[TRTLLM-14135][feat] Add Qwen-Image-Edit-2511 support

docs/source/models/supported-models.md

@@ -165,6 +165,7 @@ For full documentation, see the [Visual Generation](./visual-generation.md) page
 | `Lightricks/LTX-2` | Text-to-Video (with Audio), Image-to-Video (with Audio) |
 | `Qwen/Qwen-Image` | Text-to-Image |
 | `Qwen/Qwen-Image-2512` | Text-to-Image |
+| `Qwen/Qwen-Image-Edit-2511` | Image Editing |
 | `nvidia/Cosmos3-Nano` | Text-to-Image, Text-to-Video, Image-to-Video |
 | `nvidia/Cosmos3-Super` | Text-to-Image, Text-to-Video, Image-to-Video |
 
@@ -178,6 +179,7 @@ For full documentation, see the [Visual Generation](./visual-generation.md) page
 | **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 |
 | **Qwen-Image** [^2] | Yes | Yes | No | No | Yes | No | Yes | Yes | Yes | Yes | Yes | No |
+| **Qwen-Image-Edit-2511** | Yes | Yes | No | Yes | No | No | Yes | Yes | Yes | No | No | No |
 | **Cosmos3** | Yes | Yes | No | Yes | Yes | Yes | Yes | Yes | Yes | No | No | Yes |
 
 [^vg1]: FLUX models use embedded guidance and do not have a separate negative prompt path, so CFG parallelism is not applicable.

docs/source/models/visual-generation.md

@@ -37,6 +37,7 @@ TensorRT-LLM **VisualGen** provides a unified inference stack for diffusion mode
 | `Lightricks/LTX-2` | Text-to-Video (with Audio), Image-to-Video (with Audio) |
 | `Qwen/Qwen-Image` | Text-to-Image |
 | `Qwen/Qwen-Image-2512` | Text-to-Image |
+| `Qwen/Qwen-Image-Edit-2511` | Image Editing |
 | `nvidia/Cosmos3-Nano` | Text-to-Image, Text-to-Video, Image-to-Video |
 | `nvidia/Cosmos3-Super` | Text-to-Image, Text-to-Video, Image-to-Video |
 
@@ -53,6 +54,7 @@ Models are auto-detected from the checkpoint directory. Diffusers-format models
 | **Wan 2.2** | Yes | Yes | Yes [^3] | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | No |
 | **LTX-2** | Yes | Yes | Yes [^4] | Yes | Yes | Yes | No | No | Yes | Yes | Yes | Yes | No | No |
 | **Qwen-Image** [^5] | Yes | Yes | No | No | No | Yes | No | Yes | Yes | Yes | Yes | Yes | No | No |
+| **Qwen-Image-Edit-2511** | Yes | Yes | No | No | Yes | No | No | Yes | Yes | Yes | No | No | No | No |
 | **Cosmos3** | Yes | Yes | No | No | Yes | Yes | Yes | Yes | Yes | Yes | No | No | Yes | No |
 
 [^1]: FLUX models use embedded guidance and do not have a separate negative prompt path, so CFG parallelism is not applicable.

examples/visual_gen/README.md

@@ -23,6 +23,7 @@ python models/flux1.py
 python models/flux2.py
 python models/cosmos3_ti2v.py --prompt "A robot arm picks fruit in a grocery store"
 python models/qwen_image.py
+python models/qwen_image_edit.py --image /path/to/source.png --prompt "Make the image look like a watercolor painting"
 
