Merge pull request #418 from AI-Hypercomputer:add-sd15-support

PiperOrigin-RevId: 933185351

Author: Google-ML-Automation

src/maxdiffusion/checkpointing/base_stable_diffusion_checkpointer.py

@@ -209,6 +209,7 @@ def load_diffusers_checkpoint(self):
           split_head_dim=self.config.split_head_dim,
           norm_num_groups=self.config.norm_num_groups,
           attention_kernel=self.config.attention,
+          flash_min_seq_length=getattr(self.config, "flash_min_seq_length", 4096),
           flash_block_sizes=flash_block_sizes,
           mesh=self.mesh,
           precision=precision,
@@ -220,6 +221,7 @@ def load_diffusers_checkpoint(self):
             split_head_dim=self.config.split_head_dim,
             norm_num_groups=self.config.norm_num_groups,
             attention_kernel=self.config.attention,
+            flash_min_seq_length=getattr(self.config, "flash_min_seq_length", 4096),
             flash_block_sizes=flash_block_sizes,
             dtype=self.activations_dtype,
             weights_dtype=self.weights_dtype,
@@ -279,6 +281,7 @@ def load_checkpoint(self, step=None, scheduler_class=None):
           split_head_dim=self.config.split_head_dim,
           norm_num_groups=self.config.norm_num_groups,
           attention_kernel=self.config.attention,
+          flash_min_seq_length=getattr(self.config, "flash_min_seq_length", 4096),
           flash_block_sizes=flash_block_sizes,
           mesh=self.mesh,
           precision=precision,

src/maxdiffusion/configs/README.md

@@ -2,6 +2,13 @@
 
 This directory contains model configuration for different Stable Diffusion models.
 
+## Stable Diffusion 1.5
+
+base15.yml - used for training and inference using [stable-diffusion-v1-5](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5).
+The upstream checkpoint ships PyTorch weights only, so this config sets `from_pt: True`; point
+`pretrained_model_name_or_path` at a local diffusers snapshot for offline runs. It defaults to the
+checkpoint's PNDM scheduler (epsilon prediction) to match the reference inference path.
+
 ## Stable Diffusion 2.1
 
 base21.yml - used for training and inference using [stable-diffusion-2-1](https://huggingface.co/stabilityai/stable-diffusion-2-1)

src/maxdiffusion/configs/base15.yml

@@ -0,0 +1,279 @@
+# Copyright 2025 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Stable Diffusion 1.5 base config.
+#
+# SD 1.5 shares the same architecture as SD 1.4 (CLIP ViT-L/14 text encoder,
+# 860M UNet, AutoencoderKL) and only differs by the trained weights, so this
+# config mirrors base14.yml and points at the v1-5 checkpoint. The upstream
+# checkpoint only ships PyTorch weights, so from_pt is True by default; override
+# pretrained_model_name_or_path to a local diffusers snapshot for offline runs.
+
+# This sentinel is a reminder to choose a real run name.
+run_name: ''
+
+metrics_file: "" # for testing, local file that stores scalar metrics. If empty, no metrics are written.
+# If true save metrics such as loss and TFLOPS to GCS in {base_output_directory}/{run_name}/metrics/
+write_metrics: True
+gcs_metrics: True
+
+# For testing, local file that stores function timing metrics such as state creation and compilation.
+# If empty, no metrics are written.
+timing_metrics_file: ""
+write_timing_metrics: True
+
+# If true save config to GCS in {base_output_directory}/{run_name}/
+save_config_to_gcs: False
+log_period: 10000000000  # Flushes Tensorboard
+
+pretrained_model_name_or_path: 'stable-diffusion-v1-5/stable-diffusion-v1-5'
+unet_checkpoint: ''
+# The canonical v1-5 repo only publishes the main (PyTorch) revision.
+revision: 'main'
+
+# This will convert the weights to this dtype.
+weights_dtype: 'float32'
+# This sets the layer's dtype in the model. Ex: nn.Dense(dtype=activations_dtype)
+activations_dtype: 'bfloat16'
+
+# matmul and conv precision from https://jax.readthedocs.io/en/latest/jax.lax.html#jax.lax.Precision
+# Options are "DEFAULT", "HIGH", "HIGHEST"
+# fp32 activations and fp32 weights with HIGHEST will provide the best precision
+# at the cost of time.
+precision: "DEFAULT"
+
+# if False state is not jitted and instead replicate is called. This is good for debugging on single host
+# It must be True for multi-host.
+jit_initializers: True
+
+# Set true to load weights from pytorch. The v1-5 checkpoint is PyTorch-only.
+from_pt: True
+split_head_dim: True
+attention: 'tokamax_flash' # Supported attention: dot_product, flash, tokamax_flash
+# Minimum Q/K/V sequence length required to use flash attention. For SD 1.5
+# 1024x1024 inference, the two largest self-attention lengths are 16384 and
+# 4096, while cross-attention falls back to dot_product because text KV length
+# is 77.
+flash_min_seq_length: 4096
+# If mask_padding_tokens is True, we pass in segment ids to splash attention to avoid attending to padding tokens.
+# Else we do not pass in segment ids and on vpu bound hardware like trillium this is faster.
+# However, when padding tokens are significant, this will lead to worse quality and should be set to True.
+mask_padding_tokens: True
+# Maxdiffusion has 2 types of attention sharding strategies:
+# 1. attention_sharding_uniform = True : same sequence sharding rules applied for q in both (self and cross attention)
+# 2. attention_sharding_uniform = False : Heads are sharded uniformly across devices for self attention
+#    while sequence is sharded in cross attention q.
+attention_sharding_uniform: True
+flash_block_sizes: {
+  "block_q" : 2048,
+  "block_kv_compute" : 1024,
+  "block_kv" : 2048,
+  "block_q_dkv" : 2048,
+  "block_kv_dkv" : 2048,
+  "block_kv_dkv_compute" : 1024
+}
+# GroupNorm groups
+norm_num_groups: 32
+
+# If train_new_unet, unet weights will be randomly initialized to train the unet from scratch
+# else they will be loaded from pretrained_model_name_or_path
+train_new_unet: False
+
+# train text_encoder
+train_text_encoder: False
+text_encoder_learning_rate: 4.25e-6
+
+# https://arxiv.org/pdf/2305.08891.pdf
+snr_gamma: -1.0
+
+timestep_bias: {
+  # a value of later will increase the frequence of the model's final training steps.
+  # none, earlier, later, range
+  strategy: "none",
+  # multiplier for bias, a value of 2.0 will double the weight of the bias, 0.5 will halve it.
+  multiplier: 1.0,
+  # when using strategy=range, the beginning (inclusive) timestep to bias.
+  begin: 0,
+  # when using strategy=range, the final step (inclusive) to bias.
+  end: 1000,
+  # portion of timesteps to bias.
+  # 0.5 will bias one half of the timesteps. Value of strategy determines
+  # whether the biased portions are in the earlier or later timesteps.
+  portion: 0.25
+}
+
+# SD 1.5 uses a PNDM sampler with epsilon prediction and leading timestep
+# spacing. These mirror the checkpoint's scheduler_config.json so generation
+# matches the diffusers/reference defaults.
+diffusion_scheduler_config: {
+  _class_name: 'FlaxPNDMScheduler',
+  prediction_type: 'epsilon',
+  rescale_zero_terminal_snr: False,
+  timestep_spacing: 'leading'
+}
+
+# Hardware
+hardware: 'tpu' # Supported hardware types are 'tpu', 'gpu'
+skip_jax_distributed_system: False
+
+base_output_directory: ""
+
+# Parallelism
+mesh_axes: ['data', 'fsdp', 'context', 'tensor']
+
+# batch : batch dimension of data and activations
+# hidden :
+# embed : attention qkv dense layer hidden dim named as embed
+# heads : attention head dim = num_heads * head_dim
+# length : attention sequence length
+# temb_in : dense.shape[0] of resnet dense before conv
+# out_c : dense.shape[1] of resnet dense before conv
+# out_channels : conv.shape[-1] activation
+# keep_1 : conv.shape[0] weight
+# keep_2 : conv.shape[1] weight
+# conv_in : conv.shape[2] weight
+# conv_out : conv.shape[-1] weight
+logical_axis_rules: [
+                      ['batch', 'data'],
+                      ['activation_batch', ['data','fsdp']],
+                      ['activation_heads', 'tensor'],
+                      ['activation_kv', 'tensor'],
+                      ['embed','fsdp'],
+                      ['heads', 'tensor'],
+                      ['conv_batch', ['data','fsdp']],
+                      ['out_channels', 'tensor'],
+                      ['conv_out', 'fsdp'],
+                    ]
+data_sharding: [['data', 'fsdp', 'context', 'tensor']]
+
+# One axis for each parallelism type may hold a placeholder (-1)
+# value to auto-shard based on available slices and devices.
+# By default, product of the DCN axes should equal number of slices
+# and product of the ICI axes should equal number of devices per slice.
+dcn_data_parallelism: -1  # recommended DCN axis to be auto-sharded
+dcn_fsdp_parallelism: 1
+dcn_context_parallelism: 1
+dcn_tensor_parallelism: 1
+ici_data_parallelism: -1 # recommended ICI axis to be auto-sharded for TPUv5e
+ici_fsdp_parallelism: 1  # recommended ICI axis to be auto-sharded
+ici_context_parallelism: 1
+ici_tensor_parallelism: 1
+
+allow_split_physical_axes: False
+
+# Dataset
+# Replace with dataset path or train_data_dir. One has to be set.
+dataset_name: 'diffusers/pokemon-gpt4-captions'
+train_split: 'train'
+dataset_type: 'tf'
+cache_latents_text_encoder_outputs: True
+# cache_latents_text_encoder_outputs only apply to dataset_type="tf",
+# only apply to small dataset that fits in memory
+# prepare image latents and text encoder outputs
+# Reduce memory consumption and reduce step time during training
+# transformed dataset is saved at dataset_save_location
+dataset_save_location: '/tmp/pokemon-gpt4-captions_sd15'
+train_data_dir: ''
+dataset_config_name: ''
+jax_cache_dir: ''
+hf_data_dir: ''
+hf_train_files: ''
+hf_access_token: ''
+image_column: 'image'
+caption_column: 'text'
+resolution: 512
+center_crop: False
+random_flip: False
+# If cache_latents_text_encoder_outputs is True
+# the num_proc is set to 1
+tokenize_captions_num_proc: 4
+transform_images_num_proc: 4
+reuse_example_batch: False
+enable_data_shuffling: True
+
+# checkpoint every number of samples, -1 means don't checkpoint.
+checkpoint_every: -1
+# enables one replica to read the ckpt then broadcast to the rest
+enable_single_replica_ckpt_restoring: False
+
+# Training loop
+learning_rate: 1.e-7
+scale_lr: False
+max_train_samples: -1
+# max_train_steps takes priority over num_train_epochs.
+max_train_steps: 800
+seed: 0
+# Output directory
+# Create a GCS bucket, e.g. my-maxdiffusion-outputs and set this to "gs://my-maxdiffusion-outputs/"
+output_dir: ''
+per_device_batch_size: 1
+
+warmup_steps_fraction: 0.0
+learning_rate_schedule_steps: -1 # By default the length of the schedule is set to the number of steps.
+
+# However you may choose a longer schedule (learning_rate_schedule_steps > steps), in which case the training will end before
+# dropping fully down. Or you may choose a shorter schedule, where the unspecified steps will have a learning rate of 0.
+
+# AdamW optimizer parameters
+adam_b1: 0.9 # Exponential decay rate to track the first moment of past gradients.
+adam_b2: 0.999 # Exponential decay rate to track the second moment of past gradients.
+adam_eps: 1.e-8 # A small constant applied to denominator outside of the square root.
+adam_weight_decay: 1.e-2 # AdamW Weight decay
+opt_enable_grad_clipping: False
+max_grad_value: 1.0
+opt_enable_grad_global_norm_clipping: False
+max_grad_norm: 1.0
+
+enable_profiler: False
+# Skip first n steps for profiling, to omit things like compilation and to give
+# the iteration time a chance to stabilize.
+skip_first_n_steps_for_profiler: 1
+profiler_steps: 5
+
+# Generation parameters
+prompt: "A magical castle in the middle of a forest, artistic drawing"
+negative_prompt: "purple, red"
+guidance_scale: 7.5
+# Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+guidance_rescale: 0.0
+# SD 1.5 reference inference default.
+num_inference_steps: 20
+
+enable_mllog: False
+
+# controlnet
+controlnet_model_name_or_path: 'lllyasviel/sd-controlnet-canny'
+controlnet_from_pt: True
+controlnet_conditioning_scale: 1.0
+controlnet_image: 'https://huggingface.co/datasets/YiYiXu/test-doc-assets/resolve/main/blog_post_cell_10_output_0.jpeg'
+
+# dreambooth - this script always uses prior preservation.
+instance_data_dir: ''
+class_data_dir: ''
+instance_prompt: ''
+class_prompt: ''
+prior_loss_weight: 1.0
+num_class_images: 100
+# If true, set dataset_save_location.
+cache_dreambooth_dataset: False
+quantization: ''
+# Shard the range finding operation for quantization. By default this is set to number of slices.
+quantization_local_shard_count: -1
+use_qwix_quantization: False
+compile_topology_num_slices: -1 # Number of target slices, set to a positive integer.
+
+# ML Diagnostics settings
+enable_ml_diagnostics: False
+profiler_gcs_path: ""
+enable_ondemand_xprof: False