174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429 | class FlaxT5Attention(BaseJAXAttentionModule):
config: T5Config
has_relative_attention_bias: bool = False
causal: bool = False
dtype: jnp.dtype = jnp.bfloat16 # the dtype of the computation
def setup(self):
self.relative_attention_num_buckets = self.config.relative_attention_num_buckets
self.relative_attention_max_distance = self.config.relative_attention_max_distance
self.d_model = self.config.d_model
self.key_value_proj_dim = self.config.d_kv
self.n_heads = self.config.num_heads
self.dropout = self.config.dropout_rate
self.inner_dim = self.n_heads * self.key_value_proj_dim
q_init_std = self.config.initializer_factor * ((self.inner_dim * self.key_value_proj_dim) ** -0.5)
kv_init_std = self.config.initializer_factor * (self.inner_dim ** -0.5)
o_init_std = self.config.initializer_factor * (self.inner_dim ** -0.5)
self.q = Linear(
self.inner_dim,
use_bias=False,
kernel_init=jax.nn.initializers.normal(q_init_std),
dtype=self.dtype,
)
self.k = Linear(
self.inner_dim,
use_bias=False,
kernel_init=jax.nn.initializers.normal(kv_init_std),
dtype=self.dtype,
)
self.v = Linear(
self.inner_dim,
use_bias=False,
kernel_init=jax.nn.initializers.normal(kv_init_std),
dtype=self.dtype,
)
self.o = Linear(
self.d_model,
use_bias=False,
kernel_init=jax.nn.initializers.normal(o_init_std),
dtype=self.dtype,
)
if self.has_relative_attention_bias:
self.relative_attention_bias = nn.Embed(
self.relative_attention_num_buckets,
self.n_heads,
embedding_init=jax.nn.initializers.normal(kv_init_std),
dtype=self.dtype,
)
@staticmethod
def _relative_position_bucket(relative_position, bidirectional=True, num_buckets=32, max_distance=128):
relative_buckets = 0
if bidirectional:
num_buckets //= 2
relative_buckets += (relative_position > 0) * num_buckets
relative_position = jnp.abs(relative_position)
else:
relative_position = -jnp.clip(relative_position, a_max=0)
# now relative_position is in the range [0, inf)
# half of the buckets are for exact increments in positions
max_exact = num_buckets // 2
is_small = relative_position < max_exact
relative_position_if_large = max_exact + (
jnp.log(relative_position / max_exact) / jnp.log(max_distance / max_exact) * (num_buckets - max_exact)
)
relative_position_if_large = jnp.clip(relative_position_if_large, a_max=num_buckets - 1)
relative_buckets += jnp.where(is_small, relative_position, relative_position_if_large)
return relative_buckets.astype("i4")
def compute_bias(self, query_length, key_length):
"""Compute binned relative position bias"""
context_position = jnp.arange(query_length, dtype="i4")[:, None]
memory_position = jnp.arange(key_length, dtype="i4")[None, :]
relative_position = memory_position - context_position
relative_position_bucket = self._relative_position_bucket(
relative_position,
bidirectional=(not self.causal),
num_buckets=self.relative_attention_num_buckets,
max_distance=self.relative_attention_max_distance,
)
values = self.relative_attention_bias(relative_position_bucket)
values = values.transpose((2, 0, 1))[None, :, :, :]
return values
def _split_heads(self, hidden_states):
return hidden_states.reshape(hidden_states.shape[:2] + (self.n_heads, self.key_value_proj_dim))
def _merge_heads(self, hidden_states):
return hidden_states.reshape(hidden_states.shape[:2] + (self.inner_dim,))
def _create_position_bias(
self, key_states, query_states, attention_mask, init_cache, seq_length, causal_attention_mask_shift
):
cache_is_filled = self.causal and self.has_variable("cache", "cached_key") and (not init_cache)
key_length = key_states.shape[1]
query_length = key_length if cache_is_filled else query_states.shape[1]
if self.has_relative_attention_bias:
position_bias = self.compute_bias(query_length, key_length)
elif attention_mask is not None:
position_bias = jnp.zeros_like(attention_mask)
else:
position_bias = jnp.zeros((1, self.n_heads, query_length, key_length), dtype=self.dtype)
# if key and values are already calculated, only the last query position bias should be taken
if cache_is_filled:
max_decoder_length = self.variables["cache"]["cached_key"].shape[1]
position_bias = jax.lax.dynamic_slice(
position_bias,
(0, 0, causal_attention_mask_shift, 0),
(1, self.n_heads, seq_length, max_decoder_length),
)
return position_bias
def __call__(
self,
hidden_states,
attention_mask=None,
key_value_states=None,
position_bias=None,
use_cache=False,
output_attentions=False,
deterministic=True,
init_cache=False,
):
batch_size, seq_length = hidden_states.shape[:2]
# q, k, v projections
query_states = self.q(hidden_states) # (batch_size, n_heads, seq_length, dim_per_head)
key_states = self.k(hidden_states) if key_value_states is None else self.k(key_value_states)
value_states = self.v(hidden_states) if key_value_states is None else self.v(key_value_states)
# reshape to (batch_size, seq_length, n_heads, head_dim)
query_states = self._split_heads(query_states)
key_states = self._split_heads(key_states)
value_states = self._split_heads(value_states)
# if self.config.use_sharding_constraint:
# query_states = with_sharding_constraint(
# query_states, PartitionSpec(("dp", "fsdp"), "sp" if query_states.shape[1] != 1 else None, "tp", None)
# )
# key_states = with_sharding_constraint(
# key_states, PartitionSpec(("dp", "fsdp"), "sp", "tp", None)
# )
# value_states = with_sharding_constraint(
# value_states, PartitionSpec(("dp", "fsdp"), "sp", "tp", None)
# )
# counter-act scaling in dot_product_attention_weights function
query_states *= jnp.sqrt(query_states.shape[-1])
# for fast decoding causal attention mask should be shifted
causal_attention_mask_shift = (
self.variables["cache"]["cache_index"] if (self.has_variable("cache", "cached_key") and self.causal) else 0
)
# create causal attention_mask; attention_mask has to be defined when model is causal
if self.causal:
causal_attention_mask = make_causal_mask(attention_mask, dtype="bool")
# fast decoding for generate requires special attention_mask
if self.has_variable("cache", "cached_key"):
max_decoder_length = self.variables["cache"]["cached_key"].shape[1]
causal_attention_mask = jax.lax.dynamic_slice(
causal_attention_mask,
(0, 0, causal_attention_mask_shift, 0),
(1, 1, seq_length, max_decoder_length),
)
# broadcast causal attention mask & attention mask to fit for merge
causal_attention_mask = jnp.broadcast_to(
causal_attention_mask, (batch_size,) + causal_attention_mask.shape[1:]
)
attention_mask = jnp.broadcast_to(
jnp.expand_dims(attention_mask, axis=(-3, -2)), causal_attention_mask.shape
)
attention_mask = combine_masks(attention_mask, causal_attention_mask)
elif attention_mask is not None:
attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
# During fast autoregressive decoding, we feed one position at a time,
# and cache the keys and values step by step.
if self.causal and (self.has_variable("cache", "cached_key") or init_cache):
key_states, value_states, attention_attention_mask = self._concatenate_to_cache(
key_states, value_states, query_states, attention_mask
)
# replace masked positions with -10_000
if attention_mask is not None:
mask_value = jnp.finfo(self.dtype).min
attention_mask = jax.lax.select(
attention_mask > 0,
jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
jnp.full(attention_mask.shape, mask_value).astype(self.dtype),
)
if position_bias is None:
# compute position bias (only for first layer)
position_bias = self._create_position_bias(
key_states, query_states, attention_mask, init_cache, seq_length, causal_attention_mask_shift
)
if attention_mask is not None:
position_bias = position_bias + attention_mask
# create dropout rng
dropout_rng = None
if not deterministic and self.dropout > 0.0:
dropout_rng = self.make_rng("dropout")
# Softmax(QK^T)
attn_weights = dot_product_attention_weights(
query_states,
key_states,
bias=position_bias,
dropout_rng=dropout_rng,
dropout_rate=self.dropout,
broadcast_dropout=True,
deterministic=deterministic,
dtype=self.dtype,
)
attn_weights = with_sharding_constraint(attn_weights, PartitionSpec(
("dp", "fsdp"), "sp" if query_states.shape[1] != 1 else None, None, None
))
# multiply with value states
attn_output = jnp.einsum("...hqk,...khd->...qhd", attn_weights, value_states)
# bring back to (batch_size, seq_length, d_model)
attn_output = self._merge_heads(attn_output)
if self.config.shard_attention_computation:
attn_output = with_sharding_constraint(
attn_output, PartitionSpec(
("dp", "fsdp"),
"sp" if attn_output.shape[1] != 1 else None,
"tp"
)
)
# apply output matrix
attn_output = self.o(attn_output)
outputs = (attn_output, position_bias)
if output_attentions:
outputs = outputs + (attn_weights,)
return outputs
|