sketchssm 0.1.0__tar.gz

sketchssm-0.1.0/LICENSE

@@ -0,0 +1,202 @@
+
+                                 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.

sketchssm-0.1.0/PKG-INFO

@@ -0,0 +1,181 @@
+Metadata-Version: 2.4
+Name: sketchssm
+Version: 0.1.0
+Summary: SketchSSM decode kernels and offline calibration
+License-Expression: Apache-2.0
+Project-URL: Homepage, https://github.com/SNU-ARC/SketchSSM
+Project-URL: Repository, https://github.com/SNU-ARC/SketchSSM
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: numpy>=1.26
+Requires-Dist: torch>=2.6
+Requires-Dist: ninja
+Provides-Extra: calibration
+Requires-Dist: PyYAML>=6; extra == "calibration"
+Requires-Dist: transformers; extra == "calibration"
+Requires-Dist: datasets; extra == "calibration"
+Requires-Dist: safetensors; extra == "calibration"
+Requires-Dist: huggingface-hub; extra == "calibration"
+Provides-Extra: collection
+Requires-Dist: sketchssm[calibration]; extra == "collection"
+Provides-Extra: test
+Requires-Dist: pytest; extra == "test"
+Dynamic: license-file
+
+# SketchSSM
+
+**SketchSSM: Write to the Full State, Read from a Compact Sketch**
+
+SketchSSM speeds up decoding in linear-attention and state-space layers
+(Mamba-2, Gated DeltaNet, KDA). It keeps updating the full recurrent state, but
+between periodic exact flushes it reads a compact low-rank sketch of the state
+instead of the full state, which cuts state-read traffic. The sketch basis and
+per-head ranks come from a one-time offline calibration of each model.
+
+[Paper](https://arxiv.org/abs/2609.33051)
+
+This repository contains:
+
+- [`sketchssm/kernels/`](https://github.com/SNU-ARC/SketchSSM/blob/main/sketchssm/kernels/README.md): the CUDA decode kernels
+  (`sketchssm.kernels`), which vLLM uses when this package is installed.
+- [`sketchssm/calibration/`](https://github.com/SNU-ARC/SketchSSM/blob/main/sketchssm/calibration/README.md): calibrates a model
+  and packages the result into one portable `calibration.pt`.
+- [`vllm/`](https://github.com/SNU-ARC/SketchSSM/blob/main/vllm/README.md): vLLM v0.30.0 with SketchSSM decode kernels.
+- [`benchmarks/`](https://github.com/SNU-ARC/SketchSSM/blob/main/benchmarks/README.md): per-layer decode latency in vLLM.
+
+## Installation
+
+Python 3.10 or newer with PyTorch. Git LFS is not needed: calibrations for
+serving are downloaded from the Hugging Face Hub when used.
+
+```bash
+git clone https://github.com/SNU-ARC/SketchSSM.git
+cd SketchSSM
+python -m pip install ".[calibration]"                         # kernels + calibration
+(cd vllm && VLLM_USE_PRECOMPILED=1 python -m pip install -e .)  # serving
+```
+
+## Quick start: serve with vLLM
+
+Pass a calibration (a Hugging Face repo id or a local file) and a rank budget,
+the mean sketch rank per state head:
+
+```bash
+vllm serve nvidia/NVIDIA-Nemotron-Nano-9B-v2 --trust-remote-code \
+  --sketchssm ominn/SketchSSM-Nemotron-Nano-9B-v2-BF16 --sketchssm-mean-rank 6 \
+  --mamba-ssm-cache-dtype float32 --no-enable-prefix-caching
+```
+
+The SketchSSM CUDA kernels are compiled just in time the first time a layer
+shape is used, which needs the CUDA toolkit's `nvcc` and `ninja`
+(`pip install ninja`) on `PATH`; without them vLLM uses the portable Triton
+kernels. The per-head ranks and frames for the requested budget are derived when the
+model loads. Provided calibrations:
+
+| Model | Calibration | Collected with weights | Calibrated budgets |
+| --- | --- | --- | --- |
+| Nemotron Nano 9B v2 | [`ominn/SketchSSM-Nemotron-Nano-9B-v2-BF16`](https://huggingface.co/ominn/SketchSSM-Nemotron-Nano-9B-v2-BF16) | BF16 [`nvidia/NVIDIA-Nemotron-Nano-9B-v2`](https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2) | 4, 6, 10, 21 |
+| Nemotron 3 Super | [`ominn/SketchSSM-Nemotron-3-Super-NVFP4`](https://huggingface.co/ominn/SketchSSM-Nemotron-3-Super-NVFP4) | NVFP4 [`nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4`](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4) | 2, 3, 5, 9, 20 |
+| Qwen3.8 Flash-Next | [`ominn/SketchSSM-Qwen3.8-Flash-Next-NVFP4`](https://huggingface.co/ominn/SketchSSM-Qwen3.8-Flash-Next-NVFP4) | NVFP4 [`RadixArk/Qwen3.8-Flash-Next-NVFP4`](https://huggingface.co/RadixArk/Qwen3.8-Flash-Next-NVFP4) | 3, 4, 7, 11, 26 |
+| GLM 5.3 Flash | [`ominn/SketchSSM-GLM-5.3-Flash-NVFP4`](https://huggingface.co/ominn/SketchSSM-GLM-5.3-Flash-NVFP4) | NVFP4 [`RedHatAI/GLM-5.3-Flash-NVFP4`](https://huggingface.co/RedHatAI/GLM-5.3-Flash-NVFP4) | 3, 4, 7, 12, 28 |
+| Qwen3.5 9B | [`ominn/SketchSSM-Qwen3.5-9B-BF16`](https://huggingface.co/ominn/SketchSSM-Qwen3.5-9B-BF16) | BF16 [`Qwen/Qwen3.5-9B`](https://huggingface.co/Qwen/Qwen3.5-9B) | 3, 4, 7, 11, 26 |
