glc-loader 1.1.0__tar.gz

glc_loader-1.1.0/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 2026 Alibaba Cloud
+
+   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.

glc_loader-1.1.0/MANIFEST.in

@@ -0,0 +1,18 @@
+# The sdist must carry the package and nothing else.  `release/` also holds
+# built model artifacts (release/public-models-*, release/week1-*, ...) that
+# are gigabytes each and are not part of this distribution.
+include README.md
+include pyproject.toml
+recursive-include glc_loader *.py *.cu *.metal
+recursive-include glc_serve *.py *.cu *.cuh *.cpp *.h *.json
+global-exclude *.pyc
+global-exclude .DS_Store
+prune glc_loader/__pycache__
+prune glc_loader/metal/__pycache__
+prune glc_loader/_tbe_mma_build
+prune glc_serve/__pycache__
+prune public-models-20260719
+prune qwen-certified-20260717
+prune rung3-four-stage-20260719
+prune week1-composed-artifact-20260718
+prune week1-qwen-flagship-20260725

glc_loader-1.1.0/NOTICE

@@ -0,0 +1,4 @@
+GeoRefine codec software
+Copyright 2026 Tsotchke Corporation
+
+This package is licensed under the Apache License, Version 2.0.

glc_loader-1.1.0/PKG-INFO

@@ -0,0 +1,292 @@
+Metadata-Version: 2.1
+Name: glc-loader
+Version: 1.1.0
+Summary: Read, verify and serve GLC-FWP1 and GLC-TBE compressed model artifacts without the repository that built them
+Author: Tsotchke Corporation
+License: Apache-2.0
+Project-URL: Homepage, https://github.com/Tsotchke-Corporation/GeoRefine
+Project-URL: Model, https://huggingface.co/Tsotchke-Corporation/Qwen3.8-27B-GeoRefine-TBE
+Keywords: compression,lossless,transformer,inference,vram
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+License-File: NOTICE
+Requires-Dist: torch>=2.1.0
+Requires-Dist: safetensors>=0.4.0
+Requires-Dist: transformers>=4.56.0
+Provides-Extra: cuda
+Requires-Dist: triton>=2.1.0; extra == "cuda"
+Requires-Dist: ninja>=1.11; extra == "cuda"
+Provides-Extra: metal
+Requires-Dist: mlx>=0.26.0; extra == "metal"
+Requires-Dist: mlx-lm>=0.26.0; extra == "metal"
+Provides-Extra: qwen38
+Requires-Dist: transformers>=5.13.0; extra == "qwen38"
+Requires-Dist: huggingface_hub>=0.30.0; extra == "qwen38"
+Requires-Dist: pillow>=10.0; extra == "qwen38"
+Requires-Dist: torchvision>=0.26.0; extra == "qwen38"
+Provides-Extra: dev
+Requires-Dist: pytest>=7.0.0; extra == "dev"
+
+# glc-loader
+
+Read, verify and serve GLC compressed model artifacts — with no checkout of
+the repository that built them.
+
+The compression is **bit-exact**: the weights this package hands back are the
+same bytes the source checkpoint had, and `glc-loader verify --deep` proves
+that on your machine, from the artifact alone. Runtime speed depends on the
+backend and hardware; see [Hardware reality](#hardware-reality).
+
+---
+
+## Install
+
+```bash
+pip install glc-loader
+```
+
+That is enough to inspect, verify, expand and CPU-load any artifact. Two
+optional extras add accelerated serving; **neither is required to import the
+package**, and neither is a dependency of the other:
+
+```bash
+pip install "glc-loader[cuda]"    # + triton and Ninja for CUDA JIT backends
+pip install "glc-loader[metal]"   # + mlx, mlx-lm, for Apple Silicon
+pip install "glc-loader[qwen38]"  # Qwen3.8 multimodal runtime and processor
+```
+
+Requires Python ≥ 3.10, `torch` ≥ 2.1, `safetensors` ≥ 0.4, `transformers`
+≥ 4.56.
+
+*Tested against* torch 2.6.0, safetensors 0.8.0, transformers 5.13.0 on
+macOS/arm64, Python 3.11. The version floors above are **declared**, not
+measured — they are the lowest releases whose APIs this code uses, not the
+lowest that have been run.
+
+---
+
+## The three commands
+
+### `glc-loader info <artifact-dir>`
+
+What the artifact is, where it came from, its ratios **at all three scopes**,
+and — the part that usually goes missing — what it does **not** certify.
+
+```bash
+glc-loader info ./my-artifact
+```
+
+Ratios come back as three separate blocks, each with its own `basis` string
+and a `status` of `measured` or `unmeasured`. A scope this artifact does not
+carry comes back `null`. It is never filled in from another scope:
+
+| scope | what it counts |
+|---|---|
+| `stored` | bytes on disk |
+| `served_resident` | bytes the device holds for weights |
+| `whole_process` | measured peak process VRAM, dense vs coded — the only one that includes activations, the KV cache and allocator slack |
+
+`whole_process` is normally `unmeasured`. The transcode step leaves it blank
+by construction; it is filled only by a certify run on the target card. If
+`info` tells you it is unmeasured, that is the truth, not a gap in the tool.
+
+### `glc-loader verify <artifact-dir>`
+
+Integrity, with exit codes you can put in CI:
+
+| exit | meaning |
+|---|---|
+| 0 | verified |
+| 1 | an integrity check **failed** — a digest or a decoded tensor did not match |
+| 2 | malformed, unreadable, or not a GeoRefine container at all |
+| 3 | verified bit-exact, but the artifact **expands** (stores more bytes than the dense weights it replaces) |
+| 4 | the format is recognised but this directory **cannot be verified from what it ships** |
+
+```bash
+glc-loader verify ./my-artifact          # digests
+glc-loader verify ./my-artifact --deep   # + decode every container
+```
+
+Without `--deep` this checks digests only, and says so in the output. With
+`--deep` it decodes every stored container and checks the result against the
+source digest recorded before encoding — an end-to-end proof, reproducible on
+your machine, that the bytes this artifact produces are the bytes the source
+model had.
+
+**Exit 4 is deliberate.** A `glc_tbe_transcode` v1 directory is a correct
+artifact with no self-contained verifier: it ships a manifest and two tensor
+blobs, and nothing binds the manifest to the blobs, so checking one against
+the other would be checking the manifest against itself. Reporting that as
+exit 2 ("malformed") would be a lie about a well-formed artifact, and
+reporting it as 0 would be a lie about what was checked.
+
+### `glc-loader load <artifact-dir>`
+
+Smoke-load the artifact, generate a few tokens, print the serving receipt.
+
+```bash
+glc-loader load ./my-artifact --device cuda --prompt "The capital of France is"
+```
+
+This is a smoke test, not a benchmark. The `tokens_per_second` it prints is
+one greedy run on whatever device you gave it.
+
+Two more commands exist for `GLC-RELEASE/1` artifacts specifically:
+`glc-loader expand --out ./dense` writes a plain Hugging Face checkpoint for
+any other engine, and `glc-loader generate --prompt ...` is the older,
+more configurable generation path.
+
+---
+
+## From Python
+
+```python
+from glc_loader import load_model
+
+model, tokenizer, receipt = load_model("./my-artifact", device="cuda")
+print(receipt["backend"])
+```
+
+For a `georefine.tbe.v2` container:
+
+```python
+from glc_loader import load_standalone, single_device_map
+
+model, tokenizer, receipt = load_standalone(
+    "./my-artifact", device_map=single_device_map("cuda:0"),
+)
+```
+
+---
+
+## Hardware reality
+
+### Qwen3.8-27B serve-v1 compatibility prototype
+
+The `georefine.tbe.serve.v1` bundle has a source-free compressed PyTorch path:
+
+```python
+from glc_loader import load_compressed_transformers
+
+model = load_compressed_transformers(
+    "Tsotchke-Corporation/Qwen3.8-27B-GeoRefine-TBE",
+    expect_manifest_sha256="10028416e07c802b47e4dc93a86e9f0dd35ce6021f0e07b96c70bd9cb7a72bc3",
+)
+```
+
+Install `glc-loader[qwen38]` for the required Qwen Transformers class and Hub
+download support. The bundle remains encoded in safetensors arrays (`planes`,
+`smb`, `esc`, `sbbase`). `PortableTBELinear` decodes one weight transiently per
+call, performs a PyTorch linear operation, and discards the dense weight. This
+is a correctness and portability fallback, **not** the measured fused-kernel
+speed path. The base Qwen model's 1,184 tensors map exactly to the manifest;
+the remaining 15 MTP weights are retained for native speculative inference.
+The full 27B graph loaded on Linux CPU and produced finite text and minimal
+image logits while retaining encoded weights. Full generation on this CPU
+reference path is unverified, and its forward speed is not a serving claim.
+
+**Decode speed is a property of the engine, not of the format.** The same
+GLC-TBE weights have been measured on two decode paths:
+
+| decode path | card | measured scope |
+|---|---|---|
+| portable PyTorch reference | Linux CPU | full 27B text/image forwards passed; too slow for serving |
+| standard `glc-serve` HTTP, exact | Blackwell RTX PRO 6000 | 7.99 vs dense 17.50 tok/s end to end; 8/8 text + image parity |
+| standard `glc-serve` HTTP, fused | Blackwell RTX PRO 6000 | 14.95 vs dense 17.50 tok/s; 5/8 text parity |
+| `FastSession` whole-step CUDA graph, **bundled SM120 tune** | Blackwell RTX PRO 6000 | **37.880 vs dense 28.288 decode tok/s (1.339×)**; 8/8 text token-ID + image parity |
+
+The standard HTTP and `FastSession` rows use eight fixed greedy prompts and
+one synthetic image on one card. The `FastSession` result was reproduced from
+an installed wheel and its packaged tune; it is not a cross-platform speed
+guarantee or a general capability certificate.
+
+Earlier standalone-loader measurements were **0.94×** dense on RTX PRO and
+**0.57–0.62×** on A100. Those were single runs with roughly ±0.05 spread on a
+different decode path; they do not measure the packaged `FastSession` engine.
+
+The engine row is Qwen3.8-27B, TBE bundle, batch 1, greedy AR on one card
+(receipts in `KERNEL_SPEED_20260925.md`): **37.84 tok/s**, against 28.29 tok/s
+for the bf16 parent in the same engine and 27.4–27.6 tok/s for llama.cpp bf16
+(tg128) on the same card. Peak VRAM (NVML) was 49.9–50.0 GB, against
+65.4–72.8 GB for the dense engine. The engine's output is bitwise equal to the
+bf16 parent on its gate probes (G3a 711/711 rows).
+
+Memory is the saving that holds on every path. For orientation, one real
+artifact (`unsloth/Llama-3.2-1B-Instruct`, GLC-FWP1) reports a weight ratio of
+**1.279×** at the served-graph scope — 1.93 GB resident where the dense
+checkpoint needs 2.47 GB. Your artifact's own numbers are in
+`glc-loader info`; do not carry that one across.
+
+On the tested RTX PRO 6000, choose `FastSession` for the measured fast path.
+The standard HTTP route and portable CPU path have different speed contracts.
+
+```python
+from huggingface_hub import snapshot_download
+from glc_serve.fastserve import FastSession
+
+bundle = snapshot_download("Tsotchke-Corporation/Qwen3.8-27B-GeoRefine-TBE")
+session = FastSession.load(
+    bundle=bundle, tune=None,  # bundled tune: RTX PRO 6000 Blackwell only
+    server_flags=["--gate", "full", "--manifest-sha256",
+                  "10028416e07c802b47e4dc93a86e9f0dd35ce6021f0e07b96c70bd9cb7a72bc3"],
+)
+state = session.prefill(messages=[{"role": "user", "content": "Hello"}])
+ids = list(session.generate(state, max_tokens=48, temperature=0.0))
+print(session.tokenizer.decode(ids, skip_special_tokens=True))
+```
+
+The public Hugging Face manifest hash above differs from the internal GCS
+source manifest because private build paths were removed. Allow room for the
+40 GB bundle and about 50 GB peak GPU memory on the measured SM120 path. An
+RTX PRO 6000 is the only card whose tune and full-model speed have been
+verified in this wheel; other devices need an explicit tune and separate
+correctness and performance measurements.
+
+---
+
+## CUDA package scope
+
+This wheel includes the TBE MMA source (`tbe_mma_kernel.cu`), `FastDecoder` and MIV-TBE CUDA
+sources, plus the measured RTX PRO 6000 SM120 tune. They JIT-build in a user
+cache; a CUDA toolkit, C++ compiler, Python headers and Ninja are required.
+`FastSession.load` selects the bundled tune only on the measured SM120 card.
+Other GPUs require an explicit tune and their own exactness and speed tests.
+The experimental FWP1 Triton source is still artifact-local (`fwp1_kernels.py`) and is not in
+this wheel; that optional backend refuses or uses its documented fallback.
+
+---
+
+## Which formats this reads
+
+| format | self-contained? | `verify` | `load` |
+|---|---|---|---|
+| `GLC-RELEASE/1` (FWP1) | yes | digests, and `--deep` bit-exactness | yes |
+| `georefine.tbe.v2` | yes | `SHA256SUMS`, and `--deep` bit-exactness | yes, on a supported device |
+| `glc_tbe_transcode` v1 | **no** — no config, no tokenizer, no format tag | exit 4, with the reason | refused |
+| anything else | — | exit 2 | refused |
+
+Format is decided by a **declared tag** — `MANIFEST.json`'s `"format"`, or
+`compression_info.json`'s `"artifact_format"` — never by which files happen to
+be on disk. A plain Hugging Face checkpoint ships `config.json` and
+`model.safetensors`, and a loader that inferred "container" from those would
+claim every model on the hub. Point this at an ordinary checkpoint and it
+declines, by design.
+
+---
+
+## Licence
+
+The compressed weights inherit the **source model's** licence. `glc-loader
+info` reports whatever licence the artifact declares or ships, and reports
+`null` with a note when it declares none — which is the common case today.
+Check the source model before redistributing.
+
+This package is proprietary to Tsotchke Corporation.