# 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 math
import typing as tp
from functools import partial
import chex
import jax
import jax.numpy as jnp
from eformer import common_types
from eformer.escale import apply_logical_sharding
from eformer.pytree import auto_pytree
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 (
AttentionLayerOutput,
BaseModelOutput,
DecoderLayerOutput,
EncoderLayerOutput,
ModelOutput,
VLMCausalLMOutput,
)
from easydel.infra.utils import ACT2FN, ArrayParam, auto_remat, get_dot_general_by_bits
from easydel.layers.attention import AttentionModule, FlexibleAttentionModule
from easydel.layers.attention_unified import UnifiedAttention
from easydel.layers.base_modules import BaseCausalLMModule, BaseSequenceClassificationModule
from easydel.layers.caching import (
RaggedPagesCache,
RaggedPagesCacheView,
RaggedPagesMetadata,
TransformerCache,
TransformerCacheView,
TransformerMetadata,
)
from easydel.layers.linear import ColumnParallelLinear, RowParallelLinear
from easydel.layers.norms import RMSNorm as Llama4TextRMSNorm
from easydel.utils.compiling_utils import ejit
from .llama4_configuration import Llama4Config, Llama4TextConfig, Llama4VisionConfig
[docs]@auto_pytree
class Llama4CausalLMOutputWithPast(ModelOutput):
"""
Base class for Llama4Vision causal language model (or autoregressive) outputs.
Args:
loss (`chex.Array` of shape `(1,)`, *optional*, returned when `labels` is provided):
Language modeling loss (for next-token prediction).
logits (`chex.Array` of shape `(batch_size, sequence_length, config.vocab_size)`):
Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
past_key_values (`tuple(tuple(chex.Array))`, *optional*, returned when `use_cache=True` is passed or
when `config.use_cache=True`):
Tuple of `tuple(chex.Array)` of length `config.n_layers`, with each tuple having 2 tensors of shape
`(batch_size, num_heads, sequence_length, embed_size_per_head)`)
Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
`past_key_values` input) to speed up sequential decoding.
hidden_states (`tuple(chex.Array)`, *optional*, returned when `output_hidden_states=True` is passed or
when `config.output_hidden_states=True`):
Tuple of `chex.Array` (one for the output of the embeddings, if the model has an embedding layer, +
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
attentions (`tuple(chex.Array)`, *optional*, returned when `output_attentions=True` is passed or when
`config.output_attentions=True`):
Tuple of `chex.Array` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
sequence_length)`.
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
heads.
image_hidden_states (`chex.Array`, *optional*):
A `chex.Array` of size (batch_size * num_patches, num_images, sequence_length, hidden_size)`.
image_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.
"""
loss: chex.Array | None = None
logits: chex.Array = None
past_key_values: TransformerCache | None = None
hidden_states: tuple[chex.Array] | None = None
attentions: tuple[chex.Array] | None = None
image_hidden_states: Float[Array, "batch seq_len hidden_dim"] | None = None
[docs]def bmm(inputs, kernel, precision):
"""Batch matrix multiplication helper that works for 2D or higher-rank inputs."""
subscript = "...ik,...kj->...ij" if inputs.ndim > 1 else "...k,...kj->...j"
return jnp.einsum(
subscript,
inputs,
kernel,
precision=precision,
optimize=True,
)
@ejit(static_argnums=(0, 1, 2, 3))
def _vision_freqs(idx, hidden_size, num_attention_heads, rope_theta):
"""Compute rotary frequencies for the vision transformer grid."""
img_idx = jnp.arange(idx**2, dtype="i4").reshape(idx**2, 1)
img_idx = jnp.concatenate([img_idx, img_idx[:1]], axis=0)
img_idx = img_idx.at[-1, -1].set(-2)
frequencies_x = img_idx % idx
frequencies_y = img_idx // idx
freq_dim = hidden_size // num_attention_heads // 2
rope_arange = jnp.arange(0, freq_dim, 2)
rope_arange_sliced = rope_arange[: (freq_dim // 2)]
rope_freq = 1.0 / (rope_theta ** (rope_arange_sliced.astype("f4") / freq_dim))
rope_freq_broadcast = rope_freq[None, None, :]
freqs_x = jnp.repeat((frequencies_x + 1).astype("f4")[..., None] * rope_freq_broadcast, 2, axis=-1)
freqs_y = jnp.repeat((frequencies_y + 1).astype("f4")[..., None] * rope_freq_broadcast, 2, axis=-1)
freqs = jnp.concatenate([freqs_x, freqs_y], axis=-1)[..., ::2]
freqs = jnp.where(img_idx.reshape(-1, 1, 1) < 0, 0.0, freqs)
return jnp.exp(1j * freqs)
def _create_chunked_attention_mask(
attention_chunk_size: int,
start: int,
end: int,
):
"""Create a chunked causal attention mask for sliding window attention."""
blcok_position = jnp.abs(
(jnp.arange(start, end)[None, :] // attention_chunk_size)
- jnp.arange(start, end)[:, None] // attention_chunk_size
)
token_position = jnp.arange(start, end)[None, :] - jnp.arange(start, end)[:, None]
return ((blcok_position == 0) & (token_position <= 0)).astype("b1")
[docs]class Llama4TextExperts(nn.Module):
"""Mixture of Experts module for Llama4 text models.
Implements a sparse mixture of experts with top-k routing,
enabling efficient scaling and specialization of model capacity.