+
+A calibration matches the weights it was collected with. Serve it with those
+weights, or make a new calibration for other weights. The calibrated budgets
+were checked against the stored allocation tables; other budgets use the same
+allocation rule. See [`vllm/README.md`](https://github.com/SNU-ARC/SketchSSM/blob/main/vllm/README.md) for all options and
+constraints.
+
+## Make a calibration for your model
+
+Write a configuration for your model (start from a nearby
+`sketchssm/calibration/example/<model>/collect.yaml`, following the
+[new-model guide](https://github.com/SNU-ARC/SketchSSM/blob/main/sketchssm/calibration/docs/new_model.md)), then run the
+calibration stages:
+
+```bash
+python -m sketchssm.calibration calibrate --config my_model.yaml --out outputs/my_model
+```
+
+1. `generate`: generate continuations from WikiText-2 prompts.
+2. `covariance`: replay them and collect state and query covariances.
+3. `basis`: fit one shared sketch basis per head group.
+4. `paired`: score each rank on validation text with output-error and gradient pairs.
+5. `allocate`: assign each head a rank or a dense fallback for every requested budget.
+
+Add `--stage <name>` to run one stage or `--resume` to continue. Package the
+result into a single portable file, then serve it locally or share it on the Hub:
+
+```bash
+python -m sketchssm.calibration package --bundle outputs/my_model --out outputs/my_model/calibration.pt
+
+vllm serve <your-model> --sketchssm outputs/my_model/calibration.pt --sketchssm-mean-rank 6 \
+  --mamba-ssm-cache-dtype float32 --no-enable-prefix-caching
+python -m sketchssm.calibration manifest --bundle outputs/my_model \
+  --calibration outputs/my_model/calibration.pt --precision bf16 --out outputs/my_model/manifest.json
+hf upload <user>/<repo> outputs/my_model/calibration.pt calibration.pt  # then --sketchssm <user>/<repo>
+hf upload <user>/<repo> outputs/my_model/manifest.json manifest.json
+```
+
+`manifest` records the file hash, the base checkpoint and, for every calibrated
+budget, that the tables and frames derived from `calibration.pt` equal the
+bundle's.
+
+The [offline calibration guide](https://github.com/SNU-ARC/SketchSSM/blob/main/sketchssm/calibration/README.md) covers the
+procedure, configuration, file formats and a small CPU-only example.
+
+## Reproduce the provided calibrations
+
+The provided calibrations were collected with the checkpoints and revisions
+pinned in `sketchssm/calibration/example/<model>/config.yaml`. Running
+`calibrate` with that model's `collect.yaml` regenerates them:
+
+```bash
+python -m sketchssm.calibration calibrate \
+  --config sketchssm/calibration/example/nemotron_super/collect.yaml --out outputs/nemotron_super
+python -m sketchssm.calibration package --bundle outputs/nemotron_super \
+  --out outputs/nemotron_super/calibration.pt
+```
+
+Recollection is not needed to try other rank budgets. Each Hub
+`calibration.pt` already contains the sketch basis and the allocation scores,
+so the per-head table and frames for any budget are derived from it directly:
+
+```bash
+hf download ominn/SketchSSM-Nemotron-3-Super-NVFP4 calibration.pt --local-dir outputs/super
+python -m sketchssm.calibration export --calibration outputs/super/calibration.pt \
+  --mean-rank 5 --out outputs/super_g5_frames.pt
+```
+
+See the [calibration data guide](https://github.com/SNU-ARC/SketchSSM/blob/main/sketchssm/calibration/example/README.md) for
+the collection recipe and the expected results of each model.
+
+## Benchmarks
+
+See [`benchmarks/`](https://github.com/SNU-ARC/SketchSSM/blob/main/benchmarks/README.md) to measure the per-layer recurrent
+decode latency of a model in vLLM, and [`vllm/README.md`](https://github.com/SNU-ARC/SketchSSM/blob/main/vllm/README.md) for
+kernel microbenchmarks.
+
+## Citation
+
+If you use SketchSSM in your research, please cite:
+
+```bibtex
+@misc{kwon2026sketchssmwritestateread,
+  title={SketchSSM: Write to the Full State, Read from a Compact Sketch},
+  author={Omin Kwon and JoongWon Shin and Minseo Kim and Kurt Keutzer and Sehoon Kim and Jae W. Lee},
+  year={2026},
+  eprint={2609.33051},
+  archivePrefix={arXiv},
+  primaryClass={cs.LG},
+  url={https://arxiv.org/abs/2609.33051},
+}
+```
+
+## License
+
+SketchSSM is released under the [Apache License 2.0](https://github.com/SNU-ARC/SketchSSM/blob/main/LICENSE). The
+[`vllm/`](https://github.com/SNU-ARC/SketchSSM/blob/main/vllm/) directory is a fork of vLLM, also under Apache-2.0. Calibration
+files are derived from the base models' weights and are also subject to those
+models' licenses.

