Spaces:
Running
on
T4
Running
on
T4
Update to fix Collab launch
Browse files- app.py +36 -0
- audiocraft/__init__.py +1 -1
- audiocraft/models/lm.py +2 -1
- audiocraft/models/musicgen.py +85 -11
- audiocraft/modules/transformer.py +67 -24
app.py
CHANGED
@@ -402,6 +402,27 @@ def ui(**kwargs):
|
|
402 |
if __name__ == "__main__":
|
403 |
parser = argparse.ArgumentParser()
|
404 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
405 |
parser.add_argument(
|
406 |
'--share', action='store_true', help='Share the gradio UI'
|
407 |
)
|
@@ -418,6 +439,21 @@ if __name__ == "__main__":
|
|
418 |
)
|
419 |
|
420 |
args = parser.parse_args()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
421 |
UNLOAD_MODEL = args.unload_model
|
422 |
MOVE_TO_CPU = args.unload_to_cpu
|
423 |
if args.cache:
|
|
|
402 |
if __name__ == "__main__":
|
403 |
parser = argparse.ArgumentParser()
|
404 |
|
405 |
+
parser.add_argument(
|
406 |
+
'--listen',
|
407 |
+
type=str,
|
408 |
+
default='0.0.0.0' if 'SPACE_ID' in os.environ else '127.0.0.1',
|
409 |
+
help='IP to listen on for connections to Gradio',
|
410 |
+
)
|
411 |
+
parser.add_argument(
|
412 |
+
'--username', type=str, default='', help='Username for authentication'
|
413 |
+
)
|
414 |
+
parser.add_argument(
|
415 |
+
'--password', type=str, default='', help='Password for authentication'
|
416 |
+
)
|
417 |
+
parser.add_argument(
|
418 |
+
'--server_port',
|
419 |
+
type=int,
|
420 |
+
default=0,
|
421 |
+
help='Port to run the server listener on',
|
422 |
+
)
|
423 |
+
parser.add_argument(
|
424 |
+
'--inbrowser', action='store_true', help='Open in browser'
|
425 |
+
)
|
426 |
parser.add_argument(
|
427 |
'--share', action='store_true', help='Share the gradio UI'
|
428 |
)
|
|
|
439 |
)
|
440 |
|
441 |
args = parser.parse_args()
|
442 |
+
|
443 |
+
launch_kwargs = {}
|
444 |
+
launch_kwargs['server_name'] = args.listen
|
445 |
+
|
446 |
+
if args.username and args.password:
|
447 |
+
launch_kwargs['auth'] = (args.username, args.password)
|
448 |
+
if args.server_port:
|
449 |
+
launch_kwargs['server_port'] = args.server_port
|
450 |
+
if args.inbrowser:
|
451 |
+
launch_kwargs['inbrowser'] = args.inbrowser
|
452 |
+
if args.share:
|
453 |
+
launch_kwargs['share'] = args.share
|
454 |
+
launch_kwargs['favicon_path']= "./assets/favicon.ico"
|
455 |
+
|
456 |
+
|
457 |
UNLOAD_MODEL = args.unload_model
|
458 |
MOVE_TO_CPU = args.unload_to_cpu
|
459 |
if args.cache:
|
audiocraft/__init__.py
CHANGED
@@ -7,4 +7,4 @@
|
|
7 |
# flake8: noqa
|
8 |
from . import data, modules, models
|
9 |
|
10 |
-
__version__ = '0.0.
|
|
|
7 |
# flake8: noqa
|
8 |
from . import data, modules, models
|
9 |
|
10 |
+
__version__ = '0.0.2a2'
|
audiocraft/models/lm.py
CHANGED
@@ -363,7 +363,8 @@ class LMModel(StreamingModule):
|
|
363 |
logits = logits.permute(0, 1, 3, 2) # [B, K, card, T]
|
364 |
logits = logits[..., -1] # [B x K x card]
|
365 |
|
366 |
-
if
|
|
|
367 |
probs = torch.softmax(logits / temp, dim=-1)
|
368 |
if top_p > 0.0:
|
369 |
next_token = utils.sample_top_p(probs, p=top_p)
|
|
|
363 |
logits = logits.permute(0, 1, 3, 2) # [B, K, card, T]
|
364 |
logits = logits[..., -1] # [B x K x card]
|
365 |
|
366 |
+
# Apply softmax for sampling if temp > 0. Else, do greedy sampling to avoid zero division error.
|
367 |
+
if use_sampling and temp > 0.0:
|
368 |
probs = torch.softmax(logits / temp, dim=-1)
|
369 |
if top_p > 0.0:
|
370 |
next_token = utils.sample_top_p(probs, p=top_p)
|
audiocraft/models/musicgen.py
CHANGED
@@ -36,13 +36,16 @@ class MusicGen:
|
|
36 |
used to map audio to invertible discrete representations.
|
37 |
lm (LMModel): Language model over discrete representations.
|
38 |
"""
|
39 |
-
def __init__(self, name: str, compression_model: CompressionModel, lm: LMModel):
|
40 |
self.name = name
|
41 |
self.compression_model = compression_model
|
42 |
self.lm = lm
|
|
|
|
|
43 |
self.device = next(iter(lm.parameters())).device
|
44 |
self.generation_params: dict = {}
|
45 |
-
self.set_generation_params(duration=
|
|
|
46 |
if self.device.type == 'cpu':
|
47 |
self.autocast = TorchAutocast(enabled=False)
|
48 |
else:
|
@@ -65,7 +68,7 @@ class MusicGen:
|
|
65 |
return self.compression_model.channels
|
66 |
|
67 |
@staticmethod
|
68 |
-
def get_pretrained(name: str = 'melody', device=
|
69 |
"""Return pretrained model, we provide four models:
|
70 |
- small (300M), text to music, # see: https://huggingface.co/facebook/musicgen-small
|
71 |
- medium (1.5B), text to music, # see: https://huggingface.co/facebook/musicgen-medium
|
@@ -73,6 +76,12 @@ class MusicGen:
|
|
73 |
- large (3.3B), text to music, # see: https://huggingface.co/facebook/musicgen-large
|
74 |
"""
|
75 |
|
|
|
|
|
|
|
|
|
|
|
|
|
76 |
if name == 'debug':
|
77 |
# used only for unit tests
|
78 |
compression_model = get_debug_compression_model(device)
|
@@ -97,7 +106,7 @@ class MusicGen:
|
|
97 |
def set_generation_params(self, use_sampling: bool = True, top_k: int = 250,
|
98 |
top_p: float = 0.0, temperature: float = 1.0,
|
99 |
duration: float = 30.0, cfg_coef: float = 3.0,
|
100 |
-
two_step_cfg: bool = False, rep_penalty: float = None):
|
101 |
"""Set the generation parameters for MusicGen.
|
102 |
|
103 |
Args:
|
@@ -112,9 +121,11 @@ class MusicGen:
|
|
112 |
are padded but seems to have little impact in practice.
|
113 |
rep_penalty (float, optional): If set, use repetition penalty during generation. Not Implemented.
|
114 |
"""
|
115 |
-
assert
|
|
|
|
|
116 |
self.generation_params = {
|
117 |
-
'max_gen_len': int(duration * self.frame_rate),
|
118 |
'use_sampling': use_sampling,
|
119 |
'temp': temperature,
|
120 |
'top_k': top_k,
|
@@ -123,6 +134,10 @@ class MusicGen:
|
|
123 |
'two_step_cfg': two_step_cfg,
|
124 |
}
|
125 |
|
|
|
|
|
|
|
|
|
126 |
def generate_unconditional(self, num_samples: int, progress: bool = False) -> torch.Tensor:
|
127 |
"""Generate samples in an unconditional manner.
|
128 |
|
@@ -317,20 +332,79 @@ class MusicGen:
|
|
317 |
Returns:
|
318 |
torch.Tensor: Generated audio, of shape [B, C, T], T is defined by the generation params.
|
319 |
"""
|
|
|
|
|
|
|
|
|
320 |
def _progress_callback(generated_tokens: int, tokens_to_generate: int):
|
321 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
322 |
|
323 |
if prompt_tokens is not None:
|
324 |
-
assert
|
325 |
"Prompt is longer than audio to generate"
|
326 |
|
327 |
callback = None
|
328 |
if progress:
|
329 |
callback = _progress_callback
|
330 |
|
331 |
-
|
332 |
-
|
333 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
334 |
|
335 |
# generate audio
|
336 |
assert gen_tokens.dim() == 3
|
|
|
36 |
used to map audio to invertible discrete representations.
|
37 |
lm (LMModel): Language model over discrete representations.
|
38 |
"""
|
39 |
+
def __init__(self, name: str, compression_model: CompressionModel, lm: LMModel, max_duration: float = 30):
|
40 |
self.name = name
|
41 |
self.compression_model = compression_model
|
42 |
self.lm = lm
|
43 |
+
self.max_duration = max_duration
|
44 |
+
self.duration = 15.0 # default duration
|
45 |
self.device = next(iter(lm.parameters())).device
|
46 |
self.generation_params: dict = {}
|
47 |
+
self.set_generation_params(duration=self.duration) # 15 seconds by default
|
48 |
+
self._progress_callback: tp.Optional[tp.Callable[[int, int], None]] = None
|
49 |
if self.device.type == 'cpu':
|
50 |
self.autocast = TorchAutocast(enabled=False)
|
51 |
else:
|
|
|
68 |
return self.compression_model.channels
|
69 |
|
70 |
@staticmethod
|
71 |
+
def get_pretrained(name: str = 'melody', device=None):
|
72 |
"""Return pretrained model, we provide four models:
|
73 |
- small (300M), text to music, # see: https://huggingface.co/facebook/musicgen-small
|
74 |
- medium (1.5B), text to music, # see: https://huggingface.co/facebook/musicgen-medium
|
|
|
76 |
- large (3.3B), text to music, # see: https://huggingface.co/facebook/musicgen-large
|
77 |
"""
|
78 |
|
79 |
+
if device is None:
|
80 |
+
if torch.cuda.device_count():
|
81 |
+
device = 'cuda'
|
82 |
+
else:
|
83 |
+
device = 'cpu'
|
84 |
+
|
85 |
if name == 'debug':
|
86 |
# used only for unit tests
|
87 |
compression_model = get_debug_compression_model(device)
|
|
|
106 |
def set_generation_params(self, use_sampling: bool = True, top_k: int = 250,
|
107 |
top_p: float = 0.0, temperature: float = 1.0,
|
108 |
duration: float = 30.0, cfg_coef: float = 3.0,
|
109 |
+
two_step_cfg: bool = False, extend_stride: float = 18, rep_penalty: float = None):
|
110 |
"""Set the generation parameters for MusicGen.
|
111 |
|
112 |
Args:
|
|
|
121 |
are padded but seems to have little impact in practice.
|
122 |
rep_penalty (float, optional): If set, use repetition penalty during generation. Not Implemented.
|
123 |
"""
|
124 |
+
assert extend_stride < self.max_duration, "Cannot stride by more than max generation duration."
|
125 |
+
self.extend_stride = extend_stride
|
126 |
+
self.duration = duration
|
127 |
self.generation_params = {
|
128 |
+
#'max_gen_len': int(duration * self.frame_rate),
|
129 |
'use_sampling': use_sampling,
|
130 |
'temp': temperature,
|
131 |
'top_k': top_k,
|
|
|
134 |
'two_step_cfg': two_step_cfg,
|
135 |
}
|
136 |
|
137 |
+
def set_custom_progress_callback(self, progress_callback: tp.Optional[tp.Callable[[int, int], None]] = None):
|
138 |
+
"""Override the default progress callback."""
|
139 |
+
self._progress_callback = progress_callback
|
140 |
+
|
141 |
def generate_unconditional(self, num_samples: int, progress: bool = False) -> torch.Tensor:
|
142 |
"""Generate samples in an unconditional manner.
|
143 |
|
|
|
332 |
Returns:
|
333 |
torch.Tensor: Generated audio, of shape [B, C, T], T is defined by the generation params.
|
334 |
"""
|
335 |
+
total_gen_len = int(self.duration * self.frame_rate)
|
336 |
+
max_prompt_len = int(min(self.duration, self.max_duration) * self.frame_rate)
|
337 |
+
current_gen_offset: int = 0
|
338 |
+
|
339 |
def _progress_callback(generated_tokens: int, tokens_to_generate: int):
|
340 |
+
generated_tokens += current_gen_offset
|
341 |
+
if self._progress_callback is not None:
|
342 |
+
# Note that total_gen_len might be quite wrong depending on the
|
343 |
+
# codebook pattern used, but with delay it is almost accurate.
|
344 |
+
self._progress_callback(generated_tokens, total_gen_len)
|
345 |
+
else:
|
346 |
+
print(f'{generated_tokens: 6d} / {total_gen_len: 6d}', end='\r')
|
347 |
|
348 |
if prompt_tokens is not None:
|
349 |
+
assert max_prompt_len >= prompt_tokens.shape[-1], \
|
350 |
"Prompt is longer than audio to generate"
|
351 |
|
352 |
callback = None
|
353 |
if progress:
|
354 |
callback = _progress_callback
|
355 |
|
356 |
+
if self.duration <= self.max_duration:
|
357 |
+
# generate by sampling from LM, simple case.
|
358 |
+
with self.autocast:
|
359 |
+
gen_tokens = self.lm.generate(
|
360 |
+
prompt_tokens, attributes,
|
361 |
+
callback=callback, max_gen_len=total_gen_len, **self.generation_params)
|
362 |
+
|
363 |
+
else:
|
364 |
+
# now this gets a bit messier, we need to handle prompts,
|
365 |
+
# melody conditioning etc.
|
366 |
+
ref_wavs = [attr.wav['self_wav'] for attr in attributes]
|
367 |
+
all_tokens = []
|
368 |
+
if prompt_tokens is None:
|
369 |
+
prompt_length = 0
|
370 |
+
else:
|
371 |
+
all_tokens.append(prompt_tokens)
|
372 |
+
prompt_length = prompt_tokens.shape[-1]
|
373 |
+
|
374 |
+
stride_tokens = int(self.frame_rate * self.extend_stride)
|
375 |
+
|
376 |
+
while current_gen_offset + prompt_length < total_gen_len:
|
377 |
+
time_offset = current_gen_offset / self.frame_rate
|
378 |
+
chunk_duration = min(self.duration - time_offset, self.max_duration)
|
379 |
+
max_gen_len = int(chunk_duration * self.frame_rate)
|
380 |
+
for attr, ref_wav in zip(attributes, ref_wavs):
|
381 |
+
wav_length = ref_wav.length.item()
|
382 |
+
if wav_length == 0:
|
383 |
+
continue
|
384 |
+
# We will extend the wav periodically if it not long enough.
|
385 |
+
# we have to do it here rather than in conditioners.py as otherwise
|
386 |
+
# we wouldn't have the full wav.
|
387 |
+
initial_position = int(time_offset * self.sample_rate)
|
388 |
+
wav_target_length = int(self.max_duration * self.sample_rate)
|
389 |
+
print(initial_position / self.sample_rate, wav_target_length / self.sample_rate)
|
390 |
+
positions = torch.arange(initial_position,
|
391 |
+
initial_position + wav_target_length, device=self.device)
|
392 |
+
attr.wav['self_wav'] = WavCondition(
|
393 |
+
ref_wav[0][:, positions % wav_length],
|
394 |
+
torch.full_like(ref_wav[1], wav_target_length))
|
395 |
+
with self.autocast:
|
396 |
+
gen_tokens = self.lm.generate(
|
397 |
+
prompt_tokens, attributes,
|
398 |
+
callback=callback, max_gen_len=max_gen_len, **self.generation_params)
|
399 |
+
if prompt_tokens is None:
|
400 |
+
all_tokens.append(gen_tokens)
|
401 |
+
else:
|
402 |
+
all_tokens.append(gen_tokens[:, :, prompt_tokens.shape[-1]:])
|
403 |
+
prompt_tokens = gen_tokens[:, :, stride_tokens:]
|
404 |
+
prompt_length = prompt_tokens.shape[-1]
|
405 |
+
current_gen_offset += stride_tokens
|
406 |
+
|
407 |
+
gen_tokens = torch.cat(all_tokens, dim=-1)
|
408 |
|
409 |
# generate audio
|
410 |
assert gen_tokens.dim() == 3
|
audiocraft/modules/transformer.py
CHANGED
@@ -25,6 +25,22 @@ from xformers import ops
|
|
25 |
from .rope import RotaryEmbedding
|
26 |
from .streaming import StreamingModule
|
27 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
28 |
|
29 |
def _is_profiled() -> bool:
|
30 |
# Return true if we are currently running with a xformers profiler activated.
|
@@ -75,14 +91,22 @@ def create_sin_embedding(positions: torch.Tensor, dim: int, max_period: float =
|
|
75 |
|
76 |
def expand_repeated_kv(x: torch.Tensor, n_rep: int) -> torch.Tensor:
|
77 |
"""torch.repeat_interleave(x, dim=2, repeats=n_rep) from xlformers"""
|
78 |
-
bs, slen, n_kv_heads, head_dim = x.shape
|
79 |
if n_rep == 1:
|
80 |
return x
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
86 |
|
87 |
|
88 |
class LayerScale(nn.Module):
|
@@ -210,6 +234,7 @@ class StreamingMultiheadAttention(StreamingModule):
|
|
210 |
# Return a causal mask, accounting for potentially stored past keys/values
|
211 |
# We actually return a bias for the attention score, as this has the same
|
212 |
# convention both in the builtin MHA in Pytorch, and Xformers functions.
|
|
|
213 |
if self.memory_efficient:
|
214 |
from xformers.ops import LowerTriangularMask
|
215 |
if current_steps == 1:
|
@@ -222,7 +247,7 @@ class StreamingMultiheadAttention(StreamingModule):
|
|
222 |
return LowerTriangularMask()
|
223 |
if self._streaming_state:
|
224 |
past_keys = self._streaming_state['past_keys']
|
225 |
-
past_steps = past_keys.shape[
|
226 |
else:
|
227 |
past_steps = 0
|
228 |
|
@@ -239,6 +264,7 @@ class StreamingMultiheadAttention(StreamingModule):
|
|
239 |
torch.full([], float('-inf'), device=device, dtype=dtype))
|
240 |
|
241 |
def _complete_kv(self, k, v):
|
|
|
242 |
if self.cross_attention:
|
243 |
# With cross attention we assume all keys and values
|
244 |
# are already available, and streaming is with respect
|
@@ -247,20 +273,20 @@ class StreamingMultiheadAttention(StreamingModule):
|
|
247 |
# Complete the key/value pair using the streaming state.
|
248 |
if self._streaming_state:
|
249 |
pk = self._streaming_state['past_keys']
|
250 |
-
nk = torch.cat([pk, k], dim=
|
251 |
if v is k:
|
252 |
nv = nk
|
253 |
else:
|
254 |
pv = self._streaming_state['past_values']
|
255 |
-
nv = torch.cat([pv, v], dim=
|
256 |
else:
|
257 |
nk = k
|
258 |
nv = v
|
259 |
|
260 |
-
assert nk.shape[
|
261 |
offset = 0
|
262 |
if self.past_context is not None:
|
263 |
-
offset = max(0, nk.shape[
|
264 |
if self._is_streaming:
|
265 |
self._streaming_state['past_keys'] = nk[:, offset:]
|
266 |
if v is not k:
|
@@ -272,6 +298,8 @@ class StreamingMultiheadAttention(StreamingModule):
|
|
272 |
return nk, nv
|
273 |
|
274 |
def _apply_rope(self, query: torch.Tensor, key: torch.Tensor):
|
|
|
|
|
275 |
# Apply rope embeddings to query and key tensors.
|
276 |
assert self.rope is not None
|
277 |
if 'past_keys' in self._streaming_state:
|
@@ -292,6 +320,11 @@ class StreamingMultiheadAttention(StreamingModule):
|
|
292 |
assert not is_causal, ("new param added in torch 2.0.1 not supported, "
|
293 |
"use the causal args in the constructor.")
|
294 |
|
|
|
|
|
|
|
|
|
|
|
295 |
dtype = query.dtype
|
296 |
if self._is_streaming:
|
297 |
assert self.causal or self.cross_attention, \
|
@@ -324,8 +357,7 @@ class StreamingMultiheadAttention(StreamingModule):
|
|
324 |
if self.qk_layer_norm is True:
|
325 |
q = self.q_layer_norm(q)
|
326 |
k = self.k_layer_norm(k)
|
327 |
-
|
328 |
-
q, k, v = [rearrange(x, "b t (h d) -> b t h d", h=self.num_heads) for x in [q, k, v]]
|
329 |
else:
|
330 |
if not _is_profiled():
|
331 |
# profiling breaks that propertysomehow.
|
@@ -333,7 +365,11 @@ class StreamingMultiheadAttention(StreamingModule):
|
|
333 |
assert value is key, "specialized implementation"
|
334 |
projected = nn.functional.linear(query, self.in_proj_weight, self.in_proj_bias)
|
335 |
if self.kv_repeat == 1:
|
336 |
-
|
|
|
|
|
|
|
|
|
337 |
q, k, v = ops.unbind(packed, dim=2)
|
338 |
else:
|
339 |
embed_dim = self.embed_dim
|
@@ -344,16 +380,16 @@ class StreamingMultiheadAttention(StreamingModule):
|
|
344 |
end = start + per_head_dim * kv_heads
|
345 |
k = projected[:, :, start: end]
|
346 |
v = projected[:, :, end:]
|
347 |
-
q = rearrange(q, "b t (h d) ->
|
348 |
-
k = rearrange(k, "b t (h d) ->
|
349 |
-
v = rearrange(v, "b t (h d) ->
|
350 |
|
351 |
if self.qk_layer_norm is True:
|
352 |
assert self.kv_repeat == 1
|
353 |
-
q, k = [rearrange(x, "
|
354 |
q = self.q_layer_norm(q)
|
355 |
k = self.k_layer_norm(k)
|
356 |
-
q, k = [rearrange(x, "b t (h d) ->
|
357 |
if self.rope:
|
358 |
q, k = self._apply_rope(q, k)
|
359 |
k, v = self._complete_kv(k, v)
|
@@ -364,7 +400,11 @@ class StreamingMultiheadAttention(StreamingModule):
|
|
364 |
q, k, v = [x.float() for x in [q, k, v]]
|
365 |
if self.memory_efficient:
|
366 |
p = self.dropout if self.training else 0
|
367 |
-
|
|
|
|
|
|
|
|
|
368 |
else:
|
369 |
# We include the dot product as float32, for consistency
|
370 |
# with the other implementations that include that step
|
@@ -374,18 +414,21 @@ class StreamingMultiheadAttention(StreamingModule):
|
|
374 |
# extend a bit the range of operations done in float32,
|
375 |
# although this should make no difference.
|
376 |
q = q / q.shape[-1] ** 0.5
|
|
|
|
|
377 |
if self._is_streaming and self.safe_streaming and q.device.type == 'cuda':
|
378 |
with torch.autocast(device_type=q.device.type, dtype=torch.float32):
|
379 |
-
pre_w = torch.einsum("
|
380 |
else:
|
381 |
-
pre_w = torch.einsum("
|
382 |
if attn_mask is not None:
|
383 |
pre_w = pre_w + attn_mask
|
384 |
w = torch.softmax(pre_w, dim=-1)
|
385 |
w = F.dropout(w, self.dropout, training=self.training).to(v)
|
386 |
-
|
|
|
387 |
x = x.to(dtype)
|
388 |
-
x = rearrange(x, "
|
389 |
x = self.out_proj(x)
|
390 |
else:
|
391 |
key, value = self._complete_kv(key, value)
|
|
|
25 |
from .rope import RotaryEmbedding
|
26 |
from .streaming import StreamingModule
|
27 |
|
28 |
+
_efficient_attention_backend: str = 'torch'
|
29 |
+
|
30 |
+
|
31 |
+
def set_efficient_attention_backend(backend: str = 'torch'):
|
32 |
+
# Using torch by default, it seems a bit faster on older P100 GPUs (~20% faster).
|
33 |
+
global _efficient_attention_backend
|
34 |
+
assert _efficient_attention_backend in ['xformers', 'torch']
|
35 |
+
_efficient_attention_backend = backend
|
36 |
+
|
37 |
+
|
38 |
+
def _get_attention_time_dimension() -> int:
|
39 |
+
if _efficient_attention_backend == 'torch':
|
40 |
+
return 2
|
41 |
+
else:
|
42 |
+
return 1
|
43 |
+
|
44 |
|
45 |
def _is_profiled() -> bool:
|
46 |
# Return true if we are currently running with a xformers profiler activated.
|
|
|
91 |
|
92 |
def expand_repeated_kv(x: torch.Tensor, n_rep: int) -> torch.Tensor:
|
93 |
"""torch.repeat_interleave(x, dim=2, repeats=n_rep) from xlformers"""
|
|
|
94 |
if n_rep == 1:
|
95 |
return x
|
96 |
+
if _efficient_attention_backend == 'torch':
|
97 |
+
bs, n_kv_heads, slen, head_dim = x.shape
|
98 |
+
return (
|
99 |
+
x[:, :, None, :, :]
|
100 |
+
.expand(bs, n_kv_heads, n_rep, slen, head_dim)
|
101 |
+
.reshape(bs, n_kv_heads * n_rep, slen, head_dim)
|
102 |
+
)
|
103 |
+
else:
|
104 |
+
bs, slen, n_kv_heads, head_dim = x.shape
|
105 |
+
return (
|
106 |
+
x[:, :, :, None, :]
|
107 |
+
.expand(bs, slen, n_kv_heads, n_rep, head_dim)
|
108 |
+
.reshape(bs, slen, n_kv_heads * n_rep, head_dim)
|
109 |
+
)
|
110 |
|
111 |
|
112 |
class LayerScale(nn.Module):
|
|
|
234 |
# Return a causal mask, accounting for potentially stored past keys/values
|
235 |
# We actually return a bias for the attention score, as this has the same
|
236 |
# convention both in the builtin MHA in Pytorch, and Xformers functions.
|
237 |
+
time_dim = _get_attention_time_dimension()
|
238 |
if self.memory_efficient:
|
239 |
from xformers.ops import LowerTriangularMask
|
240 |
if current_steps == 1:
|
|
|
247 |
return LowerTriangularMask()
|
248 |
if self._streaming_state:
|
249 |
past_keys = self._streaming_state['past_keys']
|
250 |
+
past_steps = past_keys.shape[time_dim]
|
251 |
else:
|
252 |
past_steps = 0
|
253 |
|
|
|
264 |
torch.full([], float('-inf'), device=device, dtype=dtype))
|
265 |
|
266 |
def _complete_kv(self, k, v):
|
267 |
+
time_dim = _get_attention_time_dimension()
|
268 |
if self.cross_attention:
|
269 |
# With cross attention we assume all keys and values
|
270 |
# are already available, and streaming is with respect
|
|
|
273 |
# Complete the key/value pair using the streaming state.
|
274 |
if self._streaming_state:
|
275 |
pk = self._streaming_state['past_keys']
|
276 |
+
nk = torch.cat([pk, k], dim=time_dim)
|
277 |
if v is k:
|
278 |
nv = nk
|
279 |
else:
|
280 |
pv = self._streaming_state['past_values']
|
281 |
+
nv = torch.cat([pv, v], dim=time_dim)
|
282 |
else:
|
283 |
nk = k
|
284 |
nv = v
|
285 |
|
286 |
+
assert nk.shape[time_dim] == nv.shape[time_dim]
|
287 |
offset = 0
|
288 |
if self.past_context is not None:
|
289 |
+
offset = max(0, nk.shape[time_dim] - self.past_context)
|
290 |
if self._is_streaming:
|
291 |
self._streaming_state['past_keys'] = nk[:, offset:]
|
292 |
if v is not k:
|
|
|
298 |
return nk, nv
|
299 |
|
300 |
def _apply_rope(self, query: torch.Tensor, key: torch.Tensor):
|
301 |
+
# TODO: fix and verify layout.
|
302 |
+
assert _efficient_attention_backend == 'xformers', 'Rope not supported with torch attn.'
|
303 |
# Apply rope embeddings to query and key tensors.
|
304 |
assert self.rope is not None
|
305 |
if 'past_keys' in self._streaming_state:
|
|
|
320 |
assert not is_causal, ("new param added in torch 2.0.1 not supported, "
|
321 |
"use the causal args in the constructor.")
|
322 |
|
323 |
+
time_dim = _get_attention_time_dimension()
|
324 |
+
if time_dim == 2:
|
325 |
+
layout = "b h t d"
|
326 |
+
else:
|
327 |
+
layout = "b t h d"
|
328 |
dtype = query.dtype
|
329 |
if self._is_streaming:
|
330 |
assert self.causal or self.cross_attention, \
|
|
|
357 |
if self.qk_layer_norm is True:
|
358 |
q = self.q_layer_norm(q)
|
359 |
k = self.k_layer_norm(k)
|
360 |
+
q, k, v = [rearrange(x, f"b t (h d) -> {layout}", h=self.num_heads) for x in [q, k, v]]
|
|
|
361 |
else:
|
362 |
if not _is_profiled():
|
363 |
# profiling breaks that propertysomehow.
|
|
|
365 |
assert value is key, "specialized implementation"
|
366 |
projected = nn.functional.linear(query, self.in_proj_weight, self.in_proj_bias)
|
367 |
if self.kv_repeat == 1:
|
368 |
+
if time_dim == 2:
|
369 |
+
bound_layout = "b h p t d"
|
370 |
+
else:
|
371 |
+
bound_layout = "b t p h d"
|
372 |
+
packed = rearrange(projected, f"b t (p h d) -> {bound_layout}", p=3, h=self.num_heads)
|
373 |
q, k, v = ops.unbind(packed, dim=2)
|
374 |
else:
|
375 |
embed_dim = self.embed_dim
|
|
|
380 |
end = start + per_head_dim * kv_heads
|
381 |
k = projected[:, :, start: end]
|
382 |
v = projected[:, :, end:]
|
383 |
+
q = rearrange(q, f"b t (h d) -> {layout}", h=self.num_heads)
|
384 |
+
k = rearrange(k, f"b t (h d) -> {layout}", h=kv_heads)
|
385 |
+
v = rearrange(v, f"b t (h d) -> {layout}", h=kv_heads)
|
386 |
|
387 |
if self.qk_layer_norm is True:
|
388 |
assert self.kv_repeat == 1
|
389 |
+
q, k = [rearrange(x, f"{layout} -> b t (h d)") for x in [q, k]]
|
390 |
q = self.q_layer_norm(q)
|
391 |
k = self.k_layer_norm(k)
|
392 |
+
q, k = [rearrange(x, f"b t (h d) -> {layout}", h=self.num_heads) for x in [q, k]]
|
393 |
if self.rope:
|
394 |
q, k = self._apply_rope(q, k)
|
395 |
k, v = self._complete_kv(k, v)
|
|
|
400 |
q, k, v = [x.float() for x in [q, k, v]]
|
401 |
if self.memory_efficient:
|
402 |
p = self.dropout if self.training else 0
|
403 |
+
if _efficient_attention_backend == 'torch':
|
404 |
+
x = torch.nn.functional.scaled_dot_product_attention(
|
405 |
+
q, k, v, is_causal=attn_mask is not None, dropout_p=p)
|
406 |
+
else:
|
407 |
+
x = ops.memory_efficient_attention(q, k, v, attn_mask, p=p)
|
408 |
else:
|
409 |
# We include the dot product as float32, for consistency
|
410 |
# with the other implementations that include that step
|
|
|
414 |
# extend a bit the range of operations done in float32,
|
415 |
# although this should make no difference.
|
416 |
q = q / q.shape[-1] ** 0.5
|
417 |
+
key_layout = layout.replace('t', 'k')
|
418 |
+
query_layout = layout
|
419 |
if self._is_streaming and self.safe_streaming and q.device.type == 'cuda':
|
420 |
with torch.autocast(device_type=q.device.type, dtype=torch.float32):
|
421 |
+
pre_w = torch.einsum(f"{query_layout},{key_layout}-> b h t k", q, k)
|
422 |
else:
|
423 |
+
pre_w = torch.einsum(f"{query_layout},{key_layout}-> b h t k", q, k)
|
424 |
if attn_mask is not None:
|
425 |
pre_w = pre_w + attn_mask
|
426 |
w = torch.softmax(pre_w, dim=-1)
|
427 |
w = F.dropout(w, self.dropout, training=self.training).to(v)
|
428 |
+
# Key and value have the same format.
|
429 |
+
x = torch.einsum(f"b h t k, {key_layout} -> {layout}", w, v)
|
430 |
x = x.to(dtype)
|
431 |
+
x = rearrange(x, f"{layout} -> b t (h d)", h=self.num_heads)
|
432 |
x = self.out_proj(x)
|
433 |
else:
|
434 |
key, value = self._complete_kv(key, value)
|