flashrt-structures 0.2.0__tar.gz

flashrt_structures-0.2.0/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.

flashrt_structures-0.2.0/PKG-INFO

@@ -0,0 +1,140 @@
+Metadata-Version: 2.4
+Name: flashrt-structures
+Version: 0.2.0
+Summary: Attach FlashRT's verified acceleration structures to an unmodified PyTorch host (lerobot, Isaac-GR00T, openpi, transformers, diffusers, vLLM, SGLang) with no fork and no edit to the host.
+License: Apache-2.0
+Project-URL: Homepage, https://github.com/flashrt-project/FlashRT-Structures
+Project-URL: Repository, https://github.com/flashrt-project/FlashRT-Structures
+Project-URL: Issues, https://github.com/flashrt-project/FlashRT-Structures/issues
+Project-URL: Engine, https://github.com/flashrt-project/FlashRT
+Classifier: Development Status :: 4 - Beta
+Classifier: Environment :: GPU :: NVIDIA CUDA
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: Operating System :: POSIX :: Linux
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: flash-rt>=0.1.0
+Requires-Dist: numpy
+Requires-Dist: pyyaml
+Provides-Extra: hub
+Requires-Dist: kernels>=0.12; extra == "hub"
+Provides-Extra: torch
+Requires-Dist: torch; extra == "torch"
+Requires-Dist: safetensors; extra == "torch"
+Dynamic: license-file
+
+# FlashRT Structures
+
+Attach [FlashRT](https://github.com/flashrt-project/FlashRT)'s verified
+acceleration structures to an **unmodified** PyTorch host: `lerobot`,
+Isaac-GR00T, `openpi`, `transformers`, `diffusers`, and inside vLLM /
+SGLang. No fork, no edit to the host's source. Kernels arrive from the
+Hugging Face kernel hub at bind time.
+
+```python
+import flashrt_structures as structures
+
+plan = structures.attach(model, forward)   # discover → calibrate → gate → activate
+print(structures.explain(plan))            # bound / routed / kept-at-host / refused, with reasons
+
+loop = structures.decode_loop(model, max_len=4096)      # serving door
+out  = loop.generate(input_ids, max_new_tokens=256)
+```
+
+Measured on VLA, VLM, LLM and video models across RTX 5090 and Jetson AGX
+Thor; the films are in the
+[walkthrough](https://huggingface.co/spaces/liangsu9988/fast-kernels-are-not-fast-pipelines).
+
+## What this package is, and what it is not
+
+A *structure* is a versioned specification of one model region: boundary
+tensors, framework-neutral weight slots, calibration points, gates, and a
+plain-torch reference that is the gate's ground truth. Those
+specifications, and the per-host *bindings* that say where the positions
+sit on a concrete host, live in the FlashRT repository as
+`flash_rt.catalog` and ship in the pure-Python `flash-rt` wheel.
+
+This package is everything that acts on them for a PyTorch host:
+
+| layer | here |
+|---|---|
+| `impls/` | executable forms per structure family: FP8 / NVFP4 / weight-only projections and FFNs, fused norm producers, attention cores, fused decoder and vision towers, the whole-step decode loop, graph-lowering pins |
+| `adapters/` | host adapters: `transformers`, `diffusers`, Gemma / Qwen attention, gated-delta, and the vLLM / SGLang engine hooks |
+| `discover.py`, `autobuild.py`, `points.py`, `schemes.py`, `gates.py` | structural discovery, the one-pass discover → calibrate → bind assembly, calibration collection, quantisation schemes, accuracy judgment |
+| `guard.py`, `swap.py`, `frontdoor.py`, `stages.py`, `recipe.py` | the runtime contract and ledger, attach / detach, the one-call door, graph capture with declared swap windows, recipes |
