Upload ultravox_model.py with huggingface_hub
Browse files- ultravox_model.py +92 -145
ultravox_model.py
CHANGED
|
@@ -33,13 +33,10 @@ SHARED_PRETRAINED_KWARGS = [
|
|
| 33 |
class UltravoxModel(transformers.LlamaPreTrainedModel, GenerationMixin):
|
| 34 |
"""
|
| 35 |
The Ultravox model which consists of an audio encoder and a language model.
|
| 36 |
-
|
| 37 |
Audio input is processed by the audio encoder, then every `stack_factor` frames are stacked together and
|
| 38 |
projected to the language model's embedding space using a few linear layers.
|
| 39 |
The text is embedded by the language model as usual and then the audio and text embeddings are merged together.
|
| 40 |
-
|
| 41 |
A special token `<|audio|>` is used to indicate the start of the audio embeddings in the merged embeddings.
|
| 42 |
-
|
| 43 |
Parameters:
|
| 44 |
config: Model configuration class with all the parameters of the model.
|
| 45 |
"""
|
|
@@ -59,11 +56,11 @@ class UltravoxModel(transformers.LlamaPreTrainedModel, GenerationMixin):
|
|
| 59 |
self.keep_params: Set[str] = set()
|
| 60 |
self.vocab_size = config.vocab_size
|
| 61 |
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
self.audio_tower_context_length = self.audio_tower.max_context_length
|
| 66 |
|
|
|
|
| 67 |
self.language_model = self._create_language_model(config)
|
| 68 |
|
| 69 |
if self.language_model._tied_weights_keys is not None:
|
|
@@ -72,16 +69,9 @@ class UltravoxModel(transformers.LlamaPreTrainedModel, GenerationMixin):
|
|
| 72 |
]
|
| 73 |
|
| 74 |
# Determine no_split_modules dynamically to use with FSDP auto_wrap policy.
|
|
|
|
| 75 |
# This would be something like ["LlamaDecoderLayer"] as we don't split audio encoder layers.
|
| 76 |
-
|
| 77 |
-
# 1. Get the names the language model *wants* to keep intact
|
| 78 |
-
candidate_names = set(
|
| 79 |
-
getattr(self.language_model, "_no_split_modules", []) or []
|
| 80 |
-
)
|
| 81 |
-
# 2. Names that actually exist in the current model
|
| 82 |
-
present_names = {m.__class__.__name__ for m in self.modules()}
|
| 83 |
-
# 3. Keep only those that are both requested and present
|
| 84 |
-
self._no_split_modules = list(candidate_names & present_names)
|
| 85 |
|
| 86 |
self.loss_config = LossConfig()
|
| 87 |
self.post_init()
|
|
@@ -159,17 +149,13 @@ class UltravoxModel(transformers.LlamaPreTrainedModel, GenerationMixin):
|
|
| 159 |
self, labels: Optional[torch.Tensor]
|
| 160 |
) -> Tuple[torch.Tensor, torch.Tensor]:
|
| 161 |
"""Get boolean masks for positions where we want to compute KL divergence.
|
| 162 |
-
|
| 163 |
For each label position, we want the position before it since that's where
|
| 164 |
the model makes the prediction for that label.
|
| 165 |
-
|
| 166 |
Additionally, we want to identify the position right before the EOT token
|
| 167 |
(the last token with label != -100).
|
| 168 |
-
|
| 169 |
Args:
|
| 170 |
labels: Tensor of shape (B, T) where B is batch size and T is sequence length,
|
| 171 |
with -100 for masked positions and token ids for label positions
|
| 172 |
-
|
| 173 |
Returns:
|
| 174 |
Tuple containing:
|
| 175 |
- pred_mask: Boolean tensor of shape (B, T) that's True for positions where we want to compute KL divergence
|
|
@@ -239,32 +225,27 @@ class UltravoxModel(transformers.LlamaPreTrainedModel, GenerationMixin):
|
|
| 239 |
)
|
| 240 |
|
| 241 |
# Compute the KL divergence loss for EOT token positions if any exist
|
| 242 |
-
|
| 243 |
-
|
| 244 |
-
|
| 245 |
-
|
| 246 |
-
|
| 247 |
-
|
| 248 |
-
|
| 249 |
-
|
| 250 |
-
|
| 251 |
-
|
| 252 |
-
|
| 253 |
-
reduction="batchmean",
|
| 254 |
-
)
|
| 255 |
-
kl_loss += self.loss_config.eot_loss_weight * eot_loss
|
| 256 |
|
| 257 |
-
return kl_loss
|
| 258 |
|
| 259 |
def _audio_iter(
|
| 260 |
self, audio_batch_size: torch.Tensor
|
| 261 |
) -> Generator[Tuple[int, int], None, None]:
|
| 262 |
"""
|
| 263 |
Iterate over the audio batch size and yield the batch index and audio index of each audio item.
|
| 264 |
-
|
| 265 |
Args:
|
| 266 |
audio_batch_size: A tensor of shape (B,) where B is the batch size.
|
| 267 |
-
|
| 268 |
Returns:
|
| 269 |
A generator that yields a tuple of (start index, length) for each audio item.
|
| 270 |
"""
|
|
@@ -277,8 +258,8 @@ class UltravoxModel(transformers.LlamaPreTrainedModel, GenerationMixin):
|
|
| 277 |
def forward(
|
| 278 |
self,
|
| 279 |
input_ids: torch.Tensor,
|
| 280 |
-
audio_values: Optional[torch.
|
| 281 |
-
inputs_embeds: Optional[torch.
|
| 282 |
labels: Optional[torch.Tensor] = None,
|
| 283 |
attention_mask: Optional[torch.Tensor] = None,
|
| 284 |
audio_token_start_idx: Optional[torch.Tensor] = None,
|
|
@@ -291,16 +272,14 @@ class UltravoxModel(transformers.LlamaPreTrainedModel, GenerationMixin):
|
|
| 291 |
alt_attention_mask: Optional[torch.Tensor] = None,
|
| 292 |
alt_labels: Optional[torch.Tensor] = None,
|
| 293 |
**kwargs,
|
| 294 |
-
) -> transformers.modeling_outputs.CausalLMOutputWithPast:
|
| 295 |
"""
|
| 296 |
Forward pass for the Ultravox model.
|
| 297 |
-
|
| 298 |
`input_ids` are the tokenized text input. They are embedded by the language model as usual.
|
| 299 |
`audio_values` are processed by the audio encoder and then every `stack_factor` frames are stacked together and
|
| 300 |
projected to the language model's embedding space using a few linear layers.
|
| 301 |
The audio and text embeddings are merged together. A special token `<|audio|>` is used to indicate the start
|
| 302 |
of the audio embeddings in the merged embeddings.
|
| 303 |
-
|
| 304 |
Args:
|
| 305 |
input_ids: The tokenized text input.
|
| 306 |
audio_values: The processed audio values.
|
|
@@ -316,14 +295,36 @@ class UltravoxModel(transformers.LlamaPreTrainedModel, GenerationMixin):
|
|
| 316 |
inputs_embeds = self.get_input_embeddings().forward(input_ids)
|
| 317 |
|
| 318 |
if audio_values is not None and len(audio_values) > 0:
|
| 319 |
-
|
| 320 |
-
|
| 321 |
-
|
| 322 |
-
|
| 323 |
-
|
| 324 |
-
|
| 325 |
-
|
| 326 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 327 |
|
| 328 |
lm_output = self.language_model.forward(
|
| 329 |
inputs_embeds=inputs_embeds,
|
|
@@ -334,9 +335,9 @@ class UltravoxModel(transformers.LlamaPreTrainedModel, GenerationMixin):
|
|
| 334 |
)
|
| 335 |
if self.training:
|
| 336 |
if self.loss_config.loss_function == LossFunction.CrossEntropy:
|
| 337 |
-
|
| 338 |
elif self.loss_config.loss_function == LossFunction.KL_Divergence:
|
| 339 |
-
|
| 340 |
lm_output=lm_output,
|
| 341 |
labels=labels,
|
| 342 |
past_key_values=past_key_values,
|
|
@@ -349,82 +350,52 @@ class UltravoxModel(transformers.LlamaPreTrainedModel, GenerationMixin):
|
|
| 349 |
raise ValueError(
|
| 350 |
f"Unsupported loss function: {self.loss_config.loss_function}"
|
| 351 |
)
|
| 352 |
-
|
|
|
|
| 353 |
|
| 354 |
-
def
|
| 355 |
-
self,
|
| 356 |
-
inputs_embeds: Optional[torch.Tensor] = None,
|
| 357 |
-
audio_values: Optional[torch.Tensor] = None,
|
| 358 |
-
audio_token_start_idx: Optional[torch.Tensor] = None,
|
| 359 |
-
audio_lens: Optional[torch.Tensor] = None,
|
| 360 |
-
audio_token_len: Optional[torch.Tensor] = None,
|
| 361 |
-
audio_batch_size: Optional[torch.Tensor] = None,
|
| 362 |
-
) -> torch.Tensor:
|
| 363 |
-
assert (
|
| 364 |
-
inputs_embeds is not None
|
| 365 |
-
and audio_values is not None
|
| 366 |
-
and audio_token_start_idx is not None
|
| 367 |
-
and audio_token_len is not None
|
| 368 |
-
and audio_lens is not None
|
| 369 |
-
and audio_batch_size is not None
|
| 370 |
-
), "inputs_embeds/audio_values/audio_token_start_idx/audio_token_len/audio_lens/audio_batch_size must be provided."
