Skip to content

modules.mistral.modelling_mistral_flax

FlaxMistralAttention

Bases: BaseJAXAttentionModule

Source code in src/python/easydel/modules/mistral/modelling_mistral_flax.py
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
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
430
431
432
433
434
435
436
437
438
439
440
441
class FlaxMistralAttention(BaseJAXAttentionModule):
    config: MistralConfig
    dtype: jnp.dtype = jnp.float32
    param_dtype: jnp.dtype = jnp.float32
    precision: Optional[Union[jax.lax.Precision, str]] = None

    def setup(self):
        config = self.config
        self.hidden_size = config.hidden_size
        self.head_dim = self.config.hidden_size // self.config.num_attention_heads
        self.num_key_value_groups = self.config.num_attention_heads // self.config.num_key_value_heads

        if self.num_key_value_groups == 1:
            assert self.config.num_attention_heads == self.config.num_key_value_heads
        self.q_proj = Linear(
            config.num_attention_heads * self.head_dim,
            dtype=self.dtype,
            param_dtype=self.param_dtype,
            use_bias=self.config.attention_bias,
            kernel_init=jax.nn.initializers.normal(
                self.config.initializer_range),
            precision=self.precision,
            **get_dot_general_by_bits(self.config.bits, self.config.easy_method)
        )
        self.k_proj = Linear(
            config.num_key_value_heads * self.head_dim,
            dtype=self.dtype,
            param_dtype=self.param_dtype,
            use_bias=self.config.attention_bias,
            kernel_init=jax.nn.initializers.normal(
                self.config.initializer_range),
            precision=self.precision,
            **get_dot_general_by_bits(self.config.bits, self.config.easy_method)
        )
        self.v_proj = Linear(
            config.num_key_value_heads * self.head_dim,
            dtype=self.dtype,
            param_dtype=self.param_dtype,
            use_bias=self.config.attention_bias,
            kernel_init=jax.nn.initializers.normal(
                self.config.initializer_range),
            precision=self.precision,
            **get_dot_general_by_bits(self.config.bits, self.config.easy_method)
        )
        self.o_proj = Linear(
            config.hidden_size,
            dtype=self.dtype,
            param_dtype=self.param_dtype,
            use_bias=False,
            kernel_init=jax.nn.initializers.normal(
                self.config.initializer_range),
            precision=self.precision,
            **get_dot_general_by_bits(self.config.bits, self.config.easy_method)
        )

        self.rotary = FlaxMistralRotaryEmbedding(self.dtype)
        self.attention_performer = AttentionModule(
            use_sharding_constraint=self.config.use_sharding_constraint,
            block_k_major=self.config.block_k_major,
            block_b=self.config.block_b,
            block_q=self.config.block_q,
            block_k=self.config.block_k,
            block_q_major_dkv=self.config.block_q_major_dkv,
            block_k_major_dkv=self.config.block_k_major_dkv,
            block_k_major_dq=self.config.block_k_major_dq,
            block_k_dkv=self.config.block_k_dkv,
            block_q_dkv=self.config.block_q_dkv,
            block_q_dq=self.config.block_q_dq,
            block_k_dq=self.config.block_k_dq,
            num_attention_heads=self.config.num_attention_heads,
            attention_dropout=self.config.attention_dropout,
            head_dims=self.head_dim,
            attention_partition_spec=self.config.attention_partition_spec,
            shard_attention_computation=self.config.shard_attention_computation,
            precision=self.precision,
            force_float32_tpu=True,
            attn_mechanism=self.config.attn_mechanism,
            dtype=self.dtype,
            bias_partition_spec=self.config.bias_partition_spec,
            key_partition_spec=self.config.key_partition_spec,
            query_partition_spec=self.config.query_partition_spec,
            generation_query_partition_spec=self.config.generation_query_partition_spec,
            generation_bias_partition_spec=self.config.generation_bias_partition_spec,
            generation_attention_partition_spec=self.config.generation_attention_partition_spec,
            value_partition_spec=self.config.value_partition_spec,
            scan_ring_attention=self.config.scan_ring_attention,
            mesh=self.config.jax_mesh(),
            sm_scale=1 / math.sqrt(self.head_dim),
            axis_name=self.config.attention_axis_name
        )

    def _merge_heads(self, hidden_states):
        return hidden_states.reshape(hidden_states.shape[:2] + (self.hidden_size,))

    @staticmethod
    def _transpose_sequence_head(query, key, value):
        """
        The _transpose_sequence_head function transposes the query, key and value matrices.

        :param query: Get the attention weights for each of the heads
        :param key: Determine the number of heads
        :param value: Store the values of the input
        :return: The transpose of the query, key and value matrices

        """
        return jnp.transpose(query, (0, 2, 1, 3)), jnp.transpose(key, (0, 2, 1, 3)), jnp.transpose(value, (0, 2, 1, 3))

    def apply_rotary(self, batch_size, sequence_length, query, key, value, freq_cis, position_ids):
        """
        The apply_rotary function is a modified version of the apply_attention function in the BertModel class.
        The main difference is that it takes in an additional argument, freq_cis, which are used to calculate
        the rotary attention weights. The other differences are minor and mostly related to reshaping tensors.

        :param self: Access variables that belong to the class
        :param batch_size: Reshape the query, key and value tensors
        :param sequence_length: Reshape the query, key and value tensors
        :param query: Calculate the attention weights
        :param key: Calculate the attention
        :param value: Compute the attention weights
        :param freq_cis: Calculate the frequency of each word in the vocabulary
        :param position_ids: Identify the position of each token in the sequence
        :return: A tuple of 3 tensors: query, key and value

        """
        query = query.reshape(
            batch_size,
            sequence_length,
            self.config.num_attention_heads,
            self.head_dim
        )
        key = key.reshape(
            batch_size,
            sequence_length,
            self.config.num_key_value_heads,
            self.head_dim
        )
        value = value.reshape(
            batch_size,
            sequence_length,
            self.config.num_key_value_heads,
            self.head_dim
        )

        query, key, value = self._transpose_sequence_head(query, key, value)
        query, key = self.rotary(
            position_ids=position_ids, query=query, key=key, freq_cis=freq_cis
        )
        key = repeat_kv_bnsh(key, self.num_key_value_groups)
        value = repeat_kv_bnsh(value, self.num_key_value_groups)
        return self._transpose_sequence_head(query, key, value)

    def __call__(
            self,
            hidden_states: chex.Array,
            freq_cis: Tuple[chex.Array, chex.Array],
            attention_mask: chex.Array,
            position_ids: chex.Array,
            causal_mask: chex.Array,
            segment_ids: Optional[chex.Array] = None,
            deterministic: bool = True,
            init_cache: bool = False,
            output_attentions: bool = False,
            fcm_mask=None,
    ):
        """

        The __call__ function is the main function of a JAX module. It defines how the module behaves when called
        with inputs. The __call__ function can be thought of as a "forward pass" through the model,
        and it should return all outputs that are needed for training or inference.

        :param self: Access variables that belong to the class
        :param hidden_states: chex.Array: Pass the hidden states of the previous layer
        :param freq_cis: Tuple[chex.Array, chex.Array],: Pass in the frequency coefficients for each position
        :param attention_mask: chex.Array: Mask out certain tokens in the input sequence
        :param position_ids: chex.Array: Determine the position of each token in a sequence
        :param causal_mask: chex.Array: Mask out the future tokens in the decoder
        :param deterministic: bool: Determine whether to use dropout or not
        :param init_cache: bool: Initialize the cache
        :param output_attentions: bool: Determine whether to return the attention weights or not
        :param fcm_mask: Mask out the attention weights between the input and output tokens
        :param : Determine if the attention is causal or not
        :return: A tuple of two arrays

        """
        batch_size, sequence_length = hidden_states.shape[:2]
        query_states, key_states, value_states = self.q_proj(hidden_states), self.k_proj(hidden_states), self.v_proj(
            hidden_states)

        query_states = query_states.reshape(
            batch_size, sequence_length, self.config.num_attention_heads, self.head_dim)
        key_states = key_states.reshape(
            batch_size, sequence_length, self.config.num_key_value_heads, self.head_dim)
        value_states = value_states.reshape(
            batch_size, sequence_length, self.config.num_key_value_heads, self.head_dim)

        query_states, key_states, value_states = self.apply_rotary(
            query=query_states,
            key=key_states,
            value=value_states,
            position_ids=position_ids,
            freq_cis=freq_cis,
            batch_size=batch_size,
            sequence_length=sequence_length
        )

        assert_msg = (
            "num_attention_heads repeat wont work likely\n"
            f"INFO :\n\trepeat_kv_bnsh Used with num_key_value_groups = {self.num_key_value_groups}\n\t"
            f"NH : {self.config.num_attention_heads} KVH : {self.config.num_attention_heads}"
        )

        assert query_states.shape[-2] == self.config.num_attention_heads, assert_msg
        assert key_states.shape[-2] == self.config.num_attention_heads, assert_msg
        assert value_states.shape[-2] == self.config.num_attention_heads, assert_msg

        query_length, key_length = query_states.shape[1], key_states.shape[1]

        if self.has_variable("cache", "cached_key"):
            mask_shift = self.variables["cache"]["cache_index"]
            max_decoder_length = self.variables["cache"]["cached_key"].shape[1]
            causal_mask = lax.dynamic_slice(
                causal_mask, (0, 0, mask_shift, 0), (1, 1,
                                                     query_length, max_decoder_length)
            )
        else:
            causal_mask = causal_mask[:, :, :query_length, :key_length]

        batch_size = hidden_states.shape[0]
        causal_mask = jnp.broadcast_to(
            causal_mask, (batch_size,) + causal_mask.shape[1:])
        if attention_mask.ndim == 2:
            attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
        attention_mask = jnp.broadcast_to(
            attention_mask, causal_mask.shape
        )
        attention_mask = combine_masks(attention_mask, causal_mask, fcm_mask)

        dropout_rng = None

        if not deterministic and self.config.attention_dropout > 0.0:
            dropout_rng = self.make_rng("dropout")
        if self.has_variable("cache", "cached_key") or init_cache:
            key_states, value_states, attention_mask = self._concatenate_to_cache(
                key_states,
                value_states,
                query_states,
                attention_mask
            )
        # 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)
        #     )
        attention_bias = lax.select(
            attention_mask > 0,
            jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
            jnp.full(attention_mask.shape, jnp.finfo(
                self.dtype).min).astype(self.dtype),
        )

        query_length, key_length = query_states.shape[1], key_states.shape[1]

        attentions = self.attention_performer.__call__(
            query_states=query_states,
            key_states=key_states,
            value_states=value_states,
            bias=attention_bias,
            attention_mask=attention_mask,
            causal=True,
            dropout_rng=dropout_rng,
            deterministic=deterministic,
            query_sequence_length=query_length,
            key_value_sequence_length=key_length,
            uses_cache=self.has_variable("cache", "cached_key") or init_cache,
            segment_ids=segment_ids,
            causal_mask=causal_mask
        )

        attn_output = self._merge_heads(attentions.attention_outputs)
        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"
                )
            )
        attn_output = self.o_proj(attn_output)

        outputs = (
            attn_output, attentions.attention_weights
        ) if output_attentions else (
            attn_output,
        )
        return outputs

__call__(hidden_states, freq_cis, attention_mask, position_ids, causal_mask, segment_ids=None, deterministic=True, init_cache=False, output_attentions=False, fcm_mask=None)

The call function is the main function of a JAX module. It defines how the module behaves when called with inputs. The call function can be thought of as a "forward pass" through the model, and it should return all outputs that are needed for training or inference.

Parameters:

Name Type Description Default
self

Access variables that belong to the class

required
hidden_states Array

chex.Array: Pass the hidden states of the previous layer

required
freq_cis Tuple[Array, Array]

Tuple[chex.Array, chex.Array],: Pass in the frequency coefficients for each position

required
attention_mask Array

chex.Array: Mask out certain tokens in the input sequence

required
position_ids Array

chex.Array: Determine the position of each token in a sequence

required
causal_mask Array

chex.Array: Mask out the future tokens in the decoder

required
deterministic bool

bool: Determine whether to use dropout or not

True
init_cache bool

bool: Initialize the cache

False
output_attentions bool

bool: Determine whether to return the attention weights or not

False
fcm_mask

Mask out the attention weights between the input and output tokens

None

Determine if the attention is causal or not

required

Returns:

Type Description

A tuple of two arrays

Source code in src/python/easydel/modules/mistral/modelling_mistral_flax.py
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
430
431
432
433
434
435
436
437
438
439
440
441
def __call__(
        self,
        hidden_states: chex.Array,
        freq_cis: Tuple[chex.Array, chex.Array],
        attention_mask: chex.Array,
        position_ids: chex.Array,
        causal_mask: chex.Array,
        segment_ids: Optional[chex.Array] = None,
        deterministic: bool = True,
        init_cache: bool = False,
        output_attentions: bool = False,
        fcm_mask=None,
):
    """

    The __call__ function is the main function of a JAX module. It defines how the module behaves when called
    with inputs. The __call__ function can be thought of as a "forward pass" through the model,
    and it should return all outputs that are needed for training or inference.

    :param self: Access variables that belong to the class
    :param hidden_states: chex.Array: Pass the hidden states of the previous layer
    :param freq_cis: Tuple[chex.Array, chex.Array],: Pass in the frequency coefficients for each position
    :param attention_mask: chex.Array: Mask out certain tokens in the input sequence
    :param position_ids: chex.Array: Determine the position of each token in a sequence
    :param causal_mask: chex.Array: Mask out the future tokens in the decoder
    :param deterministic: bool: Determine whether to use dropout or not
    :param init_cache: bool: Initialize the cache
    :param output_attentions: bool: Determine whether to return the attention weights or not
    :param fcm_mask: Mask out the attention weights between the input and output tokens
    :param : Determine if the attention is causal or not
    :return: A tuple of two arrays

    """
    batch_size, sequence_length = hidden_states.shape[:2]
    query_states, key_states, value_states = self.q_proj(hidden_states), self.k_proj(hidden_states), self.v_proj(
        hidden_states)

    query_states = query_states.reshape(
        batch_size, sequence_length, self.config.num_attention_heads, self.head_dim)
    key_states = key_states.reshape(
        batch_size, sequence_length, self.config.num_key_value_heads, self.head_dim)
    value_states = value_states.reshape(
        batch_size, sequence_length, self.config.num_key_value_heads, self.head_dim)

    query_states, key_states, value_states = self.apply_rotary(
        query=query_states,
        key=key_states,
        value=value_states,
        position_ids=position_ids,
        freq_cis=freq_cis,
        batch_size=batch_size,
        sequence_length=sequence_length
    )

    assert_msg = (
        "num_attention_heads repeat wont work likely\n"
        f"INFO :\n\trepeat_kv_bnsh Used with num_key_value_groups = {self.num_key_value_groups}\n\t"
        f"NH : {self.config.num_attention_heads} KVH : {self.config.num_attention_heads}"
    )

    assert query_states.shape[-2] == self.config.num_attention_heads, assert_msg
    assert key_states.shape[-2] == self.config.num_attention_heads, assert_msg
    assert value_states.shape[-2] == self.config.num_attention_heads, assert_msg

    query_length, key_length = query_states.shape[1], key_states.shape[1]

    if self.has_variable("cache", "cached_key"):
        mask_shift = self.variables["cache"]["cache_index"]
        max_decoder_length = self.variables["cache"]["cached_key"].shape[1]
        causal_mask = lax.dynamic_slice(
            causal_mask, (0, 0, mask_shift, 0), (1, 1,
                                                 query_length, max_decoder_length)
        )
    else:
        causal_mask = causal_mask[:, :, :query_length, :key_length]

    batch_size = hidden_states.shape[0]
    causal_mask = jnp.broadcast_to(
        causal_mask, (batch_size,) + causal_mask.shape[1:])
    if attention_mask.ndim == 2:
        attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
    attention_mask = jnp.broadcast_to(
        attention_mask, causal_mask.shape
    )
    attention_mask = combine_masks(attention_mask, causal_mask, fcm_mask)

    dropout_rng = None

    if not deterministic and self.config.attention_dropout > 0.0:
        dropout_rng = self.make_rng("dropout")
    if self.has_variable("cache", "cached_key") or init_cache:
        key_states, value_states, attention_mask = self._concatenate_to_cache(
            key_states,
            value_states,
            query_states,
            attention_mask
        )
    # 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)
    #     )
    attention_bias = lax.select(
        attention_mask > 0,
        jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
        jnp.full(attention_mask.shape, jnp.finfo(
            self.dtype).min).astype(self.dtype),
    )

    query_length, key_length = query_states.shape[1], key_states.shape[1]

    attentions = self.attention_performer.__call__(
        query_states=query_states,
        key_states=key_states,
        value_states=value_states,
        bias=attention_bias,
        attention_mask=attention_mask,
        causal=True,
        dropout_rng=dropout_rng,
        deterministic=deterministic,
        query_sequence_length=query_length,
        key_value_sequence_length=key_length,
        uses_cache=self.has_variable("cache", "cached_key") or init_cache,
        segment_ids=segment_ids,
        causal_mask=causal_mask
    )

    attn_output = self._merge_heads(attentions.attention_outputs)
    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"
            )
        )
    attn_output = self.o_proj(attn_output)

    outputs = (
        attn_output, attentions.attention_weights
    ) if output_attentions else (
        attn_output,
    )
    return outputs

apply_rotary(batch_size, sequence_length, query, key, value, freq_cis, position_ids)

The apply_rotary function is a modified version of the apply_attention function in the BertModel class. The main difference is that it takes in an additional argument, freq_cis, which are used to calculate the rotary attention weights. The other differences are minor and mostly related to reshaping tensors.

Parameters:

Name Type Description Default
self

Access variables that belong to the class

required
batch_size

Reshape the query, key and value tensors

required
sequence_length

Reshape the query, key and value tensors

required
query

Calculate the attention weights

required
key

Calculate the attention

required
value

Compute the attention weights

required
freq_cis

Calculate the frequency of each word in the vocabulary

required
position_ids

Identify the position of each token in the sequence

required

Returns:

Type Description

A tuple of 3 tensors: query, key and value

Source code in src/python/easydel/modules/mistral/modelling_mistral_flax.py
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
def apply_rotary(self, batch_size, sequence_length, query, key, value, freq_cis, position_ids):
    """
    The apply_rotary function is a modified version of the apply_attention function in the BertModel class.
    The main difference is that it takes in an additional argument, freq_cis, which are used to calculate
    the rotary attention weights. The other differences are minor and mostly related to reshaping tensors.

    :param self: Access variables that belong to the class
    :param batch_size: Reshape the query, key and value tensors
    :param sequence_length: Reshape the query, key and value tensors
    :param query: Calculate the attention weights
    :param key: Calculate the attention
    :param value: Compute the attention weights
    :param freq_cis: Calculate the frequency of each word in the vocabulary
    :param position_ids: Identify the position of each token in the sequence
    :return: A tuple of 3 tensors: query, key and value

    """
    query = query.reshape(
        batch_size,
        sequence_length,
        self.config.num_attention_heads,
        self.head_dim
    )
    key = key.reshape(
        batch_size,
        sequence_length,
        self.config.num_key_value_heads,
        self.head_dim
    )
    value = value.reshape(
        batch_size,
        sequence_length,
        self.config.num_key_value_heads,
        self.head_dim
    )

    query, key, value = self._transpose_sequence_head(query, key, value)
    query, key = self.rotary(
        position_ids=position_ids, query=query, key=key, freq_cis=freq_cis
    )
    key = repeat_kv_bnsh(key, self.num_key_value_groups)
    value = repeat_kv_bnsh(value, self.num_key_value_groups)
    return self._transpose_sequence_head(query, key, value)

FlaxMistralDecoderLayer

Bases: Module

Source code in src/python/easydel/modules/mistral/modelling_mistral_flax.py
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
class FlaxMistralDecoderLayer(nn.Module):
    config: MistralConfig
    dtype: jnp.dtype = jnp.bfloat16
    param_dtype: jnp.dtype = jnp.bfloat16
    precision: Optional[jax.lax.Precision] = jax.lax.Precision("fastest")

    def setup(self) -> None:
        attn_block = FlaxMistralAttention
        mlp_block = FlaxMistralMLP

        if self.config.gradient_checkpointing != "":
            # hidden_states: chex.Array,
            # freq_cis: Tuple[chex.Array, chex.Array],
            # attention_mask: chex.Array,
            # position_ids: chex.Array,
            # causal_mask: chex.Array,
            # segment_ids: Optional[chex.Array] = None,
            # deterministic: bool = True,
            # init_cache: bool = False,
            # output_attentions: bool = False,
            # fcm_mask = None,
            attn_block = re_mat(
                attn_block,
                policy=get_gradient_checkpoint_policy(self.config.gradient_checkpointing),
                static_argnums=(1, 3, 4, 6, 7, 8)
            )
            mlp_block = re_mat(
                mlp_block,
                policy=get_gradient_checkpoint_policy(self.config.gradient_checkpointing),
                static_argnums=(1,)
            )
        self.self_attn = attn_block(
            config=self.config,
            dtype=self.dtype,
            param_dtype=self.param_dtype,
            precision=self.precision
        )
        self.mlp = mlp_block(
            config=self.config,
            dtype=self.dtype,
            param_dtype=self.param_dtype,
            precision=self.precision
        )
        self.input_layernorm = MistralRMSNorm(
            dim=self.config.hidden_size,
            eps=self.config.rms_norm_eps,
            dtype=self.dtype,
            param_dtype=self.param_dtype
        )
        self.post_attention_layernorm = MistralRMSNorm(
            dim=self.config.hidden_size,
            eps=self.config.rms_norm_eps,
            dtype=self.dtype,
            param_dtype=self.param_dtype
        )

    def __call__(
            self,
            hidden_states: chex.Array,
            freq_cis: Tuple[chex.Array, chex.Array],
            attention_mask: chex.Array,
            causal_mask: chex.Array,
            position_ids: chex.Array,
            segment_ids: Optional[chex.Array] = None,
            deterministic: bool = True,
            init_cache: bool = False,
            output_attentions: bool = True
    ):
        """
        The __call__ function is the main function of a TransformerEncoderLayer.
        It takes in the following arguments:
            hidden_states (chex.Array): The input to the encoder layer, which is also its output after being processed
            by all sublayers.
            freq_cis (chex.Array): A tensor containing frequency-domain representations of each token's context vector,
            used for computing self-attention weights and biases in a more efficient manner than using position
            embeddings or sinusoidal positional encoding vectors would allow for [2].

        :param self: Represent the instance of the class
        :param hidden_states: chex.Array: Represent the input to the encoder layer
        :param freq_cis: Tuple[chex.Array, chex.Array],: Pass the frequency information to the attention layer
        :param attention_mask: chex.Array: Mask out the attention weights for certain positions
        :param causal_mask: chex.Array: Mask the future tokens
        :param position_ids: chex.Array: Indicate the position of each token in the sequence
        :param deterministic: bool: Determine whether to use dropout or not
        :param init_cache: bool: Initialize the cache for the self-attention layer
        :param output_attentions: bool: Determine whether to return the attention weights or not
        :return: A tuple of hidden_states and attention_output

        """

        # hidden_states: chex.Array,
        # freq_cis: Tuple[chex.Array, chex.Array],
        # attention_mask: chex.Array,
        # position_ids: chex.Array,
        # causal_mask: chex.Array,
        # segment_ids: Optional[chex.Array] = None,
        # deterministic: bool = True,
        # init_cache: bool = False,
        # output_attentions: bool = False,
        # fcm_mask = None,
        residual = hidden_states
        attention_output = self.self_attn(
            self.input_layernorm(hidden_states),
            freq_cis,
            attention_mask,
            position_ids,
            causal_mask,
            segment_ids,
            deterministic,
            init_cache,
            output_attentions,
            None
        )

        hidden_states = attention_output[0] + residual
        ffd_inp = self.post_attention_layernorm(hidden_states)
        if self.config.use_scan_mlp:
            feed_forward_hidden_states = block_wise_ffn(
                self.mlp,
                ffd_inp,
                self.config.scan_mlp_chunk_size,
                deterministic,
            )
        else:
            feed_forward_hidden_states = self.mlp(
                ffd_inp,
                deterministic,
            )

        hidden_states = hidden_states + feed_forward_hidden_states
        outputs = (hidden_states,)
        if output_attentions:
            outputs += attention_output[1],
        return outputs

__call__(hidden_states, freq_cis, attention_mask, causal_mask, position_ids, segment_ids=None, deterministic=True, init_cache=False, output_attentions=True)

The call function is the main function of a TransformerEncoderLayer. It takes in the following arguments: hidden_states (chex.Array): The input to the encoder layer, which is also its output after being processed by all sublayers. freq_cis (chex.Array): A tensor containing frequency-domain representations of each token's context vector, used for computing self-attention weights and biases in a more efficient manner than using position embeddings or sinusoidal positional encoding vectors would allow for [2].

Parameters:

Name Type Description Default
self

Represent the instance of the class

required
hidden_states Array

chex.Array: Represent the input to the encoder layer

required
freq_cis Tuple[Array, Array]

Tuple[chex.Array, chex.Array],: Pass the frequency information to the attention layer

required
attention_mask Array

chex.Array: Mask out the attention weights for certain positions

required
causal_mask Array

chex.Array: Mask the future tokens

required
position_ids Array

chex.Array: Indicate the position of each token in the sequence

required
deterministic bool

bool: Determine whether to use dropout or not

True
init_cache bool

bool: Initialize the cache for the self-attention layer

False
output_attentions bool

bool: Determine whether to return the attention weights or not

True

Returns:

Type Description

A tuple of hidden_states and attention_output

Source code in src/python/easydel/modules/mistral/modelling_mistral_flax.py
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
def __call__(
        self,
        hidden_states: chex.Array,
        freq_cis: Tuple[chex.Array, chex.Array],
        attention_mask: chex.Array,
        causal_mask: chex.Array,
        position_ids: chex.Array,
        segment_ids: Optional[chex.Array] = None,
        deterministic: bool = True,
        init_cache: bool = False,
        output_attentions: bool = True
):
    """
    The __call__ function is the main function of a TransformerEncoderLayer.
    It takes in the following arguments:
        hidden_states (chex.Array): The input to the encoder layer, which is also its output after being processed
        by all sublayers.
        freq_cis (chex.Array): A tensor containing frequency-domain representations of each token's context vector,
        used for computing self-attention weights and biases in a more efficient manner than using position
        embeddings or sinusoidal positional encoding vectors would allow for [2].

    :param self: Represent the instance of the class
    :param hidden_states: chex.Array: Represent the input to the encoder layer
    :param freq_cis: Tuple[chex.Array, chex.Array],: Pass the frequency information to the attention layer
    :param attention_mask: chex.Array: Mask out the attention weights for certain positions
    :param causal_mask: chex.Array: Mask the future tokens
    :param position_ids: chex.Array: Indicate the position of each token in the sequence
    :param deterministic: bool: Determine whether to use dropout or not
    :param init_cache: bool: Initialize the cache for the self-attention layer
    :param output_attentions: bool: Determine whether to return the attention weights or not
    :return: A tuple of hidden_states and attention_output

    """

    # hidden_states: chex.Array,
    # freq_cis: Tuple[chex.Array, chex.Array],
    # attention_mask: chex.Array,
    # position_ids: chex.Array,
    # causal_mask: chex.Array,
    # segment_ids: Optional[chex.Array] = None,
    # deterministic: bool = True,
    # init_cache: bool = False,
    # output_attentions: bool = False,
    # fcm_mask = None,
    residual = hidden_states
    attention_output = self.self_attn(
        self.input_layernorm(hidden_states),
        freq_cis,
        attention_mask,
        position_ids,
        causal_mask,
        segment_ids,
        deterministic,
        init_cache,
        output_attentions,
        None
    )

    hidden_states = attention_output[0] + residual
    ffd_inp = self.post_attention_layernorm(hidden_states)
    if self.config.use_scan_mlp:
        feed_forward_hidden_states = block_wise_ffn(
            self.mlp,
            ffd_inp,
            self.config.scan_mlp_chunk_size,
            deterministic,
        )
    else:
        feed_forward_hidden_states = self.mlp(
            ffd_inp,
            deterministic,
        )

    hidden_states = hidden_states + feed_forward_hidden_states
    outputs = (hidden_states,)
    if output_attentions:
        outputs += attention_output[1],
    return outputs

FlaxMistralForCausalLMModule

Bases: Module

Source code in src/python/easydel/modules/mistral/modelling_mistral_flax.py
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
class FlaxMistralForCausalLMModule(nn.Module):
    config: MistralConfig
    dtype: jnp.dtype = jnp.bfloat16
    param_dtype: jnp.dtype = jnp.bfloat16
    precision: Optional[Union[jax.lax.Precision, str]] = None

    def setup(self):
        self.model: FlaxMistralModule = FlaxMistralModule(
            self.config,
            dtype=self.dtype,
            param_dtype=self.param_dtype,
            precision=self.precision,
        )

        self.lm_head = Linear(
            self.config.vocab_size,
            dtype=self.dtype,
            param_dtype=self.param_dtype,
            use_bias=False,
            kernel_init=jax.nn.initializers.normal(
                stddev=self.config.initializer_range),
            precision=self.precision,
            **get_dot_general_by_bits(self.config.bits, self.config.easy_method)
        )

    def __call__(
            self,
            input_ids: chex.Array,
            attention_mask: chex.Array,
            position_ids: chex.Array,
            deterministic: bool = True,
            inputs_embeds: chex.Array = None,
            init_cache: bool = False,
            output_attentions: bool = False,
            output_hidden_states: bool = False,
            return_dict: bool = True,
    ):
        """
            The __call__ function is the main function of a Flax module. It defines how the model will be called,
            and what it returns. In this case, we are calling our Transformer model with input_ids and attention_mask
            as inputs (these are defined in __init__). We also have some optional arguments that can be passed to
            the call function: deterministic (whether to use dropout), inputs_embeds (if you want to pass your own embeddings),
            output_attentions and output_hidden states which return additional outputs from the transformer layers if set True. Finally,

            :param self: Refer to the object itself
            :param input_ids: chex.Array: Pass in the input tokens
            :param attention_mask: chex.Array: Mask out the padding tokens
            :param position_ids: chex.Array: Specify the position of each token in the sequence
            :param deterministic: bool: Determine whether to use dropout in the model
            :param inputs_embeds: chex.Array: Pass in the embeddings of the input tokens
            :param init_cache: bool: Initialize the cache for the decoder
            :param output_attentions: bool: Return the attention weights
            :param output_hidden_states: bool: Return the hidden states of all layers
            :param return_dict: bool: Return a dictionary of the outputs or just the logits
            :param : Determine whether to return the logits or not
            :return: A tuple of (lm_logits, hidden_states, attentions)

        """
        batch_size, seq_length = input_ids.shape

        if attention_mask is None:
            attention_mask = jnp.ones_like(input_ids)
        if position_ids is None:
            position_ids = jnp.broadcast_to(
                jnp.clip(jnp.cumsum(attention_mask, axis=-1) - 1, a_min=0),
                (batch_size, seq_length)
            )
        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            deterministic=deterministic,
            inputs_embeds=inputs_embeds,
            init_cache=init_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict
        )

        hidden_states = outputs[0]

        if self.config.tie_word_embeddings:
            shared_kernel = self.transformer.variables["params"]["embed_tokens"]["embedding"]
            shared_kernel = fjformer.linen.linen.control_quantization(shared_kernel, self.param_dtype).T
            lm_logits = self.lm_head.apply(
                {"params": {"kernel": shared_kernel}}, hidden_states)
        else:
            lm_logits = self.lm_head(hidden_states)

        # lm_logits = lm_logits.astype(jnp.float32)

        if not return_dict:
            return (lm_logits,) + outputs[1:]

        return FlaxCausalLMOutput(logits=lm_logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions)

__call__(input_ids, attention_mask, position_ids, deterministic=True, inputs_embeds=None, init_cache=False, output_attentions=False, output_hidden_states=False, return_dict=True)

The call function is the main function of a Flax module. It defines how the model will be called, and what it returns. In this case, we are calling our Transformer model with input_ids and attention_mask as inputs (these are defined in init). We also have some optional arguments that can be passed to the call function: deterministic (whether to use dropout), inputs_embeds (if you want to pass your own embeddings), output_attentions and output_hidden states which return additional outputs from the transformer layers if set True. Finally,

Parameters:

Name Type Description Default
self

Refer to the object itself

required
input_ids Array

chex.Array: Pass in the input tokens

required
attention_mask Array

chex.Array: Mask out the padding tokens

required
position_ids Array

chex.Array: Specify the position of each token in the sequence

required
deterministic bool

bool: Determine whether to use dropout in the model

True
inputs_embeds Array

chex.Array: Pass in the embeddings of the input tokens

None
init_cache bool

bool: Initialize the cache for the decoder

False
output_attentions bool

bool: Return the attention weights

False
output_hidden_states bool

bool: Return the hidden states of all layers

False
return_dict bool

bool: Return a dictionary of the outputs or just the logits

True

Determine whether to return the logits or not

required

Returns:

Type Description

A tuple of (lm_logits, hidden_states, attentions)

Source code in src/python/easydel/modules/mistral/modelling_mistral_flax.py
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
def __call__(
        self,
        input_ids: chex.Array,
        attention_mask: chex.Array,
        position_ids: chex.Array,
        deterministic: bool = True,
        inputs_embeds: chex.Array = None,
        init_cache: bool = False,
        output_attentions: bool = False,
        output_hidden_states: bool = False,
        return_dict: bool = True,
):
    """
        The __call__ function is the main function of a Flax module. It defines how the model will be called,
        and what it returns. In this case, we are calling our Transformer model with input_ids and attention_mask
        as inputs (these are defined in __init__). We also have some optional arguments that can be passed to
        the call function: deterministic (whether to use dropout), inputs_embeds (if you want to pass your own embeddings),
        output_attentions and output_hidden states which return additional outputs from the transformer layers if set True. Finally,

        :param self: Refer to the object itself
        :param input_ids: chex.Array: Pass in the input tokens
        :param attention_mask: chex.Array: Mask out the padding tokens
        :param position_ids: chex.Array: Specify the position of each token in the sequence
        :param deterministic: bool: Determine whether to use dropout in the model
        :param inputs_embeds: chex.Array: Pass in the embeddings of the input tokens
        :param init_cache: bool: Initialize the cache for the decoder
        :param output_attentions: bool: Return the attention weights
        :param output_hidden_states: bool: Return the hidden states of all layers
        :param return_dict: bool: Return a dictionary of the outputs or just the logits
        :param : Determine whether to return the logits or not
        :return: A tuple of (lm_logits, hidden_states, attentions)

    """
    batch_size, seq_length = input_ids.shape

    if attention_mask is None:
        attention_mask = jnp.ones_like(input_ids)
    if position_ids is None:
        position_ids = jnp.broadcast_to(
            jnp.clip(jnp.cumsum(attention_mask, axis=-1) - 1, a_min=0),
            (batch_size, seq_length)
        )
    outputs = self.model(
        input_ids=input_ids,
        attention_mask=attention_mask,
        position_ids=position_ids,
        deterministic=deterministic,
        inputs_embeds=inputs_embeds,
        init_cache=init_cache,
        output_attentions=output_attentions,
        output_hidden_states=output_hidden_states,
        return_dict=return_dict
    )

    hidden_states = outputs[0]

    if self.config.tie_word_embeddings:
        shared_kernel = self.transformer.variables["params"]["embed_tokens"]["embedding"]
        shared_kernel = fjformer.linen.linen.control_quantization(shared_kernel, self.param_dtype).T
        lm_logits = self.lm_head.apply(
            {"params": {"kernel": shared_kernel}}, hidden_states)
    else:
        lm_logits = self.lm_head(hidden_states)

    # lm_logits = lm_logits.astype(jnp.float32)

    if not return_dict:
        return (lm_logits,) + outputs[1:]

    return FlaxCausalLMOutput(logits=lm_logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions)

FlaxMistralModule

Bases: Module

Source code in src/python/easydel/modules/mistral/modelling_mistral_flax.py
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
class FlaxMistralModule(nn.Module):
    config: MistralConfig
    dtype: jnp.dtype = jnp.bfloat16
    param_dtype: jnp.dtype = jnp.bfloat16
    precision: Optional[Union[jax.lax.Precision, str]] = None

    def setup(self):

        self.embed_tokens = nn.Embed(
            self.config.vocab_size,
            self.config.hidden_size,
            embedding_init=jax.nn.initializers.normal(
                stddev=self.config.initializer_range),
            dtype=self.dtype,
            param_dtype=self.param_dtype,
        )

        self.layers = FlaxMistralDecoratorCollection(
            self.config,
            dtype=self.dtype,
            param_dtype=self.param_dtype,
            precision=self.precision
        )
        self.norm = MistralRMSNorm(
            self.config.hidden_size,
            eps=self.config.rms_norm_eps,
            dtype=self.dtype,
            param_dtype=self.param_dtype
        )

        initial_rope_kwargs = dict(
            rope_type="none"
        )
        if self.config.rope_scaling is not None:
            scaling_type = self.config.rope_scaling["type"]
            scaling_factor = self.config.rope_scaling["factor"]
            initial_rope_kwargs = dict(
                scaling_factor=scaling_factor,
                rope_type=scaling_type
            )
        self.freq_cis = precompute_freq_cis(
            max_position_embeddings=(
                getattr(self.config, "freq_max_position_embeddings", self.config.max_position_embeddings)
            ),
            dim=self.config.hidden_size // self.config.num_attention_heads,
            base=self.config.rope_theta,
            **initial_rope_kwargs
        )
        self.causal_mask = flax.linen.make_causal_mask(
            jnp.ones(
                (1, getattr(self.config, "c_max_position_embeddings", self.config.max_position_embeddings)),
                dtype="bool"
            ), dtype="bool"
        )

    def __call__(
            self,
            input_ids: Optional[chex.Array] = None,
            attention_mask: Optional[chex.Array] = None,
            position_ids: Optional[chex.Array] = None,
            deterministic: bool = True,
            inputs_embeds: chex.Array = None,
            init_cache: bool = False,
            output_attentions: bool = False,
            output_hidden_states: bool = False,
            return_dict: bool = True,
    ) -> typing.Union[Tuple[Array, ...], FlaxBaseModelOutput]:
        """
        The __call__ function is the main function of a Flax model.
        It takes in input_ids, attention_mask, and position_ids as inputs to the model.
        The output is a tuple containing: last hidden state (hidden states), all hidden states (if output_hidden_states=True), attentions (if output attentions=True).


        :param self: Represent the instance of the class
        :param input_ids: chex.Array: Pass in the input ids
        :param attention_mask: chex.Array: Mask out the attention weights for certain tokens
        :param position_ids: chex.Array: Determine the position of each token in a sequence
        :param deterministic: bool: Determine whether to use dropout or not
        :param inputs_embeds: chex.Array: Pass in the embedding of the input_ids
        :param init_cache: bool: Initialize the cache for the decoder
        :param output_attentions: bool: Determine whether to return the attention weights or not
        :param output_hidden_states: bool: Return all hidden states or just the last one
        :param return_dict: bool: Return a dictionary of the outputs or not
        :param : Determine whether the model is in training mode or not
        :return: A tuple of the hidden states, all hidden states, and attentions

        """
        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids.astype("i4"))
        if attention_mask.ndim == 2:
            b, s = attention_mask.shape
            attention_mask = attention_mask.reshape(b, 1, 1, s)

        outputs = self.layers(
            hidden_states=inputs_embeds,
            attention_mask=attention_mask,
            position_ids=position_ids,
            freq_cis=self.freq_cis,
            init_cache=init_cache,
            output_attentions=output_attentions,
            deterministic=deterministic,
            causal_mask=self.causal_mask
        )

        hidden_states = outputs[0]
        hidden_states = self.norm(hidden_states)

        if output_hidden_states:
            all_hidden_states = outputs[1] + (hidden_states,)
            outputs = (hidden_states, all_hidden_states) + outputs[2:]
        else:
            outputs = (hidden_states,) + outputs[1:]

        if not return_dict:
            return tuple(value for value in outputs if value is not None)

        return FlaxBaseModelOutput(
            last_hidden_state=hidden_states,
            hidden_states=outputs[1],
            attentions=outputs[-1],
        )

__call__(input_ids=None, attention_mask=None, position_ids=None, deterministic=True, inputs_embeds=None, init_cache=False, output_attentions=False, output_hidden_states=False, return_dict=True)

The call function is the main function of a Flax model. It takes in input_ids, attention_mask, and position_ids as inputs to the model. The output is a tuple containing: last hidden state (hidden states), all hidden states (if output_hidden_states=True), attentions (if output attentions=True).

Parameters:

Name Type Description Default
self

Represent the instance of the class

required
input_ids Optional[Array]

chex.Array: Pass in the input ids

None
attention_mask Optional[Array]

chex.Array: Mask out the attention weights for certain tokens

None
position_ids Optional[Array]

chex.Array: Determine the position of each token in a sequence

None
deterministic bool

bool: Determine whether to use dropout or not

True
inputs_embeds Array

chex.Array: Pass in the embedding of the input_ids

None
init_cache bool

bool: Initialize the cache for the decoder

False
output_attentions bool

bool: Determine whether to return the attention weights or not

False
output_hidden_states bool

bool: Return all hidden states or just the last one

False
return_dict bool

bool: Return a dictionary of the outputs or not

True

Determine whether the model is in training mode or not

required

Returns:

Type Description
Union[Tuple[Array, ...], FlaxBaseModelOutput]

A tuple of the hidden states, all hidden states, and attentions

Source code in src/python/easydel/modules/mistral/modelling_mistral_flax.py
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
def __call__(
        self,
        input_ids: Optional[chex.Array] = None,
        attention_mask: Optional[chex.Array] = None,
        position_ids: Optional[chex.Array] = None,
        deterministic: bool = True,
        inputs_embeds: chex.Array = None,
        init_cache: bool = False,
        output_attentions: bool = False,
        output_hidden_states: bool = False,
        return_dict: bool = True,
) -> typing.Union[Tuple[Array, ...], FlaxBaseModelOutput]:
    """
    The __call__ function is the main function of a Flax model.
    It takes in input_ids, attention_mask, and position_ids as inputs to the model.
    The output is a tuple containing: last hidden state (hidden states), all hidden states (if output_hidden_states=True), attentions (if output attentions=True).


    :param self: Represent the instance of the class
    :param input_ids: chex.Array: Pass in the input ids
    :param attention_mask: chex.Array: Mask out the attention weights for certain tokens
    :param position_ids: chex.Array: Determine the position of each token in a sequence
    :param deterministic: bool: Determine whether to use dropout or not
    :param inputs_embeds: chex.Array: Pass in the embedding of the input_ids
    :param init_cache: bool: Initialize the cache for the decoder
    :param output_attentions: bool: Determine whether to return the attention weights or not
    :param output_hidden_states: bool: Return all hidden states or just the last one
    :param return_dict: bool: Return a dictionary of the outputs or not
    :param : Determine whether the model is in training mode or not
    :return: A tuple of the hidden states, all hidden states, and attentions

    """
    if inputs_embeds is None:
        inputs_embeds = self.embed_tokens(input_ids.astype("i4"))
    if attention_mask.ndim == 2:
        b, s = attention_mask.shape
        attention_mask = attention_mask.reshape(b, 1, 1, s)

    outputs = self.layers(
        hidden_states=inputs_embeds,
        attention_mask=attention_mask,
        position_ids=position_ids,
        freq_cis=self.freq_cis,
        init_cache=init_cache,
        output_attentions=output_attentions,
        deterministic=deterministic,
        causal_mask=self.causal_mask
    )

    hidden_states = outputs[0]
    hidden_states = self.norm(hidden_states)

    if output_hidden_states:
        all_hidden_states = outputs[1] + (hidden_states,)
        outputs = (hidden_states, all_hidden_states) + outputs[2:]
    else:
        outputs = (hidden_states,) + outputs[1:]

    if not return_dict:
        return tuple(value for value in outputs if value is not None)

    return FlaxBaseModelOutput(
        last_hidden_state=hidden_states,
        hidden_states=outputs[1],
        attentions=outputs[-1],
    )

FlaxMistralPretrainedModel

Bases: EasyDeLFlaxPretrainedModel

Source code in src/python/easydel/modules/mistral/modelling_mistral_flax.py
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
class FlaxMistralPretrainedModel(EasyDeLFlaxPretrainedModel):
    config_class = MistralConfig
    base_model_prefix = 'mistral'
    module_class: nn.Module = None

    def __init__(self,
                 config: MistralConfig,
                 input_shape: Tuple = (1, 1),
                 seed: int = 0,
                 dtype: jnp.dtype = jnp.bfloat16,
                 _do_init: bool = True,
                 **kwargs
                 ):
        module = self.module_class(config=config, dtype=dtype, **kwargs)
        super().__init__(config, module, input_shape=input_shape,
                         seed=seed, dtype=dtype, _do_init=_do_init)

    def init_weights(
            self,
            rng: jax.random.PRNGKey,
            input_shape: Tuple,
            params: flax.core.FrozenDict = None
    ) -> flax.core.FrozenDict:
        """
        The init_weights function is used to initialize the weights of a model.
        It takes in an rng, which is a random number generator key that can be used to generate random numbers.
        The input_shape parameter specifies the shape of the inputs that will be fed into this model.
        The params parameter allows you to pass in pre-trained weights for your model, if you have them available.

        :param self: Access variables that belong to the class
        :param rng: jax.random.PRNGKey: Initialize the weights of the model
        :param input_shape: Tuple: Initialize the input_ids, attention_mask and position_ids
        :param params: flax.core.FrozenDict: Pass in the parameters of a pre-trained model
        :return: A frozendict of parameters

        """
        input_ids = jnp.zeros(input_shape, dtype="i4")
        attention_mask = jnp.ones_like(input_ids)
        position_ids = jnp.broadcast_to(jnp.arange(
            jnp.atleast_2d(input_ids).shape[-1]), input_shape)
        params_rng, dropout_rng = jax.random.split(rng)
        rng_s = {"params": params_rng, "dropout": dropout_rng}

        if self.config.add_cross_attention:
            encoder_hidden_states = jnp.zeros(
                input_shape + (self.config.hidden_size,))
            encoder_attention_mask = attention_mask
            module_init_outputs = self.module.init(
                rng_s,
                input_ids,
                attention_mask,
                position_ids,
                encoder_hidden_states,
                encoder_attention_mask,
                return_dict=False,
            )
        else:
            module_init_outputs = self.module.init(
                rng_s, input_ids, attention_mask, position_ids, return_dict=False
            )

        random_params = module_init_outputs["params"]

        if params is not None:
            random_params = flatten_dict(unfreeze(random_params))
            params = flatten_dict(unfreeze(params))
            for missing_key in self._missing_keys:
                params[missing_key] = random_params[missing_key]
            self._missing_keys = set()
            return freeze(unflatten_dict(params))
        else:
            return random_params

    def init_cache(self, batch_size, max_length):

        input_ids = jnp.ones((batch_size, max_length))
        attention_mask = jnp.ones_like(input_ids)
        position_ids = jnp.broadcast_to(jnp.arange(
            jnp.atleast_2d(input_ids).shape[-1]), input_ids.shape)

        init_variables = self.module.init(
            jax.random.PRNGKey(0),
            input_ids,
            attention_mask,
            position_ids,
            return_dict=False,
            init_cache=True
        )
        return init_variables["cache"]

    def __call__(
            self,
            input_ids,
            attention_mask=None,
            position_ids=None,
            params: dict = None,
            past_key_values: dict = None,
            dropout_rng: jax.random.PRNGKey = None,
            train: bool = False,
            output_attentions: Optional[bool] = None,
            output_hidden_states: Optional[bool] = None,
            return_dict: Optional[bool] = None,
            add_params_field: bool = False,
            **kwargs
    ):
        """
        The __call__ function is the main function of a JAX module.
        It takes as input:
        - The parameters of the model (self.params)
        - The inputs to the model (input_ids, attention_mask, position_ids)
        - Whether we are training (train=True/False) and whether we want to return all hidden states and
        attentions weights at each layer in addition to just the last layer output (output_hidden_states=True/False).

        :param self: Represent the instance of the class
        :param input_ids: Pass the input sequence to the model
        :param attention_mask: Mask out the padding tokens
        :param position_ids: Specify the position of each token in the sequence
        :param params: dict: Pass in the parameters of the model
        :param past_key_values: dict: Pass the past key values to the model
        :param dropout_rng: jax.random.PRNGKey: Pass in a random number generator key to the model
        :param train: bool: Determine whether to use dropout or not
        :param output_attentions: Optional[bool]: Determine whether to return the attention weights
        :param output_hidden_states: Optional[bool]: Determine whether to return the hidden states of all layers
        :param return_dict: Optional[bool]: Return a dictionary of the outputs
        :param add_params_field: bool: Add a params field to the inputs dictionary
        :return: A tuple of (last_hidden_state, past_key_values)

        """
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.return_dict
        batch_size, sequence_length = input_ids.shape

        if position_ids is None:
            if past_key_values is not None:
                raise ValueError(
                    "Make sure to provide `position_ids` when passing `past_key_values`.")

            position_ids = jnp.broadcast_to(jnp.arange(sequence_length)[
                                            None, :], (batch_size, sequence_length))

        if attention_mask is None:
            attention_mask = jnp.ones((batch_size, sequence_length))

        rng_s = {}
        if dropout_rng is not None:
            rng_s["dropout"] = dropout_rng

        inputs = {
            "params": params or self.params} if add_params_field else params or self.params

        if self.config.bits is not None:
            rng_s['params'] = jax.random.key(0)
        if past_key_values:
            inputs["cache"] = past_key_values
            mutable = ["cache"]
        else:
            mutable = False

        outputs = self.module.apply(
            inputs,
            jnp.array(input_ids, dtype="i4"),
            jnp.array(attention_mask, dtype="i4"),
            jnp.array(position_ids, dtype="i4"),
            not train,
            None,
            False,
            output_attentions,
            output_hidden_states,
            return_dict,
            rngs=rng_s,
            mutable=mutable,
        )

        if past_key_values is not None and return_dict:
            outputs, past_key_values = outputs
            outputs["past_key_values"] = unfreeze(past_key_values["cache"])
            return outputs
        elif past_key_values is not None and not return_dict:
            outputs, past_key_values = outputs
            outputs = outputs[:1] + \
                      (unfreeze(past_key_values["cache"]),) + outputs[1:]

        return outputs

__call__(input_ids, attention_mask=None, position_ids=None, params=None, past_key_values=None, dropout_rng=None, train=False, output_attentions=None, output_hidden_states=None, return_dict=None, add_params_field=False, **kwargs)

The call function is the main function of a JAX module. It takes as input: - The parameters of the model (self.params) - The inputs to the model (input_ids, attention_mask, position_ids) - Whether we are training (train=True/False) and whether we want to return all hidden states and attentions weights at each layer in addition to just the last layer output (output_hidden_states=True/False).

Parameters:

Name Type Description Default
self

Represent the instance of the class

required
input_ids

Pass the input sequence to the model

required
attention_mask

Mask out the padding tokens

None
position_ids

Specify the position of each token in the sequence

None
params dict

dict: Pass in the parameters of the model

None
past_key_values dict

dict: Pass the past key values to the model

None
dropout_rng PRNGKey

jax.random.PRNGKey: Pass in a random number generator key to the model

None
train bool

bool: Determine whether to use dropout or not

False
output_attentions Optional[bool]

Optional[bool]: Determine whether to return the attention weights

None
output_hidden_states Optional[bool]

Optional[bool]: Determine whether to return the hidden states of all layers

None
return_dict Optional[bool]

Optional[bool]: Return a dictionary of the outputs

None
add_params_field bool

bool: Add a params field to the inputs dictionary

False

Returns:

Type Description

A tuple of (last_hidden_state, past_key_values)

Source code in src/python/easydel/modules/mistral/modelling_mistral_flax.py
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
def __call__(
        self,
        input_ids,
        attention_mask=None,
        position_ids=None,
        params: dict = None,
        past_key_values: dict = None,
        dropout_rng: jax.random.PRNGKey = None,
        train: bool = False,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        add_params_field: bool = False,
        **kwargs
):
    """
    The __call__ function is the main function of a JAX module.
    It takes as input:
    - The parameters of the model (self.params)
    - The inputs to the model (input_ids, attention_mask, position_ids)
    - Whether we are training (train=True/False) and whether we want to return all hidden states and
    attentions weights at each layer in addition to just the last layer output (output_hidden_states=True/False).

    :param self: Represent the instance of the class
    :param input_ids: Pass the input sequence to the model
    :param attention_mask: Mask out the padding tokens
    :param position_ids: Specify the position of each token in the sequence
    :param params: dict: Pass in the parameters of the model
    :param past_key_values: dict: Pass the past key values to the model
    :param dropout_rng: jax.random.PRNGKey: Pass in a random number generator key to the model
    :param train: bool: Determine whether to use dropout or not
    :param output_attentions: Optional[bool]: Determine whether to return the attention weights
    :param output_hidden_states: Optional[bool]: Determine whether to return the hidden states of all layers
    :param return_dict: Optional[bool]: Return a dictionary of the outputs
    :param add_params_field: bool: Add a params field to the inputs dictionary
    :return: A tuple of (last_hidden_state, past_key_values)

    """
    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
    output_hidden_states = (
        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
    )
    return_dict = return_dict if return_dict is not None else self.config.return_dict
    batch_size, sequence_length = input_ids.shape

    if position_ids is None:
        if past_key_values is not None:
            raise ValueError(
                "Make sure to provide `position_ids` when passing `past_key_values`.")

        position_ids = jnp.broadcast_to(jnp.arange(sequence_length)[
                                        None, :], (batch_size, sequence_length))

    if attention_mask is None:
        attention_mask = jnp.ones((batch_size, sequence_length))

    rng_s = {}
    if dropout_rng is not None:
        rng_s["dropout"] = dropout_rng

    inputs = {
        "params": params or self.params} if add_params_field else params or self.params

    if self.config.bits is not None:
        rng_s['params'] = jax.random.key(0)
    if past_key_values:
        inputs["cache"] = past_key_values
        mutable = ["cache"]
    else:
        mutable = False

    outputs = self.module.apply(
        inputs,
        jnp.array(input_ids, dtype="i4"),
        jnp.array(attention_mask, dtype="i4"),
        jnp.array(position_ids, dtype="i4"),
        not train,
        None,
        False,
        output_attentions,
        output_hidden_states,
        return_dict,
        rngs=rng_s,
        mutable=mutable,
    )

    if past_key_values is not None and return_dict:
        outputs, past_key_values = outputs
        outputs["past_key_values"] = unfreeze(past_key_values["cache"])
        return outputs
    elif past_key_values is not None and not return_dict:
        outputs, past_key_values = outputs
        outputs = outputs[:1] + \
                  (unfreeze(past_key_values["cache"]),) + outputs[1:]

    return outputs

init_weights(rng, input_shape, params=None)

The init_weights function is used to initialize the weights of a model. It takes in an rng, which is a random number generator key that can be used to generate random numbers. The input_shape parameter specifies the shape of the inputs that will be fed into this model. The params parameter allows you to pass in pre-trained weights for your model, if you have them available.

Parameters:

Name Type Description Default
self

Access variables that belong to the class

required
rng PRNGKey

jax.random.PRNGKey: Initialize the weights of the model

required
input_shape Tuple

Tuple: Initialize the input_ids, attention_mask and position_ids

required
params FrozenDict

flax.core.FrozenDict: Pass in the parameters of a pre-trained model

None

Returns:

Type Description
FrozenDict

A frozendict of parameters

Source code in src/python/easydel/modules/mistral/modelling_mistral_flax.py
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
def init_weights(
        self,
        rng: jax.random.PRNGKey,
        input_shape: Tuple,
        params: flax.core.FrozenDict = None
) -> flax.core.FrozenDict:
    """
    The init_weights function is used to initialize the weights of a model.
    It takes in an rng, which is a random number generator key that can be used to generate random numbers.
    The input_shape parameter specifies the shape of the inputs that will be fed into this model.
    The params parameter allows you to pass in pre-trained weights for your model, if you have them available.

    :param self: Access variables that belong to the class
    :param rng: jax.random.PRNGKey: Initialize the weights of the model
    :param input_shape: Tuple: Initialize the input_ids, attention_mask and position_ids
    :param params: flax.core.FrozenDict: Pass in the parameters of a pre-trained model
    :return: A frozendict of parameters

    """
    input_ids = jnp.zeros(input_shape, dtype="i4")
    attention_mask = jnp.ones_like(input_ids)
    position_ids = jnp.broadcast_to(jnp.arange(
        jnp.atleast_2d(input_ids).shape[-1]), input_shape)
    params_rng, dropout_rng = jax.random.split(rng)
    rng_s = {"params": params_rng, "dropout": dropout_rng}

    if self.config.add_cross_attention:
        encoder_hidden_states = jnp.zeros(
            input_shape + (self.config.hidden_size,))
        encoder_attention_mask = attention_mask
        module_init_outputs = self.module.init(
            rng_s,
            input_ids,
            attention_mask,
            position_ids,
            encoder_hidden_states,
            encoder_attention_mask,
            return_dict=False,
        )
    else:
        module_init_outputs = self.module.init(
            rng_s, input_ids, attention_mask, position_ids, return_dict=False
        )

    random_params = module_init_outputs["params"]

    if params is not None:
        random_params = flatten_dict(unfreeze(random_params))
        params = flatten_dict(unfreeze(params))
        for missing_key in self._missing_keys:
            params[missing_key] = random_params[missing_key]
        self._missing_keys = set()
        return freeze(unflatten_dict(params))
    else:
        return random_params

matmul_4d_loop(x, y)

Computes the matrix product of two 4D arrays x and y using a loop.

Source code in src/python/easydel/modules/mistral/modelling_mistral_flax.py
37
38
39
40
41
42
43
44
45
def matmul_4d_loop(x, y):
    """Computes the matrix product of two 4D arrays x and y using a loop."""
    result = jnp.zeros(*x.shape[:-2] + x.shape[-2] + y.shape[-1])
    for i in range(x.shape[0]):
        for j in range(y.shape[1]):
            for key in range(x.shape[2]):
                for l in range(y.shape[3]):
                    result[i, j, key, l] += x[i, j, key, :] * y[key, l, :, :]
    return result