sketchssm-0.1.0/README.md

@@ -0,0 +1,156 @@
+# SketchSSM
+
+**SketchSSM: Write to the Full State, Read from a Compact Sketch**
+
+SketchSSM speeds up decoding in linear-attention and state-space layers
+(Mamba-2, Gated DeltaNet, KDA). It keeps updating the full recurrent state, but
+between periodic exact flushes it reads a compact low-rank sketch of the state
+instead of the full state, which cuts state-read traffic. The sketch basis and
+per-head ranks come from a one-time offline calibration of each model.
+
+[Paper](https://arxiv.org/abs/2609.33051)
+
+This repository contains:
+
+- [`sketchssm/kernels/`](https://github.com/SNU-ARC/SketchSSM/blob/main/sketchssm/kernels/README.md): the CUDA decode kernels
+  (`sketchssm.kernels`), which vLLM uses when this package is installed.
+- [`sketchssm/calibration/`](https://github.com/SNU-ARC/SketchSSM/blob/main/sketchssm/calibration/README.md): calibrates a model
+  and packages the result into one portable `calibration.pt`.
+- [`vllm/`](https://github.com/SNU-ARC/SketchSSM/blob/main/vllm/README.md): vLLM v0.30.0 with SketchSSM decode kernels.
+- [`benchmarks/`](https://github.com/SNU-ARC/SketchSSM/blob/main/benchmarks/README.md): per-layer decode latency in vLLM.
+
+## Installation
+
+Python 3.10 or newer with PyTorch. Git LFS is not needed: calibrations for
+serving are downloaded from the Hugging Face Hub when used.
+
+```bash
+git clone https://github.com/SNU-ARC/SketchSSM.git
+cd SketchSSM
+python -m pip install ".[calibration]"                         # kernels + calibration
+(cd vllm && VLLM_USE_PRECOMPILED=1 python -m pip install -e .)  # serving
+```
+
+## Quick start: serve with vLLM
+
+Pass a calibration (a Hugging Face repo id or a local file) and a rank budget,
+the mean sketch rank per state head:
+
+```bash
+vllm serve nvidia/NVIDIA-Nemotron-Nano-9B-v2 --trust-remote-code \
+  --sketchssm ominn/SketchSSM-Nemotron-Nano-9B-v2-BF16 --sketchssm-mean-rank 6 \
+  --mamba-ssm-cache-dtype float32 --no-enable-prefix-caching
+```
+
+The SketchSSM CUDA kernels are compiled just in time the first time a layer
+shape is used, which needs the CUDA toolkit's `nvcc` and `ninja`
+(`pip install ninja`) on `PATH`; without them vLLM uses the portable Triton
+kernels. The per-head ranks and frames for the requested budget are derived when the
+model loads. Provided calibrations:
+
+| Model | Calibration | Collected with weights | Calibrated budgets |
+| --- | --- | --- | --- |
+| Nemotron Nano 9B v2 | [`ominn/SketchSSM-Nemotron-Nano-9B-v2-BF16`](https://huggingface.co/ominn/SketchSSM-Nemotron-Nano-9B-v2-BF16) | BF16 [`nvidia/NVIDIA-Nemotron-Nano-9B-v2`](https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2) | 4, 6, 10, 21 |
+| Nemotron 3 Super | [`ominn/SketchSSM-Nemotron-3-Super-NVFP4`](https://huggingface.co/ominn/SketchSSM-Nemotron-3-Super-NVFP4) | NVFP4 [`nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4`](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4) | 2, 3, 5, 9, 20 |
+| Qwen3.8 Flash-Next | [`ominn/SketchSSM-Qwen3.8-Flash-Next-NVFP4`](https://huggingface.co/ominn/SketchSSM-Qwen3.8-Flash-Next-NVFP4) | NVFP4 [`RadixArk/Qwen3.8-Flash-Next-NVFP4`](https://huggingface.co/RadixArk/Qwen3.8-Flash-Next-NVFP4) | 3, 4, 7, 11, 26 |
+| GLM 5.3 Flash | [`ominn/SketchSSM-GLM-5.3-Flash-NVFP4`](https://huggingface.co/ominn/SketchSSM-GLM-5.3-Flash-NVFP4) | NVFP4 [`RedHatAI/GLM-5.3-Flash-NVFP4`](https://huggingface.co/RedHatAI/GLM-5.3-Flash-NVFP4) | 3, 4, 7, 12, 28 |
+| Qwen3.5 9B | [`ominn/SketchSSM-Qwen3.5-9B-BF16`](https://huggingface.co/ominn/SketchSSM-Qwen3.5-9B-BF16) | BF16 [`Qwen/Qwen3.5-9B`](https://huggingface.co/Qwen/Qwen3.5-9B) | 3, 4, 7, 11, 26 |
+
+A calibration matches the weights it was collected with. Serve it with those
+weights, or make a new calibration for other weights. The calibrated budgets
+were checked against the stored allocation tables; other budgets use the same
+allocation rule. See [`vllm/README.md`](https://github.com/SNU-ARC/SketchSSM/blob/main/vllm/README.md) for all options and
+constraints.
+
+## Make a calibration for your model
+
+Write a configuration for your model (start from a nearby
