Source code for easydel.modules.cohere.modeling_cohere_flax

# Copyright 2023 The EASYDEL Author @erfanzar (Erfan Zare Chavoshi).
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.


import typing as tp
from functools import partial

import chex
import jax
import jax.numpy as jnp
from flax import nnx as nn

from easydel.infra.base_module import EasyDeLBaseModule
from easydel.infra.factory import register_module
from easydel.infra.modeling_outputs import (
	FlaxBaseModelOutput,
	FlaxCausalLMOutput,
	FlaxSequenceClassifierOutput,
)
from easydel.infra.utils import (
	auto_remat,
	block_wise_ffn,
	control_mlp_sharding,
	get_dot_general_by_bits,
)
from easydel.layers.attention import FlaxAttentionModule, FlexibleAttentionModule
from easydel.layers.caching import TransformerCache, TransformerCacheView
from easydel.modules.cohere.cohere_configuration import CohereConfig as CohereConfig


[docs]def repeat_kv(x: chex.Array, n_rep: int) -> chex.Array: bs, s, n_kv_heads, head_dim = x.shape if n_rep == 1: return x x = x[:, :, jnp.newaxis, :, :] x = jnp.repeat(x, n_rep, axis=2) return x.reshape(bs, s, n_kv_heads * n_rep, head_dim)
[docs]class RMSNorm(nn.Module): def __init__( self, dim: tp.Union[int, tuple], eps: float = 1e-6, dtype: jnp.dtype = jnp.float32, param_dtype: jnp.dtype = jnp.float32, do_t: bool = False, rngs: nn.Rngs = None, ): super().__init__() if rngs is None: rngs = nn.Rngs(0) self.dim = dim self.eps = eps self.dtype = dtype self.param_dtype = param_dtype self.do_t = do_t self.kernel = nn.Param( nn.initializers.ones( key=rngs.params(), shape=(self.dim,) if isinstance(self.dim, int) else self.dim, dtype=self.param_dtype, ), ) def _norm(self, x: jnp.ndarray) -> jnp.ndarray: return x * jax.lax.rsqrt(jnp.square(x).mean(-1, keepdims=True) + self.eps) def __call__(self, x: jnp.ndarray) -> jnp.ndarray: if self.dtype in [ jnp.float8_e4m3b11fnuz, jnp.float8_e4m3fn, jnp.float8_e4m3fnuz, jnp.float8_e5m2, jnp.float8_e5m2fnuz, ]: x = x.astype(jnp.float32) else: x = x.astype(jnp.promote_types(self.dtype, jnp.float32)) output = self._norm(x).astype(self.dtype) weight = self.kernel.value.astype(self.dtype) if self.do_t: weight = weight.T return output * weight
[docs]class CohereAttention(FlaxAttentionModule): def __init__( self, config: CohereConfig, dtype: jnp.dtype = jnp.float32, param_dtype: jnp.dtype = jnp.float32, precision: jax.lax.PrecisionLike = None, *, rngs: nn.Rngs, ) -> None: super().__init__(config=config) self.dtype = dtype self.param_dtype = param_dtype self.precision = precision self.rngs = rngs self.hidden_size = config.hidden_size 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 if config.use_qk_norm: self.q_norm = RMSNorm( dim=(self.head_dim, self.config.num_attention_heads), eps=config.layer_norm_eps, dtype=self.dtype, param_dtype=self.param_dtype, do_t=True, ) self.k_norm = RMSNorm( dim=( self.head_dim, self.config.num_key_value_heads, ), eps=config.layer_norm_eps, dtype=self.dtype, param_dtype=self.param_dtype, do_t=True, ) linear_class = partial( nn.Linear, dtype=dtype, param_dtype=param_dtype, use_bias=False, kernel_init=jax.nn.initializers.normal(config.initializer_range), precision=self.precision, **get_dot_general_by_bits(config.bits, config.easy_method), ) self.q_proj = linear_class( config.hidden_size, config.num_attention_heads * self.head_dim, rngs=rngs, ) self.k_proj = linear_class( config.hidden_size, config.num_key_value_heads * self.head_dim, rngs=rngs, ) self.v_proj = linear_class( config.hidden_size, config.num_key_value_heads * self.head_dim, rngs=rngs, ) self.o_proj = linear_class( config.num_attention_heads * self.head_dim, config.hidden_size, rngs=rngs, ) self.rotary = self.config.get_basic_rope( self.dtype, self.head_dim, self.head_dim, True, ) self.attention_performer = FlexibleAttentionModule( dropout_prob=config.attention_dropout, base_config=config, softmax_scale=self.head_dim**-0.5, ) def __call__( self, hidden_states: chex.Array, attention_mask: chex.Array, position_ids: chex.Array, causal_mask: chex.Array, cache_view: tp.Optional[TransformerCacheView] = None, segment_ids: tp.Optional[chex.Array] = None, output_attentions: bool = False, fcm_mask: tp.Optional[chex.Array] = None, frequencies: tp.Optional[chex.Array] = None, ): 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, ) if self.config.use_qk_norm: query_states = self.q_norm(query_states) key_states = self.k_norm(key_states) value_states = value_states.reshape( batch_size, sequence_length, self.config.num_key_value_heads, self.head_dim, ) query_states, key_states = self.rotary( query=query_states, key=key_states, positions=position_ids, frequencies=frequencies, ) ( key_states, value_states, attention_mask, init_attention_bias, ) = self.concatenate( query=query_states, key=key_states, cache_view=cache_view, value=value_states, attention_mask=attention_mask, causal_mask=causal_mask, fcm_mask=fcm_mask, ) attentions = self.attention_performer.forward( query_states=query_states, key_states=key_states, value_states=value_states, bias=None, init_bias=init_attention_bias, attention_mask=attention_mask, segment_ids=segment_ids, causal=True, dropout_rng=self.rngs.params(), ) attn_output = self.shard_attention_prod( self._merge_heads(attentions.attention_outputs) ) attn_output = self.o_proj(attn_output) outputs = ( (attn_output, attentions.attention_weights) if output_attentions else (attn_output, None) ) return outputs
[docs]class CohereMLP(nn.Module): def __init__( self, config: CohereConfig, dtype: jnp.dtype = jnp.float32, param_dtype: jnp.dtype = jnp.float32, precision: jax.lax.PrecisionLike = None, *, rngs: nn.Rngs, ): self.config = config self.dtype = dtype self.param_dtype = param_dtype self.precision = precision linear_class = partial( nn.Linear, dtype=dtype, param_dtype=param_dtype, use_bias=False, kernel_init=jax.nn.initializers.normal(config.initializer_range), precision=self.precision, rngs=rngs, **get_dot_general_by_bits(config.bits, config.easy_method), ) self.gate_proj = linear_class( config.hidden_size, config.intermediate_size, rngs=rngs, ) self.down_proj = linear_class( config.intermediate_size, config.hidden_size, rngs=rngs, ) self.up_proj = linear_class( config.hidden_size, config.intermediate_size, rngs=rngs, ) def __call__(self, hidden_states: jnp.ndarray) -> jnp.ndarray: hidden_states = control_mlp_sharding(hidden_states, self.config.partition_axis) hidden_states = self.down_proj( jax.nn.silu(self.gate_proj(hidden_states)) * self.up_proj(hidden_states) ) return hidden_states
[docs]class CohereBlock(nn.Module): def __init__( self, config: CohereConfig, dtype: jnp.dtype = jnp.float32, param_dtype: jnp.dtype = jnp.float32, precision: jax.lax.PrecisionLike = None, *, rngs: nn.Rngs, ) -> None: super().__init__() self.config = config self.dtype = dtype self.param_dtype = param_dtype self.precision = precision self.rngs = rngs attn_block = CohereAttention mlp_block = CohereMLP attn_block, mlp_block = auto_remat( attn_block, mlp_block, policy=config.gradient_checkpointing, ) self.self_attn = attn_block( config, dtype=dtype, param_dtype=param_dtype, precision=precision, rngs=rngs, ) self.mlp = mlp_block( config, dtype=dtype, param_dtype=param_dtype, precision=precision, rngs=rngs, ) self.input_layernorm = RMSNorm( self.config.hidden_size, eps=self.config.layer_norm_eps, dtype=dtype, param_dtype=param_dtype, rngs=rngs, ) def __call__( self, hidden_states: chex.Array, attention_mask: chex.Array, position_ids: chex.Array, causal_mask: chex.Array, cache_view: tp.Optional[TransformerCacheView] = None, segment_ids: tp.Optional[chex.Array] = None, output_attentions: bool = False, fcm_mask: tp.Optional[chex.Array] = None, frequencies: tp.Optional[chex.Array] = None, ): """ Forward pass of the module block. Args: hidden_states (chex.Array): Input hidden states. attention_mask (chex.Array): Mask to apply on the attention scores. position_ids (chex.Array): Position indices for the tokens. causal_mask (chex.Array): Causal mask for ensuring autoregressive behavior. segment_ids (tp.Optional[chex.Array]): Segment IDs for segment-based attention (optional). deterministic (bool): If True, disables dropout for deterministic behavior. init_cache (bool): If True, initializes cache for caching keys and values. output_attentions (bool): If True, outputs attention weights alongside the hidden states. fcm_mask (tp.Optional[chex.Array]): fcm mask to be combined with attn mask and causal mask. Returns: tp.Tuple[chex.Array, chex.Array]: A tuple containing the attention output and the attention weights. """ residual = hidden_states hidden_states = self.input_layernorm(hidden_states) attn_outputs = self.self_attn( hidden_states, attention_mask, position_ids, causal_mask, cache_view, segment_ids, output_attentions, fcm_mask, frequencies, ) attn_output = attn_outputs[0] feed_forward_input = hidden_states if self.config.use_scan_mlp: feed_forward_hidden_states = block_wise_ffn( self.mlp, feed_forward_input, self.config.scan_mlp_chunk_size, ) else: feed_forward_hidden_states = self.mlp( feed_forward_input, ) hidden_states = attn_output + feed_forward_hidden_states + residual return (hidden_states,) + attn_outputs[1:]
[docs]@register_module( "base-module", config=CohereConfig, model_type="cohere", embedding_layer_names=["embed_tokens"], ) class CohereModel(EasyDeLBaseModule): def __init__( self, config: CohereConfig, dtype: jnp.dtype = jnp.float32, param_dtype: jnp.dtype = jnp.float32, precision: jax.lax.PrecisionLike = None, *, rngs: nn.Rngs, ): super().__init__( config=config, dtype=dtype, param_dtype=param_dtype, precision=precision, rngs=rngs, ) self.embed_tokens = nn.Embed( config.vocab_size, config.hidden_size, embedding_init=nn.initializers.normal(stddev=config.initializer_range), dtype=dtype, param_dtype=param_dtype, rngs=rngs, ) self.layers = [ CohereBlock( config=config, dtype=dtype, param_dtype=param_dtype, precision=precision, rngs=rngs, ) for _ in range(config.num_hidden_layers) ] self.norm = RMSNorm( self.config.hidden_size, eps=self.config.layer_norm_eps, dtype=dtype, param_dtype=param_dtype, ) def __call__( self, input_ids: tp.Optional[chex.Array] = None, inputs_embeds: tp.Optional[chex.Array] = None, attention_mask: tp.Optional[chex.Array] = None, position_ids: tp.Optional[chex.Array] = None, segment_ids: tp.Optional[chex.Array] = None, output_attentions: tp.Optional[bool] = None, output_hidden_states: tp.Optional[bool] = None, past_key_values: tp.Optional[TransformerCache] = None, return_dict: bool = True, ) -> tp.Union[FlaxBaseModelOutput, tp.Tuple]: if (input_ids is None) ^ (inputs_embeds is not None): raise ValueError( "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one" ) if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids.astype("i4")) batch_size, sequence_length, _ = inputs_embeds.shape all_attentions = () if output_attentions else None all_hidden_states = () if output_hidden_states else None assert sequence_length <= self.config.max_position_embeddings, ( f"Maximum Position Embedding Reached ! (Excepted <= {self.config.max_position_embeddings} got {sequence_length})" ) if attention_mask is None: attention_mask = jnp.ones((batch_size, sequence_length), "b1") else: if attention_mask.dtype != jnp.bool: attention_mask = jnp.astype(attention_mask == 1, "b1") if position_ids is None: position_ids = jnp.broadcast_to( jnp.clip(jnp.cumsum(attention_mask, axis=-1) - 1, a_min=0), (batch_size, sequence_length), ).astype(jnp.int32) hidden_states = inputs_embeds if past_key_values is None: past_key_values = TransformerCache.init_empty(len(self.layers)) for idx, block in enumerate(self.layers): if output_hidden_states: all_hidden_states += (hidden_states,) layer_outputs = block( hidden_states=hidden_states, attention_mask=attention_mask, position_ids=position_ids, cache_view=past_key_values.views[idx], causal_mask=self.causal_mask, output_attentions=output_attentions, segment_ids=segment_ids, frequencies=self.frequencies, ) hidden_states = layer_outputs[0] if output_attentions: all_attentions += (layer_outputs[1],) hidden_states = self.norm(hidden_states) if output_hidden_states: all_hidden_states = hidden_states[1] + (hidden_states,) outputs = (hidden_states, all_hidden_states, all_attentions, past_key_values) if not return_dict: return tuple(v for v in outputs if v is not None) return FlaxBaseModelOutput( last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions, past_key_values=past_key_values, )
[docs]@register_module( "causal-language-model", config=CohereConfig, model_type="cohere", embedding_layer_names=["embed_tokens"], ) class CohereForCausalLM(EasyDeLBaseModule): def __init__( self, config: CohereConfig, dtype: jnp.dtype = jnp.float32, param_dtype: jnp.dtype = jnp.float32, precision: jax.lax.PrecisionLike = None, *, rngs: nn.Rngs, ): super().__init__( config=config, dtype=dtype, param_dtype=param_dtype, precision=precision, rngs=rngs, ) self.model = CohereModel( config=config, dtype=dtype, param_dtype=param_dtype, precision=precision, rngs=rngs, ) self.lm_head = nn.Linear( config.hidden_size, config.vocab_size, dtype=dtype, param_dtype=param_dtype, use_bias=False, kernel_init=nn.initializers.normal(stddev=config.initializer_range), precision=precision, rngs=rngs, **get_dot_general_by_bits(config.bits, config.easy_method), ) self.logit_scale = self.config.logit_scale def __call__( self, input_ids: tp.Optional[chex.Array] = None, inputs_embeds: tp.Optional[chex.Array] = None, attention_mask: tp.Optional[chex.Array] = None, position_ids: tp.Optional[chex.Array] = None, segment_ids: tp.Optional[chex.Array] = None, output_attentions: tp.Optional[bool] = None, output_hidden_states: tp.Optional[bool] = None, past_key_values: tp.Optional[TransformerCache] = None, return_dict: bool = True, ) -> tp.Union[FlaxCausalLMOutput, tp.Tuple]: """ Forward pass through the Cohere module. Args: input_ids (chex.Array): Input tensor containing token IDs. attention_mask (chex.Array): Mask for attention. position_ids (chex.Array): Positional indices. segment_ids (tp.Optional[chex.Array]): Segment IDs for different input parts. inputs_embeds (tp.Optional[chex.Array]): Embedded input tensor. output_attentions (tp.Optional[bool]): If True, output attention weights. output_hidden_states (tp.Optional[bool]): If True, output hidden states. init_cache (bool): If True, initialize cache for decoding. deterministic (bool): If True, disable dropout. return_dict (bool): If True, return a dictionary of outputs. Returns: FlaxCausalLMOutput | tp.Tuple: Model output, either as a named tuple or a standard tuple. """ batch_size, sequence_length = input_ids.shape if attention_mask is None: attention_mask = jnp.ones((batch_size, sequence_length), "b1") else: if attention_mask.dtype != jnp.bool: attention_mask = jnp.astype(attention_mask == 1, "b1") if position_ids is None: position_ids = jnp.broadcast_to( jnp.clip(jnp.cumsum(attention_mask, axis=-1) - 1, a_min=0), (batch_size, sequence_length), ) outputs = self.model( input_ids=input_ids, attention_mask=attention_mask, position_ids=position_ids, output_attentions=output_attentions, output_hidden_states=output_hidden_states, past_key_values=past_key_values, return_dict=return_dict, inputs_embeds=inputs_embeds, segment_ids=segment_ids, ) hidden_states = outputs[0] if self.config.tie_word_embeddings: lm_logits = jax.lax.dot_general( hidden_states, self.model.embed_tokens.embedding.value.T, (((hidden_states.ndim - 1), (0,)), ((), ())), ) else: lm_logits = self.lm_head(hidden_states) lm_logits = (lm_logits * self.logit_scale).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, past_key_values=outputs.past_key_values, )
[docs]@register_module( "sequence-classification", config=CohereConfig, model_type="cohere", embedding_layer_names=["embed_tokens"], ) class CohereForSequenceClassification(EasyDeLBaseModule): def __init__( self, config: CohereConfig, dtype: jnp.dtype = jnp.float32, param_dtype: jnp.dtype = jnp.float32, precision: jax.lax.PrecisionLike = None, *, rngs: nn.Rngs, ): super().__init__( config=config, dtype=dtype, param_dtype=param_dtype, precision=precision, rngs=rngs, ) self.model = CohereModel( config=config, dtype=dtype, param_dtype=param_dtype, precision=precision, rngs=rngs, ) assert hasattr(config, "num_labels"), ( "in order to use `SequenceClassification` Models in `EasyDeL` you first need to attach `num_labels` to model `config`" ) self.score = nn.Linear( config.hidden_size, config.num_labels, dtype=dtype, param_dtype=param_dtype, use_bias=False, kernel_init=jax.nn.initializers.normal(stddev=config.initializer_range), precision=precision, rngs=rngs, ) def __call__( self, input_ids: tp.Optional[chex.Array] = None, inputs_embeds: tp.Optional[chex.Array] = None, attention_mask: tp.Optional[chex.Array] = None, position_ids: tp.Optional[chex.Array] = None, segment_ids: tp.Optional[chex.Array] = None, output_attentions: tp.Optional[bool] = None, output_hidden_states: tp.Optional[bool] = None, past_key_values: tp.Optional[TransformerCache] = None, return_dict: bool = True, ) -> tp.Union[FlaxSequenceClassifierOutput, tp.Tuple]: transformer_outputs = self.model( input_ids=input_ids, attention_mask=attention_mask, position_ids=position_ids, past_key_values=past_key_values, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, inputs_embeds=inputs_embeds, segment_ids=segment_ids, ) hidden_states = transformer_outputs[0] logits = self.score(hidden_states) if input_ids is not None: batch_size = input_ids.shape[0] else: batch_size = inputs_embeds.shape[0] if self.config.pad_token_id is None and batch_size != 1: raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.") if self.config.pad_token_id is None: sequence_lengths = -1 else: if input_ids is not None: sequence_lengths = ( jnp.argmax(jnp.equal(input_ids, self.config.pad_token_id).astype("i4"), -1) - 1 ) sequence_lengths = sequence_lengths % input_ids.shape[-1] else: sequence_lengths = -1 pooled_logits = logits[jnp.arange(batch_size), sequence_lengths] if not return_dict: output = (pooled_logits,) + transformer_outputs[1:] return output return FlaxSequenceClassifierOutput( logits=pooled_logits, past_key_values=past_key_values, hidden_states=transformer_outputs.hidden_states, attentions=transformer_outputs.attentions, )