feat: add LoRA dropout, LoRA+ asymmetric LR, and curriculum timestep sampling

- LoRA dropout: applied to the LoRA path only (not frozen base weights),
  0.05–0.1 helps regularize on small datasets (arXiv:2404.09610)
- LoRA+: separate optimizer param groups for lora_A and lora_B with
  configurable LR ratio; ratio=16 enables LoRA+ (arXiv:2402.12354)
- Curriculum mode: logit_normal for first N% of steps then uniform,
  directly addresses early convergence + fine-detail degradation at
  boundaries (arXiv:2603.12517)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
2026-04-06 12:43:18 +02:00
parent 5baa070e61
commit eb63c1ead7
3 changed files with 104 additions and 26 deletions
+54 -13
View File
@@ -271,17 +271,33 @@ class SelvaLoraTrainer:
"tooltip": "Path to a step checkpoint (.pt) to resume training from.",
}),
"seed": ("INT", {"default": 42}),
"timestep_mode": (["uniform", "logit_normal"], {
"timestep_mode": (["uniform", "logit_normal", "curriculum"], {
"default": "uniform",
"tooltip": "How to sample training timesteps. "
"uniform samples all timesteps equally (default, matches original MMAudio training). "
"logit_normal concentrates steps near t=0.5 — reaches lower loss but perceptual improvement is dataset-dependent.",
"uniform: all timesteps equally (matches original MMAudio). "
"logit_normal: concentrates near t=0.5. "
"curriculum: logit_normal for first curriculum_switch% of steps then uniform (recommended for small datasets).",
}),
"logit_normal_sigma": ("FLOAT", {
"default": 1.0, "min": 0.1, "max": 3.0, "step": 0.1,
"tooltip": "Spread of the logit-normal distribution. "
"1.0 = moderate peak at t=0.5. Higher approaches uniform. "
"Only used when timestep_mode=logit_normal.",
"Used with logit_normal and curriculum modes.",
}),
"curriculum_switch": ("FLOAT", {
"default": 0.6, "min": 0.1, "max": 0.9, "step": 0.05,
"tooltip": "Fraction of steps to run logit_normal before switching to uniform. "
"0.6 = switch at 60% of total steps. Only used with timestep_mode=curriculum.",
}),
"lora_dropout": ("FLOAT", {
"default": 0.0, "min": 0.0, "max": 0.3, "step": 0.01,
"tooltip": "Dropout applied to the LoRA path only (not the frozen base weights). "
"0=disabled. 0.050.1 helps regularize on small datasets (arXiv:2404.09610).",
}),
"lora_plus_ratio": ("FLOAT", {
"default": 1.0, "min": 1.0, "max": 32.0, "step": 1.0,
"tooltip": "LoRA+ LR ratio: lr_B = lr × ratio. "
"1.0 = standard LoRA. 16.0 = LoRA+ (arXiv:2402.12354).",
}),
},
}
@@ -305,7 +321,8 @@ class SelvaLoraTrainer:
def train(self, model, data_dir, output_dir, steps, rank, lr,
alpha=0.0, target="attn.qkv", batch_size=4, warmup_steps=100,
grad_accum=1, save_every=500, resume_path="", seed=42,
timestep_mode="uniform", logit_normal_sigma=1.0):
timestep_mode="uniform", logit_normal_sigma=1.0, curriculum_switch=0.6,
lora_dropout=0.0, lora_plus_ratio=1.0):
torch.manual_seed(seed)
random.seed(seed)
@@ -442,7 +459,8 @@ class SelvaLoraTrainer:
data_dir, output_dir, steps, rank, lr,
alpha_val, target_suffixes, batch_size, warmup_steps,
grad_accum, save_every, resume_path, seed,
timestep_mode, logit_normal_sigma,
timestep_mode, logit_normal_sigma, curriculum_switch,
lora_dropout, lora_plus_ratio,
)
def _train_inner(
@@ -451,19 +469,21 @@ class SelvaLoraTrainer:
data_dir, output_dir, steps, rank, lr,
alpha_val, target_suffixes, batch_size, warmup_steps,
grad_accum, save_every, resume_path, seed,
timestep_mode="uniform", logit_normal_sigma=1.0,
timestep_mode="uniform", logit_normal_sigma=1.0, curriculum_switch=0.6,
lora_dropout=0.0, lora_plus_ratio=1.0,
):
# --- Prepare generator copy with LoRA ---
generator = copy.deepcopy(model["generator"]).to(device, dtype)
n_lora = apply_lora(generator, rank=rank, alpha=alpha_val,
target_suffixes=target_suffixes)
target_suffixes=target_suffixes, dropout=lora_dropout)
if n_lora == 0:
raise RuntimeError(
f"[LoRA Trainer] No layers matched target={target_suffixes}. "
"Check the 'target' field."
