Diagnose Qwen30 Fixed-PD admission state

This commit is contained in:
2026-07-19 20:41:20 +08:00
parent 409d83a876
commit a6c9beaa89
4 changed files with 330 additions and 0 deletions

View File

@@ -0,0 +1,45 @@
# 实验Qwen30 Fixed-PD TTFT admission diagnosis
> **状态:** 已批准,运行中
>
> 用户要求分析 Frontier 在 Qwen30 Fixed-PD 高压 case 上 56--58% TTFT
> selection regret 的根因,并给出简洁结论。
## Claim 与决策
- **Parent claim** Frontier 在 capacity knee 附近的配置排序是否会因 state transition error 失效。
- **目的:** 区分 TP2/TP4 conditional prefill-time 错误、mixed-step composition 错误与 admission queue feedback。
- **Competing hypotheses** H1Frontier 把 TP4 prefill execution 相对 TP2 算慢H2decode service time 的绝对误差使 `arrival_rate × residence_time` 越过 MNS cap首次调度等待被阈值放大H3即使固定 admission statemixed prefill/decode composition 仍反序。
- **事前预测:** H1 下去掉 queue 后 TP2 仍有更低 prefill timeH2 下去掉 queue 后 TP4 恢复更快,且只有 simulator 的 required concurrency 超过 MNSH3 下 state-matched stage contrast 仍支持 TP2。
- **判定规则:** 只有 state ledger/scorer 等价、queue counterfactual 和真机 Running/Waiting 同时支持时才归因 H2否则保留 H1/H3 并补最小 telemetry。
## Setup
- **自变量:** config 为 Frontier winner `TP2/MNS64`、real mean winner `TP4/MNS32` 与 real p90 winner `TP4/MNS64`
- **控制变量:** Qwen3-30B-A3B BF16、community vLLM 0.20、H20、Fixed-PD 4096→256、1.125 req/s/GPU、MBT8192、piecewise graph、原 measured profiles/collectives 和原 257-request traces全部冻结。
- **Workload** uniform open-loop arrivalglobal rate 随 TP 为 2.25/4.5 req/sprefix cache off每个真机 cell 三次 fresh-server。
- **Baselines** 完整 12-cell frozen real/sim surface真机三轮 pooled metrics。
- **Metrics** TTFT=`first scheduling delay + prefill execution`request execution/residence time`arrival_rate × service_time` 相对 MNSRunning/Waitingstage ledger composition。
## 预期产物与 review
- **预期数据:** 三个 scorer-equivalent Frontier state replaysservice/admission decompositionH1--H3 verdict。
- **Figure prototype** `../../runs/frontier-fidelity-envelope-v1/qwen30-fixed-pd-ttft-admission-mock.png`;左图区分 TTFT execution 与 queue右图显示 required concurrency 是否跨越 MNS。
- **人工 review** 已批准(用户要求直接分析清楚该 case
- **Review 意见:** 先复用 existing artifacts 和 CPU replay只有现有真机 periodic queue proxy 不足时才增加 GPU telemetry。
## 复现信息
- **Code** AITuner `feature/sim`Frontier frozen campaign commit 见原 provenance。
- **Environment** dash0Frontier replay CPU-onlyGPU visibility disabled。
- **输入:** `/home/admin/cpfs/wjh/aituner/qwen30-fixed-pressure-surface-20260719-r1`
- **产物路径:** `/home/admin/cpfs/wjh/aituner/qwen30-fixed-pd-ttft-diagnosis-20260719-r1`
- **已知 deviation** 原真机日志只有 10 秒 periodic Running/Waiting没有 per-iteration ledger它可验证 steady queue 是否积压,但不用于细粒度 stage timing。
## 结果
- **观察事实:** 待完成。
- **异常:** 待完成。
- **含义:** 待完成。
- **Claim update** unchanged
- **下一步:** 复跑三项 Frontier state ledger并核对 request metrics byte-identical。

