宜博 / 2080tiserver

Commit 4c014c9fe3a04610dcc0f5ae8a2ca83813e6ca2e

Authored by yibo
1 parent 8b1a0d22

raylight: 修复 /health 排队计数误把已取消任务算进 queued 

Q.qsize() 是底层 PriorityQueue 的原始元素数，取消一条排队中的任务只是把
JOBS 里的 status 标成 cancelled，不会把它从队列摘除，worker 要等真正
Q.get() 到它才跳过丢弃。堆量取消场景下 queued 数字会明显偏高（比如
"排队 253"）。改成跟 running 同口径，直接数 JOBS 里 status=='queued' 的
条目。

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SBpJTe8WfuYkKwjMaRSDg6
Inline Side-by-side
Showing 2 changed files with 14 additions and 4 deletions
raylight/scripts/server.py
  View file @4c014c9
... ... @@ -568,7 +568,12 @@ def _comfyui_alive() -&gt; bool:
568 568  
569 569 @app.get("/health")
570 570 def health():
  571 + # queued 统计真正 status=='queued' 的任务，不用 Q.qsize()——取消一条排队中的任务只是把
  572 + # JOBS 里那条的 status 标成 cancelled，不会把它从 Q 摘除，worker 要等真正 Q.get() 到它才
  573 + # 会跳过丢弃（见 worker() 里的注释）；qsize() 在那之前会把这些已取消的僵尸 entry 也算
  574 + # 进去，堆量取消时数字会明显偏高，跟 running 的统计口径（数 JOBS 状态）保持一致才准确。
571 575 running = sum(1 for j in JOBS.values() if j.status == "running")
  576 + queued = sum(1 for j in JOBS.values() if j.status == "queued")
572 577 comfy_alive = _comfyui_alive()
573 578 ready = 1 if (STATE["model"] == "ready" and comfy_alive) else 0
574 579 model_state = STATE["model"] if comfy_alive else "comfyui_unreachable"
... ... @@ -578,14 +583,14 @@ def health():
578 583 "gpu": "0,1,2,3",
579 584 "partition": "raylight-fsdp", # 不再是路由依据，纯展示——4卡是一个集群，没有单卡分区概念了
580 585 "offload": "fsdp",
581   - "queued": Q.qsize(),
  586 + "queued": queued,
582 587 "running": running,
583 588 "backend": BACKEND_URL,
584 589 }
585 590 return {
586 591 "status": "ok" if ready else "unavailable",
587 592 "ready": ready, "total": 1,
588   - "queued": Q.qsize(), "running": running,
  593 + "queued": queued, "running": running,
589 594 "backends": [backend],
590 595 "gpus": _gpu_status(),
591 596 }
... ...
raylight/server.py
  View file @4c014c9
... ... @@ -568,7 +568,12 @@ def _comfyui_alive() -&gt; bool:
568 568  
569 569 @app.get("/health")
570 570 def health():
  571 + # queued 统计真正 status=='queued' 的任务，不用 Q.qsize()——取消一条排队中的任务只是把
  572 + # JOBS 里那条的 status 标成 cancelled，不会把它从 Q 摘除，worker 要等真正 Q.get() 到它才
  573 + # 会跳过丢弃（见 worker() 里的注释）；qsize() 在那之前会把这些已取消的僵尸 entry 也算
  574 + # 进去，堆量取消时数字会明显偏高，跟 running 的统计口径（数 JOBS 状态）保持一致才准确。
571 575 running = sum(1 for j in JOBS.values() if j.status == "running")
  576 + queued = sum(1 for j in JOBS.values() if j.status == "queued")
572 577 comfy_alive = _comfyui_alive()
573 578 ready = 1 if (STATE["model"] == "ready" and comfy_alive) else 0
574 579 model_state = STATE["model"] if comfy_alive else "comfyui_unreachable"
... ... @@ -578,14 +583,14 @@ def health():
578 583 "gpu": "0,1,2,3",
579 584 "partition": "raylight-fsdp", # 不再是路由依据，纯展示——4卡是一个集群，没有单卡分区概念了
580 585 "offload": "fsdp",
581   - "queued": Q.qsize(),
  586 + "queued": queued,
582 587 "running": running,
583 588 "backend": BACKEND_URL,
584 589 }
585 590 return {
586 591 "status": "ok" if ready else "unavailable",
587 592 "ready": ready, "total": 1,
588   - "queued": Q.qsize(), "running": running,
  593 + "queued": queued, "running": running,
589 594 "backends": [backend],
590 595 "gpus": _gpu_status(),
591 596 }
... ...