Helw150
commited on
Commit
•
49f38f9
1
Parent(s):
b79a517
Add Batch Support
Browse files- modeling_diva.py +58 -26
- test.py +28 -0
modeling_diva.py
CHANGED
@@ -44,7 +44,7 @@ class WhisperConnector(nn.Module):
|
|
44 |
|
45 |
class DiVAModel(PreTrainedModel):
|
46 |
config_class = DiVAConfig
|
47 |
-
|
48 |
def __init__(
|
49 |
self, via_path=None, config_dict={}, device_map=None, speech_encoder_device=None
|
50 |
):
|
@@ -105,10 +105,9 @@ class DiVAModel(PreTrainedModel):
|
|
105 |
)
|
106 |
self.speech_encoder_device = speech_encoder_device
|
107 |
|
108 |
-
|
109 |
-
def can_generate(cls):
|
110 |
return False
|
111 |
-
|
112 |
@classmethod
|
113 |
def from_pretrained(
|
114 |
cls,
|
@@ -182,8 +181,14 @@ class DiVAModel(PreTrainedModel):
|
|
182 |
|
183 |
return outputs
|
184 |
|
|
|
185 |
def generate(
|
186 |
-
self,
|
|
|
|
|
|
|
|
|
|
|
187 |
):
|
188 |
inputs = self.processor(audio, return_tensors="pt", sampling_rate=16_000)
|
189 |
input_features = inputs.input_features.to(self.speech_encoder_device)
|
@@ -193,29 +198,45 @@ class DiVAModel(PreTrainedModel):
|
|
193 |
virt_tokens = self.connector(
|
194 |
hidden_states,
|
195 |
output_device=self.llama_decoder.model.embed_tokens.weight.device,
|
196 |
-
)
|
|
|
197 |
|
198 |
if text_prompt != None and text_prompt != "":
|
199 |
user_prompt_text = torch.tensor(
|
200 |
-
self.tokenizer(
|
|
|
|
|
|
|
|
|
|
|
201 |
device=self.pre_user_suffix.device,
|
202 |
)
|
203 |
prefix = torch.cat(
|
204 |
-
[
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
205 |
)
|
206 |
else:
|
207 |
prefix = self.prefix
|
208 |
-
prefix_embed = self.llama_decoder.model.embed_tokens(prefix)
|
209 |
suffix = self.final_header
|
210 |
-
suffix_embed = self.llama_decoder.model.embed_tokens(suffix)
|
211 |
-
inputs_embeds = torch.cat(
|
212 |
-
|
213 |
-
|
214 |
-
outs = []
|
215 |
outputs = None
|
216 |
greedy = 1
|
217 |
i = 0
|
218 |
-
while
|
219 |
past_key_values = outputs.past_key_values if outputs else None
|
220 |
outputs = self.llama_decoder(
|
221 |
inputs_embeds=inputs_embeds.to(
|
@@ -225,7 +246,7 @@ class DiVAModel(PreTrainedModel):
|
|
225 |
output_hidden_states=True,
|
226 |
past_key_values=past_key_values,
|
227 |
)
|
228 |
-
next_token_logits = outputs.logits[
|
229 |
|
230 |
if logits_processor:
|
231 |
local_outs = torch.tensor(outs) if outs != [] else suffix
|
@@ -240,16 +261,23 @@ class DiVAModel(PreTrainedModel):
|
|
240 |
probs = F.softmax(logits, dim=-1)
|
241 |
greedy = torch.multinomial(probs, num_samples=1)[0]
|
242 |
else:
|
243 |
-
greedy = next_token_logits.argmax()
|
244 |
-
|
245 |
-
|
|
|
|
|
|
|
|
|
246 |
inputs_embeds = next_embed
|
247 |
-
return self.tokenizer.