|
| 371 |
-
assert (
|
| 372 |
-
len(audio_token_start_idx)
|
| 373 |
-
== len(audio_token_len)
|
| 374 |
-
== len(audio_lens)
|
| 375 |
-
== len(audio_values)
|
| 376 |
-
), "audio_token_start_idx/audio_token_len/audio_lens/audio_values must have the same batch size."
|
| 377 |
-
assert len(audio_batch_size) == len(
|
| 378 |
-
inputs_embeds
|
| 379 |
-
), "audio_batch_size and inputs_embeds must have the same batch size."
|
| 380 |
-
|
| 381 |
-
# B x A/3200 x (D=max-audio-length-in-batch)
|
| 382 |
-
audio_tower_output = self.audio_tower.forward(
|
| 383 |
-
audio_values.to(self.audio_tower.dtype),
|
| 384 |
-
audio_len=audio_lens,
|
| 385 |
-
).last_hidden_state
|
| 386 |
-
audio_tower_output = audio_tower_output.to(inputs_embeds.dtype)
|
| 387 |
-
audio_embeds = self.multi_modal_projector.forward(audio_tower_output)
|
| 388 |
-
|
| 389 |
-
# combine audio and text embeddings
|
| 390 |
-
for i_b, i_a in self._audio_iter(audio_batch_size):
|
| 391 |
-
start_idx = audio_token_start_idx[i_a]
|
| 392 |
-
token_len = audio_token_len[i_a]
|
| 393 |
-
item_embedding = audio_embeds[i_a][:token_len]
|
| 394 |
-
inputs_embeds[i_b][start_idx : start_idx + token_len] = item_embedding
|
| 395 |
-
|
| 396 |
-
return inputs_embeds
|
| 397 |
-
|
| 398 |
-
def generate(
|
| 399 |
self,
|
| 400 |
input_ids: torch.Tensor,
|
| 401 |
-
audio_values: Optional[torch.
|
| 402 |
-
inputs_embeds: Optional[torch.Tensor] = None,
|
| 403 |
audio_token_start_idx: Optional[torch.Tensor] = None,
|
| 404 |
-
audio_lens: Optional[torch.Tensor] = None,
|
| 405 |
audio_token_len: Optional[torch.Tensor] = None,
|
|
|
|
| 406 |
audio_batch_size: Optional[torch.Tensor] = None,
|
|
|
|
|
|
|
|
|
|
|
|
|
| 407 |
**kwargs,
|
| 408 |
-
) ->
|
| 409 |
-
|
| 410 |
-
inputs_embeds = self.get_input_embeddings().forward(input_ids)
|
| 411 |
-
|
| 412 |
-
if audio_values is not None and len(audio_values) > 0:
|
| 413 |
-
inputs_embeds = self._prepare_audio_embeds(
|
| 414 |
-
inputs_embeds=inputs_embeds,
|
| 415 |
-
audio_values=audio_values,
|
| 416 |
-
audio_token_start_idx=audio_token_start_idx,
|
| 417 |
-
audio_lens=audio_lens,
|
| 418 |
-
audio_token_len=audio_token_len,
|
| 419 |
-
audio_batch_size=audio_batch_size,
|
| 420 |
-
)
|
| 421 |
-
|
| 422 |
-
return self.language_model.generate(
|
| 423 |
input_ids=input_ids,
|
|
|
|
|
|
|
| 424 |
inputs_embeds=inputs_embeds,
|
|
|
|
| 425 |
**kwargs,
|
| 426 |
)
|
| 427 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 428 |
@classmethod
|
| 429 |
def _create_multi_modal_projector(
|
| 430 |
cls, config: UltravoxConfig
|
|
@@ -454,9 +425,6 @@ class UltravoxModel(transformers.LlamaPreTrainedModel, GenerationMixin):
|
|
| 454 |
audio_tower.init_latency_mask(
|
| 455 |
config.audio_latency_block_size, dtype=config.torch_dtype
|
| 456 |
)
|
| 457 |
-
audio_tower.init_latency_mask(
|
| 458 |
-
config.audio_latency_block_size, dtype=config.torch_dtype
|
| 459 |
-
)
|
| 460 |
else:
|
| 461 |
assert config.audio_latency_block_size in (
|
| 462 |
None,
|
|
@@ -539,9 +507,7 @@ class UltravoxModel(transformers.LlamaPreTrainedModel, GenerationMixin):
|
|
| 539 |
)
|
| 540 |
)
|
| 541 |
|
| 542 |
-
if
|
| 543 |
-
self.audio_tower, peft.PeftModel
|
| 544 |
-
):
|
| 545 |
self.audio_tower = self.audio_tower.merge_and_unload()
|
| 546 |
# no need to download base audio model weights anymore, so we can remove the id
|
| 547 |
self.config.audio_model_id = None
|
|
@@ -607,33 +573,18 @@ class UltravoxModel(transformers.LlamaPreTrainedModel, GenerationMixin):
|
|
| 607 |
)
|
| 608 |
|
| 609 |
lm_trainable_params, lm_all_params = count_params(self.language_model)
|
| 610 |
-
|
| 611 |
-
audio_trainable_params, audio_all_params = count_params(self.audio_tower)
|
| 612 |
-
else:
|
| 613 |
-
audio_trainable_params, audio_all_params = 0, 0
|
| 614 |
|
| 615 |
projector_trainable_params = (
|
| 616 |
trainable_params - lm_trainable_params - audio_trainable_params
|
| 617 |
)
|
| 618 |
projector_all_params = all_param - lm_all_params - audio_all_params
|
| 619 |
|
| 620 |
-
# Calculate percentages only if the total parameters are non-zero
|
| 621 |
-
audio_percent = (
|
| 622 |
-
0.0
|
| 623 |
-
if audio_all_params == 0
|
| 624 |
-
else 100 * audio_trainable_params / audio_all_params
|
| 625 |
-
)
|
| 626 |
-
projector_percent = (
|
| 627 |
-
0.0
|
| 628 |
-
if projector_all_params == 0
|
| 629 |
-
else 100 * projector_trainable_params / projector_all_params
|
| 630 |
-
)
|
| 631 |
-
|
| 632 |
logging.info(
|
| 633 |
f"Trainable%: "
|
| 634 |
f" LLM: {100 * lm_trainable_params / lm_all_params:.1f}%"
|
| 635 |
-
f" || Audio Encoder: {
|
| 636 |
-
f" || Projector: {
|
| 637 |
)
|
| 638 |
|
| 639 |
|
|
@@ -770,7 +721,6 @@ class UltravoxProjector(nn.Module):
|
|
| 770 |
Takes in audio features from the audio tower and projects them to the text model's embedding space.
|
| 771 |
It reduces the number of frames by a factor of `stack_factor` and increases the number of channels by the same factor.
|
| 772 |
If the number of audio frames are not a multiple of the stack factor, the last few frames will be padded with zeros.
|
| 773 |
-
|
| 774 |
Input shape:
|
| 775 |
audio_features: B, T*S, C
|
| 776 |
Output shape:
|
|
@@ -784,7 +734,6 @@ class UltravoxProjector(nn.Module):
|
|
| 784 |
C: number of channels out of the encoder (aka audio tower)
|
| 785 |
H: hidden size of the projector (config.hidden_size)
|
| 786 |
D: dimension of the text model (config.text_config.hidden_size)
|
| 787 |
-
|
| 788 |
"""
|
| 789 |
# B, F, C -> B, T, C*S
|
| 790 |
audio_features = self._pad_and_stack(audio_features)
|
|
@@ -805,13 +754,11 @@ class ModifiedWhisperEncoder(
|
|
| 805 |
):
|
| 806 |
"""
|
| 807 |
Encoder portion of OpenAI's Whisper model.
