pare-quant 0.1.0__tar.gz

pare_quant-0.1.0/.gitignore

@@ -0,0 +1,53 @@
+# Python
+__pycache__/
+*.py[cod]
+*.pyo
+*.pyd
+.Python
+*.egg-info/
+dist/
+build/
+*.egg
+*.whl
+
+# Virtual environments
+.venv/
+venv/
+env/
+.env
+
+# Testing
+.pytest_cache/
+.coverage
+htmlcov/
+*.xml
+
+# Jupyter
+.ipynb_checkpoints/
+*.ipynb_checkpoints
+
+# Claude
+.claude/
+
+# IDE
+.vscode/
+.idea/
+*.swp
+*.swo
+
+# pyenv
+.python-version
+
+# macOS
+.DS_Store
+.AppleDouble
+.LSOverride
+
+# Build / packaging
+*.so
+*.dylib
+MANIFEST
+
+# Results (large binary outputs — not tracked)
+results/*.json
+results/*.log

pare_quant-0.1.0/LICENSE

@@ -0,0 +1,176 @@
+                                 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 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
+      transformations 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, as submitted to the Licensor for inclusion
+      in the Work by You, whether in the original License, and any additions
+      to, modifications of, or is a Contribution.
+
+      "Contributor" shall mean Licensor and any Legal Entity on behalf of
+      whom a Contribution has been received by the Licensor and included
+      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 the 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 your
+          distribution, You must include a readable copy of the
+          attribution notices contained within such NOTICE file, 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 in addition 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 license statement for Your modifications and
+      may provide additional grant of rights to use, copy, modify, merge,
+      publish, distribute, sublicense, and/or sell copies of the
+      Contribution, either on an unrestricted basis or subject to terms
+      and conditions.
+
+   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 reproducing 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 exemplary 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 all other
+      commercial damages or losses), even if such Contributor has been
+      advised of the possibility of such damages.
+
+   9. Accepting Warranty or 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 offer such obligations 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 warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   Copyright 2026 Yasmin Moslem
+
+   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.

