Merge remote-tracking branch 'origin/main' into elisatsai_ring_attention

# Conflicts:
#	src/maxdiffusion/pipelines/wan/wan_pipeline.py
#	src/maxdiffusion/pipelines/wan/wan_pipeline_2_1.py

Author: eltsai

src/maxdiffusion/common_types.py

@@ -51,6 +51,7 @@
 
 WAN2_1 = "wan2.1"
 WAN2_2 = "wan2.2"
+LTX2_VIDEO = "ltx2_video"
 
 WAN_MODEL = WAN2_1
 

src/maxdiffusion/configs/base_wan_1_3b.yml

@@ -280,6 +280,9 @@ num_frames: 81
 guidance_scale: 5.0
 flow_shift: 3.0
 
+# Diffusion CFG cache (FasterCache-style, WAN 2.1 T2V only)
+use_cfg_cache: False
+
 # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
 guidance_rescale: 0.0
 num_inference_steps: 30

src/maxdiffusion/configs/base_wan_27b.yml

@@ -303,6 +303,9 @@ guidance_scale_high: 4.0
 # timestep to switch between low noise and high noise transformer
 boundary_ratio: 0.875
 
+# Diffusion CFG cache (FasterCache-style, WAN 2.1 T2V only)
+use_cfg_cache: False
+
 # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
 guidance_rescale: 0.0
 num_inference_steps: 30

src/maxdiffusion/configs/base_wan_i2v_14b.yml

@@ -286,6 +286,9 @@ num_frames: 81
 guidance_scale: 5.0
 flow_shift: 5.0
 
+# Diffusion CFG cache (FasterCache-style, WAN 2.1 T2V only)
+use_cfg_cache: False
+
 # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
 guidance_rescale: 0.0
 num_inference_steps: 50

src/maxdiffusion/configs/base_wan_i2v_27b.yml

@@ -298,6 +298,9 @@ guidance_scale_high: 4.0
 # timestep to switch between low noise and high noise transformer
 boundary_ratio: 0.875
 
+# Diffusion CFG cache (FasterCache-style)
+use_cfg_cache: False
+
 # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
 guidance_rescale: 0.0
 num_inference_steps: 50

src/maxdiffusion/configs/ltx2_video.yml

@@ -0,0 +1,87 @@
+#hardware
+hardware: 'tpu'
+skip_jax_distributed_system: False
+attention: 'flash'
+attention_sharding_uniform: True 
+precision: 'bf16'
+data_sharding: ['data', 'fsdp', 'context', 'tensor']
+remat_policy: "NONE"
+names_which_can_be_saved: []
+names_which_can_be_offloaded: []
+
+jax_cache_dir: ''
+weights_dtype: 'bfloat16'
+activations_dtype: 'bfloat16'
+
+run_name: ''
+output_dir: ''
+config_path: ''
+save_config_to_gcs: False
+
+frame_rate: 30
+max_sequence_length: 1024
+sampler: "from_checkpoint"
+
+# Generation parameters
+dataset_name: ''
+dataset_save_location: ''
+global_batch_size_to_train_on: 1
+num_inference_steps: 40
+guidance_scale: 3.0
+fps: 24
+prompt: "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window."
+negative_prompt: "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
+height: 512
+width: 768
+num_frames: 121
+decode_timestep: 0.05
+decode_noise_scale: 0.025
+quantization: "int8"
+seed: 10
+#parallelism
+mesh_axes: ['data', 'fsdp', 'context', 'tensor']
+logical_axis_rules: [
+                      ['batch', 'data'],
+                      ['activation_heads', 'fsdp'],
+                      ['activation_batch', 'data'],
+                      ['activation_kv', 'tensor'],
+                      ['mlp','tensor'],
+                      ['embed','fsdp'],
+                      ['heads', 'tensor'],
+                      ['norm', 'fsdp'],
+                      ['conv_batch', ['data','fsdp']],
+                      ['out_channels', 'tensor'],
+                      ['conv_out', 'fsdp'],
+                      ['conv_in', 'fsdp']
+                    ]
+dcn_data_parallelism: 1
+dcn_fsdp_parallelism: -1
+dcn_context_parallelism: 1
+dcn_tensor_parallelism: 1
+ici_data_parallelism: 1
+ici_fsdp_parallelism: -1
+ici_context_parallelism: 1
+ici_tensor_parallelism: 1
+enable_profiler: False
+
+replicate_vae: False
+
+allow_split_physical_axes: False
+learning_rate_schedule_steps: -1
+max_train_steps: 500
+pretrained_model_name_or_path: 'Lightricks/LTX-2'
+model_name: "ltx2_video"
+model_type: "T2V"
+unet_checkpoint: ''
+checkpoint_dir: ""
+cache_latents_text_encoder_outputs: True
+per_device_batch_size: 1
+compile_topology_num_slices: -1 
+quantization_local_shard_count: -1
+use_qwix_quantization: False
+weight_quantization_calibration_method: "absmax"
+act_quantization_calibration_method: "absmax"
+bwd_quantization_calibration_method: "absmax"
+qwix_module_path: ".*"
+jit_initializers: True 
+enable_single_replica_ckpt_restoring: False

