Implementation:Predibase Lorax Flash DBRX Modeling

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

Overview

Optimized DBRX transformer implementation for LoRax inference serving with flash attention, LoRA adapter support, and Mixture of Experts (MoE) feed-forward layers using the Megablocks library for block-sparse operations.

Description

This module implements the DBRX Mixture of Experts architecture adapted for high-throughput inference in the LoRax serving framework. The main components are:

• DbrxConfig / DbrxAttentionConfig / DbrxFFNConfig -- Configuration classes that define model hyperparameters including attention settings (clip_qkv, kv_n_heads, rope_theta) and FFN settings (moe_num_experts, moe_top_k, ffn_hidden_size).
• DbrxAttention -- Multi-head attention with grouped-query attention (GQA), rotary positional embeddings (RoPE), QKV clipping, and paged attention for efficient KV cache management. Supports LoRA adapters via TensorParallelMultiAdapterLinear.
• DbrxNormAttentionNorm -- Wraps attention with pre- and post-attention layer normalization, providing residual connections.
• BlockSparseMoE -- The primary MoE feed-forward layer that uses Megablocks for block-sparse matrix operations, enabling efficient routing of tokens to experts without dropping tokens. Falls back to dense computation for small batch sizes (256 tokens or fewer).
• DenseMoE -- A fallback dense MoE implementation used when quantization is enabled, iterating over experts individually.
• DbrxLayer -- A single transformer layer combining DbrxNormAttentionNorm and the MoE feed-forward module.
• DbrxModel -- The full transformer model stacking embedding, multiple DbrxLayer instances, and final layer normalization.
• FlashDbrxForCausalLM -- The top-level causal language model wrapping DbrxModel with an LM head supporting multi-adapter inference.

Usage

Used internally by the LoRax server when serving DBRX-based models. Loaded via the model registry when the model config type matches. Requires the megablocks and stk libraries for block-sparse MoE computation.

Code Reference

Source Location

• Repository: Predibase_Lorax
• File: server/lorax_server/models/custom_modeling/flash_dbrx_modeling.py
• Lines: 1-1032

Signature

class FlashDbrxForCausalLM(torch.nn.Module):
    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,
        adapter_data: AdapterBatchData,
        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_dbrx_modeling import FlashDbrxForCausalLM

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
adapter_data	AdapterBatchData	Yes	LoRA adapter configuration for 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]	Speculative decoding logits, or None

Usage Examples

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

Related Pages

• Environment:Predibase_Lorax_CUDA_GPU_Runtime