"""
def __init__(
self,
config: Llama4Config,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.num_experts = config.num_local_experts
self.intermediate_size = config.intermediate_size
self.hidden_size = config.hidden_size
self.expert_dim = self.intermediate_size
self.gate_up_proj = ArrayParam.bound(
shape=(self.num_experts, self.hidden_size, 2 * self.expert_dim),
dtype=self.param_dtype,
init_method="normal",
init_kwargs={"stddev": config.initializer_range},
key=rngs.params(),
)
self.down_proj = ArrayParam.bound(
shape=(self.num_experts, self.expert_dim, self.hidden_size),
dtype=self.param_dtype,
init_method="normal",
init_kwargs={"stddev": config.initializer_range},
key=rngs.params(),
)
self.activation_fn = ACT2FN[self.config.hidden_act]
def __call__(self, hidden_states: Float[Array, "batch seq_len hidden_dim"]) -> jnp.ndarray:
hidden_states = hidden_states.reshape(self.num_experts, -1, self.hidden_size)
gate_up = bmm(hidden_states, self.gate_up_proj, self.precision)
gate, up = jnp.split(gate_up, 2, axis=-1)
next_states = bmm((up * self.activation_fn(gate)), self.down_proj, self.precision)
return next_states.reshape(-1, self.hidden_size)
[docs]class Llama4TextL2Norm(nn.Module):
"""L2 normalization layer for Llama4 text models.
Normalizes inputs using L2 norm with learned scaling parameters,
providing stable gradients during training.
"""
kernel_init = staticmethod(nn.initializers.ones)
def __init__(self, eps: float = 1e-6) -> None:
self.eps = eps
def _norm(self, x: jnp.ndarray) -> jnp.ndarray:
return x * jax.lax.rsqrt(jnp.square(x).mean(-1, keepdims=True) + self.eps)
@jax.named_scope("easydel-L2norm")
def __call__(self, x: jnp.ndarray) -> jnp.ndarray:
return self._norm(x.astype(jnp.float32)).astype(x.dtype)
[docs]class Llama4TextMLP(nn.Module):
"""Multi-Layer Perceptron for Llama4 text models.
Implements feedforward network with SwiGLU activation function
for improved representation learning.
"""
def __init__(
self,
config: Llama4Config,
intermediate_size=None,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
if intermediate_size is None:
intermediate_size = config.intermediate_size
column_parallel_linear = partial(
ColumnParallelLinear,
dtype=dtype,
param_dtype=param_dtype,
use_bias=False,
kernel_init=jax.nn.initializers.normal(config.initializer_range),
precision=precision,
rngs=rngs,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
row_parallel_linear = partial(
RowParallelLinear,
dtype=dtype,
param_dtype=param_dtype,
use_bias=False,
kernel_init=jax.nn.initializers.normal(config.initializer_range),
precision=precision,
rngs=rngs,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
self.gate_proj = column_parallel_linear(config.hidden_size, intermediate_size)
self.down_proj = row_parallel_linear(intermediate_size, config.hidden_size)
self.up_proj = column_parallel_linear(config.hidden_size, intermediate_size)
self.activation_fn = ACT2FN[self.config.hidden_act]
def __call__(self, hidden_states: Float[Array, "batch seq_len hidden_dim"]) -> jnp.ndarray:
gate = checkpoint_name(self.activation_fn(self.gate_proj(hidden_states)), "mlp_gate")
up = checkpoint_name(self.up_proj(hidden_states), "mlp_up")
hidden_states = checkpoint_name(self.down_proj(gate * up), "mlp_down")
return checkpoint_name(hidden_states, "mlp_output")
[docs]class Llama4TextMoe(nn.Module):
"""Mixture of Experts layer for Llama4 text models.
Routes inputs to specialized expert networks based on learned routing,
allowing for conditional computation and increased model capacity.
"""
def __init__(
self,
config: Llama4Config,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
self.top_k = config.num_experts_per_tok
self.hidden_dim = config.hidden_size
self.num_experts = config.num_local_experts
self.experts = Llama4TextExperts(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.router = ColumnParallelLinear(
config.hidden_size,
config.num_local_experts,
use_bias=False,
precision=precision,
rngs=rngs,
dtype=dtype,
param_dtype=param_dtype,
kernel_init=jax.nn.initializers.normal(config.initializer_range),
**get_dot_general_by_bits(config.bits, config.easy_method),
)
self.shared_expert = Llama4TextMLP(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
def __call__(
self, hidden_states: Float[Array, "batch seq_len hidden_dim"]
) -> Float[Array, "batch seq_len hidden_dim"]:
batch, seq_len, hidden_dim = hidden_states.shape
assert hidden_dim == self.hidden_dim, "Input hidden_dim mismatch"
# Reshape to [batch*seq_len, hidden_dim]
flattened_hidden_states = hidden_states.reshape(-1, self.hidden_dim)
tokens_per_expert = flattened_hidden_states.shape[0]
router_logits = checkpoint_name(self.router(flattened_hidden_states), "moe_router_logits")
router_top_value, router_indices_topk = jax.lax.top_k(router_logits, self.top_k)
scores_base = jnp.full_like(router_logits, -jnp.inf)
token_idx = jnp.arange(tokens_per_expert)[:, None]
expert_idx = router_indices_topk
scores_scattered = scores_base.at[token_idx, expert_idx].set(router_top_value)
router_scores = jax.nn.sigmoid(scores_scattered.astype(jnp.float32)).astype(hidden_states.dtype)
out = self.shared_expert(flattened_hidden_states)
expert_outputs = jnp.zeros_like(out)
for expert_idx in range(self.num_experts):
expert_mask = router_scores[:, expert_idx : expert_idx + 1]
expert_inputs = flattened_hidden_states * expert_mask
expert_output = self.experts(expert_inputs)
expert_outputs = expert_outputs + expert_output
final_output = checkpoint_name(out + expert_outputs, "moe_expert_output")
final_output = final_output.reshape(batch, seq_len, hidden_dim)
router_scores_transposed = router_scores.T
return final_output, router_scores_transposed
[docs]class Llama4TextAttention(UnifiedAttention):
"""Attention module for the Llama4 text decoder with optional sliding windows."""
