sync from registry v1
Browse files- LIMITS.md +24 -3
- registry.yaml +3 -3
LIMITS.md
CHANGED
|
@@ -87,9 +87,23 @@ families): use `self_from_rows=True` when a strictly consistent Maxwell matrix i
|
|
| 87 |
|
| 88 |
## 4. Deployed Spaces (HF free tier, cpu-basic)
|
| 89 |
|
| 90 |
-
|
| 91 |
-
`
|
| 92 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 93 |
|
| 94 |
## 5. Known biases and reporting (review fixes, 2026-10-08)
|
| 95 |
|
|
@@ -108,6 +122,13 @@ the environment for private Spaces) and the same stress cases through gradio_cli
|
|
| 108 |
|
| 109 |
## 6. Flagship tiers: `all-data-v3s` (coarse prior, default) and `all-data-v3-rich` (fine prior), 2-thread latency
|
| 110 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 111 |
Naming (2026-10-09): **coarse prior** = one gap/3 solve on the conductor-anchored v3 mesh (was "fast"), **fine prior** =
|
| 112 |
gap/3 + gap/6 with Richardson extrapolation (was "accurate"); `predict(model="coarse" | "fine")`, the old words are kept as
|
| 113 |
aliases and in the model ids. The chip-anchored baselines are the earlier standard (8k) and earlier 16k priors (the latter
|
|
|
|
| 87 |
|
| 88 |
## 4. Deployed Spaces (HF free tier, cpu-basic)
|
| 89 |
|
| 90 |
+
Measured 2026-10-10 after the fine-tier switch to `all-data-flagship` (private Spaces, `examples/client.py` path
|
| 91 |
+
`/predict_gds`, one request at a time; server time = prepare + predict inside the Space; wall adds upload / queue). Every
|
| 92 |
+
returned matrix equals the local prediction of the same bundle to <= 3e-5 relative (max over entries).
|
| 93 |
+
|
| 94 |
+
| layout | coarse tier (`all-data-v3s`), server s | fine tier (`all-data-flagship`), server s | note |
|
| 95 |
+
|---|---|---|---|
|
| 96 |
+
| Xmon example | 1.2 | 1.5 | |
|
| 97 |
+
| interdigital example | 2.4 | 0.8 | |
|
| 98 |
+
| TransmonPocket example | 19.7 (routed to the fine tier) | 19.4 | capped coarse mesh < 0.38 panels/gap: large-layout routing |
|
| 99 |
+
| shielded concentric example | 4.8 | 79.6 | flagship H-matrix solve, Mac 1 thread 14.6 s |
|
| 100 |
+
| two pads, size 20x (stress) | 61.7 (routed) | 60.2 | |
|
| 101 |
+
| comb, 200 fingers (stress) | 224.4 (routed) | 220.0 | |
|
| 102 |
+
|
| 103 |
+
The Space CPU is 3.4-5.4x slower than the Mac at 1 thread on the flagship solves (comb 200 fingers 64.7 s, two pads 20x 13.5 s, shielded concentric 14.6 s locally). Paper-models Space: same matrices and times
|
| 104 |
+
(e.g. TransmonPocket 18.0 s fine). Gradio has no request timeout; the queue runs 2 requests at once (`TLSQ_CONCURRENCY`)
|
| 105 |
+
but at most one flagship-prior solve (`TLSQ_HEAVY_CONCURRENCY`, default 1; fine tier and coarse requests routed to it),
|
| 106 |
+
so two large solves (up to 8.3 GB each, grid_qubit_9) cannot exceed the 16 GB of cpu-basic.
|
| 107 |
|
| 108 |
## 5. Known biases and reporting (review fixes, 2026-10-08)
|
| 109 |
|
|
|
|
| 122 |
|
| 123 |
## 6. Flagship tiers: `all-data-v3s` (coarse prior, default) and `all-data-v3-rich` (fine prior), 2-thread latency
|
| 124 |
|
| 125 |
+
**Update 2026-10-10:** the fine tier is now `all-data-flagship` (same network as `all-data-v3-rich`, flagship prior;
|
| 126 |
+
deploy/MODEL_SELECTION.md Results 3, PRIOR_SCHEME.md section 8.1). Mac M5 Pro, 2 threads, whole `predict`: median /
|
| 127 |
+
p90 / max 0.87 / 11.6 / 14.0 s on the 212 held-out designs and 1.51 / 27.6 / 78.4 s on 1,158 held-out designs (former
|
| 128 |
+
fine tier 4.64 / 8.96 / 18.0 s); peak RSS <= 8.3 GB on the stress and grid layouts (grid_qubit_9, 220k panels), every
|
| 129 |
+
one returns a matrix (the resolution floor L/8192 limits comb_f400 / pads_s20 / pads_s50 to 4.65 / 4.64 / 1.97 panels
|
| 130 |
+
per gap). The table below is the former fine tier (`all-data-v3-rich`) and stays as its record.
|
| 131 |
+
|
| 132 |
Naming (2026-10-09): **coarse prior** = one gap/3 solve on the conductor-anchored v3 mesh (was "fast"), **fine prior** =
|
| 133 |
gap/3 + gap/6 with Richardson extrapolation (was "accurate"); `predict(model="coarse" | "fine")`, the old words are kept as
|
| 134 |
aliases and in the model ids. The chip-anchored baselines are the earlier standard (8k) and earlier 16k priors (the latter
|
registry.yaml
CHANGED
|
@@ -94,9 +94,9 @@ defaults:
|
|
| 94 |
tiers: # shown in the flagship UI; keys = tier names (aliases fast -> coarse, accurate -> fine still accepted by predict);
|
| 95 |
# held-out = OOD-60 + pocket3 random-36 + Qiskit Metal new 116, pooled mutual entries
|
| 96 |
coarse: {latency_s: 4, median_pct: 1.75, mae_ff: 0.534, rmse_ff: 1.173, set: "212 held-out designs (OOD-60, random TransmonPocket-3, 6 new Qiskit Metal classes)",
|
| 97 |
-
note: "On this free CPU Space:
|
| 98 |
-
fine: {latency_s:
|
| 99 |
-
note: "Flagship prior: strip-merged Richardson (6 and 3 panels across the smallest gap on every layout), H-matrix solve for large systems.
|
| 100 |
|
| 101 |
models:
|
| 102 |
# ---------------------------------------------------------------- learned, released
|
|
|
|
| 94 |
tiers: # shown in the flagship UI; keys = tier names (aliases fast -> coarse, accurate -> fine still accepted by predict);
|
| 95 |
# held-out = OOD-60 + pocket3 random-36 + Qiskit Metal new 116, pooled mutual entries
|
| 96 |
coarse: {latency_s: 4, median_pct: 1.75, mae_ff: 0.534, rmse_ff: 1.173, set: "212 held-out designs (OOD-60, random TransmonPocket-3, 6 new Qiskit Metal classes)",
|
| 97 |
+
note: "On this free CPU Space: 1-5 s for single components (measured on the examples, 2026-10-10). Layouts whose coarse mesh would fall below the training resolution (e.g. the TransmonPocket example, long combs) are predicted with the fine tier instead (reported under model used)."}
|
| 98 |
+
fine: {latency_s: 10, median_pct: 1.13, mae_ff: 0.374, rmse_ff: 0.836, set: "212 held-out designs (OOD-60, random TransmonPocket-3, 6 new Qiskit Metal classes)",
|
| 99 |
+
note: "Flagship prior: strip-merged Richardson (6 and 3 panels across the smallest gap on every layout), H-matrix solve for large systems. On this free CPU Space (2026-10-10): 1-2 s for the Xmon and interdigital examples, 20 s for the TransmonPocket, 80 s for the shielded concentric example; large layouts several minutes (200-finger comb 220 s). One fine-tier solve runs at a time; other requests are not blocked."}
|
| 100 |
|
| 101 |
models:
|
| 102 |
# ---------------------------------------------------------------- learned, released
|