Source code for easydel.modules.internlm2.modeling_internlm2

# Copyright 2025 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 functools
from typing import ClassVar

import jax
import jax.numpy as jnp
from eformer import common_types
from eformer.escale import apply_logical_sharding
from ejkernel.types import MaskInfo
from flax import nnx as nn
from jax.ad_checkpoint import checkpoint_name
from jaxtyping import Array, Bool, Float, Int

from easydel.infra.base_module import EasyDeLBaseModule
from easydel.infra.factory import TaskType, register_module
from easydel.infra.modeling_outputs import (
    BaseModelOutput,
    CausalLMOutput,
    DecoderLayerOutput,
    SequenceClassifierOutput,
)
from easydel.infra.utils import ACT2FN, auto_remat, block_wise_ffn, get_dot_general_by_bits
from easydel.layers.attention import FlexibleAttentionModule
from easydel.layers.attention_unified import UnifiedAttention
from easydel.layers.base_modules import BaseCausalLMModule
from easydel.layers.caching import (
    RaggedPagesCache,
    RaggedPagesCacheView,
    RaggedPagesMetadata,
    TransformerCache,
    TransformerCacheView,
    TransformerMetadata,
)
from easydel.layers.linear import ColumnParallelLinear, RowParallelLinear
from easydel.layers.norms import RMSNorm

from .internlm2_configuration import InternLM2Config


[docs]class InternLM2Attention(UnifiedAttention): """InternLM2 Attention with full RoPE. Inherits from UnifiedAttention. Uses combined QKV projection (wqkv) instead of separate projections. Overrides forward_standard to efficiently handle fused QKV with GQA layout. """ projection_mapping: ClassVar[dict[str, str]] = { "output_projection": "wo", "query_key_value_projection": "wqkv", } def __init__( self, config: InternLM2Config, dtype: jnp.dtype = jnp.bfloat16, param_dtype: jnp.dtype = jnp.bfloat16, precision: jax.lax.PrecisionLike = None, *, rngs: nn.Rngs, layer_idx: int, ): """Initialize InternLM2 attention.""" super().__init__( config, dtype, param_dtype, precision, rngs=rngs, layer_idx=layer_idx, attention_type="standard", causal=True, use_fused_qkv=True, use_gqa=True, ) def _create_fused_qkv_proj(self, config, dtype, param_dtype, precision, rngs): return ColumnParallelLinear( config.hidden_size, (config.num_attention_heads + 2 * config.num_key_value_heads) * self.head_dim, dtype=dtype, param_dtype=param_dtype, use_bias=config.bias, rngs=rngs, kernel_init=jax.nn.initializers.normal(config.initializer_range), precision=precision, **get_dot_general_by_bits(config.bits, config.easy_method), ) def _create_o_proj(self, config, dtype, param_dtype, precision, rngs): return RowParallelLinear( config.num_attention_heads * self.head_dim, config.hidden_size, dtype=dtype, param_dtype=param_dtype, rngs=rngs, use_bias=config.bias, kernel_init=jax.nn.initializers.normal(config.initializer_range), precision=precision, **get_dot_general_by_bits(config.bits, config.easy_method), ) def _create_rotary(self, config: InternLM2Config, dtype: jnp.dtype): """Create InternLM2-specific rotary embedding with full RoPE.""" return config.get_basic_rope( dtype=dtype, head_size=self.head_dim, rotary_dim=self.head_dim, # Full RoPE base=config.rope_theta, ) def _create_attention_performer(self, config: InternLM2Config, rngs: nn.Rngs): """Create attention performer with zero dropout.""" return FlexibleAttentionModule( rngs=rngs, base_config=config, softmax_scale=self.head_dim**-0.5, dropout_prob=0.0, )
