msc: batch 16 files (29415 KB) @ 2026-08-04T04:37:25Z

registry/events/acct1_w0_Interactive-caf2ce3073.jsonl

@@ -3,3 +3,4 @@
 {"run_id": "p1-vgg13-cifar100-base-s2", "state": "paused", "account": "acct1", "worker_id": 0, "session_id": "Interactive-caf2ce3073", "updated_at": "2026-08-04T04:23:52Z", "ts": 1785817432.3195412, "best_accuracy": 0.5996, "last_completed_epoch": 101, "demoted_broken_stub": true, "arch": "vgg13", "seed": 2, "dataset": "cifar100", "phase": "p1"}
 {"run_id": "p1-vit_tiny-cifar100-base-s1", "state": "paused", "account": "acct1", "worker_id": 0, "session_id": "Interactive-caf2ce3073", "updated_at": "2026-08-04T04:23:52Z", "ts": 1785817432.3619242, "best_accuracy": 0.5641, "last_completed_epoch": 54, "demoted_broken_stub": true, "arch": "vit_tiny", "seed": 1, "dataset": "cifar100", "phase": "p1"}
 {"run_id": "p1-mixer_nano-cifar100-base-s2", "state": "oracle_done", "account": "acct1", "worker_id": 0, "session_id": "Interactive-caf2ce3073", "updated_at": "2026-08-04T04:27:26Z", "ts": 1785817646.0024045, "arch": "mixer_nano", "seed": 2, "sample_order_hash": "80031c23f8300724c6fbc588e4bccbe41c6392041782148727bb497b0132c4ad"}
+{"run_id": "p1-mobilenetv2-cifar100-base-s2", "state": "oracle_done", "account": "acct1", "worker_id": 0, "session_id": "Interactive-caf2ce3073", "updated_at": "2026-08-04T04:33:38Z", "ts": 1785818018.9401884, "arch": "mobilenetv2", "seed": 2, "sample_order_hash": "80031c23f8300724c6fbc588e4bccbe41c6392041782148727bb497b0132c4ad"}

runs/p1-resnet20-cifar100-base-s2/per_sample/meta.json

@@ -0,0 +1,154 @@
+{
+  "run_id": "p1-resnet20-cifar100-base-s2",
+  "arch": "resnet20",
+  "family": "resnet",
+  "dataset": "cifar100",
+  "seed": 2,
+  "sample_order_hash": "80031c23f8300724c6fbc588e4bccbe41c6392041782148727bb497b0132c4ad",
+  "config_hash": "1218d3f1c0ca3daa911ffeb051e9167d525844c56744a8df2a8b33aeecf7eea0",
+  "budgets": {
+    "depth": {
+      "configs": [
+        "d1",
+        "d2",
+        "d3",
+        "d4",
+        "d5"
+      ],
+      "K": 5,
+      "fractions": [
+        0.2222222222222222,
+        0.4444444444444444,
+        0.5555555555555556,
+        0.7777777777777778,
+        1.0
+      ],
+      "requested_fractions": [
+        0.2,
+        0.4,
+        0.6,
+        0.8,
+        1.0
+      ],
+      "stage_cuts": [
+        2,
+        4,
+        5,
+        7,
+        9
+      ],
+      "n_blocks": 9,
+      "feature_dims": [
+        16,
+        32,
+        32,
+        64,
+        64
+      ],
+      "flops": [
+        20122816,
+        37116288,
+        46619008,
+        63509248,
+        82449152
+      ],
+      "rho": [
+        0.24406334706753563,
+        0.4501718586505292,
+        0.5654273800172014,
+        0.7702838229312534,
+        1.0
+      ],
+      "note": "prefix backbone + linear exit head; forward_prefix stops early. K is adaptive: a backbone with fewer blocks than requested exits carries fewer distinct depth budgets."
+    },
+    "resolution": {
+      "configs": [
+        "r16",
+        "r20",
+        "r24",
+        "r28",
+        "r32"
+      ],
+      "values": [
+        16,
+        20,
+        24,
+        28,
+        32
+      ],
+      "flops": [
+        20621824,
+        32214400,
+        46383104,
+        63127936,
+        82448896
+      ],
+      "rho": [
+        0.2501164357616141,
+        0.39071960405631145,
+        0.5625679208609415,
+        0.7656613861755044,
+        1.0
+      ],
+      "native_supported": true,
+      "native_error": null,
+      "note": "cost measured at NATIVE input size where the architecture tolerates it; otherwise an analytic quadratic-in-r model. The proxy sweep (downsample-then-upsample to 32px) shares this cost table and is labelled idealised."
+    },
+    "precision": {
+      "configs": [
+        "int4",
+        "int6",
+        "int8",
+        "fp16",
+        "fp32"
+      ],
+      "bits": [
+        4,
+        6,
+        8,
+        16,
+        32
+      ],
+      "flops": [
+        10306112,
+        15459168,
+        20612224,
+        41224448,
+        82448896
+      ],
+      "rho": [
+        0.125,
+        0.1875,
+        0.25,
+        0.5,
+        1.0
+      ],
+      "note": "analytic bit-operation model rho = bits/32. INT4/INT6 are simulated by fake quantisation; no T4 kernel exists to time. Never reported as measured latency."
+    }
+  },
+  "full_flops": 82448896,
+  "exit_count": 5,
+  "resolutions": [
+    16,
+    20,
+    24,
+    28,
+    32
+  ],
+  "precisions": [
+    "int4",
+    "int6",
+    "int8",
+    "fp16",
+    "fp32"
+  ],
+  "tau_grid": [
+    0.0,
+    0.1,
+    0.2,
+    0.3,
+    0.5
+  ],
+  "created_utc": "2026-08-04T04:37:24Z",
+  "msc_lib_version": "1.0.0"
+}

runs/p1-resnet20-cifar100-base-s2/per_sample/test.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8bcd4247cd8b559d9679b102abd208c9e8800f8e0a1489cf4df6982ea5491dc7
+size 2816091

runs/p1-resnet20-cifar100-base-s2/per_sample/train_holdout.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf277ddec19caec7c0361b821ce71e464ddef420fffa2939e59e923621fa3512
+size 1438944