+`sketchssm/calibration/example/<model>/collect.yaml`, following the
+[new-model guide](https://github.com/SNU-ARC/SketchSSM/blob/main/sketchssm/calibration/docs/new_model.md)), then run the
+calibration stages:
+
+```bash
+python -m sketchssm.calibration calibrate --config my_model.yaml --out outputs/my_model
+```
+
+1. `generate`: generate continuations from WikiText-2 prompts.
+2. `covariance`: replay them and collect state and query covariances.
+3. `basis`: fit one shared sketch basis per head group.
+4. `paired`: score each rank on validation text with output-error and gradient pairs.
+5. `allocate`: assign each head a rank or a dense fallback for every requested budget.
+
+Add `--stage <name>` to run one stage or `--resume` to continue. Package the
+result into a single portable file, then serve it locally or share it on the Hub:
+
+```bash
+python -m sketchssm.calibration package --bundle outputs/my_model --out outputs/my_model/calibration.pt
+
+vllm serve <your-model> --sketchssm outputs/my_model/calibration.pt --sketchssm-mean-rank 6 \
+  --mamba-ssm-cache-dtype float32 --no-enable-prefix-caching
+python -m sketchssm.calibration manifest --bundle outputs/my_model \
+  --calibration outputs/my_model/calibration.pt --precision bf16 --out outputs/my_model/manifest.json
+hf upload <user>/<repo> outputs/my_model/calibration.pt calibration.pt  # then --sketchssm <user>/<repo>
+hf upload <user>/<repo> outputs/my_model/manifest.json manifest.json
+```
+
+`manifest` records the file hash, the base checkpoint and, for every calibrated
+budget, that the tables and frames derived from `calibration.pt` equal the
+bundle's.
+
+The [offline calibration guide](https://github.com/SNU-ARC/SketchSSM/blob/main/sketchssm/calibration/README.md) covers the
+procedure, configuration, file formats and a small CPU-only example.
+
+## Reproduce the provided calibrations
+
+The provided calibrations were collected with the checkpoints and revisions
+pinned in `sketchssm/calibration/example/<model>/config.yaml`. Running
+`calibrate` with that model's `collect.yaml` regenerates them:
+
+```bash
+python -m sketchssm.calibration calibrate \
+  --config sketchssm/calibration/example/nemotron_super/collect.yaml --out outputs/nemotron_super
+python -m sketchssm.calibration package --bundle outputs/nemotron_super \
+  --out outputs/nemotron_super/calibration.pt
+```
+
+Recollection is not needed to try other rank budgets. Each Hub
+`calibration.pt` already contains the sketch basis and the allocation scores,
+so the per-head table and frames for any budget are derived from it directly:
+
+```bash
+hf download ominn/SketchSSM-Nemotron-3-Super-NVFP4 calibration.pt --local-dir outputs/super
+python -m sketchssm.calibration export --calibration outputs/super/calibration.pt \
+  --mean-rank 5 --out outputs/super_g5_frames.pt
+```
+
+See the [calibration data guide](https://github.com/SNU-ARC/SketchSSM/blob/main/sketchssm/calibration/example/README.md) for
+the collection recipe and the expected results of each model.
+
+## Benchmarks
+
+See [`benchmarks/`](https://github.com/SNU-ARC/SketchSSM/blob/main/benchmarks/README.md) to measure the per-layer recurrent
+decode latency of a model in vLLM, and [`vllm/README.md`](https://github.com/SNU-ARC/SketchSSM/blob/main/vllm/README.md) for
+kernel microbenchmarks.
+
+## Citation
+
+If you use SketchSSM in your research, please cite:
+
+```bibtex
+@misc{kwon2026sketchssmwritestateread,
+  title={SketchSSM: Write to the Full State, Read from a Compact Sketch},
+  author={Omin Kwon and JoongWon Shin and Minseo Kim and Kurt Keutzer and Sehoon Kim and Jae W. Lee},
+  year={2026},
+  eprint={2609.33051},
+  archivePrefix={arXiv},
+  primaryClass={cs.LG},
+  url={https://arxiv.org/abs/2609.33051},
+}
+```
+
+## License
+
+SketchSSM is released under the [Apache License 2.0](https://github.com/SNU-ARC/SketchSSM/blob/main/LICENSE). The
+[`vllm/`](https://github.com/SNU-ARC/SketchSSM/blob/main/vllm/) directory is a fork of vLLM, also under Apache-2.0. Calibration
+files are derived from the base models' weights and are also subject to those
+models' licenses.

