A configurable memory-efficient optimizer for PyTorch. It combines fused CUDA kernels and block-wise adaptive log-space quantization of the second moment with a family of update paths—including APOLLO low-rank projection and CAME confidence guidance—reducing optimizer memory while preserving numerical stability.

⚡ Key Features

• Blockwise Adaptive Log-Space Quantization: Quantizes the second moment (variance) in log2 space with a per-block adaptive floor and a reserved zero code, accommodating the non-negative, long-tailed nature of variance estimates.
• Fused CUDA Kernels: Combines dequantization, EMA updates, Warp-Shuffle reductions, and requantization into single kernels. It utilizes float4 vectorization to optimize memory bandwidth usage.
• Configurable First-Moment: Stores the optional first moment (beta1) using configurable quantization formats (Uniform/Dynamic Map, 4-bit/8-bit) or full FP32 precision, preserving momentum updates while keeping memory overhead minimal.
• CAME Confidence Guidance: Optional Confidence-guided Adaptive Memory Efficient Optimization (CAME) that estimates update confidence from historical momentum and adaptively suppresses unstable update directions.
• APOLLO Subspace Projection: Opt-in random subspace projection that estimates adaptive gradient scaling in a low-rank space, capturing cross-dimensional covariance information while keeping memory overhead low.
• Fira Norm-Growth Limiter: Regulates the relative growth of update norms to suppress destructive gradient spikes, available as an optional stabilizer across update paths.
• Zero CPU-GPU Sync: Eliminates implicit synchronizations (e.g., D2H copies) in the control flow, ensuring the GPU computation pipeline runs without blocking.
• Cross-Platform JIT: Uses Just-In-Time (JIT) compilation for straightforward setup across both Windows and Linux environments.

📊 Performance

• Memory Footprint: Due to Adafactor's factorized second-moment estimation, 8-bit quantization, and configurable first moments, the optimizer typically consumes less memory than AdamW8Bit.
• Training Speed: The fused kernel design and reduced synchronization overhead allow it to achieve step times comparable to other mainstream 8-bit optimizers.
• Quantization Precision: The second moment (variance) in Adafactor is strictly non-negative and spans multiple orders of magnitude. By mapping it to UINT8 in log2 space rather than linear space, the optimizer preserves relative precision for small variances, mitigating the instability that can arise from outlier gradients in standard 8-bit quantization.

📦 Installation

This project uses JIT (Just-In-Time) compilation.

Please ensure torch and ninja are installed, and a CUDA compiler (such as MSVC or GCC) is available in your environment.

If CUDA compilation fails, the optimizer will automatically fall back to the pure PyTorch implementation.

From PyPI

pip install -U adafactor8bit

From Source

pip install git+https://github.com/yanfeiwong/adafactor-8bit.git

[!IMPORTANT] First-Time Compilation: The first time you instantiate the optimizer (or run the example script), it will synchronously trigger the JIT compilation of the CUDA source code. This may take anywhere from a few seconds to a couple of minutes depending on your system, and the terminal might appear unresponsive. Once compiled, the extension is cached for reuse.

🚀 Quick Start

Using it is as simple as using a standard PyTorch optimizer.

from adafactor8bit import Adafactor8Bit

optimizer = Adafactor8Bit(model.parameters(), lr=1e-3)

[!TIP] Passing model.parameters() directly works for a quick test. In production, param_groups are recommended to protect sensitive layers (Norms, Biases) from quantization and weight decay. For sparse token embeddings (large vocabularies + small batch sizes), please refer to the Advanced Example to avoid cold-start instability.

from adafactor8bit import Adafactor8Bit

def get_param_groups(model, weight_decay=1e-2):
    decay, no_decay = [], []
    for name, param in model.named_parameters():
        if not param.requires_grad: continue
        
        # Protect 1D tensors, biases, norms, embeddings, and LM head
        if param.ndim <= 1 or "bias" in name or "norm" in name or "embed" in name or "lm_head" in name:
            no_decay.append(param)
        else:
            decay.append(param)
            
    return [
        {"params": decay, "weight_decay": weight_decay, "quantize": True},
        {"params": no_decay, "weight_decay": 0.0, "quantize": False}
    ]

model = MyModel().cuda()
optimizer = Adafactor8Bit(
    get_param_groups(model), 
    lr=1e-3, 
    # For continual learning or when using an external LR scheduler
    relative_step=False,     # Disable internal LR scheduling
    beta2=0.999,             # Lock EMA window to prevent "blunting" over steps
)

# Training loop...

🛠️ Advanced Example

Here we demonstrate a hybrid grouping strategy for complex hybrid architectures (e.g., Vision-Language Models, Diffusion UNets) to achieve stable and efficient training.