def __init__(
self,
config: Llama4TextConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
):
self.use_rope = (layer_idx + 1) % 4 != 0
self.attn_scale = config.attn_scale
self.floor_scale = config.floor_scale
self.attn_temperature_tuning = config.attn_temperature_tuning
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
layer_idx=layer_idx,
attention_type="standard",
causal=False,
)
self.qk_norm = Llama4TextL2Norm() if config.use_qk_norm and self.use_rope else None
self._cached_position_ids: Int[Array, "batch seq_len"] | None = None
def _create_attention_performer(self, config: Llama4TextConfig, rngs: nn.Rngs):
return FlexibleAttentionModule(
rngs=rngs,
base_config=config,
softmax_scale=self.head_dim**-0.5,
dropout_prob=0.0,
)
def _create_rotary(self, config: Llama4TextConfig, dtype: jnp.dtype):
# RoPE is handled via custom complex rotary frequencies when enabled.
return None if not self.use_rope else super()._create_rotary(config, dtype)
def _apply_rotary(
self,
query_states: Float[Array, "batch seq_len num_heads head_dim"],
key_states: Float[Array, "batch seq_len num_kv_heads head_dim"],
position_ids: Int[Array, "batch seq_len"],
frequencies: Float[Array, "seq_len head_dim"] | None = None,
) -> tuple[Float[Array, "batch seq_len num_heads head_dim"], Float[Array, "batch seq_len num_kv_heads head_dim"]]:
if not self.use_rope:
return query_states, key_states
if frequencies is not None:
return self.apply_complex_rotary(query_states, key_states, frequencies)
return super()._apply_rotary(query_states, key_states, position_ids, frequencies)
def _postprocess_qkv(
self,
query_states: Float[Array, "batch seq_len num_heads head_dim"],
key_states: Float[Array, "batch seq_len num_kv_heads head_dim"],
value_states: Float[Array, "batch seq_len num_kv_heads head_dim"],
) -> tuple[
Float[Array, "batch seq_len num_heads head_dim"],
Float[Array, "batch seq_len num_kv_heads head_dim"],
Float[Array, "batch seq_len num_kv_heads head_dim"],
]:
if self.qk_norm is not None:
query_states = self.qk_norm(query_states)
key_states = self.qk_norm(key_states)
if self.attn_temperature_tuning and not self.use_rope and self._cached_position_ids is not None:
attn_scales = (
jnp.log(jnp.floor((self._cached_position_ids.astype("f4") + 1.0) / self.floor_scale) + 1.0)
* self.attn_scale
+ 1.0
)
attn_scales = attn_scales.reshape((*attn_scales.shape, 1, 1))
query_states = (query_states * attn_scales).astype(query_states.dtype)
return query_states, key_states, value_states
[docs] def forward(
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,
alibi: Float[Array, "batch_or_1 heads qseq_len_or_1 kvseq_len_or_1"] | None = None,
) -> AttentionLayerOutput:
self._cached_position_ids = position_ids if (self.attn_temperature_tuning and not self.use_rope) else None
try:
return super().forward(
hidden_states,
mask_info,
position_ids,
mode,
cache_view,
cache_metadata,
output_attentions,
frequencies,
alibi,
)
finally:
self._cached_position_ids = None
[docs]class Llama4TextDecoderLayer(nn.Module):
"""Single Llama4 text decoder block combining attention and MLP."""
def __init__(
self,
config: Llama4TextConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
):
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
attn_block = Llama4TextAttention
mlp_block = Llama4TextMLP
moe_block = Llama4TextMoe
attn_block, mlp_block, moe_block = auto_remat(
attn_block,
mlp_block,
moe_block,
policy=config.gradient_checkpointing,
save_names=config.gradient_checkpointing_targets,
exclude_names=config.gradient_checkpointing_targets,
)
self.self_attn = attn_block(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
layer_idx=layer_idx,
)
self.use_chunked_attention = int((layer_idx + 1) % 4 != 0) # <=> use rope
self.is_moe_layer = layer_idx in config.moe_layers
if self.is_moe_layer: # the 128E model interleaves dense / sparse
self.feed_forward = moe_block(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
else:
self.feed_forward = mlp_block(
config=config,
intermediate_size=config.intermediate_size_mlp,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.input_layernorm = Llama4TextRMSNorm(
dim=config.hidden_size,
eps=config.rms_norm_eps,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.post_attention_layernorm = Llama4TextRMSNorm(
dim=config.hidden_size,
eps=config.rms_norm_eps,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.layer_idx = layer_idx
def __call__(
self,
hidden_states: Float[Array, "batch seq_len hidden_dim"],
mask_info: MaskInfo,
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,
output_router_logits: bool = False,
frequencies: Float[Array, "seq_len head_dim"] | None = None,
):
attn_outputs = self.self_attn(
self.input_layernorm(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.post_attention_layernorm(hidden_states)
# TODO: Support Chunked MLP for LLaMA4
# 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)
if self.is_moe_layer:
feed_forward_hidden_states, router_logits = feed_forward_hidden_states
else:
router_logits = None
hidden_states = hidden_states + feed_forward_hidden_states.reshape(feed_forward_input.shape)
return DecoderLayerOutput(
hidden_states=hidden_states,
attention_weight=attn_outputs.attention_weight,
router_logits=router_logits if output_router_logits else None,
cache_view=attn_outputs.cache_view,
)
[docs]@register_module(TaskType.BASE_MODULE, config=Llama4TextConfig, model_type="llama4_text")
class Llama4TextModel(EasyDeLBaseModule):
"""Decoder-only Llama4 text model built from embeddings and decoder blocks."""
