freesolo-flash-dev 1.2.113__tar.gz → 1.2.116__tar.gz

freesolo_flash_dev-1.2.116/CHANGELOG.md

@@ -0,0 +1,196 @@
+# Changelog
+
+Notable changes to `freesolo-flash`. The dev-channel package `freesolo-flash-dev` ships in
+lockstep at the same version.
+
+Releases happen by merging a version bump from `dev` into `main`, which publishes to PyPI
+(`.github/workflows/publish.yml`). Versions are `MAJOR.MINOR.PATCH`.
+
+This file starts at 1.1.35. Earlier releases are not reconstructed here; use
+`git log` for anything before that.
+
+## Unreleased
+
+### Added
+
+- A `flash-cli` console script, the same entry point as `flash` under a name nothing else
+  claims. The `server` and `dev` extras install `runpod-flash`, which declares its own `flash`
+  script; whichever distribution is installed last wins, so on a control-plane host
+  `pip install 'freesolo-flash[server]'` could leave `flash` running RunPod's CLI, which exits 0
+  and does nothing. `flash` is unchanged and remains correct on a base (client-only) install.
+
+### Changed
+
+- Active training and hosted serving now expose exactly `Qwen/Qwen3.5-9B`,
+  `Qwen/Qwen3.8-27B`, and `Qwen/Qwen3.6-35B-A3B`. Qwen3.8 uses the immutable BF16 revision
+  `1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0` and official native block-FP8 serving revision
+  `017b9c7af6b5689d5dd426a76e0bc077eb5ca20a`. The Qwen3.5 0.8B, 2B, and 4B tiers and
+  `Qwen/Qwen3.6-27B` cannot activate, while historical status, accepted billing retry,
+  cancellation, cleanup, hosted revocation, and exact identity-based undeploy remain available.
+  Qwen3.6 27B adapters are not compatible with Qwen3.8. Customer-owned serving remains 9B-only.
+- `[train] init_from_adapter` now works for every source/target algorithm pair. SFT was rejected as
+  a warm-start target, so an adapter could be continued only by GRPO or OPD; that restriction
+  described the retired trl SFT backend, and the verl backend that replaced it loads a warm-start
+  adapter exactly like the other two. All nine combinations of `sft`, `grpo`, and `opd` are
+  supported, including same-algorithm continuation. A warm-started SFT run also inherits its
+  source's base-model pin instead of resolving its own, so it stays deployable and no longer breaks
+  when the base model's hub tip moves.
+- Managed attempts now pin, verify, and report one immutable Flash source archive across retries and recovery.
+- Multi-turn GRPO now clones each sibling rollout from the task state that produced its prepared
+  prompt. Environments that record randomized episode choices on the task during `start_episode`
+  are stepped and scored against the same episode the model saw, without sharing mutable state
+  between sibling rollouts.
+- A RunPod attempt's stall window is now measured from evidence timestamps rather than from when
+  the control plane happened to look at it. The window was seeded from the moment polling started
+  and was reset again by the first status read, and a heartbeat was credited at the time it was
+  read rather than the time it was written. Every one of those is a wall-clock reading, so each
+  reattach handed a worker that was already wedged a complete fresh setup window, and a supervisor
+  that reattached repeatedly could keep a dead worker on a paid GPU until the run's wall deadline.
+  The window is now anchored to the persisted launch, a heartbeat is credited at its own timestamp
+  clamped to that launch, and only a genuine status transition counts as progress. Instance
+  polling (Vast/Lambda) already worked this way; this brings RunPod onto the same rule. A job that
+  is still queued is unaffected: the capacity wait keeps its own exemption, so a worker granted
+  late still gets its full cold-start budget.
+- An attempt's terminal artifacts (its strict marker plus `metrics.json`) are now interpreted by
+  one shared protocol instead of being decoded separately by live instance polling and by each
+  recovery path. The two decoders disagreed about identical bytes. A marker that could not be tied
+  to the attempt was a terminal failure to the live poller but indistinguishable from "no marker
+  yet" to recovery, so the same corrupt artifact produced a different story depending on which
+  layer happened to observe it; recovery now names it rather than logging silence. Neither path
+  ever adopted such a marker as completed work, and neither does now. Recovery also computed a
+  fresh observation window for the marker and then another for `metrics.json`, so the real ceiling
+  was their sum and moved with however long the first read took; both reads now share one window,
+  which makes the bound the cutoff rather than the sum.
+- CI runs the offline test suite on both supported interpreters (3.11 and 3.12) rather than
+  3.11 alone, and sets `HF_HUB_OFFLINE` / `TRANSFORMERS_OFFLINE` so a test that reaches the
+  network fails deterministically instead of depending on runner connectivity.
