V1 metrics): nan tok/s
+
+--- 批量大小 bs=2 ---
+
+
+执行时间: 0.0193 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 2
+吞吐(生成tokens/秒): 103.58
+TTFT (V1 metrics): 0.0140 s
+解码吞吐 (V1 metrics): nan tok/s
+
+--- 批量大小 bs=4 ---
+
+
+执行时间: 0.0242 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 4
+吞吐(生成tokens/秒): 165.24
+TTFT (V1 metrics): 0.0194 s
+解码吞吐 (V1 metrics): nan tok/s
+
+--- 批量大小 bs=8 ---
+
+
+执行时间: 0.0370 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 8
+吞吐(生成tokens/秒): 216.13
+TTFT (V1 metrics): 0.0202 s
+解码吞吐 (V1 metrics): nan tok/s
+
+--- 批量大小 bs=16 ---
+
+
+执行时间: 0.0466 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 16
+吞吐(生成tokens/秒): 343.63
+TTFT (V1 metrics): 0.0267 s
+解码吞吐 (V1 metrics): nan tok/s
+
+--- 批量大小 bs=32 ---
+
+
+执行时间: 0.0620 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 32
+吞吐(生成tokens/秒): 516.16
+TTFT (V1 metrics): 0.0279 s
+解码吞吐 (V1 metrics): nan tok/s
+
+--- 批量大小 bs=64 ---
+
+
+执行时间: 0.1069 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 64
+吞吐(生成tokens/秒): 598.81
+TTFT (V1 metrics): 0.0496 s
+解码吞吐 (V1 metrics): nan tok/s
+
+--- 批量大小 bs=128 ---
+
+
+执行时间: 0.1979 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 128
+吞吐(生成tokens/秒): 646.67
+TTFT (V1 metrics): 0.0973 s
+解码吞吐 (V1 metrics): nan tok/s
+
+--- 批量大小 bs=256 ---
+
+
+执行时间: 0.3878 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 256
+吞吐(生成tokens/秒): 660.13
+TTFT (V1 metrics): 0.1947 s
+解码吞吐 (V1 metrics): nan tok/s
+
+--- 批量大小 bs=512 ---
+
+
+执行时间: 0.9331 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 512
+吞吐(生成tokens/秒): 548.69
+TTFT (V1 metrics): 0.3830 s
+解码吞吐 (V1 metrics): nan tok/s
+
+--- 批量大小 bs=1024 ---
+
+
+[rank0]:[W813 20:37:02.968083070 ProcessGroupNCCL.cpp:1479] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
+执行时间: 1.4194 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 1024
+吞吐(生成tokens/秒): 721.42
+TTFT (V1 metrics): 0.7078 s
+解码吞吐 (V1 metrics): nan tok/s
+
+完成。提示：在 Nsight Systems 中可通过 NVTX 区间快速定位各场景/批量的调用。
+GPU 3: General Metrics for NVIDIA AD10x (any frequency)
+Generating '/tmp/nsys-report-97f6.qdstrm'
+
+
+[3/8] Executing 'nvtx_sum' stats report
+
+ Time (%)  Total Time (ns)  Instances      Avg (ns)          Med (ns)         Min (ns)        Max (ns)     StdDev (ns)   Style                   Range                 
+ --------  ---------------  ---------  ----------------  ----------------  --------------  --------------  -----------  -------  --------------------------------------
+     91.5   34,993,424,316          1  34,993,424,316.0  34,993,424,316.0  34,993,424,316  34,993,424,316          0.0  PushPop  :LLM_init                             
+      3.7    1,419,326,629          1   1,419,326,629.0   1,419,326,629.0   1,419,326,629   1,419,326,629          0.0  PushPop  :generate [prefill640_decode1] bs=1024
+      2.4      933,046,428          1     933,046,428.0     933,046,428.0     933,046,428     933,046,428          0.0  PushPop  :generate [prefill640_decode1] bs=512 
+      1.0      387,636,888          1     387,636,888.0     387,636,888.0     387,636,888     387,636,888          0.0  PushPop  :generate [prefill640_decode1] bs=256 
+      0.5      197,863,320          1     197,863,320.0     197,863,320.0     197,863,320     197,863,320          0.0  PushPop  :generate [prefill640_decode1] bs=128 
+      0.3      106,809,096          1     106,809,096.0     106,809,096.0     106,809,096     106,809,096          0.0  PushPop  :generate [prefill640_decode1] bs=64  
+      0.2       61,927,285          1      61,927,285.0      61,927,285.0      61,927,285      61,927,285          0.0  PushPop  :generate [prefill640_decode1] bs=32  
+      0.1       46,494,927          1      46,494,927.0      46,494,927.0      46,494,927      46,494,927          0.0  PushPop  :generate [prefill640_decode1] bs=16  
+      0.1       36,948,348          1      36,948,348.0      36,948,348.0      36,948,348      36,948,348          0.0  PushPop  :generate [prefill640_decode1] bs=8   
+      0.1       24,170,392          1      24,170,392.0      24,170,392.0      24,170,392      24,170,392          0.0  PushPop  :generate [prefill640_decode1] bs=4   
+      0.1       19,267,324          1      19,267,324.0      19,267,324.0      19,267,324      19,267,324          0.0  PushPop  :generate [prefill640_decode1] bs=2   
+      0.0       10,500,291          1      10,500,291.0      10,500,291.0      10,500,291      10,500,291          0.0  PushPop  :generate [prefill640_decode1] bs=1   
+      0.0           99,802          2          49,901.0          49,901.0          44,120          55,682      8,175.6  PushPop  CCCL:cub::DeviceSegmentedRadixSort    
+
+[4/8] Executing 'osrt_sum' stats report
+
+ Time (%)  Total Time (ns)  Num Calls      Avg (ns)        Med (ns)    Min (ns)      Max (ns)       StdDev (ns)              Name         
+ --------  ---------------  ---------  ----------------  ------------  ---------  --------------  ----------------  ----------------------
+     31.9  315,906,533,849     13,750      22,975,020.6      26,070.5      1,015  23,666,668,799     454,327,983.2  pthread_cond_timedwait
+     26.2  259,640,374,556     18,725      13,865,974.6  10,062,170.0      1,054  26,101,659,008     334,515,177.4  epoll_wait            
+     23.8  236,353,668,764         22  10,743,348,580.2   8,483,110.0     20,022  23,667,768,753  12,042,371,346.2  pthread_cond_wait     
+      7.8   77,638,705,690      9,817       7,908,597.9       1,366.0      1,000  10,010,042,018     161,807,052.9  poll                  
+      6.1   60,833,119,112         28   2,172,611,396.9      27,845.0      9,196  10,000,124,017   3,889,257,455.2  sem_timedwait         
+      3.6   35,208,374,909     31,030       1,134,656.0       2,282.0      1,000  21,397,007,347     131,447,597.0  read                  
+      0.4    3,887,231,292        311      12,499,136.0   7,213,077.0     36,367     322,090,803      29,803,312.7  sem_wait              
+      0.0      474,955,889    193,339           2,456.6       1,418.0      1,001      94,954,678         215,965.8  munmap                
+      0.0      265,992,610      8,412          31,620.6       9,835.5      1,001      17,044,888         308,579.4  ioctl                 
+      0.0      222,714,707        369         603,562.9       2,105.0      1,033      20,700,195       3,290,651.8  fopen                 
+      0.0      121,534,035         24       5,063,918.1   5,063,775.0  5,040,945       5,076,986           7,972.2  nanosleep             
+      0.0       81,450,488     30,629           2,659.3       2,313.0      1,000          92,932           1,646.5  open64                
+      0.0       74,702,014        133         561,669.3       1,892.0      1,000      66,577,673       5,774,257.5  waitpid               
+      0.0       74,234,231         96         773,273.2       3,712.0      1,003      22,285,364       3,739,689.4  open                  
+      0.0       66,582,083         37       1,799,515.8     480,507.0      3,919      10,336,124       3,413,414.9  pthread_join          
+      0.0       63,806,761         10       6,380,676.1      20,040.5      8,233      63,627,951      20,114,646.1  connect               
+      0.0       46,888,026     12,161           3,855.6       2,570.0      1,000       1,071,189          14,531.9  mmap64                
+      0.0       32,641,092        147         222,048.2      68,846.0     55,054       7,991,497       1,073,129.2  sleep                 
+      0.0       22,603,585     11,001           2,054.7       1,785.0      1,000          53,255           2,018.4  pthread_cond_signal   
+      0.0       17,553,779      2,362           7,431.7       4,197.0      1,033       2,743,900          58,994.9  recv                  
+      0.0       15,451,862      2,365           6,533.6       5,474.0      1,382          75,703           5,847.3  send                  
+      0.0       14,269,726      1,497           9,532.2       7,368.0      1,000         543,140          16,462.6  pthread_mutex_lock    
+      0.0       13,536,775      4,653           2,909.3       2,179.0      1,006         121,272           4,763.9  write                 
+      0.0        6,900,801        131          52,677.9      46,855.0     21,276         130,729          19,309.4  pthread_create        
+      0.0        5,993,355      4,277           1,401.3       1,350.0      1,000          17,968             515.6  epoll_ctl             
+      0.0        5,355,828        887           6,038.1       3,980.0      1,016          43,090           7,169.9  fgets                 
+      0.0        1,980,565        344           5,757.5       5,562.5      2,076          45,811           2,514.7  fopen64               
+      0.0        1,764,653      1,470           1,200.4       1,068.0      1,000           9,187             564.1  fclose                
+      0.0        1,642,199         16         102,637.4      81,769.0      3,412         331,583         101,402.9  pthread_rwlock_wrlock 
+      0.0        1,173,080        195           6,015.8       3,724.0      1,150         124,348          12,102.3  mmap                  
+      0.0        1,159,048         61          19,000.8       3,157.0      1,938         319,207          49,006.4  futex                 
+      0.0          984,289          1         984,289.0     984,289.0    984,289         984,289               0.0  fork                  
+      0.0          778,197          9          86,466.3      47,775.0      4,586         209,086          76,949.1  pthread_rwlock_rdlock 
+      0.0          360,142         65           5,540.6       4,626.0      1,914          16,148           2,939.0  pipe2                 
+      0.0          236,233         41           5,761.8       4,572.0      2,034          12,662           3,154.3  socket                
+      0.0          164,351         25           6,574.0       2,657.0      1,052          62,060          13,072.9  bind                  
+      0.0          112,444         29           3,877.4       3,400.0      1,217          10,296           2,247.2  pthread_cond_broadcast
+      0.0           69,279          7           9,897.0       7,229.0      3,601          30,113           9,209.7  fread                 
+      0.0           42,530          5           8,506.0       7,533.0      3,596          14,826           4,155.8  accept4               
+      0.0           42,447         27           1,572.1       1,691.0      1,025           1,981             319.5  sigaction             
+      0.0           38,331         20           1,916.6       1,877.5      1,040           2,780             659.7  dup2                  
+      0.0           37,721         23           1,640.0       1,115.0      1,012           5,412           1,165.0  fcntl                 
+      0.0           37,457         15           2,497.1       2,022.0      1,236           5,421           1,128.1  stat                  
+      0.0           26,843          8           3,355.4       3,101.5      1,029           7,199           2,500.2  fflush                
+      0.0           26,508         10           2,650.8       1,797.0      1,023           7,028           1,852.3  pread                 
+      0.0           19,153          4           4,788.3       5,290.5      3,065           5,507           1,166.3  lstat                 
+      0.0           18,990          5           3,798.0       3,258.0      1,533           8,332           2,664.1  fwrite                
+      0.0           14,221          4           3,555.3       3,514.0      1,783           5,410           1,691.6  flock                 
+      0.0           14,049          3           4,683.0       4,386.0      4,334           5,329             560.1  fputs_unlocked        
+      0.0           13,821          5           2,764.2       2,680.0      2,269           3,685             568.1  mprotect              
+      0.0           12,045          7           1,720.7       1,457.0      1,329           2,524             461.1  listen                
+      0.0            9,226          3           3,075.3       2,227.0      1,736           5,263           1,910.4  fstat                 
+      0.0            7,086          3           2,362.0       1,954.0      1,531           3,601           1,093.7  flockfile             
+      0.0            5,765          1           5,765.0       5,765.0      5,765           5,765               0.0  kill                  
+      0.0            4,207          1           4,207.0       4,207.0      4,207           4,207               0.0  fputs                 
+      0.0            3,713          2           1,856.5       1,856.5      1,397           2,316             649.8  openat64              
+      0.0            1,334          1           1,334.0       1,334.0      1,334           1,334               0.0  pthread_mutex_trylock 
+
+[5/8] Executing 'cuda_api_sum' stats report
+
+ Time (%)  Total Time (ns)  Num Calls   Avg (ns)     Med (ns)    Min (ns)   Max (ns)    StdDev (ns)                     Name                   
+ --------  ---------------  ---------  -----------  -----------  --------  -----------  -----------  ------------------------------------------
+     31.1      698,997,620      2,853    245,004.4      8,462.0     3,452  100,828,816  2,074,316.5  cudaMemcpyAsync                           
+     25.2      565,993,891     37,122     15,246.9      7,556.0       745   58,676,993    479,701.4  cudaLaunchKernel                          
+      9.5      214,098,539      1,943    110,189.7     70,984.0    41,015    1,466,127    183,798.1  cudaGraphInstantiateWithFlags_v11040      
+      9.0      203,147,257      1,162    174,825.5      3,666.0     1,917  144,202,830  4,251,626.0  cudaStreamSynchronize                     
+      6.1      136,018,801      2,136     63,679.2     30,541.5     2,774   71,669,241  1,550,140.5  cudaDeviceSynchronize                     
+      5.7      127,417,770      7,569     16,834.2     15,015.0     9,892      226,168      6,274.9  cudaGraphLaunch_v10000                    
+      4.6      102,280,934     32,723      3,125.7      3,725.0       617      163,047      2,253.8  cuLaunchKernel                            
+      2.4       54,199,644        222    244,142.5    119,315.0    63,972    2,400,814    360,610.3  cudaFree                                  
+      1.9       41,686,022        348    119,787.4    110,186.0     4,423    1,038,043     58,088.4  cudaMalloc                                
+      1.1       24,826,279         10  2,482,627.9  2,644,171.0    54,814    4,281,423  1,392,790.3  cuLibraryLoadData                         
+      0.7       15,112,641      5,582      2,707.4      1,648.5       157       37,305      2,202.8  cudaMemsetAsync                           
+      0.5       11,339,187        169     67,095.8     76,671.0    26,797      270,701     31,403.2  cuModuleLoadData                          
+      0.5       10,277,547      9,570      1,073.9        392.0       261    4,171,570     44,224.2  cuKernelGetFunction                       
+      0.4        9,823,524      9,870        995.3        941.0       307        9,908        277.0  cudaStreamIsCapturing_v10000              
+      0.4        9,016,831     18,895        477.2        454.0       307        6,000        104.4  cudaStreamGetCaptureInfo_v2_v11030        
+      0.3        7,557,078      1,943      3,889.4      3,805.0     2,935       11,648        648.9  cudaStreamBeginCapture_v10000             
+      0.3        7,189,458      1,943      3,700.2      3,649.0     2,354       11,574        558.8  cudaGraphDestroy_v10000                   
+      0.1        2,831,657        128     22,122.3      2,001.0     1,222      925,085    116,327.5  cudaStreamCreateWithPriority              
+      0.1        2,563,056      1,943      1,319.1      1,298.0     1,039        7,104        215.9  cudaStreamEndCapture_v10000               
+      0.1        1,525,954      1,943        785.4        713.0       606        8,744        295.8  cudaGraphGetNodes_v10000                  
+      0.1        1,502,447         14    107,317.6      5,457.0     4,455    1,421,402    378,224.4  cudaHostAlloc                             
+      0.0          232,971          8     29,121.4     27,970.5    12,623       64,091     17,478.2  cudaMemGetInfo                            
+      0.0          138,843        810        171.4        142.0        80        1,544        110.5  cuGetProcAddress_v2                       
+      0.0           18,339         16      1,146.2        740.5       378        6,387      1,432.6  cuLibraryGetKernel                        
+      0.0           16,302         19        858.0        353.0       278        4,451      1,090.8  cudaEventCreateWithFlags                  
+      0.0            8,126         14        580.4        559.5       305        1,104        204.3  cudaThreadExchangeStreamCaptureMode_v10010
+      0.0            4,845          1      4,845.0      4,845.0     4,845        4,845          0.0  cudaEventRecord                           
+      0.0            4,537          3      1,512.3      1,306.0     1,229        2,002        425.8  cuInit                                    
+      0.0            3,158          1      3,158.0      3,158.0     3,158        3,158          0.0  cudaStreamWaitEvent                       
+      0.0            2,649          4        662.3        662.5       142        1,182        567.7  cuModuleGetLoadingMode                    
+      0.0            1,385          1      1,385.0      1,385.0     1,385        1,385          0.0  cudaEventDestroy                          
+      0.0            1,056          2        528.0        528.0       260          796        379.0  cudaGetDriverEntryPoint_v11030            
+
+[6/8] Executing 'cuda_gpu_kern_sum' stats report
+
+ Time (%)  Total Time (ns)  Instances   Avg (ns)     Med (ns)    Min (ns)   Max (ns)   StdDev (ns)                                                  Name                                                
+ --------  ---------------  ---------  -----------  -----------  ---------  ---------  -----------  ----------------------------------------------------------------------------------------------------
+     17.7      192,070,994      3,840     50,018.5     22,048.5      7,616    540,899     42,110.2  void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_32x6_tn_align8>(T1::Param…
+     14.5      157,599,907        924    170,562.7    162,689.0     39,968  1,415,303    225,612.7  ampere_bf16_s1688gemm_bf16_128x64_sliced1x2_ldg8_f2f_tn                                             
+     12.7      137,760,865        689    199,943.2     60,640.0     10,688    521,827    229,882.5  void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_16x16_128x2_tn_align8>(T1::Par…
+     12.0      130,433,090      7,336     17,779.9     22,656.0      6,848     25,440      6,404.2  void flash::flash_fwd_splitkv_kernel<Flash_fwd_kernel_traits<(int)128, (int)64, (int)128, (int)4, (…
+      7.2       78,491,480         28  2,803,267.1  2,805,165.5  2,788,142  2,808,782      5,107.2  ampere_bf16_s16816gemm_bf16_128x64_ldg8_f2f_tn                                                      
+      6.4       69,018,876      3,220     21,434.4     21,376.0     21,280     23,456        199.8  void flash::flash_fwd_splitkv_kernel<Flash_fwd_kernel_traits<(int)128, (int)64, (int)128, (int)4, (…
+      3.6       39,090,744          8  4,886,343.0  4,840,279.0  4,794,230  5,090,744    113,470.9  void at_cuda_detail::cub::DeviceSegmentedRadixSortKernel<at_cuda_detail::cub::DeviceRadixSortPolicy…
+      3.0       32,798,268        140    234,273.3    229,890.0    122,272    377,250     83,663.6  ampere_bf16_s1688gemm_bf16_64x128_sliced1x2_ldg8_f2f_tn                                             
+      2.1       22,635,863        784     28,872.3     12,480.0     11,584     62,177     20,637.7  ampere_bf16_s16816gemm_bf16_64x64_ldg8_f2f_stages_64x5_tn                                           
+      2.0       21,990,540      1,960     11,219.7      4,416.5      1,055    461,410     54,161.7  triton_poi_fused_mul_silu_1                                                                         
+      1.9       20,356,642          4  5,089,160.5  5,082,872.5  4,909,239  5,281,658    194,105.4  void at_cuda_detail::cub::DeviceSegmentedRadixSortKernel<at_cuda_detail::cub::DeviceRadixSortPolicy…
+      1.3       14,291,715         28    510,418.4    511,890.5    469,186    513,250      8,099.5  void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<signed char>, std::a…
+      1.2       12,995,135        142     91,515.0     51,536.5     50,401  3,036,431    336,087.1  ampere_bf16_s1688gemm_bf16_128x128_ldg8_f2f_stages_32x1_tn                                          
+      1.0       11,064,209      7,336      1,508.2      1,504.0      1,023      2,304        145.4  void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0…
+      0.9       10,243,687      1,960      5,226.4      3,488.0      1,504    111,552     12,600.6  triton_red_fused__to_copy_add_mean_mul_pow_rsqrt_2                                                  
+      0.9        9,917,616        100     99,176.2      8,896.0      7,008    499,075    172,432.9  std::enable_if<!T7, void>::type internal::gemvx::kernel<int, int, __nv_bfloat16, __nv_bfloat16, __n…
+      0.9        9,765,487          4  2,441,371.8  2,442,108.0  2,393,515  2,487,756     51,684.9  void at::native::<unnamed>::cunn_SoftMaxForward<(int)4, float, float, float, at::native::<unnamed>:…
+      0.8        9,150,993         28    326,821.2    326,786.0    324,833    330,529      1,012.2  ampere_bf16_s1688gemm_bf16_128x128_ldg8_relu_f2f_stages_32x1_tn                                     
+      0.8        8,402,949        224     37,513.2     37,504.0     36,417     38,720        433.0  void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x128_32x6_tn_align8>(T1::Para…
+      0.7        7,801,348          2  3,900,674.0  3,900,674.0  3,710,257  4,091,091    269,290.3  void at::native::_scatter_gather_elementwise_kernel<(int)128, (int)8, void at::native::_cuda_scatte…
+      0.7        7,230,573        336     21,519.6     21,440.0     21,056     22,592        353.9  ampere_bf16_s16816gemm_bf16_128x64_ldg8_relu_f2f_stages_64x3_tn                                     
+      0.6        6,377,065      1,960      3,253.6      2,047.0      1,535     79,073      9,036.2  triton_red_fused__to_copy_add_mean_mul_pow_rsqrt_0                                                  
+      0.6        6,104,261        476     12,824.1     12,800.0     11,648     14,688        734.1  ampere_bf16_s16816gemm_bf16_64x64_ldg8_relu_f2f_stages_64x5_tn                                      
+      0.5        5,896,026          4  1,474,006.5  1,473,894.0  1,473,351  1,474,887        642.5  void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::masked_fill_kernel(at…
+      0.5        5,005,878      3,220      1,554.6      1,537.0      1,471      1,792         48.7  void vllm::merge_attn_states_kernel<__nv_bfloat16, (unsigned int)128>(T1 *, float *, const T1 *, co…
+      0.4        4,533,198        266     17,042.1      6,783.0      5,376    713,379     86,005.8  void at::native::reduce_kernel<(int)512, (int)1, at::native::ReduceOp<float, at::native::ArgMaxOps<…
+      0.4        4,394,579        140     31,389.9     27,296.0     26,272     49,792      8,301.1  ampere_bf16_s1688gemm_bf16_128x64_sliced1x2_ldg8_relu_f2f_tn                                        
+      0.4        4,086,818      1,890      2,162.3      1,824.0      1,344     22,016      2,254.8  triton_poi_fused_cat_3                                                                              
+      0.4        3,998,356          2  1,999,178.0  1,999,178.0  1,997,546  2,000,810      2,308.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::BinaryFunctor<float, float, floa…
+      0.3        3,593,749         56     64,174.1     64,144.0     63,169     65,280        415.2  void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_64x1_tn_align8>(T1::Para…
+      0.3        3,432,593          4    858,148.3    858,260.0    855,589    860,484      2,519.2  void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_nocast<at::n…
+      0.3        3,410,108      4,153        821.1        800.0        767      1,280         73.4  void at::native::vectorized_elementwise_kernel<(int)2, at::native::FillFunctor<long>, std::array<ch…
+      0.3        3,365,160        264     12,746.8      5,088.0      2,111  1,005,636     86,907.4  void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::TensorIterator…
+      0.3        3,200,784          2  1,600,392.0  1,600,392.0  1,565,032  1,635,752     50,006.6  void at::native::tensor_kernel_scan_innermost_dim<float, std::plus<float>>(T1 *, const T1 *, unsign…
+      0.3        2,854,457      1,512      1,887.9      1,728.0      1,311      2,944        453.9  void cublasLt::splitKreduce_kernel<(int)32, (int)16, int, __nv_bfloat16, __nv_bfloat16, float, (boo…
+      0.2        2,697,774      1,890      1,427.4      1,152.0        863     16,704      1,774.4  triton_poi_fused_view_5                                                                             
+      0.2        2,650,361      1,890      1,402.3      1,344.0      1,215      6,688        549.7  triton_poi_fused_cat_4                                                                              
+      0.2        2,583,787          2  1,291,893.5  1,291,893.5  1,291,526  1,292,261        519.7  at::native::<unnamed>::fill_reverse_indices_kernel(long *, int, at::cuda::detail::IntDivider<unsign…
+      0.2        2,581,068          2  1,290,534.0  1,290,534.0  1,289,766  1,291,302      1,086.1  void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_nocast<at::n…
+      0.2        2,423,217        112     21,635.9     21,648.0      9,472     34,464     12,027.3  void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align8>(T1::Par…
+      0.1        1,370,950          2    685,475.0    685,475.0    672,835    698,115     17,875.7  void at::native::<unnamed>::distribution_elementwise_grid_stride_kernel<float, (int)4, void at::nat…
+      0.1        1,299,525        398      3,265.1      3,072.0      2,848      7,488        437.7  void at::native::index_elementwise_kernel<(int)128, (int)4, void at::native::gpu_index_kernel<void …
+      0.1        1,149,285         28     41,045.9     41,024.0     40,288     42,689        427.9  ampere_bf16_s1688gemm_bf16_64x64_sliced1x4_ldg8_f2f_tn                                              
+      0.1          956,389         28     34,156.8     34,767.5     17,888     35,200      3,196.0  std::enable_if<!T7, void>::type internal::gemvx::kernel<int, int, __nv_bfloat16, float, float, floa…
+      0.1          653,123         28     23,325.8     23,263.5     23,009     24,224        280.2  ampere_bf16_s16816gemm_bf16_128x64_ldg8_f2f_stages_32x6_tn                                          
+      0.0          523,203          1    523,203.0    523,203.0    523,203    523,203          0.0  void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_16x16_128x1_tn_align8>(T1::Par…
+      0.0          365,189        262      1,393.9      1,408.0      1,183      1,792        172.1  void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::TensorIterator…
+      0.0          331,960        264      1,257.4      1,248.0      1,120      1,697         54.2  void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::TensorIterator…
+      0.0          295,934        168      1,761.5      1,760.0      1,536      2,048        122.7  void cublasLt::splitKreduce_kernel<(int)32, (int)16, int, __nv_bfloat16, __nv_bfloat16, float, (boo…
+      0.0          258,083         70      3,686.9      3,040.0      1,984     36,192      4,006.4  triton_red_fused__to_copy_add_embedding_mean_mul_pow_rsqrt_0                                        
+      0.0          239,642        262        914.7        896.0        895        993         29.3  void at::native::unrolled_elementwise_kernel<at::native::CUDAFunctorOnSelf_add<int>, std::array<cha…
+      0.0          232,625        262        887.9        865.0        863        960         30.4  void at::native::unrolled_elementwise_kernel<at::native::FillFunctor<int>, std::array<char *, (unsi…
+      0.0          192,230         70      2,746.1      2,112.0      1,632     40,192      4,554.3  triton_poi_fused_cat_1                                                                              
+      0.0          157,665          1    157,665.0    157,665.0    157,665    157,665          0.0  void at::native::<unnamed>::CatArrayBatchedCopy_aligned16_contig<at::native::<unnamed>::OpaqueType<…
+      0.0          143,357        159        901.6        896.0        863        992         29.4  void at::native::vectorized_elementwise_kernel<(int)2, at::native::FillFunctor<int>, std::array<cha…
+      0.0          100,929         70      1,441.8      1,344.0      1,216      8,928        910.7  triton_poi_fused_cat_2                                                                              
+      0.0           99,267         70      1,418.1      1,184.5        864     13,728      1,523.5  triton_poi_fused_view_3                                                                             
+      0.0           95,967        107        896.9        896.0        863      1,824         94.4  void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<int>, std::array<cha…
+      0.0           79,489          1     79,489.0     79,489.0     79,489     79,489          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::bfloat16_copy_kernel_cuda(at::Te…
+      0.0           64,187         58      1,106.7        896.0        865     11,328      1,367.7  void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<c10::BFloat16>, std:…
+      0.0           43,840          1     43,840.0     43,840.0     43,840     43,840          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::sin_kernel_cuda(at::TensorIterat…
+      0.0           36,673         28      1,309.8      1,312.0      1,280      1,345         15.0  void cublasLt::splitKreduce_kernel<(int)32, (int)16, int, float, __nv_bfloat16, float, (bool)0, __n…
+      0.0           26,400          1     26,400.0     26,400.0     26,400     26,400          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::cos_kernel_cuda(at::TensorIterat…
+      0.0           19,360          1     19,360.0     19,360.0     19,360     19,360          0.0  void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_nocast<at::n…
+      0.0           11,484         11      1,044.0        895.0        767      1,567        287.4  void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<float>, std::array<c…
+      0.0           10,848          2      5,424.0      5,424.0      5,216      5,632        294.2  void at::native::_scatter_gather_elementwise_kernel<(int)128, (int)8, void at::native::_cuda_scatte…
+      0.0            8,705          2      4,352.5      4,352.5      4,160      4,545        272.2  void at::native::<unnamed>::distribution_elementwise_grid_stride_kernel<float, (int)4, void at::nat…
+      0.0            3,648          2      1,824.0      1,824.0      1,632      2,016        271.5  void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_nocast<at::n…
+      0.0            3,455          2      1,727.5      1,727.5      1,695      1,760         46.0  void at::native::vectorized_elementwise_kernel<(int)2, at::native::CUDAFunctorOnOther_add<long>, st…
+      0.0            3,135          2      1,567.5      1,567.5      1,503      1,632         91.2  void at::native::vectorized_elementwise_kernel<(int)2, at::native::<unnamed>::where_kernel_impl(at:…
+      0.0            3,104          2      1,552.0      1,552.0      1,344      1,760        294.2  void at::native::vectorized_elementwise_kernel<(int)4, void at::native::compare_scalar_kernel<float…
+      0.0            3,040          2      1,520.0      1,520.0      1,440      1,600        113.1  void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnOther_add<float>, s…
+      0.0            2,975          2      1,487.5      1,487.5        992      1,983        700.7  void <unnamed>::elementwise_kernel_with_index<int, at::native::arange_cuda_out(const c10::Scalar &,…
+      0.0            2,336          1      2,336.0      2,336.0      2,336      2,336          0.0  void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl<at::native::…
+      0.0            1,184          1      1,184.0      1,184.0      1,184      1,184          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::reciprocal_kernel_cuda(at::Tenso…
+      0.0            1,024          1      1,024.0      1,024.0      1,024      1,024          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, float, floa…
+      0.0            1,024          1      1,024.0      1,024.0      1,024      1,024          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::BUnaryFunctor<float, float, floa…
+      0.0              928          1        928.0        928.0        928        928          0.0  void at::native::vectorized_elementwise_kernel<(int)2, at::native::FillFunctor<double>, std::array<…
+
+[7/8] Executing 'cuda_gpu_mem_time_sum' stats report
+
+ Time (%)  Total Time (ns)  Count   Avg (ns)     Med (ns)    Min (ns)    Max (ns)    StdDev (ns)            Operation           
+ --------  ---------------  -----  -----------  -----------  ---------  -----------  -----------  ------------------------------
+     98.7      554,163,217  2,587    214,210.8        352.0        288  100,447,087  2,156,381.0  [CUDA memcpy Host-to-Device]  
+      1.0        5,437,594      4  1,359,398.5  1,359,767.0  1,356,454    1,361,606      2,174.7  [CUDA memcpy Device-to-Device]
+      0.3        1,567,258  2,670        587.0        480.0        288        2,240        244.5  [CUDA memset]                 
+      0.1          296,577    262      1,132.0      1,120.0        864        1,536         80.4  [CUDA memcpy Device-to-Host]  
+
+[8/8] Executing 'cuda_gpu_mem_size_sum' stats report
+
+ Total (MB)  Count  Avg (MB)  Med (MB)  Min (MB)  Max (MB)  StdDev (MB)            Operation           
+ ----------  -----  --------  --------  --------  --------  -----------  ------------------------------
+  3,090.501  2,587     1.195     0.001     0.000   466.747       10.389  [CUDA memcpy Host-to-Device]  
+  2,489.319      4   622.330   622.330   622.330   622.330        0.000  [CUDA memcpy Device-to-Device]
+      2.029  2,670     0.001     0.001     0.000     0.006        0.001  [CUDA memset]                 
+      0.008    262     0.000     0.000     0.000     0.000        0.000  [CUDA memcpy Device-to-Host]  
+
+Generated:
+    /data/cy/kv_cache_vs_util/sim_traverse_bs/traverse_bs_util_sim_prefill.nsys-rep
+    /data/cy/kv_cache_vs_util/sim_traverse_bs/traverse_bs_util_sim_prefill.sqlite

sim_traverse_bs/traverse_bs_util_sim_prefill.nsys-rep

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dd4985005949c5e5485a73c83659883635c2cdf0895cadd4deb604d1bc89a15b
+size 17188445

sim_traverse_bs/traverse_bs_util_sim_prefill.py

@@ -0,0 +1,268 @@
+import os
+import time
+import statistics
+from typing import List, Tuple, Dict
+
+import torch
+import torch.cuda.nvtx as nvtx
+
+from vllm import LLM, SamplingParams
+from transformers import AutoTokenizer
+
+# ========= 强制使用 vLLM V1 =========
+os.environ.setdefault("VLLM_USE_V1", "1")
+os.environ.setdefault("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
+
+# 可选：打开 V1 metrics 统计
+os.environ.setdefault("VLLM_LOGGING_LEVEL", "INFO")
+
+# ========= 试图导入 V1 metrics 类型（兼容不同版本）=========
+try:
+    from vllm.v1.metrics.reader import Counter, Gauge, Histogram, Vector  # type: ignore
+except Exception:
+    Counter = Gauge = Histogram = Vector = type("X", (), {})  # dummy
+
+# ========= 配置 =========
+MODEL_NAME = "Qwen/Qwen2-1.5B"
+DTYPE = "bfloat16"
+TP = 1
+GPU_MEM_UTIL = 0.90
+TRUST_REMOTE_CODE = True
+
+# 场景：prefill=输入tokens，decode=输出tokens
+SCENARIOS = [
+    {"name": "prefill640_decode1", "prompt_tokens": 1152, "max_new_tokens": 1},
+    # {"name": "prefill1_decode512", "prompt_tokens": 1,   "max_new_tokens": 512},
+    # {"name": "prefill640_decode512", "prompt_tokens": 640, "max_new_tokens": 512},
+]
+
+BATCH_SIZES = [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024]
+# BATCH_SIZES = [1, 1024]
+
+SEED = 1234
+TEMPERATURE = 0.0
+TOP_P = 1.0
+WARMUP_PER_BS = 1  # 每个批次做一次预热
+
+# ========= 构造“精确 token 数量”的 prompt =========
+def build_exact_token_prompt(tokenizer, target_len: int) -> str:
+    if target_len <= 1:
+        # 最小化 prompt：用一个简单 token（避免空串导致0 token）
+        ids = tokenizer("A", add_special_tokens=False)["input_ids"]
+        if len(ids) >= 1:
+            return tokenizer.decode(ids[:1], skip_special_tokens=True, clean_up_tokenization_spaces=False)
+
+    base_text = (
+        "You are a helpful assistant. "
+        "Please analyze the following input and respond succinctly. "
+    )
+    chunk = " ".join(["data"] * 100) + ". "
+    text = base_text + chunk * 200  # 足够长的文本
+
+    lo, hi = 0, len(text)
+    target_ids = None
+    while lo <= hi:
+        mid = (lo + hi) // 2
+        ids = tokenizer(text[:mid], add_special_tokens=False)["input_ids"]
+        if len(ids) == target_len:
+            target_ids = ids
+            break
+        if len(ids) < target_len:
+            lo = mid + 1
+        else:
+            hi = mid - 1
+
+    if target_ids is None:
+        ids = tokenizer(text[:lo], add_special_tokens=False)["input_ids"]
+        if len(ids) > target_len:
+            target_ids = ids[:target_len]
+        else:
+            filler = " data"
+            while len(ids) < target_len:
+                ids = tokenizer(tokenizer.decode(ids) + filler, add_special_tokens=False)["input_ids"]
+            target_ids = ids[:target_len]
+
+    prompt = tokenizer.decode(target_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)
+    # 断言精确长度
+    assert len(tokenizer(prompt, add_special_tokens=False)["input_ids"]) == target_len
+    return prompt
+
+# ========= V1 metrics 抽取工具 =========
+TTFT_METRIC_NAME = "vllm:time_to_first_token_seconds"
+TPOT_METRIC_NAME = "vllm:time_per_output_token_seconds"  # per-output-token latency
+
+def _iter_children_of_vector(vec_obj):
+    for attr in ("children", "metrics", "series", "values", "samples", "items"):
+        if hasattr(vec_obj, attr):
+            val = getattr(vec_obj, attr)
+            if isinstance(val, dict):
+                for v in val.values():
+                    yield v
+            else:
+                try:
+                    for v in val:
+                        yield v
+                except TypeError:
+                    pass
+
+def _collect_hist_sum_count(metrics, metric_name: str):
+    total_sum = 0.0
+    total_count = 0.0
+    for m in metrics:
+        mname = getattr(m, "name", None)
+        if mname != metric_name:
+            continue
+        # 直接 Histogram
+        if isinstance(m, Histogram) or m.__class__.__name__ == "Histogram":
+            total_sum += float(getattr(m, "sum", 0.0))
+            total_count += float(getattr(m, "count", 0.0))
+            continue
+        # Vector[Histogram]
+        if isinstance(m, Vector) or m.__class__.__name__ == "Vector":
+            for child in _iter_children_of_vector(m):
+                if isinstance(child, Histogram) or child.__class__.__name__ == "Histogram":
+                    total_sum += float(getattr(child, "sum", 0.0))
+                    total_count += float(getattr(child, "count", 0.0))
+    return total_sum, total_count
+
+def _metrics_snapshot(llm) -> Dict[str, float]:
+    try:
+        mets = llm.get_metrics()  # V1: 返回 Metric 列表（包含 Histogram/Vector 等）
+    except Exception:
+        return {"ttft_sum": 0.0, "ttft_cnt": 0.0, "tpot_sum": 0.0, "tpot_cnt": 0.0}
+    ttft_sum, ttft_cnt = _collect_hist_sum_count(mets, TTFT_METRIC_NAME)
+    tpot_sum, tpot_cnt = _collect_hist_sum_count(mets, TPOT_METRIC_NAME)
+    return {"ttft_sum": ttft_sum, "ttft_cnt": ttft_cnt, "tpot_sum": tpot_sum, "tpot_cnt": tpot_cnt}
+
+def _metrics_delta(before: dict, after: dict):
+    return {
+        "ttft_sum": after["ttft_sum"] - before["ttft_sum"],
+        "ttft_cnt": after["ttft_cnt"] - before["ttft_cnt"],
+        "tpot_sum": after["tpot_sum"] - before["tpot_sum"],
+        "tpot_cnt": after["tpot_cnt"] - before["tpot_cnt"],
+    }
+
+# ========= 带 NVTX 的 generate 包装 =========
+def decorated_generate(llm: LLM, prompts: List[str], params: SamplingParams):
+    return llm.generate(prompts, params)
+
+# ========= 统计格式化 =========
+def fmt_stats(x: List[float]) -> Tuple[float, float, float]:
+    xs = [v for v in x if (v == v)]  # 过滤 NaN
+    if not xs:
+        return (float("nan"), float("nan"), float("nan"))
+    return (statistics.mean(xs), statistics.median(xs), statistics.quantiles(xs, n=10)[-1])  # p90
+
+def main():
+    print("--- vLLM V1 基准测试（含 NVTX 标记）---")
+    print(f"模型: {MODEL_NAME}")
+    print(f"批量大小: {BATCH_SIZES}")
+    print(f"场景: {[s['name'] for s in SCENARIOS]}")
+    print("-" * 60)
+
+    if not torch.cuda.is_available():
+        print("错误：需要 CUDA GPU。")
+        return
+
+    print("加载分词器/模型中...")
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, use_fast=True, trust_remote_code=TRUST_REMOTE_CODE)
+
+    # 用 NVTX 标记模型加载阶段
+    nvtx.range_push("LLM_init")
+    llm = LLM(
+        model=MODEL_NAME,
+        tensor_parallel_size=TP,
+        dtype=DTYPE,
+        trust_remote_code=TRUST_REMOTE_CODE,
+        gpu_memory_utilization=GPU_MEM_UTIL,
+        max_num_seqs=1024,         # 足够覆盖本次扫描
+        max_model_len=4096,
+        disable_log_stats=False,  # 开启 V1 metrics 收集
+    )
+    nvtx.range_pop()
+    print("模型加载完成。")
+
+    for sc in SCENARIOS:
+        name = sc["name"]
+        prompt_tokens = sc["prompt_tokens"]
+        max_new_tokens = sc["max_new_tokens"]
+
+        print(f"\n===== 场景：{name} | prefill={prompt_tokens}, decode={max_new_tokens} =====")
+
+        # 准备精确长度 prompt
+        prompt_text = build_exact_token_prompt(tokenizer, prompt_tokens)
+
+        # 采样参数（贪心）
+        sampling_params = SamplingParams(
+            max_tokens=max_new_tokens,
+            temperature=TEMPERATURE,
+            top_p=TOP_P,
+            seed=SEED,
+            n=1,
+        )
+
+        # 记录每个 bs 的结果（便于后续统计或外部解析）
+        for bs in BATCH_SIZES:
+            print(f"\n--- 批量大小 bs={bs} ---")
+
+            prompts = [prompt_text] * bs
+
+            # 预热
+            # print("预热中...")
+            # nvtx.range_push(f"WARMUP [{name}] bs={bs}")
+            # _ = decorated_generate(llm, [prompts[0]], sampling_params)
+            # torch.cuda.synchronize()
+            # nvtx.range_pop()
+            if bs== 1:
+                print("预热中...")
+                for _ in range(WARMUP_PER_BS):
+                    _ = decorated_generate(llm, [prompts[0]], sampling_params)
+                torch.cuda.synchronize()
+
+            # 正式计时与 V1 metrics
+            # nvtx.range_push(f"RUN [{name}] bs={bs}")
+            torch.cuda.synchronize()
+            snap_before = _metrics_snapshot(llm)
+            t0 = time.perf_counter()
+
+            nvtx.range_push(f"generate [{name}] bs={bs}")
+            outputs = decorated_generate(llm, prompts, sampling_params)
+            nvtx.range_pop()  # generate
+
+            torch.cuda.synchronize()
+            t1 = time.perf_counter()
+            snap_after = _metrics_snapshot(llm)
+            # nvtx.range_pop()  # RUN
+
+            duration = t1 - t0
+
+            # 统计 token 与吞吐
+            total_output_tokens = sum(len(o.outputs[0].token_ids) for o in outputs)
+            avg_prompt_tokens = sum(len(o.prompt_token_ids) for o in outputs) / bs
+            throughput = total_output_tokens / duration if duration > 0 else float("inf")
+
+            # 解析 V1 TTFT / 解码吞吐
+            delta = _metrics_delta(snap_before, snap_after)
+            if delta["ttft_cnt"] > 0:
+                ttft = delta["ttft_sum"] / delta["ttft_cnt"]
+            else:
+                ttft = float("nan")
+
+            if delta["tpot_cnt"] > 0:
+                avg_tpot = delta["tpot_sum"] / delta["tpot_cnt"]  # seconds/token
+                decode_tps = 1.0 / avg_tpot
+            else:
+                decode_tps = float("nan")
+
+            print(f"执行时间: {duration:.4f} s")
+            print(f"实际平均输入 tokens: {avg_prompt_tokens:.2f}（目标 {prompt_tokens}）")
+            print(f"生成总 tokens: {total_output_tokens}")
+            print(f"吞吐(生成tokens/秒): {throughput:.2f}")
+            print(f"TTFT (V1 metrics): {ttft:.4f} s")
+            print(f"解码吞吐 (V1 metrics): {decode_tps:.2f} tok/s")
+
+    print("\n完成。提示：在 Nsight Systems 中可通过 NVTX 区间快速定位各场景/批量的调用。")
+
+if __name__ == "__main__":
+    print(f"CUDA_VISIBLE_DEVICES = {os.getenv('CUDA_VISIBLE_DEVICES')}")
+    main()