sketchssm-0.1.0/pyproject.toml

@@ -0,0 +1,42 @@
+[build-system]
+requires = ["setuptools>=77"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "sketchssm"
+version = "0.1.0"
+description = "SketchSSM decode kernels and offline calibration"
+readme = "README.md"
+license = "Apache-2.0"
+license-files = ["LICENSE"]
+requires-python = ">=3.10"
+# The kernels are built just-in-time, which also needs nvcc.
+dependencies = ["numpy>=1.26", "torch>=2.6", "ninja"]
+
+[project.urls]
+Homepage = "https://github.com/SNU-ARC/SketchSSM"
+Repository = "https://github.com/SNU-ARC/SketchSSM"
+
+[project.optional-dependencies]
+calibration = ["PyYAML>=6", "transformers", "datasets", "safetensors", "huggingface-hub"]
+# Older name of the calibration extra.
+collection = ["sketchssm[calibration]"]
+test = ["pytest"]
+
+[project.scripts]
+sketchssm-calibrate = "sketchssm.calibration.__main__:main"
+
+[tool.setuptools.packages.find]
+include = ["sketchssm*"]
+namespaces = false
+
+[tool.setuptools.package-data]
+"sketchssm.kernels" = [
+    "csrc/*/*.cuh", "csrc/*/*.cu", "csrc/*/*.cpp", "configs/*/*.json", "docs/*.md",
+    "README.md",
+]
+
+[tool.pytest.ini_options]
+markers = [
+    "slow: long-window sweeps and other expensive kernel cases (deselect with -m 'not slow')",
+]

sketchssm-0.1.0/setup.cfg

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build = 
+tag_date = 0
+

sketchssm-0.1.0/sketchssm/__init__.py

@@ -0,0 +1,9 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the SketchSSM project
+"""SketchSSM: write to the full state, read from a compact sketch.
+
+``sketchssm.kernels`` holds the CUDA decode kernels and
+``sketchssm.calibration`` the offline calibration (basis and rank allocation).
+"""
+
+__version__ = "0.1.0"

sketchssm-0.1.0/sketchssm/calibration/__init__.py

@@ -0,0 +1,2 @@
+"""Group-basis calibration and per-head allocation for SketchSSM."""
+__version__ = "0.1.0"