pare_quant-0.1.0/PKG-INFO

@@ -0,0 +1,194 @@
+Metadata-Version: 2.4
+Name: pare-quant
+Version: 0.1.0
+Summary: Production-ready quantization for large language and multimodal models
+Author: Yasmin Moslem
+License: Apache-2.0
+License-File: LICENSE
+Keywords: awq,gptq,llm,machine-learning,multimodal,nlp,quantization
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Science/Research
+Classifier: License :: OSI Approved :: Apache Software License
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.11
+Requires-Dist: numpy>=1.24.0
+Requires-Dist: safetensors>=0.4.0
+Requires-Dist: torch>=2.1.0
+Provides-Extra: all
+Requires-Dist: datasets>=2.18.0; extra == 'all'
+Requires-Dist: transformers>=4.40.0; extra == 'all'
+Requires-Dist: triton>=3.0.0; extra == 'all'
+Provides-Extra: calibration
+Requires-Dist: datasets>=2.18.0; extra == 'calibration'
+Provides-Extra: dev
+Requires-Dist: ipykernel>=6.0; extra == 'dev'
+Requires-Dist: jupyter>=1.0; extra == 'dev'
+Requires-Dist: mkdocs-material>=9.5; extra == 'dev'
+Requires-Dist: mkdocstrings[python]>=0.25; extra == 'dev'
+Requires-Dist: mypy>=1.9.0; extra == 'dev'
+Requires-Dist: pytest-cov; extra == 'dev'
+Requires-Dist: pytest>=8.0; extra == 'dev'
+Requires-Dist: ruff>=0.4.0; extra == 'dev'
+Provides-Extra: eval
+Requires-Dist: datasets>=2.18.0; extra == 'eval'
+Provides-Extra: kernels
+Requires-Dist: triton>=3.0.0; extra == 'kernels'
+Provides-Extra: transformers
+Requires-Dist: transformers>=4.40.0; extra == 'transformers'
+Description-Content-Type: text/markdown
+
+# Pare
+
+Quantize any LLM in one line. Switch between GPTQ, AWQ, SmoothQuant, and RTN by changing a config field — same API, same model, same output format.
+
+---
+
+## Pick your trade-off
+
+WikiText-2 perplexity (PPL ↓), A40 46 GB:
+
+| Method | Llama-2-7B | Llama-3-8B | Qwen2.5-7B |
+|--------|-----------|-----------|-----------|
+| FP16 baseline | 5.47 | 6.14 | 6.85 |
+| RTN INT8 | 5.48 (+0.01) | 6.14 (+0.01) | 6.85 (+0.00) |
+| SmoothQuant INT8 | 5.58 (+0.11) | 6.25 (+0.11) | 6.96 (+0.11) |
+| AWQ INT4 | 5.67 (+0.20) | 6.67 (+0.53) | 7.13 (+0.28) |
+| GPTQ INT4 | 5.74 (+0.27) | 8.75 (+2.61) | 7.04 (+0.19) |
+
+Throughput on Llama-2-7B (BS=1): FP16 33 tok/s · RTN/SmoothQuant ~2.3 tok/s · AWQ/GPTQ ~1.2 tok/s †
+
+† Dequantize-on-the-fly. With the optional Triton kernel: **8.8× faster at BS=1, 2.8× at BS=4**.
+
+---
+
+## Quickstart
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from pare import quantize, QuantConfig
+from pare.calibration.data import load_wikitext2_calibration
+
+model     = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf")
+tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
+calib     = load_wikitext2_calibration(tokenizer, n_samples=128, seq_len=2048)
+# or use your own: a list of tokenized tensors of shape (seq_len,)
+
+# Default is AWQ. Change scheme= to switch methods.
+config = QuantConfig(bits=4, scheme="awq", group_size=128)   # ← swap to "gptq", "rtn", "smoothquant"
+model  = quantize(model, config, calibration_data=calib, device="cuda")
+```
+
+Save and reload:
+
+```python
+from pare import save_quantized, load_quantized
+
+save_quantized(model, "llama2-awq-int4/")
+# [pare] Saved 224 quantized layers to llama2-awq-int4  (3821 MB)
+
+from transformers import AutoConfig, AutoModelForCausalLM
+config = AutoConfig.from_pretrained("meta-llama/Llama-2-7b-hf")
+model  = AutoModelForCausalLM.from_config(config)   # architecture only, no weights
+model  = load_quantized(model, "llama2-awq-int4/")
+```
+
+---
+
+## Installation
+
+```bash
+pip install pare-quant                   # core
+pip install "pare-quant[all]"            # + transformers, datasets, Triton kernel
+```
+
+Python ≥ 3.11 · PyTorch ≥ 2.1
+
+---
+
+## Methods
+
+| `scheme=` | Calibration | Quality | When to use |
+|-----------|-------------|---------|-------------|
+| `"awq"` ★ | Yes | ★★★★ | **Default.** Best robustness across architectures; recommended starting point |
+| `"gptq"` | Yes | ★★★★★ | Highest quality when used with `act_order=True`; can underperform AWQ without it |
+| `"smoothquant"` | Yes | ★★★★ | INT8 W+A; closest to FP16 PPL; no INT4 |
+| `"rtn"` | No | ★★★ | No calibration needed; good baseline or for NF4/FP8 |
+
+★ Default: `QuantConfig()` uses AWQ. AWQ consistently outperforms GPTQ (without `act_order=True`) on modern architectures — on Llama-3-8B the gap is 6.67 vs 8.75 PPL. GPTQ with `act_order=True` is the highest-quality option but requires more tuning.
+
+All schemes support `bits=4` or `bits=8`. Use `group_size=128` (default) for best INT4 quality.
+
+### Additional options
+
+**`act_order=True`** — Sort quantization by activation magnitude (improves GPTQ quality on modern architectures):
+```python
+QuantConfig(bits=4, scheme="gptq", group_size=128, act_order=True)
+```
+
+**Mixed-precision** — Automatically promote sensitive layers to higher bits:
+```python
+QuantConfig(bits=4, scheme="awq", sensitive_bits=8, sensitivity_threshold=0.05)
+# [pare] 12 of 224 layers promoted to INT8 based on activation-weighted error
+```
+
+**NF4** — Normal float 4-bit codebook (QLoRA-compatible base model format):
+```python
+from pare.core.dtype import QuantDtype
+QuantConfig(bits=4, dtype=QuantDtype.NF4, scheme="rtn")
+```
+
+**FP8** — 8-bit float for A100/H100:
+```python
+QuantConfig(bits=8, dtype=QuantDtype.FP8_E4M3, scheme="rtn")
+```
+
+---
+
+## Inference speedup (Triton kernel)
+
+The optional Triton INT4 kernel fuses dequantization into the matmul, avoiding materialising the full FP16 weight matrix. Applies to INT4 schemes (AWQ, GPTQ, RTN). Enable per-layer after quantization:
+
+```python
+from pare.layers.linear import QuantizedLinear
+
+for m in model.modules():
+    if isinstance(m, QuantizedLinear):
+        m.use_kernel = True
+```
+
+| Batch size | Without kernel | With kernel | Speedup |
+|------------|---------------|-------------|---------|
+| 1 (decode) | 2.09 ms/layer | 0.24 ms/layer | **8.8×** |
+| 4 | 2.18 ms/layer | 0.78 ms/layer | **2.8×** |
+| 16 | 2.66 ms/layer | 3.18 ms/layer | 0.8× |
+
+Requires `pip install triton>=3.0`.
+
+---
+
+## Hardware
+
+| | Minimum |
+|-|---------|
+| Quantizing a 7B model | 20 GB VRAM (layerwise strategy peaks at ~2 GB) |
+| RTN / GPTQ / AWQ / NF4 | Any CUDA GPU |
+| SmoothQuant W+A | Any CUDA GPU |
+| FP8 | PyTorch ≥ 2.1 (A100 via software; H100 native) |
+| Triton kernel | CUDA GPU + `triton ≥ 3.0` |
+
+---
+
+## Citation
+
+```bibtex
+@misc{moslem2026pare,
+  author = {Moslem, Yasmin},
+  title  = {Pare: Production-ready quantization for large language and multimodal models},
+  year   = {2026},
+  url    = {https://github.com/TinyAdapt/Pare},
+}
+```
+
+Apache 2.0

