d2Cache

d2Cache is a framework for efficient inference and caching in Diffusion Language Models (MDLMs). This repository contains several advanced model implementations and optimized inference strategies.

🚀 New Models and Usage

We have introduced several new model variants and optimized parallel versions. To use these models, you must configure the required environment variables and use the specific model parameter in eval.py.

1. Model Summary & Parameters

Model Category	model Argument	Description	Required Environment Variables
DParallel	dparallel_llada-inst	Parallel Optimized LLaDA	LLADA_INST_PATH
	dparallel_dream-inst	Parallel Optimized Dream	DREAM_BASE_PATH
D2F (LoRA)	d2f_llada-inst	LLaDA with D2F LoRA	LLADA_INST_PATH, D2F_LLADA_INST_PATH
	d2f_dream-inst	Dream with D2F LoRA	DREAM_BASE_PATH, D2F_DREAM_INST_PATH
Fast dLLM v2	fast_dllm_v2_1.5b-inst	Optimized 1.5B Variant	FAST_DLLM_V2_1_5_PATH, DREAM_BASE_PATH (Tokenizer)
	fast_dllm_v2_7b-inst	Optimized 7B Variant	FAST_DLLM_V2_7_PATH, DREAM_BASE_PATH (Tokenizer)

---

2. Environment Variables Configuration

Set the following variables based on the model you intend to use:

# === Base Model Paths ===
export LLADA_INST_PATH=/path/to/LLaDA-8B-Instruct
export DREAM_BASE_PATH=/path/to/Dream-Base-Model

# === D2F LoRA Adapter Paths (Required for D2F models) ===
export D2F_LLADA_INST_PATH=/path/to/d2f-llada-lora
export D2F_DREAM_INST_PATH=/path/to/d2f-dream-lora

# === Fast dLLM v2 Model Paths (Required for Fast dLLM models) ===
export FAST_DLLM_V2_1_5_PATH=/path/to/fast-dllm-v2-1.5b
export FAST_DLLM_V2_7_PATH=/path/to/fast-dllm-v2-7b

---

3. Quick Start Commands

A. Running DParallel LLaDA

export LLADA_INST_PATH=/path/to/LLaDA-8B-Instruct

accelerate launch eval.py \
    model=dparallel_llada-inst \
    dataset.name=gsm8k \
    batch_size=1

B. Running D2F LLaDA (with Base Model and LoRA)

export LLADA_INST_PATH=/path/to/LLaDA-8B-Instruct
export D2F_LLADA_INST_PATH=/path/to/d2f-llada-lora

accelerate launch eval.py \
    model=d2f_llada-inst \
    dataset.name=gsm8k \
    batch_size=1

C. Running Fast dLLM v2 7B (with Tokenizer Path)

export FAST_DLLM_V2_7_PATH=/path/to/fast-dllm-v2-7b
export DREAM_BASE_PATH=/path/to/Dream-Base-Model  # Required for tokenizer

accelerate launch eval.py \
    model=fast_dllm_v2_7b-inst \
    dataset.name=gsm8k \
    batch_size=1

---

🛠 Advanced Configuration

All configurations are managed via Hydra. You can override any parameters from the command line:

• Cache Strategy: cache=prefix (options: prefix, d2cache, dllm)
• Generation Strategy: generation=vanilla (options: vanilla, eb_sampler, pc_sampler, klass, wino, daedal)
• Evaluation Limit: eval_args.limit=10

For more detailed information on adding new models, please refer to AddModel.md.

📄 License

This project is licensed under the Apache 2.0 License - see the LICENSE file for details.