simulst 0.1.0__tar.gz

simulst-0.1.0/LICENSE

@@ -0,0 +1,201 @@
+                                 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.

simulst-0.1.0/MANIFEST.in

@@ -0,0 +1 @@
+include examples/openai_client.py

simulst-0.1.0/PKG-INFO

@@ -0,0 +1,394 @@
+Metadata-Version: 2.4
+Name: simulst
+Version: 0.1.0
+Summary: Unified speech+LLM model for vLLM: serve the simultaneous-interpretation stack as one model
+Author-email: pokoto <40024614+lvbotenbest@users.noreply.github.com>
+License-Expression: Apache-2.0
+Project-URL: Homepage, https://github.com/lvbotenbest/simulst
+Project-URL: Source, https://github.com/lvbotenbest/simulst
+Keywords: vllm,speech,simultaneous-interpretation,multimodal,asr
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Developers
+Classifier: Programming Language :: Python :: 3
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: <3.15,>=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Provides-Extra: server
+Requires-Dist: vllm[audio]>=0.31; extra == "server"
+Provides-Extra: audio
+Requires-Dist: av; extra == "audio"
+Requires-Dist: scipy; extra == "audio"
+Requires-Dist: soundfile; extra == "audio"
+Requires-Dist: soxr; extra == "audio"
+Requires-Dist: mistral-common[audio]>=1.11.6; extra == "audio"
+Provides-Extra: export
+Requires-Dist: safetensors>=0.6.2; extra == "export"
+Provides-Extra: client
+Requires-Dist: numpy>=1.26; extra == "client"
+Requires-Dist: openai<3,>=2.25; extra == "client"
+Dynamic: license-file
+
+# simulst
+
+Serve a speech simultaneous-interpretation stack — audio encoder, adapter, and
+LLM — as a **single vLLM model**.
+
+```bash
+# Existing compatible vLLM GPU environment: keep its torch/CUDA stack unchanged.
+python -m pip install --no-deps simulst
+
+# Or, in a fresh Python 3.12 environment whose GPU stack pip may manage:
+python -m pip install "simulst[server]"
+
+vllm serve /path/to/unified_checkpoint --served-model-name simulst
+```
+
+No wrapper script, no manual `import`, no second model directory. Installing the
+package is enough: it registers the architecture through vLLM's
+`vllm.general_plugins` entry point, which vLLM calls in the engine process and
+in every tensor-parallel worker before the model is built. One server can accept
+multiple concurrent HTTP requests; vLLM queues and schedules them.
+
+## Install
+
+The plugin supports CPython **3.10, 3.11, 3.12, 3.13, and 3.14**
+(`>=3.10,<3.15`) and requires upstream **vLLM >=0.31**. Python 3.12 is the
+recommended default for a new environment. Choose the installation mode that
+matches the environment:
+
+```bash
+# Existing vLLM >=0.31 GPU environment: do not let pip touch the GPU stack.
+python -m pip install --no-deps simulst
+# Add only HTTP audio decoders if that vLLM was installed without its audio extra.
+python -m pip install "simulst[audio]"
+
+# New environment whose torch/CUDA packages may be managed by pip.
+python -m pip install "simulst[server]"
+
+# Optional checkpoint-export and reference-client tools.
+python -m pip install "simulst[export,client]"
+```
+
+The base package intentionally does not depend on vLLM. The `audio` extra
+installs only the decoder libraries used by vLLM's HTTP `input_audio` path;
+it does not depend on vLLM. The `server` extra installs
+`vllm[audio]>=0.31`, including the same decoders. The missing upper bound avoids
+downgrading a newer vLLM already selected by the operator. It is not a promise
+that every future internal vLLM API will remain compatible. vLLM 0.31
+currently resolves a substantially newer stack (including torch 2.13 and
+Transformers 5) and defaults to CUDA 13 packages, with separate CUDA 12.9
+artifacts available. Do not install the `server` extra into an established GPU
+environment unless replacing that stack is intended.
+
+Verify the install:
+
+```bash
+python -c "
+from importlib.metadata import entry_points
+for g in ('vllm.general_plugins', 'vllm.logits_processors'):
+    print(g, {e.name: e.value for e in entry_points(group=g)})
+"
+```
+
+You should see `simulst_mm` and `simulst_boundary`.
+
+## Environment compatibility and troubleshooting
+
+`simulst` is pure Python, but it integrates with internal vLLM model,
+multimodal-processing, and logits-processing APIs. The supported server stack is
+therefore deliberately narrow:
+
+| component | supported / expected |
+|---|---|
+| Python | `>=3.10,<3.15` |
+| vLLM | `>=0.31` (0.31 adapted; 0.32 key interfaces source-checked) |
+| PyTorch | 2.13.0 for the vLLM 0.31.0 PyPI wheel |
+| Transformers | `>=5.10.4,<5.18` through vLLM |
+| FlashInfer core | `flashinfer-python==0.7.0.post1` through vLLM 0.31.0 |
+| `flashinfer-cubin` | optional; if installed, its version must exactly match FlashInfer |
+| `flashinfer-jit-cache` | optional; its version and CUDA suffix must match FlashInfer |
+| external `flash-attn` | not required by `simulst`; an old build may conflict with PyTorch 2.13 |
+| `deep-ep` | optional; only needed when that expert-parallel backend is selected |
+
+The vLLM 0.31 implementation and the key vLLM 0.32 AsyncLLM, multimodal, and
+Model Runner V2 interfaces have been checked. This is an API compatibility
+check, not a claim that every future `>=0.31` release has passed GPU/checkpoint
+end-to-end validation. Missing required model architecture or API capabilities
+fail at startup and require a `simulst` compatibility update.
+
+The exact torch, Transformers, FlashInfer, and CUDA packages are owned by vLLM.
+They are intentionally not pinned a second time by `simulst`. For an existing
+GPU environment, install this package with `--no-deps`. For a new environment,
+prefer a fresh Python 3.12 virtual environment and install `simulst[server]`.
+
+Inspect the active environment before troubleshooting:
+
+```bash
+python - <<'PY'
+from importlib.metadata import PackageNotFoundError, version
+import torch
+
+for name in (
+    "simulst",
+    "vllm",
+    "torch",
+    "transformers",
+    "flashinfer-python",
+    "flashinfer-cubin",
+    "flashinfer-jit-cache",
+    "flash-attn",
+    "deep-ep",
+):
+    try:
+        print(f"{name}: {version(name)}")
+    except PackageNotFoundError:
+        print(f"{name}: not installed")
+print("torch CUDA:", torch.version.cuda)
+PY
+```
+
+### `undefined symbol` from `flash_attn_2_cuda` or `deep_ep_cpp`
+
+This means a compiled extension was built against a different PyTorch ABI,
+usually because vLLM upgraded torch while an old extension remained installed.
+Neither external `flash-attn` nor `deep-ep` is required for a single-GPU default
+deployment. If those backends are not intentionally used, remove the stale
+extensions and restart vLLM:
+
+```bash
+python -m pip uninstall -y flash-attn deep-ep
+```
+
+If either backend is required, rebuild it against the exact torch and CUDA
+versions in the active environment instead of bypassing the import error.
+
+### FlashInfer package version mismatch
+
+An error such as:
+
+```text
+flashinfer-cubin version (X) does not match flashinfer version (Y)
+```
+
+means an optional precompiled artifact does not match `flashinfer-python`.
+`flashinfer-cubin` and `flashinfer-jit-cache` are not required: FlashInfer can
+compile or download kernels on first use. The simplest repair is:
+
+```bash
+python -m pip uninstall -y flashinfer-cubin flashinfer-jit-cache
+```
+
+If precompiled/offline kernels are required, install matching artifacts from the
+FlashInfer wheel index. All FlashInfer package versions must match exactly, and
+the JIT-cache index must match the torch CUDA build (`cu129` or `cu130`). Use
+`FLASHINFER_DISABLE_VERSION_CHECK=1` only while running FlashInfer's artifact
+installation command to replace a mismatched package; do not keep it in the
+vLLM service environment.
