Add files using upload-large-folder tool

.dockerignore

@@ -0,0 +1,8 @@
+.build
+.git
+.venv
+runtime-data
+model-*.safetensors
+codec.pth
+activation-scaling-report.json
+tokenizer.json

.gitattributes

@@ -1,35 +1,5 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
 *.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.wav filter=lfs diff=lfs merge=lfs -text
+*.flac filter=lfs diff=lfs merge=lfs -text
+*.mp3 filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,5 @@
+/.build/
+/.venv/
+/runtime-data/*
+!/runtime-data/references/
+!/runtime-data/references/.gitkeep

.source

@@ -0,0 +1,4 @@
+base_model=fishaudio/s2-pro@1de9996b6be38b745688de084d87a5633f714e4e
+fish_speech=e5e292632cb11e7a27b2b7487f58f612bc101e13
+fish_scales_ops=a04c1b63b1a7a670840fb3e97a82c0dbe2a35ded
+nvfp4_native_source=7a03467b90d6feff6bd196928dfe156bd173f36e

Dockerfile

@@ -0,0 +1,61 @@
+# syntax=docker/dockerfile:1.7
+
+ARG CUDA_IMAGE=nvidia/cuda:13.0.2-cudnn-devel-ubuntu24.04
+ARG UV_VERSION=0.12.3
+FROM ghcr.io/astral-sh/uv:${UV_VERSION} AS uv-bin
+FROM ${CUDA_IMAGE}
+
+ARG FISH_SCALES_REVISION=a04c1b63b1a7a670840fb3e97a82c0dbe2a35ded
+ENV DEBIAN_FRONTEND=noninteractive \
+    PYTHONDONTWRITEBYTECODE=1 \
+    PYTHONUNBUFFERED=1 \
+    VIRTUAL_ENV=/opt/venv \
+    PATH=/opt/venv/bin:/usr/local/cuda/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin \
+    CUDA_HOME=/usr/local/cuda \
+    TORCH_CUDA_ARCH_LIST=12.0a \
+    FISH_NVFP4_ROOT=/model \
+    FISH_NVFP4_BUILD_ROOT=/data/torch-extensions \
+    PYTHONPATH=/model/runtime \
+    TTS_PYTHON=/opt/venv/bin/python \
+    TEMPORARILY_DISABLE_PROTOBUF_VERSION_CHECK=true
+
+RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
+    --mount=type=cache,target=/var/lib/apt/lists,sharing=locked \
+    apt-get update && apt-get install -y --no-install-recommends \
+      build-essential ca-certificates cmake ffmpeg git git-lfs libsndfile1 \
+      ninja-build portaudio19-dev python3.12 python3.12-dev python3.12-venv && \
+    git lfs install --system
+
+COPY --from=uv-bin /uv /uvx /usr/local/bin/
+COPY runtime/torch-overrides.txt /opt/release/torch-overrides.txt
+COPY requirements.txt /opt/release/requirements.txt
+COPY vendor/fish-speech /opt/fish-speech
+
+RUN uv venv /opt/venv --python python3.12 && \
+    uv pip install --python /opt/venv/bin/python \
+      "setuptools>=75" wheel ninja packaging \
+      "torch==2.11.0" "torchaudio==2.11.0" --torch-backend=cu130 && \
+    uv pip install --python /opt/venv/bin/python \
+      --overrides /opt/release/torch-overrides.txt --torch-backend=cu130 \
+      /opt/fish-speech -r /opt/release/requirements.txt
+
+RUN git init /opt/fish-scales-ops && \
+    git -C /opt/fish-scales-ops remote add origin https://github.com/fishaudio/fish-scales-ops.git && \
+    git -C /opt/fish-scales-ops fetch --depth 1 origin "${FISH_SCALES_REVISION}" && \
+    git -C /opt/fish-scales-ops checkout --detach FETCH_HEAD && \
+    git -C /opt/fish-scales-ops submodule update --init --recursive --depth 1
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    CUTLASS_DIR=/opt/fish-scales-ops/3rdparty/cutlass \
+    MAX_JOBS=2 TORCH_CUDA_ARCH_LIST=12.0a \
+    uv pip install --python /opt/venv/bin/python \
+      --no-build-isolation /opt/fish-scales-ops/python
+
+RUN install -m 0644 /opt/fish-speech/.project-root \
+      /opt/venv/lib/python3.12/site-packages/.project-root && \
+    /opt/venv/bin/python -c \
+    "from pathlib import Path; import comfy_kitchen, fish_speech, torch, tools.server.model_manager; from fish_speech.models.dac.modded_dac import DAC; site_packages = Path('/opt/venv/lib/python3.12/site-packages'); assert (site_packages / 'fish_speech/configs/modded_dac_vq.yaml').is_file(); assert list((site_packages / 'fish_scales_ops').glob('_C*.so')); print(torch.__version__, torch.version.cuda)"
+
+WORKDIR /data
+EXPOSE 8080
+ENTRYPOINT ["/bin/bash", "/model/launch.sh"]

LICENSE.md

@@ -0,0 +1,94 @@
+# FISH AUDIO RESEARCH LICENSE AGREEMENT
+
+**Last Updated: March 7, 2026**
+
+## I. INTRODUCTION
+
+This Agreement applies to any individual person or entity ("You", "Your" or "Licensee") that uses or distributes any portion or element of the Fish Audio Materials or Derivative Works thereof for any Research, Non-Commercial, or Commercial purpose. Capitalized terms not otherwise defined herein are defined in Section V below.
+
+This Agreement is intended to allow research and non-commercial uses of the Materials free of charge. Any Commercial use of the Materials requires a separate license from Fish Audio.
+
+By clicking "I Accept" or by using, distributing, or accessing any portion or element of the Fish Audio Materials or Derivative Works, You agree that You have read, understood and are bound by the terms of this Agreement. If You are acting on behalf of a company, organization or other entity, then "You" includes you and that entity, and You agree that You: (i) are an authorized representative of such entity with the authority to bind such entity to this Agreement, and (ii) You agree to the terms of this Agreement on that entity's behalf.
+
+## II. RESEARCH & NON-COMMERCIAL USE LICENSE
+
+Subject to the terms of this Agreement, Fish Audio grants You a non-exclusive, worldwide, non-transferable, non-sublicensable, revocable and royalty-free limited license under Fish Audio's intellectual property or other rights owned by Fish Audio embodied in the Fish Audio Materials to use, reproduce, distribute, and create Derivative Works of, and make modifications to, the Fish Audio Materials for any Research or Non-Commercial Purpose.
+
+"Research Purpose" means academic or scientific advancement, and in each case, is not primarily intended for commercial advantage or monetary compensation to You or others.
+
+"Non-Commercial Purpose" means any purpose other than a Research Purpose that is not primarily intended for commercial advantage or monetary compensation to You or others, such as personal use (i.e., hobbyist) or evaluation and testing.
+
+## III. COMMERCIAL USE
+
+**Any use of the Fish Audio Materials or Derivative Works for a Commercial Purpose requires a separate written license agreement from Fish Audio.** No commercial rights are granted under this Agreement.
+
+"Commercial Purpose" means any purpose other than a Research Purpose or Non-Commercial Purpose that is primarily intended for or directed toward commercial advantage or monetary compensation to You or others, including but not limited to: (i) creating, modifying, or distributing Your product or service, including via a hosted service or application programming interface, (ii) Your business's or organization's internal operations, and (iii) any use in connection with a product or service for which You charge a fee or generate revenue, whether directly or indirectly.
+
+To obtain a commercial license, please contact Fish Audio at:
+
+- **Website:** [https://fish.audio](https://fish.audio)
+- **Email:** business@fish.audio
+
+## IV. GENERAL TERMS
+
+Your Research and Non-Commercial License under this Agreement is subject to the following terms.
+
+### a. Distribution & Attribution
+
+If You distribute or make available the Fish Audio Materials or a Derivative Work to a third party, or a product or service that uses any portion of them, You shall: (i) provide a copy of this Agreement to that third party, (ii) retain the following attribution notice within a "Notice" text file distributed as a part of such copies: "This model is licensed under the Fish Audio Research License, Copyright © 39 AI, INC. All Rights Reserved.", and (iii) prominently display "Built with Fish Audio" on a related website, user interface, blogpost, about page, or product documentation.
+
+If You create a Derivative Work, You may add your own attribution notice(s) to the "Notice" text file included with that Derivative Work, provided that You clearly indicate which attributions apply to the Fish Audio Materials and state in the "Notice" text file that You changed the Fish Audio Materials and how it was modified.
+
+### b. Use Restrictions
+
+Your use of the Fish Audio Materials and Derivative Works, including any output or results of the Fish Audio Materials or Derivative Works, must comply with applicable laws and regulations (including Trade Control Laws and equivalent regulations) and adhere to Fish Audio's Acceptable Use Policy, which is hereby incorporated by reference.
+
+Furthermore, You will not use the Fish Audio Materials or Derivative Works, or any output or results of the Fish Audio Materials or Derivative Works, to create or improve any foundational generative AI model (excluding the Models or Derivative Works).
+
+### c. Intellectual Property
+
+**(i) Trademark License.** No trademark licenses are granted under this Agreement, and in connection with the Fish Audio Materials or Derivative Works, You may not use any name or mark owned by or associated with Fish Audio or any of its Affiliates, except as required under Section IV(a) herein.
+
+**(ii) Ownership of Derivative Works.** As between You and Fish Audio, You are the owner of Derivative Works You create, subject to Fish Audio's ownership of the Fish Audio Materials and any Derivative Works made by or for Fish Audio.
+
+**(iii) Ownership of Outputs.** As between You and Fish Audio, You own any outputs generated from the Models or Derivative Works to the extent permitted by applicable law.
+
+**(iv) Disputes.** If You or Your Affiliate(s) institute litigation or other proceedings against Fish Audio (including a cross-claim or counterclaim in a lawsuit) alleging that the Fish Audio Materials, Derivative Works or associated outputs or results, or any portion of any of the foregoing, constitutes infringement of intellectual property or other rights owned or licensable by You, then any licenses granted to You under this Agreement shall terminate as of the date such litigation or claim is filed or instituted. You will indemnify and hold harmless Fish Audio from and against any claim by any third party arising out of or related to Your use or distribution of the Fish Audio Materials or Derivative Works in violation of this Agreement.
+
+**(v) Feedback.** From time to time, You may provide Fish Audio with verbal and/or written suggestions, comments or other feedback related to Fish Audio's existing or prospective technology, products or services (collectively, "Feedback"). You are not obligated to provide Fish Audio with Feedback, but to the extent that You do, You hereby grant Fish Audio a perpetual, irrevocable, royalty-free, fully-paid, sub-licensable, transferable, non-exclusive, worldwide right and license to exploit the Feedback in any manner without restriction. Your Feedback is provided "AS IS" and You make no warranties whatsoever about any Feedback.
+
+### d. Disclaimer of Warranty
+
+UNLESS REQUIRED BY APPLICABLE LAW, THE FISH AUDIO MATERIALS AND ANY OUTPUT AND RESULTS THEREFROM ARE PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED, INCLUDING, WITHOUT LIMITATION, ANY WARRANTIES OF TITLE, NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. YOU ARE SOLELY RESPONSIBLE FOR DETERMINING THE APPROPRIATENESS OR LAWFULNESS OF USING OR REDISTRIBUTING THE FISH AUDIO MATERIALS, DERIVATIVE WORKS OR ANY OUTPUT OR RESULTS AND ASSUME ANY RISKS ASSOCIATED WITH YOUR USE OF THE FISH AUDIO MATERIALS, DERIVATIVE WORKS AND ANY OUTPUT AND RESULTS.
+
+### e. Limitation of Liability
+
+IN NO EVENT WILL FISH AUDIO OR ITS AFFILIATES BE LIABLE UNDER ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, TORT, NEGLIGENCE, PRODUCTS LIABILITY, OR OTHERWISE, ARISING OUT OF THIS AGREEMENT, FOR ANY LOST PROFITS OR ANY DIRECT, INDIRECT, SPECIAL, CONSEQUENTIAL, INCIDENTAL, EXEMPLARY OR PUNITIVE DAMAGES, EVEN IF FISH AUDIO OR ITS AFFILIATES HAVE BEEN ADVISED OF THE POSSIBILITY OF ANY OF THE FOREGOING.
+
+### f. Term and Termination
+
+The term of this Agreement will commence upon Your acceptance of this Agreement or access to the Fish Audio Materials and will continue in full force and effect until terminated in accordance with the terms and conditions herein. Fish Audio may terminate this Agreement if You are in breach of any term or condition of this Agreement. Upon termination of this Agreement, You shall delete and cease use of any Fish Audio Materials or Derivative Works. Sections IV(d), (e), and (g) shall survive the termination of this Agreement.
+
+### g. Governing Law
+
+This Agreement will be governed by and construed in accordance with the laws of the United States and the State of California without regard to choice of law principles, and the UN Convention on Contracts for International Sale of Goods does not apply to this Agreement.
+
+## V. DEFINITIONS
+
+**"Affiliate(s)"** means any entity that directly or indirectly controls, is controlled by, or is under common control with the subject entity; for purposes of this definition, "control" means direct or indirect ownership or control of more than 50% of the voting interests of the subject entity.
+
+**"Agreement"** means this Fish Audio Research License Agreement.
+
+**"Derivative Work(s)"** means (a) any derivative work of the Fish Audio Materials as recognized by U.S. copyright laws and (b) any modifications to a Model, and any other model created which is based on or derived from the Model or the Model's output, including "fine tune" and "low-rank adaptation" models derived from a Model or a Model's output, but do not include the output of any Model.
+
+**"Documentation"** means any specifications, manuals, documentation, and other written information provided by Fish Audio related to the Software or Models.
+
+**"Fish Audio"** or **"we"** means 39 AI, INC. and its Affiliates.
+
+**"Model(s)"** means, collectively, Fish Audio's proprietary models and algorithms, including machine-learning models, trained model weights and other elements of the foregoing.
+
+**"Software"** means Fish Audio's proprietary software made available under this Agreement now or in the future.
+
+**"Fish Audio Materials"** means, collectively, Fish Audio's proprietary Models, Software and Documentation (and any portion or combination thereof) made available under this Agreement.
+
+**"Trade Control Laws"** means any applicable U.S. and non-U.S. export control and trade sanctions laws and regulations.