)
print(f"[LoRA Trainer] Wrapped {n_lora} layers (rank={rank}, alpha={alpha_val})", flush=True)
print(f"[LoRA Trainer] Wrapped {n_lora} layers "
f"(rank={rank}, alpha={alpha_val}, dropout={lora_dropout})", flush=True)
for name, p in generator.named_parameters():
p.requires_grad_("lora_" in name)
@@ -475,8 +495,16 @@ class SelvaLoraTrainer:
)
# --- Optimizer + scheduler ---
lora_params = [p for p in generator.parameters() if p.requires_grad]
optimizer = torch.optim.AdamW(lora_params, lr=lr, weight_decay=1e-2)
# LoRA+: split A and B into separate param groups with different LRs.
# ratio=1.0 = standard LoRA (same LR for both). ratio=16 = LoRA+.
lora_A_params = [p for n, p in generator.named_parameters() if "lora_A" in n and p.requires_grad]
lora_B_params = [p for n, p in generator.named_parameters() if "lora_B" in n and p.requires_grad]
optimizer = torch.optim.AdamW([
{"params": lora_A_params, "lr": lr},
{"params": lora_B_params, "lr": lr * lora_plus_ratio},
], weight_decay=1e-2)
if lora_plus_ratio != 1.0:
print(f"[LoRA Trainer] LoRA+: lr_A={lr:.2e} lr_B={lr * lora_plus_ratio:.2e}", flush=True)
def lr_lambda(s):
return s / max(1, warmup_steps) if s < warmup_steps else 1.0
@@ -518,8 +546,15 @@ class SelvaLoraTrainer:
"steps": steps,
"timestep_mode": timestep_mode,
"logit_normal_sigma": logit_normal_sigma,
"curriculum_switch": curriculum_switch,
"lora_dropout": lora_dropout,
"lora_plus_ratio": lora_plus_ratio,
}
# For curriculum mode: compute the step at which we switch from logit_normal to uniform
curriculum_switch_step = start_step + int((steps - start_step) * curriculum_switch)
_curriculum_switched = False
print(f"\n[LoRA Trainer] Training {remaining} steps "
f"(step {start_step + 1}{steps}, batch_size={batch_size}, "
f"timestep_mode={timestep_mode})\n", flush=True)
@@ -538,11 +573,17 @@ class SelvaLoraTrainer:
generator.normalize(x1)
if timestep_mode == "logit_normal":
if timestep_mode == "logit_normal" or (
timestep_mode == "curriculum" and step <= curriculum_switch_step
):
u = torch.randn(batch_size, device=device, dtype=dtype) * logit_normal_sigma
t = torch.sigmoid(u)
else:
t = torch.rand(batch_size, device=device, dtype=dtype)
if timestep_mode == "curriculum" and step == curriculum_switch_step + 1 and not _curriculum_switched:
print(f"[LoRA Trainer] Curriculum switch: logit_normal → uniform at step {step}", flush=True)
_curriculum_switched = True
x0 = torch.randn_like(x1)
xt = fm.get_conditional_flow(x0, x1, t)
@@ -552,7 +593,7 @@ class SelvaLoraTrainer:
running_loss += loss.item() * grad_accum
if step % grad_accum == 0:
torch.nn.utils.clip_grad_norm_(lora_params, max_norm=1.0)
torch.nn.utils.clip_grad_norm_(lora_A_params + lora_B_params, max_norm=1.0)
optimizer.step()
scheduler.step()
optimizer.zero_grad()
+11 -5
View File
@@ -25,13 +25,14 @@ import torch.nn as nn
class LoRALinear(nn.Module):
"""nn.Linear with a frozen base weight and trainable low-rank A/B matrices.
