Update ptxla training (huggingface#9864)

* update ptxla example

---------

Co-authored-by: Juan Acevedo <[email protected]>
Co-authored-by: Pei Zhang <[email protected]>
Co-authored-by: Pei Zhang <[email protected]>
Co-authored-by: Sayak Paul <[email protected]>
Co-authored-by: Pei Zhang <[email protected]>
Co-authored-by: hlky <[email protected]>

Author: 6 people

examples/research_projects/pytorch_xla/README.md

@@ -7,13 +7,14 @@ It has been tested on v4 and v5p TPU versions. Training code has been tested on
 This script implements Distributed Data Parallel using GSPMD feature in XLA compiler
 where we shard the input batches over the TPU devices. 
 
-As of 9-11-2024, these are some expected step times.
+As of 10-31-2024, these are some expected step times.
 
 | accelerator | global batch size | step time (seconds) |
 | ----------- | ----------------- | --------- |
-| v5p-128 | 1024 | 0.245 |
-| v5p-256 | 2048 | 0.234 |
-| v5p-512 | 4096 | 0.2498 |
+| v5p-512 | 16384 | 1.01 |
+| v5p-256 | 8192 | 1.01 |
+| v5p-128 | 4096 | 1.0 |
+| v5p-64 | 2048 | 1.01 |
 
 ## Create TPU
 
@@ -43,8 +44,9 @@ Install PyTorch and PyTorch/XLA nightly versions:
 gcloud compute tpus tpu-vm ssh ${TPU_NAME} \
 --project=${PROJECT_ID} --zone=${ZONE} --worker=all \
 --command='
-pip3 install --pre torch==2.5.0.dev20240905+cpu torchvision==0.20.0.dev20240905+cpu --index-url https://download.pytorch.org/whl/nightly/cpu
-pip3 install "torch_xla[tpu] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.5.0.dev20240905-cp310-cp310-linux_x86_64.whl" -f https://storage.googleapis.com/libtpu-releases/index.html
+pip3 install --pre torch==2.6.0.dev20241031+cpu torchvision --index-url https://download.pytorch.org/whl/nightly/cpu
+pip3 install "torch_xla[tpu] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.6.0.dev20241031.cxx11-cp310-cp310-linux_x86_64.whl" -f https://storage.googleapis.com/libtpu-releases/index.html
+pip install torch_xla[pallas] -f https://storage.googleapis.com/jax-releases/jax_nightly_releases.html -f https://storage.googleapis.com/jax-releases/jaxlib_nightly_releases.html
 '
 ```
 
@@ -88,17 +90,18 @@ are fixed.
 gcloud compute tpus tpu-vm ssh ${TPU_NAME} \
 --project=${PROJECT_ID} --zone=${ZONE} --worker=all \
 --command='
-export XLA_DISABLE_FUNCTIONALIZATION=1 
+export XLA_DISABLE_FUNCTIONALIZATION=0
 export PROFILE_DIR=/tmp/
 export CACHE_DIR=/tmp/
 export DATASET_NAME=lambdalabs/naruto-blip-captions
 export PER_HOST_BATCH_SIZE=32 # This is known to work on TPU v4. Can set this to 64 for TPU v5p
 export TRAIN_STEPS=50
 export OUTPUT_DIR=/tmp/trained-model/
-python diffusers/examples/research_projects/pytorch_xla/train_text_to_image_xla.py --pretrained_model_name_or_path=stabilityai/stable-diffusion-2-base --dataset_name=$DATASET_NAME --resolution=512 --center_crop --random_flip --train_batch_size=$PER_HOST_BATCH_SIZE  --max_train_steps=$TRAIN_STEPS --learning_rate=1e-06 --mixed_precision=bf16 --profile_duration=80000 --output_dir=$OUTPUT_DIR --dataloader_num_workers=4 --loader_prefetch_size=4 --device_prefetch_size=4'
-   
+python diffusers/examples/research_projects/pytorch_xla/train_text_to_image_xla.py --pretrained_model_name_or_path=stabilityai/stable-diffusion-2-base --dataset_name=$DATASET_NAME --resolution=512 --center_crop --random_flip --train_batch_size=$PER_HOST_BATCH_SIZE  --max_train_steps=$TRAIN_STEPS --learning_rate=1e-06 --mixed_precision=bf16 --profile_duration=80000 --output_dir=$OUTPUT_DIR --dataloader_num_workers=8 --loader_prefetch_size=4 --device_prefetch_size=4'
 ```
 
+Pass `--print_loss` if you would like to see the loss printed at every step. Be aware that printing the loss at every step disrupts the optimized flow execution, thus the step time will be longer. 
+
 ### Environment Envs Explained
 
 *   `XLA_DISABLE_FUNCTIONALIZATION`: To optimize the performance for AdamW optimizer.

examples/research_projects/pytorch_xla/train_text_to_image_xla.py

@@ -140,33 +140,43 @@ def run_optimizer(self):
         self.optimizer.step()
 
     def start_training(self):
-        times = []
-        last_time = time.time()
-        step = 0
-        while True:
-            if self.global_step >= self.args.max_train_steps:
-                xm.mark_step()
-                break
-            if step == 4 and PROFILE_DIR is not None:
-                xm.wait_device_ops()
-                xp.trace_detached(f"localhost:{PORT}", PROFILE_DIR, duration_ms=args.profile_duration)
+        dataloader_exception = False
+        measure_start_step = args.measure_start_step
+        assert measure_start_step < self.args.max_train_steps
+        total_time = 0
+        for step in range(0, self.args.max_train_steps):
             try:
                 batch = next(self.dataloader)
             except Exception as e:
+                dataloader_exception = True
                 print(e)
                 break
+            if step == measure_start_step and PROFILE_DIR is not None:
+                xm.wait_device_ops()
+                xp.trace_detached(f"localhost:{PORT}", PROFILE_DIR, duration_ms=args.profile_duration)
+                last_time = time.time()
             loss = self.step_fn(batch["pixel_values"], batch["input_ids"])
-            step_time = time.time() - last_time
-            if step >= 10:
-                times.append(step_time)
-            print(f"step: {step}, step_time: {step_time}")
-            if step % 5 == 0:
-                print(f"step: {step}, loss: {loss}")
-            last_time = time.time()
             self.global_step += 1
-            step += 1
-        # print(f"Average step time: {sum(times)/len(times)}")
-        xm.wait_device_ops()
+
+            def print_loss_closure(step, loss):
+                print(f"Step: {step}, Loss: {loss}")
+
+            if args.print_loss:
+                xm.add_step_closure(
+                    print_loss_closure,
+                    args=(
+                        self.global_step,
+                        loss,
+                    ),
+                )
+        xm.mark_step()
+        if not dataloader_exception:
+            xm.wait_device_ops()
+            total_time = time.time() - last_time
+            print(f"Average step time: {total_time/(self.args.max_train_steps-measure_start_step)}")
+        else:
+            print("dataloader exception happen, skip result")
+            return
 
     def step_fn(
         self,
@@ -180,7 +190,10 @@ def step_fn(
             noise = torch.randn_like(latents).to(self.device, dtype=self.weight_dtype)
             bsz = latents.shape[0]
             timesteps = torch.randint(
-                0, self.noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device
+                0,
+                self.noise_scheduler.config.num_train_timesteps,
+                (bsz,),
+                device=latents.device,
             )
             timesteps = timesteps.long()
 
@@ -224,9 +237,6 @@ def step_fn(
 
 def parse_args():
     parser = argparse.ArgumentParser(description="Simple example of a training script.")
-    parser.add_argument(
-        "--input_perturbation", type=float, default=0, help="The scale of input perturbation. Recommended 0.1."
-    )
     parser.add_argument("--profile_duration", type=int, default=10000, help="Profile duration in ms")
     parser.add_argument(
         "--pretrained_model_name_or_path",
@@ -258,12 +268,6 @@ def parse_args():
             " or to a folder containing files that 🤗 Datasets can understand."
         ),
     )
-    parser.add_argument(
-        "--dataset_config_name",
-        type=str,
-        default=None,
-        help="The config of the Dataset, leave as None if there's only one config.",
-    )
     parser.add_argument(
         "--train_data_dir",
         type=str,
@@ -283,15 +287,6 @@ def parse_args():
         default="text",
         help="The column of the dataset containing a caption or a list of captions.",
     )
-    parser.add_argument(
-        "--max_train_samples",
-        type=int,
-        default=None,
-        help=(
-            "For debugging purposes or quicker training, truncate the number of training examples to this "
-            "value if set."
-        ),
-    )
     parser.add_argument(
         "--output_dir",
         type=str,
@@ -304,7 +299,6 @@ def parse_args():
         default=None,
         help="The directory where the downloaded models and datasets will be stored.",
     )
-    parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
     parser.add_argument(
         "--resolution",
         type=int,
@@ -374,12 +368,19 @@ def parse_args():
         default=1,
         help=("Number of subprocesses to use for data loading to cpu."),
     )
+    parser.add_argument(
+        "--loader_prefetch_factor",
+        type=int,
+        default=2,
+        help=("Number of batches loaded in advance by each worker."),
+    )
     parser.add_argument(
         "--device_prefetch_size",
         type=int,
         default=1,
         help=("Number of subprocesses to use for data loading to tpu from cpu. "),
     )
+    parser.add_argument("--measure_start_step", type=int, default=10, help="Step to start profiling.")
     parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
     parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
     parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
@@ -394,12 +395,8 @@ def parse_args():
         "--mixed_precision",
         type=str,
         default=None,
-        choices=["no", "fp16", "bf16"],
-        help=(
-            "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
-            " 1.10.and an Nvidia Ampere GPU.  Default to the value of accelerate config of the current system or the"
-            " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
-        ),
+        choices=["no", "bf16"],
+        help=("Whether to use mixed precision. Bf16 requires PyTorch >= 1.10"),
     )
     parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
     parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
@@ -409,6 +406,12 @@ def parse_args():
         default=None,
         help="The name of the repository to keep in sync with the local `output_dir`.",
     )
+    parser.add_argument(
+        "--print_loss",
+        default=False,
+        action="store_true",
+        help=("Print loss at every step."),
+    )
 
     args = parser.parse_args()
 
@@ -436,7 +439,6 @@ def load_dataset(args):
         # Downloading and loading a dataset from the hub.
         dataset = datasets.load_dataset(
             args.dataset_name,
-            args.dataset_config_name,
             cache_dir=args.cache_dir,
             data_dir=args.train_data_dir,
         )
@@ -481,9 +483,7 @@ def main(args):
     _ = xp.start_server(PORT)
 
     num_devices = xr.global_runtime_device_count()
-    device_ids = np.arange(num_devices)
-    mesh_shape = (num_devices, 1)
-    mesh = xs.Mesh(device_ids, mesh_shape, ("x", "y"))
+    mesh = xs.get_1d_mesh("data")
     xs.set_global_mesh(mesh)
 
     text_encoder = CLIPTextModel.from_pretrained(
@@ -520,6 +520,7 @@ def main(args):
     from torch_xla.distributed.fsdp.utils import apply_xla_patch_to_nn_linear
 
     unet = apply_xla_patch_to_nn_linear(unet, xs.xla_patched_nn_linear_forward)
+    unet.enable_xla_flash_attention(partition_spec=("data", None, None, None))
 
     vae.requires_grad_(False)
     text_encoder.requires_grad_(False)
@@ -530,15 +531,12 @@ def main(args):
     # as these weights are only used for inference, keeping weights in full
     # precision is not required.
     weight_dtype = torch.float32
-    if args.mixed_precision == "fp16":
-        weight_dtype = torch.float16
-    elif args.mixed_precision == "bf16":
+    if args.mixed_precision == "bf16":
         weight_dtype = torch.bfloat16
 
     device = xm.xla_device()
-    print("device: ", device)
-    print("weight_dtype: ", weight_dtype)
 
+    # Move text_encode and vae to device and cast to weight_dtype
     text_encoder = text_encoder.to(device, dtype=weight_dtype)
     vae = vae.to(device, dtype=weight_dtype)
     unet = unet.to(device, dtype=weight_dtype)
@@ -606,24 +604,27 @@ def collate_fn(examples):
         collate_fn=collate_fn,
         num_workers=args.dataloader_num_workers,
         batch_size=args.train_batch_size,
+        prefetch_factor=args.loader_prefetch_factor,
     )
 
     train_dataloader = pl.MpDeviceLoader(
         train_dataloader,
         device,
         input_sharding={
-            "pixel_values": xs.ShardingSpec(mesh, ("x", None, None, None), minibatch=True),
-            "input_ids": xs.ShardingSpec(mesh, ("x", None), minibatch=True),
+            "pixel_values": xs.ShardingSpec(mesh, ("data", None, None, None), minibatch=True),
+            "input_ids": xs.ShardingSpec(mesh, ("data", None), minibatch=True),
         },
         loader_prefetch_size=args.loader_prefetch_size,
         device_prefetch_size=args.device_prefetch_size,
     )
 
+    num_hosts = xr.process_count()
+    num_devices_per_host = num_devices // num_hosts
     if xm.is_master_ordinal():
         print("***** Running training *****")
-        print(f"Instantaneous batch size per device = {args.train_batch_size}")
+        print(f"Instantaneous batch size per device = {args.train_batch_size // num_devices_per_host }")
         print(
-            f"Total train batch size (w. parallel, distributed & accumulation) = {args.train_batch_size * num_devices}"
+            f"Total train batch size (w. parallel, distributed & accumulation) = {args.train_batch_size * num_hosts}"
         )
         print(f"  Total optimization steps = {args.max_train_steps}")