mirror of
https://github.com/comfyanonymous/ComfyUI.git
synced 2026-07-18 12:28:17 +08:00
252 lines
8.2 KiB
Python
252 lines
8.2 KiB
Python
"""
|
||
TINT4Linear — INT4 weight-only Linear with LoRA + QuaRot.
|
||
|
||
Backed by torchao Int4PlainInt32Tensor. Activations stay FP16.
|
||
LoRA entries are injected at forward time — no kernel modification needed.
|
||
QuaRot activation rotation is applied online; LoRA A/w2 matrices are
|
||
rotated into the QuaRot basis inside forward() automatically unless
|
||
_lora_prerotated is set (for callers that pre-rotate at injection time).
|
||
"""
|
||
|
||
import torch
|
||
import torch.nn as nn
|
||
import torch.nn.functional as F
|
||
from torchao.quantization.quantize_.workflows.int4.int4_plain_int32_tensor import (
|
||
Int4PlainInt32Tensor,
|
||
)
|
||
|
||
|
||
class TINT4Linear(nn.Module):
|
||
"""
|
||
INT4 weight-only linear layer.
|
||
|
||
Stores raw int32 qdata + fp16 scales + int8 zero-points.
|
||
Int4PlainInt32Tensor is lazily constructed and rebuilt on device change.
|
||
|
||
Slots (set externally by loaders):
|
||
_use_quarot: bool — enable activation Hadamard rotation
|
||
_group_size: int — QuaRot group size
|
||
_hadamard_H: Tensor|None — normalized Hadamard matrix
|
||
_tint4_lora_entries: dict|None — {lora_name: [(type, ...), ...]}
|
||
_lora_prerotated: bool — True if caller already rotated A/w2
|
||
into QuaRot basis (default False).
|
||
When False, rotation is done in forward().
|
||
"""
|
||
|
||
def __init__(self, in_features: int, out_features: int,
|
||
qdata: torch.Tensor, scale: torch.Tensor,
|
||
zp: torch.Tensor, block_size: tuple,
|
||
bias: torch.Tensor | None = None):
|
||
"""
|
||
Initialize TINT4Linear from raw torchao INT4 tensors.
|
||
|
||
Args:
|
||
in_features: Input feature dimension.
|
||
out_features: Output feature dimension.
|
||
qdata: Packed int32 quantized weight data (out_f, in_f // 8).
|
||
scale: Per-block fp16 scales (n_blocks, out_f).
|
||
zp: Per-block int8 zero-points (n_blocks, out_f).
|
||
block_size: (b0, b1) tuple defining quantization block shape.
|
||
bias: Optional bias tensor (out_f,).
|
||
"""
|
||
super().__init__()
|
||
self.in_features = in_features
|
||
self.out_features = out_features
|
||
|
||
if bias is not None:
|
||
self.bias = nn.Parameter(bias)
|
||
else:
|
||
self.register_parameter('bias', None)
|
||
|
||
# ★ FIX 1: 使用 register_buffer 而不是直接用 = 赋值
|
||
# 好处：model.to(device) / state_dict() / load_state_dict() 能自动管理
|
||
# 旧代码：self._qdata = qdata → PyTorch 不知道这是持久化数据
|
||
self.register_buffer('_qdata', qdata)
|
||
self.register_buffer('_scale', scale)
|
||
self.register_buffer('_zp', zp)
|
||
self._block_size = block_size
|
||
self._qt: Int4PlainInt32Tensor | None = None
|
||
|
||
# QuaRot (optional)
|
||
self._use_quarot: bool = False
|
||
self._group_size: int = 128
|
||
self._hadamard_H: torch.Tensor | None = None
|
||
|
||
# LoRA entries slot
|
||
self._tint4_lora_entries: dict | None = None
|
||
|
||
# Pre-rotation flag
|
||
self._lora_prerotated: bool = False
|
||
|
||
@property
|
||
def weight(self) -> Int4PlainInt32Tensor:
|
||
"""Lazily construct and return the Int4PlainInt32Tensor view."""
