in server/text_generation_server/layers/gptq/exllama.py [0:0]
def __init__(self, weight: GPTQWeight, bias):
super().__init__()
global MAX_DQ, MAX_INNER, ACT_ORDER, DEVICE
assert weight.bits == 4
self.device = weight.qweight.device
self.qweight = weight.qweight
self.qzeros = weight.qzeros
self.scales = weight.scales
self.g_idx = weight.g_idx.cpu() if weight.g_idx is not None else None
self.bias = bias if bias is not None else None
if self.g_idx is not None and (
(self.g_idx == 0).all()
or torch.equal(
weight.g_idx.cpu(),
torch.tensor(
[i // weight.groupsize for i in range(weight.g_idx.shape[0])],
dtype=torch.int32,
),
)
):
self.empty_g_idx = True
self.g_idx = None
assert self.device.type == "cuda"
assert self.device.index is not None
self.q4 = ext_make_q4(
self.qweight, self.qzeros, self.scales, self.g_idx, self.device.index
)
self.height = weight.qweight.shape[0] * 8
self.width = weight.qweight.shape[1]
# Infer groupsize from height of qzeros
self.groupsize = None
if self.qzeros.shape[0] > 1:
self.groupsize = (self.qweight.shape[0] * 8) // (self.qzeros.shape[0])
if self.groupsize is not None:
assert weight.groupsize == self.groupsize
# Handle act-order matrix
if self.g_idx is not None:
if self.groupsize is None:
raise ValueError("Found group index but no groupsize. What do?")
self.act_order = True
else:
self.act_order = False
DEVICE = self.qweight.device
MAX_DQ = max(MAX_DQ, self.qweight.numel() * 8)
if self.act_order:
MAX_INNER = max(MAX_INNER, self.height, self.width)
ACT_ORDER = True