mlxtop v3.2.0

Journal

• GPU rows mark real load changes. GPU load wobbling around a line during
  processing, such as 84% → 79% → 81% or 99% → 94%, no longer writes a row
  each sample. The busy band turns on at gpu_warn_load and holds until load
  drops below gpu_warn_exit, and saturation clears 10 points below
  gpu_critical_load. A GPU BUSY assessment row now files under the GPU
  filter and replaces the duplicate "busy burst started" row.
• "GPU busy" in Settings now sets GPU BUSY. The assessment used a fixed
  80%; it now reads gpu_warn_load (default 75%), so the Settings value and
  the state agree.
• Telemetry online is recorded once. An idle server whose rate comes and
  goes between polls no longer logs "live API telemetry online" every few
  seconds. It logs again for a new provider or source, or after the server
  went away and came back.
• Choose what the Journal records. journal.prompts, journal.llm,
  journal.queue, journal.memory, journal.paging, journal.gpu and
  journal.thermal turn each kind of row off, from config.json or the new
  Journal section in Settings. System rows, assessment findings such as
  HEAVY PAGING or GPU THROTTLED, and a move to critical memory pressure are
  always recorded.

Dashboards

• Strata servers are supported. Detects the strata process or accepts
  MLXTOP_PROVIDER=strata, using port 8080 by default. Strata serves vLLM's
  gauge names on /metrics/prometheus beside its own strata: live-rate
  gauges, so mlxtop shows live decode and prefill speeds, active and waiting
  queues, KV occupancy, prefix-cache reuse, mean first-token time and engine
  GPU memory. Model names come from the metric labels and /v1/models;
  individual request history uses the client usage recorder.

• TensorFold is now supported. Detects the tensorfold or
  tensorfold-native process or accepts
  MLXTOP_PROVIDER=tensorfold, using port 8080 by default. Reads model names,
  active and queued requests, prefix-cache reuse, KV occupancy and device memory
  through the existing authenticated provider connection. Throughput is sampled
  from completed-request counters and excluded from live throughput charts and
  slowdown alarms; individual request history uses the client usage recorder.

• Sushi servers report their own generation and prefill rates. The
  mlx-serve adapter now reads the sushi: gauge names when a server
  publishes them instead of mlx_serve: (sushi serve does), beside the
  shared vLLM-shaped queues. Decode, prefill, per-request speeds and prefill
  progress work unchanged; MLXTOP_PROVIDER=mlx-serve stays the selection,
  and a running sushi process is detected as mlx-serve.

• Every Overview panel scales with the terminal, in balance. The
  throughput row held seven rows on any terminal tall enough for the full
  layout, and the request row held six until the first request arrived, so
  on a small window — a laptop beside a server, an SSH session from another
  machine — idle panels kept their space while the memory, compression and
  paging charts absorbed the whole change and grew to half the screen. All
  three rows now follow the same floor-share-ceiling shape the Journal
  already had, so the host charts keep near two fifths of the height at
  every size instead of dwarfing the rest, and each window resizes with the
  screen dimensions between compact and tall.

• The prefill rate stays after the prefill finishes. A finished
  request's prefill speed was held for three seconds and then replaced by a
  live zero, so a request that decoded for longer than that showed
  PREFILL 0.0 for the rest of its run. The last measured speed now stays,
  labeled LAST PREFILL, until a newer prefill replaces it. A fresh
  dashboard seeds its first reading from the server's cumulative finished
  prefill counters instead of showing nothing. The chart does not redraw
  that held reading on every poll, and it is never labeled LIVE.

• The cache panel keeps a reading through a long decode. Prefix-cache
  counters only move when a request completes, so the interval rate went
  blank while a request decoded. When no completed request supplied an
  interval rate, mlxtop now uses the prefix reuse of the requests still
  running.