src/maxdiffusion/generate_wan.py

@@ -123,6 +123,7 @@ def call_pipeline(config, pipeline, prompt, negative_prompt, num_inference_steps
           num_inference_steps=steps,
           guidance_scale_low=config.guidance_scale_low,
           guidance_scale_high=config.guidance_scale_high,
+          use_cfg_cache=config.use_cfg_cache,
       )
     else:
       raise ValueError(f"Unsupported model_name for I2V in config: {model_key}")
@@ -148,6 +149,7 @@ def call_pipeline(config, pipeline, prompt, negative_prompt, num_inference_steps
           num_inference_steps=steps,
           guidance_scale_low=config.guidance_scale_low,
           guidance_scale_high=config.guidance_scale_high,
+          use_cfg_cache=config.use_cfg_cache,
       )
     else:
       raise ValueError(f"Unsupported model_name for T2Vin config: {model_key}")

src/maxdiffusion/models/attention_flax.py

@@ -1033,18 +1033,28 @@ def __init__(
       )
 
   def _apply_rope(self, xq: jax.Array, xk: jax.Array, freqs_cis: jax.Array) -> Tuple[jax.Array, jax.Array]:
-    dtype = xq.dtype
-    reshape_xq = xq.astype(jnp.float32).reshape(*xq.shape[:-1], -1, 2)
-    reshape_xk = xk.astype(jnp.float32).reshape(*xk.shape[:-1], -1, 2)
+    # 1. Extract cos and sin, keeping them in native bfloat16
+    cos = jnp.real(freqs_cis).astype(xq.dtype)
+    sin = jnp.imag(freqs_cis).astype(xq.dtype)
 
-    xq_ = jax.lax.complex(reshape_xq[..., 0], reshape_xq[..., 1])
-    xk_ = jax.lax.complex(reshape_xk[..., 0], reshape_xk[..., 1])
+    # 2. Reshape the last dimension into pairs
+    xq_reshaped = xq.reshape(*xq.shape[:-1], -1, 2)
+    xk_reshaped = xk.reshape(*xk.shape[:-1], -1, 2)
 
-    xq_out_complex = xq_ * freqs_cis
-    xk_out_complex = xk_ * freqs_cis
+    # 3. Unbind the pairs
+    xq_0, xq_1 = xq_reshaped[..., 0], xq_reshaped[..., 1]
+    xk_0, xk_1 = xk_reshaped[..., 0], xk_reshaped[..., 1]
 
-    xq_out = jnp.stack([jnp.real(xq_out_complex), jnp.imag(xq_out_complex)], axis=-1).reshape(xq.shape).astype(dtype)
-    xk_out = jnp.stack([jnp.real(xk_out_complex), jnp.imag(xk_out_complex)], axis=-1).reshape(xk.shape).astype(dtype)
+    # 4. Pure real arithmetic (XLA will fuse these instantly into FMA instructions)
+    xq_out_0 = xq_0 * cos - xq_1 * sin
+    xq_out_1 = xq_0 * sin + xq_1 * cos
+
+    xk_out_0 = xk_0 * cos - xk_1 * sin
+    xk_out_1 = xk_0 * sin + xk_1 * cos
+
+    # 5. Stack and reshape back to original
+    xq_out = jnp.stack([xq_out_0, xq_out_1], axis=-1).reshape(xq.shape)
+    xk_out = jnp.stack([xk_out_0, xk_out_1], axis=-1).reshape(xk.shape)
 
