Eliovp's picture
Release Qwen_Image-2.1-MXFP4-Paiton with verified weights and benchmarks
49be4b6 verified
Raw History Blame Contribute Delete
2.05 kB
{
"id": "paiton-image-2.1",
"display_name": "Qwen_Image-2.1-MXFP4-Paiton",
"upstream": "Qwen/Qwen-Image-2.1",
"upstream_revision": "790c92633540aa0cb11d9abf19eb46d861714758",
"checkpoint_profile": "balanced",
"tensor_payload_bytes": 9313781512,
"architecture": "gfx1201",
"batch_size": 1,
"steps": 40,
"guidance": 1.0,
"native_region": "MXFP4 weight reconstruction",
"runtime": "Diffusers/BF16 with native HIP weight reconstruction",
"checkpoint_offload": false,
"status": "Hardware-validated bounded profile",
"tasks": {
"text-to-image": {
"sizes": [
[
1024,
1024
],
[
2048,
2048
]
],
"status": "passed hardware check for documented cases"
},
"rgba": {
"sizes": [
[
1024,
1024
],
[
2048,
2048
]
],
"status": "passed hardware check for documented cases"
},
"edit": {
"sizes": [
[
1024,
1024
]
],
"max_input_images": 1,
"status": "passed hardware check for documented cases",
"max_input_pixels": 4194304
}
},
"api": {
"generation": "/v1/images/generations",
"edit": "/v1/images/edits",
"model": "/v1/models",
"health": "/health"
},
"quality_scope": "Eight paired 2048-square prompts; measured loss, no unchanged-quality claim",
"performance_scope": "No demonstrated latency advantage over original Quark HIP decoder",
"memory_scope": "All components GPU resident, no CPU offload or VAE tiling; sampled peak below 30 GiB in documented workloads",
"repository": "EliovpAI/Qwen_Image-2.1-MXFP4-Paiton",
"tested_gpu": "AMD Radeon AI PRO R9700",
"hardware_compatibility": {
"checkpoint_format": "Architecture-independent packed MXFP4 storage; compatible loader required",
"packaged_native_target": "gfx1201",
"launcher_minimum_free_gib": 30,
"other_gpus": "Not qualified by this packaged launcher"
}
}