nodes/document97-Comfy-INT4-HIPtest
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Fix Dynamic LoRA shape handling

Browse Source
This commit is contained in:
document97
2026-07-13 16:57:19 +08:00
parent 6ba3f1a1eb
commit a6cdaf10e9
1 changed files with 20 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
int4_quant.py
+20 -12
View File
@@ -560,19 +560,22 @@ class Int4Ops(manual_cast):
if self.lora_patches:
x_shape = x.shape
y_shape = y.shape
x_2d = x.reshape(-1, x_shape[-1])
y_2d = y.reshape(-1, y_shape[-1])
for lora_down, lora_up, lora_start, lora_size in self.lora_patches:
lD = lora_down.to(x.device, non_blocking=True)
lU = lora_up.to(x.device, non_blocking=True)
lora_x = F.linear(x_2d.to(lD.dtype), lD)
lora_y = F.linear(lora_x, lU)
lora_y = F.linear(lora_x, lU).to(y_2d.dtype)
if lora_start is not None:
y[:, lora_start:lora_start + lora_size] = (
y[:, lora_start:lora_start + lora_size] + lora_y.to(y.dtype)
y_2d[:, lora_start:lora_start + lora_size] = (
y_2d[:, lora_start:lora_start + lora_size] + lora_y
)
else:
y = y + lora_y.to(y.dtype)
y_2d = y_2d + lora_y
y = y_2d.reshape(y_shape)
return y
class GroupNorm(manual_cast.GroupNorm): pass
@@ -808,13 +811,18 @@ class INT4ModelPatcher(comfy.model_patcher.ModelPatcher):
if mid is not None:
down_scaled = torch.mm(mid.flatten(1), down.flatten(1)) * scale
if getattr(module, "_convrot_groupsize", None) is not None and down_scaled.shape[1] % module._convrot_groupsize == 0:
try:
from .convrot import build_hadamard, rotate_weight
H = build_hadamard(module._convrot_groupsize, device=down_scaled.device, dtype=down_scaled.dtype)
down_scaled = rotate_weight(down_scaled, H, group_size=module._convrot_groupsize)
except ImportError:
pass
if (
getattr(module, "_quant_format", "convrot_w4a4") == "int8_tensorwise"
and getattr(module, "_use_convrot", False)
):
group_size = getattr(module, "_convrot_groupsize", 256)
if down_scaled.shape[1] % group_size == 0:
try:
from .convrot import build_hadamard, rotate_weight
H = build_hadamard(group_size, device=down_scaled.device, dtype=down_scaled.dtype)
down_scaled = rotate_weight(down_scaled, H, group_size=group_size)
except ImportError:
pass
start, size = None, None
if offset is not None: