hybrid decode halves (23 → 11 tok/s) whenever a foreign thread lands on a pinned spinning CPU-MoE pool CPU; please detect/warn, consider SCHED_FIFO or a yielding spin

Abierto
  • ¿Algo no está bien?
#744 0 comentarios 0 reacciones 0 asignados Ver en GitHub

Los mantenedores suelen responder en 1 día

Nadie ha tomado este issue todavía.

Evaluación

Dificultad
4/5
Tiempo estimado
3-5 días
Aptitud para principiantes
42/100
Tipo de issue
Error
Claridad
Bastante claro
Estado de actividad
Activo
Stack tecnológico
linux, python
Área
backend, performance

Línea de trabajo

Start with the hybrid CPU-MoE pool and its pinned worker threads; the issue does not name source files or tests, so first locate their implementation and the existing --moe-collect-stats path (#231). Check whether pool CPUs can be monitored for competing runnable tasks and whether barrier wait times are already recorded. Done means the issue’s detection and warning, barrier statistics, robustness, or documentation work has a defined scope and can be verified against the reported slowdown.

Escrito por el modelo de indexación a partir del texto del issue.

Descripción

Summary

With --moe-strategy hybrid on DeepSeek-V4.1-Flash, decode on one server process alternated between ~23 tok/s and ~11 tok/s for a day and a half. Same request, same flags, no log line in either mode. We traced it to any runnable foreign thread sharing a logical CPU with one of the pinned, spinning CPU-MoE pool threads: every layer's barrier waits for that one worker, so a small intruder costs a lot, and once a thread is parked there the kernel rarely moves it, so it looks like a sticky "mode". Fencing every other process off the pool's CPUs with systemd CPUAffinity removed it (0 slow requests since, under the loads that used to trigger it).

Filing because (a) nothing in the server surfaces this, so it cost us ~20 hours to find, and (b) a pool that spins on pinned cores could defend itself.

Environment

  • FreeToken nightly 0.1.3+g8f49e0af9
  • DeepSeek-V4.1-Flash-1017, ds_fp4(act_block=32), 40 MoE layers, top-6
  • 2× EPYC 7742 (2 NUMA nodes × 64 cores, SMT on, 256 logical CPUs), 1 TB DDR4-2666, RTX 3090 24 GB on node 1 (PCIe Gen4 x16), driver 595.x, Linux 7.0.0
  • numactl --physcpubind=64-127,192-255 --membind=1 ft serve --moe-strategy hybrid --disable-moe-prefill-overlap --kv-reserve-tokens 65536 --max-prefill-length 2048 --memory-ratio 0.85
  • Startup: CPU MoE executor ready: threads=63 (pinned to cores 64..126) … spin_wait=True handshake=kernel, --moe-hybrid-max-fetch auto: 18.1%, CUDA graphs [1,2,4]

Symptom

Single stream, 150-token completions, temperature 0, thinking off:

fast slow
engine gen throughput 22–24 tok/s 10.6–13 tok/s
GPU util / power ~60 %, at the 230 W cap ~30 %, 185 W
per token 44 ms 88 ms (≈ +1.1 ms per MoE layer)
prefill / TTFT unchanged unchanged
main thread, workers, coordinator 100 % busy 100 % busy

Fresh starts came up in either mode (6 of ~12 slow one afternoon). A sched_setaffinity on the main thread flipped fast→slow 4 of 6 times. Nsight (fast mode) shows the GPU idling a median 0.43 ms before each layer's combine kernel, i.e. waiting on the pool; we never got a slow-mode capture because the trigger stopped working once the box went quiet after midnight — which was the clue.

Cause (direct test on a live server, no restart)

A pinned busy thread placed by hand:

intruder decode
none 23.9 tok/s
20 %-duty thread pinned to a pool CPU (70) 19.2–19.7; back to 23.6 the moment it is killed
same thread on that core's idle SMT sibling (198) 23.0 (no effect)
100 %-duty thread on pool CPU 70 ~0.4 tok/s (150 tokens took 344 s)
20 %-duty on the coordinator CPU (127) 19.5
six unpinned ~30 %-duty threads 22.6–23.2 (kernel kept them off the pool)

Every slow period in our logs coincided with other services on the box (a cron-like scheduler, an API server at ~80 % CPU with a thread observed on CPU 65, a remote-desktop daemon at 100 % on a node-1 CPU). The kernel normally places foreign threads on idle logical CPUs, so unpinned load is harmless most of the time; once one is parked on a pool CPU nothing pulls it off a busy core while the pool spins, so the slow state persists until some rebalance — which is why affinity calls re-rolled it and why idle periods sometimes recovered it.

Things that did NOT matter: FREETOKEN_CPU_MOE_SPIN=0 (slower in both modes, still bimodal), FREETOKEN_CPU_MOE_FLAG_SYNC=0 (always the slow speed), NUMA layout (--interleave=all, 127 threads per #18: still flips), CPU clocks, temperature, DRAM bandwidth, PCIe link state, page faults, GPU power cap.

Fix that worked (host side)

systemd CPUAffinity=0-63,128-191 on every other service (user and system units, including sshd so login shells inherit it), leaving 64-127 to the pool. Since then: 0 slow requests in ~40 probes across the loads that previously triggered it.

Asks

  1. Detect and warn. The pool knows its pinned CPUs; a periodic check of /proc/*/task/*/stat (or /proc/schedstat run-queue length per CPU) for another runnable task on a pool CPU, logged as a WARN with the pid/comm, would have pointed straight at this. Even a one-time startup check would help.
  2. Per-layer barrier timing under --moe-collect-stats (#231): median/p99 wait at the done barrier per worker, so a straggler core is visible from the stats rather than from Nsight.
  3. Consider making the pool robust to it: e.g. SCHED_FIFO/SCHED_RR for pool threads when permitted (CAP_SYS_NICE, or document ulimit -r), or a spin with sched_yield/short backoff so an intruder on a pool CPU gets its 20 % without stalling the barrier; or a documented --moe-cpu-threads sizing note that leaving a few cores free does not help if other processes are not fenced.
  4. Docs: a line in the hybrid-mode guide saying "fence everything else off the pool CPUs (systemd CPUAffinity / cpuset); an unrelated process landing on one of them halves decode" would have saved us a day.

Happy to run anything; the probe scripts (pinned intruder, per-thread sched deltas, Nsight gap analysis) are small and reusable.

Lenguaje dominante
Python
Estrellas
14.2k
Forks
1.4k
Merge medio
18 h 46 min
PR fusionados (30 d)
98

Preparar el entorno

Este proyecto no incluye contenedor de desarrollo, Dockerfile ni guía de contribución, así que la configuración corre por tu cuenta: empieza por su README y consulta nuestra guía para la primera contribución para los pasos generales.

Primeros pasos

  1. Lee el issue completo y luego la guía de contribución del proyecto.
  2. Comenta en el issue que vas a ocuparte — evita que dos personas hagan lo mismo.
  3. Haz un fork del repositorio y trabaja en una rama.
  4. Abre un pull request que haga referencia al número del issue.

Más de FlashML-org/FreeToken

Todos los issues de FlashML-org/FreeToken

Issues similares

Más issues de Python

Recibe los nuevos issues en tu correo

Un resumen breve de issues de GitHub para principiantes.