[sfc-gh-mwyatt]

Add the unified RL client frontend: a single op API (client.py) that lowers every op to a canonical Request(op, job_id, body), a Transport ABC with JobHandles/Request (transport.py), a backend config (config.py), and the shared OnPremTransport base with the in-process Ray transport (onprem.py, onprem_ray.py). ray_server.py accepts (job_id, body) on step/save_checkpoint/reset_prefix_cache so the client can forward uniformly.

Includes tests/client/test_client_ops.py covering the op->Request mapping via a FakeTransport (no GPUs/backends needed).

[sfc-gh-kganesan]

Reviewed the whole three-PR stack end-to-end; full write-up on #47. Notes specific to this PR below.

Ran the tests on this branch: 13/13 pass.

Design

The three-layer split — client owns op vocab + job resolution, Transport ABC delivers Request(op, job_id, body, binary), OnPremTransport base owns creation/ordering/teardown, concretes implement only _start / _rpc / _destroy / _wait_running — is exactly the seam that lets #46 (HTTP) and #47 (Cortex) drop in without cross-cutting changes. JOB_CREATE_ORDER = ("sampling", "log_prob", "training") with training last as NCCL root is correct and documented. JobHandles.require() guard is a nice defensive fence. Docstrings across transport.py / onprem.py / client.py clearly state design intent — reviewer-friendly.

Plan-level work this PR opens (details on #47 as P3 / P4)

Merging this stack adds a second client with a different calling convention rather than replacing the first. AST scan:

arctic_platform/rl/http_client.py (ArcticRLHTTPClient) — every op is `async def`
arctic_platform/client/client.py  (ArcticRLClient)     — every op is `sync`

Every existing in-repo caller (the integrations/verl/ RemoteBackend, the recipes/rl/skyrl/ recipes) has to be rewritten before arctic_platform/rl/client.py can be deleted. Please add to UNIFICATION_NOTES.md: which caller migrates on what PR, when rl/*_client.py gets removed. Otherwise "one client" becomes "one-plus-legacy" indefinitely.

Related: ops present in arctic_platform/rl/*_client.py but missing from ArcticRLClient: sleep_/wake_{inference,training,log_prob}, empty_training_cache, weight_norm. Not on the immediate critical path but they block deleting the async clients. Enumerate in UNIFICATION_NOTES.md under a "not yet in unified client" section.

Branch hygiene

main has moved past #36 (the verl integration adapter); all three PRs in the stack are behind. This branch also carries stray commit e697039 (recipes/rl/skyrl … (#32)) which isn't in main and inflates the client-only diff (~866 LOC advertised, +4468 LOC actually on the branch). Rebase before merge, or land #32 first.

Not blocking

• RayTransport._rpc creates a new event loop per op — asyncio.run(method(*args)) at arctic_platform/client/transports/onprem_ray.py:65. Every fwd_bwd / generate / step / log_probs spins up and tears down a fresh loop; on the hot path this is real overhead. Fix: hold one loop on the transport.
• OnPremTransport.shutdown() iterates JOB_TYPES in insertion order rather than reverse creation order. Usually you want to destroy training before its samplers.
• No OnPremTransport unit tests — everything is exercised through FakeTransport. A test that pins _init_payload composition, initialize() ordering, and shutdown() traversal would prevent regressions in feat(client): add on-prem HTTP transport #46/feat(client): add Cortex transport + DSSST1 wire codec #47.
• No op-registry validation — if a client op is renamed and a transport misses the update, AttributeError at runtime instead of import time.
• ArcticRLClientConfig.checkpoint_path overlaps with save_checkpoint(path=...); docstring should clarify which wins when both are set.

None of these gate merge.

[sfc-gh-truwase]

@sfc-gh-mwyatt, I think the gsm8k recipe of either verl or skyrl would be good gating tests for these PRs.

[sfc-gh-mwyatt]

@sfc-gh-mwyatt, I think the gsm8k recipe of either verl or skyrl would be good gating tests for these PRs.

Let's add that in a fast followup after getting #45, #46, and #47 merged.

[sfc-gh-mwyatt]

main has moved past #36 (the verl integration adapter); all three PRs in the stack are behind. This branch also carries stray commit e697039 (recipes/rl/skyrl … (#32)) which isn't in main and inflates the client-only diff (~866 LOC advertised, +4468 LOC actually on the branch). Rebase before merge, or land #32 first.

@sfc-gh-kganesan I think this is your agent doing something wrong because this is not the case. Ignored, but please confirm manually on your side if you think I'm wrong.

[sfc-gh-mwyatt]

OnPremTransport.shutdown() iterates JOB_TYPES in insertion order rather than reverse creation order. Usually you want to destroy training before its samplers.

@sfc-gh-kganesan Is there a reason to do reverse order? Has this been a problem with onprem deployments previously?