Output: base(x) + (x @ A.T @ B.T) * (alpha / rank)
Output: base(x) + (dropout(x) @ A.T @ B.T) * (alpha / rank)
A is initialised with Kaiming uniform; B is initialised to zero so the
adapter contribution starts at zero and does not disturb pretrained behaviour.
Dropout is applied only to the LoRA path, not the base linear.
"""
def __init__(self, linear: nn.Linear, rank: int, alpha: float):
def __init__(self, linear: nn.Linear, rank: int, alpha: float, dropout: float = 0.0):
super().__init__()
in_f = linear.in_features
out_f = linear.out_features
@@ -46,16 +47,18 @@ class LoRALinear(nn.Module):
self.lora_A = nn.Parameter(torch.empty(rank, in_f, dtype=ref_dtype, device=ref_device))
self.lora_B = nn.Parameter(torch.zeros(out_f, rank, dtype=ref_dtype, device=ref_device))
self.scale = alpha / rank
self.dropout = nn.Dropout(p=dropout) if dropout > 0.0 else nn.Identity()
nn.init.kaiming_uniform_(self.lora_A, a=math.sqrt(5))
def forward(self, x: torch.Tensor) -> torch.Tensor:
return self.linear(x) + (x @ self.lora_A.T @ self.lora_B.T) * self.scale
return self.linear(x) + (self.dropout(x) @ self.lora_A.T @ self.lora_B.T) * self.scale
def extra_repr(self) -> str:
rank = self.lora_A.shape[0]
p = self.dropout.p if isinstance(self.dropout, nn.Dropout) else 0.0
return (f"in={self.linear.in_features}, out={self.linear.out_features}, "
f"rank={rank}, scale={self.scale:.4f}")
f"rank={rank}, scale={self.scale:.4f}, dropout={p}")
def apply_lora(
@@ -63,6 +66,7 @@ def apply_lora(
rank: int = 16,
alpha: float = None,
target_suffixes: tuple = ("attn.qkv",),
dropout: float = 0.0,
) -> int:
"""Replace matching nn.Linear layers with LoRALinear in-place.
@@ -74,6 +78,8 @@ def apply_lora(
("attn.qkv",) which targets all SelfAttention QKV
projections in the MM-DiT generator.
Add "linear1" to also wrap post-attention output projections.
dropout: Dropout probability on the LoRA path (not the base linear).
0.050.1 helps regularize on small datasets.
Returns:
Number of linear layers wrapped.
@@ -92,7 +98,7 @@ def apply_lora(
parent = model
for part in parts[:-1]:
parent = getattr(parent, part)
setattr(parent, parts[-1], LoRALinear(module, rank, alpha))
setattr(parent, parts[-1], LoRALinear(module, rank, alpha, dropout=dropout))
count += 1
return count
+39 -8
View File
@@ -167,10 +167,16 @@ def main():
help="Path to a step checkpoint (.pt) to resume training from.")
parser.add_argument("--precision", default="bf16", choices=["bf16", "fp16", "fp32"])
parser.add_argument("--seed", type=int, default=42)
parser.add_argument("--timestep_mode", default="uniform", choices=["uniform", "logit_normal"],
help="Timestep sampling distribution. uniform matches original MMAudio training. logit_normal reaches lower loss but perceptual improvement is dataset-dependent.")
parser.add_argument("--timestep_mode", default="uniform", choices=["uniform", "logit_normal", "curriculum"],
help="Timestep sampling. uniform=original MMAudio, logit_normal=concentrated near t=0.5, curriculum=logit_normal then uniform.")
parser.add_argument("--logit_normal_sigma", type=float, default=1.0,
help="Spread of logit-normal distribution (only used with --timestep_mode logit_normal).")
help="Spread of logit-normal distribution.")
parser.add_argument("--curriculum_switch", type=float, default=0.6,
help="Fraction of steps to use logit_normal before switching to uniform (curriculum mode only).")
parser.add_argument("--lora_dropout", type=float, default=0.0,
help="Dropout on the LoRA path only. 0.050.1 helps on small datasets.")
parser.add_argument("--lora_plus_ratio", type=float, default=1.0,
help="LoRA+ LR ratio: lr_B = lr * ratio. 1.0=standard, 16.0=LoRA+.")
args = parser.parse_args()
torch.manual_seed(args.seed)
@@ -234,8 +240,9 @@ def main():
rank=args.rank,
alpha=args.alpha,
target_suffixes=tuple(args.target),
dropout=args.lora_dropout,
)
print(f"[LoRA] Wrapped {n_lora} linear layers (rank={args.rank}, target={args.target})")
print(f"[LoRA] Wrapped {n_lora} linear layers (rank={args.rank}, target={args.target}, dropout={args.lora_dropout})")
if n_lora == 0:
print("[LoRA] ERROR: no layers were wrapped — check --target names.")
sys.exit(1)
@@ -315,8 +322,16 @@ def main():
print(f"[LoRA] {len(dataset)} clip(s) ready.")
