Implementation:Predibase Lorax Flash Phi Modeling

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

Overview

Optimized Phi (1.5/2) transformer implementation for LoRax inference serving with flash attention, partial rotary embeddings, and LoRA adapter support.

Description

This module implements the Microsoft Phi architecture (Phi-1.5 and Phi-2) adapted for high-throughput inference in the LoRax serving framework. Phi models use a parallel residual connection pattern where attention and MLP outputs are computed from the same normalized input and summed. The main components are:

• FlashPhiAttention -- Multi-head attention with partial rotary embeddings (controlled by config.partial_rotary_factor), grouped-query attention support via n_head_kv, and separate Q/K/V projections loaded via TensorParallelMultiAdapterLinear. Supports LoRA adapters on q_proj, k_proj, v_proj, and dense (output projection).
• PhiMLP -- Feed-forward network with fc1 and fc2 projections, configurable activation function (supporting various GELU approximations), and LoRA adapter support on both layers.
• FlashPhiLayer -- A single transformer layer using a parallel residual pattern: both attention and MLP are applied to the same layer-normalized input, then their outputs are summed. Uses FastLayerNorm with an all-reduce for tensor parallelism.
• FlashPhiModel -- The full transformer model with token embedding (embed_tokens), stacked FlashPhiLayer instances, and final layer normalization (final_layernorm). Computes rotary cos/sin once and passes them to all layers.
• FlashPhiForCausalLM -- The top-level causal language model wrapping FlashPhiModel with a MultiAdapterHead LM head supporting LoRA on the output projection.

Usage

Used internally by the LoRax server when serving Phi-based models (Phi-1.5, Phi-2). 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_phi_modeling.py
• Lines: 1-411

Signature

class FlashPhiForCausalLM(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_phi_modeling import FlashPhiForCausalLM

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 partial 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_phi_modeling import FlashPhiForCausalLM
# Model instantiated by model registry, not directly by users

Related Pages

• Environment:Predibase_Lorax_CUDA_GPU_Runtime