Files
Xiaoze Fan 477c8601b2 refactor(quant): config, scheme and method layers for quantization (#418)
* refactor(quant): config, scheme and method layers for checkpoint quantization

QuantConfig parses the checkpoint's quantization_config into a QuantScheme per
projection; every Linear*, ParallelLMHead and MoE layer takes quant_config and a
prefix, declares its tensors through a QuantMethod and selects its kernel at
construction. MoE kernels own the expert bank layout, pack and apply, so the
offload cache, the FTW converter and the CPU executor follow one layout.

- --quant-backend layer[.kind]=name selects kernels (--nvfp4-backend forwards to it)
- --moe-strategy replaces --moe-backend; the old spelling stays as an alias
- --moe-cpu-layers auto for Windows/WSL pin budgets; --moe-cache-auto respects the kernel's slot limit
- removes the per-family quantized layer classes, bank providers, format env switches and GLM runtime fp8
2026-09-08 23:37:23 -07:00

65 lines
2.0 KiB
Python

"""The linear side: per-layer config, the kernel ABC and the Method base."""
from __future__ import annotations
from abc import ABC, abstractmethod
from dataclasses import dataclass
from typing import Any, ClassVar
import torch
from ..method import QuantMethod
from ..scheme import QuantScheme
@dataclass(frozen=True)
class LinearConfig:
"""What a linear kernel may look at to accept or decline a layer: the TP-local shape and the weight scheme."""
in_features: int
out_features: int
# fused segments along N; a plain linear is one segment
output_sizes: tuple[int, ...] = ()
scheme: QuantScheme | None = None
def __post_init__(self) -> None:
if not self.output_sizes:
object.__setattr__(self, "output_sizes", (self.out_features,))
if sum(self.output_sizes) != self.out_features:
raise ValueError(f"output_sizes {self.output_sizes} do not sum to {self.out_features}")
@classmethod
def from_layer(cls, layer: Any, scheme: QuantScheme | None) -> "LinearConfig":
return cls(layer.in_features, layer.out_features, tuple(layer.output_sizes), scheme)
class LinearKernel(ABC):
"""One backend for one linear kind: says whether it can run a layer, then finalizes and applies it."""
name: ClassVar[str]
def unusable_reason(self, cfg: LinearConfig) -> str | None:
return None
def worth_it(self, cfg: LinearConfig) -> bool:
return True
def finalize(self, layer: Any) -> None:
pass
@abstractmethod
def apply(self, layer: Any, x: torch.Tensor) -> torch.Tensor: ...
class LinearMethod(QuantMethod):
"""Declares a linear layer's weights for its kind; finalize and apply go to the kernel."""
@abstractmethod
def create_weights(self, layer: Any) -> None: ...
def finalize(self, layer: Any) -> None:
self.kernel.finalize(layer)
def apply(self, layer: Any, x: torch.Tensor) -> torch.Tensor:
return self.kernel.apply(layer, x)