[docs]class InternLM2MLP(nn.Module): """InternLM2 MLP module. Attributes: config (InternLM2Config): Configuration object for the model. dtype (jnp.dtype): Data type for computation. Default is jnp.float32. param_dtype (jnp.dtype): Data type for parameters. Default is jnp.float32. precision (jax.lax.PrecisionLike): Precision setting for JAX operations. Default is None. w1 (ParallelLinear): First linear transformation (gate projection). w3 (ParallelLinear): Second linear transformation (up projection). w2 (ParallelLinear): Third linear transformation (down projection). act_fn (callable): Activation function (e.g., SiLU). """ def __init__( self, config: InternLM2Config, dtype: jnp.dtype = jnp.bfloat16, param_dtype: jnp.dtype = jnp.bfloat16, precision: jax.lax.PrecisionLike = None, *, rngs: nn.Rngs, layer_idx: int, ): """Initializes the InternLM2MLP module. Args: config (InternLM2Config): The configuration object for the InternLM2 model. dtype (jnp.dtype): Data type for computation. Defaults to jnp.float32. param_dtype (jnp.dtype): Data type for parameters. Defaults to jnp.float32. precision (jax.lax.PrecisionLike): Precision setting for JAX operations. Defaults to None. rngs (nn.Rngs): Random number generators. """ self.config = config self.dtype = dtype self.param_dtype = param_dtype self.precision = precision linear = functools.partial( ColumnParallelLinear, 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.w1 = linear(config.hidden_size, config.intermediate_size, rngs=rngs) self.w3 = linear(config.hidden_size, config.intermediate_size, rngs=rngs) self.w2 = linear(config.intermediate_size, config.hidden_size, rngs=rngs) self.act_fn = ACT2FN[config.hidden_act] def __call__( self, hidden_states: Float[Array, "batch seq_len hidden_dim"] ) -> Float[Array, "batch seq_len hidden_dim"]: """Forward pass of the MLP module. Args: hidden_states (jnp.ndarray): Input hidden states. Returns: jnp.ndarray: Output hidden states after MLP transformation. """ hidden_states = apply_logical_sharding( hidden_states, dynamic_axes=common_types.HiddenStateSharding, partition_manager=self.config.partition_manager, ) w1 = self.act_fn(checkpoint_name(self.w1(hidden_states), name="mlp_gate")) w3 = checkpoint_name(self.w3(hidden_states), name="mlp_up") hidden_states = checkpoint_name(self.w2(w1 * w3), name="mlp_down") hidden_states = apply_logical_sharding( hidden_states, dynamic_axes=common_types.HiddenStateSharding, partition_manager=self.config.partition_manager, ) return hidden_states
[docs]class InternLM2Block(nn.Module): """InternLM2 Transformer Block. This module combines the self-attention layer and the MLP layer with residual connections and layer normalization. Attributes: config (InternLM2Config): Configuration object for the model. dtype (jnp.dtype): Data type for computation. Default is jnp.float32. param_dtype (jnp.dtype): Data type for parameters. Default is jnp.float32. precision (jax.lax.PrecisionLike): Precision setting for JAX operations. Default is None. attention (InternLM2Attention): The self-attention module. feed_forward (InternLM2MLP): The feed-forward (MLP) module. attention_norm (RMSNorm): Layer normalization before the attention layer. ffn_norm (RMSNorm): Layer normalization before the MLP layer. """ def __init__( self, config: InternLM2Config, dtype: jnp.dtype = jnp.bfloat16, param_dtype: jnp.dtype = jnp.bfloat16, precision: jax.lax.PrecisionLike = None, *, rngs: nn.Rngs, layer_idx: int, ): """Initializes the InternLM2Block module. Args: config (InternLM2Config): The configuration object for the InternLM2 model. dtype (jnp.dtype): Data type for computation. Defaults to jnp.float32. param_dtype (jnp.dtype): Data type for parameters. Defaults to jnp.float32. precision (jax.lax.PrecisionLike): Precision setting for JAX operations. Defaults to None. rngs (nn.Rngs): Random number generators. """ self.config = config self.dtype = dtype self.param_dtype = param_dtype self.precision = precision attn_block = InternLM2Attention mlp_block = InternLM2MLP attn_block, mlp_block = auto_remat( attn_block, mlp_block, policy=config.gradient_checkpointing, save_names=config.gradient_checkpointing_targets, exclude_names=config.gradient_checkpointing_targets, ) self.attention = attn_block( config=config, dtype=dtype, param_dtype=param_dtype, precision=precision, rngs=rngs, layer_idx=layer_idx, ) self.feed_forward = mlp_block( config=config, dtype=dtype, param_dtype=param_dtype, precision=precision, rngs=rngs, layer_idx=layer_idx, ) self.attention_norm = RMSNorm( dim=config.hidden_size, eps=config.rms_norm_eps, dtype=dtype, param_dtype=param_dtype, rngs=rngs, ) self.ffn_norm = RMSNorm( dim=config.hidden_size, eps=config.rms_norm_eps, dtype=dtype, param_dtype=param_dtype, rngs=rngs, ) def __call__( self, hidden_states: Float[Array, "batch seq_len hidden_dim"], mask_info: MaskInfo | None, position_ids: Int[Array, "batch seq_len"], mode: common_types.RUNTIME_MODE_TYPES, # type:ignore cache_view: TransformerCacheView | RaggedPagesCacheView | None = None, cache_metadata: TransformerMetadata | RaggedPagesMetadata | None = None, output_attentions: bool = False, frequencies: Float[Array, "seq_len head_dim"] | None = None, ): """Forward pass of the InternLM2Block. Applies self-attention, followed by a residual connection and layer normalization, and then applies the MLP layer, followed by another residual connection and layer normalization. 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 (tp.Optional[chex.Array | bool]): Causal mask for autoregressive behavior. cache_view (tp.Optional[TransformerCacheView | RaggedPagesCacheView]): Cache view for attention KVs. cache_metadata (tp.Optional[TransformerMetadata | RaggedPagesMetadata]): Metadata for paged attention. segment_ids (tp.Optional[chex.Array]): Segment IDs (unused in standard InternLM2). output_attentions (bool): Whether to return attention weights. Default is False. fcm_mask (tp.Optional[chex.Array]): Flash Chunking Mask (FCM) for attention. frequencies (tp.Optional[chex.Array]): Precomputed rotary frequency embeddings. Returns: tp.Union[tp.Tuple[chex.Array, chex.Array], tp.Tuple[chex.Array]]: A tuple containing the output hidden states. If `output_attentions` is True, it also includes the attention weights. """ attn_outputs = self.attention( self.attention_norm(hidden_states), mask_info, position_ids, mode, cache_view, cache_metadata, output_attentions, frequencies, ) hidden_states = hidden_states + attn_outputs.attention_output feed_forward_input = self.ffn_norm(hidden_states) if self.config.use_scan_mlp: feed_forward_hidden_states = block_wise_ffn( self.feed_forward, feed_forward_input, self.config.scan_mlp_chunk_size ) else: feed_forward_hidden_states = self.feed_forward(feed_forward_input) hidden_states = hidden_states + feed_forward_hidden_states hidden_states = apply_logical_sharding( hidden_states, dynamic_axes=common_types.HiddenStateSharding, partition_manager=self.config.partition_manager, ) return DecoderLayerOutput( hidden_states=hidden_states, attention_weight=attn_outputs.attention_weight, cache_view=attn_outputs.cache_view, )
[docs]@register_module(TaskType.BASE_MODULE, config=InternLM2Config, model_type="internlm2") class InternLM2Model(EasyDeLBaseModule): """The base InternLM2 model transformer. This class represents the core transformer architecture of the InternLM2 model, consisting of embedding layers, multiple transformer blocks, and a final layer normalization. Attributes: config (InternLM2Config): Configuration object for the model. dtype (jnp.dtype): Data type for computation. Default is jnp.float32. param_dtype (jnp.dtype): Data type for parameters. Default is jnp.float32. precision (jax.lax.PrecisionLike): Precision setting for JAX operations. Default is None. embed_tokens (nn.Embed): Embedding layer for input