Implementation:Predibase Lorax Flash Granite Modeling

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

Overview

Optimized IBM Granite transformer implementation for LoRax inference serving with flash attention, LoRA adapter support, and Granite-specific scaling multipliers. Extends the Flash Llama implementation.

Description

This module implements the IBM Granite architecture adapted for high-throughput inference in the LoRax serving framework. Granite extends the Llama architecture with additional scaling multipliers for embeddings, attention, residuals, and logits. The main components are:

• GraniteConfig -- Configuration class extending LlamaConfig with Granite-specific parameters: embedding_multiplier (scales token embeddings), logits_scaling (scales output logits), residual_multiplier (scales residual connections), and attention_multiplier (replaces the standard 1/sqrt(d) attention scaling). Also adds attention_bias and mlp_bias flags.
• FlashGraniteAttention -- Extends FlashLlamaAttention, overriding the softmax_scale with the configurable attention_multiplier and using a custom load_attention function that supports optional attention bias.
• GraniteMLP -- Extends LlamaMLP with optional bias support on gate_proj, up_proj, and down_proj projections. Uses fused gate-up projection with LoRA adapter support on GATE_PROJ, UP_PROJ, and DOWN_PROJ.
• FlashGraniteLayer -- Extends FlashLlamaLayer, applying the residual_multiplier to both attention and MLP outputs before adding them to the residual stream.
• FlashGraniteForCausalLM -- Extends FlashLlamaForCausalLM, applying the embedding_multiplier to the embedding weights at initialization and scaling output logits by logits_scaling during forward.

Usage

Used internally by the LoRax server when serving IBM Granite-based models. 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_granite_modeling.py
• Lines: 1-303

Signature

class FlashGraniteForCausalLM(FlashLlamaForCausalLM):
    def __init__(self, prefix: str, config, weights, create_layer_fn=None):
        ...

    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,
        cross_attention_states: Optional[torch.Tensor] = None,
        skip_lm_head: bool = False,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
        ...

Import

from lorax_server.models.custom_modeling.flash_granite_modeling import FlashGraniteForCausalLM

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
cross_attention_states	Optional[torch.Tensor]	No	Cross-attention states (passed through to parent class)
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, scaled by logits_scaling (or hidden states if skip_lm_head is True)
speculative_logits	Optional[torch.Tensor]	Speculative decoding logits scaled by logits_scaling, or None

Usage Examples

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

Related Pages

• Environment:Predibase_Lorax_CUDA_GPU_Runtime