Upload logs/2026-09-19_adamw_chain.log with huggingface_hub

logs/2026-09-19_adamw_chain.log

@@ -0,0 +1,19 @@
+[04:48:18] waiting for build...
+[04:48:19] binary 28300512 flags=CP_FUSED_ROPE=1|CP_ROPE_DEBUG=|CP_ATTN_SPLIT=1
+[04:48:19] waiting for downloads...
+[04:48:19] downloads done
+[04:48:19] building ablmix16 (must match p1mix token-for-token)
+
+  USABLE STEPS at window 16,384 x grad_accum 32: 1220
+    (naive total//global_batch would say 1220 -- use 1220, it accounts for the
+     per-shard tail the loader discards, and for the silent wrap)
+  wrote Data_Loader/Data/ablmix16/MANIFEST.txt
+[04:48:19] mix VERIFIED token-identical to p1mix (b3=320011377)
+[04:48:19] === LAUNCHING: 8e-5 constant, AdamW only ===
+[04:48:19] [const] === phase 1: 4k -> 16k, arm=const opt=adamw on GPU 0 ===
+[04:48:19] [const] LR      : max 8e-5  min 8e-5  warmup 12 (1% of 1200)  stable 1186
+[04:48:19] [const] optim   : AdamW only (ZeRO-1 v2)
+[04:48:19] [const] parent  : varybatch   data: Data_Loader/Data/ablmix16   steps: 1200
+[04:48:19] [const] PROBE 3 steps
+[04:49:25] [const] PROBE OK: 3 steps, peak mem: 78665MB, step-0 lr=6.6667e-06 (expect ~8.0e-05 at warmup 1/12)
+[04:49:25] [const] TRAINING 1200 steps