|
||
if self._qt is None:
|
||
self._qt = Int4PlainInt32Tensor(
|
||
self._qdata, self._scale, self._zp,
|
||
self._block_size, [self.out_features, self.in_features],
|
||
)
|
||
return self._qt
|
||
|
||
@weight.setter
|
||
def weight(self, value: Int4PlainInt32Tensor) -> None:
|
||
"""Replace the underlying Int4PlainInt32Tensor."""
|
||
if isinstance(value, Int4PlainInt32Tensor):
|
||
self._qt = value
|
||
|
||
def _rotate_into_quarot(self, tensor: torch.Tensor) -> torch.Tensor:
|
||
"""
|
||
Rotate a LoRA A or w2 matrix into the QuaRot Hadamard basis.
|
||
|
||
Skipped when _lora_prerotated is True (caller already rotated).
|
||
Returns unchanged tensor if QuaRot is disabled or dimensions
|
||
are not divisible by _group_size.
|
||
"""
|
||
if self._lora_prerotated:
|
||
return tensor
|
||
if not self._use_quarot or self._hadamard_H is None:
|
||
return tensor
|
||
if tensor.shape[1] % self._group_size != 0:
|
||
return tensor
|
||
Ht = self._hadamard_H.T.to(device=tensor.device, dtype=tensor.dtype)
|
||
ng = tensor.shape[1] // self._group_size
|
||
return (tensor.reshape(tensor.shape[0], ng, self._group_size)
|
||
@ Ht).reshape(tensor.shape[0], -1)
|
||
|
||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||
"""
|
||
Forward pass with optional QuaRot activation rotation and LoRA injection.
|
||
|
||
Args:
|
||
x: Input tensor, shape (..., in_features).
|
||
|
||
Returns:
|
||
Output tensor, shape (..., out_features).
|
||
|
||
Raises:
|
||
ValueError: If activation dim is not divisible by QuaRot group_size,
|
||
or if LoRA output shapes are incompatible.
|
||
"""
|
||
x_flat = x.reshape(-1, x.shape[-1])
|
||
|
||
# QuaRot activation rotation (online)
|
||
if self._use_quarot and self._hadamard_H is not None:
|
||
if x_flat.shape[-1] % self._group_size != 0:
|
||
raise ValueError(
|
||
f"activation dim {x_flat.shape[-1]} not divisible "
|
||
f"by QuaRot group_size {self._group_size}"
|
||
)
|
||
H_dev = self._hadamard_H.to(device=x.device, dtype=x.dtype)
|
||
n_groups = x_flat.shape[-1] // self._group_size
|
||
x_g = x_flat.view(-1, n_groups, self._group_size)
|
||
x_flat = torch.matmul(x_g, H_dev).view(-1, n_groups * self._group_size)
|
||
|
||
dev = x.device
|
||
|
||
# Rebuild Int4PlainInt32Tensor on device change
|
||
if self._qt is None or self._qt.device != dev:
|
||
self._qt = Int4PlainInt32Tensor(
|
||
self._qdata.to(dev), self._scale.to(dev), self._zp.to(dev),
|
||
self._block_size, [self.out_features, self.in_features],
|
||
)
|
||
|
||
out = F.linear(x_flat, self._qt, None)
|
||
|
||
# LoRA forward injection
|
||
entries = self._tint4_lora_entries
|
||
if entries is not None and entries:
|
||
cd = (x.dtype if x.dtype in (torch.float16, torch.bfloat16)
|
||
else torch.float16)
|
||
