traverse_bs_util_std_kvcache.py

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()