| 1 | from dataclasses import dataclass |
| 2 | |
| 3 | from ltx_core.loader.module_ops import ModuleOps |
| 4 | from ltx_core.loader.sd_ops import SDOps |
| 5 | from ltx_core.quantization.fp8_cast import TRANSFORMER_LINEAR_DOWNCAST_MAP, UPCAST_DURING_INFERENCE |
| 6 | from ltx_core.quantization.fp8_scaled_mm import FP8_PREPARE_MODULE_OPS, FP8_TRANSPOSE_SD_OPS |
| 7 | |
| 8 | |
| 9 | @dataclass(frozen=True) |
| 10 | class QuantizationPolicy: |
| 11 | """Configuration for model quantization during loading. |
| 12 | Attributes: |
| 13 | sd_ops: State dict operations for weight transformation. |
| 14 | module_ops: Post-load module transformations. |
| 15 | """ |
| 16 | |
| 17 | sd_ops: SDOps | None = None |
| 18 | module_ops: tuple[ModuleOps, ...] = () |
| 19 | |
| 20 | @classmethod |
| 21 | def fp8_cast(cls) -> "QuantizationPolicy": |
| 22 | """Create policy using FP8 casting with upcasting during inference.""" |
| 23 | return cls( |
| 24 | sd_ops=TRANSFORMER_LINEAR_DOWNCAST_MAP, |
| 25 | module_ops=(UPCAST_DURING_INFERENCE,), |
| 26 | ) |
| 27 | |
| 28 | @classmethod |
| 29 | def fp8_scaled_mm(cls) -> "QuantizationPolicy": |
| 30 | """Create policy using FP8 scaled matrix multiplication.""" |
| 31 | try: |
| 32 | import tensorrt_llm # noqa: F401, PLC0415 |
| 33 | except ImportError as e: |
| 34 | raise ImportError("tensorrt_llm is not installed, skipping FP8 scaled MM quantization") from e |
| 35 | |
| 36 | return cls( |
| 37 | sd_ops=FP8_TRANSPOSE_SD_OPS, |
| 38 | module_ops=(FP8_PREPARE_MODULE_OPS,), |
| 39 | ) |
| 40 |