[submit] 670004fca640 — val by hotkey 5GW4nyVgq36T…

configs/beat/v9_4750.json

@@ -0,0 +1,41 @@
+{
+  "vocab_size": 50257,
+  "dim": 1024,
+  "n_layers": 16,
+  "n_heads": 16,
+  "head_dim": 64,
+  "ffn_mult": 2.6667,
+  "max_seq_len": 1024,
+  "seq_len": 512,
+  "batch_size": 1024,
+  "micro_batch_size": 128,
+  "total_steps": 4750,
+  "warmup_steps": 700,
+  "max_lr": 0.003,
+  "min_lr": 4e-06,
+  "weight_decay": 0.1,
+  "beta1": 0.9,
+  "beta2": 0.95,
+  "grad_clip": 1.0,
+  "schedule": "wsd",
+  "stable_frac": 0.5,
+  "decay_frac": 0.5,
+  "decay_curve": "1-sqrt",
+  "embed_lr": 0.009,
+  "embed_optimizer": "adamw",
+  "optimizer": "muon",
+  "muon_lr": 0.04,
+  "muon_momentum": 0.95,
+  "muon_ns_steps": 5,
+  "muon_weight_decay": 0.05,
+  "muon_momentum_start": null,
+  "data_seed": 42,
+  "init_seed": 42,
+  "use_bf16": true,
+  "compile": true,
+  "log_every": 100,
+  "ema_decay": 0.0,
+  "ema_start_frac": 0.0,
+  "deterministic": false,
+  "compile_mode": "max-autotune-no-cudagraphs"
+}

model/__init__.py

@@ -1,4 +1,4 @@
-from ._v7lite import KarpaBase, KarpaConfig, RalphBase, RalphConfig
+from ._v9ve import KarpaBase, KarpaConfig, RalphBase, RalphConfig
 
 # RalphBase/RalphConfig are canonical; KarpaBase/KarpaConfig are back-compat
 # aliases retained through the karpa->ralph rebrand (see ralph_base.py).