View File

@@ -0,0 +1,56 @@
#!/usr/bin/env python3
"""Render the schematic for Qwen30 Fixed-PD TTFT admission diagnosis."""
from pathlib import Path
import matplotlib.pyplot as plt
import numpy as np
OUTPUT = Path(__file__).with_name("qwen30-fixed-pd-ttft-admission-mock.png")
def main() -> None:
figure, axes = plt.subplots(1, 2, figsize=(10.8, 4.2))
labels = ["TP2/MNS64", "TP4/MNS64"]
x = np.arange(len(labels))
axes[0].bar(x, [190, 135], label="Prefill execution (mock)", color="#4c78a8")
axes[0].bar(
x,
[30, 6000],
bottom=[190, 135],
label="Admission wait (mock)",
color="#f58518",
)
axes[0].set_xticks(x, labels)
axes[0].set_yscale("log")
axes[0].set_ylabel("TTFT decomposition (ms, log scale)")
axes[0].set_title("(a) Execution error or queue amplification?")
axes[0].legend(frameon=False)
width = 0.34
axes[1].bar(x - width / 2, [8, 15], width, label="Real required slots (mock)")
axes[1].bar(x + width / 2, [58, 80], width, label="Frontier required slots (mock)")
axes[1].axhline(64, color="#d62728", linestyle="--", label="MNS=64")
axes[1].set_xticks(x, labels)
axes[1].set_ylabel("arrival rate × residence time")
axes[1].set_title("(b) Does modeled state cross admission cap?")
axes[1].legend(frameon=False)
axes[1].text(
0.02,
0.97,
"SCHEMATIC / MOCK DATA",
transform=axes[1].transAxes,
va="top",
fontsize=9,
color="#8c2d04",
)
figure.suptitle("Qwen30 Fixed-PD TTFT diagnosis", fontsize=13)
figure.tight_layout()
figure.savefig(OUTPUT, dpi=180, bbox_inches="tight")
if __name__ == "__main__":
main()

Binary file not shown.