📌 The following strategies are applied:

Implementation:

from adafactor8bit import Adafactor8Bit

# Define learning rates
lr = 1e-3
lr_adam = 1e-4 # For Embedding and N-D layers, we use an Adam-style learning rate

apollo_targets = ["attn", "mlp"]

def get_param_groups(model, lr_adam, weight_decay, apollo_rank=256):
    group_1d, group_embed, group_lm_head, group_apollo, group_2d, group_nd = [], [], [], [], [], []

    for name, param in model.named_parameters():
        if not param.requires_grad: continue
        
        is_1d = param.ndim <= 1 or "bias" in name or "norm" in name
        # Match true Token Embeddings, excluding Position and Time Embeddings
        is_embedding = ("embed" in name.lower() 
                        and "position" not in name.lower() 
                        and "pos_embed" not in name.lower()
                        and "time" not in name.lower())
        is_lm_head = param.ndim == 2 and "lm_head" in name.lower()
        is_apollo_target = param.ndim == 2 and any(t in name for t in apollo_targets)
        
        if is_1d:
            group_1d.append(param)
        elif is_embedding:
            group_embed.append(param)
        elif is_lm_head:
            group_lm_head.append(param)
        elif is_apollo_target:
            group_apollo.append(param)
        elif param.ndim == 2:
            group_2d.append(param)
        else:
            group_nd.append(param)

    # Common configuration for Adam-style groups (element-wise, no RMS scaling, no clipping)
    adam_style = {
        "factored": False,         # full-rank V
        "scale_parameter": False,  # no parameter RMS scaling
        "d": 0.0,                  # no RMS clipping
        "beta3": None,             # CAME requires factored=True
        "apollo_rank": 0,
        "lr": lr_adam,  
    }

    return [
        # 1. 1D / Sensitive: FP32, No Weight Decay
        {"params": group_1d, "weight_decay": 0.0, "quantize": False, "apollo_rank": 0},
        
        # 2. Embeddings: Momentum-free Adam-style scaling
        {
            "params": group_embed, 
            "weight_decay": 0.0, 
            "quantize": False,
            "beta1": None,             # Momentum-free
            **adam_style,
        },

        # 3. LM Head: Adam-style with momentum (avoid factored distortion)
        {
            "params": group_lm_head, 
            "weight_decay": 0.0, 
            "quantize": True,
            "beta1": 0.9,
            **adam_style,
        },
        
        # 4. 2D Weights (APOLLO targets): 8-bit quantization, APOLLO low-rank projection
        {
            "params": group_apollo, 
            "weight_decay": weight_decay, 
            "quantize": True, 
            "apollo_rank": apollo_rank,
            "beta1": 0.9,              # Remove if minimizing optimizer memory is the priority.
        },

        # 5. 2D Weights (Others): Default Adafactor path (factored), 8-bit quantization
        {
            "params": group_2d, 
            "weight_decay": weight_decay, 
            "quantize": True, 
        },
        
        # 6. >2D Weights: 8-bit quantization, Weight Decay, Full-Rank
        {
            "params": group_nd, 
            "weight_decay": weight_decay, 
            "quantize": True, 
            "beta1": 0.9,
            **adam_style,
        },
    ]

model = MyModel().cuda()
optimizer = Adafactor8Bit(
    get_param_groups(model, lr_adam=lr_adam, weight_decay=1e-2, apollo_rank=256),
    lr=lr,
    # For continual learning or when using an external LR scheduler
    relative_step=False,              # Disable internal LR scheduling
    beta2=0.999,                      # Lock EMA window to prevent "blunting" over steps
    enable_fira_for_adafactor=True    # Enable Fira Limiter for all non-APOLLO paths
)

# Training loop...

[!NOTE] For more complete examples, please refer to the examples folder.

⚙️ Advanced Configuration

Continual Learning (beta2 & relative_step)

By default, Adafactor's second-moment decay rate dynamically decays with the training step, and the internal learning rate schedule (relative_step) scales the learning rate accordingly.

For endless fine-tuning or lifelong learning, this often leads to overly small learning rates and "blunted" second-moment estimates. To avoid these issues and keep the optimizer responsive:

• Set relative_step=False to disable the built-in LR schedule (allowing you to use an external scheduler).
• Set beta2=0.999 to lock the EMA window (similar to Adam).

Decoupled Weight Decay (scale_weight_decay=False)

By default, Adafactor's weight decay is coupled with the parameter's RMS scale.

• If you prefer the AdamW-style decoupled weight decay, set scale_weight_decay=False.

Fira Limiter

The Norm-Growth Limiter limits the relative increase of update norms to suppress destructive gradient spikes.

• enable_fira_for_apollo: Defaults to True. The APOLLO path applies the limiter by default to guard against sudden gradient rises in the low-rank subspace.
• enable_fira_for_adafactor: Defaults to False. Enables the limiter for the non-APOLLO paths.
• fira_margin: Defaults to 0.01. The tolerance margin for norm growth, shared by both switches. The limiter activates only when the current update norm grows by more than this margin (e.g., 0.01 = 1%) compared to the previous step.

Full-Rank and Factorized Variance (factored)

By default, Adafactor factorizes the second moment of $\ge$ 2D tensors into row and column statistics (factored=True) to minimize state memory. Setting factored=False switches to full element-wise variance (similar to RMSProp), while still retaining Adafactor's update RMS clipping mechanism (controlled by the d parameter). This configuration can be useful in the following scenarios:

• Convolution Weights (>2D): Setting factored=False maintains independent variance for each spatial position in the convolution kernel, enabling finer per-element gradient scaling.
• Sparse Embeddings: Combining factored=False, scale_parameter=False, d=0.0 and a lower learning rate creates a momentum-free adaptive optimizer. This allows for fine-grained, per-token updates while avoiding cold-token interference.

No-Compiler Environments (use_cuda_kernel=False)

If you are in an environment without a CUDA compiler and want to bypass JIT compilation entirely:

• Set use_cuda_kernel=False to fall back to the pure PyTorch implementation.

🌌 APOLLO Low-Rank Subspace Projection

Enable the APOLLO path to compute gradient scaling factors in a memory-efficient low-rank subspace. Compared to Adafactor's standard row/column factorization (which assumes spatial independence), APOLLO uses random subspace projection to capture cross-dimensional covariance information while keeping memory overhead low.

• apollo_rank: The target rank for the projection subspace. The default is 0 (disabled).

  • The official APOLLO GitHub repository recommends a rank of 256 for 1B and 7B models.
  • The LLaMA-Factory default is 16.
  • Setting this to 1 (APOLLO-Mini style) minimizes the low-rank state (saves even more VRAM than the Adafactor path). The original APOLLO-Mini relies on the first-moment (beta1) to smooth out projection noise. To replicate this, set beta1=0.9 alongside apollo_rank=1.

• apollo_scale: Heuristic scale parameter used to compensate for approximation error introduced by low-rank gradient scaling. The implementation applies sqrt(apollo_scale) to the resulting scaling factor. Defaults to 1.0.

• apollo_scale_type: Determines how the scaling factor is applied. 'channel' applies it per channel (Standard APOLLO), while 'tensor' applies it globally (APOLLO-Mini).

• apollo_update_proj_gap: Steps between projection matrix refreshes. Defaults to 200.

• apollo_factorize (Experimental): Applies Adafactor-style row/column factorization within the low-rank subspace to further reduce state memory overhead.

🧊 CAME Confidence-Guided Updates

Enable the CAME (Confidence-guided Adaptive Memory Efficient Optimization) path to add a confidence estimation stage after momentum accumulation:

Adaptive Scaling ($V$) → Momentum Accumulation ($M$) → Confidence Weighting ($C$)

Key Parameters & Tuning

The confidence stage measures the consistency between the current update direction and historical momentum, adaptively suppressing highly oscillatory updates.

