t5 encoder fp8 seems to be working

2026-04-12 13:50:34 -04:00
parent 2818b41004
commit fcf0be38bc
13 changed files with 505 additions and 67 deletions
@@ -0,0 +1,35 @@
+{
+  "_comment": "Wan2.2 i2v MoE 4-step distill, GGUF quantized. Uses QuantStack/Wan2.2-I2V-A14B-GGUF checkpoints instead of fp8 safetensors. GGUF does not support block-level offload so offload_granularity is set to 'model' — the entire DIT is moved to GPU when active. With Q4_K_M (~9.65 GB per expert) this fits comfortably in 24+ GB VRAM. high_noise_quantized_ckpt / low_noise_quantized_ckpt are filled in at runtime by server/video_models/wan22.py. IMPORTANT: GGUF dequantizes to fp16, so you must set DTYPE=FP16 in the container environment.",
+
+  "infer_steps": 4,
+  "target_video_length": 81,
+  "text_len": 512,
+
+  "resize_mode": "adaptive",
+  "resolution": "480p",
+  "target_height": 480,
+  "target_width": 480,
+  "fps": 16,
+
+  "self_attn_1_type": "flash_attn3",
+  "cross_attn_1_type": "flash_attn3",
+  "cross_attn_2_type": "flash_attn3",
+
+  "sample_guide_scale": [3.5, 3.5],
+  "sample_shift": 5.0,
+  "enable_cfg": false,
+
+  "cpu_offload": true,
+  "offload_granularity": "model",
+  "t5_cpu_offload": true,
+  "vae_cpu_offload": false,
+
+  "use_image_encoder": false,
+
+  "boundary_step_index": 2,
+  "denoising_step_list": [1000, 750, 500, 250],
+
+  "dit_quantized": true,
+  "dit_quant_scheme": "gguf-Q4_K_M",
+  "t5_quantized": false
+}