Add Blockwise-OAT baseline artifacts (benchmarks/verify_policy_bs1.json)

benchmarks/verify_policy_bs1.json

@@ -0,0 +1,73 @@
+{
+  "ar_shapes": {
+    "action": [
+      1,
+      16,
+      7
+    ],
+    "action_pred": [
+      1,
+      32,
+      7
+    ]
+  },
+  "batch_size": 1,
+  "benchmark": {
+    "ar_mean_sec": 0.022294499054551123,
+    "blockwise_mean_sec": 0.019297172129154206,
+    "prefix_len": 4,
+    "speedup": 1.155324671684332,
+    "total_tokens": 8
+  },
+  "blockwise_shapes": {
+    "action": [
+      1,
+      16,
+      7
+    ],
+    "action_pred": [
+      1,
+      32,
+      7
+    ],
+    "action_tokens": [
+      1,
+      8
+    ]
+  },
+  "checkpoint": "output/baselines/original_oat/hf/policy_ep-0250_sr-0.596.ckpt",
+  "checks": {
+    "blockwise_numerics_finite": true,
+    "blockwise_shapes_match_ar": true,
+    "fallback_shapes_match_ar": true,
+    "speedup_gt_1": true,
+    "tail_param_ratio_ok": true
+  },
+  "device": "cuda:0",
+  "fallback_shapes": {
+    "action": [
+      1,
+      16,
+      7
+    ],
+    "action_pred": [
+      1,
+      32,
+      7
+    ]
+  },
+  "param_counts": {
+    "ar_model": 5023488,
+    "min_ratio_required": 0.35,
+    "tail_decoder": 4525568,
+    "tail_to_ar_ratio": 0.9008816185088926
+  },
+  "passed": true,
+  "prefix_len": 4,
+  "refine_iters": 1,
+  "sim_eval_commands": {
+    "ar": "python scripts/eval_policy_sim.py --checkpoint output/baselines/original_oat/hf/policy_ep-0250_sr-0.596.ckpt --output_dir output/eval_blockwise_verify/ar --n-test-per-task 5 --overwrite",
+    "blockwise": "python scripts/eval_policy_sim.py --checkpoint output/baselines/original_oat/hf/policy_ep-0250_sr-0.596.ckpt --output_dir output/eval_blockwise_verify/blockwise --n-test-per-task 5 --overwrite --use-blockwise --blockwise-prefix-len 4 --blockwise-refine-iters 1"
+  },
+  "tail_checkpoint": "output/blockwise/original_oat_tail_p4_r1.pt"
+}