+
+### GPU memory utilization check fails
+
+If vLLM reports that free memory is below its requested utilization, first stop
+other GPU processes or lower the target, for example:
+
+```bash
+vllm serve /path/to/unified_checkpoint \
+  --served-model-name simulst \
+  --gpu-memory-utilization 0.88
+```
+
+### Network interface errors from Gloo or NCCL
+
+Single-node vLLM still initializes distributed communication. If an inherited
+`GLOO_SOCKET_IFNAME` or `NCCL_SOCKET_IFNAME` names a missing interface, select a
+real interface shown by `ip -br addr`. `lo` is suitable only for single-node
+runs; multi-node deployments need a routable interface shared by all nodes.
+
+## Preparing a checkpoint
+
+The package expects one directory containing the LLM weights, the audio tower
+weights under an `audio_tower.` prefix, and a `config.json` whose
+`architectures` is `["HYV3SpeechForCausalLM"]`. If you have the two halves
+exported separately, use an existing PyTorch/Transformers export environment,
+install the safetensors helper, and merge them:
+
+```bash
+python -m pip install "simulst[export]"
+python -m simulst_mm.export.merge_export \
+  --llm    /path/to/hf_export \
+  --audio  /path/to/hf_export_audio \
+  --target /path/to/unified_checkpoint \
+  --encoder-path /path/to/pretrained_audio_encoder   # optional
+```
+
+The LLM shards are **hard-linked**, not copied — tensor names are rewritten at
+load time by a `WeightsMapper`, so there is no reason to rewrite tens of GB on
+disk. A copy is used automatically when the paths are on different filesystems.
+
+`--encoder-path` is only read for `preprocessor_config.json` and the encoder
+shape, and falls back to built-in defaults when absent. Nothing outside the
+merged directory is read at serving time.
+
+The merge also installs an audio-aware `chat_template.jinja` and a
+`generation_config.json` for `/v1/chat/completions`. Pass
+`--no-openai-assets` only if the checkpoint must retain its original template.
+An existing merged checkpoint can be updated in place with:
+
+```bash
+python -m simulst_mm.chat_template --model /path/to/unified_checkpoint
+```
+
+## OpenAI-compatible HTTP deployment and concurrency
+
+Activate the same Python environment that contains vLLM and `simulst`, then
+start one vLLM server with the unified checkpoint:
+
+```bash
+vllm serve /path/to/unified_checkpoint \
+  --served-model-name simulst \
+  --host 0.0.0.0 \
+  --port 8000
+```
+
+That command already supports concurrent HTTP requests. Multiple clients send
+independent requests to `/v1/chat/completions`; the one vLLM engine queues them
+and applies continuous batching. Do not start one model process per request, and
+do not add `AsyncTurnRunner` around the HTTP server.
+
+For example, a two-GPU deployment with capacity for up to eight scheduled
+sequences can be started as follows when those settings match the checkpoint and
+hardware:
+
+```bash
+CUDA_VISIBLE_DEVICES=0,1 vllm serve /path/to/unified_checkpoint \
+  --served-model-name simulst \
+  --tensor-parallel-size 2 \
+  --enable-expert-parallel \
+  --max-num-seqs 8 \
+  --host 0.0.0.0 \
+  --port 8000
+```
+
+`--tensor-parallel-size` and `--enable-expert-parallel` configure model
+parallelism, not HTTP concurrency. `--max-num-seqs` is scheduler capacity, not a
+hard HTTP connection limit; excess requests may wait. Available admission and
+scheduler flags vary by vLLM release, so inspect the installed CLI before adding
+version-specific tuning:
+
+```bash
+vllm serve --help | grep -E "max-num-seqs|max-num-active-seqs"
+```
+
+To run the repository's reference streaming client, use a source checkout
+(`examples/` is not installed by the wheel):
+
+```bash
+python -m pip install "simulst[client]"
+# Refresh an older unified checkpoint once to install the current chat template.
+python -m simulst_mm.chat_template --model /path/to/unified_checkpoint
+# Default: source text plus translation; appends control token 120808.
+python examples/openai_client.py --audio talk.wav --direction en-zh
+# Translation-only model: do not append the control token.