for lora_entries in entries.values():
|
||
for e in lora_entries:
|
||
# ── LoKr path ──
|
||
# ★ FIX 2: 用 torch.kron 替代 repeat_interleave
|
||
# 旧方案要求 w2 维度能被 factor 整除（例如 64÷60=不整除→报错）
|
||
# 新方案 kron 展开 + pad/trim 到目标维度，兼容任意分解
|
||
if isinstance(e[0], str) and e[0] == "lokr":
|
||
_, w1, w2, mult, factor = e[:5]
|
||
sl = e[5] if len(e) > 5 else None
|
||
se = e[6] if len(e) > 6 else None
|
||
|
||
# QuaRot: 先旋转 w2（小矩阵，转发负担小）
|
||
# 数学上等价于旋转完整 kron 展开（Hadamard 是分块对角的）
|
||
w2 = self._rotate_into_quarot(w2)
|
||
w1d = w1.to(device=dev, dtype=cd)
|
||
w2d = w2.to(device=dev, dtype=cd)
|
||
|
||
# kron 展开为完整 delta（代替 repeat_interleave）
|
||
dw = torch.kron(w1d, w2d)
|
||
|
||
# Pad/trim to layer (or slice) dimensions
|
||
tgt_out = (se - sl) if sl is not None else self.out_features
|
||
if dw.shape[0] < tgt_out:
|
||
dw = dw.repeat(
|
||
(tgt_out + dw.shape[0] - 1) // dw.shape[0], 1)
|
||
if dw.shape[0] > tgt_out:
|
||
dw = dw[:tgt_out, :]
|
||
if dw.shape[1] < self.in_features:
|
||
dw = dw.repeat(
|
||
1, (self.in_features + dw.shape[1] - 1) // dw.shape[1])
|
||
if dw.shape[1] > self.in_features:
|
||
dw = dw[:, :self.in_features]
|
||
|
||
dw = dw.contiguous().mul_(mult)
|
||
lo = x_flat.to(cd) @ dw.T
|
||
if sl is not None:
|
||
if lo.shape[1] != (se - sl):
|
||
raise ValueError(
|
||
f"LoKr output width {lo.shape[1]} "
|
||
f"doesn't match target slice width "
|
||
f"{se - sl}"
|
||
)
|
||
out[:, sl:se] += lo
|
||
elif lo.shape == out.shape:
|
||
out += lo
|
||
else:
|
||
raise ValueError(
|
||
f"LoKr output shape {lo.shape} "
|
||
f"incompatible with {out.shape}"
|
||
)
|
||
continue
|
||
|
||
# Standard LoRA path
|
||
A, B, mult = e[:3]
|
||
sl = e[3] if len(e) > 3 else None
|
||
se = e[4] if len(e) > 4 else None
|
||
A = self._rotate_into_quarot(A)
|
||
Ad = A.to(device=dev, dtype=cd)
|
||
Bd = B.to(device=dev, dtype=cd)
|
||
lo = (x_flat.to(cd) @ Ad.T) @ Bd.T * mult
|
||
if sl is not None:
|
||
if lo.shape[1] != (se - sl):
|
||
raise ValueError(
|
||
f"LoRA output width {lo.shape[1]} "
|
||
f"doesn't match target slice width "
|
||
f"{se - sl}"
|
||
)
|
||
out[:, sl:se] += lo
|
||
elif lo.shape[1] == out.shape[1]:
|
||
out += lo
|
||
else:
|
||
raise ValueError(
|
||
f"LoRA output shape {lo.shape} "
|
||
f"incompatible with {out.shape}"
|
||
)
|
||
|
||
if self.bias is not None:
|
||
out += self.bias.to(device=dev, dtype=out.dtype)
|
||
|
||
return out.reshape(*x.shape[:-1], out.shape[-1])
|
||
|
||
def release(self) -> None:
|
||
"""Drop cached Int4PlainInt32Tensor to free VRAM."""
|
||
self._qt = None
|
||
|
||
def __del__(self):
|
||
"""Cleanup raw tensor references on deletion."""
|
||
self._qdata = None
|
||
self._scale = None
|
||
self._zp = None
|
||
self._qt = None
|
||
self._tint4_lora_entries = None
|