+| `examples/` | the explicit pipeline: seat tables, calibration hooks and binder calls written out by hand for GR00T N1.7 and π0.5 |
+
+Structure specs, references and bindings are **not** here. A new structure
+or a new host binding is a change to `flash_rt/catalog/` in FlashRT; a new
+executable form, adapter or door is a change here.
+
+## Install
+
+```bash
+pip install flashrt-structures            # pulls flash-rt (pure Python) for the catalog
+pip install "flashrt-structures[hub]"     # + the kernel hub client, needed to bind
+```
+
+`flash-rt` must be a version that ships the catalog as `flash_rt.catalog`:
+FlashRT `main` after the structures split, or a wheel newer than 0.1.0
+(the 0.1.0 wheel on PyPI predates the split and still carries the layer
+inside `flash_rt.structures`). An older flash-rt makes `import
+flashrt_structures` fail with a message that says exactly this.
+
+Bring your own torch. Which kernels exist is decided by the torch version:
+the hub's published face is thickest at `torch 2.11 / cu128` on x86-64 and
+much thinner at the newest release. If binds refuse on a fresh install,
+check the torch version first.
+
+Three boundaries produce refusals that look like bugs and are not:
+
+- **Kernel availability follows the hub's build matrix, not this
+  package's.** The wheel installs on any torch; the kernels do not exist
+  for every torch.
+- **`HF_HUB_OFFLINE=1` makes every kernel unavailable, even with a fully
+  warm cache**, because a version specifier has to resolve refs online.
+  Air-gapped deployments should stage packages and point at them with
+  `LOCAL_KERNELS=<repo>=<path>` rather than switching the hub offline.
+- **aarch64 (Jetson Thor, sm_110) is not covered by the current
+  qualification pass.** Nothing in the wheel is architecture-bound, but
+  the engine adapters were last verified against vLLM 0.26 on Thor in
+  August 2026, not in the release run.
+
+## Where to read next
+
+- [`docs/hosts.md`](docs/hosts.md) — which door your host takes, how to
+  tell a seated run from a refused one from a door that never fired, how
+  to read a refusal, and what has been measured where.
+- [`docs/structures.md`](docs/structures.md) — the norm: what a structure
+  is, the three-layer split, calibration reuse, accuracy bands, the
+  runtime contract and ledger, and the norms that came from being wrong.
+- [`docs/serving_engines.md`](docs/serving_engines.md) — attaching inside
+  vLLM or SGLang without forking either.
+- [`docs/adopt_in_20_lines.md`](docs/adopt_in_20_lines.md) — a measured
+  Qwen3-VL-8B adoption, twenty lines end to end.
+- [`examples/README.md`](examples/README.md) — the explicit pipeline
+  against the automatic one, measured on two GR00T N1.7 hosts.
+- [`docs/structure_release_qualification.md`](docs/structure_release_qualification.md)
+  — how a structure moves from a reusable boundary to a released hardware
+  route.
+- [`AGENTS.md`](AGENTS.md) — the operating procedure for producing
+  structures; [`docs/structure_contributing.md`](docs/structure_contributing.md)
+  — the contribution boundary and PR self-review checklist.
+
+## Tests
+
+```bash
+pip install -e . pytest
+pytest tests            # the CPU set; two *_gpu tests need a CUDA device and hub access
+```
+
+## License
+
+Apache-2.0, same as FlashRT.