     return xq_out, xk_out
 

src/maxdiffusion/models/embeddings_flax.py

@@ -501,3 +501,78 @@ def __call__(self, timestep, guidance, pooled_projection):
     conditioning = time_guidance_emb + pooled_projections
 
     return conditioning
+
+
+class NNXTimesteps(nnx.Module):
+
+  def __init__(self, num_channels: int, flip_sin_to_cos: bool, downscale_freq_shift: float, scale: int = 1):
+    self.num_channels = num_channels
+    self.flip_sin_to_cos = flip_sin_to_cos
+    self.downscale_freq_shift = downscale_freq_shift
+    self.scale = scale
+
+  def __call__(self, timesteps: jax.Array) -> jax.Array:
+    return get_sinusoidal_embeddings(
+        timesteps=timesteps,
+        embedding_dim=self.num_channels,
+        freq_shift=self.downscale_freq_shift,
+        flip_sin_to_cos=self.flip_sin_to_cos,
+        scale=self.scale,
+    )
+
+
+class NNXPixArtAlphaCombinedTimestepSizeEmbeddings(nnx.Module):
+
+  def __init__(
+      self,
+      rngs: nnx.Rngs,
+      embedding_dim: int,
+      size_emb_dim: int,
+      use_additional_conditions: bool = False,
+      dtype: jnp.dtype = jnp.float32,
+      weights_dtype: jnp.dtype = jnp.float32,
+  ):
+    self.outdim = size_emb_dim
+    self.use_additional_conditions = use_additional_conditions
+
+    self.time_proj = NNXTimesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
+    self.timestep_embedder = NNXTimestepEmbedding(
+        rngs=rngs, in_channels=256, time_embed_dim=embedding_dim, dtype=dtype, weights_dtype=weights_dtype
+    )
+
+    if use_additional_conditions:
+      self.additional_condition_proj = NNXTimesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
+      self.resolution_embedder = NNXTimestepEmbedding(
+          rngs=rngs, in_channels=256, time_embed_dim=size_emb_dim, dtype=dtype, weights_dtype=weights_dtype
+      )
+      self.aspect_ratio_embedder = NNXTimestepEmbedding(
+          rngs=rngs, in_channels=256, time_embed_dim=size_emb_dim, dtype=dtype, weights_dtype=weights_dtype
+      )
+
+  def __call__(
+      self,
+      timestep: jax.Array,
+      resolution: Optional[jax.Array] = None,
+      aspect_ratio: Optional[jax.Array] = None,
+      hidden_dtype: jnp.dtype = jnp.float32,
+  ) -> jax.Array:
+    timesteps_proj = self.time_proj(timestep)
+    timesteps_emb = self.timestep_embedder(timesteps_proj.astype(hidden_dtype))
+
+    if self.use_additional_conditions:
+      if resolution is None or aspect_ratio is None:
+        raise ValueError("resolution and aspect_ratio must be provided when use_additional_conditions is True")
+
+      resolution_emb = self.additional_condition_proj(resolution.flatten()).astype(hidden_dtype)
+      resolution_emb = self.resolution_embedder(resolution_emb)
+      resolution_emb = resolution_emb.reshape(timestep.shape[0], -1)
+
+      aspect_ratio_emb = self.additional_condition_proj(aspect_ratio.flatten()).astype(hidden_dtype)
+      aspect_ratio_emb = self.aspect_ratio_embedder(aspect_ratio_emb)
+      aspect_ratio_emb = aspect_ratio_emb.reshape(timestep.shape[0], -1)
+
+      conditioning = timesteps_emb + jnp.concatenate([resolution_emb, aspect_ratio_emb], axis=1)
+    else:
+      conditioning = timesteps_emb
+
+    return conditioning

src/maxdiffusion/models/ltx2/__init__.py

@@ -0,0 +1,15 @@
+"""
+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.
+"""