def __init__(
self,
config: Llama4TextConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
embed_block = auto_remat(
nn.Embed,
policy=self.config.gradient_checkpointing,
save_names=config.gradient_checkpointing_targets,
exclude_names=config.gradient_checkpointing_targets,
)
self.embed_tokens = embed_block(
num_embeddings=self.config.vocab_size,
features=self.config.hidden_size,
dtype=dtype,
param_dtype=param_dtype,
embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
rngs=rngs,
)
self.layers = [
Llama4TextDecoderLayer(
config=config,
layer_idx=layer_idx,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
for layer_idx in range(self.config.num_hidden_layers)
]
self.norm = Llama4TextRMSNorm(
dim=self.config.hidden_size,
eps=self.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 through the Llama model.
Args:
input_ids (chex.Array, optional): Input token IDs, shape (batch_size, sequence_length).
inputs_embeds (chex.Array, optional): Input embeddings, shape (batch_size, sequence_length, hidden_size).
attention_mask (chex.Array, optional): Mask to avoid attention on padding tokens.
position_ids (chex.Array, optional): Indices of positions of each input sequence token.
segment_ids (chex.Array, optional): Segment token indices for segment embeddings.
past_key_values (TransformerCache | RaggedPagesCache, optional):
Cache containing precomputed key/value states.
cache_metadata (TransformerMetadata | RaggedPagesMetadata, optional): Metadata for cache handling.
output_attentions (bool, optional): Whether to return attention weights.
output_hidden_states (bool, optional): Whether to return hidden states of all layers.
Returns:
Union[BaseModelOutput, Tuple]: Model outputs (last hidden state, optional hidden states, optional attentions)
"""
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"))
sequence_length = inputs_embeds.shape[1]
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 ! "
f"(Excepted <= {self.config.max_position_embeddings} got {sequence_length})"
)
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
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,
)
mask_info = mask_info.apply_chunked(self.config.attention_chunk_size)
# causal_mask = jnp.expand_dims(
# _create_chunked_attention_mask(
# self.config.attention_chunk_size,
# 0,
# sequence_length,
# ),
# (0, 1),
# )
frequencies = self.compute_complex_rotary(position_ids)
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=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.embed_tokens
[docs]@register_module(TaskType.CAUSAL_LM, config=Llama4TextConfig, model_type="llama4_text")
class Llama4ForCausalLM(BaseCausalLMModule[Llama4TextModel, Llama4TextConfig]):
"""Llama4 model with a Causal Language Modeling head."""
_task_type = TaskType.CAUSAL_LM
_model_type = "llama4_text"
_config_class = Llama4TextConfig
def __init__(
self,
config: Llama4TextConfig,
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=Llama4TextModel,
base_model_name="model",
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
lm_head_bias=False,
)
[docs]@register_module(TaskType.SEQUENCE_CLASSIFICATION, config=Llama4TextConfig, model_type="llama4_text")
class Llama4ForSequenceClassification(BaseSequenceClassificationModule[Llama4TextModel, Llama4TextConfig]):
"""Llama4 model for sequence classification tasks."""
_task_type = TaskType.SEQUENCE_CLASSIFICATION
_model_type = "llama4_text"
_config_class = Llama4TextConfig
def __init__(
self,
config: Llama4TextConfig,
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=Llama4TextModel,
base_model_name="model",
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
classifier_name="score",
classifier_bias=False,
)
[docs]class Llama4MultiModalProjector(nn.Module):
"""Multi-modal projector for Llama4 vision-language models.
Projects vision features into the text embedding space using MLP layers,
enabling cross-modal understanding and generation.
"""
def __init__(
self,
config,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
self.linear_1 = RowParallelLinear(
config.vision_config.vision_output_dim,
config.get_text_config().hidden_size,
use_bias=False,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
kernel_init=jax.nn.initializers.normal(0.01),
)
def __call__(self, hidden_states: Float[Array, "batch seq_len hidden_dim"]) -> chex.Array:
return self.linear_1(hidden_states)
[docs]def pixel_shuffle(input_tensor, shuffle_ratio):
"""Rearrange flattened vision tokens to a denser spatial grid."""
batch_size, num_patches, channels = input_tensor.shape
patch_size = int(math.sqrt(num_patches))
input_tensor = input_tensor.reshape(batch_size, patch_size, patch_size, -1)
batch_size, height, width, channels = input_tensor.shape
reshaped_tensor = input_tensor.reshape(
batch_size,
height,
int(width * shuffle_ratio),
int(channels / shuffle_ratio),
)
reshaped_tensor = jnp.transpose(reshaped_tensor, (0, 2, 1, 3))
reshaped_tensor = reshaped_tensor.reshape(
batch_size,
int(height * shuffle_ratio),
int(width * shuffle_ratio),
int(channels / (shuffle_ratio**2)),
)
reshaped_tensor = jnp.transpose(reshaped_tensor, (0, 2, 1, 3))
output_tensor = reshaped_tensor.reshape(batch_size, -1, reshaped_tensor.shape[-1])
return output_tensor
[docs]class Llama4VisionPixelShuffleMLP(nn.Module):
"""Pixel shuffle MLP for Llama4 vision models.
Performs spatial downsampling of vision features through pixel shuffling
and MLP transformations for efficient processing.
"""
def __init__(
self,
config,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
self.pixel_shuffle_ratio = config.pixel_shuffle_ratio
self.inner_dim = int(config.projector_input_dim // (self.pixel_shuffle_ratio**2))
self.output_dim = config.projector_output_dim
self.mlp = Llama4VisionMLP2(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
def __call__(self, encoded_patches: chex.Array) -> chex.Array:
encoded_patches = pixel_shuffle(encoded_patches, self.pixel_shuffle_ratio)
return self.mlp(encoded_patches)
[docs]def reshape_for_broadcast(frequencies: jax.Array, query: jax.Array) -> jax.Array:
"""Reshape rotary frequencies so they broadcast over the complex query tensor."""
ndim = query.ndim
return jnp.reshape(
frequencies,
[d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(query.shape)],
)
[docs]def vision_apply_rotary_emb(
query: jax.Array,
key: jax.Array,
frequencies: jax.Array,
) -> tuple[jax.Array, jax.Array]:
"""Apply rotary position embeddings to complex-valued vision queries and keys."""
query_dtype = query.dtype
key_dtype = key.dtype
query_reshaped = query.astype(jnp.float32).reshape((*query.shape[:-1], -1, 2))
key_reshaped = key.astype(jnp.float32).reshape((*key.shape[:-1], -1, 2))
query_complex = jax.lax.complex(query_reshaped[..., 0], query_reshaped[..., 1])
key_complex = jax.lax.complex(key_reshaped[..., 0], key_reshaped[..., 1])
frequencies_broadcast = reshape_for_broadcast(frequencies, query_complex)
query_rotated = query_complex * frequencies_broadcast
key_rotated = key_complex * frequencies_broadcast
query_out_real_imag = jnp.stack(
[jnp.real(query_rotated), jnp.imag(query_rotated)],
axis=-1,
)
key_out_real_imag = jnp.stack([jnp.real(key_rotated), jnp.imag(key_rotated)], axis=-1)
query_out = query_out_real_imag.reshape(query.shape)
key_out = key_out_real_imag.reshape(key.shape)
return query_out.astype(query_dtype), key_out.astype(key_dtype)
[docs]class Llama4VisionAttention(AttentionModule):
"""Attention module for the Llama4 vision transformer."""
def __init__(
self,
config: Llama4VisionConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
):
super().__init__(config=config)
self.layer_idx = layer_idx
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
self.embed_dim = config.hidden_size
self.num_heads = config.num_attention_heads
self.head_dim = config.hidden_size // config.num_attention_heads
self.num_key_value_groups = 1
self.attention_dropout = config.attention_dropout
linear_class = partial(
ColumnParallelLinear,
dtype=dtype,
param_dtype=param_dtype,
use_bias=True,
kernel_init=jax.nn.initializers.normal(config.initializer_range),
precision=precision,
rngs=rngs,
**get_dot_general_by_bits(config.bits, config.easy_method),
)
self.q_proj = linear_class(self.embed_dim, self.num_heads * self.head_dim)
self.k_proj = linear_class(self.embed_dim, self.num_heads * self.head_dim)
self.v_proj = linear_class(self.embed_dim, self.num_heads * self.head_dim)
self.o_proj = linear_class(self.num_heads * self.head_dim, self.embed_dim)
self.attention_performer = FlexibleAttentionModule(
rngs=rngs,
base_config=self.config,
softmax_scale=self.head_dim**-0.5,
dropout_prob=0.0,
)
def __call__(
self,
hidden_states: Float[Array, "batch seq_len hidden_dim"],
frequencies: Float[Array, "seq_len head_dim"] | None = None,
output_attentions: bool = False,
) -> AttentionLayerOutput:
input_shape = hidden_states.shape[:-1]
hidden_shape = (*input_shape, -1, self.head_dim)
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(*hidden_shape)
key_states = key_states.reshape(*hidden_shape)
value_states = value_states.reshape(*hidden_shape)
query_states, key_states = vision_apply_rotary_emb(
query_states,
key_states,
frequencies=frequencies,
)
attentions = self.attention_performer.forward(
query_states=query_states,
key_states=key_states,
value_states=value_states,
mode=common_types.MODE_TRAIN,
bias=None,
cache_metadata=None,
cache_view=None,
init_bias=None,
mask_info=None,
causal=False,
)
attn_output = attentions.attention_outputs.reshape(*input_shape, -1)
attn_output = self.shard_attention_prod(attn_output)
attn_output = self.o_proj(attn_output)
return AttentionLayerOutput(
attention_output=attn_output,
attention_weight=attentions.attention_weights if output_attentions else None,
)
[docs]class Llama4VisionMLP2(nn.Module):
"""Two-layer MLP module for Llama4 vision models.
Implements a simple two-layer feedforward network with GELU activation
for vision feature transformation.
"""
def __init__(
self,
config,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
self.activation_fn = ACT2FN["gelu"]
linear_class = partial(
ColumnParallelLinear,
use_bias=False,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
kernel_init=jax.nn.initializers.normal(0.01),
)
self.hidden_size = config.hidden_size
self.intermediate_size = config.intermediate_size
self.fc1 = linear_class(self.intermediate_size, config.projector_input_dim)
self.fc2 = linear_class(config.projector_output_dim, config.projector_output_dim)
def __call__(self, hidden_states: Float[Array, "batch seq_len hidden_dim"]) -> chex.Array:
hidden_states = self.fc2(self.activation_fn(self.fc1(hidden_states)))
return self.activation_fn(hidden_states)
[docs]class Llama4VisionMLP(nn.Module):
"""MLP module for Llama4 vision transformer.
Standard feedforward network with GELU activation for vision
feature transformation within transformer blocks.
"""
def __init__(
self,
config,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.rngs = rngs
linear_class = partial(
ColumnParallelLinear,
use_bias=True,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
kernel_init=jax.nn.initializers.normal(0.01),
)
self.fc1 = linear_class(config.hidden_size, config.intermediate_size)
self.fc2 = linear_class(config.intermediate_size, config.hidden_size)
self.activation_fn = ACT2FN["gelu"]
def __call__(self, hidden_states: Float[Array, "batch seq_len hidden_dim"]) -> chex.Array:
hidden_states = self.fc1(hidden_states)
hidden_states = self.activation_fn(hidden_states)
hidden_states = self.fc2(hidden_states)
return hidden_states
[docs]class Llama4VisionEncoderLayer(nn.Module):
"""Single encoder layer for Llama4 vision models.
Combines self-attention and feedforward networks with layer normalization
and residual connections for vision feature encoding.
"""
def __init__(
self,
config: Llama4VisionConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
layer_idx: int,
):
self.config = config
self.layer_idx = layer_idx
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
attn_block = Llama4VisionAttention
mlp_block = Llama4VisionMLP
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.self_attn = attn_block(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
layer_idx=layer_idx,
)
self.mlp = mlp_block(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.input_layernorm = nn.LayerNorm(
num_features=config.hidden_size,
epsilon=0.00001,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.post_attention_layernorm = nn.LayerNorm(
num_features=config.hidden_size,
epsilon=0.00001,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.layer_idx = layer_idx
def __call__(
self,
hidden_states: Float[Array, "batch seq_len hidden_dim"],
output_attentions: bool = False,
frequencies: Float[Array, "seq_len head_dim"] | None = None,
):
residual = hidden_states
hidden_states = self.input_layernorm(hidden_states)
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
attn_outputs = self.self_attn(
hidden_states,
frequencies,
output_attentions,
)
hidden_states = residual + attn_outputs.attention_output
residual = hidden_states
hidden_states = self.post_attention_layernorm(hidden_states)
hidden_states = self.mlp(hidden_states)
hidden_states = residual + hidden_states
hidden_states = apply_logical_sharding(
hidden_states,
dynamic_axes=common_types.HiddenStateSharding,
partition_manager=self.config.partition_manager,
)
return EncoderLayerOutput(
hidden_states=hidden_states,
attention_weight=attn_outputs.attention_weight,
)
[docs]class Llama4VisionEncoder(nn.Module):
"""Vision encoder stack for Llama4 models.
Stacks multiple vision encoder layers to progressively encode
visual features for downstream processing.
"""
def __init__(
self,
config: Llama4VisionConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
self.config = config
self.dtype = dtype
self.param_dtype = param_dtype
self.precision = precision
self.layers = [
Llama4VisionEncoderLayer(
config=config,
layer_idx=layer_idx,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
for layer_idx in range(self.config.num_hidden_layers)
]
def __call__(
self,
hidden_states: jax.Array,
frequencies: jax.Array,
attention_mask: jax.Array | None = None,
output_attentions: bool | None = None,
output_hidden_states: bool | None = None,
) -> BaseModelOutput:
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
)
encoder_states = () if output_hidden_states else None
all_attentions = () if output_attentions else None
for encoder_layer in self.layers:
if output_hidden_states:
encoder_states = (*encoder_states, hidden_states)
layer_outputs = encoder_layer(
hidden_states=hidden_states,
output_attentions=output_attentions,
frequencies=frequencies,
)
if output_attentions:
all_attentions = (*all_attentions, layer_outputs.attention_weight)
hidden_states = layer_outputs.hidden_states
if output_hidden_states:
encoder_states = (*encoder_states, hidden_states)
return BaseModelOutput(
last_hidden_state=hidden_states,
hidden_states=encoder_states,
attentions=all_attentions,
)
[docs]class Llama4UnfoldConvolution(nn.Module):
"""Unfold convolution module for Llama4 vision models.
Implements patch extraction with optional convolution,
converting images into sequences of patch embeddings.
"""
def __init__(
self,
config: Llama4VisionConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
patch_size_val = config.patch_size
if isinstance(patch_size_val, int):
self.kernel_size: tuple[int, int] = (patch_size_val, patch_size_val)
else:
self.kernel_size: tuple[int, int] = patch_size_val
self.stride = config.patch_size
if isinstance(self.stride, int):
self.stride = (self.stride, self.stride)
self.num_channels: int = config.num_channels
self.hidden_size: int = config.hidden_size
# Linear layer similar to PyTorch's version
in_features = self.num_channels * self.kernel_size[0] * self.kernel_size[1]
self.linear = ColumnParallelLinear(
in_features=in_features,
out_features=self.hidden_size,
use_bias=False,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
def __call__(self, hidden_states: Float[Array, "batch seq_len hidden_dim"]) -> jax.Array:
batch_size = hidden_states.shape[0]
hidden_states_nhwc = jnp.transpose(hidden_states, (0, 2, 3, 1))
patches = jax.lax.conv_general_dilated_patches(
lhs=hidden_states_nhwc,
filter_shape=self.kernel_size,
window_strides=self.stride,
padding="VALID",
dimension_numbers=("NHWC", "HWIO", "NHWC"),
)
num_patches = patches.shape[1] * patches.shape[2]
patches_reshaped = jnp.reshape(patches, (batch_size, num_patches, -1))
hidden_states = self.linear(patches_reshaped)
return hidden_states
[docs]@register_module(TaskType.BASE_VISION, config=Llama4VisionConfig, model_type="llama4_vision")
@register_module(TaskType.BASE_MODULE, config=Llama4VisionConfig, model_type="llama4_vision")
class Llama4VisionModel(EasyDeLBaseModule):
"""Vision transformer for Llama4 including patchify stem, transformer blocks, and final norm."""
def __init__(
self,
config: Llama4VisionConfig,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.image_size = config.image_size
self.patch_size = config.patch_size
self.hidden_size = config.hidden_size
self.num_channels = config.num_channels
self.num_patches = (self.image_size // self.patch_size) ** 2 + 1
self.scale = config.hidden_size**-0.5
self.patch_embedding = Llama4UnfoldConvolution(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.class_embedding = ArrayParam.bound(
shape=(self.hidden_size,),
dtype=param_dtype,
init_method="normal",
init_kwargs={"stddev": self.scale},
key=rngs.params(),
)
self.positional_embedding_vlm = ArrayParam.bound(
shape=(self.num_patches, self.hidden_size),
dtype=param_dtype,
init_method="normal",
init_kwargs={"stddev": self.scale},
key=rngs.params(),
)
self.layernorm_pre = nn.LayerNorm(
num_features=self.hidden_size,
epsilon=0.00001,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
self.layernorm_post = nn.LayerNorm(
num_features=self.hidden_size,
epsilon=0.00001,
dtype=dtype,
param_dtype=param_dtype,
rngs=rngs,
)
# encoders
self.model = Llama4VisionEncoder(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.vision_adapter = Llama4VisionPixelShuffleMLP(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.vision_idx = self.config.image_size // self.config.patch_size
def __call__(
self,
pixel_values: jax.Array,
attention_mask: jax.Array | None = None,
output_attentions: bool | None = None,
output_hidden_states: bool | None = None,
) -> BaseModelOutput:
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
)
batch_size_times_num_tiles = pixel_values.shape[0]
num_concurrent_media = 1
num_chunks = 1
hidden_states = self.patch_embedding(pixel_values)
_, num_patches, hidden_dim = hidden_states.shape
# Add cls token
hidden_states = hidden_states.reshape(
batch_size_times_num_tiles * num_concurrent_media * num_chunks,
num_patches,
hidden_dim,
)
class_embedding = jnp.broadcast_to(
self.class_embedding.value,
(hidden_states.shape[0], 1, hidden_states.shape[-1]),
)
hidden_states = jnp.concatenate([hidden_states, class_embedding], axis=1)
num_patches += 1
# Position embeddings
hidden_states = hidden_states.reshape(
batch_size_times_num_tiles * num_concurrent_media,
num_chunks,
num_patches,
hidden_dim,
)
hidden_states = hidden_states + self.positional_embedding_vlm
hidden_states = self.layernorm_pre(hidden_states)
hidden_states = hidden_states.reshape(batch_size_times_num_tiles, -1, hidden_dim)
output = self.model(
hidden_states,
output_hidden_states=output_hidden_states,
output_attentions=output_attentions,
frequencies=_vision_freqs(
self.vision_idx,
self.config.hidden_size,
self.config.num_attention_heads,
self.config.rope_theta,
),
)
hidden_states = output.last_hidden_state
hidden_states = self.layernorm_post(hidden_states)
hidden_states = hidden_states[:, :-1, :]
hidden_states = self.vision_adapter(hidden_states)
all_hidden_states = output.hidden_states if output_hidden_states else None
return BaseModelOutput(
last_hidden_state=hidden_states,
hidden_states=all_hidden_states,
attentions=output.attentions,
)
[docs] def get_encoder(self):
"""
Returns the encoder part of the model's graph definition.
This vision model acts as the encoder.
"""
return self
[docs] def get_decoder(self):
"""
Returns the decoder part of the model's graph definition.
This is an encoder-only model and does not have a decoder.
"""
raise NotImplementedError("This is an encoder-only model and does not have a decoder.")
[docs] def get_lm_head(self):
"""
Returns the language model head of the module.
This vision model does not have a language model head.
"""
raise NotImplementedError("This vision model does not have a language model head.")
[docs] def get_embedding(self):
"""
Returns the embedding layer of the module.
"""
return self.patch_embedding
[docs]@register_module(TaskType.IMAGE_TEXT_TO_TEXT, config=Llama4Config, model_type="llama4")
class Llama4ForConditionalGeneration(EasyDeLBaseModule):
"""Llama4 Vision model for conditional text generation based on image inputs.
Combines a vision tower and a language model with a multi-modal projector.
Note: Llama4 has a unique architecture where the language_model is already
a complete Llama4ForCausalLM (with its own lm_head), unlike other VLMs where
the base model doesn't include the lm_head.
Attributes:
config (Llama4Config): Configuration object.
dtype (jnp.dtype): Data type for computation.
param_dtype (jnp.dtype): Data type for parameters.
precision (jax.lax.PrecisionLike): JAX precision level.
rngs (nn.Rngs): Random number generators.
Class Attributes:
_task_type: IMAGE_TEXT_TO_TEXT task type
_model_type: "llama4" model identifier
_supports_video: True (Llama4 supports video input)
_uses_mrope: False (uses standard RoPE)
"""
# Class attributes for VLM capabilities
_task_type = TaskType.IMAGE_TEXT_TO_TEXT
_model_type = "llama4"
_config_class = Llama4Config
_auto_register = False # Already registered via decorator
_supports_video = True
_uses_mrope = False
# Component name mapping
_vision_tower_name = "vision_model"
_projector_name = "multi_modal_projector"
_language_model_name = "language_model"
loss_type = "ForCausalLM"
def __init__(
self,
config: Llama4Config,
dtype: jnp.dtype = jnp.bfloat16,
param_dtype: jnp.dtype = jnp.bfloat16,
precision: jax.lax.PrecisionLike = None,
*,
rngs: nn.Rngs,
):
"""Initializes the Llama4ForConditionalGeneration model."""
super().__init__(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.vision_model = Llama4VisionModel(
config=config.vision_config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.multi_modal_projector = Llama4MultiModalProjector(
config=config,
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.language_model = Llama4ForCausalLM(
config=config.get_text_config(),
dtype=dtype,
param_dtype=param_dtype,
precision=precision,
rngs=rngs,
)
self.vocab_size = config.get_text_config().vocab_size
self.pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else -1
[docs] def get_image_features(self, pixel_values: chex.Array, **kwargs) -> chex.Array:
"""Extracts and projects image features from the vision tower.
Args:
pixel_values (chex.Array): Input pixel values for the images.
Returns:
chex.Array: Processed image features ready for the language model.
"""
kwargs = {k: v for k, v in kwargs.items() if v is not None}
image_outputs = self.vision_model(
pixel_values,
output_hidden_states=False,
**kwargs,
)
hidden_states = image_outputs.last_hidden_state
return hidden_states
def __call__(
self,
input_ids: Int[Array, "batch seq_len"] = None,
pixel_values: chex.Array = 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,
inputs_embeds: Float[Array, "batch seq_len hidden_dim"] | None = None,
output_attentions: bool | None = None,
output_hidden_states: bool | None = None,
**lm_kwargs,
):
if (input_ids is None) ^ (inputs_embeds is not None):
raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
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
)
if input_ids is not None and self.config.image_token_id >= self.vocab_size:
special_image_mask = input_ids == self.config.image_token_id
llm_input_ids = input_ids
llm_input_ids = jnp.where(special_image_mask, 0, llm_input_ids)
else:
llm_input_ids = input_ids
if inputs_embeds is None:
inputs_embeds = self.language_model.model.embed_tokens(llm_input_ids)
if pixel_values is not None:
orgshape = inputs_embeds.shape
image_features = self.get_image_features(pixel_values)
vision_flat = image_features.reshape(-1, image_features.shape[-1])
projected_vision_flat = self.multi_modal_projector(vision_flat)
final_mask = jnp.expand_dims(input_ids == self.config.image_token_id, axis=-1)
inputs_embeds_flat = inputs_embeds.reshape(-1, inputs_embeds.shape[-1])
final_mask_1d = final_mask[..., 0].reshape(-1)
num_projected_tokens = projected_vision_flat.shape[0]
image_token_indices = jnp.where(
final_mask_1d,
size=num_projected_tokens,
fill_value=-1,
)[0]
inputs_embeds_updated_flat = inputs_embeds_flat.at[image_token_indices].set(projected_vision_flat)
inputs_embeds = inputs_embeds_updated_flat.reshape(orgshape)
outputs = self.language_model(
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,
**lm_kwargs,
)
return VLMCausalLMOutput(
logits=outputs.logits,
past_key_values=outputs.past_key_values,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
image_hidden_states=image_features if pixel_values is not None else None,
)
[docs] def init_cache(
self,
batch_size,
max_length,
starts=None,
shardings=None,
pad_token_id=None,
):
return self.language_model.init_cache(batch_size, max_length, starts, shardings, pad_token_id)
def _get_compile_model_kwargs(
self,
batch_size: int,
input_tokens_length: int,
input_sharding: jax.sharding.PartitionSpec,
rngs: jax.random.PRNGKey,
vision_included: bool = False,
vision_batch_size: int = 1,
vision_channels: int = 3,
vision_height: int | None = None,
vision_width: int | None = None,
required_props: tp.Mapping[str, dict[str, tp.Any]] | None = None,
**kwargs,
):
"""Helper function to get keyword arguments for model compilation, potentially including vision inputs.
Args:
batch_size (int): Batch size for text inputs.
input_tokens_length (int): Sequence length for text inputs.
input_sharding (jax.sharding.PartitionSpec): Sharding specification for text inputs.
rngs (jax.random.PRNGKey): Random number generator key.
vision_included (bool): Whether to include dummy vision inputs. Defaults to False.
vision_batch_size (int): Batch size for vision inputs. Defaults to 1.
vision_channels (int): Number of channels for vision inputs. Defaults to 3.
vision_height (Optional[int]): Height for vision inputs (defaults to config).
vision_width (Optional[int]): Width for vision inputs (defaults to config).
required_props (Optional[Mapping[str, Dict[str, Any]]]): Required properties.
**kwargs: Additional arguments passed to the language model's compile kwargs method.
Returns:
dict: Keyword arguments for model compilation.
"""
basics = self.language_model._get_compile_model_kwargs(
batch_size=batch_size,
input_tokens_length=input_tokens_length,
input_sharding=input_sharding,
rngs=rngs,
vision_included=vision_included,
vision_batch_size=vision_batch_size,
vision_channels=vision_channels,
vision_height=vision_height,
vision_width=vision_width,
required_props=required_props,
**kwargs,
)
if vision_included:
pixel_values = jnp.ones(
(
vision_batch_size or 1,
vision_channels or 3,
self.config.vision_config.image_size,
self.config.vision_config.image_size,
),
dtype="f4",
)
basics.update({"pixel_values": pixel_values})
return basics
[docs] def get_encoder(self):
"""Returns the encoder part of the model (vision tower)."""
return self.vision_model
[docs] def get_decoder(self):
"""Returns the decoder part of the model."""
return self.language_model.get_decoder()
[docs] def get_lm_head(self):
"""Returns the language model head."""
return self.language_model.get_lm_head()
[docs] def get_embedding(self):
"""Returns the embedding layer."""
return self.language_model.get_embedding()
[docs] def get_vision_tower(self) -> nn.Module:
"""Returns the vision tower component."""
return self.vision_model
[docs] def get_projector(self) -> nn.Module:
"""Returns the multimodal projector component."""
return self.multi_modal_projector
[docs] def get_language_model(self) -> nn.Module:
"""Returns the language model component."""
return self.language_model