|
| 808 |
-
|
| 809 |
This implementation is a slightly modified version of HF Transformers' Whisper Encoder, with only a few fixes:
|
| 810 |
1. base_model_prefix updated to allow for doing `.from_pretrained` directly on the encoder
|
| 811 |
2. allow less than 30 second of audio padding to be passed in:
|
| 812 |
- relaxed ValueError check for `input_features` length to be less than or equal to `expected_seq_length` instead of strictly equal
|
| 813 |
- embed_pos is now sliced to match the length of `inputs_embeds`
|
| 814 |
-
|
| 815 |
Original: https://github.com/huggingface/transformers/blob/main/src/transformers/models/whisper/modeling_whisper.py
|
| 816 |
"""
|
| 817 |
|
|
@@ -913,7 +860,7 @@ class ModifiedWhisperEncoder(
|
|
| 913 |
# This masking ensures consistent behavior between training and inference
|
| 914 |
# by preventing the model from attending to padding tokens in both cases
|
| 915 |
attention_mask = None
|
| 916 |
-
if audio_len
|
| 917 |
audio_feature_len = self._get_feat_extract_output_lengths(audio_len)
|
| 918 |
max_seq_len = hidden_states.shape[1]
|
| 919 |
attention_mask = torch.arange(max_seq_len, device=hidden_states.device)[
|
|
|
|
| 33 |
class UltravoxModel(transformers.LlamaPreTrainedModel, GenerationMixin):
|
| 34 |
"""
|
| 35 |
The Ultravox model which consists of an audio encoder and a language model.
|
|
|
|
| 36 |
Audio input is processed by the audio encoder, then every `stack_factor` frames are stacked together and
|
| 37 |
projected to the language model's embedding space using a few linear layers.
|
| 38 |
The text is embedded by the language model as usual and then the audio and text embeddings are merged together.
|
|
|
|
| 39 |
A special token `<|audio|>` is used to indicate the start of the audio embeddings in the merged embeddings.
|
|
|
|
| 40 |
Parameters:
|
| 41 |
config: Model configuration class with all the parameters of the model.
|
| 42 |
"""
|
|
|
|
| 56 |
self.keep_params: Set[str] = set()
|
| 57 |
self.vocab_size = config.vocab_size
|
| 58 |
|
| 59 |
+
self.audio_tower = self._create_audio_tower(config)
|
| 60 |
+
self.audio_tower_context_length: Optional[int] = None
|
| 61 |
+
self.audio_tower_context_length = self.audio_tower.max_context_length
|
|
|
|
| 62 |
|
| 63 |
+
self.multi_modal_projector = self._create_multi_modal_projector(config)
|
| 64 |
self.language_model = self._create_language_model(config)
|
| 65 |
|
| 66 |
if self.language_model._tied_weights_keys is not None:
|
|
|
|
| 69 |
]
|
| 70 |
|
| 71 |
# Determine no_split_modules dynamically to use with FSDP auto_wrap policy.
|
| 72 |
+
# FSDP throws an error if some of the layer types are not found in the model.
|
| 73 |
# This would be something like ["LlamaDecoderLayer"] as we don't split audio encoder layers.
|
| 74 |
+
self._no_split_modules = self.language_model._no_split_modules
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 75 |
|
| 76 |
self.loss_config = LossConfig()
|
| 77 |
self.post_init()
|
|
|
|
| 149 |
self, labels: Optional[torch.Tensor]
|
| 150 |
) -> Tuple[torch.Tensor, torch.Tensor]:
|
| 151 |
"""Get boolean masks for positions where we want to compute KL divergence.
|
|
|
|
| 152 |
For each label position, we want the position before it since that's where
|
| 153 |
the model makes the prediction for that label.
|
|
|
|
| 154 |
Additionally, we want to identify the position right before the EOT token
|
| 155 |
(the last token with label != -100).