• beta3: EMA decay coefficient for the confidence matrix. Requires beta1 (momentum) and factored=True. Defaults to None (disabled).
• Learning Rate: The official CAME implementation recommends 0.5–0.9× the AdamW learning rate (see official tuning guide). To align with the original CAME behavior in this library, disable Adafactor's RMS scaling (scale_parameter=False) and set d=1.0 (which corresponds to CAME's clip_threshold).
• Warmup: Since the confidence matrix is zero-initialized without bias correction, a learning rate warmup is recommended to safely establish the confidence baseline.
• Choosing beta3: beta3 should generally be larger than beta2 so the confidence estimate evolves more slowly than the variance estimate. A practical starting range is 0.9995–0.99995 when beta2=0.999.

Configuration Example

To replicate "vanilla" CAME (stripping Adafactor's native modifications), you can use the following configuration:

{
    "params": param_group,
    "lr": lr,                           # Original CAME recommends 0.5-0.9x AdamW LR
    "beta1": 0.9,
    "beta2": 0.999,
    "beta3": 0.9999,                    # Enable CAME confidence guidance
    "apollo_rank": 0,                   # Set to 0 for vanilla CAME. (Set >0 to enable APOLLO+CAME fusion)
    "weight_decay": weight_decay,
    "scale_weight_decay": False,
    "scale_parameter": False,           # Disable Adafactor RMS scaling to align with vanilla CAME
    "d": 1.0,
    "relative_step": False,
},

📈 Learning Rate Guide for Beginners

If you are migrating from optimizers like AdamW, Adafactor's learning rate behavior might feel a bit different. This is mainly due to the scale_parameter option.

• scale_parameter=True (default) Because of RMS scaling, a very small lr (e.g., 1e-5) often leads to extremely slow progress. Start with lr=1e-3 and adjust in the range 1e-4–5e-3 if needed.

• scale_parameter=False Disables RMS scaling, making the update scale more similar to AdamW. Use the learning rates you're familiar with for AdamW and tune as usual. (Note: the second moment is still factorized, so behavior is not identical.)

These are safe starting points. Always validate on your own task and batch size.

📌 Implementation Differences from Reference Optimizers

To maintain a unified and robust framework, some algorithmic paths have implementation-level differences compared to their original standalone repositories:

• APOLLO: Uses deterministic per-state seed allocation and an optimizer-level seed counter for projection matrix management.
• Adafactor: When beta1 is enabled, the momentum update ordering differs from the Hugging Face implementation.

🎓 Acknowledgements

This project builds upon the foundational work of several researchers and open-source communities. Sincere thanks to the following for their invaluable contributions:

Core Algorithm & Optimizer Design

• Noam Shazeer & Mitchell Stern for proposing the original Adafactor algorithm (Adafactor: Adaptive Learning Rates with Sublinear Memory Cost).
• Tim Dettmers for the inspiration from 8-bit block-wise quantization (8-BIT OPTIMIZERS VIA BLOCK-WISE QUANTIZATION) and the bitsandbytes library.
• Hanqing Zhu, Zhenyu Zhang, et al. for the APOLLO algorithm (APOLLO: SGD-Like Memory, AdamW-level Performance).
• Xi Chen, Kaituo Feng, et al. for the Norm-Growth Limiter mechanism in Fira (Fira: Can We Achieve Full-rank Training of LLMs Under Low-rank Constraint?).
• Yang Luo, et al. for the confidence-guided strategy in CAME (CAME: Confidence-guided Adaptive Memory Efficient Optimization).

Quantization & Implementation

• The QLoRA Team for their work on memory-efficient fine-tuning and quantization techniques (QLoRA: Efficient Finetuning of Quantized LLMs).
• The PyTorch AO Team for their work on 4-bit optimizer states, validating distribution-aware quantization for optimizer moments.
• The PyTorch Team for providing the foundational optimizer implementation and the C++ Extension toolchain.

Technical Review & Discussion

• Qwen, ChatGLM, and DeepSeek (large language models) for valuable technical discussions and code reviews on CUDA low-level optimization, memory safety mechanisms, and cross-platform compilation pipeline design.

🏛️ License

The project is released under the MIT License.

⭐ Star the Project

If this optimizer has been useful in your work, consider giving the repository a star. It helps others discover the project and supports future development.