MANIFEST.json

@@ -0,0 +1,1134 @@
+{
+  "schema_version": 1,
+  "algorithm": "sha256",
+  "excluded": [
+    "MANIFEST.json"
+  ],
+  "files": [
+    {
+      "path": ".dockerignore",
+      "size": 107,
+      "sha256": "8846b754cd333873df23ed2c6447c3c26657f21008bf901866561d1a58587cda"
+    },
+    {
+      "path": ".gitattributes",
+      "size": 219,
+      "sha256": "ad359bc56ca7ad35381312a9987f07c468b574209288305d5d3df127c8ac6e8b"
+    },
+    {
+      "path": ".gitignore",
+      "size": 95,
+      "sha256": "d9c2f19352ae8ea0dd69c2e66c46a3ecae6dad27c3e0fc043295ca7acf9753a9"
+    },
+    {
+      "path": ".source",
+      "size": 240,
+      "sha256": "e5d3907e41ff60c4e50108a4fcc4b64de346d5a02aa45819211f0f364bfd8799"
+    },
+    {
+      "path": "Dockerfile",
+      "size": 2981,
+      "sha256": "35a9862345533b12de2bbc90f7933e8558da522ac9c072c51e2cc2c1a55eb67b"
+    },
+    {
+      "path": "LICENSE.md",
+      "size": 10360,
+      "sha256": "aa7d9206e9d710590987a3636934f643529c00cd490323594e6206aaa0c32d80"
+    },
+    {
+      "path": "Notice",
+      "size": 994,
+      "sha256": "49f2d4dcecbb95e10aaba411e6237fadbe0fd1aeccc863adef3c2a7ff8b1adfe"
+    },
+    {
+      "path": "README.md",
+      "size": 10945,
+      "sha256": "c010e235b428b3a45f118df4e2eb9053278fbc85a5543deb542a60f5e239f18e"
+    },
+    {
+      "path": "THIRD_PARTY_NOTICES.md",
+      "size": 1039,
+      "sha256": "fe4c78dfee8f6ccf99bed12fbf918ce9ca4a24d2f80e6279cbccec1959ce0def"
+    },
+    {
+      "path": "activation-scaling-report.json",
+      "size": 2657835,
+      "sha256": "20024a68428ee44e4aa1b8101cc6357310d7670ebdd2ab87e91eb4f8f54d43ef"
+    },
+    {
+      "path": "chat_template.jinja",
+      "size": 4116,
+      "sha256": "87a2728cb8dc9fe424d624542f6060ec05a1d285ebbec578bb078900e33396b5"
+    },
+    {
+      "path": "client.py",
+      "size": 3640,
+      "sha256": "45a7831d9e8482728bcad99065c9deb2d66c0133d625eb830313cc9cd8aa83db"
+    },
+    {
+      "path": "codec.pth",
+      "size": 1871099728,
+      "sha256": "74fc41c5a7151c6f350af8bd7e5d6e3accfcc7f3dfbfac23afd35af07052bb2f"
+    },
+    {
+      "path": "compose.yaml",
+      "size": 612,
+      "sha256": "98f2090a2588f19628fdacefd0efb50d6b272d7f0f0933e8f201b3bed54efd51"
+    },
+    {
+      "path": "config.json",
+      "size": 2139,
+      "sha256": "bd1b81bf140b1eac0a15fe75319f5552902ad3b9903d557ddada358aef743125"
+    },
+    {
+      "path": "install.sh",
+      "size": 2521,
+      "sha256": "0eb32f6024fefbd3c3bf5130aecd0764dec9018656519d3532d6148405d451cc"
+    },
+    {
+      "path": "launch.sh",
+      "size": 1311,
+      "sha256": "83fd98f286bdf924247bd217da125169c1f6527c8375d1011b595112336c6d9c"
+    },
+    {
+      "path": "model-00001-of-00003.safetensors",
+      "size": 1992555752,
+      "sha256": "b304711a5a7a102dc27a49616048e72aca0c695ef4648e860a4b9b5f2c39d694"
+    },
+    {
+      "path": "model-00002-of-00003.safetensors",
+      "size": 1984565944,
+      "sha256": "ab75ab15859d88276b88d9c4672f196489a71336fb062335f7833da8e35e277b"
+    },
+    {
+      "path": "model-00003-of-00003.safetensors",
+      "size": 926458160,
+      "sha256": "0187275d6a87c4879b24b54462d8b3f487318ab402213f27bc9cca017b7d3816"
+    },
+    {
+      "path": "model.safetensors.index.json",
+      "size": 48027,
+      "sha256": "25229032b7d5a846fb2f946c46a35229f79babc569dad1dbc577705d892b2fba"
+    },
+    {
+      "path": "quantization.json",
+      "size": 70173,
+      "sha256": "9dc6d4d1bb010db508d9d5a0a9b520ff10c19e8e5c7f1d1eeb7ee9d4ee1e465b"
+    },
+    {
+      "path": "requirements.txt",
+      "size": 150,
+      "sha256": "b91057af47321ec40b3fb5137ad769a0510ea879054a7cc02ffb8a4c552cb477"
+    },
+    {
+      "path": "runtime/experimental/__init__.py",
+      "size": 76,
+      "sha256": "e2111b3a6982e6f1a84fe32ea7095a6670909abc4ae1b3c68626fc55db636d8e"
+    },
+    {
+      "path": "runtime/experimental/codec.py",
+      "size": 11733,
+      "sha256": "5bf0c02babbfefe1a736f3e844f6a46aac18d9653ba7444fd324d19abaf9dbfe"
+    },
+    {
+      "path": "runtime/experimental/fp8/__init__.py",
+      "size": 303,
+      "sha256": "752f7212156ef486d5015f55fc0098e6240260e07b08417f8f32db71c6899a4e"
+    },
+    {
+      "path": "runtime/experimental/fp8/modules.py",
+      "size": 10873,
+      "sha256": "3c4b2f71d0c3f078ad5f295c6ada96a5eb4c9c63ea518177985bdd6112766c5a"
+    },
+    {
+      "path": "runtime/experimental/fp8/server.py",
+      "size": 3551,
+      "sha256": "e421337cdddf0b3fe4c835fac47c97d56ea977a13c0455aec19eadf6aef00aa6"
+    },
+    {
+      "path": "runtime/experimental/nvfp4/__init__.py",
+      "size": 910,
+      "sha256": "267c7f107de1c3038459f1c20d72c3a92e56da01009a766bd178e7dfc4d0087d"
+    },
+    {
+      "path": "runtime/experimental/nvfp4/checkpoint.py",
+      "size": 7342,
+      "sha256": "41870d3effbcf4e5647ba62e107bbacf0db172e963b0cf14a9503349858b7114"
+    },
+    {
+      "path": "runtime/experimental/nvfp4/modules.py",
+      "size": 70345,
+      "sha256": "275be371fed819985805cf496735c32e679779fe1ccf0096f26e948bd2a668f7"
+    },
+    {
+      "path": "runtime/experimental/nvfp4/server.py",
+      "size": 3787,
+      "sha256": "bf84ec5ff00916e88b504eaef2cba5d0175322bb2a19b446a43e94ea7bf89ac1"
+    },
+    {
+      "path": "runtime/native/LICENSE",
+      "size": 11364,
+      "sha256": "cd1c9ea1573c9047e9b5a3676ecc1c8f755e56f249ac1ebd8b5ceda7899037da"
+    },
+    {
+      "path": "runtime/native/direct_w4a4_m1/direct_w4a4_m1.cpp",
+      "size": 4196,
+      "sha256": "47205d814c92081d85dd70161706efb10ce46fa7ae0da93fe43c329376ffb649"
+    },
+    {
+      "path": "runtime/native/direct_w4a4_m1/direct_w4a4_m1.cu",
+      "size": 5609,
+      "sha256": "23dd124a573bae75efdb9e676b8bab8b23400e6bc59f684caa17e449d8592781"
+    },
+    {
+      "path": "runtime/native/direct_w4a4_m1/direct_w4a4_m1.h",
+      "size": 428,
+      "sha256": "80e09accc0eff8ff12ea72e355b0f4872824fa05cb70a4a57f35fb187128d8fd"
+    },
+    {
+      "path": "runtime/native/rmsnorm_nvfp4_m1/rmsnorm_nvfp4_m1.cpp",
+      "size": 1230,
+      "sha256": "ab6126591d36c783ca08bb5ce0a0e0564a2085d3927631406e9b41e73197d242"
+    },
+    {
+      "path": "runtime/native/rmsnorm_nvfp4_m1/rmsnorm_nvfp4_m1.cu",
+      "size": 5848,
+      "sha256": "26f219fe25c6a51904e8fd11c1a05f762b20518042a095b0745e55e80f743a47"
+    },
+    {
+      "path": "runtime/native/rmsnorm_nvfp4_m1/rmsnorm_nvfp4_m1.h",
+      "size": 199,
+      "sha256": "eb50d22a5765fb178e48d87d3bad214c41097502446a2002094ffbd11e329676"
+    },
+    {
+      "path": "runtime/native/silu_product_nvfp4_m1/silu_product_nvfp4_m1.cpp",
+      "size": 3572,
+      "sha256": "eadd24b46567ea2a605b260552f07157e6a09901b1313afa5922f877781a4c2f"
+    },
+    {
+      "path": "runtime/native/silu_product_nvfp4_m1/silu_product_nvfp4_m1.cu",
+      "size": 10912,
+      "sha256": "ca9bd7e7a6251d6eb729333167305c60b184d0732f50294125c8b42ac424f9e4"
+    },
+    {
+      "path": "runtime/native/silu_product_nvfp4_m1/silu_product_nvfp4_m1.h",
+      "size": 507,
+      "sha256": "968d4dbb617bc9d650a739a07a025f0e8c577e52c29ab41f024855304724b813"
+    },
+    {
+      "path": "runtime/native/smallm_gemv/smallm_gemv.cpp",
+      "size": 2970,
+      "sha256": "ae1bf8fa6400c55c9f148bcc6939ce39c6d7fcf576de5f5c7831680a1f0da626"
+    },
+    {
+      "path": "runtime/native/smallm_gemv/smallm_gemv.cu",
+      "size": 5918,
+      "sha256": "5de1f50d0c10ab87f87025de29600db5212d94b2558eae5bda292c3aac083521"
+    },
+    {
+      "path": "runtime/native/smallm_gemv/smallm_gemv.h",
+      "size": 316,
+      "sha256": "8ef4a719d3799c5afb226e40d961d82dd97412bf709fac084a9bdf71a4eb2d27"
+    },
+    {
+      "path": "runtime/server.py",
+      "size": 5276,
+      "sha256": "2975cb70ebc20828e39719306d1e36075141531ea457aa2b0b12f700dfea68d1"
+    },
+    {
+      "path": "runtime/torch-overrides.txt",
+      "size": 141,
+      "sha256": "1e4323fd7ec8624ef33657e15571562d214879275875ff59a464952d2585a02d"
+    },
+    {
+      "path": "runtime/web/index.html",
+      "size": 11668,
+      "sha256": "74653d8ba1b8fbfb48f5db12d85aeae7511119079043cfa18bb874841bfb2bc5"
+    },
+    {
+      "path": "runtime-data/references/.gitkeep",
+      "size": 1,
+      "sha256": "01ba4719c80b6fe911b091a7c05124b64eeece964e09c058ef8f9805daca546b"
+    },
+    {
+      "path": "special_tokens_map.json",
+      "size": 101864,
+      "sha256": "c2ff18fde6e43b7408435bc8ed079af74531befba549358a97cbc59ce606bc6b"
+    },
+    {
+      "path": "tokenizer.json",
+      "size": 12217872,
+      "sha256": "f24e08099d45a8adf3f52f5f0b03276e433bb9d689bb15fcbcc48ce58744588b"
+    },
+    {
+      "path": "tokenizer_config.json",
+      "size": 860832,
+      "sha256": "b8d149343ae425b0da67e6708686aceb51be7815d9792f265fc12ff04d5e9856"
+    },
+    {
+      "path": "validate_release.py",
+      "size": 11355,
+      "sha256": "c37f3e9c97882396db85679a8790f51b7dc50eee75848334b8191e6e275deb21"
+    },
+    {
+      "path": "vendor/fish-speech/.dockerignore",
+      "size": 1589,
+      "sha256": "d2e3fb35eae76b30d8ea06950affe3fd5174542249fd052f61644e8c0f7760bd"
+    },
+    {
+      "path": "vendor/fish-speech/.github/ISSUE_TEMPLATE/bug_report.yml",
+      "size": 2820,
+      "sha256": "1a105d8adfe941ab402fe7b5c67b37a65a13e44ca155917ec171e719d4212974"
+    },
+    {
+      "path": "vendor/fish-speech/.github/ISSUE_TEMPLATE/config.yml",
+      "size": 207,
+      "sha256": "53700059c64ff2b2a68b3cbe3ae5aec2e94783aba2414b7be14bc7a4e504fcea"
+    },
+    {
+      "path": "vendor/fish-speech/.github/ISSUE_TEMPLATE/feature_request.yml",
+      "size": 3013,
+      "sha256": "7aa0fa3e8b6a4695b90a26fe0ca0b766f81e5008aebef796d1316ffbcf0cec58"
+    },
+    {
+      "path": "vendor/fish-speech/.github/pull_request_template.md",
+      "size": 157,
+      "sha256": "600549340bc95f15c3f95c6d015fa814027220a99aaf1734dba344da4615f9c4"
+    },
+    {
+      "path": "vendor/fish-speech/.github/workflows/build-docker-image.yml",
+      "size": 2617,
+      "sha256": "2823c0767aebc69469286f50bdf387fa77cf08ac7246970cc65b612ce5910b67"
+    },
+    {
+      "path": "vendor/fish-speech/.github/workflows/docs.yml",
+      "size": 822,
+      "sha256": "8ae583f964ad63254123abbbfa8b11179cd763445012831b56a95ceac6c2c0ec"
+    },
+    {
+      "path": "vendor/fish-speech/.github/workflows/stale.yml",
+      "size": 959,
+      "sha256": "059b20b9316f59ec485f770fdd43f81bcd7b213fcc7d57c6e3763c6ad16d1e55"
+    },
+    {
+      "path": "vendor/fish-speech/.gitignore",
+      "size": 1344,
+      "sha256": "826527d0096c82edc0e6a076976134ba21b9eaf884da295b44788f7bd0f7a193"
+    },
+    {
+      "path": "vendor/fish-speech/.pre-commit-config.yaml",
+      "size": 550,
+      "sha256": "6d3e3c6ff61b53f62370995a0177d6fcf337d0b430ee858a407c3930185d23cf"
+    },
+    {
+      "path": "vendor/fish-speech/.project-root",
+      "size": 0,
+      "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
+    },
+    {
+      "path": "vendor/fish-speech/.readthedocs.yaml",
+      "size": 438,
+      "sha256": "2bc0776dddcc9829597377de94866048f00dfe926bf0b4d44d484c59c0ce0827"
+    },
+    {
+      "path": "vendor/fish-speech/API_FLAGS.txt",
+      "size": 204,
+      "sha256": "bd47758241d9a11d4b8b88aefd6500560500eb8db5b596c5f0241c6d4b7b08df"
+    },
+    {
+      "path": "vendor/fish-speech/LICENSE",
+      "size": 10360,
+      "sha256": "aa7d9206e9d710590987a3636934f643529c00cd490323594e6206aaa0c32d80"
+    },
+    {
+      "path": "vendor/fish-speech/README.md",
+      "size": 12065,
+      "sha256": "23193581f6c9a0dd174cd1427e670d64a0aeda11394f627a124b78876a23d544"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/.gitignore",
+      "size": 253,
+      "sha256": "fe718e7babb14f3cbad2d97f08889b9ce5215ed3fe0e43b2b8cfbfb3b9b844e8"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/README.md",
+      "size": 2520,
+      "sha256": "14e649116424a26ec2a1e47477c37383e4288f690a17167d2d2b88d3c4a5f725"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/eslint.config.js",
+      "size": 616,
+      "sha256": "4efe97b16d1200fac0eaf07aa00930a8668b1f96a0be1891ace8c1e712ff0ccb"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/index.html",
+      "size": 285,
+      "sha256": "f344a5836a0826a031127fdc7bc1f9f79e8ca928c45f79076d460b36df301610"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/package-lock.json",
+      "size": 121456,
+      "sha256": "307a2e98994212382c86e9356642f0dacfd405fc0cc74405ff0ca4dec43ab2d9"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/package.json",
+      "size": 1290,
+      "sha256": "a569905880fa7d29b8d5aadd08cccfb4b478bb3987057167201cb97c04054aaf"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/public/vite.svg",
+      "size": 1498,
+      "sha256": "fff702862e14c3ce019b81d86e07a0734764cac22ca5dbbe8542d4a682e482b2"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/src/App.tsx",
+      "size": 42735,
+      "sha256": "c5af7ee6cef89fab8f5a8271afaec8a84314b9b8168e883d4f9c4ff6b6c9169d"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/src/assets/react.svg",
+      "size": 4127,
+      "sha256": "9d72e876c70d7c7d4cbfe6c315747b66308176577329d2dc59b75191bdc21a4e"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/src/components/ui/alert.tsx",
+      "size": 1116,
+      "sha256": "2d5d2bf02f63e9b0e7f0e35a05d288f640e75b8c0c432166ac5aafb381d877c5"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/src/components/ui/badge.tsx",
+      "size": 862,
+      "sha256": "dece5fb21fe58f1cac9d6787e148c89e6c528db1f6c0906f4367f46fb6f750e9"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/src/components/ui/button.tsx",
+      "size": 1625,
+      "sha256": "2acdf5422a5886858c53f6116c984fb052f1f281fc7e1a02a842ab57b80e4ae4"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/src/components/ui/card.tsx",
+      "size": 1041,
+      "sha256": "17bb4d1be2607c0468369dd8ccab60f7b69a7e997d7131c1554b26ec23b71a65"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/src/components/ui/collapsible.tsx",
+      "size": 313,
+      "sha256": "0e665a7b8c5b0cca799e13704afb704e26f6afc81e70b132c903a163289acb08"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/src/components/ui/dialog.tsx",
+      "size": 2369,
+      "sha256": "37e74fb20d49ab9001261e46831d28b88bdc1695a421764e107071d6b8800f75"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/src/components/ui/label.tsx",
+      "size": 366,
+      "sha256": "2364eaa6c73a1662a1f6a6c771d1b806913ee77b6678481c2791796325050ce0"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/src/components/ui/scroll-area.tsx",
+      "size": 1318,
+      "sha256": "642b84426acec1cfefc23a472c6ddd051f9d43995215ceb0a52b83d7bf1e1099"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/src/components/ui/separator.tsx",
+      "size": 588,
+      "sha256": "3efb4936a73d93406b0cb07da3bdf070e21e2c93f44be61ff422e1eab220b090"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/src/components/ui/slider.tsx",
+      "size": 937,
+      "sha256": "edbc70edafa017bd6beb6fe169fe9e160d56a0a90aecd70a07d8f70245f841f9"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/src/components/ui/switch.tsx",
+      "size": 1041,
+      "sha256": "ebcf059e99fbbd83da8199fb26895353d718769eee111156a31dcb7249885d64"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/src/components/ui/textarea.tsx",
+      "size": 548,
+      "sha256": "3600836a7ae332cbc4a5609df662c3a328fde91066d3058779b8557d0e6a0c59"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/src/components/ui/toggle-group.tsx",
+      "size": 1423,
+      "sha256": "186970df1f46976a563b5a58ea52722cef6165f3a2547afec886bbaf2c27829e"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/src/index.css",
+      "size": 2003,
+      "sha256": "15688edb964294fa722b43c9acf444d3e61e4302d4102d5740ad7c7700ce8ad9"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/src/lib/utils.ts",
+      "size": 168,
+      "sha256": "bb1155a89f94a70f6e1c64138929f034af72405f698e3346b9c571c6a81e1938"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/src/main.tsx",
+      "size": 230,
+      "sha256": "6e9e5807fcbd48b75a96db5cbef36c996262196be42e6d4760dc86babbe61ad2"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/tsconfig.app.json",
+      "size": 816,
+      "sha256": "b75e309a7934571963d315a8f2376c6761adbdc8d7ab177d79ead9ff7fdc5a08"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/tsconfig.json",
+      "size": 119,
+      "sha256": "770b4140bbb581e2dfd9ea9946ffc9c75a1d86ba7d2db5f77c83e37cbdf9d808"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/tsconfig.node.json",
+      "size": 725,
+      "sha256": "8783c2ca6cf9b6cf07578deb90c2a42c09460bdfa725de6add8e1ace9ce486ee"
+    },
+    {
+      "path": "vendor/fish-speech/awesome_webui/vite.config.ts",
+      "size": 3278,
+      "sha256": "8d9fddde878fbd0e6fad17325272c81d1ba3ff40ef8c34e904e7553e9e78190e"
+    },
+    {
+      "path": "vendor/fish-speech/compose.base.yml",
+      "size": 633,
+      "sha256": "9741535b4d5bc58c088dd9a8ed6a308fa6dc845a78ec17b98e268a7ead7e4e7b"
+    },
+    {
+      "path": "vendor/fish-speech/compose.rocm.yml",
+      "size": 1318,
+      "sha256": "e6ce9cb3cfa06d81a537662e544bf6c1a6d1ee0bf57eb3ab126a8201ab4b16a9"
+    },
+    {
+      "path": "vendor/fish-speech/compose.yml",
+      "size": 476,
+      "sha256": "9a069086b7edc69c27c9863199962c3c3cfc3ade329678cd288ad2d0891cc3e4"
+    },
+    {
+      "path": "vendor/fish-speech/docker/Dockerfile",
+      "size": 12538,
+      "sha256": "3df60e1f23454cf31fed92e3e018eb772f1d6677eecce24985a74078b904d5e2"
+    },
+    {
+      "path": "vendor/fish-speech/docker/Dockerfile.rocm",
+      "size": 4025,
+      "sha256": "7b5fe82284aa9b28601c5952d4eafaa3b0f01652303b6718c41884a4e7050a79"
+    },
+    {
+      "path": "vendor/fish-speech/dockerfile.dev",
+      "size": 1074,
+      "sha256": "52cf29b9ae3bd05fe1d58473c9538fd632a8b2e8c0dd12d63e220e4a7a9648b5"
+    },
+    {
+      "path": "vendor/fish-speech/docs/CNAME",
+      "size": 18,
+      "sha256": "94b6e87dc6e8704a583456b9d2db4a39559c5ae6a65e1d92f977776863175a00"
+    },
+    {
+      "path": "vendor/fish-speech/docs/README.ar.md",
+      "size": 15648,
+      "sha256": "19b3439785d22912f794449de6f489c879cd4c93632bc8857aede5d0ee00c4d9"
+    },
+    {
+      "path": "vendor/fish-speech/docs/README.es.md",
+      "size": 9421,
+      "sha256": "69c381dba1b8d401a55e6d67c6a332938d2c058fa20f9e543187a92330f44c21"
+    },
+    {
+      "path": "vendor/fish-speech/docs/README.ja.md",
+      "size": 13744,
+      "sha256": "1a3b6726e5c2b9c6af0e172a52ae21b136c58b64fb599698aa0dcb48dd90c6eb"
+    },
+    {
+      "path": "vendor/fish-speech/docs/README.ko.md",
+      "size": 13034,
+      "sha256": "415f15e48f2b1f97619b91a9602dcd7348d81df84d69cd3cf9a2a4159cac6443"
+    },
+    {
+      "path": "vendor/fish-speech/docs/README.pt-BR.md",
+      "size": 13135,
+      "sha256": "90233321ef5f1d753f9dc8162a13f0a21a86370d0e32e5b732d232c7dc0e33d8"
+    },
+    {
+      "path": "vendor/fish-speech/docs/README.zh.md",
+      "size": 11662,
+      "sha256": "c395ac54df44cf7362bfd800e3d2db170ad5ec60a5908cea15bde87b1c953bda"
+    },
+    {
+      "path": "vendor/fish-speech/docs/ar/finetune.md",
+      "size": 6082,
+      "sha256": "e94b0cdcea11ce741c17d883b32beabb541e1653f267fe59a9efd5931b0f7cdb"
+    },
+    {
+      "path": "vendor/fish-speech/docs/ar/index.md",
+      "size": 13191,
+      "sha256": "29995c610cc69e982e905948463592273c712d15fac87622c54483e2199aeb9e"
+    },
+    {
+      "path": "vendor/fish-speech/docs/ar/inference.md",
+      "size": 3189,
+      "sha256": "a0dc48584eeff9452d4051b7ca3cc86c4b118aa796dbd4aa2acded07571a1404"
+    },
+    {
+      "path": "vendor/fish-speech/docs/ar/install.md",
+      "size": 7525,
+      "sha256": "837fd7cde51ce62c685aaac3945d6aa8d28818a08c7a51f4d0c9546ce1ec836e"
+    },
+    {
+      "path": "vendor/fish-speech/docs/assets/chattemplate.png",
+      "size": 3001037,
+      "sha256": "125061f3dbb707d7a494d991e6854a47dc7bd5e853590f0223bac154243b4778"
+    },
+    {
+      "path": "vendor/fish-speech/docs/assets/logo.svg",
+      "size": 1787,
+      "sha256": "15ae95ec011b44031163c7d92044152dbdba9055d13ee050992ad93455b18ab9"
+    },
+    {
+      "path": "vendor/fish-speech/docs/assets/totalability.png",
+      "size": 3564498,
+      "sha256": "b21736aa07c0cb1a1a0ae55ae0b97ab2a6ecbee34c0b7bc38076c83420617bcd"
+    },
+    {
+      "path": "vendor/fish-speech/docs/en/finetune.md",
+      "size": 4450,
+      "sha256": "af5355d64cef5bfbe652763790b45817b4027991ac866461b04cade72a9de32c"
+    },
+    {
+      "path": "vendor/fish-speech/docs/en/index.md",
+      "size": 10688,
+      "sha256": "cd8ddcff0de3c2b549a992f22118029e289fe0e213de83ac30700b4b8204b798"
+    },
+    {
+      "path": "vendor/fish-speech/docs/en/inference.md",
+      "size": 2400,
+      "sha256": "bbe2b3e9e2355c480477f4bfdf053f348917a42665cdb52464ede8514aa337bc"
+    },
+    {
+      "path": "vendor/fish-speech/docs/en/install.md",
+      "size": 7135,
+      "sha256": "9930c9099eced275c6dc7ccecae026411ba9de8ab61195651a927213030b2a76"
+    },
+    {
+      "path": "vendor/fish-speech/docs/en/server.md",
+      "size": 2053,
+      "sha256": "910820d8c593bc5722fae1108c21f583f588fbde12b5bb63fc3c2f472bce5f78"
+    },
+    {
+      "path": "vendor/fish-speech/docs/es/finetune.md",
+      "size": 4661,
+      "sha256": "3a498c00f147f31155c29007d9572db985c7c49d78febb3f2136e3d862d5c7c3"
+    },
+    {
+      "path": "vendor/fish-speech/docs/es/index.md",
+      "size": 11320,
+      "sha256": "0c476da28d0074d2c0614b0ac8a297bb7f9734579d258e46497fa5e27446cfec"
+    },
+    {
+      "path": "vendor/fish-speech/docs/es/inference.md",
+      "size": 1652,
+      "sha256": "697399235eb16525cb9f11db052c6c662301cbcae70811cb0139ff88354f2ace"
+    },
+    {
+      "path": "vendor/fish-speech/docs/es/install.md",
+      "size": 6258,
+      "sha256": "e9445f4d9abeccf9870d9912d1a3ed287282d87620c94507727566c6150e5d82"
+    },
+    {
+      "path": "vendor/fish-speech/docs/es/server.md",
+      "size": 2291,
+      "sha256": "a5ef0ccf911c67ada2f414c3e4b10cc17d89944c163ea838f01c2997ef288695"
+    },
+    {
+      "path": "vendor/fish-speech/docs/ja/finetune.md",
+      "size": 5665,
+      "sha256": "ee03602acc09a7a23f0bd51e9215f64b1d64d03a6c46e92fc13e3234ebec979c"
+    },
+    {
+      "path": "vendor/fish-speech/docs/ja/index.md",
+      "size": 11910,
+      "sha256": "b55777fc674b3ea5c419a52a4702479120cba951d7a0a196cb38c68ec5dd90cc"
+    },
+    {
+      "path": "vendor/fish-speech/docs/ja/inference.md",
+      "size": 3127,
+      "sha256": "a4bb49bd21855128b2cb157c0733fab0ef62c62c850b3b97b2dc9724eee1ef43"
+    },
+    {
+      "path": "vendor/fish-speech/docs/ja/install.md",
+      "size": 7155,
+      "sha256": "9032434192c823e8f3c6749723aa0dbe2c9861dde5b45675d03e3024c95915f9"
+    },
+    {
+      "path": "vendor/fish-speech/docs/ko/finetune.md",
+      "size": 5064,
+      "sha256": "62a5e0e92fc398e4f9bf2736ee9fe2543d2b4b845e3297dc33c8ea4a295906bb"
+    },
+    {
+      "path": "vendor/fish-speech/docs/ko/index.md",
+      "size": 10976,
+      "sha256": "457f171814d81819ed4bb776244858c9195ee686fc1bfac5e8e1a27d5ac89021"
+    },
+    {
+      "path": "vendor/fish-speech/docs/ko/inference.md",
+      "size": 2707,
+      "sha256": "aba1a3997e6b7b5d140c79f83976827b420fd3a1db69c63772a8fceb884120d9"
+    },
+    {
+      "path": "vendor/fish-speech/docs/ko/install.md",
+      "size": 6154,
+      "sha256": "4ef88b9b1c66fa5c6249c76242cf4c28bd253c8a7e75b34f2c0975b8aa3b679b"
+    },
+    {
+      "path": "vendor/fish-speech/docs/pt/finetune.md",
+      "size": 4808,
+      "sha256": "26fe44a755f77a9ecdba86de8170c32845d011b7e9c9e6a28e0c099acd0e31b7"
+    },
+    {
+      "path": "vendor/fish-speech/docs/pt/index.md",
+      "size": 10937,
+      "sha256": "d491cbe173ed548bbcac3b722855b0d2b34aec59e5c36e10954719efc8db7c9c"
+    },
+    {
+      "path": "vendor/fish-speech/docs/pt/inference.md",
+      "size": 2643,
+      "sha256": "0a00d7066248b42fd8f6c3ef1b60bb96303ee7e8ed60b4365c1573b3f042ea34"
+    },
+    {
+      "path": "vendor/fish-speech/docs/pt/install.md",
+      "size": 6239,
+      "sha256": "b46c769edb2a7f9a3159ada928bd85ed5baa318ef0c8c3b5b49fcb15236686b8"
+    },
+    {
+      "path": "vendor/fish-speech/docs/requirements.txt",
+      "size": 58,
+      "sha256": "d1ad209ce393487abd53afa357470c9d12e17f048689e30e20dfce5e03f42127"
+    },
+    {
+      "path": "vendor/fish-speech/docs/stylesheets/extra.css",
+      "size": 35,
+      "sha256": "8d429df9c7907d7e4fee12af1af8bfec0a29f6142cc9b5cb62c8b1de29e0898c"
+    },
+    {
+      "path": "vendor/fish-speech/docs/zh/finetune.md",
+      "size": 4081,
+      "sha256": "e0622704c25cd0941b5672a9370101b2123dcb39be48f6c1764204fa864b2f9d"
+    },
+    {
+      "path": "vendor/fish-speech/docs/zh/index.md",
+      "size": 10114,
+      "sha256": "0560d6e67a097ee92d02615b893ae306c13b9d71b15e4edef6abcd70ae46f94c"
+    },
+    {
+      "path": "vendor/fish-speech/docs/zh/inference.md",
+      "size": 2300,
+      "sha256": "19afd07e8af2a5e403adf8e3df2ff5468e8e3ab5e85d6b545707578855e547ec"
+    },
+    {
+      "path": "vendor/fish-speech/docs/zh/install.md",
+      "size": 5166,
+      "sha256": "c5b4953f4f0cc8654f49866442c9a2ce8655b5703730014f103a053e11e4c4a6"
+    },
+    {
+      "path": "vendor/fish-speech/entrypoint.sh",
+      "size": 171,
+      "sha256": "e16d821b9e21b0ba8915b497aed9a4612cf99801181f2fe1f9a87d895e71cf50"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/callbacks/__init__.py",
+      "size": 141,
+      "sha256": "2e60f60140744e0acbd2d5e8b7d4bc95ea5593586b1b4d5ebbe6a58ea7ef4fbb"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/callbacks/grad_norm.py",
+      "size": 3436,
+      "sha256": "a73b94c549d75806bd53a5ca72f4299c6a146613261502a2992a981fe59a8c84"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/callbacks/progress_bar.py",
+      "size": 518,
+      "sha256": "caa68bcfaa4894ffe23312648ccdbf65ed24698dbc017eb413fc9a5da94d930b"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/configs/base.yaml",
+      "size": 2618,
+      "sha256": "3131447a82b4f8e8176c653d1a7dc4021cf7d4ede12c320a3a76fa210d072b55"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/configs/lora/r_32_alpha_16_fast.yaml",
+      "size": 182,
+      "sha256": "cdc5f6c8a8af194869ae544215b86e5a3a334f47266df47e922c2c008cb19bea"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/configs/lora/r_8_alpha_16.yaml",
+      "size": 98,
+      "sha256": "6bd41590dd80eb8ff497c5d47f05757e05b97998bdd76e0be691ddb0df598082"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/configs/modded_dac_vq.yaml",
+      "size": 1376,
+      "sha256": "73321408579c372149620d877f0dfb841cf70465758a535f7243e1cb6553d56a"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/configs/text2semantic_finetune.yaml",
+      "size": 2074,
+      "sha256": "dcfb2cf20f62506a04078df72a1ecf52cea2bbb23db86f900439a02a7f6f3370"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/content_sequence.py",
+      "size": 14345,
+      "sha256": "429218fb07feb5910d3b232bf6289555dc1c99e2a8622a28963737c3990a0c50"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/conversation.py",
+      "size": 5602,
+      "sha256": "1baece5e55f9e3b27f655a3ac8c3ceef441ff87354db0b91a840cac137c4e3c2"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/datasets/concat_repeat.py",
+      "size": 1498,
+      "sha256": "cbda40b72dd26b78327c011149ab9b98c351754cf2030e33ce80f25f3b861b14"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/datasets/protos/text-data.proto",
+      "size": 392,
+      "sha256": "d6bb8c95bc7925f43efd0033cf03e7059e73add38f4d1944a2f560a51c279f7c"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/datasets/protos/text_data_pb2.py",
+      "size": 1759,
+      "sha256": "54f5f84e4290f21e2a53cc5063228c5de52e9736aac971e22c38cf7a5b2cf750"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/datasets/protos/text_data_stream.py",
+      "size": 781,
+      "sha256": "bfd8b78db202e8a15348bbd256ca3cfb79810ee7d23b013b200af8fb9457810d"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/datasets/semantic.py",
+      "size": 20572,
+      "sha256": "ef9332053ea341bb4e04d9cafedfa2e9771d5a9e522abf868ad3dec58eb0beea"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/datasets/vqgan.py",
+      "size": 3933,
+      "sha256": "d8612b491cabf5e9f6094617559d8d5a74c223fd2b08e7135e7f88a34b885b67"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/i18n/README.md",
+      "size": 1483,
+      "sha256": "a57432fe33c55f24c14129e2e094851f526db28a8e7bbbce7d6a51b4e3c6cfac"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/i18n/__init__.py",
+      "size": 43,
+      "sha256": "bbdaff1c2b53998714eb82fad8683d94b00dbc6c9b7dbd3e113c3c39bf9f9f16"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/i18n/core.py",
+      "size": 1036,
+      "sha256": "f0ab3cc06a20ef753e019aa5cc696d7f1eb52802c64e1eee0885e74070e70cec"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/i18n/locale/ar_SA.json",
+      "size": 9843,
+      "sha256": "0eeeb96d4ee6e6405c415d490651566f2d9275172c8ffcae75c23637f410a1be"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/i18n/locale/en_US.json",
+      "size": 8103,
+      "sha256": "c41c66facb7f4e4641501828830f532a20f216ad2b258adddb91b96a3d8a0c98"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/i18n/locale/es_ES.json",
+      "size": 9012,
+      "sha256": "dcf80e38f59d415171271e3f12b4324bdba0ff27f1222a8504e0917df5c226b4"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/i18n/locale/ja_JP.json",
+      "size": 9496,
+      "sha256": "225330a3fb51f0be2c2aa54afe6522ced6ad1ecec57d15cd4cc695de54768b04"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/i18n/locale/ko_KR.json",
