[Cherry-Pick]【TI-Consisent】Added Metric logits_stats to the ZMQ branch (#6979)

[liuruyan]

Motivation

背景：出于对训推一致性丰富检测指标及长期CI/CE监控考量，对sample之后的logits添加logits_stats(min/max/mean/std)，用来保证确定性及稳定性。

Modifications

数据结构及接口：由于logprob与logits_stat同样是输出的重要检测指标且均从logits计算而来，暂时实现先将logits_stat存入LogprobsTensors数据结构保存，并升级logprob传播链路上相关接口，支持同时透传logits_stats

class LogprobsTensors(NamedTuple):
    """ """

    # [num_reqs, max_num_logprobs + 1]
    logprob_token_ids: paddle.Tensor
    # [num_reqs, max_num_logprobs + 1]
    logprobs: paddle.Tensor
    # [num_reqs]
    selected_token_ranks: paddle.Tensor
    # Logits statistics for each sequence (optional)
    logits_min: Optional[paddle.Tensor] = None  # [num_reqs]
    logits_max: Optional[paddle.Tensor] = None  # [num_reqs]
    logits_mean: Optional[paddle.Tensor] = None  # [num_reqs]
    logits_std: Optional[paddle.Tensor] = None
    ...

FLAG：添加与enable_logprob同级别model_config：self.compute_logits_stats = False，且在server启动时支持配置--compute-logits-stats

注：由于改变了返回字段，导致有些单测无法通过，所以改动单测文件，返回值中提出新增字段（logits_stats ）

Usage or Command

• 本功能暂时只支持ZMQ，流式与非流式测试均可正常返回
• 启动FD服务时需要同时开启--compute-logits-stats,--enable-logprob

export FD_USE_GET_SAVE_OUTPUT_V1=1 
python -m fastdeploy.entrypoints.openai.api_server \
       --enable-logprob \
       --compute-logits-stats \
       ...  # more setting

• 发送请求时需要指定logprobs=True,top_logprobs=0

response = client.chat.completions.create(
    model="null",
    messages=[
        {"role": "system", "content": "I'm a helpful AI assistant."},
        {"role": "user", "content": "把李白的静夜思改写为现代诗"},
    ],
    stream=True,  # False
    max_tokens=100,
    logprobs=True,
    top_logprobs=0
)

Accuracy Tests

本PR不涉及精度修改，稳定性测试结果如下：

1. server 实例

export CUDA_VISIBLE_DEVICES=5
export FD_USE_GET_SAVE_OUTPUT_V1=1
python -m fastdeploy.entrypoints.openai.api_server \
       --tensor-parallel-size 1 \
       --load-choices "default_v1" \
       --port 8137 \
       --engine-worker-queue-port "8978" \
       --metrics-port 8977 \
       --max-model-len 8192 \
       --max-num-seqs 32 \
       --gpu-memory-utilization 0.9 \
       --reasoning-parser ernie5 \
       --quantization "block_wise_fp8" \
       --model "$MODEL_PATH" \
       --graph-optimization-config '{"use_cudagraph":true}' \
       --no-enable-prefix-caching \
       --enable-logprob \
       --compute-logits-stats \

2. 流式请求

import openai
host = "0.0.0.0"
port = "8137"
response = client.chat.completions.create(
    model="null",
    messages=[
        {"role": "system", "content": "I'm a helpful AI assistant."},
        {"role": "user", "content": "把李白的静夜思改写为现代诗"},
    ],
    stream=True,
    max_tokens=100,
    seed=0,
    top_p=1e-5,
    logprobs=True,
    top_logprobs=0
)
for chunk in response:
    print(chunk)
    if chunk.choices[0].delta:
        print(chunk.choices[0].delta.reasoning_content, end='')
print('\n')

3. 非流式请求

import openai
host = "0.0.0.0"
port = "8137"
client = openai.Client(base_url=f"http://{host}:{port}/v1", api_key="null")
response = client.chat.completions.create(
    model="null",
    messages=[
        {"role": "system", "content": "I'm a helpful AI assistant."},
        {"role": "user", "content": "把李白的静夜思改写为现代诗"},
    ],
    stream=False,
    max_tokens=100,
    seed=0,
    top_p=1e-5,
    logprobs=True,
    top_logprobs=0
)
print(response)

4. 实验结果：

• 各自多次实验所有Token logits指标均可对齐

• 流式及非流式所有Token logits指标均可对齐

• 流式request返回体：

• 非流式request返回体：

Checklist

• Add at least a tag in the PR title.
  • Tag list: [[FDConfig],[APIServer],[Engine], [Scheduler], [PD Disaggregation], [Executor], [Graph Optimization], [Speculative Decoding], [RL], [Models], [Quantization], [Loader], [OP], [KVCache], [DataProcessor], [BugFix], [Docs], [CI], [Optimization], [Feature], [Benchmark], [Others], [XPU], [HPU], [GCU], [DCU], [Iluvatar], [Metax]]
  • You can add new tags based on the PR content, but the semantics must be clear.
• Format your code, run pre-commit before commit.
• Add unit tests. Please write the reason in this PR if no unit tests.
• Provide accuracy results.
• If the current PR is submitting to the release branch, make sure the PR has been submitted to the develop branch, then cherry-pick it to the release branch with the [Cherry-Pick] PR tag.

[paddle-bot]

Thanks for your contribution!

[codecov-commenter]

Codecov Report

❌ Patch coverage is 65.45455% with 19 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (release/2.5@d12aa28). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/entrypoints/openai/serving_chat.py	50.00%	11 Missing and 4 partials ⚠️
fastdeploy/output/token_processor.py	42.85%	4 Missing ⚠️

Additional details and impacted files

@@              Coverage Diff               @@
##             release/2.5    #6978   +/-   ##
==============================================
  Coverage               ?   68.40%           
==============================================
  Files                  ?      390           
  Lines                  ?    54120           
  Branches               ?     8523           
==============================================
  Hits                   ?    37021           
  Misses                 ?    14408           
  Partials               ?     2691           

Flag	Coverage Δ
GPU	68.40% <65.45%> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[EmmonsCurse]

LGTM～ Skip coverage check as it mainly relies on end-to-end tests.

[liuruyan]

/re-run run_tests_with_coverage