Fix backward for quantization
Browse files- modeling_chatglm.py +6 -8
- quantization.py +2 -2
modeling_chatglm.py
CHANGED
@@ -134,11 +134,11 @@ def load_tf_weights_in_chatglm_6b(model, config, tf_checkpoint_path):
|
|
134 |
|
135 |
|
136 |
class PrefixEncoder(torch.nn.Module):
|
137 |
-
|
138 |
The torch.nn model to encode the prefix
|
139 |
Input shape: (batch-size, prefix-length)
|
140 |
Output shape: (batch-size, prefix-length, 2*layers*hidden)
|
141 |
-
|
142 |
def __init__(self, config):
|
143 |
super().__init__()
|
144 |
self.prefix_projection = config.prefix_projection
|
@@ -148,7 +148,7 @@ class PrefixEncoder(torch.nn.Module):
|
|
148 |
self.trans = torch.nn.Sequential(
|
149 |
torch.nn.Linear(config.hidden_size, config.hidden_size),
|
150 |
torch.nn.Tanh(),
|
151 |
-
torch.nn.Linear(config.hidden_size, config.num_layers
|
152 |
)
|
153 |
else:
|
154 |
self.embedding = torch.nn.Embedding(config.pre_seq_len, config.num_layers * config.hidden_size * 2)
|
@@ -814,7 +814,7 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
|
|
814 |
self.num_attention_heads,
|
815 |
self.hidden_size // self.num_attention_heads
|
816 |
)
|
817 |
-
#seq_len, b, nh, hidden_size
|
818 |
past_key_values = self.dropout(past_key_values)
|
819 |
past_key_values = past_key_values.permute([2, 1, 0, 3, 4]).split(2)
|
820 |
# past_key_values = [(v[0], v[1]) for v in past_key_values]
|
@@ -909,7 +909,8 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
|
|
909 |
)
|
910 |
|
911 |
if self.pre_seq_len is not None:
|
912 |
-
prefix_attention_mask = torch.ones(batch_size, 1, input_ids.size(-1), self.pre_seq_len).to(
|
|
|
913 |
prefix_attention_mask = (prefix_attention_mask < 0.5).bool()
|
914 |
attention_mask = torch.cat((prefix_attention_mask, attention_mask), dim=3)
|
915 |
|
@@ -942,9 +943,6 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
|
|
942 |
else:
|
943 |
attention_mask = attention_mask.to(input_ids.device)
|
944 |
|
945 |
-
if self.training:
|
946 |
-
hidden_states = hidden_states.requires_grad_(True)
|
947 |
-
|
948 |
for i, layer in enumerate(self.layers):
|
949 |
|
950 |
if output_hidden_states:
|
|
|
134 |
|
135 |
|
136 |
class PrefixEncoder(torch.nn.Module):
|
137 |
+
"""
|
138 |
The torch.nn model to encode the prefix
|
139 |
Input shape: (batch-size, prefix-length)
|
140 |
Output shape: (batch-size, prefix-length, 2*layers*hidden)
|
141 |
+
"""
|
142 |
def __init__(self, config):
|
143 |
super().__init__()
|
144 |
self.prefix_projection = config.prefix_projection
|
|
|
148 |
self.trans = torch.nn.Sequential(
|
149 |
torch.nn.Linear(config.hidden_size, config.hidden_size),
|
150 |
torch.nn.Tanh(),
|
151 |
+
torch.nn.Linear(config.hidden_size, config.num_layers * config.hidden_size * 2)
|
152 |
)
|
153 |
else:
|
154 |
self.embedding = torch.nn.Embedding(config.pre_seq_len, config.num_layers * config.hidden_size * 2)
|
|
|
814 |
self.num_attention_heads,
|
815 |
self.hidden_size // self.num_attention_heads
|
816 |
)
|
817 |
+
# seq_len, b, nh, hidden_size
|
818 |
past_key_values = self.dropout(past_key_values)
|
819 |
past_key_values = past_key_values.permute([2, 1, 0, 3, 4]).split(2)
|
820 |
# past_key_values = [(v[0], v[1]) for v in past_key_values]
|
|
|
909 |
)
|
910 |
|
911 |
if self.pre_seq_len is not None:
|
912 |
+
prefix_attention_mask = torch.ones(batch_size, 1, input_ids.size(-1), self.pre_seq_len).to(
|
913 |
+
attention_mask.device)
|
914 |
prefix_attention_mask = (prefix_attention_mask < 0.5).bool()
|
915 |
attention_mask = torch.cat((prefix_attention_mask, attention_mask), dim=3)
|
916 |
|
|
|
943 |
else:
|
944 |
attention_mask = attention_mask.to(input_ids.device)
|
945 |
|
|
|
|
|
|
|
946 |
for i, layer in enumerate(self.layers):
|
947 |
|
948 |
if output_hidden_states:
|
quantization.py
CHANGED
@@ -14,11 +14,11 @@ class W8A16Linear(torch.autograd.Function):
|
|
14 |
@staticmethod
|
15 |
def forward(ctx, inp: torch.Tensor, quant_w: torch.Tensor, scale_w: torch.Tensor, weight_bit_width):
|
16 |
ctx.inp_shape = inp.size()
|
17 |
-
ctx.weight_shape = quant_w.size()
|
18 |
ctx.weight_bit_width = weight_bit_width
|
19 |
out_features = quant_w.size(0)
|
20 |
inp = inp.contiguous().view(-1, inp.size(-1))
|
21 |
weight = extract_weight_to_half(quant_w, scale_w, weight_bit_width)
|
|
|
22 |
output = inp.mm(weight.t())
|
23 |
ctx.save_for_backward(inp, quant_w, scale_w)
|
24 |
return output.view(*(ctx.inp_shape[:-1] + (out_features,)))
|
@@ -30,7 +30,7 @@ class W8A16Linear(torch.autograd.Function):
|
|
30 |
grad_output = grad_output.contiguous().view(-1, weight.size(0))
|
31 |
grad_input = grad_output.mm(weight)
|
32 |
grad_weight = grad_output.t().mm(inp)
|
33 |
-
return grad_input.view(ctx.inp_shape), grad_weight.view(ctx.weight_shape), None
|
34 |
|
35 |
|
36 |
class Kernel:
|
|
|
14 |
@staticmethod
|
15 |
def forward(ctx, inp: torch.Tensor, quant_w: torch.Tensor, scale_w: torch.Tensor, weight_bit_width):
|
16 |
ctx.inp_shape = inp.size()
|
|
|
17 |
ctx.weight_bit_width = weight_bit_width
|
18 |
out_features = quant_w.size(0)
|
19 |
inp = inp.contiguous().view(-1, inp.size(-1))
|
20 |
weight = extract_weight_to_half(quant_w, scale_w, weight_bit_width)
|
21 |
+
ctx.weight_shape = weight.size()
|
22 |
output = inp.mm(weight.t())
|
23 |
ctx.save_for_backward(inp, quant_w, scale_w)
|
24 |
return output.view(*(ctx.inp_shape[:-1] + (out_features,)))
|
|
|
30 |
grad_output = grad_output.contiguous().view(-1, weight.size(0))
|
31 |
grad_input = grad_output.mm(weight)
|
32 |
grad_weight = grad_output.t().mm(inp)
|
33 |
+
return grad_input.view(ctx.inp_shape), grad_weight.view(ctx.weight_shape), None, None
|
34 |
|
35 |
|
36 |
class Kernel:
|