flashrt_structures-0.2.0/README.md

@@ -0,0 +1,108 @@
+# FlashRT Structures
+
+Attach [FlashRT](https://github.com/flashrt-project/FlashRT)'s verified
+acceleration structures to an **unmodified** PyTorch host: `lerobot`,
+Isaac-GR00T, `openpi`, `transformers`, `diffusers`, and inside vLLM /
+SGLang. No fork, no edit to the host's source. Kernels arrive from the
+Hugging Face kernel hub at bind time.
+
+```python
+import flashrt_structures as structures
+
+plan = structures.attach(model, forward)   # discover → calibrate → gate → activate
+print(structures.explain(plan))            # bound / routed / kept-at-host / refused, with reasons
+
+loop = structures.decode_loop(model, max_len=4096)      # serving door
+out  = loop.generate(input_ids, max_new_tokens=256)
+```
+
+Measured on VLA, VLM, LLM and video models across RTX 5090 and Jetson AGX
+Thor; the films are in the
+[walkthrough](https://huggingface.co/spaces/liangsu9988/fast-kernels-are-not-fast-pipelines).
+
+## What this package is, and what it is not
+
+A *structure* is a versioned specification of one model region: boundary
+tensors, framework-neutral weight slots, calibration points, gates, and a
+plain-torch reference that is the gate's ground truth. Those
+specifications, and the per-host *bindings* that say where the positions
+sit on a concrete host, live in the FlashRT repository as
+`flash_rt.catalog` and ship in the pure-Python `flash-rt` wheel.
+
+This package is everything that acts on them for a PyTorch host:
+
+| layer | here |
+|---|---|
+| `impls/` | executable forms per structure family: FP8 / NVFP4 / weight-only projections and FFNs, fused norm producers, attention cores, fused decoder and vision towers, the whole-step decode loop, graph-lowering pins |
+| `adapters/` | host adapters: `transformers`, `diffusers`, Gemma / Qwen attention, gated-delta, and the vLLM / SGLang engine hooks |
+| `discover.py`, `autobuild.py`, `points.py`, `schemes.py`, `gates.py` | structural discovery, the one-pass discover → calibrate → bind assembly, calibration collection, quantisation schemes, accuracy judgment |
+| `guard.py`, `swap.py`, `frontdoor.py`, `stages.py`, `recipe.py` | the runtime contract and ledger, attach / detach, the one-call door, graph capture with declared swap windows, recipes |
+| `examples/` | the explicit pipeline: seat tables, calibration hooks and binder calls written out by hand for GR00T N1.7 and π0.5 |
+
+Structure specs, references and bindings are **not** here. A new structure
+or a new host binding is a change to `flash_rt/catalog/` in FlashRT; a new
+executable form, adapter or door is a change here.
+
+## Install
+
+```bash
+pip install flashrt-structures            # pulls flash-rt (pure Python) for the catalog
+pip install "flashrt-structures[hub]"     # + the kernel hub client, needed to bind
+```
+
+`flash-rt` must be a version that ships the catalog as `flash_rt.catalog`:
+FlashRT `main` after the structures split, or a wheel newer than 0.1.0
+(the 0.1.0 wheel on PyPI predates the split and still carries the layer
+inside `flash_rt.structures`). An older flash-rt makes `import
+flashrt_structures` fail with a message that says exactly this.
+
+Bring your own torch. Which kernels exist is decided by the torch version:
+the hub's published face is thickest at `torch 2.11 / cu128` on x86-64 and
+much thinner at the newest release. If binds refuse on a fresh install,
+check the torch version first.
+
+Three boundaries produce refusals that look like bugs and are not:
+
+- **Kernel availability follows the hub's build matrix, not this
+  package's.** The wheel installs on any torch; the kernels do not exist
+  for every torch.
+- **`HF_HUB_OFFLINE=1` makes every kernel unavailable, even with a fully
+  warm cache**, because a version specifier has to resolve refs online.
+  Air-gapped deployments should stage packages and point at them with
+  `LOCAL_KERNELS=<repo>=<path>` rather than switching the hub offline.
+- **aarch64 (Jetson Thor, sm_110) is not covered by the current
+  qualification pass.** Nothing in the wheel is architecture-bound, but
+  the engine adapters were last verified against vLLM 0.26 on Thor in
+  August 2026, not in the release run.
+
+## Where to read next
+
+- [`docs/hosts.md`](docs/hosts.md) — which door your host takes, how to
+  tell a seated run from a refused one from a door that never fired, how
+  to read a refusal, and what has been measured where.
+- [`docs/structures.md`](docs/structures.md) — the norm: what a structure
+  is, the three-layer split, calibration reuse, accuracy bands, the
+  runtime contract and ledger, and the norms that came from being wrong.
+- [`docs/serving_engines.md`](docs/serving_engines.md) — attaching inside
+  vLLM or SGLang without forking either.
+- [`docs/adopt_in_20_lines.md`](docs/adopt_in_20_lines.md) — a measured
+  Qwen3-VL-8B adoption, twenty lines end to end.
+- [`examples/README.md`](examples/README.md) — the explicit pipeline
+  against the automatic one, measured on two GR00T N1.7 hosts.
+- [`docs/structure_release_qualification.md`](docs/structure_release_qualification.md)
+  — how a structure moves from a reusable boundary to a released hardware
+  route.
+- [`AGENTS.md`](AGENTS.md) — the operating procedure for producing
+  structures; [`docs/structure_contributing.md`](docs/structure_contributing.md)
+  — the contribution boundary and PR self-review checklist.
+
+## Tests
+
+```bash
+pip install -e . pytest
+pytest tests            # the CPU set; two *_gpu tests need a CUDA device and hub access
+```
+
+## License
+
+Apache-2.0, same as FlashRT.