+python examples/openai_client.py --audio talk.wav --direction en-zh \
+  --output-mode translation
+```
+
+The example writes each raw server JSON response to stdout. It is only a
+protocol reference; production conversation and result handling belong in the
+downstream inference service. `/v1/completions` cannot carry audio and is not
+supported for this path.
+
+## In-process usage
+
+Pre-tokenised callers may still provide one expanded run of audio placeholder
+tokens per chunk. The simultaneous-interpretation protocol (turn structure,
+WAIT/EMIT policy, rolling window) remains in the client.
+
+```python
+from vllm import LLM, SamplingParams
+from vllm.inputs import TokensPrompt
+
+llm = LLM(
+    model="/path/to/unified_checkpoint",
+    tensor_parallel_size=2,
+    enable_expert_parallel=True,
+    dtype="bfloat16",
+    enforce_eager=True,
+)
+
+prompt = TokensPrompt(
+    prompt_token_ids=token_ids,                        # placeholders already present
+    multi_modal_data={"audio": [chunk_waveform, ...]},  # one entry per chunk
+)
+llm.generate([prompt], SamplingParams(temperature=0.7, max_tokens=128))
+```
+
+## In-process AsyncLLM concurrency without HTTP
+
+This helper is for applications that create vLLM `AsyncLLM` directly instead
+of running `vllm serve`. HTTP clients do not need it: send concurrent requests
+to the server as described above.
+
+`AsyncTurnRunner` provides a synchronous one-input/one-output facade over one
+shared `AsyncLLM`. Each caller blocks for its own final output, while calls from
+different threads run concurrently and may finish out of submission order:
+
+```python
+from concurrent.futures import ThreadPoolExecutor
+
+from simulst_mm.serving import AsyncTurnRunner
+
+# The factory runs on the runner's private asyncio thread and must return a
+# configured vLLM AsyncLLM instance.
+runner = AsyncTurnRunner(make_async_llm, timeout_seconds=900)
+try:
+    with ThreadPoolExecutor(max_workers=8) as pool:
+        futures = [
+            pool.submit(runner.submit, prompt, sampling_params, owner=session_id)
+            for session_id, prompt in independent_turns
+        ]
+        outputs = [future.result() for future in futures]
+finally:
+    runner.close()
+```
+
+Prompts and audio arrays are treated as immutable for the duration of a request.
+Owner cancellation affects only that owner's active requests. This helper does
+not implement conversation history, HTTP/SSE/WebSocket transport, or
+multi-instance routing; those belong in a higher-level serving application.
+
+## Design notes
+
+**One chunk per encoder call.** Whisper-style feature extraction pads to the
+longest waveform in the batch, so batching changes the padded mel width, which
+in bf16 selects different kernels and shifts embeddings by ~0.03. Encoding one
+chunk at a time is bit-exact across repeated calls (12/12 identical); batching
+is not (0/12). The processor therefore extracts one waveform per call.
+
+**Two placeholder shapes are supported.** The OpenAI chat template emits one
+audio token per chunk and the processor expands it to the required run. Existing
+`TokensPrompt` callers may supply those runs pre-expanded; the processor detects
+that shape and uses vLLM's left-to-right scanner so runs of the same token remain
+associated with the correct chunks.
+
+**Boundary protection travels with the model.** A vLLM 0.31 Model Runner V2
+logits processor masks end-of-turn and EOS for sampling requests unless the
+boundary is already the argmax, so greedy boundary decisions are preserved.
+State is keyed by persistent request slot and expanded through vLLM's row mapping,
+including speculative-decoding rows. The entry point is global, so the processor
+inspects the architecture and disables itself for every other model.
+
+## Verification
+
+Parity against a reference deployment that keeps the tower outside vLLM:
+
+| check | result |
+|---|---|
+| audio embeddings, bit-level | 12/12 identical, max abs diff 0 |
+| generated token ids, 8 turns | identical on every turn |
+| prefix cache, 40 turns | 0.9432 steady (0.977 by turn 40) |
+| boundary processor | 5/5 checks pass |
+
+## License
+
+See `LICENSE`.