tpu-optimized-llm / model /transformer.py
Threatthriver's picture
Upload folder using huggingface_hub
f24563f verified
"""
Transformer blocks for the LLM model.
"""
import jax
import jax.numpy as jnp
import flax.linen as nn
from typing import Optional, Tuple, Dict, Any, Callable, Union
import math
from model.attention import MultiHeadAttention, MultiQueryAttention, RotaryMultiQueryAttention
class FeedForward(nn.Module):
"""
Feed-forward network with SwiGLU activation.
Attributes:
dim: Input and output dimension
hidden_dim: Hidden dimension
dropout_rate: Dropout probability
dtype: Data type for computations
"""
dim: int
hidden_dim: int
dropout_rate: float = 0.0
dtype: jnp.dtype = jnp.float32
def setup(self):
self.gate_proj = nn.Dense(
features=self.hidden_dim,
dtype=self.dtype,
kernel_init=nn.initializers.normal(stddev=0.02),
name="gate_proj"
)
self.up_proj = nn.Dense(
features=self.hidden_dim,
dtype=self.dtype,
kernel_init=nn.initializers.normal(stddev=0.02),
name="up_proj"
)
self.down_proj = nn.Dense(
features=self.dim,
dtype=self.dtype,
kernel_init=nn.initializers.normal(stddev=0.02),
name="down_proj"
)
self.dropout = nn.Dropout(rate=self.dropout_rate)
def __call__(self, x: jnp.ndarray, deterministic: bool = True) -> jnp.ndarray:
"""
Apply feed-forward network.
Args:
x: Input tensor [batch_size, seq_len, dim]
deterministic: Whether to use deterministic operations (no dropout)
Returns:
Output tensor [batch_size, seq_len, dim]
"""
# SwiGLU activation
gate = self.gate_proj(x)
gate = jax.nn.silu(gate)
up = self.up_proj(x)
# Element-wise multiplication
hidden = gate * up
# Project back to input dimension
output = self.down_proj(hidden)
# Apply dropout
output = self.dropout(output, deterministic=deterministic)
return output
class TransformerBlock(nn.Module):
"""
Transformer block with attention and feed-forward network.
Attributes:
dim: Hidden dimension
num_heads: Number of attention heads
hidden_dim: Hidden dimension in feed-forward network
dropout_rate: Dropout probability
attention_dropout_rate: Dropout probability for attention
layer_norm_epsilon: Epsilon for layer normalization
dtype: Data type for computations
"""
dim: int
num_heads: int
hidden_dim: int
dropout_rate: float = 0.0
attention_dropout_rate: float = 0.0
layer_norm_epsilon: float = 1e-5
dtype: jnp.dtype = jnp.float32
def setup(self):
# Layer normalization
self.input_layernorm = nn.LayerNorm(
epsilon=self.layer_norm_epsilon,
dtype=self.dtype,
name="input_layernorm"
)
self.post_attention_layernorm = nn.LayerNorm(
epsilon=self.layer_norm_epsilon,
dtype=self.dtype,
name="post_attention_layernorm"
)
# Attention
self.attention = MultiHeadAttention(
dim=self.dim,
num_heads=self.num_heads,
dropout_rate=self.attention_dropout_rate,
dtype=self.dtype,
name="attention"
)
# Feed-forward network
self.feed_forward = FeedForward(
dim=self.dim,
hidden_dim=self.hidden_dim,
dropout_rate=self.dropout_rate,
dtype=self.dtype,
name="feed_forward"
)
# Dropout
self.dropout = nn.Dropout(rate=self.dropout_rate)
def __call__(
self,
hidden_states: jnp.ndarray,
attention_mask: Optional[jnp.ndarray] = None,
position_ids: Optional[jnp.ndarray] = None,
past_key_value: Optional[Tuple[jnp.ndarray, jnp.ndarray]] = None,
output_attentions: bool = False,
use_cache: bool = False,
deterministic: bool = True,
) -> Tuple[jnp.ndarray, ...]:
"""
Apply transformer block.
Args:
hidden_states: Input tensor [batch_size, seq_len, dim]
attention_mask: Attention mask [batch_size, 1, seq_len, seq_len]
position_ids: Position indices [batch_size, seq_len]
past_key_value: Cached key and value tensors for incremental decoding
output_attentions: Whether to return attention weights
use_cache: Whether to use cached key and values
deterministic: Whether to use deterministic operations (no dropout)
Returns:
Tuple of (output, attention_weights, present_key_value)
"""
# Self-attention
residual = hidden_states
hidden_states = self.input_layernorm(hidden_states)
attention_outputs = self.attention(
hidden_states=hidden_states,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_value=past_key_value,
output_attentions=output_attentions,
use_cache=use_cache,
deterministic=deterministic,
)
hidden_states = attention_outputs[0]
hidden_states = self.dropout(hidden_states, deterministic=deterministic)
hidden_states = residual + hidden_states
# Feed-forward network
residual = hidden_states
hidden_states = self.post_attention_layernorm(hidden_states)
hidden_states = self.feed_forward(hidden_states, deterministic=deterministic)
hidden_states = residual + hidden_states
outputs = (hidden_states,) + attention_outputs[1:]
return outputs
class TransformerLayer(nn.Module):
"""
Transformer layer with multi-query attention and feed-forward network.
Attributes:
dim: Hidden dimension
num_query_heads: Number of query heads
num_kv_heads: Number of key-value heads
hidden_dim: Hidden dimension in feed-forward network
max_seq_len: Maximum sequence length for RoPE
dropout_rate: Dropout probability
attention_dropout_rate: Dropout probability for attention
layer_norm_epsilon: Epsilon for layer normalization
use_rope: Whether to use rotary position embeddings
dtype: Data type for computations
"""
dim: int
num_query_heads: int
num_kv_heads: int = 1
hidden_dim: int = None
max_seq_len: int = 4096
dropout_rate: float = 0.0
attention_dropout_rate: float = 0.0
layer_norm_epsilon: float = 1e-5
use_rope: bool = True
dtype: jnp.dtype = jnp.float32
def setup(self):
# Set hidden dimension if not provided
if self.hidden_dim is None:
self.actual_hidden_dim = 4 * self.dim
else:
self.actual_hidden_dim = self.hidden_dim
# Layer normalization
self.input_layernorm = nn.LayerNorm(
epsilon=self.layer_norm_epsilon,
dtype=self.dtype,
name="input_layernorm"
)
self.post_attention_layernorm = nn.LayerNorm(
epsilon=self.layer_norm_epsilon,
dtype=self.dtype,
name="post_attention_layernorm"
)
# Attention
if self.use_rope:
self.attention = RotaryMultiQueryAttention(
dim=self.dim,
num_query_heads=self.num_query_heads,
num_kv_heads=self.num_kv_heads,
max_seq_len=self.max_seq_len,
dropout_rate=self.attention_dropout_rate,
dtype=self.dtype,
name="attention"
)
else:
self.attention = MultiQueryAttention(
dim=self.dim,
num_query_heads=self.num_query_heads,
num_kv_heads=self.num_kv_heads,
dropout_rate=self.attention_dropout_rate,
dtype=self.dtype,
name="attention"
)
# Feed-forward network
self.feed_forward = FeedForward(
dim=self.dim,
hidden_dim=self.actual_hidden_dim,
dropout_rate=self.dropout_rate,
dtype=self.dtype,
name="feed_forward"
)
# Dropout
self.dropout = nn.Dropout(rate=self.dropout_rate)
def __call__(
self,
hidden_states: jnp.ndarray,
attention_mask: Optional[jnp.ndarray] = None,
position_ids: Optional[jnp.ndarray] = None,
past_key_value: Optional[Tuple[jnp.ndarray, jnp.ndarray]] = None,
output_attentions: bool = False,
use_cache: bool = False,
deterministic: bool = True,
) -> Tuple[jnp.ndarray, ...]:
"""
Apply transformer layer.
Args:
hidden_states: Input tensor [batch_size, seq_len, dim]
attention_mask: Attention mask [batch_size, 1, seq_len, seq_len]
position_ids: Position indices [batch_size, seq_len]
past_key_value: Cached key and value tensors for incremental decoding
output_attentions: Whether to return attention weights
use_cache: Whether to use cached key and values
deterministic: Whether to use deterministic operations (no dropout)
Returns:
Tuple of (output, attention_weights, present_key_value)
"""
# Self-attention
residual = hidden_states
hidden_states = self.input_layernorm(hidden_states)
attention_outputs = self.attention(
hidden_states=hidden_states,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_value=past_key_value,
output_attentions=output_attentions,
use_cache=use_cache,
deterministic=deterministic,
)
hidden_states = attention_outputs[0]
hidden_states = self.dropout(hidden_states, deterministic=deterministic)
hidden_states = residual + hidden_states
# Feed-forward network
residual = hidden_states
hidden_states = self.post_attention_layernorm(hidden_states)
hidden_states = self.feed_forward(hidden_states, deterministic=deterministic)
hidden_states = residual + hidden_states
outputs = (hidden_states,) + attention_outputs[1:]
return outputs