bhetherman
9debc56137
Benchmarks the dense 5B Turbo model (Q8_0 GGUF + fp8 T5) as a lower-VRAM alternative to the 14B MoE pipeline. Includes dtype patches for dense WanModel, Wan 2.2 VAE config (48 channels, 16x spatial), and Blackwell fp8 workaround. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
41 lines
977 B
JSON
41 lines
977 B
JSON
{
|
|
"_comment": "LightX2V config for Wan2.2-TI2V-5B-Turbo (dense, GGUF). Single DIT checkpoint (not MoE). dit_quantized_ckpt is filled in at runtime by setup_model.py / test_i2v.py.",
|
|
|
|
"infer_steps": 4,
|
|
"target_video_length": 81,
|
|
"text_len": 512,
|
|
|
|
"resize_mode": "adaptive",
|
|
"resolution": "480p",
|
|
"target_height": 480,
|
|
"target_width": 480,
|
|
"fps": 16,
|
|
|
|
"vae_stride": [4, 16, 16],
|
|
"num_channels_latents": 48,
|
|
|
|
"self_attn_1_type": "torch_sdpa",
|
|
"cross_attn_1_type": "torch_sdpa",
|
|
"cross_attn_2_type": "torch_sdpa",
|
|
"modulate_type": "torch",
|
|
"rope_type": "torch",
|
|
|
|
"sample_guide_scale": 1.0,
|
|
"sample_shift": 5.0,
|
|
"enable_cfg": false,
|
|
|
|
"cpu_offload": false,
|
|
"offload_granularity": "model",
|
|
"t5_cpu_offload": true,
|
|
"vae_cpu_offload": false,
|
|
|
|
"use_image_encoder": false,
|
|
|
|
"denoising_step_list": [1000, 750, 500, 250],
|
|
|
|
"dit_quantized": true,
|
|
"dit_quant_scheme": "gguf-Q8_0",
|
|
"t5_quantized": true,
|
|
"t5_quant_scheme": "fp8-sgl"
|
|
}
|