+      "size": 8787,
+      "sha256": "792f1862d348506b334e444b1adba83d148da67638a7f3918efa9ef594886bfb"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/i18n/locale/pt_BR.json",
+      "size": 9615,
+      "sha256": "44d54552e9ff585bbbaaa8389a79200fa7e68b59881121b08cee10bda96c1ae8"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/i18n/locale/zh_CN.json",
+      "size": 7773,
+      "sha256": "ec1739836b8ade678540c692903489b992171e1afb171fcff0e9dbdae2c3ddbf"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/i18n/scan.py",
+      "size": 3751,
+      "sha256": "d6675728f791d1dddb4dc1595a15fa38359425b6e23873bbdcc0431099539946"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/inference_engine/__init__.py",
+      "size": 6261,
+      "sha256": "2e511bb701b2e1a07770700fb0d33a1c51b7066a418e8897467729913c6c59d0"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/inference_engine/reference_loader.py",
+      "size": 9567,
+      "sha256": "7b9fd6ce1b5aa0b08a2d6ec6630549168ea57933e5f35027b6ac22efd1e7b7bb"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/inference_engine/utils.py",
+      "size": 685,
+      "sha256": "22f32b647e2d73795dddbde9cb1e21884663627bd1a2c78e38381a81c0b933fe"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/inference_engine/vq_manager.py",
+      "size": 1952,
+      "sha256": "881f9fa54661076b976e917caf9da1c04b7d476b77c9cf92f1edc22c2125e932"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/models/dac/__init__.py",
+      "size": 0,
+      "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/models/dac/inference.py",
+      "size": 3895,
+      "sha256": "1914f547001ec23d25cdff3a495dc40e2434f4657d8fe3bda837d88001520331"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/models/dac/modded_dac.py",
+      "size": 35062,
+      "sha256": "a08407421ee85d8af28377d6a14d989b5a17a985c9719f8b5701cd0a845840c0"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/models/dac/rvq.py",
+      "size": 13143,
+      "sha256": "a4d38e529846473c712dd1b2f5eaa889eb0233fd56228799c168060f335c0246"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/models/text2semantic/__init__.py",
+      "size": 0,
+      "sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/models/text2semantic/inference.py",
+      "size": 31250,
+      "sha256": "0c2e6746cd1fcc5e776cc37d6430a33c3fda1fc1c08b347860903f345a1c0779"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/models/text2semantic/lit_module.py",
+      "size": 6935,
+      "sha256": "c1c6d1d39d20e6e94a0aa75148e6d416534188ac595b6ec22d5ff131b7f31915"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/models/text2semantic/llama.py",
+      "size": 36007,
+      "sha256": "b7dc3c039ddcbc05e445293e5d4babb1e80340b8a2944747fab0cf44c0919852"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/models/text2semantic/lora.py",
+      "size": 4084,
+      "sha256": "551e140a0cf78a779ab86cdeb4b5c71af330282651ba9cc401937fba4ed23dc7"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/scheduler.py",
+      "size": 1101,
+      "sha256": "c69b2d01f5e5b1dc7e37f84370925814a88b6ff716e664dd5c00b2b8e0eb7088"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/text/__init__.py",
+      "size": 56,
+      "sha256": "584dd80b9ec25e96e1bdae800874e8d0edd3355c2732e52a0644a906e32b4e26"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/text/clean.py",
+      "size": 832,
+      "sha256": "3b9b1ec80e2d4f8aa71d386f19ab362bea7603485b63836b38a202d137e36d6e"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/tokenizer.py",
+      "size": 3948,
+      "sha256": "a25d136cdf3d150e862a808fe9badf8d7ab473509f4068c6115772ff98b566da"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/train.py",
+      "size": 4470,
+      "sha256": "32d55a2cfb4c50498b4ecca9b0edc201a835ecf1ee8b232973d58e477bec9e40"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/utils/__init__.py",
+      "size": 706,
+      "sha256": "dd8206cec25b8fce284d1f96428173fe9ced66d5081e9d7adcaff1d170fcb715"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/utils/braceexpand.py",
+      "size": 6724,
+      "sha256": "5ffe38c16f8b24fdb590d6f5f63eb70d24b3168a6482c48d92a9ac1252d24ba5"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/utils/context.py",
+      "size": 287,
+      "sha256": "1b9032235a9abe15febc17ac00d53431bf7ba23058f349f089706bccd15d4391"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/utils/file.py",
+      "size": 3354,
+      "sha256": "d6d13187c1487698a4d5aa5c4b01e52d6da1b579931294b465534956c10b8286"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/utils/instantiators.py",
+      "size": 1514,
+      "sha256": "2a45f5f174f3d7712dde02b308610b5c8d59b05fce9ecf2b750c4f9b846fc61e"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/utils/logger.py",
+      "size": 2467,
+      "sha256": "5959ccc8442a94eda7bede31c92905a2071ea5e34335c1d68d3618fa6d1fa382"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/utils/logging_utils.py",
+      "size": 1384,
+      "sha256": "049de1f2117eeb631bab16a38666602a0964a633dfcff193153afaf364887985"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/utils/rich_utils.py",
+      "size": 3105,
+      "sha256": "447454aeadb1635775120db886d1679b28f44515b6351770c3dc73d235a742df"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/utils/schema.py",
+      "size": 3912,
+      "sha256": "8f4e8b272c29fc390272ce5c15f74268679d3f4434e684a6519141692c7f7c53"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/utils/spectrogram.py",
+      "size": 3325,
+      "sha256": "9663d20ae9c79e5a517516be9aefa25485292a69f53426b847d812c7f13bbbd2"
+    },
+    {
+      "path": "vendor/fish-speech/fish_speech/utils/utils.py",
+      "size": 4283,
+      "sha256": "df0d1dac088ba396d59cead71fc7b3b0801a02448fddcab082931cafb9044454"
+    },
+    {
+      "path": "vendor/fish-speech/inference.ipynb",
+      "size": 5021,
+      "sha256": "3ad1d80be8824a5fb980e38e9efc940e71d2b082f73b4efeacb49a4561077e1d"
+    },
+    {
+      "path": "vendor/fish-speech/mkdocs.yml",
+      "size": 4108,
+      "sha256": "17a35919a0f090a195205cdb68288874d1f09a4c1454acc04bd1eb60a5830e6b"
+    },
+    {
+      "path": "vendor/fish-speech/pyproject.toml",
+      "size": 2966,
+      "sha256": "6157b9769168546b14f76831da5549ec1674d7f338185f4119e72c8dbb0eab59"
+    },
+    {
+      "path": "vendor/fish-speech/pyrightconfig.json",
+      "size": 63,
+      "sha256": "f78eaa043520319af6a9619303228999027e636d3307baf174638c7b08777fad"
+    },
+    {
+      "path": "vendor/fish-speech/tools/api_client.py",
+      "size": 7704,
+      "sha256": "47cc75a9fd146763581dafb12e991fa7969e915aae78a0690d90b476371732a9"
+    },
+    {
+      "path": "vendor/fish-speech/tools/api_server.py",
+      "size": 4408,
+      "sha256": "0e8c70f6184cdb0b64464a7f3204284051d86ba132c478ff86ddb649cfbd89dc"
+    },
+    {
+      "path": "vendor/fish-speech/tools/llama/build_dataset.py",
+      "size": 4910,
+      "sha256": "b115a58c3e3e588ce2106f2e0db7ee3a26773ac2b5520d7244a36ef5c0ab8c7c"
+    },
+    {
+      "path": "vendor/fish-speech/tools/llama/eval_in_context.py",
+      "size": 4641,
+      "sha256": "e46a8072380ad934dfadad4c621ab2749646376eef5e5b6a873ea15d29f75b59"
+    },
+    {
+      "path": "vendor/fish-speech/tools/llama/merge_lora.py",
+      "size": 3369,
+      "sha256": "6f1de34312b56697c275e47d35104e336b5905db6804bd93ea4561fc92e1ea65"
+    },
+    {
+      "path": "vendor/fish-speech/tools/llama/quantize.py",
+      "size": 16589,
+      "sha256": "3a07f3fbc44240c744725c75ec3587d48857de7e039847a9c79e6e4be7fd07f3"
+    },
+    {
+      "path": "vendor/fish-speech/tools/run_webui.py",
+      "size": 3377,
+      "sha256": "20c3b10f10dbf45659dbfbfd5bff2fbf32f73ace966013a80d7909362d8dd5a3"
+    },
+    {
+      "path": "vendor/fish-speech/tools/server/api_utils.py",
+      "size": 4451,
+      "sha256": "6bbb1414eefd42f6be053e0d27a381df8c921adcb0e4c4e12d02e4bd596bf5ec"
+    },
+    {
+      "path": "vendor/fish-speech/tools/server/exception_handler.py",
+      "size": 729,
+      "sha256": "09471ac426e926d9c824e3f67c30f0e2884b17a2e502492f1632035c8075c171"
+    },
+    {
+      "path": "vendor/fish-speech/tools/server/inference.py",
+      "size": 1352,
+      "sha256": "445898e3c2f97ce6fedd34b906b4bf0d53dd7dd896a27830d4f9484e142f26ae"
+    },
+    {
+      "path": "vendor/fish-speech/tools/server/model_manager.py",
+      "size": 2974,
+      "sha256": "6135527b37067f4cc7ae4a782ee6584b8881d1b7fbd9b7fabecfebd8ebda7e03"
+    },
+    {
+      "path": "vendor/fish-speech/tools/server/model_utils.py",
+      "size": 2643,
+      "sha256": "4680f7c92d04524b72a506cdf4525c4bd5dc5fa967e9a9cef0d018a1c6e2a893"
+    },
+    {
+      "path": "vendor/fish-speech/tools/server/views.py",
+      "size": 17012,
+      "sha256": "25234db4aa025bde24d26a69a89fa8c5eb91b0c24dac2265533a4d4f458104e2"
+    },
+    {
+      "path": "vendor/fish-speech/tools/vqgan/create_train_split.py",
+      "size": 3008,
+      "sha256": "ac096e4fb7ae098c6d79bd4e91a79e9d610eb141ba0f492b036566badb61fad6"
+    },
+    {
+      "path": "vendor/fish-speech/tools/vqgan/extract_vq.py",
+      "size": 7116,
+      "sha256": "3552649e3827fd2ac388fd3207aed3e095e107bb10f304e8cdd5414c71d849f6"
+    },
+    {
+      "path": "vendor/fish-speech/tools/webui/__init__.py",
+      "size": 6167,
+      "sha256": "8638f675fa822782a290e28907a3e9b73cf3b376fea68298baa45f92e0b3bb7a"
+    },
+    {
+      "path": "vendor/fish-speech/tools/webui/inference.py",
+      "size": 2109,
+      "sha256": "01f14ab86b3fe5e883dffd0729a0a3e4af67f9834d2ac3c529568b8ca33aba15"
+    },
+    {
+      "path": "vendor/fish-speech/tools/webui/variables.py",
+      "size": 605,
+      "sha256": "3e10038f814557353737e3892dcb07cad2e0073188129010b9b4d817c12f4fbd"
+    },
+    {
+      "path": "vendor/fish-speech/uv.lock",
+      "size": 1225073,
+      "sha256": "59711c5afba9fa5b53e7c31bf869d1ee7993d60c3a11b36a73095d2f135f0807"
+    }
+  ]
+}