tokens. layers (tp.Sequence[InternLM2Block]): Sequence of transformer blocks. norm (RMSNorm): Final layer normalization. gradient_checkpointing (EasyDeLGradientCheckPointers): Gradient checkpointing configuration. scan_layers (bool): Whether to use JAX scan for layer processing. blocks_class (InternLM2Block): The class used for the transformer blocks. """ def __init__( self, config: InternLM2Config, dtype: jnp.dtype = jnp.bfloat16, param_dtype: jnp.dtype = jnp.bfloat16, precision: jax.lax.PrecisionLike = None, *, rngs: nn.Rngs, ): """Initializes the InternLM2Model. Args: config (InternLM2Config): The configuration object for the InternLM2 model. dtype (jnp.dtype): Data type for computation. Defaults to jnp.float32. param_dtype (jnp.dtype): Data type for parameters. Defaults to jnp.float32. precision (jax.lax.PrecisionLike): Precision setting for JAX operations. Defaults to None. rngs (nn.Rngs): Random number generators. """ super().__init__( config=config, dtype=dtype, param_dtype=param_dtype, precision=precision, rngs=rngs, ) self.tok_embeddings = nn.Embed( config.vocab_size, config.hidden_size, embedding_init=jax.nn.initializers.normal(stddev=config.initializer_range), dtype=dtype, param_dtype=param_dtype, rngs=rngs, ) self.layers = [ InternLM2Block( config=config, layer_idx=i, rngs=rngs, dtype=dtype, param_dtype=param_dtype, precision=precision, ) for i in range(config.num_hidden_layers) ] self.norm = RMSNorm( dim=config.hidden_size, eps=config.rms_norm_eps, dtype=dtype, param_dtype=param_dtype, rngs=rngs, ) def __call__( self, input_ids: Int[Array, "batch seq_len"] | None = None, inputs_embeds: Float[Array, "batch seq_len hidden_dim"] | None = None, attention_mask: Bool[Array, "batch seq_len"] | None = None, mask_info: MaskInfo | None = None, position_ids: Int[Array, "batch seq_len"] | None = None, mode: common_types.RUNTIME_MODE_TYPES | None = None, # type:ignore past_key_values: TransformerCache | RaggedPagesCache | None = None, cache_metadata: TransformerMetadata | RaggedPagesMetadata | None = None, output_attentions: bool | None = None, output_hidden_states: bool | None = None, ) -> BaseModelOutput: """Forward pass of the InternLM2Model. Args: input_ids (tp.Optional[chex.Array]): Input token IDs. Shape: (batch_size, sequence_length). inputs_embeds (tp.Optional[chex.Array]): Input embeddings. Either `input_ids` or `inputs_embeds` must be provided. attention_mask (tp.Optional[chex.Array]): Mask to avoid performing attention on padding token indices. Shape: (batch_size, sequence_length). position_ids (tp.Optional[chex.Array]): Position indices for the tokens. Shape: (batch_size, sequence_length). segment_ids (tp.Optional[chex.Array]): Segment IDs (unused). output_attentions (tp.Optional[bool]): Whether to return attention weights. Defaults to `config.output_attentions`. output_hidden_states (tp.Optional[bool]): Whether to return hidden states for all layers. Defaults to `config.output_hidden_states`. past_key_values (tp.Optional[TransformerCache | RaggedPagesCache]): Precomputed key/value states for attention. cache_metadata (tp.Optional[TransformerMetadata | RaggedPagesMetadata]): Metadata for paged attention. Returns: BaseModelOutput: The model's output. returns a `BaseModelOutput` object containing `last_hidden_state`, `hidden_states` (optional), and `attentions` (optional). Raises: ValueError: If neither `input_ids` nor `inputs_embeds` is provided. """ if inputs_embeds is None and input_ids is not None: inputs_embeds = self.tok_embeddings(input_ids.astype("i4")) else: raise ValueError("you should specify inputs_embeds or input_ids one of them") sequence_length = inputs_embeds.shape[1] mask_info = MaskInfo.dynamic_init( mask_info=mask_info, input_ids=input_ids, inputs_embeds=inputs_embeds, attention_mask=attention_mask, ) if