[new-model] Add initial Qwen-Image-2.1 inference (#1918)

Co-authored-by: Aryan Kumar <156166681+aryan5v@users.noreply.github.com>

examples/inference/basic/basic_qwen_image21.py

@@ -0,0 +1,73 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Qwen-Image-2.1 generation, editing, references and transparent PNG output."""
+
+import argparse
+
+from fastvideo import VideoGenerator
+from fastvideo.api import (
+    EngineConfig,
+    GenerationRequest,
+    GeneratorConfig,
+    InputConfig,
+    OffloadConfig,
+    OutputConfig,
+    ParallelismConfig,
+    PipelineSelection,
+    SamplingConfig,
+)
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--model-path", default="Qwen/Qwen-Image-2.1")
+    parser.add_argument("--revision")
+    parser.add_argument("--prompt", default="A red ceramic teapot, isolated on a transparent background.")
+    parser.add_argument("--reference", action="append", default=[], help="Ordered reference image; repeat up to ten times")
+    parser.add_argument("--negative-prompt", default=None)
+    parser.add_argument("--true-cfg-scale", type=float, default=1.0)
+    parser.add_argument("--height", type=int, default=1024)
+    parser.add_argument("--width", type=int, default=1024)
+    parser.add_argument("--reference-resolution", type=int, default=1024)
+    parser.add_argument("--steps", type=int, default=40)
+    parser.add_argument("--seed", type=int, default=42)
+    parser.add_argument("--use-kv-cache", action=argparse.BooleanOptionalAction, default=True)
+    parser.add_argument("--output", default="outputs/qwen_image21/image.png")
+    args = parser.parse_args()
+    if len(args.reference) > 10:
+        parser.error("At most ten ordered reference images are supported")
+    workload = "i2i" if args.reference else "t2i"
+    generator = VideoGenerator.from_config(
+        GeneratorConfig(
+            model_path=args.model_path,
+            revision=args.revision,
+            engine=EngineConfig(
+                num_gpus=1,
+                parallelism=ParallelismConfig(tp_size=1, sp_size=1),
+                offload=OffloadConfig(dit=True, dit_layerwise=True, text_encoder=True, vae=True,
+                                      lazy_module_load=True),
+            ),
+            pipeline=PipelineSelection(
+                workload_type=workload,
+                experimental={
+                    "reference_resolution": args.reference_resolution,
+                    "use_kv_cache": args.use_kv_cache,
+                },
+            ),
+        ))
+    try:
+        generator.generate(
+            GenerationRequest(
+                prompt=args.prompt,
+                negative_prompt=args.negative_prompt,
+                inputs=InputConfig(references=args.reference or None),
+                sampling=SamplingConfig(height=args.height, width=args.width, num_frames=1, fps=1,
+                                        num_inference_steps=args.steps, guidance_scale=1.0,
+                                        true_cfg_scale=args.true_cfg_scale, seed=args.seed),
+                output=OutputConfig(output_path=args.output, save_video=True, return_frames=False),
+            ))
+    finally:
+        generator.shutdown()
+
+
+if __name__ == "__main__":
+    main()

examples/inference/basic/qwen_image21_edit.yaml

@@ -0,0 +1,39 @@
+# Replace source.png and mask.png with your own images. The mask is an ordered
+# reference interpreted by the model, not a hard pixel-preservation constraint.
+generator:
+  model_path: Qwen/Qwen-Image-2.1
+  engine:
+    num_gpus: 1
+    parallelism:
+      tp_size: 1
+      sp_size: 1
+    offload:
+      dit: true
+      dit_layerwise: true
+      text_encoder: true
+      vae: true
+      lazy_module_load: true
+  pipeline:
+    workload_type: i2i
+    experimental:
+      reference_resolution: 1024
+      use_kv_cache: true
+request:
+  prompt: In image 1, turn the object highlighted by the mask in image 2 blue.
+  inputs:
+    references:
+      - source.png
+      - mask.png
+  sampling:
+    height: 1024
+    width: 1024
+    num_frames: 1
+    fps: 1
+    num_inference_steps: 40
+    guidance_scale: 1.0
+    true_cfg_scale: 1.0
+    seed: 42
+  output:
+    output_path: outputs/qwen_image21/edit.png
+    save_video: true
+    return_frames: false

examples/inference/basic/qwen_image21_t2i.yaml

@@ -0,0 +1,40 @@
+# Run on an allocated CUDA GPU after installing FastVideo and providing weights.
+# fastvideo generate --config examples/inference/basic/qwen_image21_t2i.yaml
+generator:
+  model_path: Qwen/Qwen-Image-2.1
+  engine:
+    num_gpus: 1
+    compile:
+      enabled: false
+      backend: inductor
+      fullgraph: false
+    parallelism:
+      tp_size: 1
+      sp_size: 1
+    offload:
+      dit: true
+      dit_layerwise: true
+      text_encoder: true
+      vae: true
+      lazy_module_load: true
+  pipeline:
+    workload_type: t2i
+    experimental:
+      kv_cache_device: cuda
+      reference_resolution: 1024
+      use_kv_cache: true
+request:
+  prompt: A red ceramic teapot, isolated on a transparent background.
+  sampling:
+    height: 1024
+    width: 1024
+    num_frames: 1
+    fps: 1
+    num_inference_steps: 40
+    guidance_scale: 1.0
+    true_cfg_scale: 1.0
+    seed: 42
+  output:
+    output_path: outputs/qwen_image21/teapot.png
+    save_video: true
+    return_frames: false