Notice

@@ -0,0 +1,20 @@
+This model is licensed under the Fish Audio Research License, Copyright © 39 AI, INC. All Rights Reserved.
+
+Built with Fish Audio.
+
+Changes made by ajh-code:
+
+- Quantized 60 slow-transformer gate/up projections in layers 3–32 to packed
+  NVFP4 E2M1 weights with a native W4A16 decode path and W4A4 larger-M path.
+- Quantized the other 120 slow-transformer projections to native MXFP8 W8A8.
+- Folded calibrated English channel scales into selected normalization and
+  packed-weight tensors.
+- Added a portable mixed-precision loader, Blackwell kernels, a TTS API/web
+  interface, compact inference-only DAC buffers, and reference-encoder staging.
+- Kept embeddings, normalization arithmetic, output heads, the fast
+  transformer, KV cache, sampling, and DAC arithmetic at BF16 or their original
+  higher precision.
+
+The original model is fishaudio/s2-pro, revision
+1de9996b6be38b745688de084d87a5633f714e4e. The bundled Fish Speech source is
+revision e5e292632cb11e7a27b2b7487f58f612bc101e13.

README.md

@@ -0,0 +1,273 @@
+---
+license: other
+license_name: fish-audio-research-license
+license_link: https://huggingface.co/fishaudio/s2-pro/blob/1de9996b6be38b745688de084d87a5633f714e4e/LICENSE.md
+library_name: fish-speech
+pipeline_tag: text-to-speech
+base_model: fishaudio/s2-pro
+base_model_relation: quantized
+language:
+  - en
+tags:
+  - fish-speech
+  - fish-audio
+  - text-to-speech
+  - voice-cloning
+  - quantization
+  - nvfp4
+  - mxfp8
+  - blackwell
+  - sm120
+  - english
+  - v1
+---
+
+<h1 align="center">V1 · Fish Audio S2-Pro · NVFP4 Balanced</h1>
+
+<p align="center"><strong>A complete Blackwell-ready S2-Pro download:</strong> mixed native NVFP4/MXFP8 transformer weights, BF16 codec, tokenizer, pinned Fish Speech source, API server, web UI, and reproducibility checks.</p>
+
+<p align="center"><strong>Balanced V1 prioritizes useful English quantization fidelity and VRAM reduction. This is not yet an XPO3 release.</strong></p>
+
+<p align="center"><strong>An XPO3 version is coming soon.</strong> Follow <a href="https://huggingface.co/ajh-code">ajh-code on Hugging Face</a> and <a href="https://arands.com">Arands.com</a> for release updates.</p>
+
+<p align="center">
+  <a href="https://huggingface.co/fishaudio/s2-pro">Original S2-Pro</a>
+  · <a href="https://github.com/fishaudio/fish-speech">Fish Speech</a>
+  · <a href="https://arands.com">Arands.com · updates</a>
+</p>
+
+---
+
+## Download
+
+| Component | Purpose | Size |
+|:--|:--|--:|
+| `model-*.safetensors` | Complete mixed NVFP4/MXFP8 transformer checkpoint | 4.90 GB |
+| `codec.pth` | Complete BF16 S2-Pro DAC codec | 1.87 GB |
+| Tokenizer, runtime, and pinned source | No separate base-model or codec download | ~24 MB |
+| Complete repository | Weights, codec, runtime, source, and metadata | 6.80 GB |
+
+All model and codec weights required by the server are in this repository.
+The root `config.json` preserves the S2-Pro architecture metadata and adds the
+mixed-precision policy, while Hugging Face metadata records this repository as
+a quantization of [`fishaudio/s2-pro`](https://huggingface.co/fishaudio/s2-pro).
+
+## Quick start
+
+Tested on Linux x86-64, Python 3.12, CUDA 13.0, PyTorch `2.11.0+cu130`,
+`comfy-kitchen==0.2.22`, and NVIDIA Blackwell SM120. The current native path is
+for GeForce RTX 50-series/SM120 GPUs; it is not a generic CUDA fallback.
+
+```bash
+hf download ajh-code/Fish-Audio-S2-Pro-NVFP4-Balanced \
+  --local-dir fish-audio-s2-pro-nvfp4-balanced
+cd fish-audio-s2-pro-nvfp4-balanced
+./install.sh
+./launch.sh
+```
+
+Open <http://127.0.0.1:8080/ui> for the bundled zero-shot web interface. The
+API listens on all interfaces by default; set `TTS_HOST=127.0.0.1` if it should
+not be reachable from the local network. Protect or firewall the service before
+exposing it beyond a trusted network.
+
+Docker Compose is the recommended clean deployment when Docker, the NVIDIA
+Container Toolkit, and a compatible driver are already configured:
+
+```bash
+docker compose up --build
+```
+
+This path passed a clean outer-Docker build, SM120 runtime launch, route/UI
+checks, and a zero-shot API smoke test on an RTX 5080 with CUDA 13.0.
+
+The image excludes the 6.8 GB model payload and mounts the downloaded repository
+read-only, so rebuilding the runtime does not duplicate the weights inside the
+image.
+
+## Zero-shot voice cloning
+
+Use a clean, consented 10–30 second reference with one speaker and supply its
+exact transcript:
+
+```bash
+python client.py \
+  --url http://127.0.0.1:8080/v1/tts \
+  --reference-audio reference.wav \
+  --reference-text "The exact words spoken in reference.wav." \
+  --text "A few notes as this story begins." \
+  --seed 42 \
+  --output result.wav
+```
+
+Equivalent JSON API call in Python:
+
+```python
+import base64
+from pathlib import Path
+
+import requests
+
+payload = {
+    "text": "A few notes as this story begins.",
+    "references": [{
+        "audio": base64.b64encode(Path("reference.wav").read_bytes()).decode(),
+        "text": "The exact words spoken in reference.wav.",
+    }],
+    "reference_id": None,
+    "format": "wav",
+    "streaming": False,
+    "normalize": True,
+    "max_new_tokens": 1024,
+    "chunk_length": 200,
+    "top_p": 0.9,
+    "temperature": 0.9,
+    "repetition_penalty": 1.1,
+    "seed": 42,
+    "use_memory_cache": "off",
+}
+response = requests.post("http://127.0.0.1:8080/v1/tts", json=payload, timeout=600)
+response.raise_for_status()
+Path("result.wav").write_bytes(response.content)
+```
+
+Useful endpoints:
+
+| Endpoint | Purpose |
+|:--|:--|
+| `GET /ui` | Bundled zero-shot web interface |
+| `GET /v1/health` | Service health |
+| `GET /v1/model` | Active release, quantization, and sampling metadata |
+| `POST /v1/tts` | Fish Speech-compatible TTS request; returns audio |
+
+Set `TTS_API_KEY` before launch to require bearer authentication. For an API
+key named `secret`, send `Authorization: Bearer secret`.
+
+## Quantization policy
+
+S2-Pro has 180 projections in its 36-layer slow transformer. V1 uses:
+
+| Scope | Stored/executed precision | Count |
+|:--|:--|--:|
+| Gate/up in layers 3–32 | packed NVFP4 E2M1; W4A16 at `M=1`, W4A4 above `M=1` | 60 |
+| Other slow-transformer projections | native dynamic MXFP8 W8A8 | 120 |
+| Embeddings, tied text output, fast transformer/output, norms, RoPE, KV cache, sampling | BF16/original precision | — |
+| DAC codec arithmetic | BF16 | — |
+
+The English calibration scale is folded into the selected norm and packed
+gate/up tensors. It adds no runtime tensor or operation. The original BF16
+slow-projection weights are not retained as a second copy. V1's transformer
+checkpoint is 46.25% smaller than the original transformer shards.
+
+This is intentionally described as a mixed NVFP4/MXFP8 checkpoint. It is not a
+claim that every operation, activation, or weight in the end-to-end TTS stack
+runs at FP4.
+
+## Measured performance
+
+Measurements below are local RTX 5080 results with the bundled compact BF16
+codec path and a 3072-token cache. They are not universal performance claims.
+
+| Measurement | V1 result |
+|:--|--:|
+| Loaded PyTorch allocation | 5.350 GiB |
+| 15-sample short zero-shot peak | 5.607–5.769 GiB |
+| Held-out ~39-second generation peak | 7.561 GiB |
+| Median short/control real-time factor | about 1.05–1.06 |
+| Median short/control time to first playable audio | about 5–6.5 s |
+| Semantic generation throughput | about 20.4–20.6 frames/s |
+
+The matching compact-runtime BF16 control loaded at 9.273 GiB, so V1 reduced
+loaded PyTorch allocation by 42.31%. V1 is near real time on the RTX 5080, but
+the current ordinary single-speaker API waits for a complete semantic segment
+before playable audio. This release therefore does **not** claim agent-grade
+low-latency streaming. An RTX 5060 Ti focused zero-shot run measured about 2.06
+RTF and is not a real-time path.
+
+## Limited blind voice-cloning test
+
+The first blind comparison is encouraging, but deliberately small. It used one
+listener, one consented English reference speaker, and four matched BF16/V1
+pairs: conversational, reflective, question-shaped, and long narrative prompts
+at seeds 7, 17, 123, and 42. Both models used the same reference, runtime path,
+`temperature=0.9`, `top_p=0.9`, and `top_k=30`.
+
+| Blind result | BF16 | Balanced V1 |
+|:--|--:|--:|
+| Speaker-likeness scores | all four 5/5 | all four 5/5 |
+| Mean reference-style likeness | 4.50 / 5 | 4.50 / 5 |
+| Mean naturalness | 3.75 / 5 | 3.75 / 5 |
+| Pair preference | 1 | 2 |
+
+The fourth pair was tied. The only severe artifact reported in the set was a
+deterministic BF16 pitch squeak in the long seed-42 sample; its V1 counterpart
+did not contain that excursion.
+
+This test suggests that the quant did not cause a detectable speaker-identity
+loss for that reference. It is **not** a general MOS study or broad cloning
+qualification: more listeners, speakers, accents, recording conditions, and
+languages are still needed. The release therefore reports the result without
+claiming parity in every voice-cloning setting.
+
+## Validated scope
+
+| Gate | Result |
+|:--|:--|
+| Native execution | 60 NVFP4 and 120 MXFP8 projections execute through native SM120 paths |
+| Standalone packaging | Fresh load from these shards, without BF16 source projections, matched a frozen 64-frame code canary bit exactly |
+| English automated gates | Passed fixed-input signal/spectral, ASR, speaker-embedding, short/control, and held-out long-termination screens |
+| Blind English clone identity | Limited four-pair test above: every BF16 and V1 sample scored 5/5 speaker likeness; preferences were V1 2, BF16 1, tie 1 |
+| Multilingual | Not qualified; use an MXFP8 or BF16 model when language coverage matters |
+| Hardware | NVIDIA Blackwell SM120 only in V1 |
+
+The blind result supports quantization fidelity for that English reference; it
+does not establish universal cloning quality across voices, recording
+conditions, accents, or languages.
+
+## Known limitations
+
+- Fish S2-Pro itself sounded substantially flatter and less expressive than
+  VoxCPM2 in our reference comparison. BF16 shared this behavior, so V1 does
+  not treat it as NVFP4-specific damage and does not claim to fix it.
+- Inline emotion/style instructions change output trajectories but did not
+  reliably repair the perceived flatness in the tested voice.
+- The objective speaker embedding saturated near 0.99 and failed to predict
+  human preference; human listening remains required for new voices.
+- One matched long BF16 sample produced a deterministic pitch squeak while its
+  V1 counterpart did not. This is evidence from one seed, not a claim that V1
+  is generally more artifact-free than BF16.
+- Long-form peak memory is materially higher than loaded memory. Do not market
+  V1 as a sub-6-GiB operational model for arbitrary request lengths.
+
+These bounded claims are why this package is **Balanced V1**, not an XPO3
+speed/quality/size release. Follow
+[`ajh-code`](https://huggingface.co/ajh-code) for the upcoming XPO3 version.
+
+## Validate the download
+
+```bash
+python validate_release.py
+```
+
+`MANIFEST.json` records the byte size and SHA-256 of every distributed file
+except itself. Validation also checks the safetensors index/header mapping,
+the 60/120 NVFP4/MXFP8 tensor counts, source pins, license/notice files, and
+runtime payload. Hashing the 6.8 GB package takes a little while.
+
+For an additional hash check every time the service loads:
+
+```bash
+TTS_VERIFY_CHECKSUMS=1 ./launch.sh
+```
+
+## License and attribution
+
+**Built with Fish Audio.** This derivative is governed by the
+[Fish Audio Research License](./LICENSE.md). Research and non-commercial use
+are permitted subject to its terms. **Commercial use requires a separate
+written license from Fish Audio; no commercial rights are granted by this
+repository.** See [`Notice`](./Notice) for the required attribution and exact
+change statement, and [`THIRD_PARTY_NOTICES.md`](./THIRD_PARTY_NOTICES.md) for
+runtime dependencies.
+
+Use only voices and recordings you have the right and consent to use.

THIRD_PARTY_NOTICES.md

@@ -0,0 +1,28 @@
+# Third-party notices
+
+## Fish Audio S2-Pro and Fish Speech
+
+The model weights, codec, tokenizer, configuration, and bundled Fish Speech
+source derive from Fish Audio materials. They are governed by the Fish Audio
+Research License in `LICENSE.md`. The required Fish attribution and the changes
+made for this derivative are recorded in `Notice`.
+
+## comfy-kitchen
+
+The runtime depends on `comfy-kitchen==0.2.22`, distributed separately under
+Apache-2.0. Project: <https://github.com/Comfy-Org/comfy-kitchen>.
+
+## fish-scales-ops
+
+The runtime builds pinned revision
+`a04c1b63b1a7a670840fb3e97a82c0dbe2a35ded` of `fish-scales-ops`, distributed
+separately under Apache-2.0. Project:
+<https://github.com/fishaudio/fish-scales-ops>.
+
+## Native NVFP4 helper kernels
+
+The source under `runtime/native/` was adapted from the Apache-2.0 Mage-VL
+NVFP4 runtime at revision `7a03467b90d6feff6bd196928dfe156bd173f36e`.
+Its Apache-2.0 license is included at `runtime/native/LICENSE`.
+
+No third-party project listed here endorses this derivative release.

