Update TensorBoard training metrics

benchmark_cache/shard_scheduler.json

@@ -22,8 +22,8 @@
       "tokens_per_second": 598315.063954628,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 76227.7126162269,
-      "required_pages_per_second_upper_estimate": 0.005273610031452144
+      "required_token_positions_per_second": 76620.97844411158,
+      "required_pages_per_second_upper_estimate": 0.005300817073927645
     },
     {
       "source": "rewrite6",
@@ -36,8 +36,8 @@
       "tokens_per_second": 8533.957213384612,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 9528.464077028362,
-      "required_pages_per_second_upper_estimate": 0.003989368932095423
+      "required_token_positions_per_second": 9577.622305513947,
+      "required_pages_per_second_upper_estimate": 0.004009950455821794
     },
     {
       "source": "openmath",
@@ -50,8 +50,8 @@
       "tokens_per_second": 380144.35761751566,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 19056.928154056724,
-      "required_pages_per_second_upper_estimate": 0.003969055569498561
+      "required_token_positions_per_second": 19155.244611027894,
+      "required_pages_per_second_upper_estimate": 0.003989532294706328
     },
     {
       "source": "ultra_qa",
@@ -64,8 +64,8 @@
       "tokens_per_second": 43742.16556775435,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 28585.392231085087,
-      "required_pages_per_second_upper_estimate": 0.0017136506168868785
+      "required_token_positions_per_second": 28732.86691654184,
+      "required_pages_per_second_upper_estimate": 0.0017224914990992047
     },
     {
       "source": "ultra_style",
@@ -78,8 +78,8 @@
       "tokens_per_second": 22018.86276523721,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 1,
-      "required_token_positions_per_second": 28585.392231085087,
-      "required_pages_per_second_upper_estimate": 0.002222388512114404
+      "required_token_positions_per_second": 28732.86691654184,
+      "required_pages_per_second_upper_estimate": 0.0022338540202360793
     },
     {
       "source": "cortex",
@@ -92,8 +92,8 @@
       "tokens_per_second": 910315.184299418,
       "note": "Bounded fresh sample; full-page latency is an extrapolation.",
       "estimated_pages_for_buffer": 23,
-      "required_token_positions_per_second": 28585.392231085087,
-      "required_pages_per_second_upper_estimate": 0.36666271893748265
+      "required_token_positions_per_second": 28732.86691654184,
+      "required_pages_per_second_upper_estimate": 0.3685543658565416
     }
   ],
   "gpu_measurement": "Selected forward/backward estimate; replaced by live successful update timings.",
@@ -158,24 +158,24 @@
     "prefetch_batches": 63,
     "prefetch_capacity_limit": 512,
     "batch_memory_estimate_bytes": 720750,
-    "producer_batch_seconds": 0.14673761170394073,
-    "producer_p95_seconds": 0.3523163350037066,
-    "gpu_update_seconds": 0.967207298626264,
-    "producer_batches_per_second": 6.8148853479884215,
-    "gpu_batches_per_second": 1.0339045222470011,
-    "producer_headroom": 6.5914068478755885,
+    "producer_batch_seconds": 0.09744666843766936,
+    "producer_p95_seconds": 0.3500157450034749,
+    "gpu_update_seconds": 0.9622429978987836,
+    "producer_batches_per_second": 10.262023484565185,
+    "gpu_batches_per_second": 1.0392385314142738,
+    "producer_headroom": 9.874560242295725,
     "sustainable": true,
-    "buffer_seconds": 60.93405981345463,
+    "buffer_seconds": 60.62130886762337,
     "buffer_drain_seconds": null,
     "shard_target_mib": 16,
-    "upload_pages_per_commit": 75,
+    "upload_pages_per_commit": 87,
     "max_pending_pages": 4096,
-    "generation_bytes_per_second": 311757.3046127759,
+    "generation_bytes_per_second": 348840.80833099736,
     "upload_bytes_per_second": 4794636.860081473,
     "upload_sustainable": true,
     "logical_page_rows": 4096,
     "logical_cortex_page_rows": 128,
-    "estimated_full_shards_per_hour": 66.89586023127993,
+    "estimated_full_shards_per_hour": 74.85311686942521,
     "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-v1/1791039222312944374-63c10d34/train.jsonl

@@ -374,3 +374,4 @@
 {"step": 641800, "loss": 0.5942527055740356, "lr": 0.0001, "grad_norm": 0.1982612733062987, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 93942907709, "tokens_per_second": 129289.82179239734, "data_wait_s": 1.2977994629181921e-05, "gpu_peak_gb": 89.37861251831055}
 {"step": 641900, "loss": 0.5198739171028137, "lr": 0.0001, "grad_norm": 0.1751701157317937, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 93957429030, "tokens_per_second": 151414.70536737356, "data_wait_s": 8.953997166827321e-06, "gpu_peak_gb": 89.37861251831055}
 {"step": 642000, "loss": 0.5328903794288635, "lr": 0.0001, "grad_norm": 0.16293173309692888, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 93971947121, "tokens_per_second": 150037.77706439354, "data_wait_s": 1.1133000953122973e-05, "gpu_peak_gb": 89.37861251831055}
+{"step": 642100, "loss": 0.5516042709350586, "lr": 0.0001, "grad_norm": 0.188053310168775, "recovery_retries": 0, "loss_scale": 1.0, "microbatch": 90, "input_bytes_and_specials": 93986618751, "tokens_per_second": 153111.63567327755, "data_wait_s": 1.1608004570007324e-05, "gpu_peak_gb": 89.37861251831055}

runs/void-byte-v1/events.out.tfevents.1791039222.f2a393d4-e1e4-407b-a6aa-034aa2ada5ac-vwfgg.353.0.1791039222312944374-63c10d34

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:eaf1e60017b47457c8dd0ac278185dbe33a0ee6231221aefc9d356ef3d195f0e
-size 802797
+oid sha256:273abd1e6c359399b985a85d2cf929b789520bae76566ac9dd3ac9250e423a2a
+size 804915