better

RWKV-v4neo/src/dataset.py

@@ -24,7 +24,7 @@ class MyDataset(Dataset):
 
             if args.my_pile_stage > 0:
                 assert self.data_size == 332115325534 and self.vocab_size == 50277 and args.ctx_len == 1024
-                self.samples_per_epoch = args.epoch_steps * int(args.devices) * args.micro_bsz
+                self.samples_per_epoch = args.epoch_steps * args.real_bsz
                 assert self.samples_per_epoch == 40320
                 print(f"########## Pile 20b-tokenized stage {args.my_pile_stage} ##########")
                 self.magic_prime = 324331313

RWKV-v4neo/src/model.py

@@ -280,10 +280,16 @@ class RWKV(pl.LightningModule):
             lr_2x = set()
             lr_3x = set()
             for n, p in self.named_parameters():
-                if ("time_mix" in n) and (self.args.my_pile_stage == 2):
-                    lr_2x.add(n)
+                if "time_mix" in n:
+                    if self.args.my_pile_stage == 2:
+                        lr_2x.add(n)
+                    else:
+                        lr_1x.add(n)
                 elif "time_decay" in n:
-                    lr_2x.add(n)
+                    if self.args.my_pile_stage == 2:
+                        lr_3x.add(n)
+                    else:
+                        lr_2x.add(n)
                 elif "time_first" in n:
                     lr_3x.add(n)
                 else:
@@ -295,11 +301,18 @@ class RWKV(pl.LightningModule):
             # print('2x', lr_2x)
             # print('3x', lr_3x)
             param_dict = {n: p for n, p in self.named_parameters()}
-            optim_groups = [
-                {"params": [param_dict[n] for n in lr_1x], "weight_decay": 0.0, "my_lr_scale": 1.0},
-                {"params": [param_dict[n] for n in lr_2x], "weight_decay": 0.0, "my_lr_scale": 2.0},
-                {"params": [param_dict[n] for n in lr_3x], "weight_decay": 0.0, "my_lr_scale": 3.0},
-            ]
+            if self.args.my_pile_stage == 2:
+                optim_groups = [
+                    {"params": [param_dict[n] for n in lr_1x], "weight_decay": 0.0, "my_lr_scale": 1.0},
+                    {"params": [param_dict[n] for n in lr_2x], "weight_decay": 0.0, "my_lr_scale": 2.0},
+                    {"params": [param_dict[n] for n in lr_3x], "weight_decay": 0.0, "my_lr_scale": 5.0},
+                ]
+            else:
+                optim_groups = [
+                    {"params": [param_dict[n] for n in lr_1x], "weight_decay": 0.0, "my_lr_scale": 1.0},
+                    {"params": [param_dict[n] for n in lr_2x], "weight_decay": 0.0, "my_lr_scale": 2.0},
+                    {"params": [param_dict[n] for n in lr_3x], "weight_decay": 0.0, "my_lr_scale": 3.0},
+                ]
         else:
             optim_groups = [
                 {"params": [p for n, p in self.named_parameters()], "weight_decay": 0.0},

RWKV-v4neo/src/trainer.py

@@ -34,13 +34,7 @@ class train_callback(pl.Callback):
 
         for param_group in trainer.optimizers[0].param_groups:
             if args.layerwise_lr > 0:
-                if self.args.my_pile_stage != 2:
-                    param_group["lr"] = lr * param_group["my_lr_scale"]
-                else:
-                    if param_group["my_lr_scale"] > 1:
-                        param_group["lr"] = lr * 5
-                    else:
-                        param_group["lr"] = lr
+                param_group["lr"] = lr * param_group["my_lr_scale"]
                 # print(param_group["lr"], param_group["my_lr_scale"])
             else:
                 param_group["lr"] = lr