aggregate speed bench: speed tier hf+sd 1.47x; accuracy tier hf+sd+t80+avg (0.91x, cos 0.94, avg free); compile demoted

Full-ep_step wall times on quiet A6000 (warm s2000, B24), res parity across
all 8 configs. compile only 1.12x at this shape (historical 1.46x was a
different workload split); FULL cmp_sdpa saves 4% over eager at t80 — not
worth the guard complexity. tforce_sdpa added (flash baked into compiled
graph, flag-free so grad paths never see SDPA). bp_lm --tie probe in flight.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn

1 files changed, 51 insertions, 0 deletions

diff --git a/ep_run/agg_bench.py b/ep_run/agg_bench.py
new file mode 100644
index 0000000..d44b5a8
--- /dev/null
+++ b/ep_run/agg_bench.py
@@ -0,0 +1,51 @@
+"""THE AGGREGATE SPEED BENCH: full ep_step wall-time for every combination of the speed levers, on a
+quiet local GPU, warm s2000 operator, B24 (production shape). Configs:
+  base           : orig track, t2sel40, eager, manual attn      (the historical default)
+  hf             : +holofast
+  sd             : +sdpa (eager relax)
+  hf+sd          : the exact-math pack at t2sel40
+  hf+sd+t80      : the accuracy pack (cos 0.89->0.94)
+  hf+sd+t80+avg  : + holoavg (trend/plateau estimator)
+  cmp            : --compile alone (manual attn in graph)
+  cmp(sdpa)+hf+t80+avg : FULL STACK (flash baked into compiled graph)
+Reports median full-step time of 3 (after 1 warmup step each) + the step's res, as parity smoke."""
+import time, torch
+import lt_ep_train as L
+
+torch.manual_seed(0)
+blk = L.EQBlock(512, 16, 256, 256, c=1.0, attn_mode='thick'); blk.qknorm = True
+ck = torch.load('runs/redx_traj/s2000.pt', map_location=L.dev)
+with torch.no_grad():
+    for p, w in zip(blk.allp, ck['allp']):
+        p.copy_(w.to(L.dev))
+blk.track = True
+torch.manual_seed(42)
+idx, y = L.get_batch('train', 24, 256)
+
+CFG = [
+    ('base',            dict(hf=False, sd=False, t2=40,  avg=False, cmp=False)),
+    ('hf',              dict(hf=True,  sd=False, t2=40,  avg=False, cmp=False)),
+    ('sd',              dict(hf=False, sd=True,  t2=40,  avg=False, cmp=False)),
+    ('hf+sd',           dict(hf=True,  sd=True,  t2=40,  avg=False, cmp=False)),
+    ('hf+sd+t80',       dict(hf=True,  sd=True,  t2=80,  avg=False, cmp=False)),
+    ('hf+sd+t80+avg',   dict(hf=True,  sd=True,  t2=80,  avg=True,  cmp=False)),
+    ('cmp',             dict(hf=False, sd=False, t2=40,  avg=False, cmp=True)),
+    ('FULL(cmp_sdpa)',  dict(hf=True,  sd=True,  t2=80,  avg=True,  cmp=True)),
+]
+
+for name, c in CFG:
+    blk.holofast, blk.sdpa, blk.holoavg = c['hf'], c['sd'], c['avg']
+    blk._cstep = None
+    if c['cmp']:
+        _tf = blk.tforce_sdpa if c['sd'] else blk.tforce
+        blk._cstep = torch.compile(lambda z, xin, _tf=_tf: z + 0.1 * _tf(z, xin))
+    ts = []
+    for rep in range(4):                                   # rep 0 = warmup (compile/JIT)
+        torch.cuda.synchronize(); t = time.time()
+        _, res = L.ep_step(blk, idx, y, 150, 20, 0.1, 0.02, jacreg=0.1, holo=2, hr=0.02,
+                           t1max=300, res_est=1e-4, t2sel=c['t2'], corr_every=1, res_gate=0.0, resreg=0.2)
+        torch.cuda.synchronize(); ts.append(time.time() - t)
+    med = sorted(ts[1:])[1]
+    print(f"{name:>16}: {med:6.2f}s/step  (res {res:.1e})", flush=True)
+blk._cstep = None
+print("AGG_BENCH_DONE", flush=True)