Fix: fastwam training lr_eta_min

lightx2v_train/configs/train/fastwam/libero_plus_uncond_2cam224_4gpu.yaml

@@ -31,6 +31,8 @@ training:
   save_final: true
   lr_scheduler: cosine
   lr_warmup_iters: 8742
+  # Cosine floor = optimizer.learning_rate * lr_eta_min_ratio; set to 0 to decay to zero.
+  lr_eta_min_ratio: 0.01
   optimizer:
     learning_rate: 1.0e-4
     weight_decay: 1.0e-2

lightx2v_train/configs/train/fastwam/libero_plus_uncond_2cam224_8gpu.yaml

@@ -31,6 +31,8 @@ training:
   save_final: true
   lr_scheduler: cosine
   lr_warmup_iters: 8742
+  # Cosine floor = optimizer.learning_rate * lr_eta_min_ratio; set to 0 to decay to zero.
+  lr_eta_min_ratio: 0.01
   optimizer:
     learning_rate: 1.0e-4
     weight_decay: 1.0e-2

lightx2v_train/configs/train/fastwam/libero_uncond_2cam224.yaml

@@ -29,6 +29,8 @@ training:
   save_final: true
   lr_scheduler: cosine
   lr_warmup_iters: 1085
+  # Cosine floor = optimizer.learning_rate * lr_eta_min_ratio; set to 0 to decay to zero.
+  lr_eta_min_ratio: 0.01
   optimizer:
     learning_rate: 1.0e-4
     weight_decay: 1.0e-2

lightx2v_train/lightx2v_train/trainers/fastwam.py

@@ -9,7 +9,7 @@
 from PIL import Image, ImageDraw
 from loguru import logger
 from torch.nn.parallel import DistributedDataParallel
-from torch.optim.lr_scheduler import CosineAnnealingLR, LinearLR, SequentialLR
+from torch.optim.lr_scheduler import ConstantLR, CosineAnnealingLR, LinearLR, SequentialLR
 
 from lightx2v_train.runtime.checkpoint import find_latest_checkpoint, parse_checkpoint_iteration, prune_checkpoints
 from lightx2v_train.runtime.distributed import (
@@ -87,6 +87,7 @@ def __init__(self, config):
         self.save_final = bool(self.training_config.get("save_final", True))
         self.lr_scheduler_name = self.training_config.get("lr_scheduler", "constant")
         self.lr_warmup_iters = int(self.training_config.get("lr_warmup_iters", 0))
+        self.lr_eta_min_ratio = float(self.training_config.get("lr_eta_min_ratio", 0.01))
         self.train_log_every_iters = max(1, int(self.logging_config.get("train_log_every_iters", 10)))
 
         zero1_config = self.config.get("distributed", {}).get("zero1", {})
@@ -153,23 +154,37 @@ def _build_optimizer(self):
         return torch.optim.AdamW(self.trainable_params, **optimizer_kwargs)
 
     def _build_lr_scheduler(self):
+        # LinearLR changes the optimizer's current LR during construction. Capture
+        # the configured base LR first so cosine eta_min is not scaled by warmup.
+        base_lr = float(self.optimizer.param_groups[0]["lr"])
+        remaining_iters = self.max_train_iters - self.lr_warmup_iters
+        if self.lr_scheduler_name == "cosine":
+            main_scheduler = CosineAnnealingLR(
+                self.optimizer,
+                T_max=remaining_iters,
+                eta_min=base_lr * self.lr_eta_min_ratio,
+            )
+        else:
+            main_scheduler = ConstantLR(
+                self.optimizer,
+                factor=1.0,
+                total_iters=remaining_iters,
+            )
+
+        if self.lr_warmup_iters == 0:
+            return main_scheduler
+
         warmup_scheduler = LinearLR(
             self.optimizer,
-            start_factor=1e-8,
+            start_factor=1.0 / self.lr_warmup_iters,
             end_factor=1.0,
             total_iters=self.lr_warmup_iters,
         )
-        cosine_scheduler = CosineAnnealingLR(
-            self.optimizer,
-            T_max=self.max_train_iters - self.lr_warmup_iters,
-            eta_min=self.optimizer.param_groups[0]["lr"] * 0.01,
-        )
-        scheduler = SequentialLR(
+        return SequentialLR(
             self.optimizer,
-            schedulers=[warmup_scheduler, cosine_scheduler],
+            schedulers=[warmup_scheduler, main_scheduler],
             milestones=[self.lr_warmup_iters],
         )
-        return scheduler
 
     def setup(self, resume_ckpt_path=None):
         self.model.set_dit_only_trainable()