ericsorides
commited on
Commit
·
7f122be
1
Parent(s):
09a25ba
Model without scatter and with tree
Browse files- README.md +7 -1
- model.onnx +2 -2
README.md
CHANGED
@@ -99,8 +99,11 @@ def generate_text(model_path, prompt, tokenizer, max_gen_tokens, total_sequence,
|
|
99 |
inputs_dict = {}
|
100 |
inputs_dict['input_ids'] = actual_input[:, :window].reshape(1, window).numpy()
|
101 |
inputs_dict['attention_mask'] = first_attention
|
|
|
|
|
|
|
102 |
for name in inputs_names:
|
103 |
-
if name == 'input_ids' or name == 'attention_mask': continue
|
104 |
inputs_dict[name] = np.zeros([1, n_heads, context-window, 128], dtype="float16")
|
105 |
index = 0
|
106 |
new_token = np.array([10])
|
@@ -132,6 +135,9 @@ def generate_text(model_path, prompt, tokenizer, max_gen_tokens, total_sequence,
|
|
132 |
inputs_dict['input_ids']= total_input[:, j:next_index].reshape(1, window)
|
133 |
elif name == 'attention_mask':
|
134 |
inputs_dict['attention_mask'] = np.concatenate((np.zeros((1, total_sequence-next_index), dtype = 'int64'), np.ones((1, next_index), dtype = 'int64')), axis=1)
|
|
|
|
|
|
|
135 |
else:
|
136 |
old_name = name.replace("past_key_values", "present")
|
137 |
inputs_dict[name] = outs_dictionary[old_name][:, :, next_index-old_j:context-window+(next_index - old_j), :]
|
|
|
99 |
inputs_dict = {}
|
100 |
inputs_dict['input_ids'] = actual_input[:, :window].reshape(1, window).numpy()
|
101 |
inputs_dict['attention_mask'] = first_attention
|
102 |
+
index_pos = sum(first_attention[0])
|
103 |
+
inputs_dict['position_ids'] = np.concatenate((np.zeros([1, total_sequence - index_pos], dtype = 'int64'), np.arange(index_pos, dtype = 'int64').reshape(1, index_pos)), axis=1)
|
104 |
+
inputs_dict['tree_attention'] = np.triu(-65504*np.ones(total_sequence), k= 1).astype('float16').reshape(1, 1, total_sequence, total_sequence)
|
105 |
for name in inputs_names:
|
106 |
+
if name == 'input_ids' or name == 'attention_mask' or name == 'position_ids' or name == 'tree_attention': continue
|
107 |
inputs_dict[name] = np.zeros([1, n_heads, context-window, 128], dtype="float16")
|
108 |
index = 0
|
109 |
new_token = np.array([10])
|
|
|
135 |
inputs_dict['input_ids']= total_input[:, j:next_index].reshape(1, window)
|
136 |
elif name == 'attention_mask':
|
137 |
inputs_dict['attention_mask'] = np.concatenate((np.zeros((1, total_sequence-next_index), dtype = 'int64'), np.ones((1, next_index), dtype = 'int64')), axis=1)
|
138 |
+
elif name == 'position_ids':
|
139 |
+
inputs_dict['position_ids'] = np.concatenate((np.zeros([1, total_sequence - next_index], dtype = 'int64'), np.arange(next_index, dtype = 'int64').reshape(1, next_index)), axis=1)
|
140 |
+
elif name == 'tree_attention': continue
|
141 |
else:
|
142 |
old_name = name.replace("past_key_values", "present")
|
143 |
inputs_dict[name] = outs_dictionary[old_name][:, :, next_index-old_j:context-window+(next_index - old_j), :]
|
model.onnx
CHANGED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:9fb4dc57c626c6f6164d2d98e2f15de289093ddc0f74fc006726ef2fb856a0bd
|
3 |
+
size 1934549
|