Review PR

Review PR (From PR URL)

Reviews the changes in a specific GitHub pull request. The primary focus is code correctness and design-pattern consistency with the existing codebase. PR formatting, commit conventions, and user-facing documentation are checked but should not dominate the review. See CONTRIBUTING.md for the contribution rules referenced below.

1. Input: PR URL

Require a PR URL (for example: https://github.com/RLinf/RLinf/pull/123).

2. Fetch PR data and the main branch

Fetch PR details directly from the URL:

• Open/fetch the PR page itself for title, description, and metadata.
• Fetch unified diff via URL forms:
  • <PR_URL>.diff (preferred)
  • <PR_URL>.patch (fallback)
• If needed, fetch related pages directly from URL for comments/checks.

All cross-references must be against origin/main, not the local working tree. The current checkout may be on an unrelated branch or contain WIP changes; Glob/Grep over the working tree does not show the upstream state. Before any cross-reference lookup:

• Run git fetch origin main to refresh.
• Read files via git show origin/main:<path> (or browse https://github.com/RLinf/RLinf/blob/main/<path>).
• Use git log origin/main..<pr-head> to scope what the PR actually adds.
• Never treat Glob/Grep results over the working tree as "current state of the project" — they reflect whatever branch is checked out, not main.

If the PR page is private and URL fetch is blocked, report that access is unavailable and ask the user to provide exported diff/details.

3. Review priorities

Findings are ordered by severity. Categories are listed in priority order — most of the review should be on (a) and (b); (c)–(e) are checked but should not pad the output.

(a) Correctness, bugs, and edge cases — primary

For every changed function, branch, and config path, look for:

• Logic bugs: off-by-one, inverted conditions, wrong operator/default, swapped args, mutated shared state, missing await/sync.
• Edge cases: empty/None/NaN, single-element batches, zero-size tensors, world-size=1, first/last iter, eval-only paths, resume-from-checkpoint, multi-node vs single-node paths.
• Concurrency / distributed: collectives that must run on every rank, device placement, non-deterministic ordering across workers, blocking calls in async paths, races on Ray actors / channels.
• Resource & lifecycle: GPU memory leaks (uncleared tensors, missing enable_offload), file handles, Ray actor lifetime, missed cleanup on exceptions, double-init.
• Numerical: dtype mismatches (fp16/bf16/fp32), in-place ops on grad-required tensors, unsafe casts, accumulator precision, loss-mask correctness.
• Error handling: silent except, exceptions on hot paths, missing input validation at boundaries; conversely, over-validation of internally-trusted state.
• Refactor regressions: read both the origin/main version and the new version side-by-side; partial refactors often change semantics by accident.

Cite file:line in the diff and the matching origin/main:file:line when relevant.

(b) Design and pattern consistency vs origin/main — primary

The PR must match how RLinf already does things. Mismatches are usually defects from writing in isolation, not style nits.

• Find the closest sibling in origin/main — comparable model under rlinf/models/embodiment/, env under rlinf/envs/, runner under rlinf/runners/, worker under rlinf/workers/, advantage/loss/reward under rlinf/algorithms/. Compare structure, naming, registration, base-class usage, and config wiring.
• Registry wiring: new advantage/loss/reward must use register_advantage / register_policy_loss / register_reward; new model/env must extend SupportedModel / SupportedEnvType and update get_env_cls() and validate_cfg. Flag ad-hoc bypasses.
• Worker conventions: subclass Worker, implement initialize, use self.log_info / log_warning / log_error (not print or stdlib logging), launch via create_group(...).launch(...).
• Base class / interface: embodied policies must inherit BasePolicy and implement the documented forwards (default_forward, predict_action_batch, plus algorithm-specific). Flag re-implementations of base behavior.
• Config layout: new YAML must be copied from a sibling in examples/ and follow the same key hierarchy; no calculations or dynamic values in YAML; fields read-only in code.
• Reuse vs duplication: if the change reimplements a helper that already exists in rlinf/utils/ (placement, checkpoint, distributed, data-iter, logging), point to the existing helper with file:line.
• Simplicity: prefer the approach the codebase already uses; flag clumsy / over-engineered alternatives with a concrete simpler suggestion.
• No hardcoded paths/hacks: machine-specific paths, sleep-based sync, monkey-patches → propose a config/env-driven version.

(c) Code ↔ docs consistency — required when code OR docs change

Both directions matter, and EN/ZH parity must be checked explicitly:

• If this PR changes documentation, follow the docs-check skill to drive this review: it cross-checks docs against code and against each other (commands, config keys, paths, model/env names) and enforces EN↔ZH parity.
• Docs → code: every config key, CLI flag, env var, file path, function/class name, and supported model/env name mentioned in changed docs must exist in origin/main + this PR. Verify with git show origin/main:rlinf/.... Stale references = finding.
• Code → docs: when this PR adds, removes, or renames a public-facing config key, model, env, runner, script, env var, or supported feature, the corresponding doc page must be updated in the same PR. If missing, list the exact doc files (EN and ZH) that need edits.
• Example-config model paths ↔ docs: for changed example configs under examples/embodiment/config/*.yaml, every model-weight path field (model_path, lora_path, backbone_model_path, wan_wm_hf_ckpt_path) must satisfy all three:
  1. Form: it is /path/to/<repo-name> where the trailing path component equals the repo in its # https://huggingface.co/<org>/<repo> comment (the docs convention is hf download <org>/<repo> --local-dir <repo>, so basename == repo name). Flag stale/placeholder basenames (RLinf-Pi0-SFT, model, openpi, …) and a missing leading slash.
  2. Exists: each distinct referenced repo resolves on Hugging Face — curl -sL -o /dev/null -w '%{http_code}' https://huggingface.co/api/models/<org>/<repo> returns 200 (use /api/datasets/ for datasets). Flag 401/404. Also flag redirecting names: if the API's returned id differs from the requested name, the repo was renamed — point at the canonical id (a 200 via redirect is still a stale reference).
  3. Matches the docs' per-variant model: the repo equals the one the corresponding env recipe page prescribes for that exact environment + task suite + model family — not just any existing repo. Cross-check the recipe doc's model table/download block (e.g. LIBERO spatial/object/goal + π₀ → RLinf-Pi0-LIBERO-Spatial-Object-Goal-SFT, LIBERO-10/Long + π₀ → RLinf-Pi0-LIBERO-Long-SFT, any LIBERO suite + π₀.₅ → RLinf-Pi05-LIBERO-SFT, GR00T N1.5 per-suite RLinf-Gr00t-SFT-{Spatial,Object,Goal,10}). Watch for base-vs-adapter mixups (model_path = full base model, lora_path = LoRA adapter) and casing drift vs the doc/canonical name. Intentional user-supplied placeholders (e.g. DAgger student_model/expert_model, "any pi05 checkpoint") are not findings — confirm against the recipe page before flagging.
• EN ↔ ZH parity (do this explicitly, even when only one language was touched): paired pages under docs/source-en/ and docs/source-zh/ must agree on setup commands, paths, env vars, config keys, supported models/envs/algorithms, capability claims, reported numbers (metrics, table values, dataset sizes, trial counts), and section structure/order. If only one side is updated, name the matching file that also needs the change.
• Sibling style: cross-check with sibling pages in the same area (e.g. opensora.rst) for section naming/order, code-block conventions, table/link style.
• Each docs finding must include a concrete suggested wording / structure fix and exact file references.