|
|
|
|
| 156 |
Args:
|
| 157 |
labels: Tensor of shape (B, T) where B is batch size and T is sequence length,
|
| 158 |
with -100 for masked positions and token ids for label positions
|
|
|
|
| 159 |
Returns:
|
| 160 |
Tuple containing:
|
| 161 |
- pred_mask: Boolean tensor of shape (B, T) that's True for positions where we want to compute KL divergence
|
|
|
|
| 225 |
)
|
| 226 |
|
| 227 |
# Compute the KL divergence loss for EOT token positions if any exist
|
| 228 |
+
eot_loss = F.kl_div(
|
| 229 |
+
F.log_softmax(
|
| 230 |
+
lm_output.logits[eot_mask] / self.loss_config.kl_temperature,
|
| 231 |
+
dim=-1,
|
| 232 |
+
),
|
| 233 |
+
F.softmax(
|
| 234 |
+
alt_lm_output.logits[alt_eot_mask] / self.loss_config.kl_temperature,
|
| 235 |
+
dim=-1,
|
| 236 |
+
),
|
| 237 |
+
reduction="batchmean",
|
| 238 |
+
)
|
|
|
|
|
|
|
|
|
|
| 239 |
|
| 240 |
+
return {"loss": kl_loss + self.loss_config.eot_loss_weight * eot_loss}
|
| 241 |
|
| 242 |
def _audio_iter(
|
| 243 |
self, audio_batch_size: torch.Tensor
|
| 244 |
) -> Generator[Tuple[int, int], None, None]:
|
| 245 |
"""
|
| 246 |
Iterate over the audio batch size and yield the batch index and audio index of each audio item.
|
|
|
|
| 247 |
Args:
|
| 248 |
audio_batch_size: A tensor of shape (B,) where B is the batch size.
|
|
|
|
| 249 |
Returns:
|
| 250 |
A generator that yields a tuple of (start index, length) for each audio item.
|
| 251 |
"""
|
|
|
|
| 258 |
def forward(
|
| 259 |
self,
|
| 260 |
input_ids: torch.Tensor,
|
| 261 |
+
audio_values: Optional[torch.FloatTensor] = None,
|
| 262 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 263 |
labels: Optional[torch.Tensor] = None,
|
| 264 |
attention_mask: Optional[torch.Tensor] = None,
|
| 265 |
audio_token_start_idx: Optional[torch.Tensor] = None,
|
|
|
|
| 272 |
alt_attention_mask: Optional[torch.Tensor] = None,
|
| 273 |
alt_labels: Optional[torch.Tensor] = None,
|
| 274 |
**kwargs,
|
| 275 |
+
) -> Union[Tuple, transformers.modeling_outputs.CausalLMOutputWithPast]:
|
| 276 |
"""
|
| 277 |
Forward pass for the Ultravox model.
|
|
|
|
| 278 |
`input_ids` are the tokenized text input. They are embedded by the language model as usual.
|
| 279 |
`audio_values` are processed by the audio encoder and then every `stack_factor` frames are stacked together and
|
| 280 |
projected to the language model's embedding space using a few linear layers.
|
| 281 |
The audio and text embeddings are merged together. A special token `<|audio|>` is used to indicate the start
|
| 282 |
of the audio embeddings in the merged embeddings.
|
|
|
|
| 283 |
Args:
|
| 284 |
input_ids: The tokenized text input.
|
| 285 |
audio_values: The processed audio values.
|
|
|
|
| 295 |
inputs_embeds = self.get_input_embeddings().forward(input_ids)
|
| 296 |
|
| 297 |
if audio_values is not None and len(audio_values) > 0:
|
| 298 |
+
assert (
|
| 299 |
+
audio_token_start_idx is not None
|
| 300 |
+
and audio_token_len is not None
|
| 301 |
+
and audio_lens is not None
|
| 302 |
+
and audio_batch_size is not None
|
| 303 |
+
), "audio_token_start_idx/audio_token_len/audio_lens must be provided if audio_values are provided."
|
| 304 |
+
assert (
|
| 305 |
+
len(audio_token_start_idx)
|
| 306 |
+
== len(audio_token_len)
|
| 307 |
+
== len(audio_lens)
|
| 308 |
+
== len(audio_values)
|
| 309 |
+
), "audio_token_start_idx/audio_token_len/audio_lens/audio_values must have the same batch size."
|
| 310 |
+
assert len(audio_batch_size) == len(
|
| 311 |
+
inputs_embeds
|
| 312 |
+
), "audio_batch_size and inputs_embeds must have the same batch size."
