[Scheduler] Separate request lifecycle from scheduled work

benchmarks/cdb_runtime.py

@@ -28,7 +28,7 @@ def drive(engine):
         if batch is not None:
             steps.append(
                 dict(
-                    stage=batch.phase.value,
+                    works=sorted(work.name.lower() for work in batch.works),
                     effective_rows=batch.num_tokens,
                     submitted_rows=engine.model_runner.last_submitted_size,
                 )

benchmarks/preemption.py

@@ -98,7 +98,9 @@ def drive(engine, requests, state):
             next_request = next(pending, None)
         state["max_waiting"] = max(
             state["max_waiting"],
-            sum(request.phase.value == "waiting" for request in engine.scheduler.requests.values()),
+            sum(
+                request.status.value == "waiting" for request in engine.scheduler.requests.values()
+            ),
         )
         if engine.has_unfinished_requests():
             for output in engine.step():

docs/diagrams/README.md

@@ -34,7 +34,7 @@ do not imply that every submission executes all three operations.
 
 `scheduler-overview.svg` is the editable 1800 × 2600 source and
 `scheduler-overview.png` its 2× export. The vertical A–F spine shows admission,
-phase queues, policy/batch selection, execution, completion tracking and progress.
+WorkKey queues, policy/batch selection, execution, completion tracking and progress.
 Read top to bottom; one return arrow on the left leads from F to B. A includes waiting, safe preemption and
 restoration; KV growth pressure in C returns to the same preemption logic. C
 separates prefill granularity from batching and shows supported combinations.

docs/diagrams/index.html

@@ -28,7 +28,7 @@
 <section class="panel card" id="overview" role="tabpanel" aria-labelledby="tab-overview" hidden>
 <div class="bar"><div><h2>Scheduler overview</h2><p class="desc">Follow A–F from top to bottom, then return to B along the left edge. Admission includes waiting, preemption and restoration; C separates prefill granularity from batching.</p></div><div class="links"><a href="scheduler-overview.svg" target="_blank" rel="noopener">Open SVG</a><a href="scheduler-overview.png" target="_blank" rel="noopener">High-resolution PNG</a></div></div>
 <div class="tools"><label>Zoom <input type="range" min="60" max="200" value="100" step="10" aria-label="Zoom scheduler overview"></label><output>100%</output><button class="fit">Fit width</button><a href="#scheduler" onclick="select('scheduler')">Expand Mixed: loop scheduling</a></div>
-<div class="canvas"><img src="scheduler-overview.svg" alt="Complete scheduler paths: admission, phase queues, policy selection, ordinary/speculative execution, pending results, PD transfer and retirement."></div></section>
+<div class="canvas"><img src="scheduler-overview.svg" alt="Complete scheduler paths: admission, WorkKey queues, policy selection, ordinary/speculative execution, pending results, PD transfer and retirement."></div></section>
 <section class="panel card" id="scheduler" role="tabpanel" aria-labelledby="tab-scheduler" hidden>
 <div class="bar"><div><h2>Loop-level scheduler</h2><p class="desc">Select individual loops, mix compatible requests, then regroup from independent progress.</p></div><div class="links"><a href="loop-scheduler.svg" target="_blank" rel="noopener">Open SVG</a><a href="loop-scheduler.png" target="_blank" rel="noopener">High-resolution PNG</a></div></div>
 <div class="tools"><label>Zoom <input type="range" min="60" max="200" value="100" step="10" aria-label="Zoom scheduler diagram"></label><output>100%</output><button class="fit">Fit width</button></div>