sim_traverse_bs/traverse_bs_util_sim_prefill_1152.log

@@ -0,0 +1,381 @@
+WARNING: CPU IP/backtrace sampling not supported, disabling.
+Try the 'nsys status --environment' command to learn more.
+
+WARNING: CPU context switch tracing not supported, disabling.
+Try the 'nsys status --environment' command to learn more.
+
+INFO 08-14 10:12:47 [__init__.py:235] Automatically detected platform cuda.
+CUDA_VISIBLE_DEVICES = 3
+--- vLLM V1 基准测试（含 NVTX 标记）---
+模型: Qwen/Qwen2-1.5B
+批量大小: [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024]
+场景: ['prefill640_decode1']
+------------------------------------------------------------
+加载分词器/模型中...
+INFO 08-14 10:12:58 [config.py:1604] Using max model len 4096
+INFO 08-14 10:12:58 [config.py:2434] Chunked prefill is enabled with max_num_batched_tokens=8192.
+INFO 08-14 10:13:05 [__init__.py:235] Automatically detected platform cuda.
+INFO 08-14 10:13:07 [core.py:572] Waiting for init message from front-end.
+INFO 08-14 10:13:07 [core.py:71] Initializing a V1 LLM engine (v0.10.0) with config: model='Qwen/Qwen2-1.5B', speculative_config=None, tokenizer='Qwen/Qwen2-1.5B', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config={}, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=4096, download_dir=None, load_format=LoadFormat.AUTO, tensor_parallel_size=1, pipeline_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=False, kv_cache_dtype=auto,  device_config=cuda, decoding_config=DecodingConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_backend=''), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None), seed=0, served_model_name=Qwen/Qwen2-1.5B, num_scheduler_steps=1, multi_step_stream_outputs=True, enable_prefix_caching=True, chunked_prefill_enabled=True, use_async_output_proc=True, pooler_config=None, compilation_config={"level":3,"debug_dump_path":"","cache_dir":"","backend":"","custom_ops":[],"splitting_ops":["vllm.unified_attention","vllm.unified_attention_with_output","vllm.mamba_mixer2"],"use_inductor":true,"compile_sizes":[],"inductor_compile_config":{"enable_auto_functionalized_v2":false},"inductor_passes":{},"use_cudagraph":true,"cudagraph_num_of_warmups":1,"cudagraph_capture_sizes":[512,504,496,488,480,472,464,456,448,440,432,424,416,408,400,392,384,376,368,360,352,344,336,328,320,312,304,296,288,280,272,264,256,248,240,232,224,216,208,200,192,184,176,168,160,152,144,136,128,120,112,104,96,88,80,72,64,56,48,40,32,24,16,8,4,2,1],"cudagraph_copy_inputs":false,"full_cuda_graph":false,"max_capture_size":512,"local_cache_dir":null}
+INFO 08-14 10:13:14 [parallel_state.py:1102] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, TP rank 0, EP rank 0
+WARNING 08-14 10:13:14 [topk_topp_sampler.py:59] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
+INFO 08-14 10:13:14 [gpu_model_runner.py:1843] Starting to load model Qwen/Qwen2-1.5B...
+INFO 08-14 10:13:15 [gpu_model_runner.py:1875] Loading model from scratch...
+INFO 08-14 10:13:15 [cuda.py:290] Using Flash Attention backend on V1 engine.
+INFO 08-14 10:13:15 [weight_utils.py:296] Using model weights format ['*.safetensors']
+INFO 08-14 10:13:16 [weight_utils.py:349] No model.safetensors.index.json found in remote.
+
+
+
+
+INFO 08-14 10:13:17 [default_loader.py:262] Loading weights took 0.66 seconds
+INFO 08-14 10:13:17 [gpu_model_runner.py:1892] Model loading took 2.9105 GiB and 1.885122 seconds
+INFO 08-14 10:13:23 [backends.py:530] Using cache directory: /home/cy/.cache/vllm/torch_compile_cache/40b61c71e9/rank_0_0/backbone for vLLM's torch.compile
+INFO 08-14 10:13:23 [backends.py:541] Dynamo bytecode transform time: 6.01 s
+INFO 08-14 10:13:28 [backends.py:161] Directly load the compiled graph(s) for dynamic shape from the cache, took 4.662 s
+INFO 08-14 10:13:29 [monitor.py:34] torch.compile takes 6.01 s in total
+INFO 08-14 10:13:30 [gpu_worker.py:255] Available KV cache memory: 12.81 GiB
+INFO 08-14 10:13:30 [kv_cache_utils.py:833] GPU KV cache size: 479,536 tokens
+INFO 08-14 10:13:30 [kv_cache_utils.py:837] Maximum concurrency for 4,096 tokens per request: 117.07x
+
+INFO 08-14 10:13:33 [gpu_model_runner.py:2485] Graph capturing finished in 2 secs, took 0.49 GiB
+INFO 08-14 10:13:33 [core.py:193] init engine (profile, create kv cache, warmup model) took 15.81 seconds
+模型加载完成。
+
+===== 场景：prefill640_decode1 | prefill=1152, decode=1 =====
+
+--- 批量大小 bs=1 ---
+预热中...
+
+
+
+
+执行时间: 0.0159 s
+实际平均输入 tokens: 1152.00（目标 1152）
+生成总 tokens: 1
+吞吐(生成tokens/秒): 62.80
+TTFT (V1 metrics): 0.0147 s
+解码吞吐 (V1 metrics): nan tok/s
+
+--- 批量大小 bs=2 ---
+
+
+执行时间: 0.0213 s
+实际平均输入 tokens: 1152.00（目标 1152）
+生成总 tokens: 2
+吞吐(生成tokens/秒): 94.11
+TTFT (V1 metrics): 0.0155 s
+解码吞吐 (V1 metrics): nan tok/s
+
+--- 批量大小 bs=4 ---
+
+
+执行时间: 0.0305 s
+实际平均输入 tokens: 1152.00（目标 1152）
+生成总 tokens: 4
+吞吐(生成tokens/秒): 131.32
+TTFT (V1 metrics): 0.0172 s
+解码吞吐 (V1 metrics): nan tok/s
+
+--- 批量大小 bs=8 ---
+
+
+执行时间: 0.0400 s
+实际平均输入 tokens: 1152.00（目标 1152）
+生成总 tokens: 8
+吞吐(生成tokens/秒): 199.98
+TTFT (V1 metrics): 0.0204 s
+解码吞吐 (V1 metrics): nan tok/s
+
+--- 批量大小 bs=16 ---
+
+
+执行时间: 0.0617 s
+实际平均输入 tokens: 1152.00（目标 1152）
+生成总 tokens: 16
+吞吐(生成tokens/秒): 259.19
+TTFT (V1 metrics): 0.0285 s
+解码吞吐 (V1 metrics): nan tok/s
+
+--- 批量大小 bs=32 ---
+
+
+执行时间: 0.1109 s
+实际平均输入 tokens: 1152.00（目标 1152）
+生成总 tokens: 32
+吞吐(生成tokens/秒): 288.55
+TTFT (V1 metrics): 0.0514 s
+解码吞吐 (V1 metrics): nan tok/s
+
+--- 批量大小 bs=64 ---
+
+
+执行时间: 0.2001 s
+实际平均输入 tokens: 1152.00（目标 1152）
+生成总 tokens: 64
+吞吐(生成tokens/秒): 319.77
+TTFT (V1 metrics): 0.0997 s
+解码吞吐 (V1 metrics): nan tok/s
+
+--- 批量大小 bs=128 ---
+
+
+执行时间: 0.3899 s
+实际平均输入 tokens: 1152.00（目标 1152）
+生成总 tokens: 128
+吞吐(生成tokens/秒): 328.30
+TTFT (V1 metrics): 0.1963 s
+解码吞吐 (V1 metrics): nan tok/s
+
+--- 批量大小 bs=256 ---
+
+
+执行时间: 1.0404 s
+实际平均输入 tokens: 1152.00（目标 1152）
+生成总 tokens: 256
+吞吐(生成tokens/秒): 246.06
+TTFT (V1 metrics): 0.6454 s
+解码吞吐 (V1 metrics): nan tok/s
+
+--- 批量大小 bs=512 ---
+
+
+执行时间: 1.4178 s
+实际平均输入 tokens: 1152.00（目标 1152）
+生成总 tokens: 512
+吞吐(生成tokens/秒): 361.12
+TTFT (V1 metrics): 0.6722 s
+解码吞吐 (V1 metrics): nan tok/s
+
+--- 批量大小 bs=1024 ---
+
+
+[rank0]:[W814 10:13:41.589921824 ProcessGroupNCCL.cpp:1479] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
+执行时间: 2.4235 s
+实际平均输入 tokens: 1152.00（目标 1152）
+生成总 tokens: 1024
+吞吐(生成tokens/秒): 422.54
+TTFT (V1 metrics): 1.2129 s
+解码吞吐 (V1 metrics): nan tok/s
+
+完成。提示：在 Nsight Systems 中可通过 NVTX 区间快速定位各场景/批量的调用。
+GPU 3: General Metrics for NVIDIA AD10x (any frequency)
+Generating '/tmp/nsys-report-18f3.qdstrm'
+
+
+[3/8] Executing 'nvtx_sum' stats report
+
+ Time (%)  Total Time (ns)  Instances      Avg (ns)          Med (ns)         Min (ns)        Max (ns)     StdDev (ns)   Style                   Range                 
+ --------  ---------------  ---------  ----------------  ----------------  --------------  --------------  -----------  -------  --------------------------------------
+     88.5   44,070,406,690          1  44,070,406,690.0  44,070,406,690.0  44,070,406,690  44,070,406,690          0.0  PushPop  :LLM_init                             
+      4.9    2,423,366,261          1   2,423,366,261.0   2,423,366,261.0   2,423,366,261   2,423,366,261          0.0  PushPop  :generate [prefill640_decode1] bs=1024
+      2.8    1,417,657,581          1   1,417,657,581.0   1,417,657,581.0   1,417,657,581   1,417,657,581          0.0  PushPop  :generate [prefill640_decode1] bs=512 
+      2.1    1,040,325,509          1   1,040,325,509.0   1,040,325,509.0   1,040,325,509   1,040,325,509          0.0  PushPop  :generate [prefill640_decode1] bs=256 
+      0.8      389,808,438          1     389,808,438.0     389,808,438.0     389,808,438     389,808,438          0.0  PushPop  :generate [prefill640_decode1] bs=128 
+      0.4      200,051,514          1     200,051,514.0     200,051,514.0     200,051,514     200,051,514          0.0  PushPop  :generate [prefill640_decode1] bs=64  
+      0.2      110,723,882          1     110,723,882.0     110,723,882.0     110,723,882     110,723,882          0.0  PushPop  :generate [prefill640_decode1] bs=32  
+      0.1       61,663,721          1      61,663,721.0      61,663,721.0      61,663,721      61,663,721          0.0  PushPop  :generate [prefill640_decode1] bs=16  
+      0.1       39,934,013          1      39,934,013.0      39,934,013.0      39,934,013      39,934,013          0.0  PushPop  :generate [prefill640_decode1] bs=8   
+      0.1       30,382,401          1      30,382,401.0      30,382,401.0      30,382,401      30,382,401          0.0  PushPop  :generate [prefill640_decode1] bs=4   
+      0.0       21,185,162          1      21,185,162.0      21,185,162.0      21,185,162      21,185,162          0.0  PushPop  :generate [prefill640_decode1] bs=2   
+      0.0       15,818,426          1      15,818,426.0      15,818,426.0      15,818,426      15,818,426          0.0  PushPop  :generate [prefill640_decode1] bs=1   
+      0.0          104,750          2          52,375.0          52,375.0          47,675          57,075      6,646.8  PushPop  CCCL:cub::DeviceSegmentedRadixSort    
+
+[4/8] Executing 'osrt_sum' stats report
+
+ Time (%)  Total Time (ns)  Num Calls      Avg (ns)        Med (ns)    Min (ns)      Max (ns)       StdDev (ns)              Name         
+ --------  ---------------  ---------  ----------------  ------------  ---------  --------------  ----------------  ----------------------
+     30.5  361,093,251,838     18,099      19,951,005.7      22,399.0      1,000  26,543,007,832     444,078,095.4  pthread_cond_timedwait
+     27.4  324,888,102,277     20,985      15,481,920.5  10,061,722.0      1,004  34,595,644,343     416,535,555.9  epoll_wait            
+     22.4  265,117,572,216         21  12,624,646,296.0   9,618,795.0     18,952  26,544,576,053  13,564,859,509.6  pthread_cond_wait     
+      8.7  103,158,712,529     12,027       8,577,260.5       1,332.0      1,000  10,010,048,294     168,486,764.8  poll                  
+      6.3   74,091,760,041         28   2,646,134,287.2     172,477.0     77,298  10,000,094,040   4,119,108,342.9  sem_timedwait         
+      4.0   46,893,474,571     32,679       1,434,972.8       3,126.0      1,000  24,724,048,183     164,795,577.6  read                  
+      0.6    6,727,161,428        745       9,029,746.9   6,344,952.0     26,533     529,046,356      30,627,414.0  sem_wait              
+      0.0      480,836,474    193,689           2,482.5       1,588.0      1,001      75,149,274         170,780.2  munmap                
+      0.0      328,343,219      8,501          38,624.1      10,820.0      1,002      30,644,743         467,164.9  ioctl                 
+      0.0      251,356,732        370         679,342.5       2,868.5      1,122      19,761,752       3,411,160.1  fopen                 
+      0.0      121,530,743         24       5,063,781.0   5,062,739.5  5,054,831       5,078,176           7,766.8  nanosleep             
+      0.0      102,831,793     30,682           3,351.5       2,838.0      1,021         121,668           2,147.4  open64                
+      0.0       88,351,007     16,056           5,502.7       2,407.0      1,000      15,130,665         165,401.9  mmap64                
+      0.0       82,243,111         49       1,678,430.8       4,835.0      1,004      75,416,346      10,759,062.2  waitpid               
+      0.0       72,255,858         96         752,665.2       3,804.0      1,029      19,707,249       3,616,782.1  open                  
+      0.0       68,970,368         38       1,815,009.7     659,953.5      2,972      10,358,269       3,351,149.7  pthread_join          
+      0.0       56,563,497         10       5,656,349.7      22,802.5     10,900      56,234,234      17,771,314.4  connect               
+      0.0       30,201,734      2,482          12,168.3       8,512.0      1,026         424,091          16,296.8  pthread_mutex_lock    
+      0.0       29,341,923     15,018           1,953.8       1,793.0      1,000          71,749           1,783.3  pthread_cond_signal   
+      0.0       22,096,068      2,803           7,883.0       4,886.0      1,024       3,311,062          64,777.7  recv                  
+      0.0       17,172,528      2,802           6,128.7       5,160.0      1,857          89,996           6,209.7  send                  
+      0.0       16,852,837      5,836           2,887.7       2,233.0      1,006          79,284           4,761.4  write                 
+      0.0        9,855,547        147          67,044.5      68,806.0     55,639          86,945           4,938.1  sleep                 
+      0.0        8,928,549        131          68,156.9      65,190.0     19,928         172,451          29,775.9  pthread_create        
+      0.0        7,702,498        910           8,464.3       5,459.5      1,005          92,618          10,147.4  fgets                 
+      0.0        4,416,671      3,034           1,455.7       1,368.0      1,000          11,404             539.8  epoll_ctl             
+      0.0        2,829,014         20         141,450.7     186,919.5      9,591         225,750          82,026.0  pthread_rwlock_wrlock 
+      0.0        2,644,265        344           7,686.8       7,272.0      1,996          38,686           2,849.6  fopen64               
+      0.0        1,848,847         66          28,012.8       5,390.0      1,156         480,098          75,714.8  futex                 
+      0.0        1,696,227      1,286           1,319.0       1,127.0      1,000           9,841             674.0  fclose                
+      0.0        1,083,206        197           5,498.5       3,606.0      1,087          74,701           8,142.4  mmap                  
+      0.0        1,076,556         10         107,655.6     127,469.5     19,259         199,656          71,478.9  pthread_rwlock_rdlock 
+      0.0          806,160          1         806,160.0     806,160.0    806,160         806,160               0.0  fork                  
+      0.0          388,292         65           5,973.7       5,152.0      2,419          14,760           3,003.5  pipe2                 
+      0.0          314,102         41           7,661.0       7,552.0      1,994          18,407           4,501.9  socket                
+      0.0          240,921         24          10,038.4       3,584.5      1,030          69,128          17,724.3  bind                  
+      0.0          144,657         34           4,254.6       3,011.5      1,351          16,000           2,941.9  pthread_cond_broadcast
+      0.0           86,266          7          12,323.7      13,530.0      3,460          23,269           7,320.6  fread                 
+      0.0           60,319          5          12,063.8       8,575.0      3,768          23,707           7,966.6  accept4               
+      0.0           56,413         37           1,524.7       1,289.0      1,013           4,207             757.9  fcntl                 
+      0.0           46,861         15           3,124.1       2,439.0      1,515           7,179           1,660.3  stat                  
+      0.0           44,352         25           1,774.1       1,871.0      1,125           2,466             387.1  sigaction             
+      0.0           33,908         16           2,119.3       2,269.0      1,010           3,538             822.4  dup2                  
+      0.0           26,126         10           2,612.6       2,300.0      1,008           7,345           1,792.1  pread                 
+      0.0           24,947          8           3,118.4       2,937.5      1,040           5,596           2,129.1  fflush                
+      0.0           23,097          5           4,619.4       4,103.0      2,090           8,551           2,725.9  fwrite                
+      0.0           22,454          4           5,613.5       5,204.5      3,900           8,145           1,805.3  lstat                 
+      0.0           21,962          4           5,490.5       5,141.0      3,524           8,156           1,995.5  flock                 
+      0.0           20,689         11           1,880.8       1,816.0      1,426           2,769             414.2  listen                
+      0.0           13,565          3           4,521.7       4,506.0      4,310           4,749             219.9  fputs_unlocked        
+      0.0           11,536          5           2,307.2       2,273.0      1,638           3,118             528.8  mprotect              
+      0.0            8,918          5           1,783.6       1,795.0      1,003           3,233             906.7  fstat                 
+      0.0            6,379          1           6,379.0       6,379.0      6,379           6,379               0.0  kill                  
+      0.0            5,888          3           1,962.7       1,980.0      1,768           2,140             186.6  flockfile             
+      0.0            3,752          1           3,752.0       3,752.0      3,752           3,752               0.0  fputs                 
+      0.0            3,389          2           1,694.5       1,694.5      1,296           2,093             563.6  openat64              
+
+[5/8] Executing 'cuda_api_sum' stats report
+
+ Time (%)  Total Time (ns)  Num Calls   Avg (ns)     Med (ns)    Min (ns)   Max (ns)    StdDev (ns)                     Name                   
+ --------  ---------------  ---------  -----------  -----------  --------  -----------  -----------  ------------------------------------------
+     43.7    1,486,144,372      5,546    267,966.9      5,405.5     3,421  119,285,070  1,741,913.5  cudaMemcpyAsync                           
+     20.3      691,134,943     58,898     11,734.4      6,501.0       881   65,983,554    383,241.3  cudaLaunchKernel                          
+      8.2      277,499,678     20,155     13,768.3     13,012.0     9,102      240,985      4,245.9  cudaGraphLaunch_v10000                    
+      6.9      235,333,202      1,943    121,118.5     77,523.0    42,429    1,573,158    204,447.3  cudaGraphInstantiateWithFlags_v11040      
+      5.5      187,977,096      2,136     88,004.3     30,687.0     4,644  122,466,605  2,649,205.6  cudaDeviceSynchronize                     
+      5.5      186,726,680      1,366    136,696.0      3,126.5     2,622  143,581,903  3,899,107.7  cudaStreamSynchronize                     
+      3.4      114,031,743     33,151      3,439.8      4,167.0       689      147,601      2,360.2  cuLaunchKernel                            
+      1.6       55,915,195        222    251,870.2    119,564.5    59,545    2,410,123    388,773.3  cudaFree                                  
+      1.2       41,787,784        348    120,079.8    112,828.0     8,334    1,053,627     57,682.4  cudaMalloc                                
+      0.8       28,012,374         10  2,801,237.4  2,938,591.5    56,912    5,017,452  1,646,078.4  cuLibraryLoadData                         
+      0.6       19,603,564      6,016      3,258.6      3,252.5       182      206,497      3,546.5  cudaMemsetAsync                           
+      0.5       18,001,538     22,456        801.6        744.0       325       15,163        252.0  cudaStreamIsCapturing_v10000              
+      0.4       13,493,455        169     79,842.9     82,039.0    26,598      421,517     46,195.9  cuModuleLoadData                          
+      0.3       10,862,324      9,998      1,086.4        417.0       271    4,538,659     46,724.1  cuKernelGetFunction                       
+      0.3        9,786,234     18,895        517.9        489.0       307        7,453        151.6  cudaStreamGetCaptureInfo_v2_v11030        
+      0.2        8,477,629      1,943      4,363.2      4,336.0     3,372       11,361        550.5  cudaStreamBeginCapture_v10000             
+      0.2        7,848,956      1,943      4,039.6      3,988.0     2,469       11,052        584.9  cudaGraphDestroy_v10000                   
+      0.1        3,305,186        128     25,821.8      2,208.5     1,414    1,124,514    138,112.8  cudaStreamCreateWithPriority              
+      0.1        2,707,973      1,943      1,393.7      1,378.0     1,055        7,779        187.1  cudaStreamEndCapture_v10000               
+      0.0        1,688,826         14    120,630.4      5,489.5     3,618    1,615,171    430,163.0  cudaHostAlloc                             
+      0.0        1,671,277      1,943        860.2        794.0       637        3,051        250.4  cudaGraphGetNodes_v10000                  
+      0.0          242,364          8     30,295.5     30,087.5    10,007       68,277     20,230.1  cudaMemGetInfo                            
+      0.0          135,593        810        167.4        142.0        81        1,741        108.0  cuGetProcAddress_v2                       
+      0.0           14,928         19        785.7        434.0       292        4,526      1,066.1  cudaEventCreateWithFlags                  
+      0.0           14,127         15        941.8        817.0       490        1,683        386.4  cuLibraryGetKernel                        
+      0.0            9,003         14        643.1        654.0       387        1,028        172.1  cudaThreadExchangeStreamCaptureMode_v10010
+      0.0            6,456          1      6,456.0      6,456.0     6,456        6,456          0.0  cudaEventRecord                           
+      0.0            4,689          3      1,563.0      1,239.0     1,181        2,269        612.1  cuInit                                    
+      0.0            4,627          4      1,156.8        965.5       190        2,506      1,151.2  cuModuleGetLoadingMode                    
+      0.0            4,166          1      4,166.0      4,166.0     4,166        4,166          0.0  cudaStreamWaitEvent                       
+      0.0            1,492          1      1,492.0      1,492.0     1,492        1,492          0.0  cudaEventDestroy                          
+      0.0              923          2        461.5        461.5       352          571        154.9  cudaGetDriverEntryPoint_v11030            
+
+[6/8] Executing 'cuda_gpu_kern_sum' stats report
+
+ Time (%)  Total Time (ns)  Instances   Avg (ns)     Med (ns)    Min (ns)   Max (ns)   StdDev (ns)                                                  Name                                                
+ --------  ---------------  ---------  -----------  -----------  ---------  ---------  -----------  ----------------------------------------------------------------------------------------------------
+     39.3      732,498,877     19,488     37,587.2     37,313.0     36,128     65,280      1,363.6  void flash::flash_fwd_splitkv_kernel<Flash_fwd_kernel_traits<(int)128, (int)64, (int)128, (int)4, (…
+     18.9      351,468,468      1,122    313,251.8    497,570.0     10,560    505,058    229,889.0  void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_16x16_128x2_tn_align8>(T1::Par…
+     10.2      191,063,614      3,836     49,808.0     22,016.0      7,808    132,256     39,115.9  void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_32x6_tn_align8>(T1::Param…
+      8.5      157,619,408        924    170,583.8    162,464.5     40,128  1,414,887    225,663.1  ampere_bf16_s1688gemm_bf16_128x64_sliced1x2_ldg8_f2f_tn                                             
+      4.2       78,494,924         28  2,803,390.1  2,805,357.0  2,788,621  2,809,197      5,574.6  ampere_bf16_s16816gemm_bf16_128x64_ldg8_f2f_tn                                                      
+      2.2       40,692,415        140    290,660.1    254,817.0    186,816    423,458     91,368.5  ampere_bf16_s1688gemm_bf16_64x128_sliced1x2_ldg8_f2f_tn                                             
+      2.1       39,174,391          8  4,896,798.9  4,852,759.0  4,797,142  5,086,872    109,448.4  void at_cuda_detail::cub::DeviceSegmentedRadixSortKernel<at_cuda_detail::cub::DeviceRadixSortPolicy…
+      1.6       29,472,378     19,488      1,512.3      1,568.0      1,055      3,232        143.0  void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0…
+      1.2       22,871,420      1,960     11,669.1      4,447.5      1,055    461,155     54,297.5  triton_poi_fused_mul_silu_1                                                                         
+      1.2       22,661,811        784     28,905.4     12,512.0     11,712     62,496     20,606.8  ampere_bf16_s16816gemm_bf16_64x64_ldg8_f2f_stages_64x5_tn                                           
+      1.1       20,375,102          4  5,093,775.5  5,084,087.5  4,911,190  5,295,737    198,253.9  void at_cuda_detail::cub::DeviceSegmentedRadixSortKernel<at_cuda_detail::cub::DeviceRadixSortPolicy…
+      0.8       14,291,999         28    510,428.5    511,906.5    469,986    513,666      7,950.6  void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<signed char>, std::a…
+      0.8       14,252,929        170     83,840.8     51,520.0     44,417  3,040,014    307,906.7  ampere_bf16_s1688gemm_bf16_128x128_ldg8_f2f_stages_32x1_tn                                          
+      0.7       12,842,687        106    121,157.4      8,912.0      6,976    498,914    190,206.3  std::enable_if<!T7, void>::type internal::gemvx::kernel<int, int, __nv_bfloat16, __nv_bfloat16, __n…
+      0.6       10,367,020      1,960      5,289.3      3,488.0      1,536    111,617     12,601.5  triton_red_fused__to_copy_add_mean_mul_pow_rsqrt_2                                                  
+      0.5        9,729,741          4  2,432,435.3  2,432,139.0  2,391,627  2,473,836     44,136.5  void at::native::<unnamed>::cunn_SoftMaxForward<(int)4, float, float, float, at::native::<unnamed>:…
+      0.5        9,150,918         28    326,818.5    326,993.5    324,545    330,945      1,225.0  ampere_bf16_s1688gemm_bf16_128x128_ldg8_relu_f2f_stages_32x1_tn                                     
+      0.5        8,477,259        224     37,844.9     37,792.5     36,449     39,201        492.7  void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x128_32x6_tn_align8>(T1::Para…
+      0.4        7,797,860          2  3,898,930.0  3,898,930.0  3,708,016  4,089,844    269,993.2  void at::native::_scatter_gather_elementwise_kernel<(int)128, (int)8, void at::native::_cuda_scatte…
+      0.4        7,227,678        336     21,510.9     21,424.0     21,088     22,752        352.3  ampere_bf16_s16816gemm_bf16_128x64_ldg8_relu_f2f_stages_64x3_tn                                     
+      0.4        6,877,224        700      9,824.6      5,776.0      5,408    712,515     53,225.4  void at::native::reduce_kernel<(int)512, (int)1, at::native::ReduceOp<float, at::native::ArgMaxOps<…
+      0.3        6,458,033      1,960      3,294.9      2,048.0      1,536     79,136      9,038.3  triton_red_fused__to_copy_add_mean_mul_pow_rsqrt_0                                                  
+      0.3        6,187,093        476     12,998.1     12,896.0     12,032     14,464        534.8  ampere_bf16_s16816gemm_bf16_64x64_ldg8_relu_f2f_stages_64x5_tn                                      
+      0.3        5,894,651          4  1,473,662.8  1,473,527.0  1,472,615  1,474,982      1,009.9  void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::masked_fill_kernel(at…
+      0.2        4,390,517        140     31,360.8     27,248.5     26,080     49,569      8,300.7  ampere_bf16_s1688gemm_bf16_128x64_sliced1x2_ldg8_relu_f2f_tn                                        
+      0.2        4,165,038        698      5,967.1      3,088.0      2,080  1,005,157     53,500.9  void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::TensorIterator…
+      0.2        4,133,648      1,890      2,187.1      1,824.0      1,343     22,113      2,250.2  triton_poi_fused_cat_3                                                                              
+      0.2        3,999,603          2  1,999,801.5  1,999,801.5  1,998,857  2,000,746      1,335.7  void at::native::vectorized_elementwise_kernel<(int)4, at::native::BinaryFunctor<float, float, floa…
+      0.2        3,826,146      4,587        834.1        800.0        767      1,312         75.4  void at::native::vectorized_elementwise_kernel<(int)2, at::native::FillFunctor<long>, std::array<ch…
+      0.2        3,594,124         56     64,180.8     64,192.5     63,264     65,537        457.5  void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_64x1_tn_align8>(T1::Para…
+      0.2        3,435,441          4    858,860.3    858,356.0    855,844    862,885      3,126.9  void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_nocast<at::n…
+      0.2        3,196,367          2  1,598,183.5  1,598,183.5  1,574,055  1,622,312     34,122.9  void at::native::tensor_kernel_scan_innermost_dim<float, std::plus<float>>(T1 *, const T1 *, unsign…
+      0.2        2,907,709      1,512      1,923.1      1,856.0      1,312      2,976        434.3  void cublasLt::splitKreduce_kernel<(int)32, (int)16, int, __nv_bfloat16, __nv_bfloat16, float, (boo…
+      0.1        2,738,586      1,890      1,449.0      1,152.0        863     17,280      1,803.1  triton_poi_fused_view_5                                                                             
+      0.1        2,670,047      1,890      1,412.7      1,375.0      1,216      6,112        532.9  triton_poi_fused_cat_4                                                                              
+      0.1        2,665,755        832      3,204.0      3,136.0      2,880      7,488        322.6  void at::native::index_elementwise_kernel<(int)128, (int)4, void at::native::gpu_index_kernel<void …
+      0.1        2,582,188          2  1,291,094.0  1,291,094.0  1,290,950  1,291,238        203.6  at::native::<unnamed>::fill_reverse_indices_kernel(long *, int, at::cuda::detail::IntDivider<unsign…
+      0.1        2,578,221          2  1,289,110.5  1,289,110.5  1,287,431  1,290,790      2,375.2  void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_nocast<at::n…
+      0.1        2,425,933        112     21,660.1     21,632.0      9,535     34,625     12,038.2  void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align8>(T1::Par…
+      0.1        1,374,214          2    687,107.0    687,107.0    679,619    694,595     10,589.6  void at::native::<unnamed>::distribution_elementwise_grid_stride_kernel<float, (int)4, void at::nat…
+      0.1        1,150,824         28     41,100.9     41,056.0     40,161     43,264        550.5  ampere_bf16_s1688gemm_bf16_64x64_sliced1x4_ldg8_f2f_tn                                              
+      0.1        1,135,783        696      1,631.9      1,632.0      1,344      1,856         53.5  void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::TensorIterator…
+      0.1          957,187         28     34,185.3     34,752.5     17,888     35,328      3,202.1  std::enable_if<!T7, void>::type internal::gemvx::kernel<int, int, __nv_bfloat16, float, float, floa…
+      0.0          868,532        698      1,244.3      1,248.0      1,056      1,728         62.7  void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::TensorIterator…
+      0.0          633,512        696        910.2        896.0        895      1,024         24.6  void at::native::unrolled_elementwise_kernel<at::native::CUDAFunctorOnSelf_add<int>, std::array<cha…
+      0.0          614,815        696        883.4        865.0        863        992         24.3  void at::native::unrolled_elementwise_kernel<at::native::FillFunctor<int>, std::array<char *, (unsi…
+      0.0          456,669        515        886.7        896.0        863      1,855         47.4  void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<int>, std::array<cha…
+      0.0          296,773        168      1,766.5      1,760.0      1,536      2,016        122.4  void cublasLt::splitKreduce_kernel<(int)32, (int)16, int, __nv_bfloat16, __nv_bfloat16, float, (boo…
+      0.0          260,900         70      3,727.1      3,072.0      1,984     36,449      4,037.6  triton_red_fused__to_copy_add_embedding_mean_mul_pow_rsqrt_0                                        
+      0.0          192,163         70      2,745.2      2,144.0      1,663     39,552      4,482.5  triton_poi_fused_cat_1                                                                              
+      0.0          167,425        185        905.0        896.0        863        992         31.8  void at::native::vectorized_elementwise_kernel<(int)2, at::native::FillFunctor<int>, std::array<cha…
+      0.0          156,225          1    156,225.0    156,225.0    156,225    156,225          0.0  void at::native::<unnamed>::CatArrayBatchedCopy_aligned16_contig<at::native::<unnamed>::OpaqueType<…
+      0.0          102,214         70      1,460.2      1,360.0      1,216      9,152        936.0  triton_poi_fused_cat_2                                                                              
+      0.0          100,832         70      1,440.5      1,199.5        864     13,984      1,559.1  triton_poi_fused_view_3                                                                             
+      0.0           79,488          1     79,488.0     79,488.0     79,488     79,488          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::bfloat16_copy_kernel_cuda(at::Te…
+      0.0           64,127         58      1,105.6        896.0        864     11,392      1,376.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<c10::BFloat16>, std:…
+      0.0           44,032          1     44,032.0     44,032.0     44,032     44,032          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::sin_kernel_cuda(at::TensorIterat…
+      0.0           36,640         28      1,308.6      1,312.0      1,280      1,375         18.0  void cublasLt::splitKreduce_kernel<(int)32, (int)16, int, float, __nv_bfloat16, float, (bool)0, __n…
+      0.0           26,464          1     26,464.0     26,464.0     26,464     26,464          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::cos_kernel_cuda(at::TensorIterat…
+      0.0           19,520          1     19,520.0     19,520.0     19,520     19,520          0.0  void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_nocast<at::n…
+      0.0           11,425         11      1,038.6        864.0        800      1,504        271.4  void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<float>, std::array<c…
+      0.0           10,720          2      5,360.0      5,360.0      5,280      5,440        113.1  void at::native::_scatter_gather_elementwise_kernel<(int)128, (int)8, void at::native::_cuda_scatte…
+      0.0            8,736          2      4,368.0      4,368.0      4,224      4,512        203.6  void at::native::<unnamed>::distribution_elementwise_grid_stride_kernel<float, (int)4, void at::nat…
+      0.0            3,616          2      1,808.0      1,808.0      1,600      2,016        294.2  void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_nocast<at::n…
+      0.0            3,456          2      1,728.0      1,728.0      1,696      1,760         45.3  void at::native::vectorized_elementwise_kernel<(int)2, at::native::CUDAFunctorOnOther_add<long>, st…
+      0.0            3,200          2      1,600.0      1,600.0      1,504      1,696        135.8  void at::native::vectorized_elementwise_kernel<(int)2, at::native::<unnamed>::where_kernel_impl(at:…
+      0.0            3,008          2      1,504.0      1,504.0      1,344      1,664        226.3  void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnOther_add<float>, s…
+      0.0            2,977          2      1,488.5      1,488.5        993      1,984        700.7  void <unnamed>::elementwise_kernel_with_index<int, at::native::arange_cuda_out(const c10::Scalar &,…
+      0.0            2,912          2      1,456.0      1,456.0      1,344      1,568        158.4  void at::native::vectorized_elementwise_kernel<(int)4, void at::native::compare_scalar_kernel<float…
+      0.0            2,304          1      2,304.0      2,304.0      2,304      2,304          0.0  void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl<at::native::…
+      0.0            1,185          1      1,185.0      1,185.0      1,185      1,185          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::reciprocal_kernel_cuda(at::Tenso…
+      0.0            1,024          1      1,024.0      1,024.0      1,024      1,024          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, float, floa…
+      0.0            1,024          1      1,024.0      1,024.0      1,024      1,024          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::BUnaryFunctor<float, float, floa…
+      0.0              896          1        896.0        896.0        896        896          0.0  void at::native::vectorized_elementwise_kernel<(int)2, at::native::FillFunctor<double>, std::array<…
+
+[7/8] Executing 'cuda_gpu_mem_time_sum' stats report
+
+ Time (%)  Total Time (ns)  Count   Avg (ns)     Med (ns)    Min (ns)    Max (ns)    StdDev (ns)            Operation           
+ --------  ---------------  -----  -----------  -----------  ---------  -----------  -----------  ------------------------------
+     98.5      509,288,023  4,846    105,094.5        352.0        288  118,727,568  1,792,036.2  [CUDA memcpy Host-to-Device]  
+      1.1        5,441,785      4  1,360,446.3  1,361,446.0  1,356,198    1,362,695      2,892.8  [CUDA memcpy Device-to-Device]
+      0.3        1,631,584  3,104        525.6        384.0        289        2,016        236.9  [CUDA memset]                 
+      0.2          794,753    696      1,141.9      1,120.0        864        1,664         71.6  [CUDA memcpy Device-to-Host]  
+
+[8/8] Executing 'cuda_gpu_mem_size_sum' stats report
+
+ Total (MB)  Count  Avg (MB)  Med (MB)  Min (MB)  Max (MB)  StdDev (MB)            Operation           
+ ----------  -----  --------  --------  --------  --------  -----------  ------------------------------
+  3,090.507  4,846     0.638     0.000     0.000   466.747        7.613  [CUDA memcpy Host-to-Device]  
+  2,489.319      4   622.330   622.330   622.330   622.330        0.000  [CUDA memcpy Device-to-Device]
+      2.068  3,104     0.001     0.001     0.000     0.006        0.001  [CUDA memset]                 
+      0.008    696     0.000     0.000     0.000     0.000        0.000  [CUDA memcpy Device-to-Host]  
+
+Generated:
+    /data/cy/kv_cache_vs_util/sim_traverse_bs/traverse_bs_util_sim_prefill_1152.nsys-rep
+    /data/cy/kv_cache_vs_util/sim_traverse_bs/traverse_bs_util_sim_prefill_1152.sqlite

