Plasmarine
commited on
Commit
•
fc42df6
1
Parent(s):
d6b992d
Linformer
Browse files- modeling_cocom.py +23 -49
modeling_cocom.py
CHANGED
@@ -1,4 +1,4 @@
|
|
1 |
-
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, PreTrainedModel, PretrainedConfig, AutoModel
|
2 |
import torch
|
3 |
import math
|
4 |
from peft import get_peft_model, LoraConfig, TaskType
|
@@ -71,8 +71,7 @@ class COCOMConfig(PretrainedConfig):
|
|
71 |
lora = False,
|
72 |
training_form="both",
|
73 |
lora_r=16,
|
74 |
-
attn_implementation="
|
75 |
-
attention_window=512,
|
76 |
device_map = "cuda",
|
77 |
**kwargs):
|
78 |
super().__init__(**kwargs)
|
@@ -96,28 +95,6 @@ class COCOM(PreTrainedModel):
|
|
96 |
super().__init__(cfg)
|
97 |
# define models
|
98 |
attn_impl = cfg.attn_implementation
|
99 |
-
|
100 |
-
if cfg.attn_implementation == "longformer":
|
101 |
-
# Initialize Longformer
|
102 |
-
longformer_config = LongformerConfig.from_pretrained(cfg.decoder_model_name)
|
103 |
-
longformer_config.attention_window = 512 # Modify based on context window size
|
104 |
-
self.decoder = LongformerForCausalLM.from_pretrained(
|
105 |
-
cfg.decoder_model_name,
|
106 |
-
config=longformer_config,
|
107 |
-
torch_dtype=torch.float16,
|
108 |
-
low_cpu_mem_usage=True,
|
109 |
-
device_map=cfg.device_map
|
110 |
-
)
|
111 |
-
else:
|
112 |
-
# Original decoder initialization
|
113 |
-
self.decoder = AutoModelForCausalLM.from_pretrained(
|
114 |
-
cfg.decoder_model_name,
|
115 |
-
torch_dtype=torch.float16,
|
116 |
-
attn_implementation=attn_impl,
|
117 |
-
low_cpu_mem_usage=True,
|
118 |
-
device_map=cfg.device_map
|
119 |
-
)
|
120 |
-
|
121 |
# model could be loaded in three quantization modes: no, int4, int8
|
122 |
if cfg.quantization == "no":
|
123 |
self.decoder = AutoModelForCausalLM.from_pretrained(
|
@@ -216,20 +193,15 @@ class COCOM(PreTrainedModel):
|
|
216 |
self.compr_rate = cfg.compr_rate
|
217 |
self.local_rank = os.getenv('LOCAL_RANK', '0')
|
218 |
|
219 |
-
def compress_and_replace_emb(self, enc_input_ids, enc_attention_mask, dec_input_ids
|
220 |
indices = range(0, enc_input_ids.size(0) + 1, self.generation_top_k)
|
221 |
-
|
222 |
-
# Perform compression
|
223 |
if self.compr:
|
224 |
compressed_embs = self.compr(enc_input_ids, enc_attention_mask)
|
|
|
225 |
else:
|
226 |
compressed_embs = self.compr_decoder(enc_input_ids, enc_attention_mask)
|
227 |
-
|
228 |
-
# Replace embeddings with compressed ones
|
229 |
-
input_embeds = self.replace_embeddings(compressed_embs, dec_input_ids, dec_attention_mask, indices)
|
230 |
-
|
231 |
return input_embeds
|
232 |
-
|
233 |
|
234 |
def compr_decoder(self, input_ids, attention_mask):
|
235 |
emb = self.decoder(input_ids=input_ids, attention_mask=attention_mask, output_hidden_states=True).hidden_states[-1]
|
@@ -240,23 +212,19 @@ class COCOM(PreTrainedModel):
|
|
240 |
def replace_embeddings(self, compressed_embs, dec_input_ids, indices):
|
241 |
# Embed the decoder input
|
242 |
inputs_embeds = self.decoder.get_input_embeddings()(dec_input_ids)
|
243 |
-
|
244 |
-
# Number of compressed embeddings
|
245 |
num_embs = compressed_embs.size(1)
|
246 |
-
|
247 |
-
|
248 |
-
|
249 |
-
|
250 |
-
#
|
251 |
first_mem_token_indices = torch.argmax((dec_input_ids == self.decoder_tokenizer.mem_token_id).int(), dim=1)
|
252 |
batch_size = inputs_embeds.size(0)
|
253 |
-
|
254 |
-
# Replace memory tokens with compressed embeddings
|
255 |
for i in range(batch_size):
|
256 |
for j in range(indices[i], indices[i + 1]):
|
257 |
-
start_idx = first_mem_token_indices[i].item() + (j
|
258 |
inputs_embeds[i, start_idx:start_idx + num_embs, :] = compressed_embs[j]
|
259 |
-
|
260 |
return inputs_embeds
|
261 |
|
262 |
|
@@ -267,13 +235,19 @@ class COCOM(PreTrainedModel):
|
|
267 |
dec_attention_mask: torch.LongTensor = None,
|
268 |
labels: torch.LongTensor = None):
|
269 |
|
270 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
271 |
|
272 |
-
#
|
273 |
if (self.training_form == "compressor") and (self.compr is None):
|
274 |
-
inputs_embeds
|
275 |
|
276 |
-
#
|
277 |
decoder_outputs = self.decoder(inputs_embeds=inputs_embeds, attention_mask=dec_attention_mask, labels=labels)
|
278 |
|
279 |
return {"loss": decoder_outputs.loss, "logits": decoder_outputs.logits}
|
