[Trainer] training with proper offloading (#12189)

* Fix bypass dtype/device moving

* Force offloading mode for training

* training context var

* offloading implementation in training node

* fix wrong input type

* Support bypass load lora model, correct adapter/offloading handling

comfy/ldm/flux/math.py

@@ -29,19 +29,34 @@ def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
     return out.to(dtype=torch.float32, device=pos.device)
 
 
+def _apply_rope1(x: Tensor, freqs_cis: Tensor):
+    x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2)
+
+    x_out = freqs_cis[..., 0] * x_[..., 0]
+    x_out.addcmul_(freqs_cis[..., 1], x_[..., 1])
+
+    return x_out.reshape(*x.shape).type_as(x)
+
+
+def _apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor):
+    return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis)
+
+
 try:
     import comfy.quant_ops
-    apply_rope = comfy.quant_ops.ck.apply_rope
-    apply_rope1 = comfy.quant_ops.ck.apply_rope1
+    q_apply_rope = comfy.quant_ops.ck.apply_rope
+    q_apply_rope1 = comfy.quant_ops.ck.apply_rope1
+    def apply_rope(xq, xk, freqs_cis):
+        if comfy.model_management.in_training:
+            return _apply_rope(xq, xk, freqs_cis)
+        else:
+            return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis)
+    def apply_rope1(x, freqs_cis):
+        if comfy.model_management.in_training:
+            return _apply_rope1(x, freqs_cis)
+        else:
+            return q_apply_rope1(x, freqs_cis)
 except:
     logging.warning("No comfy kitchen, using old apply_rope functions.")
-    def apply_rope1(x: Tensor, freqs_cis: Tensor):
-        x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2)
-
-        x_out = freqs_cis[..., 0] * x_[..., 0]
-        x_out.addcmul_(freqs_cis[..., 1], x_[..., 1])
-
-        return x_out.reshape(*x.shape).type_as(x)
-
-    def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor):
-        return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis)
+    apply_rope = _apply_rope
+    apply_rope1 = _apply_rope1

comfy/model_management.py

@@ -55,6 +55,11 @@ cpu_state = CPUState.GPU
 
 total_vram = 0
 
+
+# Training Related State
+in_training = False
+
+
 def get_supported_float8_types():
     float8_types = []
     try:

comfy/sampler_helpers.py

@@ -122,20 +122,26 @@ def estimate_memory(model, noise_shape, conds):
     minimum_memory_required = model.model.memory_required([noise_shape[0]] + list(noise_shape[1:]), cond_shapes=cond_shapes_min)
     return memory_required, minimum_memory_required
 
-def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False):
+def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False, force_offload=False):
     executor = comfy.patcher_extension.WrapperExecutor.new_executor(
         _prepare_sampling,
         comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.PREPARE_SAMPLING, model_options, is_model_options=True)
     )
-    return executor.execute(model, noise_shape, conds, model_options=model_options, force_full_load=force_full_load)
+    return executor.execute(model, noise_shape, conds, model_options=model_options, force_full_load=force_full_load, force_offload=force_offload)
 
-def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False):
+def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False, force_offload=False):
     real_model: BaseModel = None
     models, inference_memory = get_additional_models(conds, model.model_dtype())
     models += get_additional_models_from_model_options(model_options)
     models += model.get_nested_additional_models()  # TODO: does this require inference_memory update?
-    memory_required, minimum_memory_required = estimate_memory(model, noise_shape, conds)
-    comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required + inference_memory, minimum_memory_required=minimum_memory_required + inference_memory, force_full_load=force_full_load)
+    if force_offload: # In training + offload enabled, we want to force prepare sampling to trigger partial load
+        memory_required = 1e20
+        minimum_memory_required = None
+    else:
+        memory_required, minimum_memory_required = estimate_memory(model, noise_shape, conds)
+        memory_required += inference_memory
+        minimum_memory_required += inference_memory
+    comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required, minimum_memory_required=minimum_memory_required, force_full_load=force_full_load)
     real_model = model.model
 
     return real_model, conds, models

comfy/weight_adapter/bypass.py

@@ -21,6 +21,7 @@ from typing import Optional, Union
 import torch
 import torch.nn as nn
 
+import comfy.model_management
 from .base import WeightAdapterBase, WeightAdapterTrainBase
 from comfy.patcher_extension import PatcherInjection
 
@@ -181,18 +182,21 @@ class BypassForwardHook:
             )
             return  # Already injected
 
-        # Move adapter weights to module's device to avoid CPU-GPU transfer on every forward
-        device = None
+        # Move adapter weights to compute device (GPU)
+        # Use get_torch_device() instead of module.weight.device because
+        # with offloading, module weights may be on CPU while compute happens on GPU
+        device = comfy.model_management.get_torch_device()
+
+        # Get dtype from module weight if available
         dtype = None
         if hasattr(self.module, "weight") and self.module.weight is not None:
-            device = self.module.weight.device
             dtype = self.module.weight.dtype
-        elif hasattr(self.module, "W_q"):  # Quantized layers might use different attr
-            device = self.module.W_q.device
-            dtype = self.module.W_q.dtype
 
-        if device is not None:
-            self._move_adapter_weights_to_device(device, dtype)
+        # Only use dtype if it's a standard float type, not quantized
+        if dtype is not None and dtype not in (torch.float32, torch.float16, torch.bfloat16):
+            dtype = None
+
+        self._move_adapter_weights_to_device(device, dtype)
 
         self.original_forward = self.module.forward
         self.module.forward = self._bypass_forward