sim_traverse_bs/traverse_bs_util_sim_prefill_1152.nsys-rep

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf9e96f6acddf12e5bda76a4f88fa873ab34c046eca72b3ca35e23ba788c9e25
+size 21975663

std_traverse_bs/terminal.log

@@ -0,0 +1,116 @@
+模型加载完成。
+
+===== 场景：prefill640_decode512 | prefill=640, decode=512 =====
+
+--- 批量大小 bs=1 ---
+Adding requests: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 431.65it/s]
+Processed prompts: 100%|███████████████████████████████████████████████████████████████████| 1/1 [00:02<00:00,  2.57s/it, est. speed input: 249.31 toks/s, output: 199.44 toks/s]
+执行时间: 2.5731 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 512
+吞吐(生成tokens/秒): 198.98
+TTFT (V1 metrics): 0.0238 s
+解码吞吐 (V1 metrics): 200.69 tok/s
+
+--- 批量大小 bs=2 ---
+Adding requests: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 202.04it/s]
+Processed prompts: 100%|███████████████████████████████████████████████████████████████████| 2/2 [00:02<00:00,  1.36s/it, est. speed input: 470.56 toks/s, output: 376.45 toks/s]
+执行时间: 2.7325 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 1024
+吞吐(生成tokens/秒): 374.75
+TTFT (V1 metrics): 0.0129 s
+解码吞吐 (V1 metrics): 188.36 tok/s
+
+--- 批量大小 bs=4 ---
+Adding requests: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:00<00:00, 210.51it/s]
+Processed prompts: 100%|███████████████████████████████████████████████████████████████████| 4/4 [00:03<00:00,  1.33it/s, est. speed input: 852.39 toks/s, output: 681.91 toks/s]
+执行时间: 3.0248 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 2048
+吞吐(生成tokens/秒): 677.07
+TTFT (V1 metrics): 0.0164 s
+解码吞吐 (V1 metrics): 170.68 tok/s
+
+--- 批量大小 bs=8 ---
+Adding requests: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 8/8 [00:00<00:00, 229.89it/s]
+Processed prompts: 100%|█████████████████████████████████████████████████████████████████| 8/8 [00:03<00:00,  2.59it/s, est. speed input: 1658.80 toks/s, output: 1327.03 toks/s]
+执行时间: 3.1236 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 4096
+吞吐(生成tokens/秒): 1311.31
+TTFT (V1 metrics): 0.0219 s
+解码吞吐 (V1 metrics): 165.93 tok/s
+
+--- 批量大小 bs=16 ---
+Adding requests: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 16/16 [00:00<00:00, 259.80it/s]
+Processed prompts: 100%|███████████████████████████████████████████████████████████████| 16/16 [00:03<00:00,  5.28it/s, est. speed input: 3376.77 toks/s, output: 2701.40 toks/s]
+执行时间: 3.0966 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 8192
+吞吐(生成tokens/秒): 2645.49
+TTFT (V1 metrics): 0.0314 s
+解码吞吐 (V1 metrics): 168.86 tok/s
+
+--- 批量大小 bs=32 ---
+Adding requests: 100%|████████████████████████████████████���█████████████████████████████████████████████████████████████████████████████████████| 32/32 [00:00<00:00, 335.64it/s]
+Processed prompts: 100%|███████████████████████████████████████████████████████████████| 32/32 [00:03<00:00,  9.64it/s, est. speed input: 6170.77 toks/s, output: 4936.59 toks/s]
+执行时间: 3.4167 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 16384
+吞吐(生成tokens/秒): 4795.34
+TTFT (V1 metrics): 0.0426 s
+解码吞吐 (V1 metrics): 154.14 tok/s
+
+--- 批量大小 bs=64 ---
+Adding requests: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 64/64 [00:00<00:00, 449.34it/s]
+Processed prompts: 100%|██████████████████████████████████████████████████████████████| 64/64 [00:03<00:00, 16.67it/s, est. speed input: 10672.90 toks/s, output: 8538.28 toks/s]
+执行时间: 3.9839 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 32768
+吞吐(生成tokens/秒): 8225.04
+TTFT (V1 metrics): 0.0638 s
+解码吞吐 (V1 metrics): 133.10 tok/s
+
+--- 批量大小 bs=128 ---
+Adding requests: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 128/128 [00:00<00:00, 489.69it/s]
+Processed prompts: 100%|███████████████████████████████████████████████████████████| 128/128 [00:04<00:00, 26.33it/s, est. speed input: 16849.84 toks/s, output: 13425.52 toks/s]
+执行时间: 5.1276 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 65272
+吞吐(生成tokens/秒): 12729.46
+TTFT (V1 metrics): 0.1208 s
+解码吞吐 (V1 metrics): 104.39 tok/s
+
+--- 批量大小 bs=256 ---
+Adding requests: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 256/256 [00:00<00:00, 351.03it/s]
+Processed prompts: 100%|███████████████████████████████████████████████████████████| 256/256 [00:07<00:00, 35.75it/s, est. speed input: 22881.00 toks/s, output: 18304.75 toks/s]
+执行时间: 7.8977 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 131072
+吞吐(生成tokens/秒): 16596.30
+TTFT (V1 metrics): 0.4649 s
+解码吞吐 (V1 metrics): 69.79 tok/s
+
+--- 批量大小 bs=512 ---
+Adding requests: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 512/512 [00:00<00:00, 535.56it/s]
+Processed prompts: 100%|███████████████████████████████████████████████████████████| 512/512 [00:12<00:00, 41.03it/s, est. speed input: 26258.02 toks/s, output: 21006.35 toks/s]
+执行时间: 13.4567 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 262144
+吞吐(生成tokens/秒): 19480.62
+TTFT (V1 metrics): 0.4882 s
+解码吞吐 (V1 metrics): 40.60 tok/s
+
+--- 批量大小 bs=1024 ---
+Adding requests: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1024/1024 [00:01<00:00, 553.17it/s]
+Processed prompts: 100%|████████��████████████████████████████████████████████████| 1024/1024 [00:24<00:00, 41.44it/s, est. speed input: 26524.06 toks/s, output: 21219.22 toks/s]
+执行时间: 26.5997 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 524288
+吞吐(生成tokens/秒): 19710.29
+TTFT (V1 metrics): 0.9604 s
+解码吞吐 (V1 metrics): 20.56 tok/s
+
+完成。提示：在 Nsight Systems 中可通过 NVTX 区间快速定位各场景/批量的调用。
+[rank0]:[W813 18:53:54.532255598 ProcessGroupNCCL.cpp:1479] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

std_traverse_bs/traverse_bs_util_std.log

@@ -0,0 +1,398 @@
+WARNING: CPU IP/backtrace sampling not supported, disabling.
+Try the 'nsys status --environment' command to learn more.
+
+WARNING: CPU context switch tracing not supported, disabling.
+Try the 'nsys status --environment' command to learn more.
+
+INFO 08-13 19:02:19 [__init__.py:235] Automatically detected platform cuda.
+CUDA_VISIBLE_DEVICES = 3
+--- vLLM V1 基准测试（含 NVTX 标记）---
+模型: Qwen/Qwen2-1.5B
+批量大小: [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024]
+场景: ['prefill640_decode512']
+------------------------------------------------------------
+加载分词器/模型中...
+INFO 08-13 19:02:29 [config.py:1604] Using max model len 4096
+INFO 08-13 19:02:29 [config.py:2434] Chunked prefill is enabled with max_num_batched_tokens=8192.
+INFO 08-13 19:02:35 [__init__.py:235] Automatically detected platform cuda.
+INFO 08-13 19:02:37 [core.py:572] Waiting for init message from front-end.
+INFO 08-13 19:02:37 [core.py:71] Initializing a V1 LLM engine (v0.10.0) with config: model='Qwen/Qwen2-1.5B', speculative_config=None, tokenizer='Qwen/Qwen2-1.5B', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config={}, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=4096, download_dir=None, load_format=LoadFormat.AUTO, tensor_parallel_size=1, pipeline_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=False, kv_cache_dtype=auto,  device_config=cuda, decoding_config=DecodingConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_backend=''), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None), seed=0, served_model_name=Qwen/Qwen2-1.5B, num_scheduler_steps=1, multi_step_stream_outputs=True, enable_prefix_caching=True, chunked_prefill_enabled=True, use_async_output_proc=True, pooler_config=None, compilation_config={"level":3,"debug_dump_path":"","cache_dir":"","backend":"","custom_ops":[],"splitting_ops":["vllm.unified_attention","vllm.unified_attention_with_output","vllm.mamba_mixer2"],"use_inductor":true,"compile_sizes":[],"inductor_compile_config":{"enable_auto_functionalized_v2":false},"inductor_passes":{},"use_cudagraph":true,"cudagraph_num_of_warmups":1,"cudagraph_capture_sizes":[512,504,496,488,480,472,464,456,448,440,432,424,416,408,400,392,384,376,368,360,352,344,336,328,320,312,304,296,288,280,272,264,256,248,240,232,224,216,208,200,192,184,176,168,160,152,144,136,128,120,112,104,96,88,80,72,64,56,48,40,32,24,16,8,4,2,1],"cudagraph_copy_inputs":false,"full_cuda_graph":false,"max_capture_size":512,"local_cache_dir":null}
+INFO 08-13 19:02:40 [parallel_state.py:1102] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, TP rank 0, EP rank 0
+WARNING 08-13 19:02:40 [topk_topp_sampler.py:59] FlashInfer is not available. Falling back to the PyTorch-native implementation of top-p & top-k sampling. For the best performance, please install FlashInfer.
+INFO 08-13 19:02:40 [gpu_model_runner.py:1843] Starting to load model Qwen/Qwen2-1.5B...
+INFO 08-13 19:02:40 [gpu_model_runner.py:1875] Loading model from scratch...
+INFO 08-13 19:02:40 [cuda.py:290] Using Flash Attention backend on V1 engine.
+INFO 08-13 19:02:40 [weight_utils.py:296] Using model weights format ['*.safetensors']
+INFO 08-13 19:02:41 [weight_utils.py:349] No model.safetensors.index.json found in remote.
+
+
+
+
+INFO 08-13 19:02:42 [default_loader.py:262] Loading weights took 0.75 seconds
+INFO 08-13 19:02:42 [gpu_model_runner.py:1892] Model loading took 2.9105 GiB and 1.965894 seconds
+INFO 08-13 19:02:48 [backends.py:530] Using cache directory: /home/cy/.cache/vllm/torch_compile_cache/40b61c71e9/rank_0_0/backbone for vLLM's torch.compile
+INFO 08-13 19:02:48 [backends.py:541] Dynamo bytecode transform time: 6.08 s
+INFO 08-13 19:02:53 [backends.py:194] Cache the graph for dynamic shape for later use
+INFO 08-13 19:03:14 [backends.py:215] Compiling a graph for dynamic shape takes 25.33 s
+INFO 08-13 19:03:21 [monitor.py:34] torch.compile takes 31.41 s in total
+INFO 08-13 19:03:22 [gpu_worker.py:255] Available KV cache memory: 12.80 GiB
+INFO 08-13 19:03:22 [kv_cache_utils.py:833] GPU KV cache size: 479,456 tokens
+INFO 08-13 19:03:22 [kv_cache_utils.py:837] Maximum concurrency for 4,096 tokens per request: 117.05x
+
+INFO 08-13 19:03:25 [gpu_model_runner.py:2485] Graph capturing finished in 2 secs, took 0.49 GiB
+INFO 08-13 19:03:25 [core.py:193] init engine (profile, create kv cache, warmup model) took 42.37 seconds
+模型加载完成。
+
+===== 场景：prefill640_decode512 | prefill=640, decode=512 =====
+
+--- 批量大小 bs=1 ---
+
+
+执行时间: 3.3360 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 512
+吞吐(生成tokens/秒): 153.48
+TTFT (V1 metrics): 0.0327 s
+解码吞吐 (V1 metrics): 154.93 tok/s
+
+--- 批量大小 bs=2 ---
+
+
+执行时间: 3.7300 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 1024
+吞吐(生成tokens/秒): 274.53
+TTFT (V1 metrics): 0.0158 s
+解码吞吐 (V1 metrics): 137.84 tok/s
+
+--- 批量大小 bs=4 ---
+
+
+执行时间: 3.6164 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 2048
+吞吐(生成tokens/秒): 566.30
+TTFT (V1 metrics): 0.0169 s
+解码吞吐 (V1 metrics): 142.63 tok/s
+
+--- 批量大小 bs=8 ---
+
+
+执行时间: 3.7265 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 4096
+吞吐(生成tokens/秒): 1099.15
+TTFT (V1 metrics): 0.0219 s
+解码吞吐 (V1 metrics): 138.89 tok/s
+
+--- 批量大小 bs=16 ---
+
+
+执行时间: 3.8919 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 8192
+吞吐(生成tokens/秒): 2104.89
+TTFT (V1 metrics): 0.0329 s
+解码吞吐 (V1 metrics): 133.82 tok/s
+
+--- 批量大小 bs=32 ---
+
+
+执行时间: 4.0341 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 16384
+吞吐(生成tokens/秒): 4061.41
+TTFT (V1 metrics): 0.0461 s
+解码吞吐 (V1 metrics): 130.12 tok/s
+
+--- 批量大小 bs=64 ---
+
+
+执行时间: 4.4199 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 32768
+吞吐(生成tokens/秒): 7413.77
+TTFT (V1 metrics): 0.0691 s
+解码吞吐 (V1 metrics): 120.00 tok/s
+
+--- 批量大小 bs=128 ---
+
+
+执行时间: 6.2947 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 65421
+吞吐(生成tokens/秒): 10393.02
+TTFT (V1 metrics): 0.1218 s
+解码吞吐 (V1 metrics): 84.64 tok/s
+
+--- 批量大小 bs=256 ---
+
+
+执行时间: 9.2625 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 131072
+吞吐(生成tokens/秒): 14150.76
+TTFT (V1 metrics): 0.4813 s
+解码吞吐 (V1 metrics): 59.00 tok/s
+
+--- 批量大小 bs=512 ---
+
+
+执行时间: 14.1481 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 262144
+吞吐(生成tokens/秒): 18528.59
+TTFT (V1 metrics): 0.4908 s
+解码吞吐 (V1 metrics): 38.46 tok/s
+
+--- 批量大小 bs=1024 ---
+
+
+[rank0]:[W813 19:04:51.071947265 ProcessGroupNCCL.cpp:1479] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
+执行时间: 27.7908 s
+实际平均输入 tokens: 640.00（目标 640）
+生成总 tokens: 524288
+吞吐(生成tokens/秒): 18865.49
+TTFT (V1 metrics): 0.9638 s
+解码吞吐 (V1 metrics): 19.72 tok/s
+
+完成。提示：在 Nsight Systems 中可通过 NVTX 区间快速定位各场景/批量的调用。
+GPU 3: General Metrics for NVIDIA AD10x (any frequency)
+Generating '/tmp/nsys-report-e7ae.qdstrm'
+
+
+[3/8] Executing 'nvtx_sum' stats report
+
+ Time (%)  Total Time (ns)  Instances      Avg (ns)          Med (ns)         Min (ns)        Max (ns)     StdDev (ns)   Style                    Range                  
+ --------  ---------------  ---------  ----------------  ----------------  --------------  --------------  -----------  -------  ----------------------------------------
+     43.1   63,806,267,138          1  63,806,267,138.0  63,806,267,138.0  63,806,267,138  63,806,267,138          0.0  PushPop  :LLM_init                               
+     18.8   27,790,304,411          1  27,790,304,411.0  27,790,304,411.0  27,790,304,411  27,790,304,411          0.0  PushPop  :generate [prefill640_decode512] bs=1024
+      9.6   14,147,468,287          1  14,147,468,287.0  14,147,468,287.0  14,147,468,287  14,147,468,287          0.0  PushPop  :generate [prefill640_decode512] bs=512 
+      6.3    9,262,392,366          1   9,262,392,366.0   9,262,392,366.0   9,262,392,366   9,262,392,366          0.0  PushPop  :generate [prefill640_decode512] bs=256 
+      4.3    6,294,556,076          1   6,294,556,076.0   6,294,556,076.0   6,294,556,076   6,294,556,076          0.0  PushPop  :generate [prefill640_decode512] bs=128 
+      3.0    4,419,734,921          1   4,419,734,921.0   4,419,734,921.0   4,419,734,921   4,419,734,921          0.0  PushPop  :generate [prefill640_decode512] bs=64  
+      2.7    4,033,922,062          1   4,033,922,062.0   4,033,922,062.0   4,033,922,062   4,033,922,062          0.0  PushPop  :generate [prefill640_decode512] bs=32  
+      2.6    3,891,757,396          1   3,891,757,396.0   3,891,757,396.0   3,891,757,396   3,891,757,396          0.0  PushPop  :generate [prefill640_decode512] bs=16  
+      2.5    3,729,817,085          1   3,729,817,085.0   3,729,817,085.0   3,729,817,085   3,729,817,085          0.0  PushPop  :generate [prefill640_decode512] bs=2   
+      2.5    3,726,348,651          1   3,726,348,651.0   3,726,348,651.0   3,726,348,651   3,726,348,651          0.0  PushPop  :generate [prefill640_decode512] bs=8   
+      2.4    3,616,307,172          1   3,616,307,172.0   3,616,307,172.0   3,616,307,172   3,616,307,172          0.0  PushPop  :generate [prefill640_decode512] bs=4   
+      2.3    3,335,871,818          1   3,335,871,818.0   3,335,871,818.0   3,335,871,818   3,335,871,818          0.0  PushPop  :generate [prefill640_decode512] bs=1   
+      0.0           88,217          2          44,108.5          44,108.5          42,206          46,011      2,690.5  PushPop  CCCL:cub::DeviceSegmentedRadixSort      
+
+[4/8] Executing 'osrt_sum' stats report
+
+ Time (%)   Total Time (ns)   Num Calls     Avg (ns)          Med (ns)      Min (ns)      Max (ns)        StdDev (ns)             Name         
+ --------  -----------------  ---------  ---------------  ----------------  ---------  ---------------  ---------------  ----------------------
+     30.2  1,744,937,975,193     54,559     31,982,587.2          36,252.0      1,000  131,944,469,826  1,264,387,583.2  pthread_cond_timedwait
+     24.4  1,409,850,293,508     90,937     15,503,593.6      10,063,900.0      1,000   89,656,403,165    542,281,129.3  epoll_wait            
+     24.3  1,405,293,461,182      2,062    681,519,622.3         169,137.0      1,777  131,946,128,306  9,309,452,118.6  pthread_cond_wait     
+      8.2    470,649,973,426         67  7,024,626,469.0  10,000,075,590.0      8,879   10,000,128,605  4,563,470,551.8  sem_timedwait         
+      7.6    440,451,323,826     42,747     10,303,678.0           1,400.0      1,000   13,214,170,856    151,629,150.0  poll                  
+      2.6    148,761,603,454     44,756      3,323,836.0           2,218.0      1,000  130,016,250,506    617,021,890.8  read                  
+      2.4    141,326,431,455     11,558     12,227,585.3       7,211,303.0     21,424      658,160,680     14,124,156.6  sem_wait              
+      0.1      5,907,485,407        725      8,148,255.7           1,042.0      1,000      442,887,066     44,916,303.4  waitpid               
+      0.0      1,252,683,205    505,872          2,476.3           1,440.0      1,006       94,581,342        133,009.9  munmap                
+      0.0        907,262,068        353      2,570,147.5       1,227,045.0      1,197       26,263,430      3,108,782.5  pthread_rwlock_wrlock 
+      0.0        708,053,362        172      4,116,589.3         615,554.5      3,112       29,565,907      7,196,056.2  pthread_join          
+      0.0        327,596,329     10,080         32,499.6          10,962.5      1,000       29,220,721        379,953.4  ioctl                 
+      0.0        261,805,946        495        528,900.9           2,972.0      1,077       19,958,090      3,072,110.4  fopen                 
+      0.0        160,631,905     36,713          4,375.3           3,385.0      1,000        1,692,951         12,061.0  mmap64                
+      0.0        150,219,603      6,263         23,985.2           8,256.0      1,010        2,585,639        128,799.4  pthread_mutex_lock    
+      0.0        126,650,615         25      5,066,024.6       5,065,286.0  5,053,603        5,077,238          7,361.7  nanosleep             
+      0.0         99,487,929     31,988          3,110.2           2,516.0      1,000           75,099          2,795.8  open64                
+      0.0         90,344,763      9,083          9,946.6           4,712.0      1,148        2,648,924         36,587.2  recv                  
+      0.0         84,823,390      9,082          9,339.7           5,735.5      1,468           89,916          8,146.6  send                  
+      0.0         78,685,390     43,321          1,816.3           1,662.0      1,000          354,334          2,967.5  pthread_cond_signal   
+      0.0         69,793,423      5,751         12,135.9           2,001.0      1,035       19,751,200        420,727.9  open                  
+      0.0         66,948,530     15,954          4,196.3           2,557.0      1,010          575,245         10,386.1  write                 
+      0.0         50,999,387         10      5,099,938.7          19,477.5      9,808       50,802,125     16,058,116.0  connect               
+      0.0         18,140,344     11,019          1,646.3           1,411.0      1,000           24,064            709.4  epoll_ctl             
+      0.0         18,080,692        280         64,573.9          51,655.5     16,288          578,620         60,626.8  pthread_create        
+      0.0          9,776,017        147         66,503.5          68,572.0     55,588           85,413          5,091.0  sleep                 
+      0.0          7,076,518         18        393,139.9         383,949.5    262,109          584,262         85,377.7  posix_spawn           
+      0.0          6,601,831        899          7,343.5           5,539.0      1,010           84,493          9,244.8  fgets                 
+      0.0          5,962,720         22        271,032.7         169,725.0     16,478          675,901        237,522.4  pthread_rwlock_rdlock 
+      0.0          3,506,178        342         10,252.0           2,354.5      1,006          207,881         30,027.2  pthread_cond_broadcast
+      0.0          3,013,330      1,417          2,126.6           1,083.0      1,008          110,583          7,412.5  fclose                
+      0.0          2,975,781      1,319          2,256.1           1,598.0      1,000           22,362          1,984.6  stat                  
+      0.0          2,885,075        692          4,169.2           4,026.5      1,238           45,085          4,258.3  fopen64               
+      0.0          2,145,928        345          6,220.1           3,291.0      1,000           51,101          8,060.8  fread                 
+      0.0          1,950,425         65         30,006.5           3,106.0      1,201          261,266         71,483.9  futex                 
+      0.0          1,877,146        336          5,586.7           4,406.0      1,002           83,435          6,148.2  mmap                  
+      0.0          1,741,859        102         17,077.0           4,155.5      1,030          432,266         67,240.1  fwrite                
+      0.0          1,680,806      1,203          1,397.2           1,251.0      1,000            8,322            493.0  fstat                 
+      0.0          1,075,803          1      1,075,803.0       1,075,803.0  1,075,803        1,075,803              0.0  fork                  
+      0.0            721,566         99          7,288.5           6,229.0      2,560           17,976          3,711.4  pipe2                 
+      0.0            595,674         19         31,351.3           5,092.0      4,006          383,239         86,162.0  putc                  
+      0.0            248,534         41          6,061.8           4,850.0      1,628           18,035          4,032.6  socket                
+      0.0            183,085        115          1,592.0           1,513.0      1,001            2,876            429.3  sigaction             
+      0.0            167,627         16         10,476.7           2,835.5      1,086           55,140         16,437.5  bind                  
+      0.0            115,447          8         14,430.9           6,418.0      3,542           41,638         15,571.8  fputs                 
+      0.0             93,879         16          5,867.4           5,035.0      1,640           14,112          3,769.0  lstat                 
+      0.0             60,214          6         10,035.7          10,003.0      9,288           10,680            531.1  getc                  
+      0.0             49,588         27          1,836.6           1,677.0      1,019            2,939            596.6  dup2                  
+      0.0             47,084         37          1,272.5           1,112.0      1,002            3,284            431.2  fcntl                 
+      0.0             37,610         24          1,567.1           1,488.5      1,021            2,512            357.1  signal                
+      0.0             35,001          5          7,000.2           7,792.0      3,662           10,502          2,759.1  accept4               
+      0.0             31,541          9          3,504.6           4,424.0      1,078            6,400          2,269.5  fflush                
+      0.0             15,432          4          3,858.0           3,690.5      2,526            5,525          1,477.8  flock                 
+      0.0             15,311         11          1,391.9           1,301.0      1,125            1,918            254.0  listen                
+      0.0             14,040          8          1,755.0           1,552.0      1,310            3,235            642.4  pread                 
+      0.0             12,841          3          4,280.3           4,425.0      3,921            4,495            313.2  fputs_unlocked        
+      0.0             12,488          5          2,497.6           2,496.0      2,206            2,947            282.0  mprotect              
+      0.0             10,666          1         10,666.0          10,666.0     10,666           10,666              0.0  dup                   
+      0.0              6,561          3          2,187.0           1,727.0      1,636            3,198            876.7  flockfile             
+      0.0              6,208          1          6,208.0           6,208.0      6,208            6,208              0.0  kill                  
+      0.0              3,788          2          1,894.0           1,894.0      1,355            2,433            762.3  openat64              
+      0.0              2,317          2          1,158.5           1,158.5      1,025            1,292            188.8  pthread_mutex_trylock 
+
+[5/8] Executing 'cuda_api_sum' stats report
+
+ Time (%)  Total Time (ns)  Num Calls   Avg (ns)     Med (ns)    Min (ns)   Max (ns)    StdDev (ns)                     Name                   
+ --------  ---------------  ---------  -----------  -----------  --------  -----------  -----------  ------------------------------------------
+     54.0   18,079,332,597     62,830    287,750.0      8,213.0     2,790  112,775,333  1,983,727.7  cudaMemcpyAsync                           
+     21.3    7,151,454,905  1,151,462      6,210.8      4,908.0       775   60,344,468    100,074.1  cudaLaunchKernel                          
+     13.9    4,659,698,239      3,031  1,537,346.8     36,322.0     1,646  137,811,508  5,892,401.8  cudaDeviceSynchronize                     
+      5.7    1,902,754,251    154,454     12,319.2     10,944.0     7,132    6,919,461     39,439.6  cudaGraphLaunch_v10000                    
+      2.4      817,695,167    151,595      5,393.9      4,998.0       606    8,288,395     63,229.0  cuLaunchKernel                            
+      0.7      225,761,461      1,943    116,192.2     74,685.0    38,766    1,503,454    195,876.4  cudaGraphInstantiateWithFlags_v11040      
+      0.4      145,677,030     27,893      5,222.7      5,401.0       170    3,577,542     23,514.0  cudaMemsetAsync                           
+      0.4      123,078,496    156,852        784.7        749.0       293       28,187        217.2  cudaStreamIsCapturing_v10000              
+      0.2       80,818,259     41,653      1,940.3      1,901.0     1,681      230,930      1,185.3  cudaEventRecord                           
+      0.2       70,123,080     11,007      6,370.8      2,991.0     1,576   11,464,127    117,490.1  cudaStreamSynchronize                     
+      0.2       54,985,787        222    247,683.7    127,094.5    70,289    2,353,583    356,501.3  cudaFree                                  
+      0.1       39,808,657        349    114,064.9    107,734.0     9,252    1,028,648     56,280.0  cudaMalloc                                
+      0.1       33,850,909     41,671        812.3        782.0       275      186,043        928.6  cudaEventCreateWithFlags                  
+      0.1       25,030,442         10  2,503,044.2  2,591,457.0    57,483    4,465,642  1,429,904.6  cuLibraryLoadData                         
+      0.1       20,138,251        281     71,666.4     73,047.0    25,875      416,759     45,705.2  cuModuleLoadData                          
+      0.1       18,115,827     41,653        434.9        403.0       338      226,830      1,887.5  cudaEventDestroy                          
+      0.1       17,924,935     16,808      1,066.5        493.0       261    6,344,359     50,228.8  cuKernelGetFunction                       