+- Every GitHub Action is pinned to a commit SHA rather than a mutable tag.
+- `mypy` runs in CI as an advisory (non-blocking) job.
+
+### Fixed
+
+- Completed verl children now finish each captured W&B run before the W&B service shuts down,
+  avoiding teardown tracebacks, and `flash env setup` now gives every starter config explicit W&B
+  project and folder-derived run metadata.
+
+- An ordered `[gpu] type` list is now priced over every class it declares acceptable, not just the
+  first. Allocation cost-ranks the whole set, so quoting the head alone priced a shape the run may
+  never be given — an authored `["B200", "H100"]` quoted roughly 3x the H100 the allocator would
+  really rent, and the submit-time affordability check runs on that estimate, so a run the
+  organization could afford could be refused before allocation ever got to choose. A single
+  authored class is still a hard pin, and a no-fit rejection now names every acceptable class
+  rather than whichever happened to be written first.
+
+- A provider preference that exhausts every configured provider no longer falls back to quoting
+  hardware this plane cannot rent. The final fallback dropped the restriction and ranked the
+  registered RunPod pool regardless of credentials, so a vast-only plane could quote a B200: the
+  estimate passed affordability and the run was recorded, only for live allocation to fail. The
+  quote now raises against the eligible set, which names the real constraint.
+
+- A GitHub token that is invalid, expired, or missing a scope now fails during preparation instead
+  of after a GPU is rented. Non-rate-limit 401 and 403 responses were left as untyped errors, which
+  the submit-time preflight treats as a transient blip worth deferring, so the run proceeded
+  through affordability and allocation and hit the same permanent credential error on the worker.
+  They are now classified permanent, like 404 and 422. The rate-limit 403 is still transient.
+
+- A slow heartbeat upload no longer stalls every other heartbeat behind it. The throttle clock only
+  advances once a commit lands, so during a slow HuggingFace commit other callers still saw an
+  upload as due and blocked on the upload lock for up to its acquire timeout — 30s, or 120s on a
+  critical stage — which stalled the liveness daemon and delayed leaving the wrapped stage by the
+  same window. A throttled heartbeat now skips while a commit is in flight, since that commit
+  already carries the stage's state. Terminal, error, initial, and forced step commits still queue
+  and land: they carry state no later heartbeat would repair.
+
+- A GRPO or OPD run whose rollout actor exited hard no longer wedges on a paid GPU until the
+  wall-time limit. The actor's `os._exit` bypassed verl's own prompt handler, so the prompt
+  stayed `running` forever and verl's unbounded replay poll waited on it at ~0% utilization. The
+  child now writes durable failure evidence, marks the prompt failed, cleans up, and only then
+  exits, and replay sampling fails the step on an explicit failure tag or a positively dead
+  actor instead of returning a survivor-only batch that would silently train on fewer rollouts
+  than were collected. A child that reaches the fit loop and then stops producing distinct
+  output, while the parent has no teacher or environment work in flight, is torn down with a
+  named cause rather than running silently to the limit; repeated identical lines count as
+  silence, and authoritative OOM, host-RAM, and infrastructure evidence still classifies the
+  exit first.
+- `flash models deploy` now warns before it moves the shared `<run-id>` model id onto a
+  different checkpoint. Every checkpoint of a run is served under that one id, so deploying
+  `<run-id>/step-50` while `step-100` was live silently changed what `<run-id>` served for
+  everyone using it — which made "deploy a second checkpoint to compare" a destructive
+  operation on the first, and read as a serving regression rather than the deploy that caused
+  it. The warning names both checkpoints and points at `flash models chat <run-id>/step-N`,
+  which addresses a checkpoint directly and does not depend on where the id points. It is
+  advisory: the deploy still proceeds, the pre-deploy read is abandoned outright once it exceeds
+  a few seconds so no shape of slow plane can hold a deploy behind it, and a control plane that
+  cannot answer (or answers with a checkpoint step this client cannot read) does not fail the
+  command. A deployment whose activation outcome was never settled still warns, but without
+  naming a checkpoint: that record's step describes the incoming attempt rather than what the id
+  currently serves, and the authoritative target is server-side only, so the warning says the
+  live checkpoint cannot be determined and points at `flash models deployments`.
