Add gradient accumulation and data parallelism

flux/dreambooth.py

@@ -7,7 +7,8 @@ import mlx.core as mx
 import mlx.nn as nn
 import mlx.optimizers as optim
 import numpy as np
-from mlx.utils import tree_reduce, tree_unflatten
+from mlx.nn.utils import average_gradients
+from mlx.utils import tree_map, tree_reduce, tree_unflatten
 from PIL import Image
 from tqdm import tqdm
 
@@ -15,13 +16,14 @@ from flux import FluxPipeline
 from flux.lora import LoRALinear
 
 
-def linear_to_lora_layers(flux):
+def linear_to_lora_layers(flux, args):
     lora_layers = []
+    rank = args.lora_rank
     for name, mod in flux.flow.named_modules():
-        if ".img_attn" not in name:
+        if ".img_attn" not in name and ".txt_attn" not in name:
             continue
         if ".qkv" in name or ".proj" in name:
-            lora_layers.append((name, LoRALinear.from_base(mod, r=32)))
+            lora_layers.append((name, LoRALinear.from_base(mod, r=rank)))
     flux.flow.update_modules(tree_unflatten(lora_layers))
 
 
@@ -60,6 +62,15 @@ def generate_latents(flux, n_images, prompt, steps, seed=None, leave=True):
     return x_t
 
 
+def iterate_batches(t5_tokens, clip_tokens, x, batch_size):
+    while True:
+        indices = mx.random.randint(0, len(x), (batch_size,))
+        t5_i = t5_tokens[indices]
+        clip_i = clip_tokens[indices]
+        x_i = x[indices]
+        yield t5_i, clip_i, x_i
+
+
 def generate_progress_images(iteration, flux, args):
     out_dir = Path(args.output_dir)
     out_dir.mkdir(parents=True, exist_ok=True)
@@ -72,7 +83,7 @@ def generate_progress_images(iteration, flux, args):
         n_images,
         args.progress_prompt,
         args.progress_steps,
-        seed=42,
+        seed=42 + mx.distributed.init().rank(),
     )
 
     # Arrange them on a grid
@@ -82,6 +93,7 @@ def generate_progress_images(iteration, flux, args):
     B, H, W, C = x.shape
     x = x.reshape(n_rows, B // n_rows, H, W, C).transpose(0, 2, 1, 3, 4)
     x = x.reshape(n_rows * H, B // n_rows * W, C)
+    x = mx.pad(x, [(4, 4), (4, 4), (0, 0)])
     x = (x * 255).astype(mx.uint8)
 
     # Save them to disc
@@ -137,9 +149,18 @@ if __name__ == "__main__":
         default=50,
         help="Save the model every CHECKPOINT_EVERY steps",
     )
+    parser.add_argument(
+        "--lora-rank", type=int, default=32, help="LoRA rank for finetuning"
+    )
     parser.add_argument(
         "--learning-rate", type=float, default="1e-6", help="Learning rate for training"
     )
+    parser.add_argument(
+        "--grad-accumulate",
+        type=int,
+        default=1,
+        help="Accumulate gradients for that many iterations before applying them",
+    )
     parser.add_argument(
         "--output-dir", default="mlx_output", help="Folder to save the checkpoints in"
     )
@@ -154,7 +175,7 @@ if __name__ == "__main__":
     flux = FluxPipeline("flux-" + args.model)
     flux.ensure_models_are_loaded()
     flux.flow.freeze()
-    linear_to_lora_layers(flux)
+    linear_to_lora_layers(flux, args)
 
     trainable_params = tree_reduce(
         lambda acc, x: acc + x.size, flux.flow.trainable_parameters(), 0
@@ -165,14 +186,61 @@ if __name__ == "__main__":
     state = [flux.flow.state, optimizer.state, mx.random.state]
 
     @partial(mx.compile, inputs=state, outputs=state)
-    def step(t5_tokens, clip_tokens, x, guidance):
+    def single_step(t5_tokens, clip_tokens, x, guidance):
         loss, grads = nn.value_and_grad(flux.flow, flux.training_loss)(
             t5_tokens, clip_tokens, x, guidance
         )
+        grads = average_gradients(grads)
         optimizer.update(flux.flow, grads)
 
         return loss
 
+    @partial(mx.compile, inputs=state, outputs=state)
+    def compute_loss_and_grads(t5_tokens, clip_tokens, x, guidance):
+        return nn.value_and_grad(flux.flow, flux.training_loss)(
+            t5_tokens, clip_tokens, x, guidance
+        )
+
+    @partial(mx.compile, inputs=state, outputs=state)
+    def compute_loss_and_accumulate_grads(
+        t5_tokens, clip_tokens, x, guidance, prev_grads
+    ):
+        loss, grads = nn.value_and_grad(flux.flow, flux.training_loss)(
+            t5_tokens, clip_tokens, x, guidance
+        )
+        grads = tree_map(lambda a, b: a + b, prev_grads, grads)
+        return loss, grads
+
+    @partial(mx.compile, inputs=state, outputs=state)
+    def grad_accumulate_and_step(t5_tokens, clip_tokens, x, guidance, prev_grads):
+        loss, grads = nn.value_and_grad(flux.flow, flux.training_loss)(
+            t5_tokens, clip_tokens, x, guidance
+        )
+        grads = tree_map(lambda a, b: a + b, prev_grads, grads)
+        grads = average_gradients(grads)
+        optimizer.update(flux.flow, grads)
+
+        return loss
+
+    def step(t5_tokens, clip_tokens, x, guidance, prev_grads, perform_step):
+        if prev_grads is None:
+            if perform_step:
+                return single_step(t5_tokens, clip_tokens, x, guidance), None
+            else:
+                return compute_loss_and_grads(t5_tokens, clip_tokens, x, guidance)
+        else:
+            if perform_step:
+                return (
+                    grad_accumulate_and_step(
+                        t5_tokens, clip_tokens, x, guidance, prev_grads
+                    ),
+                    None,
+                )
+            else:
+                return compute_loss_and_accumulate_grads(
+                    t5_tokens, clip_tokens, x, guidance, prev_grads
+                )
+
     print("Encoding training images to latent space")
     x = extract_latent_vectors(flux, args.image_folder)
     t5_tokens, clip_tokens = flux.tokenize([args.prompt] * len(x))
@@ -181,14 +249,13 @@ if __name__ == "__main__":
     # An initial generation to compare
     generate_progress_images(0, flux, args)
 
+    grads = None
     losses = []
     tic = time.time()
-    for i in range(args.iterations):
-        indices = (mx.random.uniform(shape=(args.batch_size,)) * len(x)).astype(
-            mx.uint32
-        )
-        loss = step(t5_tokens[indices], clip_tokens[indices], x[indices], guidance)
-        mx.eval(loss, state)
+    batches = iterate_batches(t5_tokens, clip_tokens, x, args.batch_size)
+    for i, batch in zip(range(args.iterations), batches):
+        loss, grads = step(*batch, guidance, grads, (i + 1) % args.grad_accumulate == 0)
+        mx.eval(loss, grads, state)
         losses.append(loss.item())
 
         if (i + 1) % 10 == 0: