Phase 1 of #7: inference/learning roles, one engine, label-routed backends - #9
Merged
Conversation
Phase 1 of #7. One engine per session now carries two named backends, and asyncflow routes per task on its label -- the mechanism shipped in asyncflow 0.5.1 (`function_task(backend=...)` -> `target_backend`), so the whole change is service-side. * session: `_create_engine` becomes `_create_backend`; the engine is created once around the 'inference' backend, a configured 'learning' backend attaches to it (asyncflow `_attach_backend`; the public spelling is an upstream ask). Build dedup, per-role locks, cancelled-caller survival, close-during-build disposal and the R8 loss mapping all keep their semantics, now keyed by role. * learn: `learn_flow` is gone. The service injects `learn_backend` (the label), and the investigator staples it onto ROSE's own registration seam -- `_register_task` forwards `decor_kwargs` into `function_task`, so no ROSE change and no engine proxy (ROSE type-checks the engine argument; a proxy does not pass). * naming: the reserved roles are `inference` and `learning` -- function, not placement ('insitu'/'exsitu' broke the moment inference ran remotely). No compatibility aliases: config keys, env vars (`DT_INFERENCE_ENDPOINT` / `DT_LEARNING_ENDPOINT`), endpoint examples, dashboard lanes, demos, docs and both bundled recordings renamed in one sweep. Dashboard VERSION 0.6.0: an old recording's role keys would draw empty lanes, and the badge says so. Verified: 164 unit + 29 integration, including the dual-endpoint learner test that proves the label routing end to end (training on the learning endpoint, inference served from the inference endpoint). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
andre-merzky
added a commit
that referenced
this pull request
Aug 24, 2026
devel brings the merged stack and the phase-1 role rename (#9). Per the rule in #7, every conflict resolves toward `inference`/`learning`: the pool cards keep their new table+graph body and say 'Pool: inference' / 'Pool: learning', `poolHistory` and the lane colour checks follow, and the page version supersedes both sides as 0.7.0. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Phase 1 of the #7 sequence, per the design comments there.
inference,learning); asyncflow 0.5.1 routes per task ontarget_backend. Service-side only: the label rides ROSE'sdecor_kwargsseam, no ROSE change.learn_flowremoved; the service injectslearn_backend(the label). Engine build machinery (dedup, per-role locks, cancelled-caller survival, close disposal, R8 loss mapping) keeps its semantics, keyed by role.DT_INFERENCE_ENDPOINT/DT_LEARNING_ENDPOINT, endpoint-name examples, dashboard lanes and both bundled recordings, demos, docs. Dashboard VERSION 0.6.0 — old recordings carry old role keys and would draw empty lanes; the badge makes that visible.inference/learning— see Change the M2 PR to use pools from Task Dispatcher #7.Verified: 164 unit + 29 integration locally; the dual-endpoint learner test proves label routing end to end (
trained_on: learning-endpoint,served_by: inference-endpoint).Phase 2 (dispatcher pools) follows per #7.
🤖 Generated with Claude Code