+      0.0        9,263,417     18,895        490.3        467.0       322        6,477        105.4  cudaStreamGetCaptureInfo_v2_v11030        
+      0.0        7,974,210      1,943      4,104.1      4,022.0     3,214        9,676        584.5  cudaStreamBeginCapture_v10000             
+      0.0        7,518,878      1,943      3,869.7      3,828.0     2,357        7,833        536.7  cudaGraphDestroy_v10000                   
+      0.0        3,416,827        128     26,694.0      2,299.0     1,471    1,153,703    140,616.2  cudaStreamCreateWithPriority              
+      0.0        2,744,082      1,943      1,412.3      1,389.0     1,050        7,178        196.6  cudaStreamEndCapture_v10000               
+      0.0        1,570,575      1,943        808.3        739.0       614        2,547        251.2  cudaGraphGetNodes_v10000                  
+      0.0        1,322,243         15     88,149.5      6,436.0     3,579    1,170,830    300,044.6  cudaHostAlloc                             
+      0.0          280,352          8     35,044.0     26,955.5    12,673      101,212     28,421.5  cudaMemGetInfo                            
+      0.0          138,906        810        171.5        140.0        79        1,705        118.0  cuGetProcAddress_v2                       
+      0.0           23,009         16      1,438.1        808.5       451        5,531      1,508.2  cuLibraryGetKernel                        
+      0.0            8,159         14        582.8        544.5       324          990        193.4  cudaThreadExchangeStreamCaptureMode_v10010
+      0.0            4,031          1      4,031.0      4,031.0     4,031        4,031          0.0  cudaStreamWaitEvent                       
+      0.0            3,969          3      1,323.0      1,051.0     1,031        1,887        488.5  cuInit                                    
+      0.0            3,693          4        923.3        916.5        75        1,785        960.8  cuModuleGetLoadingMode                    
+      0.0            1,064          2        532.0        532.0       356          708        248.9  cudaGetDriverEntryPoint_v11030            
+
+[6/8] Executing 'cuda_gpu_kern_sum' stats report
+
+ Time (%)  Total Time (ns)  Instances   Avg (ns)     Med (ns)    Min (ns)   Max (ns)   StdDev (ns)                                                  Name                                                
+ --------  ---------------  ---------  -----------  -----------  ---------  ---------  -----------  ----------------------------------------------------------------------------------------------------
+     30.6   11,421,290,038    118,048     96,751.2     42,337.0     12,320    576,069    124,175.4  void flash::flash_fwd_splitkv_kernel<Flash_fwd_kernel_traits<(int)128, (int)64, (int)128, (int)4, (…
+     20.4    7,605,306,813     28,807    264,009.0    265,858.0     32,961    765,542    115,223.0  ampere_bf16_s1688gemm_bf16_64x128_sliced1x2_ldg8_f2f_tn                                             
+      8.8    3,277,208,753     47,634     68,799.8     77,473.0        800     81,121     21,559.5  void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<int>, std::array<cha…
+      5.6    2,083,505,861      1,271  1,639,265.0  1,387,852.0     39,745  4,515,436  1,121,994.0  ampere_bf16_s1688gemm_bf16_128x128_ldg8_f2f_stages_32x1_tn                                          
+      5.1    1,891,072,505     20,210     93,571.1     20,224.0      1,055    481,762    169,672.7  triton_poi_fused_mul_silu_1                                                                         
+      4.3    1,624,497,751    101,584     15,991.7      8,544.0      6,367     81,025     11,832.8  void flash::flash_fwd_splitkv_kernel<Flash_fwd_kernel_traits<(int)128, (int)64, (int)128, (int)4, (…
+      3.3    1,229,556,187      9,203    133,603.8     42,977.0      7,648    557,317    169,265.9  void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x64_32x6_tn_align8>(T1::Param…
+      2.5      940,467,368      5,865    160,352.5     13,312.0      1,984  1,008,298    288,283.2  void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::TensorIterator…
+      2.2      808,325,696      2,044    395,462.7    496,644.0     10,592    510,980    192,958.7  void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_16x16_128x2_tn_align8>(T1::Par…
+      1.9      720,826,362      5,867    122,861.1      9,824.0      5,120    714,183    210,401.9  void at::native::reduce_kernel<(int)512, (int)1, at::native::ReduceOp<float, at::native::ArgMaxOps<…
+      1.8      666,198,895    287,392      2,318.1      1,920.0      1,536      6,368        968.2  void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_nocast<at::n…
+      1.6      609,901,106      6,048    100,843.4     32,161.0      6,912  3,159,614    305,220.1  void flash::flash_fwd_splitkv_kernel<Flash_fwd_kernel_traits<(int)128, (int)64, (int)128, (int)4, (…
+      1.5      576,753,139     13,496     42,735.1     42,624.0     26,016    102,081      3,914.6  ampere_bf16_s1688gemm_bf16_128x64_sliced1x2_ldg8_relu_f2f_tn                                        
+      1.3      487,869,104     13,020     37,470.7     37,440.0     36,640     42,816        321.1  ampere_bf16_s1688gemm_bf16_64x64_sliced1x4_ldg8_f2f_tn                                              
+      1.2      431,150,928     22,086     19,521.5      3,489.0      1,344     75,104     27,792.8  triton_poi_fused_cat_3                                                                              
+      0.9      347,479,864        341  1,019,002.5    598,630.0    373,475  2,788,687    741,262.9  ampere_bf16_s16816gemm_bf16_128x64_ldg8_f2f_tn                                                      
+      0.8      316,815,668      1,904    166,394.8    154,193.5     40,352  1,291,207    149,498.4  ampere_bf16_s1688gemm_bf16_128x64_sliced1x2_ldg8_f2f_tn                                             
+      0.7      264,631,581        521    507,930.1    507,844.0    506,244    519,973        763.9  void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_16x16_128x1_tn_align8>(T1::Par…
+      0.7      258,705,730        610    424,107.8    487,972.0      6,976    500,003    159,920.1  std::enable_if<!T7, void>::type internal::gemvx::kernel<int, int, __nv_bfloat16, __nv_bfloat16, __n…
+      0.7      248,913,221    164,164      1,516.2      1,280.0      1,023     13,249        475.7  void vllm::reshape_and_cache_flash_kernel<__nv_bfloat16, __nv_bfloat16, (vllm::Fp8KVCacheDataType)0…
+      0.6      206,205,114     57,848      3,564.6      3,584.0      3,295      3,936        112.4  void flash::flash_fwd_splitkv_combine_kernel<Flash_fwd_kernel_traits<(int)128, (int)64, (int)128, (…
+      0.5      203,900,076    143,696      1,419.0      1,344.0      1,183      2,208        221.7  void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_nocast<at::n…
+      0.5      189,670,922     16,968     11,178.2     12,256.0      1,536    111,105      5,147.2  triton_red_fused__to_copy_add_mean_mul_pow_rsqrt_2                                                  
+      0.5      173,673,906     22,086      7,863.5      2,368.0        832     25,696      9,575.0  triton_poi_fused_view_5                                                                             
+      0.5      170,442,702     61,516      2,770.7      2,432.0      1,247     17,088      1,351.4  void vllm::merge_attn_states_kernel<__nv_bfloat16, (unsigned int)128>(T1 *, float *, const T1 *, co…
+      0.3       95,480,145     22,086      4,323.1      1,376.0      1,215     15,104      5,013.1  triton_poi_fused_cat_4                                                                              
+      0.2       88,249,892     16,968      5,201.0      5,473.0      1,535     80,096      3,526.8  triton_red_fused__to_copy_add_mean_mul_pow_rsqrt_0                                                  
+      0.2       58,167,224     15,232      3,818.8      3,840.0      3,711      4,032         31.4  void flash::flash_fwd_splitkv_combine_kernel<Flash_fwd_kernel_traits<(int)128, (int)64, (int)128, (…
+      0.1       45,137,635     14,420      3,130.2      3,167.0      3,008      3,457         59.7  void flash::flash_fwd_splitkv_combine_kernel<Flash_fwd_kernel_traits<(int)128, (int)64, (int)128, (…
+      0.1       43,394,155     14,084      3,081.1      3,072.0      3,008      3,200         19.7  void flash::flash_fwd_splitkv_combine_kernel<Flash_fwd_kernel_traits<(int)128, (int)64, (int)128, (…
+      0.1       39,138,907          8  4,892,363.4  4,849,279.0  4,795,198  5,088,513    110,842.2  void at_cuda_detail::cub::DeviceSegmentedRadixSortKernel<at_cuda_detail::cub::DeviceRadixSortPolicy…
+      0.1       22,694,961        784     28,947.7     12,575.5     11,744     62,528     20,606.8  ampere_bf16_s16816gemm_bf16_64x64_ldg8_f2f_stages_64x5_tn                                           
+      0.1       20,952,447      5,999      3,492.7      3,104.0      2,751      7,392        961.1  void at::native::index_elementwise_kernel<(int)128, (int)4, void at::native::gpu_index_kernel<void …
+      0.1       20,383,074          4  5,095,768.5  5,085,968.5  4,915,359  5,295,778    195,206.9  void at_cuda_detail::cub::DeviceSegmentedRadixSortKernel<at_cuda_detail::cub::DeviceRadixSortPolicy…
+      0.0       15,859,055        818     19,387.6      3,136.0      1,600     74,881     27,973.6  triton_poi_fused_cat_1                                                                              
+      0.0       14,288,503         28    510,303.7    511,683.0    469,923    513,059      7,932.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<signed char>, std::a…
+      0.0        9,813,089          4  2,453,272.3  2,468,672.5  2,391,504  2,484,240     42,132.0  void at::native::<unnamed>::cunn_SoftMaxForward<(int)4, float, float, float, at::native::<unnamed>:…
+      0.0        9,653,448        448     21,547.9     21,345.0     21,120     24,928        817.3  ampere_bf16_s16816gemm_bf16_128x64_ldg8_f2f_stages_32x6_tn                                          
+      0.0        9,408,780      5,863      1,604.8      1,376.0      1,120      2,752        455.9  void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::TensorIterator…
+      0.0        8,489,112        224     37,897.8     37,856.0     36,545     39,328        536.2  void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_64x128_32x6_tn_align8>(T1::Para…
+      0.0        8,311,916         28    296,854.1    294,593.5    293,505    332,482      7,194.4  ampere_bf16_s1688gemm_bf16_128x128_ldg8_relu_f2f_stages_32x1_tn                                     
+      0.0        7,845,574      9,023        869.5        864.0        767      1,281         77.2  void at::native::vectorized_elementwise_kernel<(int)2, at::native::FillFunctor<long>, std::array<ch…
+      0.0        7,775,377          2  3,887,688.5  3,887,688.5  3,705,239  4,070,138    258,022.6  void at::native::_scatter_gather_elementwise_kernel<(int)128, (int)8, void at::native::_cuda_scatte…
+      0.0        7,471,240      5,865      1,273.9      1,120.0        991      2,080        282.3  void at::native::unrolled_elementwise_kernel<at::native::direct_copy_kernel_cuda(at::TensorIterator…
+      0.0        7,214,254        336     21,471.0     21,408.0     21,056     22,625        299.8  ampere_bf16_s16816gemm_bf16_128x64_ldg8_relu_f2f_stages_64x3_tn                                     
+      0.0        6,324,707        818      7,731.9      2,368.0        863     24,832      9,334.7  triton_poi_fused_view_3                                                                             
+      0.0        6,169,925        476     12,962.0     12,864.0     11,776     14,304        602.5  ampere_bf16_s16816gemm_bf16_64x64_ldg8_relu_f2f_stages_64x5_tn                                      
+      0.0        5,903,909          4  1,475,977.3  1,475,449.0  1,472,585  1,480,426      3,980.7  void at::native::vectorized_elementwise_kernel<(int)4, at::native::<unnamed>::masked_fill_kernel(at…
+      0.0        5,501,036      5,863        938.3        928.0        895      1,312         71.9  void at::native::unrolled_elementwise_kernel<at::native::CUDAFunctorOnSelf_add<int>, std::array<cha…
+      0.0        4,936,334         28    176,297.6    176,817.5    174,209    178,978      1,446.6  ampere_bf16_s1688gemm_bf16_64x128_sliced1x2_ldg8_relu_f2f_tn                                        
+      0.0        4,754,998      5,432        875.4        864.0        800      1,217         36.3  void at::native::unrolled_elementwise_kernel<at::native::FillFunctor<int>, std::array<char *, (unsi…
+      0.0        3,995,930          2  1,997,965.0  1,997,965.0  1,996,108  1,999,822      2,626.2  void at::native::vectorized_elementwise_kernel<(int)4, at::native::BinaryFunctor<float, float, floa…
+      0.0        3,598,526         56     64,259.4     64,320.5     63,104     65,313        559.8  void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_64x1_tn_align8>(T1::Para…
+      0.0        3,541,818        818      4,329.9      1,376.0      1,215     14,815      5,022.4  triton_poi_fused_cat_2                                                                              
+      0.0        3,434,102          4    858,525.5    858,069.5    855,685    862,278      2,939.8  void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_nocast<at::n…
+      0.0        3,192,725          2  1,596,362.5  1,596,362.5  1,563,050  1,629,675     47,111.0  void at::native::tensor_kernel_scan_innermost_dim<float, std::plus<float>>(T1 *, const T1 *, unsign…
+      0.0        2,897,906      1,512      1,916.6      1,824.0      1,312      2,976        438.1  void cublasLt::splitKreduce_kernel<(int)32, (int)16, int, __nv_bfloat16, __nv_bfloat16, float, (boo…
+      0.0        2,593,403        606      4,279.5      4,384.0      1,984     35,904      1,445.1  triton_red_fused__to_copy_add_embedding_mean_mul_pow_rsqrt_0                                        
+      0.0        2,582,288          2  1,291,144.0  1,291,144.0  1,290,536  1,291,752        859.8  at::native::<unnamed>::fill_reverse_indices_kernel(long *, int, at::cuda::detail::IntDivider<unsign…
+      0.0        2,580,625          2  1,290,312.5  1,290,312.5  1,288,393  1,292,232      2,714.6  void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_nocast<at::n…
+      0.0        2,421,451        112     21,620.1     21,552.0      9,376     34,400     12,006.5  void cutlass::Kernel2<cutlass_80_wmma_tensorop_bf16_s161616gemm_bf16_32x32_128x2_tn_align8>(T1::Par…
+      0.0        1,377,256          2    688,628.0    688,628.0    682,820    694,436      8,213.8  void at::native::<unnamed>::distribution_elementwise_grid_stride_kernel<float, (int)4, void at::nat…
+      0.0        1,128,024      1,252        901.0        896.0        800      1,280         45.2  void at::native::vectorized_elementwise_kernel<(int)2, at::native::FillFunctor<int>, std::array<cha…
+      0.0          957,446         28     34,194.5     34,768.5     17,983     35,232      3,184.5  std::enable_if<!T7, void>::type internal::gemvx::kernel<int, int, __nv_bfloat16, float, float, floa…
+      0.0          670,018        731        916.6        928.0        863      1,024         22.0  void at::native::unrolled_elementwise_kernel<at::native::FillFunctor<long>, std::array<char *, (uns…
+      0.0          296,898        168      1,767.3      1,760.0      1,536      2,080        121.8  void cublasLt::splitKreduce_kernel<(int)32, (int)16, int, __nv_bfloat16, __nv_bfloat16, float, (boo…
+      0.0          157,249          1    157,249.0    157,249.0    157,249    157,249          0.0  void at::native::<unnamed>::CatArrayBatchedCopy_aligned16_contig<at::native::<unnamed>::OpaqueType<…
+      0.0           90,491         86      1,052.2        927.5        895     11,488      1,139.9  void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<c10::BFloat16>, std:…
+      0.0           78,785          1     78,785.0     78,785.0     78,785     78,785          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::bfloat16_copy_kernel_cuda(at::Te…
+      0.0           43,232          1     43,232.0     43,232.0     43,232     43,232          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::sin_kernel_cuda(at::TensorIterat…
+      0.0           36,737         28      1,312.0      1,312.0      1,280      1,344         17.4  void cublasLt::splitKreduce_kernel<(int)32, (int)16, int, float, __nv_bfloat16, float, (bool)0, __n…
+      0.0           26,432          1     26,432.0     26,432.0     26,432     26,432          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::cos_kernel_cuda(at::TensorIterat…
+      0.0           19,520          1     19,520.0     19,520.0     19,520     19,520          0.0  void at::native::elementwise_kernel<(int)128, (int)2, void at::native::gpu_kernel_impl_nocast<at::n…
+      0.0           11,713         11      1,064.8        864.0        800      1,536        305.4  void at::native::vectorized_elementwise_kernel<(int)4, at::native::FillFunctor<float>, std::array<c…
+      0.0           10,624          2      5,312.0      5,312.0      5,024      5,600        407.3  void at::native::_scatter_gather_elementwise_kernel<(int)128, (int)8, void at::native::_cuda_scatte…
+      0.0            8,639          2      4,319.5      4,319.5      4,128      4,511        270.8  void at::native::<unnamed>::distribution_elementwise_grid_stride_kernel<float, (int)4, void at::nat…
+      0.0            3,616          2      1,808.0      1,808.0      1,600      2,016        294.2  void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl_nocast<at::n…
+      0.0            3,489          2      1,744.5      1,744.5      1,696      1,793         68.6  void at::native::vectorized_elementwise_kernel<(int)2, at::native::CUDAFunctorOnOther_add<long>, st…
+      0.0            3,103          2      1,551.5      1,551.5      1,503      1,600         68.6  void at::native::vectorized_elementwise_kernel<(int)2, at::native::<unnamed>::where_kernel_impl(at:…
+      0.0            2,976          2      1,488.0      1,488.0      1,376      1,600        158.4  void at::native::vectorized_elementwise_kernel<(int)4, void at::native::compare_scalar_kernel<float…
+      0.0            2,975          2      1,487.5      1,487.5        991      1,984        702.2  void <unnamed>::elementwise_kernel_with_index<int, at::native::arange_cuda_out(const c10::Scalar &,…
+      0.0            2,944          2      1,472.0      1,472.0      1,344      1,600        181.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::CUDAFunctorOnOther_add<float>, s…
+      0.0            2,400          1      2,400.0      2,400.0      2,400      2,400          0.0  void at::native::elementwise_kernel<(int)128, (int)4, void at::native::gpu_kernel_impl<at::native::…
+      0.0            1,185          1      1,185.0      1,185.0      1,185      1,185          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::reciprocal_kernel_cuda(at::Tenso…
+      0.0            1,025          1      1,025.0      1,025.0      1,025      1,025          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::AUnaryFunctor<float, float, floa…
+      0.0            1,025          1      1,025.0      1,025.0      1,025      1,025          0.0  void at::native::vectorized_elementwise_kernel<(int)4, at::native::BUnaryFunctor<float, float, floa…
+      0.0              896          1        896.0        896.0        896        896          0.0  void at::native::vectorized_elementwise_kernel<(int)2, at::native::FillFunctor<double>, std::array<…
+
+[7/8] Executing 'cuda_gpu_mem_time_sum' stats report
+
+ Time (%)  Total Time (ns)  Count   Avg (ns)  Med (ns)  Min (ns)   Max (ns)    StdDev (ns)            Operation           
+ --------  ---------------  ------  --------  --------  --------  -----------  -----------  ------------------------------
+     93.8      627,226,731  42,463  14,771.1     352.0       320  112,333,155    587,539.7  [CUDA memcpy Host-to-Device]  
+      2.8       18,735,373  14,448   1,296.7     928.0       895    1,362,505     22,615.1  [CUDA memcpy Device-to-Device]
+      2.4       16,204,705  24,393     664.3     768.0       320        8,224        282.8  [CUDA memset]                 
+      1.0        6,719,471   5,919   1,135.2   1,120.0       863        1,920        102.9  [CUDA memcpy Device-to-Host]  
+
+[8/8] Executing 'cuda_gpu_mem_size_sum' stats report
+
+ Total (MB)  Count   Avg (MB)  Med (MB)  Min (MB)  Max (MB)  StdDev (MB)            Operation           
+ ----------  ------  --------  --------  --------  --------  -----------  ------------------------------
+  4,194.770  42,463     0.099     0.000     0.000   466.747        2.582  [CUDA memcpy Host-to-Device]  
+  2,533.618  14,448     0.175     0.003     0.000   622.330       10.354  [CUDA memcpy Device-to-Device]
+     17.613  24,393     0.001     0.001     0.000     0.006        0.000  [CUDA memset]                 
+      4.192   5,919     0.001     0.000     0.000     0.004        0.001  [CUDA memcpy Device-to-Host]  
+
+Generated:
+    /data/cy/kv_cache_vs_util/std_traverse_bs/traverse_bs_util_std.nsys-rep
+    /data/cy/kv_cache_vs_util/std_traverse_bs/traverse_bs_util_std.sqlite

std_traverse_bs/traverse_bs_util_std.nsys-rep

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bdc4881551bfa6904baceba8d047d0fc2738308fbadb2d1c3d42a14840041bd6
+size 134872655