chat_template.jinja

@@ -0,0 +1,85 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set content = message.content %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is defined and message.reasoning_content is not none %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in message.content %}
+                {%- set content = message.content.split('</think>')[-1].lstrip('\n') %}
+                {%- set reasoning_content = message.content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {%- if loop.last or (not loop.last and reasoning_content) %}
+                {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
+            {%- else %}
+                {{- '<|im_start|>' + message.role + '\n' + content }}
+            {%- endif %}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- endif %}
+{%- endif %}

client.py

@@ -0,0 +1,102 @@
+#!/usr/bin/env python
+"""Call a project Fish TTS API with a zero-shot voice or saved reference ID."""
+
+from __future__ import annotations
+
+import argparse
+import base64
+from pathlib import Path
+
+import requests
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--url", default="http://127.0.0.1:8080/v1/tts")
+    parser.add_argument("--text", required=True)
+    voice = parser.add_mutually_exclusive_group()
+    voice.add_argument("--reference-audio", type=Path)
+    voice.add_argument("--reference-id")
+    parser.add_argument("--reference-text")
+    parser.add_argument("--output", type=Path, default=Path("output.wav"))
+    parser.add_argument("--api-key")
+    parser.add_argument("--seed", type=int)
+    parser.add_argument("--max-new-tokens", type=int, default=1024)
+    parser.add_argument("--chunk-length", type=int, default=200)
+    parser.add_argument("--top-p", type=float, default=0.9)
+    parser.add_argument("--temperature", type=float, default=0.9)
+    parser.add_argument("--repetition-penalty", type=float, default=1.1)
+    parser.add_argument(
+        "--normalize",
+        action=argparse.BooleanOptionalAction,
+        default=True,
+        help="Enable or disable API loudness normalization",
+    )
+    parser.add_argument(
+        "--use-memory-cache",
+        choices=("on", "off"),
+        default="on",
+        help="Reuse cached reference codes/text, or force a fresh reference encode",
+    )
+    return parser.parse_args()
+
+
+def main() -> int:
+    args = parse_args()
+    references = []
+    if args.reference_audio is not None:
+        if not args.reference_audio.is_file():
+            raise SystemExit(f"Reference audio is missing: {args.reference_audio}")
+        if not args.reference_text or not args.reference_text.strip():
+            raise SystemExit("--reference-text is required with --reference-audio")
+        references.append(
+            {
+                "audio": base64.b64encode(args.reference_audio.read_bytes()).decode(
+                    "ascii"
+                ),
+                "text": args.reference_text,
+            }
+        )
+    elif args.reference_text:
+        raise SystemExit("--reference-text requires --reference-audio")
+
+    payload = {
+        "text": args.text,
+        "references": references,
+        "reference_id": args.reference_id,
+        "format": "wav",
+        "streaming": False,
+        "normalize": args.normalize,
+        "max_new_tokens": args.max_new_tokens,
+        "chunk_length": args.chunk_length,
+        "top_p": args.top_p,
+        "temperature": args.temperature,
+        "repetition_penalty": args.repetition_penalty,
+        "seed": args.seed,
+        "use_memory_cache": args.use_memory_cache,
+    }
+    headers = {"Accept": "audio/wav"}
+    if args.api_key:
+        headers["Authorization"] = f"Bearer {args.api_key}"
+    response = requests.post(
+        args.url,
+        json=payload,
+        headers=headers,
+        timeout=600,
+    )
+    if response.status_code != 200:
+        raise SystemExit(
+            f"TTS request failed ({response.status_code}): {response.text[:2000]}"
+        )
+    content_type = response.headers.get("content-type", "")
+    if "audio/" not in content_type and not response.content.startswith(b"RIFF"):
+        raise SystemExit(f"Unexpected response content type: {content_type}")
+    args.output.parent.mkdir(parents=True, exist_ok=True)
+    args.output.write_bytes(response.content)
+    print(f"audio={args.output}")
+    print(f"bytes={len(response.content)}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

compose.yaml

@@ -0,0 +1,24 @@
+services:
+  tts:
+    build:
+      context: .
+      dockerfile: Dockerfile
+    image: fish-audio-s2-pro-nvfp4-balanced:v1-cu130
+    init: true
+    gpus: all
+    shm_size: 8gb
+    ports:
+      - "${TTS_PORT:-8080}:8080"
+    volumes:
+      - .:/model:ro
+      - ./runtime-data:/data
+    environment:
+      TTS_HOST: 0.0.0.0
+      TTS_PORT: 8080
+      TTS_DEVICE: ${TTS_DEVICE:-cuda:0}
+      TTS_CACHE_LENGTH: ${TTS_CACHE_LENGTH:-3072}
+      TTS_RUNTIME_DATA: /data
+      TTS_VERIFY_CHECKSUMS: ${TTS_VERIFY_CHECKSUMS:-0}
+      TTS_API_KEY: ${TTS_API_KEY:-}
+      MAX_JOBS: ${MAX_JOBS:-2}
+    restart: unless-stopped

config.json

@@ -0,0 +1,80 @@
+{
+  "audio_decoder_config": {
+    "attention_o_bias": false,
+    "attention_qk_norm": false,
+    "attention_qkv_bias": false,
+    "audio_hidden_dim": 5120,
+    "dim": 2560,
+    "dropout": 0.0,
+    "head_dim": 128,
+    "initializer_range": 0.01976423537605237,
+    "intermediate_size": 9728,
+    "max_seq_len": 11,
+    "model_type": "fish_qwen3_audio_decoder",
+    "moe_intermediate_size": 768,
+    "n_head": 32,
+    "n_layer": 4,
+    "n_local_heads": 8,
+    "norm_eps": 1e-06,
+    "norm_topk_prob": true,
+    "num_codebooks": 10,
+    "num_experts": 1,
+    "num_experts_per_tok": 1,
+    "rope_base": 1000000,
+    "router_gamma": 0.001,
+    "text_dim": 2560,
+    "tie_word_embeddings": false,
+    "use_aux_loss_free": false,
+    "use_bfloat16": false,
+    "use_gradient_checkpointing": true,
+    "use_moe": false,
+    "vocab_size": 4096
+  },
+  "audio_pad_token_id": 151677,
+  "dtype": "bfloat16",
+  "eos_token_id": 151645,
+  "model_type": "fish_qwen3_omni",
+  "pad_token_id": 151669,
+  "semantic_end_token_id": 155773,
+  "semantic_start_token_id": 151678,
+  "fish_s2_quantization": {
+    "format": "mixed_nvfp4_mxfp8",
+    "profile": "balanced",
+    "release": "v1",
+    "policy": "w4a16_gate_up_middle30_mxfp8_rest",
+    "nvfp4_modules": 60,
+    "mxfp8_modules": 120,
+    "runtime": "bundled",
+    "hardware_family": "sm_120"
+  },
+  "text_config": {
+    "attention_o_bias": false,
+    "attention_qk_norm": true,
+    "attention_qkv_bias": false,
+    "audio_hidden_dim": 5120,
+    "dim": 2560,
+    "dropout": 0.0,
+    "head_dim": 128,
+    "initializer_range": 0.01976423537605237,
+    "intermediate_size": 9728,
+    "max_seq_len": 32768,
+    "model_type": "fish_qwen3",
+    "moe_intermediate_size": 768,
+    "n_head": 32,
+    "n_layer": 36,
+    "n_local_heads": 8,
+    "norm_eps": 1e-06,
+    "norm_topk_prob": true,
+    "num_experts": 1,
+    "num_experts_per_tok": 1,
+    "rope_base": 1000000,
+    "router_gamma": 0.001,
+    "tie_word_embeddings": true,
+    "use_aux_loss_free": false,
+    "use_bfloat16": false,
+    "use_gradient_checkpointing": true,
+    "use_moe": false,
+    "vocab_size": 155776
+  },
+  "transformers_version": "4.57.1"
+}

install.sh

@@ -0,0 +1,66 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+release_root="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+python_bin="${PYTHON_BIN:-python3.12}"
+venv_dir="${release_root}/.venv"
+build_dir="${release_root}/.build"
+scales_revision="a04c1b63b1a7a670840fb3e97a82c0dbe2a35ded"
+scales_dir="${build_dir}/fish-scales-ops-${scales_revision}"
+cuda_root="${CUDA_HOME:-/usr/local/cuda}"
+
+for command_name in "${python_bin}" git; do
+  if ! command -v "${command_name}" >/dev/null 2>&1; then
+    echo "Missing required command: ${command_name}" >&2
+    exit 1
+  fi
+done
+if [[ ! -x "${cuda_root}/bin/nvcc" ]]; then
+  echo "CUDA compiler not found at ${cuda_root}/bin/nvcc" >&2
+  exit 1
+fi
+
+mkdir -p "${build_dir}" "${release_root}/runtime-data/torch-extensions"
+if [[ ! -x "${venv_dir}/bin/python" ]]; then
+  "${python_bin}" -m venv "${venv_dir}"
+fi
+"${venv_dir}/bin/python" -m pip install --upgrade "pip>=25" "uv==0.12.3"
+uv_bin="${venv_dir}/bin/uv"
+
+"${uv_bin}" pip install --python "${venv_dir}/bin/python" \
+  "setuptools>=75" wheel ninja packaging \
+  "torch==2.11.0" "torchaudio==2.11.0" \
+  --torch-backend=cu130
+"${uv_bin}" pip install --python "${venv_dir}/bin/python" \
+  --overrides "${release_root}/runtime/torch-overrides.txt" \
+  --torch-backend=cu130 \
+  "${release_root}/vendor/fish-speech" \
+  -r "${release_root}/requirements.txt"
+
+if [[ ! -d "${scales_dir}/.git" ]]; then
+  if [[ -e "${scales_dir}" ]]; then
+    echo "${scales_dir} exists but is not the expected Git checkout; move it aside and retry." >&2
+    exit 1
+  fi
+  git init "${scales_dir}"
+  git -C "${scales_dir}" remote add origin https://github.com/fishaudio/fish-scales-ops.git
+  git -C "${scales_dir}" fetch --depth 1 origin "${scales_revision}"
+  git -C "${scales_dir}" checkout --detach FETCH_HEAD
+  git -C "${scales_dir}" submodule update --init --recursive --depth 1
+fi
+if [[ "$(git -C "${scales_dir}" rev-parse HEAD)" != "${scales_revision}" ]]; then
+  echo "Unexpected fish-scales-ops revision in ${scales_dir}" >&2
+  exit 1
+fi
+git -C "${scales_dir}" submodule update --init --recursive --depth 1
+
+CUTLASS_DIR="${scales_dir}/3rdparty/cutlass" \
+CUDA_HOME="${cuda_root}" \
+MAX_JOBS="${MAX_JOBS:-2}" \
+TORCH_CUDA_ARCH_LIST=12.0a \
+"${uv_bin}" pip install --python "${venv_dir}/bin/python" \
+  --no-build-isolation "${scales_dir}/python"
+
+"${venv_dir}/bin/python" -c \
+  "import comfy_kitchen, fish_scales_ops, fish_speech, torch; print('installed', torch.__version__, torch.version.cuda)"
+echo "Installation complete. Start the service with ./launch.sh"

launch.sh

@@ -0,0 +1,40 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+release_root="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+python_bin="${TTS_PYTHON:-${release_root}/.venv/bin/python}"
+listen_host="${TTS_HOST:-0.0.0.0}"
+listen_port="${TTS_PORT:-8080}"
+device_name="${TTS_DEVICE:-cuda:0}"
+cache_length="${TTS_CACHE_LENGTH:-3072}"
+runtime_data="${TTS_RUNTIME_DATA:-${release_root}/runtime-data}"
+
+if [[ ! -x "${python_bin}" ]]; then
+  echo "Python environment not found at ${python_bin}; run ./install.sh first." >&2
+  exit 1
+fi
+mkdir -p "${runtime_data}/references" "${runtime_data}/torch-extensions"
+
+export FISH_NVFP4_ROOT="${release_root}"
+export FISH_NVFP4_BUILD_ROOT="${runtime_data}/torch-extensions"
+export PYTHONPATH="${release_root}/runtime${PYTHONPATH:+:${PYTHONPATH}}"
+export TEMPORARILY_DISABLE_PROTOBUF_VERSION_CHECK=true
+export TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST:-12.0a}"
+
+server_args=(
+  "${release_root}/runtime/server.py"
+  --checkpoint "${release_root}"
+  --host "${listen_host}"
+  --port "${listen_port}"
+  --device "${device_name}"
+  --cache-length "${cache_length}"
+)
+if [[ "${TTS_VERIFY_CHECKSUMS:-0}" == "1" ]]; then
+  server_args+=(--verify-checksums)
+fi
+if [[ -n "${TTS_API_KEY:-}" ]]; then
+  server_args+=(--api-key "${TTS_API_KEY}")
+fi
+
+cd "${runtime_data}"
+exec "${python_bin}" "${server_args[@]}"

