remove force gradient checkpointing (#1589)

mi804 · web-flow · commit 9725d31103aa · 2026-08-13T11:46:23.000+08:00
diff --git a/examples/lingbot_video/model_training/train.py b/examples/lingbot_video/model_training/train.py
@@ -1,4 +1,4 @@
-import torch, os, argparse, accelerate, warnings
+import torch, os, argparse, accelerate
 from diffsynth.core import UnifiedDataset
 from diffsynth.pipelines.lingbot_video import LingBotVideoPipeline
 from diffsynth.diffusion import *
@@ -26,11 +26,6 @@ def __init__(
         min_timestep_boundary=0.0,
     ):
         super().__init__()
-        # Warning
-        if not use_gradient_checkpointing:
-            warnings.warn("Gradient checkpointing is detected as disabled. To prevent out-of-memory errors, the training framework will forcibly enable gradient checkpointing.")
-            use_gradient_checkpointing = True
-
         # Load models. The Qwen3-VL processor (tokenizer + image/video processor) is
         # passed separately via `processor_config`, mirroring the inference pipeline.
         model_configs = self.parse_model_configs(model_paths, model_id_with_origin_paths, fp8_models=fp8_models, offload_models=offload_models, device=device)
diff --git a/examples/ltx2/model_training/train.py b/examples/ltx2/model_training/train.py
@@ -1,4 +1,4 @@
-import torch, os, argparse, accelerate, warnings
+import torch, os, argparse, accelerate
 from diffsynth.core import UnifiedDataset
 from diffsynth.core.data.operators import LoadAudioWithTorchaudio, ToAbsolutePath, RouteByType, SequencialProcess
 from diffsynth.pipelines.ltx2_audio_video import LTX2AudioVideoPipeline, ModelConfig
@@ -24,11 +24,6 @@ def __init__(
         task="sft",
     ):
         super().__init__()
-        # Warning
-        if not use_gradient_checkpointing:
-            warnings.warn("Gradient checkpointing is detected as disabled. To prevent out-of-memory errors, the training framework will forcibly enable gradient checkpointing.")
-            use_gradient_checkpointing = True
-
         # Load models
         model_configs = self.parse_model_configs(model_paths, model_id_with_origin_paths, fp8_models=fp8_models, offload_models=offload_models, device=device)
         tokenizer_config = ModelConfig(model_id="google/gemma-3-12b-it-qat-q4_0-unquantized") if tokenizer_path is None else ModelConfig(tokenizer_path)
diff --git a/examples/minimax_h3/model_training/train.py b/examples/minimax_h3/model_training/train.py
@@ -1,4 +1,4 @@
-import torch, os, argparse, accelerate, warnings
+import torch, os, argparse, accelerate
 from diffsynth.core import UnifiedDataset
 from diffsynth.core.data.operators import LoadAudioWithTorchaudio, ToAbsolutePath
 from diffsynth.utils.data.minimax_h3 import MiniMaxH3ReferenceLoader
@@ -31,11 +31,6 @@ def __init__(
         task="sft",
     ):
         super().__init__()
-        # Warning
-        if not use_gradient_checkpointing:
-            warnings.warn("Gradient checkpointing is detected as disabled. To prevent out-of-memory errors, the training framework will forcibly enable gradient checkpointing.")
-            use_gradient_checkpointing = True
-
         # Load models
         model_configs = self.parse_model_configs(model_paths, model_id_with_origin_paths, fp8_models=fp8_models, offload_models=offload_models, device=device)
         pipe_kwargs = {}
diff --git a/examples/mova/model_training/train.py b/examples/mova/model_training/train.py
@@ -1,4 +1,4 @@
-import torch, os, argparse, accelerate, warnings
+import torch, os, argparse, accelerate
 from diffsynth.core import UnifiedDataset
 from diffsynth.core.data.operators import LoadAudioWithTorchaudio, ToAbsolutePath, RouteByType, SequencialProcess
 from diffsynth.pipelines.mova_audio_video import MovaAudioVideoPipeline, ModelConfig
@@ -26,11 +26,6 @@ def __init__(
         min_timestep_boundary=0.0,
     ):
         super().__init__()
-        # Warning
-        if not use_gradient_checkpointing:
-            warnings.warn("Gradient checkpointing is detected as disabled. To prevent out-of-memory errors, the training framework will forcibly enable gradient checkpointing.")
-            use_gradient_checkpointing = True
-
         # Load models
         model_configs = self.parse_model_configs(model_paths, model_id_with_origin_paths, fp8_models=fp8_models, offload_models=offload_models, device=device)
         tokenizer_config = ModelConfig(model_id="google/gemma-3-12b-it-qat-q4_0-unquantized") if tokenizer_path is None else ModelConfig(tokenizer_path)
