feat: update quantization configuration to use fp8-sgl as default schemes; remove unused quantization backend logic for cleaner implementation

2025-10-21 17:18:59 +08:00
parent b450902437
commit 50844d8dfe
3 changed files with 16 additions and 49 deletions
@@ -95,15 +95,13 @@ class TeaCacheConfig:
 class QuantizationConfig:
    """Quantization configuration."""

-    quant_op: str = "none"
-    dit_quant_scheme: str = "bf16"
-    t5_quant_scheme: str = "bf16"
-    clip_quant_scheme: str = "fp16"
-    adapter_quant_scheme: str = "bf16"
+    dit_quant_scheme: str = "fp8-sgl"
+    t5_quant_scheme: str = "fp8-sgl"
+    clip_quant_scheme: str = "fp8-sgl"
+    adapter_quant_scheme: str = "fp8-sgl"

    def to_dict(self) -> Dict[str, Any]:
        return {
-            "quant_op": self.quant_op,
            "dit_quant_scheme": self.dit_quant_scheme,
            "t5_quant_scheme": self.t5_quant_scheme,
            "clip_quant_scheme": self.clip_quant_scheme,