|
| 313 |
+
|
| 314 |
+
# B x A/3200 x (D=max-audio-length-in-batch)
|
| 315 |
+
audio_tower_output = self.audio_tower.forward(
|
| 316 |
+
audio_values.to(self.audio_tower.dtype),
|
| 317 |
+
audio_len=audio_lens,
|
| 318 |
+
).last_hidden_state
|
| 319 |
+
audio_tower_output = audio_tower_output.to(inputs_embeds.dtype)
|
| 320 |
+
audio_embeds = self.multi_modal_projector.forward(audio_tower_output)
|
| 321 |
+
|
| 322 |
+
# combine audio and text embeddings
|
| 323 |
+
for i_b, i_a in self._audio_iter(audio_batch_size):
|
| 324 |
+
start_idx = audio_token_start_idx[i_a]
|
| 325 |
+
token_len = audio_token_len[i_a]
|
| 326 |
+
item_embedding = audio_embeds[i_a][:token_len]
|
| 327 |
+
inputs_embeds[i_b][start_idx : start_idx + token_len] = item_embedding
|
| 328 |
|
| 329 |
lm_output = self.language_model.forward(
|
| 330 |
inputs_embeds=inputs_embeds,
|
|
|
|
| 335 |
)
|
| 336 |
if self.training:
|
| 337 |
if self.loss_config.loss_function == LossFunction.CrossEntropy:
|
| 338 |
+
return lm_output
|
| 339 |
elif self.loss_config.loss_function == LossFunction.KL_Divergence:
|
| 340 |
+
return self._compute_kl_loss(
|
| 341 |
lm_output=lm_output,
|
| 342 |
labels=labels,
|
| 343 |
past_key_values=past_key_values,
|
|
|
|
| 350 |
raise ValueError(
|
| 351 |
f"Unsupported loss function: {self.loss_config.loss_function}"
|
| 352 |
)
|
| 353 |
+
else:
|
| 354 |
+
return lm_output
|
| 355 |
|
| 356 |
+
def prepare_inputs_for_generation(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 357 |
self,
|
| 358 |
input_ids: torch.Tensor,
|
| 359 |
+
audio_values: Optional[torch.FloatTensor] = None,
|
|
|
|
| 360 |
audio_token_start_idx: Optional[torch.Tensor] = None,
|
|
|
|
| 361 |
audio_token_len: Optional[torch.Tensor] = None,
|
| 362 |
+
audio_lens: Optional[torch.Tensor] = None,
|
| 363 |
audio_batch_size: Optional[torch.Tensor] = None,
|
| 364 |
+
past_key_values: Optional[Union[Tuple, transformers.cache_utils.Cache]] = None,
|
| 365 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 366 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
| 367 |
+
cache_position: Optional[torch.Tensor] = None,
|
| 368 |
**kwargs,
|
| 369 |
+
) -> Dict[str, Any]:
|
| 370 |
+
model_input = self.language_model.prepare_inputs_for_generation(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 371 |
input_ids=input_ids,
|
| 372 |
+
past_key_values=past_key_values,
|
| 373 |
+
attention_mask=attention_mask,
|
| 374 |
inputs_embeds=inputs_embeds,
|
| 375 |
+
cache_position=cache_position,
|
| 376 |
**kwargs,
|
| 377 |
)
|
| 378 |
|
| 379 |
+
# include audio information in model_input only when it is needed during prefilling
|
| 380 |
+
# audio_token_start_idx should always be relative to the current cache position
|
| 381 |
+
prefill_start_idx: int | torch.Tensor = (
|
| 382 |
+
0 if cache_position is None else cache_position[0]
|
| 383 |
+
)
|
| 384 |
+
if (
|
| 385 |
+
audio_values is not None
|
| 386 |
+
and audio_token_start_idx is not None
|
| 387 |
+
and prefill_start_idx <= torch.max(audio_token_start_idx)
|
| 388 |
+
):
|
| 389 |
+
model_input["audio_values"] = audio_values
|
| 390 |
+
model_input["audio_token_start_idx"] = (
|
| 391 |
+
audio_token_start_idx - prefill_start_idx
|
| 392 |
+
)
|
| 393 |
+
model_input["audio_token_len"] = audio_token_len
|
| 394 |
+
model_input["audio_batch_size"] = audio_batch_size
|
| 395 |
+
model_input["audio_lens"] = audio_lens
|
| 396 |
+
|
| 397 |
+
return model_input
|
| 398 |
+
|
| 399 |
@classmethod
|
| 400 |
def _create_multi_modal_projector(
|
| 401 |
cls, config: UltravoxConfig
|
|
|
|
| 425 |
audio_tower.init_latency_mask(
|
| 426 |
config.audio_latency_block_size, dtype=config.torch_dtype
|
| 427 |
)
|
|
|
|
|
|
|
|
|
|
| 428 |
else:
|
| 429 |
assert config.audio_latency_block_size in (
|
| 430 |
None,
|
|
|
|
| 507 |
)
|
| 508 |
)
|
| 509 |
|
| 510 |
+
if isinstance(self.audio_tower, peft.PeftModel):
|
|
|
|
|
|
|
| 511 |
self.audio_tower = self.audio_tower.merge_and_unload()
|
| 512 |
# no need to download base audio model weights anymore, so we can remove the id
|
| 513 |
self.config.audio_model_id = None
|
|
|
|
| 573 |
)
|
| 574 |
|
| 575 |
lm_trainable_params, lm_all_params = count_params(self.language_model)
|
| 576 |
+
audio_trainable_params, audio_all_params = count_params(self.audio_tower)
|
|
|
|
|
|
|
|
|
|
| 577 |
|
| 578 |
projector_trainable_params = (
|
| 579 |
trainable_params - lm_trainable_params - audio_trainable_params
|
| 580 |
)
|
| 581 |
projector_all_params = all_param - lm_all_params - audio_all_params
|
| 582 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 583 |
logging.info(
|
| 584 |
f"Trainable%: "
|
| 585 |
f" LLM: {100 * lm_trainable_params / lm_all_params:.1f}%"
|
| 586 |
+
f" || Audio Encoder: {100 * audio_trainable_params / audio_all_params:.1f}%"
|
| 587 |
+
f" || Projector: {100 * projector_trainable_params / projector_all_params:.1f}%"
|
| 588 |
)
|
| 589 |
|
| 590 |
|
|
|
|
| 721 |
Takes in audio features from the audio tower and projects them to the text model's embedding space.
|
| 722 |
It reduces the number of frames by a factor of `stack_factor` and increases the number of channels by the same factor.
|
| 723 |
If the number of audio frames are not a multiple of the stack factor, the last few frames will be padded with zeros.
|
|
|
|
| 724 |
Input shape:
|
| 725 |
audio_features: B, T*S, C
|
| 726 |
Output shape:
|
|
|
|
| 734 |
C: number of channels out of the encoder (aka audio tower)
|
| 735 |
H: hidden size of the projector (config.hidden_size)
|
| 736 |
D: dimension of the text model (config.text_config.hidden_size)
|
|
|
|
| 737 |
"""
|
| 738 |
# B, F, C -> B, T, C*S
|
| 739 |
audio_features = self._pad_and_stack(audio_features)
|
|
|
|
| 754 |
):
|
| 755 |
"""
|
| 756 |
Encoder portion of OpenAI's Whisper model.
|
|
|
|
| 757 |
This implementation is a slightly modified version of HF Transformers' Whisper Encoder, with only a few fixes:
|
| 758 |
1. base_model_prefix updated to allow for doing `.from_pretrained` directly on the encoder
|
| 759 |
2. allow less than 30 second of audio padding to be passed in:
|
| 760 |
- relaxed ValueError check for `input_features` length to be less than or equal to `expected_seq_length` instead of strictly equal
|
| 761 |
- embed_pos is now sliced to match the length of `inputs_embeds`
|
|
|
|
| 762 |
Original: https://github.com/huggingface/transformers/blob/main/src/transformers/models/whisper/modeling_whisper.py
|
| 763 |
"""
|
| 764 |
|
|
|
|
| 860 |
# This masking ensures consistent behavior between training and inference
|
| 861 |
# by preventing the model from attending to padding tokens in both cases
|
| 862 |
attention_mask = None
|
| 863 |
+
if audio_len != None:
|
| 864 |
audio_feature_len = self._get_feat_extract_output_lengths(audio_len)
|
| 865 |
max_seq_len = hidden_states.shape[1]
|
| 866 |
attention_mask = torch.arange(max_seq_len, device=hidden_states.device)[
|