Cleaned Flux2 Klein Implementation, with benchmarking done on v6 TPU

README.md

@@ -81,6 +81,7 @@ MaxDiffusion supports
     - [LTX-2 Video](#ltx-2-video)
     - [Flux](#flux)
       - [Fused Attention for GPU](#fused-attention-for-gpu)
+    - [Flux.2-klein-4B](#flux2-klein-4b)
     - [SDXL](#stable-diffusion-xl)
     - [SD 2 base](#stable-diffusion-2-base)
     - [SD 2.1](#stable-diffusion-21)
@@ -675,6 +676,46 @@ We added ring attention support for Wan models. Below are the stats for one `720
   ```bash
   python src/maxdiffusion/generate_flux.py src/maxdiffusion/configs/base_flux_schnell.yml jax_cache_dir=/tmp/cache_dir run_name=flux_test output_dir=/tmp/ prompt="photograph of an electronics chip in the shape of a race car with trillium written on its side" per_device_batch_size=1 ici_data_parallelism=1 ici_fsdp_parallelism=-1 offload_encoders=False
   ```
+
+  ## Flux.2-klein-4B
+
+  MaxDiffusion supports JAX+TPU inference for the **Flux.2-klein-4B** model. We provide both a standard one-shot generation script and an **interactive generation mode** that reuses the JAX/XLA compiled graphs for instant subsequent runs.
+
+  ### Running Inference
+
+  To generate images, run the following command:
+
+  ```bash
+  python src/maxdiffusion/generate_flux2klein.py src/maxdiffusion/configs/base_flux2klein.yml
+  ```
+
+  To enable **interactive mode** (reusing the compiled graphs to avoid JIT compilation latency on subsequent prompts):
+
+  ```bash
+  python src/maxdiffusion/generate_flux2klein.py src/maxdiffusion/configs/base_flux2klein.yml interactive=True
+  ```
+
+  ### Running Tests & Verification
+
+  All tests and verification scripts are located in `src/maxdiffusion/tests/flux2klein/`.
+
+  To run the 19-point mathematical parity verification suite against PyTorch CPU:
+
+  ```bash
+  python src/maxdiffusion/tests/flux2klein/verify_blockwise_parity.py
+  ```
+
+  To run the pure TPU latency benchmark (permanent HBM residence):
+
+  ```bash
+  python src/maxdiffusion/tests/flux2klein/benchmark_tpu_pure.py
+  ```
+
+  To run all unit and integration tests:
+
+  ```bash
+  pytest src/maxdiffusion/tests/flux2klein/
+  ```
   ## Fused Attention for GPU:
   Fused Attention for GPU is supported via TransformerEngine. Installation instructions:
 

src/maxdiffusion/configs/base_flux2klein.yml

@@ -0,0 +1,271 @@
+# Copyright 2026 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.
+
+# This sentinel is a reminder to choose a real run name.
+run_name: 'flux2klein_test_run'
+
+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
+
+timing_metrics_file: "" # for testing, local file that stores function timing metrics such as state creation, compilation. If empty, no metrics are written.
+write_timing_metrics: True 
+
+gcs_metrics: False
+# If true save config to GCS in {base_output_directory}/{run_name}/
+save_config_to_gcs: False
+log_period: 100
+
+pretrained_model_name_or_path: 'black-forest-labs/FLUX.2-klein-4B'
+clip_model_name_or_path: 'ariG23498/clip-vit-large-patch14-text-flax'
+t5xxl_model_name_or_path: 'ariG23498/t5-v1-1-xxl-flax'
+
+# Flux params
+flux_name: "flux2klein"
+use_latents: False
+max_sequence_length: 512
+time_shift: True
+base_shift: 0.5
+max_shift: 1.15
+# offloads t5 encoder after text encoding to save memory.
+offload_encoders: True
+
+
+unet_checkpoint: ''
+revision: 'refs/pr/95'
+# This will convert the weights to this dtype.
+# When running inference on TPUv5e, use weights_dtype: 'bfloat16'
+weights_dtype: 'bfloat16'
+# 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
+from_pt: True
+split_head_dim: True
+attention: 'flash' # Supported attention: dot_product, flash, cudnn_flash_te
+# 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: {}
+# GroupNorm groups
+norm_num_groups: 32
+
+# If train_new_flux, flux weights will be randomly initialized to train flux from scratch
+# else they will be loaded from pretrained_model_name_or_path
+train_new_flux: False
+
+# train text_encoder - Currently not supported for SDXL
+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
+}
+
+# Override parameters from checkpoints's scheduler.
+diffusion_scheduler_config: {
+  _class_name: 'FlaxEulerDiscreteScheduler',
+  prediction_type: 'epsilon',
+  rescale_zero_terminal_snr: False,
+  timestep_spacing: 'trailing'
+}
+
+# Output directory
+# Create a GCS bucket, e.g. my-maxtext-outputs and set this to "gs://my-maxtext-outputs/"
+base_output_directory: ""
+
+# Hardware
+hardware: 'tpu' # Supported hardware types are 'tpu', 'gpu'
+skip_jax_distributed_system: False
+
+# 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'],
+                      ['mlp','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
+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: 'tfrecord'  # Options: 'tfrecord', 'hf', 'tf', 'grain', 'synthetic'
+cache_latents_text_encoder_outputs: True
+dataset_save_location: '/tmp/pokemon-gpt4-captions_xl'
+train_data_dir: ''
+dataset_config_name: ''
+jax_cache_dir: '/tmp/jax_cache'
+hf_data_dir: ''
+hf_train_files: ''
+hf_access_token: ''
+image_column: 'image'
+caption_column: 'text'
+resolution: 512
+center_crop: False
+random_flip: False
+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-5
+scale_lr: False
+max_train_samples: -1
+# max_train_steps takes priority over num_train_epochs.
+max_train_steps: 1500
+num_train_epochs: 1
+seed: 0
+output_dir: 'output/'
+per_device_batch_size: 1
+
+warmup_steps_fraction: 0.1
+learning_rate_schedule_steps: -1 # By default the length of the schedule is set to the number of steps.
+
+# 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: 0 # 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_profiler: 5
+profiler_steps: 10
+profiler: ""
+
+# Generation parameters
+prompt: "A detailed vector illustration of a robotic hummingbird || A cinematic shot of a neon-lit cyberpunk street"
+prompt_2: "A detailed vector illustration of a robotic hummingbird || A cinematic shot of a neon-lit cyberpunk street"
+negative_prompt: ""
+do_classifier_free_guidance: True
+guidance_scale: 4.0
+guidance_rescale: 0.0
+num_inference_steps: 4
+save_final_checkpoint: False
+
+# SDXL Lightning parameters
+lightning_from_pt: True
+lightning_repo: ""
+lightning_ckpt: ""
+
+# LoRA parameters
+lora_config: {
+  lora_model_name_or_path: [],
+  weight_name: [],
+  adapter_name: [],
+  scale: [],
+  from_pt: []
+}
+
+enable_mllog: False
+
+#controlnet
+controlnet_model_name_or_path: 'diffusers/controlnet-canny-sdxl-1.0'
+controlnet_from_pt: True
+controlnet_conditioning_scale: 0.5
+controlnet_image: 'https://upload.wikimedia.org/wikipedia/commons/thumb/c/c1/Google_%22G%22_logo.svg/1024px-Google_%22G%22_logo.svg.png'
+quantization: ''
+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
+
+# Specific additions for generate_flux2klein execution
+height: 1024
+width: 1024
+batch_size: 4
+interactive: False
+