anima_closeout v11: first live run of bf16_master_weights crashed at the scheduler — ds_config['bf16'] made DeepSpeed wrap the client Adam (FP16_UnfusedOptimizer), which torch ConstantLR rejects. Fork 812449a: client-side MasterWeightsAdam (fp32 masters, still a torch Adam; engine stays on the exp004-proven no-wrapping path). Proven in pure torch incl. the exact exp004 freeze repro: -3.0 bf16 gate frozen under plain Adam, moves under MasterWeightsAdam. CELL 0 asserts the new mechanism. Caches from the 28-min build are all banked — next launch goes straight to training.

colab/anima_closeout.ipynb

@@ -128,6 +128,11 @@
     "assert \"pa.chunked_array(chunks)\" in \\\n",
     "    (FORK / \"utils\" / \"parquet_source.py\").read_text(), \\\n",
     "    \"fork main lacks the per-row-group ParquetImageReader (parquet_source.py)\"\n",
+    "# bf16_master_weights must be the CLIENT-SIDE mechanism (fork 812449a):\n",
+    "# the ds_config['bf16'] route wraps the optimizer and torch schedulers\n",
+    "# reject the wrapper ('FP16_UnfusedOptimizer is not an Optimizer')\n",
+    "assert \"MasterWeightsAdam\" in _src, \\\n",
+    "    \"fork main lacks the client-side MasterWeightsAdam (bf16_master_weights)\"\n",
     "\n",
     "if not ANIMA.get(\"DEPS_DONE\"):\n",
     "    sh(f\"pip install -q -r {FORK}/requirements.txt\")\n",