Update TensorBoard training metrics

benchmark_cache/shard_scheduler.json

@@ -22,8 +22,8 @@
       "tokens_per_second": 1062967.6168812835,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 75842.61374065137,
-      "required_pages_per_second_upper_estimate": 0.006573845537692509
+      "required_token_positions_per_second": 74734.02813480957,
+      "required_pages_per_second_upper_estimate": 0.006477756147062671
     },
     {
       "source": "rewrite6",
@@ -36,8 +36,8 @@
       "tokens_per_second": 1206672.263377472,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 9480.326717581422,
-      "required_pages_per_second_upper_estimate": 0.003781844071159016
+      "required_token_positions_per_second": 9341.753516851197,
+      "required_pages_per_second_upper_estimate": 0.0037265651495337468
     },
     {
       "source": "openmath",
@@ -50,8 +50,8 @@
       "tokens_per_second": 410042.7581134458,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 18960.653435162843,
-      "required_pages_per_second_upper_estimate": 0.004386213476121602
+      "required_token_positions_per_second": 18683.507033702394,
+      "required_pages_per_second_upper_estimate": 0.004322100533753801
     },
     {
       "source": "ultra_qa",
@@ -64,8 +64,8 @@
       "tokens_per_second": 2163486.5624068812,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 28440.98015274427,
-      "required_pages_per_second_upper_estimate": 0.0017319064730114617
+      "required_token_positions_per_second": 28025.26055055359,
+      "required_pages_per_second_upper_estimate": 0.0017065913303502369
     },
     {
       "source": "ultra_style",
@@ -78,8 +78,8 @@
       "tokens_per_second": 1546694.2737633225,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 28440.98015274427,
-      "required_pages_per_second_upper_estimate": 0.00218946538183024
+      "required_token_positions_per_second": 28025.26055055359,
+      "required_pages_per_second_upper_estimate": 0.002157462136068089
     },
     {
       "source": "cortex",
@@ -91,9 +91,9 @@
       "read_seconds": 0.042780299000014566,
       "tokens_per_second": 1015177.8395898147,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
-      "estimated_pages_for_buffer": 22,
-      "required_token_positions_per_second": 28440.98015274427,
-      "required_pages_per_second_upper_estimate": 0.35314178765963805
+      "estimated_pages_for_buffer": 21,
+      "required_token_positions_per_second": 28025.26055055359,
+      "required_pages_per_second_upper_estimate": 0.3479799415244371
     }
   ],
   "gpu_measurement": "Selected forward/backward estimate; replaced by live successful update timings.",
@@ -155,27 +155,27 @@
   "shard_benchmark_scope": "Local packing/read measurements; upload capacity measured only on real data commits. Targets larger than available pages are sample-limited.",
   "estimated_fresh_generation_batch_seconds": 0.1434574083060457,
   "selected": {
-    "prefetch_batches": 62,
+    "prefetch_batches": 61,
     "prefetch_capacity_limit": 512,
     "batch_memory_estimate_bytes": 690670,
-    "producer_batch_seconds": 1.1080608083517234,
-    "producer_p95_seconds": 0.32552214799943613,
-    "gpu_update_seconds": 0.9721183957625401,
-    "producer_batches_per_second": 0.9024775467760949,
-    "gpu_batches_per_second": 1.0286812844597897,
-    "producer_headroom": 0.8773150249836902,
-    "sustainable": false,
-    "buffer_seconds": 60.27134053727749,
-    "buffer_drain_seconds": 491.2691267147016,
+    "producer_batch_seconds": 0.06774875721097828,
+    "producer_p95_seconds": 0.2427683669993712,
+    "gpu_update_seconds": 0.9865385533214557,
+    "producer_batches_per_second": 14.760418362891473,
+    "gpu_batches_per_second": 1.0136451298666662,
+    "producer_headroom": 14.561721778146405,
+    "sustainable": true,
+    "buffer_seconds": 60.17885175260879,
+    "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 78,
+    "upload_pages_per_commit": 79,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 219705.554414035,
+    "generation_bytes_per_second": 219696.181290267,
     "upload_bytes_per_second": 2534349.6634966815,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 47.14369749370372,
+    "estimated_full_shards_per_hour": 47.14168623953827,
     "selection_policy": "Smallest shard within 90% of measured local peak, enlarged to amortize 30 seconds of measured page production; bounded to 256 MiB."
   },
   "upload_measurements": [

metrics/void-byte-v1/1790703713325746110-adad6d18/train.jsonl

@@ -162,3 +162,4 @@
 {"step": 296400, "loss": 0.608258318156004, "lr": 0.0003, "grad_norm": 0.14973363631122263, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 43385143040, "tokens_per_second": 150896.99945533354, "data_wait_s": 1.6546000551898032e-05, "gpu_peak_gb": 85.46991729736328}
 {"step": 296500, "loss": 0.6336312182247639, "lr": 0.0003, "grad_norm": 0.15525307091539312, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 43399781144, "tokens_per_second": 150483.4825429867, "data_wait_s": 1.135300044552423e-05, "gpu_peak_gb": 85.46991729736328}
 {"step": 296600, "loss": 0.6367071159183979, "lr": 0.0003, "grad_norm": 0.1494025677632923, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 43414573517, "tokens_per_second": 80757.99840457646, "data_wait_s": 1.1347001418471336e-05, "gpu_peak_gb": 85.46991729736328}
+{"step": 296700, "loss": 0.5971754863858223, "lr": 0.0003, "grad_norm": 0.15423460034309433, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 86, "input_bytes_and_specials": 43429289405, "tokens_per_second": 150648.87395211856, "data_wait_s": 8.986000466393307e-06, "gpu_peak_gb": 85.46991729736328}

runs/void-byte-v1/events.out.tfevents.1790703713.3e6d470a-8cc7-4d7e-a224-51a0f9a2ce55-85btc.38629.0.1790703713325746110-adad6d18

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5c72ce002bc9b309bbbcd2a51ef54318af94c36bff09eadc76570862ea81715c
-size 349866
+oid sha256:a15bc70eb3582f07976fecde6efadf9dd2af01b94e89f7f26bbda23133e13884
+size 351984