|
248 |
-
"<|eot_id|>", ""
|
249 |
-
)
|
250 |
|
251 |
def generate_stream(
|
252 |
-
self,
|
|
|
|
|
|
|
|
|
|
|
253 |
):
|
254 |
inputs = self.processor(audio, return_tensors="pt", sampling_rate=16_000)
|
255 |
input_features = inputs.input_features.to(self.whisper_encoder.device)
|
@@ -284,7 +312,7 @@ class DiVAModel(PreTrainedModel):
|
|
284 |
while greedy != 128009 and len(outs) < max_new_tokens:
|
285 |
past_key_values = outputs.past_key_values if outputs else None
|
286 |
outputs = self.llama_decoder(
|
287 |
-
|
288 |
self.llama_decoder.model.embed_tokens.weight.device
|
289 |
).half(),
|
290 |
return_dict=True,
|
@@ -310,5 +338,9 @@ class DiVAModel(PreTrainedModel):
|
|
310 |
outs.append(greedy)
|
311 |
next_embed = self.llama_decoder.model.embed_tokens(greedy.reshape(1, 1))
|
312 |
inputs_embeds = next_embed
|
313 |
-
yield self.tokenizer.decode(outs, skip_special_tokens=True).replace(
|
314 |
-
|
|
|
|
|
|
|
|
|
|
44 |
|
45 |
class DiVAModel(PreTrainedModel):
|
46 |
config_class = DiVAConfig
|
47 |
+
|
48 |
def __init__(
|
49 |
self, via_path=None, config_dict={}, device_map=None, speech_encoder_device=None
|
50 |
):
|
|
|
105 |
)
|
106 |
self.speech_encoder_device = speech_encoder_device
|
107 |
|
108 |
+
def can_generate(cls):
|
|
|
109 |
return False
|
110 |
+
|
111 |
@classmethod
|
112 |
def from_pretrained(
|
113 |
cls,
|
|
|
181 |
|
182 |
return outputs
|
183 |
|
184 |
+
@torch.no_grad()
|
185 |
def generate(
|
186 |
+
self,
|
187 |
+
audio,
|
188 |
+
text_prompt=None,
|
189 |
+
do_sample=False,
|
190 |
+
logits_processor=None,
|
191 |
+
max_new_tokens=128,
|
192 |
):
|
193 |
inputs = self.processor(audio, return_tensors="pt", sampling_rate=16_000)
|
194 |
input_features = inputs.input_features.to(self.speech_encoder_device)
|
|
|
198 |
virt_tokens = self.connector(
|
199 |
hidden_states,
|
200 |
output_device=self.llama_decoder.model.embed_tokens.weight.device,
|
201 |
+
)
|
202 |
+
bsz = virt_tokens.shape[0]
|
203 |
|
204 |
if text_prompt != None and text_prompt != "":
|
205 |
user_prompt_text = torch.tensor(
|
206 |
+
self.tokenizer(
|
207 |
+
text_prompt,
|
208 |
+
add_special_tokens=False,
|
209 |
+
padding=True,
|
210 |
+
padding_side="right",
|
211 |
+
)["input_ids"],
|
212 |
device=self.pre_user_suffix.device,
|
213 |
)
|
214 |
prefix = torch.cat(
|
215 |
+
[
|
216 |
+
self.pre_user_suffix.expand(
|
217 |
+
bsz,
|
218 |
+
-1,
|
219 |
+
),
|
220 |
+
user_prompt_text,
|
221 |
+
self.prefix.expand(
|
222 |
+
bsz,
|
223 |
+
-1,
|
224 |
+
),
|
225 |
+
],
|
226 |
+
axis=1,
|
227 |
)
|
228 |
else:
|
229 |
prefix = self.prefix
|
230 |
+
prefix_embed = self.llama_decoder.model.embed_tokens(prefix).expand(bsz, -1, -1)
|
231 |
suffix = self.final_header
|
232 |
+
suffix_embed = self.llama_decoder.model.embed_tokens(suffix).expand(bsz, -1, -1)
|
233 |
+
inputs_embeds = torch.cat([prefix_embed, virt_tokens, suffix_embed], axis=1)
|
234 |
+
outs = [[] for i in range(bsz)]
|
235 |
+
complete = [False] * bsz
|
|
|
236 |
outputs = None
|
237 |
greedy = 1
|
238 |
i = 0
|
239 |
+
while not all(complete) and len(outs[0]) < max_new_tokens:
|
240 |
past_key_values = outputs.past_key_values if outputs else None
|
241 |
outputs = self.llama_decoder(
|
242 |
inputs_embeds=inputs_embeds.to(
|
|
|
246 |
output_hidden_states=True,
|
247 |
past_key_values=past_key_values,
|
248 |
)
|
249 |
+
next_token_logits = outputs.logits[:, -1, :]
|
250 |
|
251 |
if logits_processor:
|
252 |
local_outs = torch.tensor(outs) if outs != [] else suffix
|
|
|
261 |
probs = F.softmax(logits, dim=-1)
|
262 |
greedy = torch.multinomial(probs, num_samples=1)[0]
|
263 |
else:
|
264 |
+
greedy = next_token_logits.argmax(dim=-1)
|
265 |
+
for token_index, out in enumerate(greedy.flatten().tolist()):
|
266 |
+
outs[token_index].append(out)
|
267 |
+
if out == 128009:
|
268 |
+
complete[token_index] = True
|
269 |
+
|
270 |
+
next_embed = self.llama_decoder.model.embed_tokens(greedy.reshape(-1, 1))
|
271 |
inputs_embeds = next_embed
|
272 |
+
return self.tokenizer.batch_decode(outs, skip_special_tokens=True)
|
|
|
|
|
273 |
|
274 |
def generate_stream(
|
275 |
+
self,
|
276 |
+
audio,
|
277 |
+
text_prompt,
|
278 |
+
do_sample=False,
|
279 |
+
logits_processor=None,
|
280 |
+
max_new_tokens=128,
|
281 |
):
|
282 |
inputs = self.processor(audio, return_tensors="pt", sampling_rate=16_000)
|
283 |
input_features = inputs.input_features.to(self.whisper_encoder.device)
|
|
|
312 |
while greedy != 128009 and len(outs) < max_new_tokens:
|
313 |
past_key_values = outputs.past_key_values if outputs else None
|
314 |
outputs = self.llama_decoder(
|
315 |
+
inputs_embeds=inputs_embeds.to(
|
316 |
self.llama_decoder.model.embed_tokens.weight.device
|
317 |
).half(),
|
318 |
return_dict=True,
|
|
|
338 |
outs.append(greedy)
|
339 |
next_embed = self.llama_decoder.model.embed_tokens(greedy.reshape(1, 1))
|
340 |
inputs_embeds = next_embed
|
341 |
+
yield self.tokenizer.decode(outs, skip_special_tokens=True).replace(
|
342 |
+
"<|eot_id|>", ""
|
343 |
+
)
|
344 |
+
return self.tokenizer.decode(outs, skip_special_tokens=True).replace(
|
345 |
+
"<|eot_id|>", ""
|
346 |
+
)
|
test.py
ADDED
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import AutoModel
|
2 |
+
import librosa
|
3 |
+
import wget
|
4 |
+
from modeling_diva import DiVAModel
|
5 |
+
|
6 |
+
filename = wget.download(
|
7 |
+
"https://github.com/ffaisal93/SD-QA/raw/refs/heads/master/dev/eng/irl/wav_eng/-1008642825401516622.wav"
|
8 |
+
)
|
9 |
+
|
10 |
+
speech_data, _ = librosa.load(filename, sr=16_000)
|
11 |
+
|
12 |
+
model = DiVAModel.from_pretrained("./")
|
13 |
+
|
14 |
+
print(model.generate([speech_data]))
|
15 |
+
print(model.generate([speech_data], ["Reply Briefly Like A Pirate"]))
|
16 |
+
|
17 |
+
filename = wget.download(
|
18 |
+
"https://github.com/ffaisal93/SD-QA/raw/refs/heads/master/dev/eng/irl/wav_eng/-2426554427049983479.wav"
|
19 |
+
)
|
20 |
+
|
21 |
+
speech_data2, _ = librosa.load(filename, sr=16_000)
|
22 |
+
|
23 |
+
print(
|
24 |
+
model.generate(
|
25 |
+
[speech_data, speech_data2],
|
26 |
+
["Reply Briefly Like A Pirate", "Reply Briefly Like A New Yorker"],
|
27 |
+
)
|
28 |
+
)
|