 # With engine config (quant, parallelism, etc.)
 python models/wan_t2v.py --visual_gen_args configs/wan2.2-t2v-fp4-1gpu.yaml
@@ -32,6 +33,7 @@ python models/flux1.py --visual_gen_args configs/flux1-dev-fp4-1gpu.yaml
 python models/flux2.py --visual_gen_args configs/flux2-dev-fp4-1gpu.yaml
 python models/cosmos3_ti2v.py --visual_gen_args configs/cosmos3-nano-1gpu.yaml --prompt "A robot arm picks fruit in a grocery store"
 python models/qwen_image.py --visual_gen_args configs/qwen-image-fp8-1gpu.yaml
+python models/qwen_image_edit.py --visual_gen_args configs/qwen-image-edit-2511-fp4-1gpu.yaml --image /path/to/source.png --prompt "Make the image look like a watercolor painting"
 ```
 
 Install deps from the repo root: `pip install -r requirements-dev.txt`.

examples/visual_gen/configs/qwen-image-edit-2511-fp4-1gpu.yaml

@@ -0,0 +1,35 @@
+# 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 Qwen-Image-Edit-2511 with selective NVFP4 dynamic quantization.
+# Q/K/V projections remain in BF16 because they preserve accuracy
+# Shared by offline examples (--visual_gen_args) and trtllm-serve.
+quant_config:
+  quant_algo: NVFP4
+  dynamic: true
+  ignore:
+    - "transformer_blocks.*.attn.to_q"
+    - "transformer_blocks.*.attn.to_k"
+    - "transformer_blocks.*.attn.to_v"
+    - "transformer_blocks.*.attn.add_q_proj"
+    - "transformer_blocks.*.attn.add_k_proj"
+    - "transformer_blocks.*.attn.add_v_proj"
+attention_config:
+  backend: VANILLA
+parallel_config:
+  cfg_size: 1
+  ulysses_size: 1
+cuda_graph_config:
+  enable: false

examples/visual_gen/configs/qwen-image-edit-2511-fp8-1gpu.yaml

@@ -0,0 +1,26 @@
+# Copyright (c) 2026, NVIDIA CORPORATION.  All rights reserved.
+#
+# 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 Qwen-Image-Edit-2511 with FP8 blockwise dynamic quantization.
+# Shared by offline examples (--visual_gen_args) and trtllm-serve.
+quant_config:
+  quant_algo: FP8_BLOCK_SCALES
+  dynamic: true
+attention_config:
+  backend: VANILLA
+parallel_config:
+  cfg_size: 1
+  ulysses_size: 1
+cuda_graph_config:
+  enable: false

examples/visual_gen/models/qwen_image_edit.py

@@ -0,0 +1,74 @@
+# Copyright (c) 2026, NVIDIA CORPORATION.  All rights reserved.
+#
+# 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.
+
+"""Qwen-Image-Edit-2511 image editing.
+
+Usage:
+    python qwen_image_edit.py --image input.png --prompt "Make the image look like a watercolor painting"
+    python qwen_image_edit.py \
+        --visual_gen_args ../configs/qwen-image-edit-2511-fp8-1gpu.yaml \
+        --image input.png
+"""
+
+import argparse
+
+from tensorrt_llm import VisualGen, VisualGenArgs
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "--model",
+        default="Qwen/Qwen-Image-Edit-2511",
+        help="Hugging Face model id or local checkpoint path.",
+    )
+    parser.add_argument(
+        "--visual_gen_args",
+        dest="visual_gen_args",
+        type=str,
+        default=None,
+        help="Path to YAML config (same as trtllm-serve --visual_gen_args)",
+    )
+    parser.add_argument(
+        "--image",
+        nargs="+",
+        required=True,
+        help="One or more input image paths or URLs.",
+    )
+    parser.add_argument(
+        "--prompt",
+        default="Make the image look like a watercolor painting while preserving the main subject.",
+        help="Text edit instruction.",
+    )
+    parser.add_argument(
+        "--output_path",
+        default="qwen_image_edit_output.png",
+        help="Edited image output path.",
+    )
+    return parser.parse_args()
+
+
+def main() -> None:
+    args = parse_args()
+    extra_args = VisualGenArgs.from_yaml(args.visual_gen_args) if args.visual_gen_args else None
+    visual_gen = VisualGen(model=args.model, args=extra_args)
+    params = visual_gen.default_params
+    params.image = args.image if len(args.image) > 1 else args.image[0]
+    output = visual_gen.generate(inputs=args.prompt, params=params)
+    saved = output.save(args.output_path)
+    print(f"Saved edited image to {saved}")
+
+
+if __name__ == "__main__":
+    main()

tensorrt_llm/_torch/visual_gen/cuda_graph_runner.py

@@ -77,6 +77,12 @@ def __init__(
 
         self.graphs: Dict[KeyType, torch.cuda.CUDAGraph] = {}
         self.graph_outputs: Dict[KeyType, Any] = {}  # weak refs
+        # Keep the capture-created output tensors alive. Some pipelines keep
+        # outputs from multiple graph keys live at once, for example true CFG
+        # combines positive and negative Qwen-Image-Edit transformer outputs.
+        # Without strong refs, a later graph capture using the same memory pool
+        # can reuse the earlier output storage and make those outputs alias.
+        self._graph_output_refs: Dict[KeyType, Any] = {}
         self.static_inputs: Dict[KeyType, Tuple[List[Any], Dict[str, Any]]] = {}
         self.memory_pool = config.cuda_graph_mem_pool
 
@@ -147,6 +153,7 @@ def capture(
 
         self.graphs[key] = graph
         self.static_inputs[key] = (static_args, static_kwargs)
+        self._graph_output_refs[key] = output
         self.graph_outputs[key] = make_weak_ref(output)
         self.memory_pool = graph.pool()
 
@@ -209,5 +216,6 @@ def clear(self):
             graph.reset()
         self.graphs.clear()
         self.graph_outputs.clear()
+        self._graph_output_refs.clear()
         self.static_inputs.clear()
         self.memory_pool = None

tensorrt_llm/_torch/visual_gen/models/__init__.py

@@ -36,14 +36,15 @@
 from .cosmos3 import Cosmos3OmniMoTPipeline
 from .flux import Flux2Pipeline, FluxPipeline
 from .ltx2 import LTX2Pipeline  # noqa: F401
-from .qwen_image import QwenImagePipeline
+from .qwen_image import QwenImageEditPlusPipeline, QwenImagePipeline
 from .wan import WanImageToVideoPipeline, WanPipeline
 
 __all__ = [
     "AutoPipeline",
     "BasePipeline",
     "FluxPipeline",
     "Flux2Pipeline",
+    "QwenImageEditPlusPipeline",
     "QwenImagePipeline",
     "WanPipeline",
     "WanImageToVideoPipeline",

tensorrt_llm/_torch/visual_gen/models/qwen_image/__init__.py

@@ -9,6 +9,7 @@
 """
 
 from .pipeline_qwen_image import QwenImagePipeline
+from .pipeline_qwen_image_edit import QwenImageEditPlusPipeline
 from .transformer_qwen_image import (
     AdaLayerNormContinuous,
     QwenEmbedRope,
@@ -25,6 +26,7 @@
 __all__ = [
     "AdaLayerNormContinuous",
     "QwenEmbedRope",
+    "QwenImageEditPlusPipeline",
     "QwenImagePipeline",
     "QwenImageTransformer2DModel",
     "QwenImageTransformerBlock",