# --- Optimizer + LR scheduler ---
lora_params = [p for p in net_generator.parameters() if p.requires_grad]
optimizer = torch.optim.AdamW(lora_params, lr=args.lr, weight_decay=1e-2)
# LoRA+: separate param groups for A and B with different LRs.
# ratio=1.0 = standard LoRA. ratio=16 = LoRA+ (arXiv:2402.12354).
lora_A_params = [p for n, p in net_generator.named_parameters() if "lora_A" in n and p.requires_grad]
lora_B_params = [p for n, p in net_generator.named_parameters() if "lora_B" in n and p.requires_grad]
optimizer = torch.optim.AdamW([
{"params": lora_A_params, "lr": args.lr},
{"params": lora_B_params, "lr": args.lr * args.lora_plus_ratio},
], weight_decay=1e-2)
if args.lora_plus_ratio != 1.0:
print(f"[LoRA] LoRA+: lr_A={args.lr:.2e} lr_B={args.lr * args.lora_plus_ratio:.2e}")
def lr_lambda(step):
if step < args.warmup_steps:
@@ -351,6 +366,9 @@ def main():
f"batch_size={args.batch_size}, lr={args.lr}, grad_accum={args.grad_accum}")
print(f"[LoRA] Checkpoints every {args.save_every} steps → {output_dir}\n")
curriculum_switch_step = start_step + int((args.steps - start_step) * args.curriculum_switch)
_curriculum_switched = False
total_loss = 0.0
for step in range(start_step + 1, args.steps + 1):
batch = random.choices(dataset, k=args.batch_size)
@@ -363,11 +381,18 @@ def main():
net_generator.normalize(x1)
if args.timestep_mode == "logit_normal":
if args.timestep_mode == "logit_normal" or (
args.timestep_mode == "curriculum" and step <= curriculum_switch_step
):
u = torch.randn(args.batch_size, device=device, dtype=dtype) * args.logit_normal_sigma
t = torch.sigmoid(u)
else:
t = torch.rand(args.batch_size, device=device, dtype=dtype)
if args.timestep_mode == "curriculum" and step == curriculum_switch_step + 1 and not _curriculum_switched:
print(f"[LoRA] Curriculum switch: logit_normal → uniform at step {step}")
_curriculum_switched = True
x0 = torch.randn_like(x1)
xt = fm.get_conditional_flow(x0, x1, t)
@@ -378,7 +403,7 @@ def main():
total_loss += loss.item() * args.grad_accum
if step % args.grad_accum == 0:
torch.nn.utils.clip_grad_norm_(lora_params, max_norm=1.0)
torch.nn.utils.clip_grad_norm_(lora_A_params + lora_B_params, max_norm=1.0)
optimizer.step()
scheduler.step()
optimizer.zero_grad()
@@ -404,6 +429,9 @@ def main():
"steps": args.steps,
"timestep_mode": args.timestep_mode,
"logit_normal_sigma": args.logit_normal_sigma,
"curriculum_switch": args.curriculum_switch,
"lora_dropout": args.lora_dropout,
"lora_plus_ratio": args.lora_plus_ratio,
},
}, ckpt_path)
print(f"[LoRA] Saved {ckpt_path}")
@@ -424,6 +452,9 @@ def main():
"steps": args.steps,
"timestep_mode": args.timestep_mode,
"logit_normal_sigma": args.logit_normal_sigma,
"curriculum_switch": args.curriculum_switch,
"lora_dropout": args.lora_dropout,
"lora_plus_ratio": args.lora_plus_ratio,
}
torch.save({"state_dict": get_lora_state_dict(net_generator), "meta": meta}, final)
(output_dir / "meta.json").write_text(json.dumps(meta, indent=2))