Metadata-Version: 2.4
Name: trimkv
Version: 0.1.5
Summary: TrimKV: token retention for memory-bounded key-value eviction
Author-email: Ngoc Bui <ngocbh.pt@gmail.com>
License: 
                                         Apache License
                                   Version 2.0, January 2004
                                http://www.apache.org/licenses/
        
           TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
        
           1. Definitions.
        
              "License" shall mean the terms and conditions for use, reproduction,
              and distribution as defined by Sections 1 through 9 of this document.
        
              "Licensor" shall mean the copyright owner or entity authorized by
              the copyright owner that is granting the License.
        
              "Legal Entity" shall mean the union of the acting entity and all
              other entities that control, are controlled by, or are under common
              control with that entity. For the purposes of this definition,
              "control" means (i) the power, direct or indirect, to cause the
              direction or management of such entity, whether by contract or
              otherwise, or (ii) ownership of fifty percent (50%) or more of the
              outstanding shares, or (iii) beneficial ownership of such entity.
        
              "You" (or "Your") shall mean an individual or Legal Entity
              exercising permissions granted by this License.
        
              "Source" form shall mean the preferred form for making modifications,
              including but not limited to software source code, documentation
              source, and configuration files.
        
              "Object" form shall mean any form resulting from mechanical
              transformation or translation of a Source form, including but
              not limited to compiled object code, generated documentation,
              and conversions to other media types.
        
              "Work" shall mean the work of authorship, whether in Source or
              Object form, made available under the License, as indicated by a
              copyright notice that is included in or attached to the work
              (an example is provided in the Appendix below).
        
              "Derivative Works" shall mean any work, whether in Source or Object
              form, that is based on (or derived from) the Work and for which the
              editorial revisions, annotations, elaborations, or other modifications
              represent, as a whole, an original work of authorship. For the purposes
              of this License, Derivative Works shall not include works that remain
              separable from, or merely link (or bind by name) to the interfaces of,
              the Work and Derivative Works thereof.
        
              "Contribution" shall mean any work of authorship, including
              the original version of the Work and any modifications or additions
              to that Work or Derivative Works thereof, that is intentionally
              submitted to Licensor for inclusion in the Work by the copyright owner
              or by an individual or Legal Entity authorized to submit on behalf of
              the copyright owner. For the purposes of this definition, "submitted"
              means any form of electronic, verbal, or written communication sent
              to the Licensor or its representatives, including but not limited to
              communication on electronic mailing lists, source code control systems,
              and issue tracking systems that are managed by, or on behalf of, the
              Licensor for the purpose of discussing and improving the Work, but
              excluding communication that is conspicuously marked or otherwise
              designated in writing by the copyright owner as "Not a Contribution."
        
              "Contributor" shall mean Licensor and any individual or Legal Entity
              on behalf of whom a Contribution has been received by Licensor and
              subsequently incorporated within the Work.
        
           2. Grant of Copyright License. Subject to the terms and conditions of
              this License, each Contributor hereby grants to You a perpetual,
              worldwide, non-exclusive, no-charge, royalty-free, irrevocable
              copyright license to reproduce, prepare Derivative Works of,
              publicly display, publicly perform, sublicense, and distribute the
              Work and such Derivative Works in Source or Object form.
        
           3. Grant of Patent License. Subject to the terms and conditions of
              this License, each Contributor hereby grants to You a perpetual,
              worldwide, non-exclusive, no-charge, royalty-free, irrevocable
              (except as stated in this section) patent license to make, have made,
              use, offer to sell, sell, import, and otherwise transfer the Work,
              where such license applies only to those patent claims licensable
              by such Contributor that are necessarily infringed by their
              Contribution(s) alone or by combination of their Contribution(s)
              with the Work to which such Contribution(s) was submitted. If You
              institute patent litigation against any entity (including a
              cross-claim or counterclaim in a lawsuit) alleging that the Work
              or a Contribution incorporated within the Work constitutes direct
              or contributory patent infringement, then any patent licenses
              granted to You under this License for that Work shall terminate
              as of the date such litigation is filed.
        
           4. Redistribution. You may reproduce and distribute copies of the
              Work or Derivative Works thereof in any medium, with or without
              modifications, and in Source or Object form, provided that You
              meet the following conditions:
        
              (a) You must give any other recipients of the Work or
                  Derivative Works a copy of this License; and
        
              (b) You must cause any modified files to carry prominent notices
                  stating that You changed the files; and
        
              (c) You must retain, in the Source form of any Derivative Works
                  that You distribute, all copyright, patent, trademark, and
                  attribution notices from the Source form of the Work,
                  excluding those notices that do not pertain to any part of
                  the Derivative Works; and
        
              (d) If the Work includes a "NOTICE" text file as part of its
                  distribution, then any Derivative Works that You distribute must
                  include a readable copy of the attribution notices contained
                  within such NOTICE file, excluding those notices that do not
                  pertain to any part of the Derivative Works, in at least one
                  of the following places: within a NOTICE text file distributed
                  as part of the Derivative Works; within the Source form or
                  documentation, if provided along with the Derivative Works; or,
                  within a display generated by the Derivative Works, if and
                  wherever such third-party notices normally appear. The contents
                  of the NOTICE file are for informational purposes only and
                  do not modify the License. You may add Your own attribution
                  notices within Derivative Works that You distribute, alongside
                  or as an addendum to the NOTICE text from the Work, provided
                  that such additional attribution notices cannot be construed
                  as modifying the License.
        
              You may add Your own copyright statement to Your modifications and
              may provide additional or different license terms and conditions
              for use, reproduction, or distribution of Your modifications, or
              for any such Derivative Works as a whole, provided Your use,
              reproduction, and distribution of the Work otherwise complies with
              the conditions stated in this License.
        
           5. Submission of Contributions. Unless You explicitly state otherwise,
              any Contribution intentionally submitted for inclusion in the Work
              by You to the Licensor shall be under the terms and conditions of
              this License, without any additional terms or conditions.
              Notwithstanding the above, nothing herein shall supersede or modify
              the terms of any separate license agreement you may have executed
              with Licensor regarding such Contributions.
        
           6. Trademarks. This License does not grant permission to use the trade
              names, trademarks, service marks, or product names of the Licensor,
              except as required for reasonable and customary use in describing the
              origin of the Work and reproducing the content of the NOTICE file.
        
           7. Disclaimer of Warranty. Unless required by applicable law or
              agreed to in writing, Licensor provides the Work (and each
              Contributor provides its Contributions) on an "AS IS" BASIS,
              WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
              implied, including, without limitation, any warranties or conditions
              of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
              PARTICULAR PURPOSE. You are solely responsible for determining the
              appropriateness of using or redistributing the Work and assume any
              risks associated with Your exercise of permissions under this License.
        
           8. Limitation of Liability. In no event and under no legal theory,
              whether in tort (including negligence), contract, or otherwise,
              unless required by applicable law (such as deliberate and grossly
              negligent acts) or agreed to in writing, shall any Contributor be
              liable to You for damages, including any direct, indirect, special,
              incidental, or consequential damages of any character arising as a
              result of this License or out of the use or inability to use the
              Work (including but not limited to damages for loss of goodwill,
              work stoppage, computer failure or malfunction, or any and all
              other commercial damages or losses), even if such Contributor
              has been advised of the possibility of such damages.
        
           9. Accepting Warranty or Additional Liability. While redistributing
              the Work or Derivative Works thereof, You may choose to offer,
              and charge a fee for, acceptance of support, warranty, indemnity,
              or other liability obligations and/or rights consistent with this
              License. However, in accepting such obligations, You may act only
              on Your own behalf and on Your sole responsibility, not on behalf
              of any other Contributor, and only if You agree to indemnify,
              defend, and hold each Contributor harmless for any liability
              incurred by, or claims asserted against, such Contributor by reason
              of your accepting any such warranty or additional liability.
        
           END OF TERMS AND CONDITIONS
        
           APPENDIX: How to apply the Apache License to your work.
        
              To apply the Apache License to your work, attach the following
              boilerplate notice, with the fields enclosed by brackets "[]"
              replaced with your own identifying information. (Don't include
              the brackets!)  The text should be enclosed in the appropriate
              comment syntax for the file format. We also recommend that a
              file or class name and description of purpose be included on the
              same "printed page" as the copyright notice for easier
              identification within third-party archives.
        
           Copyright [yyyy] [name of copyright owner]
        
           Licensed under the Apache License, Version 2.0 (the "License");
           you may not use this file except in compliance with the License.
           You may obtain a copy of the License at
        
               http://www.apache.org/licenses/LICENSE-2.0
        
           Unless required by applicable law or agreed to in writing, software
           distributed under the License is distributed on an "AS IS" BASIS,
           WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
           See the License for the specific language governing permissions and
           limitations under the License.
        
Project-URL: Homepage, https://github.com/ngocbh/trimkv
Project-URL: Bug Tracker, https://github.com/ngocbh/trimkv/issues
Project-URL: Source, https://github.com/ngocbh/trimkv
Project-URL: Paper, https://arxiv.org/abs/2405.14352
Keywords: kv-cache,compression,long-context,trimkv,dbtrimkv,transformers
Classifier: License :: OSI Approved :: Apache Software License
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.11
Classifier: Programming Language :: Python :: 3.12
Classifier: Operating System :: POSIX :: Linux
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Requires-Python: >=3.11
Description-Content-Type: text/markdown
License-File: LICENSE
Requires-Dist: torch>=2.7.0
Requires-Dist: transformers==4.57.1
Requires-Dist: accelerate
Requires-Dist: einops
Requires-Dist: numpy
Requires-Dist: huggingface_hub
Provides-Extra: wandb
Requires-Dist: wandb; extra == "wandb"
Provides-Extra: dev
Requires-Dist: pytest; extra == "dev"
Requires-Dist: pytest-xdist; extra == "dev"
Dynamic: license-file

# TrimKV: Token Retention for Memory-Bounded Key-Value Eviction

<a href="https://arxiv.org/abs/2405.14352"><img src="https://img.shields.io/badge/arxiv-2405.14352-red?style=for-the-badge"></a>

## 🚀 Updates

- **🆕 DBTrimKV — dynamic-budget variant powered by `PagedTrimKVCache`.** A single global KV budget is shared across all layers and heads and **reallocated on the fly** at every step, instead of fixing a per-head budget upfront. The retention gate's final projection is tied across layers and heads, and the runtime uses [`PagedTrimKVCache`](src/trimkv/cache_utils.py) — a paged-attention-style cache where blocks are dynamically (re)assigned to the heads that currently need them. The result: **significantly outperforms TrimKV at low KV budgets**, and matches or even beats the full KV cache — without any per-head tuning. Same training surface as TrimKV — just two env-var flips (`RETENTION_GATE=rg10`, `GLOBAL_CAPACITY=True`). Public LLM checkpoints:
  - [DBTrimKV-Qwen3-4B-Math](https://huggingface.co/ngocbh/DBTrimKV-Qwen3-4B-Math) · [DBTrimKV-Qwen3-4B-Instruct-2507](https://huggingface.co/ngocbh/DBTrimKV-Qwen3-4B-Instruct-2507)
- **🆕 First VLM release — TrimKV / DBTrimKV go multimodal.** Full Qwen3-VL / Qwen2.5-VL / LLaVA support with end-to-end training recipes in [`train/vlm/`](train/vlm/) and an evaluation harness in [`experiments/lmms-eval/`](experiments/lmms-eval/) and [`experiments/mmdu/`](experiments/mmdu/). Auto-downloading data prep for R1-Onevision, M4-Instruct, LLaVA-Video-178K, MMDU, and OpenR1-Math-220k under [`train/vlm/scripts/data/`](train/vlm/scripts/data/). Public VLM checkpoints (DBTrimKV):
  - [DBTrimKV-Qwen3-VL-8B-Thinking](https://huggingface.co/ngocbh/DBTrimKV-Qwen3-VL-8B-Thinking) · [DBTrimKV-Qwen3-VL-4B-Instruct](https://huggingface.co/ngocbh/DBTrimKV-Qwen3-VL-4B-Instruct)
- **Codebase refactor for transformers v4.57.0.** This release freezes the codebase at a version close to what produced the paper results, so all reported numbers are reproducible. If you hit issues, please open a GitHub issue.

---

## What is TrimKV?

> An efficient and learnable key–value eviction strategy designed to improve the efficiency of large language models (LLMs) in long-horizon inference.

Imagine what if our brain worked like a transformer:

<div align="center">
    <img width="1000" alt="teaser" src="https://raw.githubusercontent.com/ngocbh/trimkv/main/assets/fun.gif"/>
</div>

This is because it tried to remember every single detail (token) forever. TrimKV lets your model forget the parts that aren't very important so it doesn't melt its VRAM. Don't let the brain (or GPU) explode. 💥🧠

The core idea behind TrimKV is to learn the intrinsic importance of each key–value pair at creation time — what we call *token retention* — and then decay this importance exponentially over time to mimic standard inference running with eviction.

The retention score is **query-agnostic** and captures the **long-term utility** of tokens. This is different from attention scores, which are query-dependent: they capture short-term utility for predicting the next token, are recomputed at every step, and are highly dependent on the transient decoding state.

## TrimKV vs DBTrimKV

Both variants share the same training loop, datasets, and loss surface. They differ in **how the KV budget is allocated**, **which retention-gate parameterisation is used**, and **which cache class powers inference**:

| | TrimKV | DBTrimKV (new) |
|---|---|---|
| Budget semantics | per-layer, per-head **local** budget `M_local = M` | single **global** budget `M_global = M × num_layers × num_heads`, redistributed dynamically across layers/heads |
| Gate parameterisation | independent retention gate per head | final projection of the gate **tied across layers and heads** |
| Inference cache | `TrimKVCache` (fixed per-head allocation) | **`PagedTrimKVCache`** — paged-attention-style blocks dynamically (re)assigned to heads that currently need capacity |
| `RETENTION_GATE` flag | `rg` | `rg10` |
| `GLOBAL_CAPACITY` flag | `False` | `True` |

DBTrimKV's combination of the global retention gate with `PagedTrimKVCache` lets it run at much tighter **average** budgets while preserving accuracy — heads with high retention demand temporarily borrow capacity from heads with low demand on a per-step basis. See [`train/llm/README.md`](train/llm/README.md) and [`train/vlm/README.md`](train/vlm/README.md) for the full training surface, and [`src/trimkv/cache_utils.py`](src/trimkv/cache_utils.py) for the cache implementations.

## Why TrimKV?

It's fast

<div align="center">
    <img width="300" alt="teaser" src="https://raw.githubusercontent.com/ngocbh/trimkv/main/assets/speed.png"/>
</div>

It's smart

<div align="center">
    <img width="800" alt="teaser" src="https://raw.githubusercontent.com/ngocbh/trimkv/main/assets/performance.png"/>
</div>

---

## Getting started

### Requirements

- Python 3.11 or higher (tested with 3.12)
- PyTorch 2.7.0 or higher (tested with 2.8.0)
- FlashAttention 2.7.2.post1 or higher (tested with 2.8.0)
- Transformers 4.57.1

```sh
pip install -r requirements.txt
```

This is a minimal set of requirements for training. Additional dependencies may be needed for individual experiments; see [`examples/env.yaml`](examples/env.yaml) for a full reproducible environment.

### Installation

```sh
pip install trimkv
```

---

## Quick start

```python
import torch
from trimkv.models.qwen3 import TrimKVQwen3ForCausalLM
from trimkv.cache_utils import TrimKVCache, PagedTrimKVCache
from transformers import AutoTokenizer

# Pick any TrimKV / DBTrimKV checkpoint from the table below
model_path = "ngocbh/DBTrimKV-Qwen3-4B-Math"
download_from = "huggingface"  # also: "wandb", "local"

model = TrimKVQwen3ForCausalLM.from_pretrained(
    model_path,
    torch_dtype=torch.bfloat16,
    load_trimkv_weights=True,
    download_from=download_from,
    use_cache=True,
    device_map="cuda",
)
model.config._attn_implementation = "flash_attention_2"

tokenizer = AutoTokenizer.from_pretrained(
    model.config.base_model, use_fast=True, padding_side="left"
)

# PagedTrimKVCache is the inference-time cache used by DBTrimKV. It allocates a
# global pool of blocks and (re)assigns them to heads on the fly so heads with
# high retention demand can borrow capacity from heads with low demand.
# For (non-DB) TrimKV, swap in TrimKVCache(memory_size=..., buffer_size=..., device="cuda").
past_key_values = PagedTrimKVCache(
    num_layers=model.config.num_hidden_layers,
    num_heads=model.config.num_key_value_heads,
    max_seq_len=32768,
    memory_size=128,
    num_blocks_ratio=1.0,
    buffer_size=32,
    strategy="fixed_budget",
    device="cuda",
)

# Use model.generate as normal — pass past_key_values to enable TrimKV eviction.
```

For a runnable end-to-end example see [`examples/test_qwen3.py`](examples/test_qwen3.py). VLM checkpoints use `TrimKVQwen3VLForConditionalGeneration` from `trimkv.models.qwen3_vl` and the same `PagedTrimKVCache`, but read `model.config.text_config.num_hidden_layers` / `num_key_value_heads` instead.

---

## Training

- **LLMs (Qwen3, Qwen2, Llama, Phi-3):** [`train/llm/`](train/llm/) — DeepSpeed + 🤗 Trainer. Two recipes (`train_trimkv_long.sh` for long-context KL distillation, `train_trimkv_math.sh` for R1-style math reasoning). Same recipes train both TrimKV and DBTrimKV — flip `RETENTION_GATE` / `GLOBAL_CAPACITY` to switch.
- **VLMs (Qwen2.5-VL, Qwen3-VL, LLaVA):** [`train/vlm/`](train/vlm/) — same harness extended for visual data. Auto-downloading data prep for R1-Onevision, M4-Instruct, LLaVA-Video-178K, MMDU, and OpenR1-Math-220k under [`train/vlm/scripts/data/`](train/vlm/scripts/data/).

---

## Experiments

Per-benchmark evaluation harnesses live in [`experiments/`](experiments/) — see [experiments/README.md](experiments/README.md) for the full index.

- **Baselines:** TrimKV, DBTrimKV, R-KV, SeerAttention, SnapKV, StreamingLLM, H2O, KeyDiff, LocRet.
- **Long-horizon generation:** GSM8K, MATH-500, AIME-24, LongProc.
- **Long-context understanding:** SCBench, LongMemEval, LongBench, LongBench v2.
- **Multimodal:** lmms-eval task suite (mathvision_testmini, video_mmmu_*, mmmu_pro_vision, videomme, videomathqa_mcq, mmstar) plus MMDU.

---

## Released models

### LLM checkpoints

| Base Model | Variant | Checkpoint | Training Datasets | Max Context Len | Training $M$ |
|---|---|---|---|---|---|
| Qwen3-1.7B               | TrimKV   | [TrimKV-Qwen3-1.7B-Math](https://huggingface.co/ngocbh/TrimKV-Qwen3-1.7B-Math)               | OpenR1-Math-220k             | 16K  | 512  |
| Qwen3-4B                 | TrimKV   | [TrimKV-Qwen3-4B-Math](https://huggingface.co/ngocbh/TrimKV-Qwen3-4B-Math)                   | OpenR1-Math-220k             | 16K  | 512  |
| Qwen3-8B                 | TrimKV   | [TrimKV-Qwen3-8B-Math](https://huggingface.co/ngocbh/TrimKV-Qwen3-8B-Math)                   | OpenR1-Math-220k             | 16K  | 512  |
| Qwen3-14B                | TrimKV   | [TrimKV-Qwen3-14B-Math](https://huggingface.co/ngocbh/TrimKV-Qwen3-14B-Math)                 | OpenR1-Math-220k             | 16K  | 512  |
| Qwen3-4B-Instruct-2507   | TrimKV   | [TrimKV-Qwen3-4B-Instruct-2507](https://huggingface.co/ngocbh/TrimKV-Qwen3-4B-Instruct-2507) | Synth-Long, BookSum, Buddhi  | 128K | 4096 |
| Phi-3-mini-128k-instruct | TrimKV   | [TrimKV-Phi-3-mini-128k-instruct](https://huggingface.co/ngocbh/TrimKV-Phi-3-mini-128k-instruct) | LongAlpaca               | 128K | 2048 |
| Qwen3-4B                 | **DBTrimKV** 🆕 | [DBTrimKV-Qwen3-4B-Math](https://huggingface.co/ngocbh/DBTrimKV-Qwen3-4B-Math)               | OpenR1-Math-220k             | 32K  | 128  |
| Qwen3-4B-Instruct-2507   | **DBTrimKV** 🆕 | [DBTrimKV-Qwen3-4B-Instruct-2507](https://huggingface.co/ngocbh/DBTrimKV-Qwen3-4B-Instruct-2507) | Synth-Long, BookSum, Buddhi | 128K | 512  |

### VLM checkpoints — first multimodal release 🆕

| Base Model | Variant | Checkpoint | Training Datasets | Max Context Len | Training $M$ |
|---|---|---|---|---|---|
| Qwen3-VL-8B-Thinking | DBTrimKV | [DBTrimKV-Qwen3-VL-8B-Thinking](https://huggingface.co/ngocbh/DBTrimKV-Qwen3-VL-8B-Thinking) | R1-Onevision, M4-Instruct, LLaVA-Video-178K, MMDU, OpenR1-Math-220k | 32K | 32 |
| Qwen3-VL-4B-Instruct | DBTrimKV | [DBTrimKV-Qwen3-VL-4B-Instruct](https://huggingface.co/ngocbh/DBTrimKV-Qwen3-VL-4B-Instruct) | M4-Instruct, MMDU | 32K | 32 |

Happy to mention here if you have your own checkpoints for different settings.

---

## Acknowledgements

A large portion of this repository is adapted from or built on top of the following projects:

- [SeerAttention](https://github.com/microsoft/SeerAttention)
- [R-KV](https://github.com/Zefan-Cai/R-KV)
- [MInference](https://github.com/microsoft/MInference) (SCBench)
- [Locret](https://github.com/huangyuxiang03/Locret)
- [LongMemEval](https://github.com/xiaowu0162/LongMemEval)
- [LongProc](https://github.com/princeton-pli/LongProc)
- [LongBench](https://github.com/THUDM/LongBench)
- [MMDU](https://github.com/Liuziyu77/MMDU)
- [lmms-eval](https://github.com/EvolvingLMMs-Lab/lmms-eval)
