p4_plan_sweep: drive the shipped plan() at each candidate rate to settle the session shape before D-018

kernels/p4_plan_sweep.py

@@ -0,0 +1,63 @@
+"""Session-shape sweep on free CPU: what the shipped planner actually does at each candidate rate.
+
+The D-018 decision is not only "12,977 tok/s or 9,696 tok/s" -- it is how many billed sessions the run
+becomes and whether the remaining weekly quota covers them. Both are computed by the launcher's own `plan()`,
+so this runs that function rather than re-deriving the arithmetic in prose: import the pinned launcher, walk
+the schedule leg by leg at each rate, and print the legs, the session count, the stepping hours and the
+hours the week still owes.
+
+It mutates the module globals between walks, which is exactly what the environment variables do at launch,
+so nothing here is a model of the planner -- it is the planner.
+"""
+import hashlib
+import importlib.util
+import json
+import os
+import subprocess
+import sys
+
+REV = "add04ae7a24bd9925115773a93e870fc1dd01f59"
+SHA = "5f506471141683b710f41813cc0f19fc02442f7442d395d95dd9cba73a1b760f"
+os.chdir("/kaggle/working")
+url = f"https://huggingface.co/Cion-lab/ounce100m-code/resolve/{REV}/kernels/phase4_session.py"
+assert subprocess.run(["curl", "-sfL", url, "-o", "phase4_session.py"]).returncode == 0, "fetch launcher"
+got = hashlib.sha256(open("phase4_session.py", "rb").read()).hexdigest()
+assert got == SHA, ("LAUNCHER SHA MISMATCH", got[:16], SHA[:16])
+print("pinned launcher", got[:12], flush=True)
+
+spec = importlib.util.spec_from_file_location("phase4_session", "phase4_session.py")
+S = importlib.util.module_from_spec(spec)
+spec.loader.exec_module(S)
+
+RATES = [(9358.0, "checkpointed, the Gate 3 P3 measurement"),
+         (12792.0, "no checkpointing, the Gate 3 opts cell"),
+         (12977.0, "no checkpointing, measured again by probe v8")]
+QUOTAS = [26.3, 24.5, 22.0]
+
+for rate, why in RATES:
+    S.PLANNING_TOK_PER_S = rate
+    sec_per_step = S.TOKENS_PER_STEP / rate
+    for quota in QUOTAS:
+        S.QUOTA_LEFT_HOURS = quota
+        legs, start, guard = [], 0, 0
+        while start < S.HORIZON_STEPS and guard < 30:
+            guard += 1
+            pl = S.plan(start, 19.5)
+            n = pl["planned_steps"]
+            if not n:
+                legs.append((start, "REFUSED", pl["usable_hours"]))
+                break
+            end = S.HORIZON_STEPS if pl["stop_after_steps"] == 0 else pl["stop_after_steps"]
+            legs.append((start, end, end - start))
+            start = end
+        trained = sum(l[1] - l[0] for l in legs if isinstance(l[1], int))
+        hours = trained * sec_per_step / 3600.0
+        sessions = len(legs)
+        print("RATE %6.0f (%s) QUOTA %5.1f h -> %d sessions, %d/%d steps, %.2f h stepping; %s" % (
+            rate, why, quota, sessions, trained, S.HORIZON_STEPS, hours,
+            "FITS" if trained == S.HORIZON_STEPS and hours <= quota else "DOES NOT FIT"), flush=True)
+        print("   legs", json.dumps(legs), flush=True)
+        if trained != S.HORIZON_STEPS:
+            print("   shortfall %d steps = %.2f h at this rate" % (
+                S.HORIZON_STEPS - trained, (S.HORIZON_STEPS - trained) * sec_per_step / 3600.0), flush=True)
+print("PLAN_SWEEP_DONE", flush=True)