+
+- A response deadline did not actually bound a response that arrived slowly. The deadline is
+  checked between reads, but each read asked for a fixed number of bytes and blocked until all of
+  them arrived, so a peer trickling a short body kept one read inside the socket timeout
+  indefinitely and the check never ran: a 2s deadline took 12s on a 42-byte body, and 61s when
+  the socket timeout was the client default. Reading whatever has already arrived lets the check
+  run, which bounds every stall shape. This affected `flash env list`, whose deadline exists to
+  stop a slow environment hub hanging the command, as well as the new pre-deploy read.
+
+- A response deadline could still be overrun by a read that began near its end. The socket timeout
+  is installed once, when the request opens, and time spent connecting and waiting for headers is
+  charged to the deadline but not to that timeout, so a peer that delayed its headers and then
+  stalled mid-body held the call open past the deadline: 3.5s against a 2s budget. Each read now
+  re-caps the socket to the remaining budget, so the two bounds agree instead of stacking. On a
+  body still arriving this surfaces as a timeout rather than the "stalled" message; both bound the
+  call, and which one wins is a race not worth depending on.
+
+- Commands printed for the operator to run (the resume/cancel hand-off after `flash train`,
+  usage strings, `next:` hints) now name the executable actually invoked rather than always
+  `flash`. On a host where `runpod-flash` owns the `flash` script, the printed
+  `flash runs cancel <run-id>` exited 0 without cancelling, leaving a run billing while the
+  operator believed they had stopped it. Reached through the `python -m flash.cli` escape hatch,
+  the printed command now names the interpreter that is actually running, rather than a bare
+  `python` that is absent on a python3-only host and may resolve to a different environment
+  inside a virtualenv.
+
+- `flash env setup` scaffolded hosted-only environment instructions on every plane. A
+  self-hosted plane cannot publish to Freesolo's managed environment hub, so the generated
+  configs left `[environment] id` empty (which fails validation) beside a `flash env push`
+  command that could not work. Against a self-hosted plane the scaffold now emits the
+  `github:owner/repo@ref:path` id form, which such a plane resolves directly. Output against
+  the managed plane is unchanged.
+
+- `flash-server` reported missing operator configuration as an unhandled ASGI startup
+  exception, so the actionable message arrived under ~20 frames of starlette/contextlib. The
+  refusing half of the preflight now runs before uvicorn starts, and `flash-server` prints
+  `error: ...` and exits 3 (uvicorn's `STARTUP_FAILURE`, which this path already used, so
+  supervision keying on it is unaffected). The advisory warnings still come from the lifespan
+  alone, so a booted plane logs them once.
+- `flash env test` blamed the reward function for a gold answer that scored zero, when the
+  gold completion is equally likely to be at fault: with no `sft_completion` hook the replayed
+  answer is the dataset row's raw `output`, so a dataset whose `output` is a bare value fails a
+  grader that (correctly) requires a wrapper. The warning now names both candidates and prints
+  the exact text it scored, verbatim rather than through the whitespace-collapsing, 200-character
+  preview that hid the formatting faults it exists to reveal. The scaffolded `environment.py`
+  documents that `output` is also the gold assistant turn SFT trains on, and that per-row scorer
+  state belongs in `metadata`.
+- `flash env test` labelled the replayed dataset turns as the text that scored zero, but a
+  multi-turn environment whose `step_episode` returns a `final_response_text` has that override
+  replace the episode's response before scoring, so the diagnostic named text the grader never
+  saw. It now reports what the scorer actually received, captured at the scoring call.
+- An environment that overrode its answer to the empty string had `flash env test` report the
+  replayed turns instead: the captured text was tested for truth, so a graded `""` was
+  indistinguishable from an episode that never reached the scorer. The override propagates on
+  `is not None`, so `""` really is what the grader saw -- and an empty answer explains a zero
+  reward on its own, making it the reading a reader most needs named rather than hidden.
+- A crashed scorer went unreported on an `echo` episode -- a row with no gold answer to replay,
+  which is also the case that leaves the GRPO gate with nothing to count. `flash env test` could
+  exit `overall: PASS` with every reward coming from a scorer that never ran. The scorer's error is
+  now surfaced for any policy.
+- A scorer that crashed behind the SDK's guard reached `flash env test` as a bare `0.000000`,
+  indistinguishable from a judged-wrong answer, because the adapter kept only
+  `RewardResult.score`. The new `FreesoloEnvironment.reward_with_error` returns the reward and the
+  discarded error from a single scoring call, so a missing runtime dependency names itself without
+  re-scoring the episode -- which would bill a paid judge twice and, for a non-deterministic
+  scorer, could report an error that did not produce the printed reward.
+- Two tests resolved Hugging Face model metadata over the network and passed only on a
+  connected runner. They now stub the resolver like their siblings already did.