Skip to content

finetune

Supervised finetuning method (algorithm layer).

Classes

fastvideo.train.methods.fine_tuning.finetune.FineTuneMethod 

FineTuneMethod(*, cfg: Any, role_models: dict[str, ModelBase])

Bases: TrainingMethod

Supervised finetuning: only student participates.

Source code in fastvideo/train/methods/fine_tuning/finetune.py 
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
def __init__(
    self,
    *,
    cfg: Any,
    role_models: dict[str, ModelBase],
) -> None:
    super().__init__(cfg=cfg, role_models=role_models)

    if "student" not in role_models:
        raise ValueError("FineTuneMethod requires role 'student'")
    if not self.student._trainable:
        raise ValueError("FineTuneMethod requires student to be "
                         "trainable")
    self._attn_kind: Literal["dense", "vsa"] = (self._infer_attn_kind())

    # Initialize preprocessors on student.
    self.student.init_preprocessors(self.training_config)

    self._init_optimizers_and_schedulers()

Methods:

fastvideo.train.methods.fine_tuning.finetune.FineTuneMethod.backward 
backward(loss_map: dict[str, Tensor], outputs: dict[str, Any], *, grad_accum_rounds: int = 1) -> None

Backpropagate an accumulation-scaled loss through the student model.

Delegating to ModelBase.backward lets each model restore its forward context before the distributed wrapper synchronizes parameter gradients.

Source code in fastvideo/train/methods/fine_tuning/finetune.py 
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
def backward(
    self,
    loss_map: dict[str, torch.Tensor],
    outputs: dict[str, Any],
    *,
    grad_accum_rounds: int = 1,
) -> None:
    """Backpropagate an accumulation-scaled loss through the student model.

    Delegating to ``ModelBase.backward`` lets each model restore its forward
    context before the distributed wrapper synchronizes parameter gradients.
    """
    grad_accum_rounds = max(1, int(grad_accum_rounds))
    ctx = outputs.get("_fv_backward")
    if ctx is None:
        super().backward(
            loss_map,
            outputs,
            grad_accum_rounds=grad_accum_rounds,
        )
        return
    self.student.backward(
        loss_map["total_loss"],
        ctx,
        grad_accum_rounds=grad_accum_rounds,
    )
fastvideo.train.methods.fine_tuning.finetune.FineTuneMethod.single_train_step 
single_train_step(batch: dict[str, Any], iteration: int) -> tuple[dict[str, Tensor], dict[str, Any], dict[str, LogScalar]]

Prepare synchronized targets and compute supervised flow loss.

The returned forward context lets model-specific backward methods restore activation-checkpoint metadata during recomputation.

Source code in fastvideo/train/methods/fine_tuning/finetune.py 
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
def single_train_step(
    self,
    batch: dict[str, Any],
    iteration: int,
) -> tuple[
        dict[str, torch.Tensor],
        dict[str, Any],
        dict[str, LogScalar],
]:
    """Prepare synchronized targets and compute supervised flow loss.

    The returned forward context lets model-specific backward methods
    restore activation-checkpoint metadata during recomputation.
    """
    del iteration
    training_batch = self.student.prepare_batch(
        batch,
        generator=self.cuda_generator,
        latents_source="data",
    )

    if training_batch.latents is None:
        raise RuntimeError("prepare_batch() must set "
                           "TrainingBatch.latents")
    if training_batch.noisy_model_input is None:
        raise RuntimeError("prepare_batch() must set "
                           "TrainingBatch.noisy_model_input")
    if training_batch.noise is None:
        raise RuntimeError("prepare_batch() must set "
                           "TrainingBatch.noise")
    if training_batch.sigmas is None:
        raise RuntimeError("prepare_batch() must set "
                           "TrainingBatch.sigmas")
    if training_batch.timesteps is None:
        raise RuntimeError("prepare_batch() must set "
                           "TrainingBatch.timesteps")

    clean_latents = training_batch.latents
    noisy_latents = (training_batch.noisy_model_input.permute(0, 2, 1, 3, 4))
    noise = training_batch.noise.permute(0, 2, 1, 3, 4)
    sigmas = training_batch.sigmas
    timesteps = training_batch.timesteps

    pred = self.student.predict_noise(
        noisy_latents,
        timesteps,
        training_batch,
        conditional=True,
        attn_kind=self._attn_kind,
    )

    loss_map = _compute_finetune_loss_map(
        pred,
        clean_latents,
        noisy_latents,
        noise,
        sigmas,
        training_batch,
        precondition_outputs=bool(self.training_config.model.precondition_outputs),
    )

    attn_metadata = training_batch.attn_metadata_vsa if self._attn_kind == "vsa" else training_batch.attn_metadata

    outputs: dict[str, Any] = {
        "_fv_backward": (
            training_batch.timesteps,
            attn_metadata,
        )
    }
    metrics: dict[str, LogScalar] = {}
    return loss_map, outputs, metrics

Functions: