nexusquant-kv 0.4.0__tar.gz

nexusquant_kv-0.4.0/LICENSE

@@ -0,0 +1,98 @@
+                                 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.
+
+      "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.
+
+      "Contribution" shall mean any work of authorship submitted to the
+      Licensor for inclusion in the Work.
+
+      "Contributor" shall mean Licensor and any Legal Entity on behalf
+      of whom a Contribution has been received by the Licensor.
+
+   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
+      patent license to make, have made, use, offer to sell, sell,
+      import, and otherwise transfer the Work.
+
+   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; and
+
+      (d) If the Work includes a "NOTICE" text file, You must include
+          a readable copy of the attribution notices contained within
+          such NOTICE file.
+
+   5. Submission of Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND.
+
+   8. Limitation of Liability. In no event shall any Contributor be
+      liable for any damages arising from the Work.
+
+   9. Accepting Warranty or Additional Liability.
+
+   Copyright 2026 NexusQuant Team
+
+   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.

nexusquant_kv-0.4.0/PKG-INFO

@@ -0,0 +1,142 @@
+Metadata-Version: 2.4
+Name: nexusquant-kv
+Version: 0.4.0
+Summary: Training-free KV cache compression via E8 lattice quantization and attention-aware token eviction
+Author: Joao Marques
+License: Apache-2.0
+Project-URL: Repository, https://github.com/jagmarques/nexusquant
+Keywords: kv-cache,compression,quantization,llm,inference,e8-lattice,token-eviction,transformers,long-context,memory-efficient,attention,vector-quantization
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Science/Research
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: Apache Software License
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.9
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Operating System :: OS Independent
+Requires-Python: >=3.9
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: torch>=2.0
+Requires-Dist: numpy>=1.24
+Requires-Dist: zstandard>=0.22
+Provides-Extra: hf
+Requires-Dist: transformers>=4.40; extra == "hf"
+Provides-Extra: all
+Requires-Dist: transformers>=4.40; extra == "all"
+Requires-Dist: accelerate>=0.26; extra == "all"
+Dynamic: license-file
+
+<p align="center">
+  <strong>NexusQuant</strong>
+</p>
+<p align="center">
+  Compress your LLM's KV cache 10-33x. Training-free. One line of code.
+</p>
+<p align="center">
+  <a href="https://github.com/jagmarques/nexusquant/blob/main/LICENSE"><img src="https://img.shields.io/badge/License-Apache_2.0-blue.svg?style=flat-square" alt="License"></a>
+  <a href="https://www.python.org/downloads/"><img src="https://img.shields.io/badge/python-3.9+-blue?style=flat-square&logo=python&logoColor=white" alt="Python"></a>
+  <a href="https://github.com/jagmarques/nexusquant"><img src="https://img.shields.io/github/stars/jagmarques/nexusquant?style=social" alt="Stars"></a>
+</p>
+
+---
+
+Token eviction + E8 lattice quantization, applied once after prefill. No training, no calibration data, no model modifications.
+
+## Install
+
+```bash
+pip install nexusquant
+pip install "nexusquant[hf]"  # with HuggingFace transformers
+```
+
+## Quickstart
+
+```python
+from nexusquant import nexusquant_evict
+
+with nexusquant_evict(model, quality="balanced"):
+    output = model.generate(input_ids, max_new_tokens=512)
+```
+
+## Why
+
+| Without NexusQuant | With NexusQuant |
+|---|---|
+| 128K context → 80 GB KV cache | 128K context → 5 GB KV cache (17x) |
+| OOM at 32K on a single A100 | 500K+ tokens on one A100 |
+| Needs 8× A100 cluster for long context | Single GPU, single machine |
+| Deploy a fine-tuned retrieval model | One `with` block, no code changes |
+
+## Quality presets
+
+Measured on Mistral-7B, A100, FP16. Compression ratios include all overhead (scales, indices, metadata).
+
+| Preset | Compression | PPL degradation | Context on 80 GB |
+|---|---|---|---|
+| `high` | 10x | +0.4% | ~1.3M tokens |
+| `balanced` | 17x | +1.3% | ~2.2M tokens |
+| `max` | 33x | +2.6% | ~4.2M tokens |
+
+Validated on Mistral-7B, TinyLlama-1.1B, Llama-3-8B across academic, technical, and creative text.
+
+## How it works
+
+1. **Importance scoring** — rank tokens by cross-head attention weight (key-key dot product)
+2. **Token eviction** — drop lowest-scoring tokens; always keep BOS and a recent sliding window
+3. **RoPE removal** — undo rotary embeddings on keys so they share a common subspace, reducing quantization error ~0.7 pp
+4. **Hadamard rotation** — spread energy uniformly across dimensions so no outlier dominates the quantization scale
+5. **E8 lattice quantization** — quantize 8-float groups onto the E8 root lattice (densest sphere packing in 8D), 2 bits/dim
+6. **Delta coding + zstd** — consecutive tokens produce similar lattice indices; storing deltas then compressing with zstd yields another 2-3x on the index stream
+
+Token eviction reduces *count* (2.5x at 60% eviction). E8 quantization reduces *precision* (~7x after entropy coding). Combined: 17x.
+
+## Compared to
+
+| Method | Compression | PPL degradation | Training required |
+|---|---|---|---|
+| **NexusQuant** | **10-33x** | **+0.4-2.6%** | **No** |
+| TurboQuant (Google) | ~5-6x | ~0% | No |
+| KVTC (NVIDIA) | up to 20x | <1% | Yes (calibration, ~10 min) |
+| CommVQ (Apple) | ~8x | ~0% | Yes (full retraining) |
+| Palu | 11x | ~25% rel | Yes (calibration) |
+
+NexusQuant is the highest-compression training-free method. KVTC achieves comparable ratios with better quality but requires calibration data. Competitor numbers are from their published papers, not reproduced on our hardware.
+
+## Supported models
+
+Any HuggingFace causal LM using split-half RoPE (the standard since Llama-2):
+
+- Llama family (Llama-2, Llama-3, Llama-3.1)
+- Mistral / Mixtral
+- Qwen
+- Phi
+- Gemma
+
+Not yet supported: models with interleaved RoPE (GPT-NeoX, GPT-J).
+
+## Limitations
+
+- **Quality is text-dependent.** Creative/narrative text degrades more than structured/technical text at the same compression ratio. Test on your actual workload before deploying.
+- **Short prefixes hurt.** Prefixes under 500 tokens see more degradation than the numbers above, which were measured at 1600-3500 tokens. The importance scorer needs enough tokens to distinguish signal from noise.
+- **E8 quantization is CPU-bound.** A production deployment needs Triton/CUDA kernels for the quantization step. The current implementation writes dequantized values back to the cache for compatibility — actual GPU memory savings require native compact storage.
+- **Eviction is permanent.** Evicted tokens are gone. If your task requires precise recall of a specific token, measure eviction sensitivity on that task first.
+
+## Citation
+
+```bibtex
+@software{nexusquant2025,
+  author  = {Marques, Joao},
+  title   = {{NexusQuant}: Training-Free {KV} Cache Compression via {E8} Lattice Quantization},
+  year    = {2025},
+  url     = {https://github.com/jagmarques/nexusquant},
+  license = {Apache-2.0},
+}
+```
+
+## License
+
+Apache 2.0. See [LICENSE](LICENSE).

nexusquant_kv-0.4.0/README.md

@@ -0,0 +1,110 @@
+<p align="center">
+  <strong>NexusQuant</strong>
+</p>
+<p align="center">
+  Compress your LLM's KV cache 10-33x. Training-free. One line of code.
+</p>
+<p align="center">
+  <a href="https://github.com/jagmarques/nexusquant/blob/main/LICENSE"><img src="https://img.shields.io/badge/License-Apache_2.0-blue.svg?style=flat-square" alt="License"></a>
+  <a href="https://www.python.org/downloads/"><img src="https://img.shields.io/badge/python-3.9+-blue?style=flat-square&logo=python&logoColor=white" alt="Python"></a>
+  <a href="https://github.com/jagmarques/nexusquant"><img src="https://img.shields.io/github/stars/jagmarques/nexusquant?style=social" alt="Stars"></a>
+</p>
+
+---
+
+Token eviction + E8 lattice quantization, applied once after prefill. No training, no calibration data, no model modifications.
+
+## Install
+
+```bash
+pip install nexusquant
+pip install "nexusquant[hf]"  # with HuggingFace transformers
+```
+
+## Quickstart
+
+```python
+from nexusquant import nexusquant_evict
+
+with nexusquant_evict(model, quality="balanced"):
+    output = model.generate(input_ids, max_new_tokens=512)
+```
+
+## Why
+
+| Without NexusQuant | With NexusQuant |
+|---|---|
+| 128K context → 80 GB KV cache | 128K context → 5 GB KV cache (17x) |
+| OOM at 32K on a single A100 | 500K+ tokens on one A100 |
+| Needs 8× A100 cluster for long context | Single GPU, single machine |
+| Deploy a fine-tuned retrieval model | One `with` block, no code changes |
+
+## Quality presets
+
+Measured on Mistral-7B, A100, FP16. Compression ratios include all overhead (scales, indices, metadata).
+
+| Preset | Compression | PPL degradation | Context on 80 GB |
+|---|---|---|---|
+| `high` | 10x | +0.4% | ~1.3M tokens |
+| `balanced` | 17x | +1.3% | ~2.2M tokens |
+| `max` | 33x | +2.6% | ~4.2M tokens |
+
+Validated on Mistral-7B, TinyLlama-1.1B, Llama-3-8B across academic, technical, and creative text.
+
+## How it works
+
+1. **Importance scoring** — rank tokens by cross-head attention weight (key-key dot product)
+2. **Token eviction** — drop lowest-scoring tokens; always keep BOS and a recent sliding window
+3. **RoPE removal** — undo rotary embeddings on keys so they share a common subspace, reducing quantization error ~0.7 pp
+4. **Hadamard rotation** — spread energy uniformly across dimensions so no outlier dominates the quantization scale
+5. **E8 lattice quantization** — quantize 8-float groups onto the E8 root lattice (densest sphere packing in 8D), 2 bits/dim
+6. **Delta coding + zstd** — consecutive tokens produce similar lattice indices; storing deltas then compressing with zstd yields another 2-3x on the index stream
+
+Token eviction reduces *count* (2.5x at 60% eviction). E8 quantization reduces *precision* (~7x after entropy coding). Combined: 17x.
+
+## Compared to
+
+| Method | Compression | PPL degradation | Training required |
+|---|---|---|---|
+| **NexusQuant** | **10-33x** | **+0.4-2.6%** | **No** |
+| TurboQuant (Google) | ~5-6x | ~0% | No |
+| KVTC (NVIDIA) | up to 20x | <1% | Yes (calibration, ~10 min) |
+| CommVQ (Apple) | ~8x | ~0% | Yes (full retraining) |
+| Palu | 11x | ~25% rel | Yes (calibration) |
+
+NexusQuant is the highest-compression training-free method. KVTC achieves comparable ratios with better quality but requires calibration data. Competitor numbers are from their published papers, not reproduced on our hardware.
+
+## Supported models
+
+Any HuggingFace causal LM using split-half RoPE (the standard since Llama-2):
+
+- Llama family (Llama-2, Llama-3, Llama-3.1)
+- Mistral / Mixtral
+- Qwen
+- Phi
+- Gemma
+
+Not yet supported: models with interleaved RoPE (GPT-NeoX, GPT-J).
+
+## Limitations
+
+- **Quality is text-dependent.** Creative/narrative text degrades more than structured/technical text at the same compression ratio. Test on your actual workload before deploying.
+- **Short prefixes hurt.** Prefixes under 500 tokens see more degradation than the numbers above, which were measured at 1600-3500 tokens. The importance scorer needs enough tokens to distinguish signal from noise.
+- **E8 quantization is CPU-bound.** A production deployment needs Triton/CUDA kernels for the quantization step. The current implementation writes dequantized values back to the cache for compatibility — actual GPU memory savings require native compact storage.
+- **Eviction is permanent.** Evicted tokens are gone. If your task requires precise recall of a specific token, measure eviction sensitivity on that task first.
+
+## Citation
+
+```bibtex
+@software{nexusquant2025,
+  author  = {Marques, Joao},
+  title   = {{NexusQuant}: Training-Free {KV} Cache Compression via {E8} Lattice Quantization},
+  year    = {2025},
+  url     = {https://github.com/jagmarques/nexusquant},
+  license = {Apache-2.0},
+}
+```
+
+## License
+
+Apache 2.0. See [LICENSE](LICENSE).

nexusquant_kv-0.4.0/nexusquant/__init__.py

@@ -0,0 +1,32 @@
+"""NexusQuant: Training-Free KV Cache Compression via E8 Lattice VQ + Token Eviction.
+
+GPU-validated numbers (Mistral-7B, A100, all overhead included):
+  10.4x at +0.43% PPL (quality="high", 35% eviction)
+  16.8x at +1.34% PPL (quality="balanced", 60% eviction)
+  33.3x at +2.64% PPL (quality="max", 80% eviction)
+
+Training-free. Zero calibration. One line of code. Apache 2.0.
+
+Quick start:
+    from nexusquant import nexusquant_evict
+
+    with nexusquant_evict(model, quality="balanced"):
+        output = model.generate(input_ids, max_new_tokens=100)
+"""
+
+__version__ = "0.4.0"
+
+from nexusquant.pipeline import (
+    NexusQuantFast,
+    NexusQuantSimple,
+    NexusQuantMax,
+    NexusQuantEvict,
+    compress_kv_cache,
+)
+
+from nexusquant.integrations.huggingface import (
+    nexusquant,
+    nexusquant_simple,
+    nexusquant_max,
+    nexusquant_evict,
+)

nexusquant_kv-0.4.0/nexusquant/core/__init__.py

@@ -0,0 +1,40 @@
+from nexusquant.core.e8_lattice import E8Lattice
+from nexusquant.core.hadamard import hadamard_matrix, fht, ifht
+from nexusquant.core.dp_allocator import (
+    dp_bit_allocation,
+    calibrate_distortion_factors,
+    DISTORTION_THEORETICAL,
+    DISTORTION_EMPIRICAL,
+)
+from nexusquant.core.rope_utils import inverse_rope, forward_rope
+from nexusquant.core.token_merger import merge_tokens, merge_and_drop
+from nexusquant.core.entropy_coder import (
+    encode_e8,
+    decode_e8,
+    measure_entropy,
+    measure_e8_entropy,
+    e8_quantize_with_entropy,
+)
+from nexusquant.core.temporal_codec import (
+    compress_indices,
+    decompress_indices,
+    temporal_delta_encode,
+    temporal_delta_decode,
+    measure_compression,
+)
+from nexusquant.core.nsn import forward_nsn, inverse_nsn, NSNStats
+from nexusquant.core.tcc import (
+    forward_tcc,
+    inverse_tcc,
+    compute_compression_stats,
+    TCCCompressed,
+)
+from nexusquant.core.optimal_shrinkage import (
+    estimate_noise_sigma,
+    optimal_shrinkage_frobenius,
+    optimal_hard_threshold,
+    effective_dims,
+    variance_retained,
+    reconstruct_with_shrinkage,
+    mp_median,
+)