pare_quant-0.1.0/README.md

@@ -0,0 +1,153 @@
+# Pare
+
+Quantize any LLM in one line. Switch between GPTQ, AWQ, SmoothQuant, and RTN by changing a config field — same API, same model, same output format.
+
+---
+
+## Pick your trade-off
+
+WikiText-2 perplexity (PPL ↓), A40 46 GB:
+
+| Method | Llama-2-7B | Llama-3-8B | Qwen2.5-7B |
+|--------|-----------|-----------|-----------|
+| FP16 baseline | 5.47 | 6.14 | 6.85 |
+| RTN INT8 | 5.48 (+0.01) | 6.14 (+0.01) | 6.85 (+0.00) |
+| SmoothQuant INT8 | 5.58 (+0.11) | 6.25 (+0.11) | 6.96 (+0.11) |
+| AWQ INT4 | 5.67 (+0.20) | 6.67 (+0.53) | 7.13 (+0.28) |
+| GPTQ INT4 | 5.74 (+0.27) | 8.75 (+2.61) | 7.04 (+0.19) |
+
+Throughput on Llama-2-7B (BS=1): FP16 33 tok/s · RTN/SmoothQuant ~2.3 tok/s · AWQ/GPTQ ~1.2 tok/s †
+
+† Dequantize-on-the-fly. With the optional Triton kernel: **8.8× faster at BS=1, 2.8× at BS=4**.
+
+---
+
+## Quickstart
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from pare import quantize, QuantConfig
+from pare.calibration.data import load_wikitext2_calibration
+
+model     = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf")
+tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
+calib     = load_wikitext2_calibration(tokenizer, n_samples=128, seq_len=2048)
+# or use your own: a list of tokenized tensors of shape (seq_len,)
+
+# Default is AWQ. Change scheme= to switch methods.
+config = QuantConfig(bits=4, scheme="awq", group_size=128)   # ← swap to "gptq", "rtn", "smoothquant"
+model  = quantize(model, config, calibration_data=calib, device="cuda")
+```
+
+Save and reload:
+
+```python
+from pare import save_quantized, load_quantized
+
+save_quantized(model, "llama2-awq-int4/")
+# [pare] Saved 224 quantized layers to llama2-awq-int4  (3821 MB)
+
+from transformers import AutoConfig, AutoModelForCausalLM
+config = AutoConfig.from_pretrained("meta-llama/Llama-2-7b-hf")
+model  = AutoModelForCausalLM.from_config(config)   # architecture only, no weights
+model  = load_quantized(model, "llama2-awq-int4/")
+```
+
+---
+
+## Installation
+
+```bash
+pip install pare-quant                   # core
+pip install "pare-quant[all]"            # + transformers, datasets, Triton kernel
+```
+
+Python ≥ 3.11 · PyTorch ≥ 2.1
+
+---
+
+## Methods
+
+| `scheme=` | Calibration | Quality | When to use |
+|-----------|-------------|---------|-------------|
+| `"awq"` ★ | Yes | ★★★★ | **Default.** Best robustness across architectures; recommended starting point |
+| `"gptq"` | Yes | ★★★★★ | Highest quality when used with `act_order=True`; can underperform AWQ without it |
+| `"smoothquant"` | Yes | ★★★★ | INT8 W+A; closest to FP16 PPL; no INT4 |
+| `"rtn"` | No | ★★★ | No calibration needed; good baseline or for NF4/FP8 |
+
+★ Default: `QuantConfig()` uses AWQ. AWQ consistently outperforms GPTQ (without `act_order=True`) on modern architectures — on Llama-3-8B the gap is 6.67 vs 8.75 PPL. GPTQ with `act_order=True` is the highest-quality option but requires more tuning.
+
+All schemes support `bits=4` or `bits=8`. Use `group_size=128` (default) for best INT4 quality.
+
+### Additional options
+
+**`act_order=True`** — Sort quantization by activation magnitude (improves GPTQ quality on modern architectures):
+```python
+QuantConfig(bits=4, scheme="gptq", group_size=128, act_order=True)
+```
+
+**Mixed-precision** — Automatically promote sensitive layers to higher bits:
+```python
+QuantConfig(bits=4, scheme="awq", sensitive_bits=8, sensitivity_threshold=0.05)
+# [pare] 12 of 224 layers promoted to INT8 based on activation-weighted error
+```
+
+**NF4** — Normal float 4-bit codebook (QLoRA-compatible base model format):
+```python
+from pare.core.dtype import QuantDtype
+QuantConfig(bits=4, dtype=QuantDtype.NF4, scheme="rtn")
+```
+
+**FP8** — 8-bit float for A100/H100:
+```python
+QuantConfig(bits=8, dtype=QuantDtype.FP8_E4M3, scheme="rtn")
+```
+
+---
+
+## Inference speedup (Triton kernel)
+
+The optional Triton INT4 kernel fuses dequantization into the matmul, avoiding materialising the full FP16 weight matrix. Applies to INT4 schemes (AWQ, GPTQ, RTN). Enable per-layer after quantization:
+
+```python
+from pare.layers.linear import QuantizedLinear
+
+for m in model.modules():
+    if isinstance(m, QuantizedLinear):
+        m.use_kernel = True
+```
+
+| Batch size | Without kernel | With kernel | Speedup |
+|------------|---------------|-------------|---------|
+| 1 (decode) | 2.09 ms/layer | 0.24 ms/layer | **8.8×** |
+| 4 | 2.18 ms/layer | 0.78 ms/layer | **2.8×** |
+| 16 | 2.66 ms/layer | 3.18 ms/layer | 0.8× |
+
+Requires `pip install triton>=3.0`.
+
+---
+
+## Hardware
+
+| | Minimum |
+|-|---------|
+| Quantizing a 7B model | 20 GB VRAM (layerwise strategy peaks at ~2 GB) |
+| RTN / GPTQ / AWQ / NF4 | Any CUDA GPU |
+| SmoothQuant W+A | Any CUDA GPU |
+| FP8 | PyTorch ≥ 2.1 (A100 via software; H100 native) |
+| Triton kernel | CUDA GPU + `triton ≥ 3.0` |
+
+---
+
+## Citation
+
+```bibtex
+@misc{moslem2026pare,
+  author = {Moslem, Yasmin},
+  title  = {Pare: Production-ready quantization for large language and multimodal models},
+  year   = {2026},
+  url    = {https://github.com/TinyAdapt/Pare},
+}
+```
+
+Apache 2.0