std_traverse_bs/traverse_bs_util_std.py

@@ -0,0 +1,262 @@
+import os
+import time
+import statistics
+from typing import List, Tuple, Dict
+
+import torch
+import torch.cuda.nvtx as nvtx
+
+from vllm import LLM, SamplingParams
+from transformers import AutoTokenizer
+
+# ========= 强制使用 vLLM V1 =========
+os.environ.setdefault("VLLM_USE_V1", "1")
+os.environ.setdefault("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
+
+# 可选：打开 V1 metrics 统计
+os.environ.setdefault("VLLM_LOGGING_LEVEL", "INFO")
+
+# ========= 试图导入 V1 metrics 类型（兼容不同版本）=========
+try:
+    from vllm.v1.metrics.reader import Counter, Gauge, Histogram, Vector  # type: ignore
+except Exception:
+    Counter = Gauge = Histogram = Vector = type("X", (), {})  # dummy
+
+# ========= 配置 =========
+MODEL_NAME = "Qwen/Qwen2-1.5B"
+DTYPE = "bfloat16"
+TP = 1
+GPU_MEM_UTIL = 0.90
+TRUST_REMOTE_CODE = True
+
+# 场景：prefill=输入tokens，decode=输出tokens
+SCENARIOS = [
+    # {"name": "prefill640_decode1", "prompt_tokens": 640, "max_new_tokens": 1},
+    # {"name": "prefill1_decode512", "prompt_tokens": 1,   "max_new_tokens": 512},
+    {"name": "prefill640_decode512", "prompt_tokens": 640, "max_new_tokens": 512},
+]
+
+BATCH_SIZES = [1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024]
+
+SEED = 1234
+TEMPERATURE = 0.0
+TOP_P = 1.0
+WARMUP_PER_BS = 1  # 每个批次做一次预热
+
+# ========= 构造“精确 token 数量”的 prompt =========
+def build_exact_token_prompt(tokenizer, target_len: int) -> str:
+    if target_len <= 1:
+        # 最小化 prompt：用一个简单 token（避免空串导致0 token）
+        ids = tokenizer("A", add_special_tokens=False)["input_ids"]
+        if len(ids) >= 1:
+            return tokenizer.decode(ids[:1], skip_special_tokens=True, clean_up_tokenization_spaces=False)
+
+    base_text = (
+        "You are a helpful assistant. "
+        "Please analyze the following input and respond succinctly. "
+    )
+    chunk = " ".join(["data"] * 100) + ". "
+    text = base_text + chunk * 200  # 足够长的文本
+
+    lo, hi = 0, len(text)
+    target_ids = None
+    while lo <= hi:
+        mid = (lo + hi) // 2
+        ids = tokenizer(text[:mid], add_special_tokens=False)["input_ids"]
+        if len(ids) == target_len:
+            target_ids = ids
+            break
+        if len(ids) < target_len:
+            lo = mid + 1
+        else:
+            hi = mid - 1
+
+    if target_ids is None:
+        ids = tokenizer(text[:lo], add_special_tokens=False)["input_ids"]
+        if len(ids) > target_len:
+            target_ids = ids[:target_len]
+        else:
+            filler = " data"
+            while len(ids) < target_len:
+                ids = tokenizer(tokenizer.decode(ids) + filler, add_special_tokens=False)["input_ids"]
+            target_ids = ids[:target_len]
+
+    prompt = tokenizer.decode(target_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)
+    # 断言精确长度
+    assert len(tokenizer(prompt, add_special_tokens=False)["input_ids"]) == target_len
+    return prompt
+
+# ========= V1 metrics 抽取工具 =========
+TTFT_METRIC_NAME = "vllm:time_to_first_token_seconds"
+TPOT_METRIC_NAME = "vllm:time_per_output_token_seconds"  # per-output-token latency
+
+def _iter_children_of_vector(vec_obj):
+    for attr in ("children", "metrics", "series", "values", "samples", "items"):
+        if hasattr(vec_obj, attr):
+            val = getattr(vec_obj, attr)
+            if isinstance(val, dict):
+                for v in val.values():
+                    yield v
+            else:
+                try:
+                    for v in val:
+                        yield v
+                except TypeError:
+                    pass
+
+def _collect_hist_sum_count(metrics, metric_name: str):
+    total_sum = 0.0
+    total_count = 0.0
+    for m in metrics:
+        mname = getattr(m, "name", None)
+        if mname != metric_name:
+            continue
+        # 直接 Histogram
+        if isinstance(m, Histogram) or m.__class__.__name__ == "Histogram":
+            total_sum += float(getattr(m, "sum", 0.0))
+            total_count += float(getattr(m, "count", 0.0))
+            continue
+        # Vector[Histogram]
+        if isinstance(m, Vector) or m.__class__.__name__ == "Vector":
+            for child in _iter_children_of_vector(m):
+                if isinstance(child, Histogram) or child.__class__.__name__ == "Histogram":
+                    total_sum += float(getattr(child, "sum", 0.0))
+                    total_count += float(getattr(child, "count", 0.0))
+    return total_sum, total_count
+
+def _metrics_snapshot(llm) -> Dict[str, float]:
+    try:
+        mets = llm.get_metrics()  # V1: 返回 Metric 列表（包含 Histogram/Vector 等）
+    except Exception:
+        return {"ttft_sum": 0.0, "ttft_cnt": 0.0, "tpot_sum": 0.0, "tpot_cnt": 0.0}
+    ttft_sum, ttft_cnt = _collect_hist_sum_count(mets, TTFT_METRIC_NAME)
+    tpot_sum, tpot_cnt = _collect_hist_sum_count(mets, TPOT_METRIC_NAME)
+    return {"ttft_sum": ttft_sum, "ttft_cnt": ttft_cnt, "tpot_sum": tpot_sum, "tpot_cnt": tpot_cnt}
+
+def _metrics_delta(before: dict, after: dict):
+    return {
+        "ttft_sum": after["ttft_sum"] - before["ttft_sum"],
+        "ttft_cnt": after["ttft_cnt"] - before["ttft_cnt"],
+        "tpot_sum": after["tpot_sum"] - before["tpot_sum"],
+        "tpot_cnt": after["tpot_cnt"] - before["tpot_cnt"],
+    }
+
+# ========= 带 NVTX 的 generate 包装 =========
+def decorated_generate(llm: LLM, prompts: List[str], params: SamplingParams):
+    return llm.generate(prompts, params)
+
+# ========= 统计格式化 =========
+def fmt_stats(x: List[float]) -> Tuple[float, float, float]:
+    xs = [v for v in x if (v == v)]  # 过滤 NaN
+    if not xs:
+        return (float("nan"), float("nan"), float("nan"))
+    return (statistics.mean(xs), statistics.median(xs), statistics.quantiles(xs, n=10)[-1])  # p90
+
+def main():
+    print("--- vLLM V1 基准测试（含 NVTX 标记）---")
+    print(f"模型: {MODEL_NAME}")
+    print(f"批量大小: {BATCH_SIZES}")
+    print(f"场景: {[s['name'] for s in SCENARIOS]}")
+    print("-" * 60)
+
+    if not torch.cuda.is_available():
+        print("错误：需要 CUDA GPU。")
+        return
+
+    print("加载分词器/模型中...")
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, use_fast=True, trust_remote_code=TRUST_REMOTE_CODE)
+
+    # 用 NVTX 标记模型加载阶段
+    nvtx.range_push("LLM_init")
+    llm = LLM(
+        model=MODEL_NAME,
+        tensor_parallel_size=TP,
+        dtype=DTYPE,
+        trust_remote_code=TRUST_REMOTE_CODE,
+        gpu_memory_utilization=GPU_MEM_UTIL,
+        max_num_seqs=1024,         # 足够覆盖本次扫描
+        max_model_len=4096,
+        disable_log_stats=False,  # 开启 V1 metrics 收集
+    )
+    nvtx.range_pop()
+    print("模型加载完成。")
+
+    for sc in SCENARIOS:
+        name = sc["name"]
+        prompt_tokens = sc["prompt_tokens"]
+        max_new_tokens = sc["max_new_tokens"]
+
+        print(f"\n===== 场景：{name} | prefill={prompt_tokens}, decode={max_new_tokens} =====")
+
+        # 准备精确长度 prompt
+        prompt_text = build_exact_token_prompt(tokenizer, prompt_tokens)
+
+        # 采样参数（贪心）
+        sampling_params = SamplingParams(
+            max_tokens=max_new_tokens,
+            temperature=TEMPERATURE,
+            top_p=TOP_P,
+            seed=SEED,
+            n=1,
+        )
+
+        # 记录每个 bs 的结果（便于后续统计或外部解析）
+        for bs in BATCH_SIZES:
+            print(f"\n--- 批量大小 bs={bs} ---")
+
+            prompts = [prompt_text] * bs
+
+            # 预热
+            # print("预热中...")
+            # nvtx.range_push(f"WARMUP [{name}] bs={bs}")
+            # _ = decorated_generate(llm, [prompts[0]], sampling_params)
+            # torch.cuda.synchronize()
+            # nvtx.range_pop()
+
+            # 正式计时与 V1 metrics
+            # nvtx.range_push(f"RUN [{name}] bs={bs}")
+            torch.cuda.synchronize()
+            snap_before = _metrics_snapshot(llm)
+            t0 = time.perf_counter()
+
+            nvtx.range_push(f"generate [{name}] bs={bs}")
+            outputs = decorated_generate(llm, prompts, sampling_params)
+            nvtx.range_pop()  # generate
+
+            torch.cuda.synchronize()
+            t1 = time.perf_counter()
+            snap_after = _metrics_snapshot(llm)
+            # nvtx.range_pop()  # RUN
+
+            duration = t1 - t0
+
+            # 统计 token 与吞吐
+            total_output_tokens = sum(len(o.outputs[0].token_ids) for o in outputs)
+            avg_prompt_tokens = sum(len(o.prompt_token_ids) for o in outputs) / bs
+            throughput = total_output_tokens / duration if duration > 0 else float("inf")
+
+            # 解析 V1 TTFT / 解码吞吐
+            delta = _metrics_delta(snap_before, snap_after)
+            if delta["ttft_cnt"] > 0:
+                ttft = delta["ttft_sum"] / delta["ttft_cnt"]
+            else:
+                ttft = float("nan")
+
+            if delta["tpot_cnt"] > 0:
+                avg_tpot = delta["tpot_sum"] / delta["tpot_cnt"]  # seconds/token
+                decode_tps = 1.0 / avg_tpot
+            else:
+                decode_tps = float("nan")
+
+            print(f"执行时间: {duration:.4f} s")
+            print(f"实际平均输入 tokens: {avg_prompt_tokens:.2f}（目标 {prompt_tokens}）")
+            print(f"生成总 tokens: {total_output_tokens}")
+            print(f"吞吐(生成tokens/秒): {throughput:.2f}")
+            print(f"TTFT (V1 metrics): {ttft:.4f} s")
+            print(f"解码吞吐 (V1 metrics): {decode_tps:.2f} tok/s")
+
+    print("\n完成。提示：在 Nsight Systems 中可通过 NVTX 区间快速定位各场景/批量的调用。")
+
+if __name__ == "__main__":
+    print(f"CUDA_VISIBLE_DEVICES = {os.getenv('CUDA_VISIBLE_DEVICES')}")
+    main()

traverse_bs_util.py

@@ -0,0 +1,261 @@
+import os
+import time
+import statistics
+from typing import List, Tuple, Dict
+
+import torch
+import torch.cuda.nvtx as nvtx
+
+from vllm import LLM, SamplingParams
+from transformers import AutoTokenizer
+
+# ========= 强制使用 vLLM V1 =========
+os.environ.setdefault("VLLM_USE_V1", "1")
+os.environ.setdefault("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
+
+# 可选：打开 V1 metrics 统计
+os.environ.setdefault("VLLM_LOGGING_LEVEL", "INFO")
+
+# ========= 试图导入 V1 metrics 类型（兼容不同版本）=========
+try:
+    from vllm.v1.metrics.reader import Counter, Gauge, Histogram, Vector  # type: ignore
+except Exception:
+    Counter = Gauge = Histogram = Vector = type("X", (), {})  # dummy
+
+# ========= 配置 =========
+MODEL_NAME = "Qwen/Qwen2-1.5B"
+DTYPE = "bfloat16"
+TP = 1
+GPU_MEM_UTIL = 0.90
+TRUST_REMOTE_CODE = True
+
+# 场景：prefill=输入tokens，decode=输出tokens
+SCENARIOS = [
+    {"name": "prefill640_decode1", "prompt_tokens": 640, "max_new_tokens": 1},
+    {"name": "prefill1_decode512", "prompt_tokens": 1,   "max_new_tokens": 512},
+]
+
+BATCH_SIZES = [1, 2, 4, 8, 16, 32, 64, 128, 256]
+
+SEED = 1234
+TEMPERATURE = 0.0
+TOP_P = 1.0
+WARMUP_PER_BS = 1  # 每个批次做一次预热
+
+# ========= 构造“精确 token 数量”的 prompt =========
+def build_exact_token_prompt(tokenizer, target_len: int) -> str:
+    if target_len <= 1:
+        # 最小化 prompt：用一个简单 token（避免空串导致0 token）
+        ids = tokenizer("A", add_special_tokens=False)["input_ids"]
+        if len(ids) >= 1:
+            return tokenizer.decode(ids[:1], skip_special_tokens=True, clean_up_tokenization_spaces=False)
+
+    base_text = (
+        "You are a helpful assistant. "
+        "Please analyze the following input and respond succinctly. "
+    )
+    chunk = " ".join(["data"] * 100) + ". "
+    text = base_text + chunk * 200  # 足够长的文本
+
+    lo, hi = 0, len(text)
+    target_ids = None
+    while lo <= hi:
+        mid = (lo + hi) // 2
+        ids = tokenizer(text[:mid], add_special_tokens=False)["input_ids"]
+        if len(ids) == target_len:
+            target_ids = ids
+            break
+        if len(ids) < target_len:
+            lo = mid + 1
+        else:
+            hi = mid - 1
+
+    if target_ids is None:
+        ids = tokenizer(text[:lo], add_special_tokens=False)["input_ids"]
+        if len(ids) > target_len:
+            target_ids = ids[:target_len]
+        else:
+            filler = " data"
+            while len(ids) < target_len:
+                ids = tokenizer(tokenizer.decode(ids) + filler, add_special_tokens=False)["input_ids"]
+            target_ids = ids[:target_len]
+
+    prompt = tokenizer.decode(target_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)
+    # 断言精确长度
+    assert len(tokenizer(prompt, add_special_tokens=False)["input_ids"]) == target_len
+    return prompt
+
+# ========= V1 metrics 抽取工具 =========
+TTFT_METRIC_NAME = "vllm:time_to_first_token_seconds"
+TPOT_METRIC_NAME = "vllm:time_per_output_token_seconds"  # per-output-token latency
+
+def _iter_children_of_vector(vec_obj):
+    for attr in ("children", "metrics", "series", "values", "samples", "items"):
+        if hasattr(vec_obj, attr):
+            val = getattr(vec_obj, attr)
+            if isinstance(val, dict):
+                for v in val.values():
+                    yield v
+            else:
+                try:
+                    for v in val:
+                        yield v
+                except TypeError:
+                    pass
+
+def _collect_hist_sum_count(metrics, metric_name: str):
+    total_sum = 0.0
+    total_count = 0.0
+    for m in metrics:
+        mname = getattr(m, "name", None)
+        if mname != metric_name:
+            continue
+        # 直接 Histogram
+        if isinstance(m, Histogram) or m.__class__.__name__ == "Histogram":
+            total_sum += float(getattr(m, "sum", 0.0))
+            total_count += float(getattr(m, "count", 0.0))
+            continue
+        # Vector[Histogram]
+        if isinstance(m, Vector) or m.__class__.__name__ == "Vector":
+            for child in _iter_children_of_vector(m):
+                if isinstance(child, Histogram) or child.__class__.__name__ == "Histogram":
+                    total_sum += float(getattr(child, "sum", 0.0))
+                    total_count += float(getattr(child, "count", 0.0))
+    return total_sum, total_count
+
+def _metrics_snapshot(llm) -> Dict[str, float]:
+    try:
+        mets = llm.get_metrics()  # V1: 返回 Metric 列表（包含 Histogram/Vector 等）
+    except Exception:
+        return {"ttft_sum": 0.0, "ttft_cnt": 0.0, "tpot_sum": 0.0, "tpot_cnt": 0.0}
+    ttft_sum, ttft_cnt = _collect_hist_sum_count(mets, TTFT_METRIC_NAME)
+    tpot_sum, tpot_cnt = _collect_hist_sum_count(mets, TPOT_METRIC_NAME)
+    return {"ttft_sum": ttft_sum, "ttft_cnt": ttft_cnt, "tpot_sum": tpot_sum, "tpot_cnt": tpot_cnt}
+
+def _metrics_delta(before: dict, after: dict):
+    return {
+        "ttft_sum": after["ttft_sum"] - before["ttft_sum"],
+        "ttft_cnt": after["ttft_cnt"] - before["ttft_cnt"],
+        "tpot_sum": after["tpot_sum"] - before["tpot_sum"],
+        "tpot_cnt": after["tpot_cnt"] - before["tpot_cnt"],
+    }
+
+# ========= 带 NVTX 的 generate 包装 =========
+def decorated_generate(llm: LLM, prompts: List[str], params: SamplingParams):
+    return llm.generate(prompts, params)
+
+# ========= 统计格式化 =========
+def fmt_stats(x: List[float]) -> Tuple[float, float, float]:
+    xs = [v for v in x if (v == v)]  # 过滤 NaN
+    if not xs:
+        return (float("nan"), float("nan"), float("nan"))
+    return (statistics.mean(xs), statistics.median(xs), statistics.quantiles(xs, n=10)[-1])  # p90
+
+def main():
+    print("--- vLLM V1 基准测试（含 NVTX 标记）---")
+    print(f"模型: {MODEL_NAME}")
+    print(f"批量大小: {BATCH_SIZES}")
+    print(f"场景: {[s['name'] for s in SCENARIOS]}")
+    print("-" * 60)
+
+    if not torch.cuda.is_available():
+        print("错误：需要 CUDA GPU。")
+        return
+
+    print("加载分词器/模型中...")
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, use_fast=True, trust_remote_code=TRUST_REMOTE_CODE)
+
+    # 用 NVTX 标记模型加载阶段
+    nvtx.range_push("LLM_init")
+    llm = LLM(
+        model=MODEL_NAME,
+        tensor_parallel_size=TP,
+        dtype=DTYPE,
+        trust_remote_code=TRUST_REMOTE_CODE,
+        gpu_memory_utilization=GPU_MEM_UTIL,
+        max_num_seqs=256,         # 足够覆盖本次扫描
+        max_model_len=8192,
+        disable_log_stats=False,  # 开启 V1 metrics 收集
+    )
+    nvtx.range_pop()
+    print("模型加载完成。")
+
+    for sc in SCENARIOS:
+        name = sc["name"]
+        prompt_tokens = sc["prompt_tokens"]
+        max_new_tokens = sc["max_new_tokens"]
+
+        print(f"\n===== 场景：{name} | prefill={prompt_tokens}, decode={max_new_tokens} =====")
+
+        # 准备精确长度 prompt
+        prompt_text = build_exact_token_prompt(tokenizer, prompt_tokens)
+
+        # 采样参数（贪心）
+        sampling_params = SamplingParams(
+            max_tokens=max_new_tokens,
+            temperature=TEMPERATURE,
+            top_p=TOP_P,
+            seed=SEED,
+            n=1,
+        )
+
+        # 记录每个 bs 的结果（便于后续统计或外部解析）
+        for bs in BATCH_SIZES:
+            print(f"\n--- 批量大小 bs={bs} ---")
+
+            prompts = [prompt_text] * bs
+
+            # 预热
+            print("预热中...")
+            nvtx.range_push(f"WARMUP [{name}] bs={bs}")
+            _ = decorated_generate(llm, [prompts[0]], sampling_params)
+            torch.cuda.synchronize()
+            nvtx.range_pop()
+
+            # 正式计时与 V1 metrics
+            nvtx.range_push(f"RUN [{name}] bs={bs}")
+            torch.cuda.synchronize()
+            snap_before = _metrics_snapshot(llm)
+            t0 = time.perf_counter()
+
+            nvtx.range_push(f"generate [{name}] bs={bs}")
+            outputs = decorated_generate(llm, prompts, sampling_params)
+            nvtx.range_pop()  # generate
+
+            torch.cuda.synchronize()
+            t1 = time.perf_counter()
+            snap_after = _metrics_snapshot(llm)
+            nvtx.range_pop()  # RUN
+
+            duration = t1 - t0
+
+            # 统计 token 与吞吐
+            total_output_tokens = sum(len(o.outputs[0].token_ids) for o in outputs)
+            avg_prompt_tokens = sum(len(o.prompt_token_ids) for o in outputs) / bs
+            throughput = total_output_tokens / duration if duration > 0 else float("inf")
+
+            # 解析 V1 TTFT / 解码吞吐
+            delta = _metrics_delta(snap_before, snap_after)
+            if delta["ttft_cnt"] > 0:
+                ttft = delta["ttft_sum"] / delta["ttft_cnt"]
+            else:
+                ttft = float("nan")
+
+            if delta["tpot_cnt"] > 0:
+                avg_tpot = delta["tpot_sum"] / delta["tpot_cnt"]  # seconds/token
+                decode_tps = 1.0 / avg_tpot
+            else:
+                decode_tps = float("nan")
+
+            print(f"执行时间: {duration:.4f} s")
+            print(f"实际平均输入 tokens: {avg_prompt_tokens:.2f}（目标 {prompt_tokens}）")
+            print(f"生成总 tokens: {total_output_tokens}")
+            print(f"吞吐(生成tokens/秒): {throughput:.2f}")
+            print(f"TTFT (V1 metrics): {ttft:.4f} s")
+            print(f"解码吞吐 (V1 metrics): {decode_tps:.2f} tok/s")
+
+    print("\n完成。提示：在 Nsight Systems 中可通过 NVTX 区间快速定位各场景/批量的调用。")
+
+if __name__ == "__main__":
+    print(f"CUDA_VISIBLE_DEVICES = {os.getenv('CUDA_VISIBLE_DEVICES')}")
+    main()

traverse_bs_util_std_kvcache.py

@@ -0,0 +1,452 @@
+import os
+import time
+import statistics
+from typing import List, Tuple, Dict
+
+import torch
+import torch.cuda.nvtx as nvtx
+
+from vllm import LLM, SamplingParams
+from transformers import AutoTokenizer
+
+# ========= 强制使用 vLLM V1 =========
+os.environ.setdefault("VLLM_USE_V1", "1")
+os.environ.setdefault("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
+
+# 可选：打开 V1 metrics 统计
+os.environ.setdefault("VLLM_LOGGING_LEVEL", "INFO")
+
+# ========= 试图导入 V1 metrics 类型（兼容不同版本）=========
+try:
+    from vllm.v1.metrics.reader import Counter, Gauge, Histogram, Vector  # type: ignore
+except Exception:
+    Counter = Gauge = Histogram = Vector = type("X", (), {})  # dummy
+
+# ========= 配置 =========
+MODEL_NAME = "Qwen/Qwen2-1.5B"
+DTYPE = "bfloat16"
+TP = 1
+GPU_MEM_UTIL = 0.90
+TRUST_REMOTE_CODE = True
+
+# 场景：prefill=输入tokens，decode=输出tokens
+SCENARIOS = [
+    # {"name": "prefill640_decode1", "prompt_tokens": 640, "max_new_tokens": 1},
+    # {"name": "prefill1_decode512", "prompt_tokens": 1,   "max_new_tokens": 512},
+    {"name": "prefill640_decode512", "prompt_tokens": 640, "max_new_tokens": 512},
+]
+
+BATCH_SIZES = [1, 2, 4, 8, 16, 32, 64, 128, 256]
+
+SEED = 1234
+TEMPERATURE = 0.0
+TOP_P = 1.0
+WARMUP_PER_BS = 1  # 每个批次做一次预热
+
+# ========= 构造“精确 token 数量”的 prompt =========
+def build_exact_token_prompt(tokenizer, target_len: int) -> str:
+    if target_len <= 1:
+        # 最小化 prompt：用一个简单 token（避免空串导致0 token）
+        ids = tokenizer("A", add_special_tokens=False)["input_ids"]
+        if len(ids) >= 1:
+            return tokenizer.decode(ids[:1], skip_special_tokens=True, clean_up_tokenization_spaces=False)
+
+    base_text = (
+        "You are a helpful assistant. "
+        "Please analyze the following input and respond succinctly. "
+    )
+    chunk = " ".join(["data"] * 100) + ". "
+    text = base_text + chunk * 200  # 足够长的文本
+
+    lo, hi = 0, len(text)
+    target_ids = None
+    while lo <= hi:
+        mid = (lo + hi) // 2
+        ids = tokenizer(text[:mid], add_special_tokens=False)["input_ids"]
+        if len(ids) == target_len:
+            target_ids = ids
+            break
+        if len(ids) < target_len:
+            lo = mid + 1
+        else:
+            hi = mid - 1
+
+    if target_ids is None:
+        ids = tokenizer(text[:lo], add_special_tokens=False)["input_ids"]
+        if len(ids) > target_len:
+            target_ids = ids[:target_len]
+        else:
+            filler = " data"
+            while len(ids) < target_len:
+                ids = tokenizer(tokenizer.decode(ids) + filler, add_special_tokens=False)["input_ids"]
+            target_ids = ids[:target_len]
+
+    prompt = tokenizer.decode(target_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)
+    # 断言精确长度
+    assert len(tokenizer(prompt, add_special_tokens=False)["input_ids"]) == target_len
+    return prompt
+
+# ========= V1 metrics 抽取工具 =========
+TTFT_METRIC_NAME = "vllm:time_to_first_token_seconds"
+TPOT_METRIC_NAME = "vllm:time_per_output_token_seconds"  # per-output-token latency
+
+def _iter_children_of_vector(vec_obj):
+    for attr in ("children", "metrics", "series", "values", "samples", "items"):
+        if hasattr(vec_obj, attr):
+            val = getattr(vec_obj, attr)
+            if isinstance(val, dict):
+                for v in val.values():
+                    yield v
+            else:
+                try:
+                    for v in val:
+                        yield v
+                except TypeError:
+                    pass
+
+def _collect_hist_sum_count(metrics, metric_name: str):
+    total_sum = 0.0
+    total_count = 0.0
+    for m in metrics:
+        mname = getattr(m, "name", None)
+        if mname != metric_name:
+            continue
+        # 直接 Histogram
+        if isinstance(m, Histogram) or m.__class__.__name__ == "Histogram":
+            total_sum += float(getattr(m, "sum", 0.0))
+            total_count += float(getattr(m, "count", 0.0))
+            continue
+        # Vector[Histogram]
+        if isinstance(m, Vector) or m.__class__.__name__ == "Vector":
+            for child in _iter_children_of_vector(m):
+                if isinstance(child, Histogram) or child.__class__.__name__ == "Histogram":
+                    total_sum += float(getattr(child, "sum", 0.0))
+                    total_count += float(getattr(child, "count", 0.0))
+    return total_sum, total_count
+
+def _metrics_snapshot(llm) -> Dict[str, float]:
+    try:
+        mets = llm.get_metrics()  # V1: 返回 Metric 列表（包含 Histogram/Vector 等）
+    except Exception:
+        return {"ttft_sum": 0.0, "ttft_cnt": 0.0, "tpot_sum": 0.0, "tpot_cnt": 0.0}
+    ttft_sum, ttft_cnt = _collect_hist_sum_count(mets, TTFT_METRIC_NAME)
+    tpot_sum, tpot_cnt = _collect_hist_sum_count(mets, TPOT_METRIC_NAME)
+    return {"ttft_sum": ttft_sum, "ttft_cnt": ttft_cnt, "tpot_sum": tpot_sum, "tpot_cnt": tpot_cnt}
+
+def _metrics_delta(before: dict, after: dict):
+    return {
+        "ttft_sum": after["ttft_sum"] - before["ttft_sum"],
+        "ttft_cnt": after["ttft_cnt"] - before["ttft_cnt"],
+        "tpot_sum": after["tpot_sum"] - before["tpot_sum"],
+        "tpot_cnt": after["tpot_cnt"] - before["tpot_cnt"],
+    }
+
+# ========= 内存 / KV cache 统计辅助 =========
+def _try_import_pynvml():
+    try:
+        import pynvml  # type: ignore
+        pynvml.nvmlInit()
+        return pynvml
+    except Exception:
+        return None
+
+def _gpu_mem_stats(device_idx: int = 0):
+    # PyTorch 视角
+    try:
+        torch.cuda.synchronize(device_idx)
+        allocated = torch.cuda.memory_allocated(device_idx)
+        reserved  = torch.cuda.memory_reserved(device_idx)
+        max_alloc = torch.cuda.max_memory_allocated(device_idx)
+        max_resv  = torch.cuda.max_memory_reserved(device_idx)
+    except Exception:
+        allocated = reserved = max_alloc = max_resv = float("nan")
+
+    # NVML 视角（接近 nvidia-smi）
+    pynvml = _try_import_pynvml()
+    nvml_total = nvml_used = nvml_free = float("nan")
+    if pynvml is not None:
+        try:
+            h = pynvml.nvmlDeviceGetHandleByIndex(device_idx)
+            mem = pynvml.nvmlDeviceGetMemoryInfo(h)
+            nvml_total, nvml_used, nvml_free = float(mem.total), float(mem.used), float(mem.free)
+        except Exception:
+            pass
+
+    def _fmt(b):
+        # 以 MiB 显示
+        if b != b:  # NaN
+            return "NaN"
+        return f"{b/1024/1024:.1f} MiB"
+
+    return {
+        "torch_allocated": _fmt(allocated),
+        "torch_reserved":  _fmt(reserved),
+        "torch_max_alloc": _fmt(max_alloc),
+        "torch_max_resvd": _fmt(max_resv),
+        "nvml_total":      _fmt(nvml_total),
+        "nvml_used":       _fmt(nvml_used),
+        "nvml_free":       _fmt(nvml_free),
+    }
+
+# —— vLLM v1 metrics 中收集 Gauge/Counter 最新值（含 Vector 包装）——
+def _collect_latest_metric(metrics, metric_name: str):
+    latest = None
+    for m in metrics:
+        mname = getattr(m, "name", None)
+        is_vector = getattr(m, "__class__", type("X", (), {})).__name__ in ("Vector",)
+        if mname != metric_name and not (is_vector and any(getattr(child, "name", None) == metric_name for child in _iter_children_of_vector(m))):
+            continue
+        # 直接 Gauge/Counter
+        if m.__class__.__name__ in ("Gauge", "Counter"):
+            val = getattr(m, "value", None)
+            if val is None:
+                val = getattr(m, "_value", None)
+            if isinstance(val, (int, float)):
+                latest = float(val)
+        # Vector[Gauge/Counter]
+        if is_vector:
+            for child in _iter_children_of_vector(m):
+                if getattr(child, "name", None) == metric_name and child.__class__.__name__ in ("Gauge", "Counter"):
+                    val = getattr(child, "value", None)
+                    if val is None:
+                        val = getattr(child, "_value", None)
+                    if isinstance(val, (int, float)):
+                        latest = float(val)
+    return latest
+
+# —— 从 metrics 中“尽力而为”获取 KV cache 使用情况 —— 
+def _kv_cache_stats_from_metrics(llm):
+    try:
+        mets = llm.get_metrics()
+    except Exception:
+        return None
+
+    # 常见/候选指标名（不同版本命名可能略有差异，尽可能覆盖）
+    names_used_blocks = [
+        "vllm:kv_cache_gpu_blocks_in_use",
+        "vllm:kv_cache:gpu_blocks_in_use",
+        "vllm:cache:gpu_blocks_in_use",
+        "vllm:num_gpu_blocks_in_use",
+    ]
+    names_total_blocks = [
+        "vllm:kv_cache_gpu_blocks_total",
+        "vllm:kv_cache:gpu_blocks_total",
+        "vllm:cache:gpu_blocks_total",
+        "vllm:num_gpu_blocks_total",
+    ]
+    names_used_bytes = [
+        "vllm:kv_cache_usage_bytes",
+        "vllm:kv_cache:usage_bytes",
+        "vllm:cache:kv_bytes_used",
+    ]
+    used_blocks = None
+    total_blocks = None
+    used_bytes   = None
+
+    for n in names_used_blocks:
+        v = _collect_latest_metric(mets, n)
+        if v is not None:
+            used_blocks = v; break
+    for n in names_total_blocks:
+        v = _collect_latest_metric(mets, n)
+        if v is not None:
+            total_blocks = v; break
+    for n in names_used_bytes:
+        v = _collect_latest_metric(mets, n)
+        if v is not None:
+            used_bytes = v; break
+
+    util = None
+    if used_blocks is not None and total_blocks and total_blocks > 0:
+        util = used_blocks / total_blocks
+
+    def _fmt_bytes(b):
+        if b is None: return None
+        return f"{b/1024/1024:.1f} MiB"
+
+    return {
+        "used_blocks": used_blocks,
+        "total_blocks": total_blocks,
+        "utilization": (f"{util*100:.1f}%" if util is not None else None),
+        "used_bytes": _fmt_bytes(used_bytes),
+    }
+
+# —— 回退：从内部引擎尝试估算 KV cache（不同版本可能不兼容，失败就返回 None）——
+def _kv_cache_stats_from_engine(llm):
+    try:
+        eng = getattr(llm, "llm_engine", None)
+        if eng is None:
+            return None
+        # 常见位置：eng.scheduler.cache_config 或 eng.cache_config
+        cfg = getattr(getattr(eng, "scheduler", eng), "cache_config", None) or getattr(eng, "cache_config", None)
+        if cfg is None:
+            return None
+        num_gpu_blocks = getattr(cfg, "num_gpu_blocks", None)
+        block_size     = getattr(cfg, "block_size", None)  # 通常为 tokens/块
+        # 注：准确的字节大小依赖实现细节（层数、头数、dtype等），这里仅报告 blocks 维度，避免误导
+        return {
+            "used_blocks": None,           # 无法直接从 config 得到“已用”，只知道上限
+            "total_blocks": float(num_gpu_blocks) if num_gpu_blocks is not None else None,
+            "utilization": None,
+            "used_bytes": None,
+            "notes": f"block_size={block_size} (tokens per block)" if block_size is not None else None,
+        }
+    except Exception:
+        return None
+
+def _print_mem_and_kv(llm, header: str):
+    print(f"[MEM/KV] {header}")
+    # GPU 内存
+    mem = _gpu_mem_stats(0)
+    print(f"  PyTorch allocated/reserved: {mem['torch_allocated']} / {mem['torch_reserved']} "
+          f"(peak {mem['torch_max_alloc']} / {mem['torch_max_resvd']})")
+    if mem["nvml_total"] != "NaN":
+        print(f"  NVML total/used/free: {mem['nvml_total']} / {mem['nvml_used']} / {mem['nvml_free']}")
+
+    # KV cache（优先从 metrics；不行再从引擎配置）
+    kv = _kv_cache_stats_from_metrics(llm)
+    if kv is None:
+        kv = _kv_cache_stats_from_engine(llm)
+    if kv is None:
+        print("  KV cache: 未能获取（该 vLLM 版本可能未暴露对应指标）")
+    else:
+        parts = []
+        if kv.get("used_blocks") is not None:
+            parts.append(f"used_blocks={int(kv['used_blocks'])}")
+        if kv.get("total_blocks") is not None:
+            parts.append(f"total_blocks={int(kv['total_blocks'])}")
+        if kv.get("utilization") is not None:
+            parts.append(f"util={kv['utilization']}")
+        if kv.get("used_bytes") is not None:
+            parts.append(f"used≈{kv['used_bytes']}")
+        if kv.get("notes"):
+            parts.append(kv["notes"])
+        if parts:
+            print("  KV cache: " + ", ".join(parts))
+        else:
+            print("  KV cache: 指标存在但内容为空/不兼容")
+
+# ========= 带 NVTX 的 generate 包装 =========
+def decorated_generate(llm: LLM, prompts: List[str], params: SamplingParams):
+    return llm.generate(prompts, params)
+
+# ========= 统计格式化 =========
+def fmt_stats(x: List[float]) -> Tuple[float, float, float]:
+    xs = [v for v in x if (v == v)]  # 过滤 NaN
+    if not xs:
+        return (float("nan"), float("nan"), float("nan"))
+    return (statistics.mean(xs), statistics.median(xs), statistics.quantiles(xs, n=10)[-1])  # p90
+
+def main():
+    print("--- vLLM V1 基准测试（含 NVTX 标记）---")
+    print(f"模型: {MODEL_NAME}")
+    print(f"批量大小: {BATCH_SIZES}")
+    print(f"场景: {[s['name'] for s in SCENARIOS]}")
+    print("-" * 60)
+
+    if not torch.cuda.is_available():
+        print("错误：需要 CUDA GPU。")
+        return
+
+    print("加载分词器/模型中...")
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, use_fast=True, trust_remote_code=TRUST_REMOTE_CODE)
+
+    # 用 NVTX 标记模型加载阶段
+    nvtx.range_push("LLM_init")
+    llm = LLM(
+        model=MODEL_NAME,
+        tensor_parallel_size=TP,
+        dtype=DTYPE,
+        trust_remote_code=TRUST_REMOTE_CODE,
+        gpu_memory_utilization=GPU_MEM_UTIL,
+        max_num_seqs=256,         # 足够覆盖本次扫描
+        max_model_len=8192,
+        disable_log_stats=False,  # 开启 V1 metrics 收集
+    )
+    nvtx.range_pop()
+    print("模型加载完成。")
+    # —— 新增：打印一次基线内存/KV 状态
+    _print_mem_and_kv(llm, "after LLM_init")
+
+    for sc in SCENARIOS:
+        name = sc["name"]
+        prompt_tokens = sc["prompt_tokens"]
+        max_new_tokens = sc["max_new_tokens"]
+
+        print(f"\n===== 场景：{name} | prefill={prompt_tokens}, decode={max_new_tokens} =====")
+
+        # 准备精确长度 prompt
+        prompt_text = build_exact_token_prompt(tokenizer, prompt_tokens)
+
+        # 采样参数（贪心）
+        sampling_params = SamplingParams(
+            max_tokens=max_new_tokens,
+            temperature=TEMPERATURE,
+            top_p=TOP_P,
+            seed=SEED,
+            n=1,
+        )
+
+        # 记录每个 bs 的结果（便于后续统计或外部解析）
+        for bs in BATCH_SIZES:
+            print(f"\n--- 批量大小 bs={bs} ---")
+
+            prompts = [prompt_text] * bs
+
+            # 预热（如需）
+            # print("预热中...")
+            # nvtx.range_push(f"WARMUP [{name}] bs={bs}")
+            # _ = decorated_generate(llm, [prompts[0]], sampling_params)
+            # torch.cuda.synchronize()
+            # nvtx.range_pop()
+
+            # 正式计时与 V1 metrics
+            # nvtx.range_push(f"RUN [{name}] bs={bs}")
+            torch.cuda.synchronize()
+            snap_before = _metrics_snapshot(llm)
+            t0 = time.perf_counter()
+
+            nvtx.range_push(f"generate [{name}] bs={bs}")
+            outputs = decorated_generate(llm, prompts, sampling_params)
+            nvtx.range_pop()  # generate
+
+            torch.cuda.synchronize()
+            t1 = time.perf_counter()
+            snap_after = _metrics_snapshot(llm)
+            # nvtx.range_pop()  # RUN
+
+            duration = t1 - t0
+
+            # 统计 token 与吞吐
+            total_output_tokens = sum(len(o.outputs[0].token_ids) for o in outputs)
+            avg_prompt_tokens = sum(len(o.prompt_token_ids) for o in outputs) / bs
+            throughput = total_output_tokens / duration if duration > 0 else float("inf")
+
+            # 解析 V1 TTFT / 解码吞吐
+            delta = _metrics_delta(snap_before, snap_after)
+            if delta["ttft_cnt"] > 0:
+                ttft = delta["ttft_sum"] / delta["ttft_cnt"]
+            else:
+                ttft = float("nan")
+
+            if delta["tpot_cnt"] > 0:
+                avg_tpot = delta["tpot_sum"] / delta["tpot_cnt"]  # seconds/token
+                decode_tps = 1.0 / avg_tpot
+            else:
+                decode_tps = float("nan")
+
+            print(f"执行时间: {duration:.4f} s")
+            print(f"实际平均输入 tokens: {avg_prompt_tokens:.2f}（目标 {prompt_tokens}）")
+            print(f"生成总 tokens: {total_output_tokens}")
+            print(f"吞吐(生成tokens/秒): {throughput:.2f}")
+            print(f"TTFT (V1 metrics): {ttft:.4f} s")
+            print(f"解码吞吐 (V1 metrics): {decode_tps:.2f} tok/s")
+
+            # —— 新增：打印该批次后内存 / KV cache 状态
+            _print_mem_and_kv(llm, f"after generate [{name}] bs={bs}")
+
+    print("\n完成。提示：在 Nsight Systems 中可通过 NVTX 区间快速定位各场景/批量的调用。")
+
+if __name__ == "__main__":
+    print(f"CUDA_VISIBLE_DEVICES = {os.getenv('CUDA_VISIBLE_DEVICES')}")
+    main()

常用命令

@@ -0,0 +1,23 @@
+hf upload Hamerlate/tts . --repo-type=dataset
+
+nsys profile -w true -t cuda,nvtx,osrt --force-overwrite=true --stats=true  --gpu-metrics-device=3 -x true -o traverse_bs_util_std python3 /data/cy/kv_cache_vs_util/traverse_bs_util_std.py > traverse_bs_util_std.log 2>&1
+
+export CUDA_VISIBLE_DEVICES=3
+
+nsys profile -w true -t cuda,nvtx,osrt --force-overwrite=true --stats=true  --gpu-metrics-device=3 -x true -o traverse_bs_util_std python3 /data/cy/kv_cache_vs_util/traverse_bs_util_std.py > traverse_bs_util_std.log 2>&1
+
+nsys profile -w true -t cuda,nvtx,osrt --force-overwrite=true --stats=true  --gpu-metrics-device=3 -x true -o traverse_bs_util_std python3 /data/cy/kv_cache_vs_util/std_traverse_bs/traverse_bs_util_std.py > traverse_bs_util_std.log 2>&1
+
+nsys profile -w true -t cuda,nvtx,osrt --force-overwrite=true --stats=true  --gpu-metrics-device=3 -x true -o traverse_bs_util_sim_prefill python3 /data/cy/kv_cache_vs_util/sim_traverse_bs/traverse_bs_util_sim_prefill.py > traverse_bs_util_sim_prefill.log 2>&1
+
+nsys profile -w true -t cuda,nvtx,osrt --force-overwrite=true --stats=true  --gpu-metrics-device=3 -x true -o traverse_bs_util_sim_decoding python3 /data/cy/kv_cache_vs_util/sim_traverse_bs/traverse_bs_util_sim_decoding.py > traverse_bs_util_sim_decoding.log 2>&1
+
+hf upload Hamerlate/tts std_traverse_bs sim_traverse_bs --repo-type=dataset
+
+hf upload Hamerlate/tts sim_traverse_bs sim_traverse_bs --repo-type=dataset --exclude '*.sqlite'
+
+hf upload Hamerlate/tts sim_traverse_bs --repo-type=dataset --exclude '*.sqlite'
+
+nsys profile -w true -t cuda,nvtx,osrt --force-overwrite=true --stats=true  --gpu-metrics-device=3 -x true -o traverse_bs_util_sim_prefill_1152 python3 /data/cy/kv_cache_vs_util/sim_traverse_bs/traverse_bs_util_sim_prefill.py > traverse_bs_util_sim_prefill_1152.log 2>&1
+
+nsys profile -w true -t cuda,nvtx,osrt --force-overwrite=true --stats=true  --gpu-metrics-device=3 -x true -o traverse_bs_util_sim_decoding_1024 python3 /data/cy/kv_cache_vs_util/sim_traverse_bs/traverse_bs_util_sim_decoding.py > traverse_bs_util_sim_decoding_1024.log 2>&1