position_ids is None: position_ids = mask_info.q_position_ids assert sequence_length <= self.config.max_position_embeddings, ( f"Maximum Position Embedding Reached ! " f"(Excepted <= {self.config.max_position_embeddings} got {sequence_length})" ) hidden_states = inputs_embeds if mode is None: mode = ( common_types.MODE_DECODE if sequence_length == 1 and past_key_values is not None else common_types.MODE_TRAIN ) if past_key_values is None: past_key_values = TransformerCache.init_empty(len(self.layers)) hidden_states = apply_logical_sharding( hidden_states, dynamic_axes=common_types.HiddenStateSharding, partition_manager=self.config.partition_manager, ) all_attentions = () if output_attentions else None all_hidden_states = () if output_hidden_states else None for idx, block in enumerate(self.layers): if output_hidden_states: all_hidden_states += (hidden_states,) layer_outputs = block( hidden_states=hidden_states, mask_info=mask_info, position_ids=position_ids, mode=mode, cache_view=past_key_values.views[idx], cache_metadata=cache_metadata, output_attentions=output_attentions, frequencies=self.frequencies, ) hidden_states = layer_outputs.hidden_states if output_attentions: all_attentions += (layer_outputs.attention_weight,) past_key_values[idx] = layer_outputs.cache_view hidden_states = self.norm(hidden_states) if output_hidden_states: all_hidden_states += (hidden_states,) return BaseModelOutput( last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions, past_key_values=past_key_values, )
[docs] def get_encoder(self): """ Returns the encoder part of the model's graph definition. Decoder-Only models don't have an encoder. """ raise NotImplementedError("This is a decoder-only model and does not have an encoder.")
[docs] def get_decoder(self): """ Returns the decoder part of the model's graph definition. """ return self
[docs] def get_lm_head(self): """ Returns the language model head of the module. Base Models don't have a Language Model Head. """ raise NotImplementedError("The base model does not have a language model head.")
[docs] def get_embedding(self): """ Returns the embedding layer of the module. """ return self.tok_embeddings
[docs]@register_module(TaskType.CAUSAL_LM, config=InternLM2Config, model_type="internlm2") class InternLM2ForCausalLM(BaseCausalLMModule[InternLM2Model, InternLM2Config]): """InternLM2 model with a Causal Language Modeling head.""" _task_type = TaskType.CAUSAL_LM _model_type = "internlm2" _config_class = InternLM2Config def __init__( self, config: InternLM2Config, dtype: jnp.dtype = jnp.bfloat16, param_dtype: jnp.dtype = jnp.bfloat16, precision: jax.lax.PrecisionLike = None, *, rngs: nn.Rngs, ): super().__init__( config=config, base_model_class=InternLM2Model, base_model_name="model", dtype=dtype, param_dtype=param_dtype, precision=precision, rngs=rngs, lm_head_bias=False, lm_head_name="output", # InternLM2 uses 'output' not 'lm_head' lm_head_class=RowParallelLinear, # InternLM2 uses RowParallelLinear ) def __call__( self, input_ids: Int[Array, "batch seq_len"] | None = None, inputs_embeds: Float[Array, "batch seq_len hidden_dim"] | None = None, attention_mask: Bool[Array, "batch seq_len"] | None = None, mask_info: MaskInfo | None = None, position_ids: Int[Array, "batch seq_len"] | None = None, mode: common_types.RUNTIME_MODE_TYPES | None = None, # type:ignore past_key_values: TransformerCache | RaggedPagesCache | None = None, cache_metadata: TransformerMetadata | RaggedPagesMetadata | None = None, apply_lm_head: bool = True, output_attentions: bool | None = None, output_hidden_states: bool | None = None, ) -> CausalLMOutput: """Forward pass of the InternLM2ForCausalLM model. Args: input_ids (tp.Optional[chex.Array]): Input