After

Width:  |  Height:  |  Size: 100 KiB

View File

@@ -0,0 +1,229 @@
#!/usr/bin/env python3
"""Replay the decisive Qwen30 Fixed-PD cells with Frontier state outputs."""
from __future__ import annotations
import argparse
import importlib.util
import json
import subprocess
import sys
import time
from pathlib import Path
from typing import Any
HERE = Path(__file__).resolve().parent
CONFIGS = {
"tp2_mns64": "tp2",
"tp4_mns32": "tp4",
"tp4_mns64": "tp4",
}
def load_module(name: str, path: Path):
spec = importlib.util.spec_from_file_location(name, path)
if spec is None or spec.loader is None:
raise ImportError(path)
module = importlib.util.module_from_spec(spec)
sys.modules[spec.name] = module
spec.loader.exec_module(module)
return module
BASE = load_module(
"qwen30_fixed_pd_state_base", HERE / "run_qwen235_fixed_pd_state_replay.py"
)
Q30 = load_module(
"qwen30_fixed_pd_surface", HERE / "run_frontier_qwen30_exact_trace_surface.py"
)
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser()
parser.add_argument("--base-sim-root", type=Path, required=True)
parser.add_argument("--output-root", type=Path, required=True)
parser.add_argument("--frontier-source", type=Path, required=True)
parser.add_argument("--python-deps", type=Path, required=True)
parser.add_argument("--config", action="append", choices=CONFIGS)
parser.add_argument("--timeout-seconds", type=float, default=1800)
parser.add_argument("--resume", action="store_true")
parser.add_argument("--op-trace", action="store_true")
return parser.parse_args()
def run_cell(
*,
config: str,
base_sim_root: Path,
output_root: Path,
frontier_source: Path,
python_deps: Path,
timeout_seconds: float,
resume: bool,
op_trace: bool,
) -> dict[str, Any]:
base_run = base_sim_root / "runs" / config / CONFIGS[config]
base_command_path = base_run / "command.json"
base_result_path = base_run / "result.json"
if not base_command_path.is_file() or not base_result_path.is_file():
raise FileNotFoundError(f"frozen inputs missing under {base_run}")
base_command = json.loads(base_command_path.read_text())
base_result = json.loads(base_result_path.read_text())
if base_result.get("status") != "completed":
raise ValueError(f"frozen result is not completed: {base_result_path}")
run_root = output_root / config
result_path = run_root / "result.json"
if resume and result_path.is_file():
previous = json.loads(result_path.read_text())
if (
previous.get("status") == "PASS"
and previous.get("inputs", {}).get("base_command_sha256")
== BASE.sha256_file(base_command_path)
and previous.get("inputs", {}).get("base_result_sha256")
== BASE.sha256_file(base_result_path)
and bool(previous.get("op_trace_enabled")) == op_trace
):
return previous
if run_root.exists() and any(run_root.iterdir()):
raise FileExistsError(f"refusing non-empty output: {run_root}")
run_root.mkdir(parents=True, exist_ok=True)
command = BASE.transform_command(
base_command,
metrics_root=run_root / "frontier_metrics",
run_id=f"qwen30_fixed_pd_state_{config}",
op_trace=op_trace,
)
BASE.atomic_json(run_root / "command.json", command)
inputs = {
"base_run": str(base_run.resolve()),
"base_command_sha256": BASE.sha256_file(base_command_path),
"base_result_sha256": BASE.sha256_file(base_result_path),
"base_request_metrics_sha256": base_result["request_metrics_sha256"],
}
manifest = {
"schema": "qwen30-fixed-pd-state-replay-v1",
"config": config,
"inputs": inputs,
"command_sha256": BASE.sha256_file(run_root / "command.json"),
"frontier_git_head": subprocess.check_output(
["git", "-C", str(frontier_source), "rev-parse", "HEAD"], text=True
).strip(),
"op_trace_enabled": op_trace,
"controlled_changes": [
"metrics output directory",
"metrics run id",
"full Frontier stage/batch ledger enabled",
"individual batch metrics enabled",
]
+ (["op-level tracing enabled"] if op_trace else []),
}
BASE.atomic_json(run_root / "run_manifest.json", manifest)
started = time.monotonic()
with (run_root / "stdout.log").open("w") as stdout, (
run_root / "stderr.log"
).open("w") as stderr:
try:
completed = subprocess.run(
command,
cwd=frontier_source,
env=BASE.environment(frontier_source, python_deps),
stdout=stdout,
stderr=stderr,
timeout=timeout_seconds,
check=False,
)
returncode = int(completed.returncode)
except subprocess.TimeoutExpired:
returncode = 124
elapsed_seconds = time.monotonic() - started
if returncode != 0:
failure = {
"status": "STOP",
"config": config,
"returncode": returncode,
"elapsed_seconds": elapsed_seconds,
}
BASE.atomic_json(run_root / "failure.json", failure)
raise RuntimeError(f"state replay failed: {failure}")
fallback_evidence = Q30.collective_fallback_evidence(run_root)
if fallback_evidence:
raise RuntimeError(f"collective-profile fallback detected: {fallback_evidence}")
paths = BASE.find_state_metrics(run_root)
request_metrics_sha256 = BASE.sha256_file(paths["requests"])
scorer_equivalent = request_metrics_sha256 == base_result["request_metrics_sha256"]
if not scorer_equivalent:
raise RuntimeError(
f"observation changed scorer input for {config}: "
f"{request_metrics_sha256} != {base_result['request_metrics_sha256']}"
)
state = BASE.summarize_frontier(
system_metrics_path=paths["system"],
request_metrics_path=paths["requests"],
batch_metrics_path=paths["batches"],
ledger_path=paths["ledger"],
)
BASE.atomic_json(run_root / "common-state.json", state)
if op_trace:
matches = sorted(run_root.glob("frontier_metrics/**/op_traces.jsonl"))
if len(matches) != 1:
raise ValueError(f"expected one op trace, found {len(matches)}")
paths["op_trace"] = matches[0]
result = {
"schema": "qwen30-fixed-pd-state-replay-result-v1",
"status": "PASS",
"config": config,
"elapsed_seconds": elapsed_seconds,
"returncode": returncode,
"inputs": inputs,
"scorer_equivalence": {
"request_metrics_byte_identical": scorer_equivalent,
"request_metrics_sha256": request_metrics_sha256,
"base_metrics": base_result["metrics"],
},
"state_artifacts": {
name: {"path": str(path.resolve()), "sha256": BASE.sha256_file(path)}
for name, path in paths.items()
},
"common_state": state,
"collective_fallback_evidence": fallback_evidence,
"op_trace_enabled": op_trace,
}
BASE.atomic_json(result_path, result)
return result
def main() -> None:
args = parse_args()
for name in ("base_sim_root", "output_root", "frontier_source", "python_deps"):
setattr(args, name, getattr(args, name).resolve())
results = []
for config in tuple(args.config or CONFIGS):
result = run_cell(
config=config,
base_sim_root=args.base_sim_root,
output_root=args.output_root,
frontier_source=args.frontier_source,
python_deps=args.python_deps,
timeout_seconds=args.timeout_seconds,
resume=args.resume,
op_trace=args.op_trace,
)
results.append(result)
print(json.dumps({"config": config, "status": result["status"]}), flush=True)
BASE.atomic_json(
args.output_root / "state_replay.json",
{
"schema": "qwen30-fixed-pd-state-replay-aggregate-v1",
"status": "PASS",
"results": results,
},
)
if __name__ == "__main__":
main()