t5 encoder fp8 seems to be working

config.yml

@@ -32,16 +32,23 @@ video:
       casual gestures, natural lighting, soft focus background
     prompt_reply_words: 18        # max words lifted from reply to inject as {reply_hint}
 
-  # Model sources for the video stack. The fp8 e4m3 4-step distilled DIT
-  # weights from lightx2v/Wan2.2-Distill-Models are ~15 GB each (vs ~28 GB
-  # bf16) — that's the "save VRAM" path. T5/VAE/tokenizer still come from
-  # the Wan-AI base repo. Both repos download on first run into
-  # HF_HOME=/cache/huggingface.
+  # Model sources for the video stack. T5/VAE/tokenizer come from the
+  # Wan-AI base repo. DIT weights come from wan22_dit_repo in the format
+  # specified by wan22_dit_quant_scheme. Both repos download on first run
+  # into HF_HOME=/cache/huggingface.
+  #
+  # Supported dit_quant_scheme values:
+  #   fp8-sgl    — fp8 e4m3 safetensors (~15 GB/expert, from lightx2v/Wan2.2-Distill-Models)
+  #   gguf-Q4_K_M — GGUF 4-bit (~9.65 GB/expert, from QuantStack/Wan2.2-I2V-A14B-GGUF)
+  #   gguf-Q8_0  — GGUF 8-bit (~15.4 GB/expert)
+  #   (any gguf-<level> supported by LightX2V — see base_model.py MM_WEIGHT_REGISTER)
   models:
     wan22_base_repo: Wan-AI/Wan2.2-I2V-A14B
-    wan22_fp8_repo: lightx2v/Wan2.2-Distill-Models
+    wan22_dit_repo: QuantStack/Wan2.2-I2V-A14B-GGUF
+    wan22_dit_quant_scheme: gguf-Q4_K_M
+    wan22_t5_quantized: true
     wan22_model_cls: wan2.2_moe_distill
-    wan22_config_json: /app/configs/lightx2v/wan22_i2v_fp8_distill.json
+    wan22_config_json: /app/configs/lightx2v/wan22_i2v_gguf_distill.json
     musetalk_path: TMElyralab/MuseTalk
 
   # LoRAs applied to the fp8 base at load time via runtime switch_lora.