token IDs. Shape: (batch_size, sequence_length). inputs_embeds (tp.Optional[chex.Array]): Input embeddings. Either `input_ids` or `inputs_embeds` must be provided. attention_mask (tp.Optional[chex.Array]): Mask to avoid performing attention on padding token indices. Shape: (batch_size, sequence_length). position_ids (tp.Optional[chex.Array]): Position indices for the tokens. Shape: (batch_size, sequence_length). segment_ids (tp.Optional[chex.Array]): Segment IDs (unused). output_attentions (tp.Optional[bool]): Whether to return attention weights. Defaults to `config.output_attentions`. output_hidden_states (tp.Optional[bool]): Whether to return hidden states for all layers. Defaults to `config.output_hidden_states`. past_key_values (tp.Optional[TransformerCache | RaggedPagesCache]): Precomputed key/value states for attention. cache_metadata (tp.Optional[TransformerMetadata | RaggedPagesMetadata]): Metadata for paged attention. Returns: CausalLMOutput: The model's output. returns a `CausalLMOutput` object containing `logits`, `hidden_states` (optional), and `attentions` (optional). """ outputs = self.model( input_ids=input_ids, attention_mask=attention_mask, mask_info=mask_info, position_ids=position_ids, output_attentions=output_attentions, output_hidden_states=output_hidden_states, mode=mode, past_key_values=past_key_values, cache_metadata=cache_metadata, inputs_embeds=inputs_embeds, ) hidden_states = outputs.last_hidden_state hidden_states = apply_logical_sharding( hidden_states, dynamic_axes=common_types.HiddenStateSharding, partition_manager=self.config.partition_manager, ) lm_logits = None if apply_lm_head: lm_logits = self.apply_lm_head(hidden_states) return CausalLMOutput( logits=lm_logits, hidden_states=outputs.hidden_states, last_hidden_state=outputs.last_hidden_state, attentions=outputs.attentions, past_key_values=outputs.past_key_values, )
[docs] def get_encoder(self): """ Returns the encoder part of the model's graph definition. Decoder-Only models don't have an encoder. """ raise NotImplementedError("This is a decoder-only model and does not have an encoder.")
[docs] def get_decoder(self): """ Returns the decoder part of the model's graph definition. """ return self.model.get_decoder()
[docs] def get_lm_head(self): """ Returns the language model head of the module. """ return self.output
[docs] def get_embedding(self): """ Returns the embedding layer of the module. """ return self.model.get_embedding()
[docs]@register_module(TaskType.SEQUENCE_CLASSIFICATION, config=InternLM2Config, model_type="internlm2") class InternLM2ForSequenceClassification(EasyDeLBaseModule): """InternLM2 model with a Sequence Classification head. This model consists of the base InternLM2 transformer (`InternLM2Model`) followed by a linear layer (`score`) that projects the transformer's output hidden states (typically the hidden state of the first token) to the number of classes for classification. Attributes: config (InternLM2Config): Configuration object for the model. dtype (jnp.dtype): Data type for computation. Default is jnp.float32. param_dtype (jnp.dtype): Data type for parameters. Default is jnp.float32. precision (jax.lax.PrecisionLike): Precision setting for JAX operations. Default is None. rngs (nn.Rngs): Random number generators. module (InternLM2Model): The core InternLM2 transformer model. score (ParallelLinear): The linear layer for classification. """ def __init__( self, config: InternLM2Config, dtype: jnp.dtype = jnp.bfloat16, param_dtype: jnp.dtype = jnp.bfloat16, precision: jax.lax.PrecisionLike = None, *, rngs: nn.Rngs, ): """Initializes the InternLM2ForSequenceClassification model. Args: config (InternLM2Config): The configuration object for the InternLM2 model. dtype (jnp.dtype): Data type for computation. Defaults to jnp.float32. param_dtype (jnp.dtype): Data type for parameters. Defaults to jnp.float32. precision (jax.lax.PrecisionLike): Precision setting for JAX operations. Defaults to None. rngs (nn.Rngs): Random number generators. """ super().