Will Lin
51898f48e9
Frees v2/models/ to become the vendored-architecture namespace that mirrors fastvideo/models (part of making v2 self-contained / able to replace fastvideo). The v2 recipe layer (per-family card.py/loop.py/program.py + common.py + the build_*_engine re-exports) is the recipe, not the architectures, so it moves to v2/recipes/. The pure-numpy toy/parity implementations (ToyDiT etc.) move from v2/models/backend.py to v2/platform/backends/toy.py (alongside cpu.py/accel.py/torch_*). Mechanical: all imports are absolute, so v2.models.<x> -> v2.recipes.<x> and v2.models.backend -> v2.platform.backends.toy across v2/ + examples/ (89 files, 178 refs). No behavior change. CPU mini green (202 passed, 2 skipped); all v2 files compile.
102 lines
5.2 KiB
Python
102 lines
5.2 KiB
Python
"""DMD2 distribution-matching distillation (design_v3 §10; repo: train/methods/distribution_matching).
|
||
|
||
Roles: student (trainable) · teacher (frozen real score) · critic (trainable fake score).
|
||
The student rollout is driven through the SHARED denoise loop (not a vendored sampler) — the §10
|
||
collocation moat. Loss math (repo dmd2.py):
|
||
|
||
real_cfg_x0 = teacher_uncond_x0 + s·(teacher_cond_x0 − teacher_uncond_x0)
|
||
faker_x0 = critic_x0
|
||
grad = (faker_x0 − real_cfg_x0) / mean(|gen_x0 − real_cfg_x0|)
|
||
gen_loss = 0.5·MSE(gen_x0, (gen_x0 − grad).detach()) # generator (student) update
|
||
critic_loss = MSE(critic(noised_gen), ε − gen_x0) # fake-score flow-matching
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import numpy as np
|
||
|
||
from v2._enums import ConsistencyLevel
|
||
from v2.recipes.common import cached_text_encode
|
||
from v2.request import DiffusionParams, TaskType, make_request
|
||
from v2.training.methods.base import TrainingMethod, new_instance, predict_x0
|
||
|
||
|
||
class DMD2Method(TrainingMethod):
|
||
name = "dmd2"
|
||
consistency = ConsistencyLevel.C1 # kernel-pinned: rollout under trainer kernels
|
||
|
||
def __init__(self, student_instance, teacher_instance, critic_instance, *,
|
||
lr: float = 0.05, critic_lr: float = 0.05, guidance_scale: float = 4.5,
|
||
generator_update_interval: int = 5, rollout_steps: int = 3, **kw):
|
||
super().__init__(student_instance, lr=lr, **kw)
|
||
self.teacher = teacher_instance
|
||
self.critic = critic_instance
|
||
self.critic_lr = critic_lr
|
||
self.guidance_scale = guidance_scale
|
||
self.generator_update_interval = generator_update_interval
|
||
self.rollout_steps = rollout_steps
|
||
|
||
def get_grad_clip_targets(self, iteration: int = 0) -> dict:
|
||
t = {"critic": self.critic.component("transformer")}
|
||
if iteration % self.generator_update_interval == 0:
|
||
t["student"] = self.student_dit
|
||
return t
|
||
|
||
def _rollout_sample(self, prompt: str, seed: int) -> np.ndarray:
|
||
req = make_request(TaskType.T2V, self.student.card.model_id, prompt,
|
||
diffusion=DiffusionParams(num_steps=self.rollout_steps, seed=seed))
|
||
return np.asarray(self._rollout(req).outputs["latents"], dtype="float32")
|
||
|
||
def train_step(self, batch: dict, iteration: int) -> tuple[dict, dict]:
|
||
self.iteration = iteration
|
||
teacher = self.teacher.component("transformer")
|
||
critic = self.critic.component("transformer")
|
||
gen_losses, critic_losses, gnorm_s, gnorm_c = [], [], [], []
|
||
for i, p in enumerate(batch["prompts"]):
|
||
seed = batch.get("seeds", list(range(len(batch["prompts"]))))[i]
|
||
emb = cached_text_encode(self.student, p)
|
||
neg = cached_text_encode(self.student, "")
|
||
x0g = self._rollout_sample(p, seed) # student rollout via SHARED loop
|
||
|
||
rng = np.random.default_rng(seed ^ 0xD2)
|
||
sigma = float(rng.uniform(0.1, 0.9))
|
||
noise = rng.standard_normal(x0g.shape).astype("float32")
|
||
noised = ((1.0 - sigma) * x0g + sigma * noise).astype("float32")
|
||
|
||
# real score (teacher, CFG) + fake score (critic), both in x0-space
|
||
real_cond = predict_x0(teacher(noised, emb, sigma), noised, sigma)
|
||
real_uncond = predict_x0(teacher(noised, neg, sigma), noised, sigma)
|
||
real_cfg = real_uncond + self.guidance_scale * (real_cond - real_uncond)
|
||
faker = predict_x0(critic(noised, emb, sigma), noised, sigma)
|
||
denom = float(np.mean(np.abs(x0g - real_cfg))) + 1e-6
|
||
grad = np.nan_to_num((faker - real_cfg) / denom).astype("float32")
|
||
gen_losses.append(0.5 * float(np.mean(grad ** 2))) # DMD generator loss value
|
||
|
||
# generator (student) update toward (gen_x0 − grad), converted to velocity space
|
||
if iteration % self.generator_update_interval == 0:
|
||
gen_target_x0 = (x0g - grad).astype("float32")
|
||
v_target = ((noised - gen_target_x0) / max(sigma, 1e-3)).astype("float32")
|
||
_, gn = self.student_dit.mse_grad_step(noised, emb, sigma, v_target, self.lr)
|
||
gnorm_s.append(gn)
|
||
|
||
# critic (fake-score) flow-matching update on the generated sample
|
||
sigma2 = float(rng.uniform(0.1, 0.9))
|
||
noise2 = rng.standard_normal(x0g.shape).astype("float32")
|
||
noised2 = ((1.0 - sigma2) * x0g + sigma2 * noise2).astype("float32")
|
||
c_loss, gc = critic.mse_grad_step(noised2, emb, sigma2, (noise2 - x0g).astype("float32"),
|
||
self.critic_lr)
|
||
critic_losses.append(c_loss)
|
||
gnorm_c.append(gc)
|
||
|
||
metrics = {
|
||
"dmd_loss": float(np.mean(gen_losses)),
|
||
"critic_loss": float(np.mean(critic_losses)),
|
||
"grad_norm/critic": float(np.mean(gnorm_c)),
|
||
}
|
||
if gnorm_s:
|
||
metrics["grad_norm/student"] = float(np.mean(gnorm_s))
|
||
return ({"dmd_loss": metrics["dmd_loss"], "critic_loss": metrics["critic_loss"]}, metrics)
|
||
|
||
|
||
def build_dmd2(card, **kw) -> DMD2Method:
|
||
return DMD2Method(new_instance(card), new_instance(card), new_instance(card), **kw)
|