Implementation:Predibase Lorax Flash NeoX Modeling

Knowledge Sources	Predibase_Lorax
Domains	Model_Architecture, Inference
Last Updated	2026-02-08 00:00 GMT

Overview

Optimized GPT-NeoX transformer implementation for LoRax inference serving with flash attention and rotary positional embeddings.

Description

This module implements the GPT-NeoX architecture adapted for high-throughput inference in the LoRax serving framework. GPT-NeoX features rotary positional embeddings and supports both parallel and sequential residual connections. The main components are:

• FlashNeoxAttention -- Multi-head self-attention with rotary positional embeddings (RoPE) applied to queries and keys. Uses a fused query_key_value projection that is permuted to separate Q, K, V heads. Supports flash attention for prefill and paged attention for decode.
• FlashMLP -- Feed-forward network with dense_h_to_4h and dense_4h_to_h projections, supporting configurable activation functions including various GELU approximations.
• FlashNeoXLayer -- A single transformer layer that supports two residual connection modes: parallel residual (attention and MLP computed in parallel from the same normalized input, then summed) and sequential residual (standard pre-norm architecture). Includes input_layernorm and post_attention_layernorm.
• FlashGPTNeoXModel -- The full transformer model with token embedding (embed_in), stacked FlashNeoXLayer instances, and final layer normalization. Computes rotary cos/sin once and passes them to all layers.
• FlashGPTNeoXForCausalLM -- The top-level causal language model wrapping FlashGPTNeoXModel with an embed_out head via TensorParallelHead.

Note: This implementation does not include LoRA adapter support, unlike some other flash model implementations in the LoRax codebase.

Usage

Used internally by the LoRax server when serving GPT-NeoX-based models (such as Pythia). Loaded via the model registry when the model config type matches.

Code Reference

Source Location

• Repository: Predibase_Lorax
• File: server/lorax_server/models/custom_modeling/flash_neox_modeling.py
• Lines: 1-379

Signature

class FlashGPTNeoXForCausalLM(FlashGPTNeoXPreTrainedModel):
    def __init__(self, prefix: str, config, weights):
        ...

    def forward(
        self,
        input_ids: torch.Tensor,
        position_ids: torch.Tensor,
        cu_seqlen_prefill: Optional[torch.Tensor],
        kv_cache: List[Tuple[torch.Tensor, torch.Tensor]],
        block_tables: torch.Tensor,
        slots: torch.Tensor,
        seqlen: Seqlen,
        max_s: int,
        prefill_cache_indices: Optional[torch.Tensor] = None,
        lm_head_indices: Optional[torch.Tensor] = None,
        skip_lm_head: bool = False,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
        ...

Import

from lorax_server.models.custom_modeling.flash_neox_modeling import FlashGPTNeoXForCausalLM

I/O Contract

Inputs

Name	Type	Required	Description
input_ids	torch.Tensor	Yes	Token IDs for the input sequence
position_ids	torch.Tensor	Yes	Position indices for rotary embeddings
cu_seqlen_prefill	Optional[torch.Tensor]	Yes	Cumulative sequence lengths for flash attention prefill; None during decode
kv_cache	List[Tuple[torch.Tensor, torch.Tensor]]	Yes	Key-value cache tensors for each layer
block_tables	torch.Tensor	Yes	Block tables for paged attention
slots	torch.Tensor	Yes	Slot indices for KV cache storage
seqlen	Seqlen	Yes	Sequence length information for the batch
max_s	int	Yes	Maximum sequence length in the batch
prefill_cache_indices	Optional[torch.Tensor]	No	Indices for selective cache prefilling
lm_head_indices	Optional[torch.Tensor]	No	Indices to select specific positions for LM head
skip_lm_head	bool	No	If True, return hidden states without applying the LM head

Outputs

Name	Type	Description
logits	torch.Tensor	Next-token logits over the vocabulary (or hidden states if skip_lm_head is True)
speculative_logits	Optional[torch.Tensor]	Always None for GPT-NeoX

Usage Examples

# Internal usage within LoRax server
from lorax_server.models.custom_modeling.flash_neox_modeling import FlashGPTNeoXForCausalLM
# Model instantiated by model registry, not directly by users

Related Pages

• Environment:Predibase_Lorax_CUDA_GPU_Runtime