__init__( config=config, dtype=dtype, param_dtype=param_dtype, precision=precision, rngs=rngs, ) self.model = InternLM2Model( 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 = ColumnParallelLinear( self.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=self.precision, rngs=rngs, ) def __call__( self, input_ids: Int[Array, "batch seq_len"] | None = None, inputs_embeds: Float[Array, "batch seq_len hidden_dim"] | None = None, attention_mask: Bool[Array, "batch seq_len"] | None = None, mask_info: MaskInfo | None = None, position_ids: Int[Array, "batch seq_len"] | None = None, mode: common_types.RUNTIME_MODE_TYPES | None = None, # type:ignore past_key_values: TransformerCache | RaggedPagesCache | None = None, cache_metadata: TransformerMetadata | RaggedPagesMetadata | None = None, output_attentions: bool | None = None, output_hidden_states: bool | None = None, ) -> SequenceClassifierOutput: """Forward pass of the InternLM2ForSequenceClassification model. Args: input_ids (tp.Optional[chex.Array]): Input token IDs. Shape: (batch_size, sequence_length). inputs_embeds (tp.Optional[chex.Array]): Input embeddings. Either `input_ids` or `inputs_embeds` must be provided. attention_mask (tp.Optional[chex.Array]): Mask to avoid performing attention on padding token indices. Shape: (batch_size, sequence_length). position_ids (tp.Optional[chex.Array]): Position indices for the tokens. Shape: (batch_size, sequence_length). segment_ids (tp.Optional[chex.Array]): Segment IDs (unused). past_key_values (tp.Optional[TransformerCache | RaggedPagesCache]): Precomputed key/value states for attention. cache_metadata (tp.Optional[TransformerMetadata | RaggedPagesMetadata]): Metadata for paged attention. output_attentions (tp.Optional[bool]): Whether to return attention weights. Defaults to `config.output_attentions`. output_hidden_states (tp.Optional[bool]): Whether to return hidden states for all layers. Defaults to `config.output_hidden_states`. Returns: SequenceClassifierOutput: The model's output, returns a `SequenceClassifierOutput` object containing `logits`, `hidden_states` (optional), and `attentions` (optional). """ transformer_outputs = self.model( input_ids=input_ids, attention_mask=attention_mask, mask_info=mask_info, position_ids=position_ids, mode=mode, past_key_values=past_key_values, cache_metadata=cache_metadata, output_attentions=output_attentions, output_hidden_states=output_hidden_states, inputs_embeds=inputs_embeds, ) hidden_states = transformer_outputs.last_hidden_state 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] return SequenceClassifierOutput( logits=pooled_logits, past_key_values=past_key_values, hidden_states=transformer_outputs.hidden_states, attentions=transformer_outputs.attentions, )
[docs] def get_encoder(self): """ Returns the encoder part of the model's graph definition. Decoder-Only models don't have an encoder. """ raise NotImplementedError("This is a decoder-only model and does not have an encoder.")
[docs] def get_decoder(self): """ Returns the decoder part of the model's graph definition. """ return self.model
[docs] def get_lm_head(self): """ Returns the language model head of the module. This model has a sequence classification head, not an LM Head. """ raise NotImplementedError("This model has a sequence classification head, not a language model head.")
[docs] def get_embedding(self): """ Returns the embedding layer of the module. """ return self.model.get_embedding()