flashrt_structures-0.2.0/flashrt_structures/__init__.py

@@ -0,0 +1,174 @@
+"""FlashRT structures — verified, composable model sub-blocks.
+
+A structure is a versioned specification of one model region: boundary
+tensors, framework-neutral weight slots, a plain reference implementation
+used as ground truth, and qualification gates. This package hosts the
+structure catalog and its registry. Implementations, host adapters, and
+the qualification harness build on top of these specifications.
+"""
+
+try:
+    from flash_rt.catalog.binding import (
+        BindingSpec,
+        CoverageSegment,
+        list_bindings,
+        load_binding,
+    )
+    from flash_rt.catalog.registry import StructureSpec, list_structures, load
+except ImportError as _e:  # pragma: no cover - depends on the installed flash-rt
+    raise ImportError(
+        "flashrt-structures needs a flash-rt that ships the structure "
+        "catalog as flash_rt.catalog (FlashRT main after the structures "
+        "split; the 0.1.0 wheel on PyPI predates it and still carries the "
+        "layer as flash_rt.structures). Install FlashRT from "
+        "https://github.com/flashrt-project/FlashRT or upgrade flash-rt."
+    ) from _e
+
+
+def get(name):
+    """Explicit door: pull one structure and bind it yourself.
+
+    Mirrors ``kernels.get_kernel``: ``get("decoder_ffn").bind(module,
+    calibration=[...])`` returns a gated drop-in replacement you plug in
+    where you choose. See :mod:`flashrt_structures.handle`.
+    """
+    from flashrt_structures.handle import get as _get
+
+    return _get(name)
+
+
+def capture(fn, **kwargs):
+    """Capture door: graph a hot stage with declared swap windows.
+
+    See :func:`flashrt_structures.stages.capture`.
+    """
+    from flashrt_structures.stages import capture as _capture
+
+    return _capture(fn, **kwargs)
+
+
+def auto_swaps(model, forward, **kwargs):
+    """Distribution layer: discover, calibrate, bind — one pass, no
+    per-seam scaffolding. Returns an :class:`AutoPlan` of swaps.
+
+    See :func:`flashrt_structures.autobuild.auto_swaps`.
+    """
+    from flashrt_structures.autobuild import auto_swaps as _auto
+
+    return _auto(model, forward, **kwargs)
+
+
+def run_recipe(recipe, model, ctx=None, **kwargs):
+    """Recipe door: assemble declared levers, audit same-process on the
+    graph, certify or refuse — one call, one receipt.
+
+    See :mod:`flashrt_structures.recipe` for ``Recipe``/``Lever``/
+    ``Gates`` and the switch lifecycle.
+    """
+    from flashrt_structures.recipe import run_recipe as _run
+
+    return _run(recipe, model, ctx, **kwargs)
+
+
+def attach(model, forward, **kwargs):
+    """One-call front door: discover, calibrate, gate, activate.
+
+    See :func:`flashrt_structures.frontdoor.attach`. Imported lazily so
+    that spec-only consumers do not pay for torch-side machinery.
+    """
+    from flashrt_structures.frontdoor import attach as _attach
+
+    return _attach(model, forward, **kwargs)
+
+
+def adopt_prequantized(model, fmt="ct_nvfp4", **kwargs):
+    """Checkpoint door: adopt an already-quantized checkpoint by
+    converting its packed projections into structure impls.
+
+    See :func:`flashrt_structures.prequantized.adopt_prequantized`.
+    """
+    from flashrt_structures.prequantized import (
+        adopt_prequantized as _adopt)
+
+    return _adopt(model, fmt, **kwargs)
+
+
+def quantize_on_adopt(model, fmt="moe_experts_nvfp4", **kwargs):
+    """Checkpoint door: quantize a full-precision checkpoint that cannot
+    fit the card, converting its dominant structure family (a sparse-MoE
+    expert bank) into structure impls at load time.
+
+    See :func:`flashrt_structures.quantize_on_adopt.quantize_on_adopt`.
+    """
+    from flashrt_structures.quantize_on_adopt import (
+        quantize_on_adopt as _adopt)
+
+    return _adopt(model, fmt, **kwargs)
+
+
+def explain(plan):
+    """Coverage table for one plan: bound / routed / kept / refused,
+    each with its reason. See :mod:`flashrt_structures.explain`."""
+    from flashrt_structures.explain import explain as _explain
+
+    return _explain(plan)
+
+
+def decode_loop(model, *, max_len, compile_step=True,
+                compile_prefill=True, kv_band=None):
+    """Serving door: the whole-loop decode form (static cache + compiled
+    step + whole-step CUDA graph) over whatever structures are attached.
+
+    See :mod:`flashrt_structures.impls.decode_loop.whole_step`.
+    """
+    from flashrt_structures.impls.decode_loop.whole_step import (
+        build_decode_loop)
+
+    return build_decode_loop(model, max_len=max_len,
+                             compile_step=compile_step,
+                             compile_prefill=compile_prefill,
+                             kv_band=kv_band)
+
+
+def aot_package(module, args=(), kwargs=None,
+                package_path="module_aot.pt2", **opts):
+    """Whole-graph door: export the swapped module and AOT-compile it
+    into a reusable package (graph breaks are defects, not fallbacks).
+
+    See :mod:`flashrt_structures.aot`.
+    """
+    from flashrt_structures.aot import aot_package as _pkg
+
+    return _pkg(module, args=args, kwargs=kwargs,
+                package_path=package_path, **opts)
+
+
+def aot_load(package_path, weights=None):
+    """Load an AOT package back as a callable graph."""
+    from flashrt_structures.aot import aot_load as _load
+
+    return _load(package_path, weights=weights)
+
+
+from . import schemes  # noqa: E402  (registry: quantisation schemes)
+
+__all__ = [
+    "BindingSpec",
+    "CoverageSegment",
+    "StructureSpec",
+    "adopt_prequantized",
+    "aot_load",
+    "aot_package",
+    "decode_loop",
+    "explain",
+    "attach",
+    "capture",
+    "get",
+    "list_bindings",
+    "list_structures",
+    "load",
+    "load_binding",
+    "quantize_on_adopt",
+    "run_recipe",
+    "schemes",
+]

flashrt_structures-0.2.0/flashrt_structures/adapters/__init__.py

@@ -0,0 +1,55 @@
+"""Host-family adapters — where a structure's seam is host-specific.
+
+Importing this package registers the built-in adapters with autobuild.
+Attention seams (attention_core) live here because where the attention
+math runs differs by host family; a static module pattern cannot find
+them, so each family gets a small adapter.
+"""
+from ..autobuild import (
+    register_attention_adapter,
+    register_gated_delta_adapter,
+    register_qk_norm_rope_adapter,
+    register_qkv_rope_adapter,
+)
+from .diffusers_attention import DiffusersAttentionAdapter
+from .diffusers_rotary_attention import DiffusersRotaryAttentionAdapter
+from .factored_two_way_attention import FactoredTwoWayAttentionAdapter
+from .factored_qk_norm_rope import FactoredQkNormRopeAdapter
+from .packed_stream_qk_norm_rope import PackedStreamQkNormRopeAdapter
+from .gemma_attention import GemmaAttentionAdapter
+from .transformers_gated_delta import TransformersGatedDeltaAdapter
+from .transformers_gated_delta_fused import (
+    TransformersGatedDeltaFusedAdapter,
+)
+from .qwen_per_head_qk_norm_rope import (
+    PerHeadGqaQkNormRopeAdapter,
+    QwenPerHeadQkNormRopeAdapter,
+)
+from .packed_qkv_rope import PackedQkvRopeAdapter
+
+register_qk_norm_rope_adapter(PerHeadGqaQkNormRopeAdapter())
+register_qk_norm_rope_adapter(FactoredQkNormRopeAdapter())
+register_qk_norm_rope_adapter(PackedStreamQkNormRopeAdapter())
+register_qkv_rope_adapter(PackedQkvRopeAdapter())
+register_attention_adapter(GemmaAttentionAdapter())
+register_attention_adapter(FactoredTwoWayAttentionAdapter())
+register_attention_adapter(DiffusersRotaryAttentionAdapter())
+register_attention_adapter(DiffusersAttentionAdapter())
+# the fused-layer form is tried first; it refuses cleanly (missing
+# package entries, out-of-profile layers) and the ladder falls
+# through to the callable-slot form
+register_gated_delta_adapter(TransformersGatedDeltaFusedAdapter())
+register_gated_delta_adapter(TransformersGatedDeltaAdapter())
+
+__all__ = [
+    "DiffusersAttentionAdapter",
+    "DiffusersRotaryAttentionAdapter",
+    "GemmaAttentionAdapter",
+    "TransformersGatedDeltaAdapter",
+    "FactoredTwoWayAttentionAdapter",
+    "FactoredQkNormRopeAdapter",
+    "PackedStreamQkNormRopeAdapter",
+    "PerHeadGqaQkNormRopeAdapter",
+    "QwenPerHeadQkNormRopeAdapter",
+    "PackedQkvRopeAdapter",
+]