[None][feat] Add LTX-2.3 t2v and retake pipeline

docs/source/models/supported-models.md

@@ -218,6 +218,7 @@ For full documentation, see the [Visual Generation](./visual-generation.md) page
 | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` | Image-to-Video |
 | `Wan-AI/Wan2.2-TI2V-5B-Diffusers` | Text-to-Video, Image-to-Video |
 | `Lightricks/LTX-2` | Text-to-Video (with Audio), Image-to-Video (with Audio) |
+| `Lightricks/LTX-2.3` | Text-to-Video (with Audio), Video Retake (with Audio) |
 | `MiniMaxAI/MiniMax-H3` | Text-to-Video (with Audio), First/Last-Frame-to-Video (with Audio) |
 | `Qwen/Qwen-Image` | Text-to-Image |
 | `Qwen/Qwen-Image-2512` | Text-to-Image |
@@ -238,6 +239,7 @@ For full documentation, see the [Visual Generation](./visual-generation.md) page
 | **Wan 2.1** | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
 | **Wan 2.2** | Yes | Yes | No | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
 | **LTX-2** | Yes | Yes | No | Yes | Yes | No | No | Yes | Yes | Yes | Yes | No |
+| **LTX-2.3** | Yes | Yes | No | No | Yes | Yes | Yes | Yes | Yes | No | No | No |
 | **MiniMax-H3** | Yes | Yes | No | No | Yes | Yes | No | Yes | Yes | No | No | No |
 | **Qwen-Image** | Yes | Yes | Yes | Yes | Yes | No | Yes | Yes | Yes | Yes | Yes | No |
 | **Qwen-Image-Layered** | Yes | No | No | No | No | No | Yes | Yes | Yes | No | No | No |

docs/source/models/visual-generation.md

@@ -37,6 +37,7 @@ TensorRT-LLM **VisualGen** provides a unified inference stack for diffusion mode
 | `Wan-AI/Wan2.2-TI2V-5B-Diffusers` | Text-to-Video, Image-to-Video |
 | `FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers` | Text-to-Video (3-step distilled) |
 | `Lightricks/LTX-2` | Text-to-Video (with Audio), Image-to-Video (with Audio) |
+| `Lightricks/LTX-2.3` | Text-to-Video (with Audio), Video Retake (with Audio) |
 | `Qwen/Qwen-Image` | Text-to-Image |
 | `Qwen/Qwen-Image-2512` | Text-to-Image |
 | `Qwen/Qwen-Image-Layered` | Image-to-Image |
@@ -65,6 +66,7 @@ Models are auto-detected from the checkpoint directory. Diffusers-format models
 | **Wan 2.2** | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes |
 | **FastWan 2.2** | Yes | Yes | No | No | No | No | No | No | Yes | Yes | Yes | No | No | No | No |
 | **LTX-2** | Yes | Yes | Yes | Yes | No | Yes | Yes | No | No | Yes | Yes | Yes | Yes | No | Yes |
+| **LTX-2.3** | Yes | Yes | No | No | No | No | Yes | Yes | Yes | Yes | Yes | No | No | No | No |
 | **MiniMax-H3** | Yes | Yes | No | No | No | No | Yes | Yes | No | Yes | Yes | No | No | No | No |
 | **Qwen-Image** | Yes | Yes | Yes | Yes | No | Yes | Yes | No | Yes | Yes | Yes | Yes | Yes | No | No |
 | **Qwen-Image-Layered** | No | No | No | No | No | No | No | No | Yes | Yes | Yes | No | No | No | No |
@@ -392,7 +394,7 @@ Configured under `VisualGenArgs.parallel_config`. Modes can be combined:
 - **CFG Parallelism** (`cfg_size: 2`): Splits positive/negative guidance prompts across GPUs. FLUX uses embedded guidance without a separate negative prompt path; CFG parallelism is not applicable to FLUX or the distilled FastWan 2.2 model.
 - **Ulysses Parallelism** (`ulysses_size: N`): Splits the sequence dimension across GPUs for longer sequences.
     - **Async Ulysses A2A pipeline** (`async_ulysses: true` in `parallel_config`): Overlaps per-rank V/Q/K projection compute with the cross-rank all-to-all on a dedicated side stream. Requires `ulysses_size > 1` and an NVLink-connected GPU domain (uses PyTorch `_SymmetricMemory` with CUDA IPC for peer pushes; not currently supported across nodes without MNNVL). Currently wired for WAN and LTX-2 self-attention.
-- **Parallel VAE** (`parallel_vae_size: N`): Parallelizes the final VAE decode (constraint: `parallel_vae_size ≤ world_size`). WAN/Cosmos3 shard along a spatial axis; LTX-2 and MiniMax-H3 distribute independent tiles.
+- **Parallel VAE** (`parallel_vae_size: N`): Parallelizes the final VAE decode (constraint: `parallel_vae_size ≤ world_size`). WAN/Cosmos3 shard along a spatial axis; LTX-2, LTX-2.3, and MiniMax-H3 distribute independent tiles.
 - **Context Parallel (CP)** — Partitions the sequence into shards so that each rank computes partial attention. Requires an LSE-capable attention backend (`FA4` or `CUTEDSL`). CP can be composed with Ulysses, giving a total sequence-parallel (SP) degree = `cp_size · ulysses_size`. The CP degree depends on the implementation below:
     - **Attention2D** (`attn2d_size: [N, M]`): Shards the sequence axis across an `N × M` device mesh (CP degree = `N · M`; total SP degree = `N · M · ulysses_size`).
     - **Ring Attention** (`ring_size: N`): Shards the sequence axis across a 1D ring of `N` ranks, streaming K/V blocks (CP degree = `N`; total SP degree = `N · ulysses_size`; mutually exclusive with Attention2D).

examples/visual_gen/README.md

@@ -16,13 +16,20 @@ for feature details.
 
 ```bash
 python quickstart_example.py
+python models/ltx23.py
 python models/glm_image.py
 
 # With engine config (quant, parallelism, etc.)
 # Omit --visual_gen_args and its YAML path to use default settings.
 python models/wan_t2v.py --visual_gen_args configs/wan2.2-t2v-fp4-1gpu.yaml
 python models/wan_i2v.py --visual_gen_args configs/wan2.2-i2v-fp4-1gpu.yaml --image /path/to/image.png
 python models/ltx2.py --visual_gen_args configs/ltx2-1gpu.yaml
+python models/ltx23.py --visual_gen_args configs/ltx23-t2v-bf16-1gpu.yaml
+python models/ltx23.py --visual_gen_args configs/ltx23-t2v-fp8-1gpu.yaml
+python models/ltx23.py --visual_gen_args configs/ltx23-t2v-fp4-1gpu.yaml
+python models/ltx23_retake.py --visual_gen_args configs/ltx23-retake-bf16-1gpu.yaml --source /path/to/retake_input.mp4
+python models/ltx23_retake.py --visual_gen_args configs/ltx23-retake-fp8-1gpu.yaml --source /path/to/retake_input.mp4
+python models/ltx23_retake.py --visual_gen_args configs/ltx23-retake-fp4-1gpu.yaml --source /path/to/retake_input.mp4
 # FP8 blockwise config: configs/minimax-h3-fp8-blockwise-1gpu.yaml.
 # 8-GPU Ulysses config: configs/minimax-h3-bf16-8gpu.yaml.
 python models/minimax_h3.py --model <approved-checkpoint> --visual_gen_args configs/minimax-h3-bf16-1gpu.yaml

examples/visual_gen/configs/ltx23-retake-bf16-1gpu.yaml

@@ -0,0 +1,27 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# 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
+#
+# http://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.
+
+# 1-GPU BF16 LTX-2.3 retake.
+pipeline_config:
+  workflow: retake
+attention_config:
+  backend: VANILLA
+parallel_config:
+  cfg_size: 1
+  ulysses_size: 1
+torch_compile_config:
+  enable: true
+cuda_graph_config:
+  enable: true

examples/visual_gen/configs/ltx23-retake-bf16-2gpu.yaml

@@ -0,0 +1,28 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# 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
+#
+# http://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.
+
+# 2-GPU BF16 LTX-2.3 retake.
+pipeline_config:
+  workflow: retake
+attention_config:
+  backend: VANILLA
+parallel_config:
+  cfg_size: 1
+  ulysses_size: 2
+  parallel_vae_size: 2
+torch_compile_config:
+  enable: true
+cuda_graph_config:
+  enable: true

examples/visual_gen/configs/ltx23-retake-fp4-1gpu.yaml

@@ -0,0 +1,32 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# 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
+#
+# http://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.
+
+# 1-GPU LTX-2.3 retake with dynamic NVFP4 transformer quantization at every step
+# and FlashInfer NVFP4 Q/K + FP8 V video self-attention on SM100/SM103 GPUs.
+pipeline_config:
+  workflow: retake
+quant_config:
+  quant_algo: NVFP4
+  dynamic: true
+attention_config:
+  backend: FLASHINFER
+  quant_attention_config:
+    qk_dtype: nvfp4
+    v_dtype: fp8
+parallel_config:
+  cfg_size: 1
+  ulysses_size: 1
+cuda_graph_config:
+  enable: true

examples/visual_gen/configs/ltx23-retake-fp8-1gpu.yaml

@@ -0,0 +1,32 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# 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
+#
+# http://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.
+
+# 1-GPU LTX-2.3 retake with dynamic FP8 block-scale transformer quantization
+# and FlashInfer MXFP8 Q/K + FP8 V video self-attention on SM100/SM103 GPUs.
+pipeline_config:
+  workflow: retake
+quant_config:
+  quant_algo: FP8_BLOCK_SCALES
+  dynamic: true
+attention_config:
+  backend: FLASHINFER
+  quant_attention_config:
+    qk_dtype: mxfp8
+    v_dtype: fp8
+parallel_config:
+  cfg_size: 1
+  ulysses_size: 1
+cuda_graph_config:
+  enable: true

examples/visual_gen/configs/ltx23-t2v-bf16-1gpu.yaml

@@ -0,0 +1,25 @@
+# SPDX-FileCopyrightText: Copyright (c) 2022-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# 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
+#
+# http://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.
+
+# 1-GPU LTX-2.3 text-to-video with audio, BF16.
+attention_config:
+  backend: VANILLA
+parallel_config:
+  cfg_size: 1
+  ulysses_size: 1
+torch_compile_config:
+  enable: true
+cuda_graph_config:
+  enable: true

examples/visual_gen/configs/ltx23-t2v-bf16-2gpu.yaml

@@ -0,0 +1,26 @@
+# SPDX-FileCopyrightText: Copyright (c) 2022-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# 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
+#
+# http://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.
+
+# 2-GPU LTX-2.3 text-to-video with audio, BF16.
+attention_config:
+  backend: VANILLA
+parallel_config:
+  cfg_size: 1
+  ulysses_size: 2
+  parallel_vae_size: 2
+torch_compile_config:
+  enable: true
+cuda_graph_config:
+  enable: true

examples/visual_gen/configs/ltx23-t2v-fp4-1gpu.yaml

@@ -0,0 +1,32 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# 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
+#
+# http://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.
+
+# 1-GPU LTX-2.3 text-to-video with audio, dynamic NVFP4 transformer weights
+# and FlashInfer NVFP4 Q/K + FP8 V video self-attention on SM100/SM103 GPUs.
+quant_config:
+  quant_algo: NVFP4
+  dynamic: true
+attention_config:
+  backend: FLASHINFER
+  quant_attention_config:
+    qk_dtype: nvfp4
+    v_dtype: fp8
+parallel_config:
+  cfg_size: 1
+  ulysses_size: 1
+torch_compile_config:
+  enable: true
+cuda_graph_config:
+  enable: true

examples/visual_gen/configs/ltx23-t2v-fp8-1gpu.yaml

@@ -0,0 +1,32 @@
+# SPDX-FileCopyrightText: Copyright (c) 2022-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# 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
+#
+# http://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.
+
+# 1-GPU LTX-2.3 text-to-video with audio, dynamic FP8 transformer weights
+# and FlashInfer MXFP8 Q/K + FP8 V video self-attention on SM100/SM103 GPUs.
+quant_config:
+  quant_algo: FP8_BLOCK_SCALES
+  dynamic: true
+attention_config:
+  backend: FLASHINFER
+  quant_attention_config:
+    qk_dtype: mxfp8
+    v_dtype: fp8
+parallel_config:
+  cfg_size: 1
+  ulysses_size: 1
+torch_compile_config:
+  enable: true
+cuda_graph_config:
+  enable: true