@@ -289,7 +263,7 @@ class COCOM(PreTrainedModel):
|
|
289 |
attention_mask=dec_attention_mask.to(device),
|
290 |
do_sample=False,
|
291 |
top_p=None,
|
292 |
-
max_new_tokens=
|
293 |
)
|
294 |
decoded = self.decoder_tokenizer.batch_decode(output_ids, skip_special_tokens=True)
|
295 |
return decoded
|
|
|
1 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, PreTrainedModel, PretrainedConfig, AutoModel
|
2 |
import torch
|
3 |
import math
|
4 |
from peft import get_peft_model, LoraConfig, TaskType
|
|
|
71 |
lora = False,
|
72 |
training_form="both",
|
73 |
lora_r=16,
|
74 |
+
attn_implementation="eager",
|
|
|
75 |
device_map = "cuda",
|
76 |
**kwargs):
|
77 |
super().__init__(**kwargs)
|
|
|
95 |
super().__init__(cfg)
|
96 |
# define models
|
97 |
attn_impl = cfg.attn_implementation
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
98 |
# model could be loaded in three quantization modes: no, int4, int8
|
99 |
if cfg.quantization == "no":
|
100 |
self.decoder = AutoModelForCausalLM.from_pretrained(
|
|
|
193 |
self.compr_rate = cfg.compr_rate
|
194 |
self.local_rank = os.getenv('LOCAL_RANK', '0')
|
195 |
|
196 |
+
def compress_and_replace_emb(self, enc_input_ids, enc_attention_mask, dec_input_ids):
|
197 |
indices = range(0, enc_input_ids.size(0) + 1, self.generation_top_k)
|
|
|
|
|
198 |
if self.compr:
|
199 |
compressed_embs = self.compr(enc_input_ids, enc_attention_mask)
|
200 |
+
input_embeds = self.replace_embeddings(compressed_embs, dec_input_ids, indices)
|
201 |
else:
|
202 |
compressed_embs = self.compr_decoder(enc_input_ids, enc_attention_mask)
|
203 |
+
input_embeds = self.replace_embeddings(compressed_embs, dec_input_ids, indices)
|
|
|
|
|
|
|
204 |
return input_embeds
|
|
|
205 |
|
206 |
def compr_decoder(self, input_ids, attention_mask):
|
207 |
emb = self.decoder(input_ids=input_ids, attention_mask=attention_mask, output_hidden_states=True).hidden_states[-1]
|
|
|
212 |
def replace_embeddings(self, compressed_embs, dec_input_ids, indices):
|
213 |
# Embed the decoder input
|
214 |
inputs_embeds = self.decoder.get_input_embeddings()(dec_input_ids)
|
|
|
|
|
215 |
num_embs = compressed_embs.size(1)
|
216 |
+
if self.sep:
|
217 |
+
slot_len = num_embs + 1
|
218 |
+
else:
|
219 |
+
slot_len = num_embs
|
220 |
+
# get first mem_token inidices
|
221 |
first_mem_token_indices = torch.argmax((dec_input_ids == self.decoder_tokenizer.mem_token_id).int(), dim=1)
|
222 |
batch_size = inputs_embeds.size(0)
|
223 |
+
# for each example in batch, replace them with compressed embeddings
|
|
|
224 |
for i in range(batch_size):
|
225 |
for j in range(indices[i], indices[i + 1]):
|
226 |
+
start_idx = first_mem_token_indices[i].item() + (j-indices[i]) * slot_len
|
227 |
inputs_embeds[i, start_idx:start_idx + num_embs, :] = compressed_embs[j]
|
|
|
228 |
return inputs_embeds
|
229 |
|
230 |
|
|
|
235 |
dec_attention_mask: torch.LongTensor = None,
|
236 |
labels: torch.LongTensor = None):
|
237 |
|
238 |
+
# enc_input_ids: stores the contexts, should be flattened from all queries before input, dimention (batch_size*generation_top_k, token_length)
|
239 |
+
# enc_attention_mask: attention mask of enc_input_ids
|
240 |
+
# dec_input_ids: stores the prompts (including mem tokens), dimention (batch_size, token_length)
|
241 |
+
# dec_attention_mask: attention mask of dec_input_ids
|
242 |
+
|
243 |
+
# Perform compression with gradient tracking
|
244 |
+
inputs_embeds = self.compress_and_replace_emb(enc_input_ids, enc_attention_mask, dec_input_ids)
|
245 |
|
246 |
+
# if training_form is compressor, then detach the inputs_embeds, to make gradient not count in decoder
|
247 |
if (self.training_form == "compressor") and (self.compr is None):
|
248 |
+
inputs_embeds = inputs_embeds.detach()
|
249 |
|
250 |
+
# decoding
|
251 |
decoder_outputs = self.decoder(inputs_embeds=inputs_embeds, attention_mask=dec_attention_mask, labels=labels)
|
252 |
|
253 |
return {"loss": decoder_outputs.loss, "logits": decoder_outputs.logits}
|
|
|
263 |
attention_mask=dec_attention_mask.to(device),
|
264 |
do_sample=False,
|
265 |
top_p=None,
|
266 |
+
max_new_tokens=max_new_tokens
|
267 |
)
|
268 |
decoded = self.decoder_tokenizer.batch_decode(output_ids, skip_special_tokens=True)
|
269 |
return decoded
|