model.safetensors.index.json

@@ -0,0 +1,605 @@
+{
+  "metadata": {
+    "total_size": 4903514352
+  },
+  "weight_map": {
+    "embeddings.weight": "model-00001-of-00003.safetensors",
+    "codebook_embeddings.weight": "model-00001-of-00003.safetensors",
+    "layers.0.attention.wqkv.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.0.attention.wqkv.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.0.attention.wo.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.0.attention.wo.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.0.attention.q_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.0.attention.k_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.0.feed_forward.w1.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.0.feed_forward.w1.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.0.feed_forward.w3.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.0.feed_forward.w3.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.0.feed_forward.w2.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.0.feed_forward.w2.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.0.ffn_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.0.attention_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.1.attention.wqkv.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.1.attention.wqkv.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.1.attention.wo.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.1.attention.wo.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.1.attention.q_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.1.attention.k_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.1.feed_forward.w1.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.1.feed_forward.w1.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.1.feed_forward.w3.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.1.feed_forward.w3.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.1.feed_forward.w2.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.1.feed_forward.w2.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.1.ffn_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.1.attention_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.2.attention.wqkv.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.2.attention.wqkv.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.2.attention.wo.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.2.attention.wo.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.2.attention.q_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.2.attention.k_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.2.feed_forward.w1.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.2.feed_forward.w1.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.2.feed_forward.w3.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.2.feed_forward.w3.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.2.feed_forward.w2.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.2.feed_forward.w2.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.2.ffn_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.2.attention_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.3.attention.wqkv.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.3.attention.wqkv.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.3.attention.wo.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.3.attention.wo.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.3.attention.q_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.3.attention.k_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.3.feed_forward.w1.qdata": "model-00001-of-00003.safetensors",
+    "layers.3.feed_forward.w1.weight_block_scale": "model-00001-of-00003.safetensors",
+    "layers.3.feed_forward.w1.weight_scale": "model-00001-of-00003.safetensors",
+    "layers.3.feed_forward.w3.qdata": "model-00001-of-00003.safetensors",
+    "layers.3.feed_forward.w3.weight_block_scale": "model-00001-of-00003.safetensors",
+    "layers.3.feed_forward.w3.weight_scale": "model-00001-of-00003.safetensors",
+    "layers.3.feed_forward.w2.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.3.feed_forward.w2.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.3.ffn_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.3.attention_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.4.attention.wqkv.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.4.attention.wqkv.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.4.attention.wo.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.4.attention.wo.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.4.attention.q_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.4.attention.k_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.4.feed_forward.w1.qdata": "model-00001-of-00003.safetensors",
+    "layers.4.feed_forward.w1.weight_block_scale": "model-00001-of-00003.safetensors",
+    "layers.4.feed_forward.w1.weight_scale": "model-00001-of-00003.safetensors",
+    "layers.4.feed_forward.w3.qdata": "model-00001-of-00003.safetensors",
+    "layers.4.feed_forward.w3.weight_block_scale": "model-00001-of-00003.safetensors",
+    "layers.4.feed_forward.w3.weight_scale": "model-00001-of-00003.safetensors",
+    "layers.4.feed_forward.w2.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.4.feed_forward.w2.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.4.ffn_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.4.attention_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.5.attention.wqkv.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.5.attention.wqkv.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.5.attention.wo.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.5.attention.wo.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.5.attention.q_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.5.attention.k_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.5.feed_forward.w1.qdata": "model-00001-of-00003.safetensors",
+    "layers.5.feed_forward.w1.weight_block_scale": "model-00001-of-00003.safetensors",
+    "layers.5.feed_forward.w1.weight_scale": "model-00001-of-00003.safetensors",
+    "layers.5.feed_forward.w3.qdata": "model-00001-of-00003.safetensors",
+    "layers.5.feed_forward.w3.weight_block_scale": "model-00001-of-00003.safetensors",
+    "layers.5.feed_forward.w3.weight_scale": "model-00001-of-00003.safetensors",
+    "layers.5.feed_forward.w2.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.5.feed_forward.w2.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.5.ffn_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.5.attention_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.6.attention.wqkv.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.6.attention.wqkv.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.6.attention.wo.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.6.attention.wo.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.6.attention.q_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.6.attention.k_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.6.feed_forward.w1.qdata": "model-00001-of-00003.safetensors",
+    "layers.6.feed_forward.w1.weight_block_scale": "model-00001-of-00003.safetensors",
+    "layers.6.feed_forward.w1.weight_scale": "model-00001-of-00003.safetensors",
+    "layers.6.feed_forward.w3.qdata": "model-00001-of-00003.safetensors",
+    "layers.6.feed_forward.w3.weight_block_scale": "model-00001-of-00003.safetensors",
+    "layers.6.feed_forward.w3.weight_scale": "model-00001-of-00003.safetensors",
+    "layers.6.feed_forward.w2.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.6.feed_forward.w2.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.6.ffn_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.6.attention_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.7.attention.wqkv.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.7.attention.wqkv.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.7.attention.wo.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.7.attention.wo.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.7.attention.q_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.7.attention.k_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.7.feed_forward.w1.qdata": "model-00001-of-00003.safetensors",
+    "layers.7.feed_forward.w1.weight_block_scale": "model-00001-of-00003.safetensors",
+    "layers.7.feed_forward.w1.weight_scale": "model-00001-of-00003.safetensors",
+    "layers.7.feed_forward.w3.qdata": "model-00001-of-00003.safetensors",
+    "layers.7.feed_forward.w3.weight_block_scale": "model-00001-of-00003.safetensors",
+    "layers.7.feed_forward.w3.weight_scale": "model-00001-of-00003.safetensors",
+    "layers.7.feed_forward.w2.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.7.feed_forward.w2.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.7.ffn_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.7.attention_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.8.attention.wqkv.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.8.attention.wqkv.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.8.attention.wo.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.8.attention.wo.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.8.attention.q_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.8.attention.k_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.8.feed_forward.w1.qdata": "model-00001-of-00003.safetensors",
+    "layers.8.feed_forward.w1.weight_block_scale": "model-00001-of-00003.safetensors",
+    "layers.8.feed_forward.w1.weight_scale": "model-00001-of-00003.safetensors",
+    "layers.8.feed_forward.w3.qdata": "model-00001-of-00003.safetensors",
+    "layers.8.feed_forward.w3.weight_block_scale": "model-00001-of-00003.safetensors",
+    "layers.8.feed_forward.w3.weight_scale": "model-00001-of-00003.safetensors",
+    "layers.8.feed_forward.w2.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.8.feed_forward.w2.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.8.ffn_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.8.attention_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.9.attention.wqkv.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.9.attention.wqkv.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.9.attention.wo.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.9.attention.wo.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.9.attention.q_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.9.attention.k_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.9.feed_forward.w1.qdata": "model-00001-of-00003.safetensors",
+    "layers.9.feed_forward.w1.weight_block_scale": "model-00001-of-00003.safetensors",
+    "layers.9.feed_forward.w1.weight_scale": "model-00001-of-00003.safetensors",
+    "layers.9.feed_forward.w3.qdata": "model-00001-of-00003.safetensors",
+    "layers.9.feed_forward.w3.weight_block_scale": "model-00001-of-00003.safetensors",
+    "layers.9.feed_forward.w3.weight_scale": "model-00001-of-00003.safetensors",
+    "layers.9.feed_forward.w2.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.9.feed_forward.w2.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.9.ffn_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.9.attention_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.10.attention.wqkv.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.10.attention.wqkv.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.10.attention.wo.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.10.attention.wo.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.10.attention.q_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.10.attention.k_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.10.feed_forward.w1.qdata": "model-00001-of-00003.safetensors",
+    "layers.10.feed_forward.w1.weight_block_scale": "model-00001-of-00003.safetensors",
+    "layers.10.feed_forward.w1.weight_scale": "model-00001-of-00003.safetensors",
+    "layers.10.feed_forward.w3.qdata": "model-00001-of-00003.safetensors",
+    "layers.10.feed_forward.w3.weight_block_scale": "model-00001-of-00003.safetensors",
+    "layers.10.feed_forward.w3.weight_scale": "model-00001-of-00003.safetensors",
+    "layers.10.feed_forward.w2.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.10.feed_forward.w2.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.10.ffn_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.10.attention_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.11.attention.wqkv.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.11.attention.wqkv.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.11.attention.wo.weight_fp8": "model-00001-of-00003.safetensors",
+    "layers.11.attention.wo.weight_scale_storage": "model-00001-of-00003.safetensors",
+    "layers.11.attention.q_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.11.attention.k_norm.weight": "model-00001-of-00003.safetensors",
+    "layers.11.feed_forward.w1.qdata": "model-00002-of-00003.safetensors",
+    "layers.11.feed_forward.w1.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.11.feed_forward.w1.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.11.feed_forward.w3.qdata": "model-00002-of-00003.safetensors",
+    "layers.11.feed_forward.w3.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.11.feed_forward.w3.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.11.feed_forward.w2.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.11.feed_forward.w2.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.11.ffn_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.11.attention_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.12.attention.wqkv.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.12.attention.wqkv.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.12.attention.wo.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.12.attention.wo.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.12.attention.q_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.12.attention.k_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.12.feed_forward.w1.qdata": "model-00002-of-00003.safetensors",
+    "layers.12.feed_forward.w1.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.12.feed_forward.w1.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.12.feed_forward.w3.qdata": "model-00002-of-00003.safetensors",
+    "layers.12.feed_forward.w3.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.12.feed_forward.w3.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.12.feed_forward.w2.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.12.feed_forward.w2.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.12.ffn_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.12.attention_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.13.attention.wqkv.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.13.attention.wqkv.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.13.attention.wo.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.13.attention.wo.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.13.attention.q_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.13.attention.k_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.13.feed_forward.w1.qdata": "model-00002-of-00003.safetensors",
+    "layers.13.feed_forward.w1.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.13.feed_forward.w1.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.13.feed_forward.w3.qdata": "model-00002-of-00003.safetensors",
+    "layers.13.feed_forward.w3.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.13.feed_forward.w3.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.13.feed_forward.w2.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.13.feed_forward.w2.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.13.ffn_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.13.attention_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.14.attention.wqkv.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.14.attention.wqkv.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.14.attention.wo.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.14.attention.wo.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.14.attention.q_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.14.attention.k_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.14.feed_forward.w1.qdata": "model-00002-of-00003.safetensors",
+    "layers.14.feed_forward.w1.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.14.feed_forward.w1.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.14.feed_forward.w3.qdata": "model-00002-of-00003.safetensors",
+    "layers.14.feed_forward.w3.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.14.feed_forward.w3.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.14.feed_forward.w2.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.14.feed_forward.w2.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.14.ffn_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.14.attention_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.15.attention.wqkv.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.15.attention.wqkv.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.15.attention.wo.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.15.attention.wo.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.15.attention.q_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.15.attention.k_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.15.feed_forward.w1.qdata": "model-00002-of-00003.safetensors",
+    "layers.15.feed_forward.w1.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.15.feed_forward.w1.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.15.feed_forward.w3.qdata": "model-00002-of-00003.safetensors",
+    "layers.15.feed_forward.w3.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.15.feed_forward.w3.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.15.feed_forward.w2.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.15.feed_forward.w2.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.15.ffn_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.15.attention_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.16.attention.wqkv.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.16.attention.wqkv.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.16.attention.wo.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.16.attention.wo.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.16.attention.q_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.16.attention.k_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.16.feed_forward.w1.qdata": "model-00002-of-00003.safetensors",
+    "layers.16.feed_forward.w1.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.16.feed_forward.w1.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.16.feed_forward.w3.qdata": "model-00002-of-00003.safetensors",
+    "layers.16.feed_forward.w3.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.16.feed_forward.w3.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.16.feed_forward.w2.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.16.feed_forward.w2.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.16.ffn_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.16.attention_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.17.attention.wqkv.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.17.attention.wqkv.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.17.attention.wo.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.17.attention.wo.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.17.attention.q_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.17.attention.k_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.17.feed_forward.w1.qdata": "model-00002-of-00003.safetensors",
+    "layers.17.feed_forward.w1.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.17.feed_forward.w1.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.17.feed_forward.w3.qdata": "model-00002-of-00003.safetensors",
+    "layers.17.feed_forward.w3.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.17.feed_forward.w3.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.17.feed_forward.w2.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.17.feed_forward.w2.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.17.ffn_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.17.attention_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.18.attention.wqkv.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.18.attention.wqkv.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.18.attention.wo.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.18.attention.wo.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.18.attention.q_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.18.attention.k_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.18.feed_forward.w1.qdata": "model-00002-of-00003.safetensors",
+    "layers.18.feed_forward.w1.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.18.feed_forward.w1.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.18.feed_forward.w3.qdata": "model-00002-of-00003.safetensors",
+    "layers.18.feed_forward.w3.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.18.feed_forward.w3.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.18.feed_forward.w2.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.18.feed_forward.w2.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.18.ffn_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.18.attention_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.19.attention.wqkv.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.19.attention.wqkv.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.19.attention.wo.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.19.attention.wo.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.19.attention.q_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.19.attention.k_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.19.feed_forward.w1.qdata": "model-00002-of-00003.safetensors",
+    "layers.19.feed_forward.w1.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.19.feed_forward.w1.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.19.feed_forward.w3.qdata": "model-00002-of-00003.safetensors",
+    "layers.19.feed_forward.w3.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.19.feed_forward.w3.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.19.feed_forward.w2.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.19.feed_forward.w2.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.19.ffn_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.19.attention_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.20.attention.wqkv.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.20.attention.wqkv.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.20.attention.wo.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.20.attention.wo.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.20.attention.q_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.20.attention.k_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.20.feed_forward.w1.qdata": "model-00002-of-00003.safetensors",
+    "layers.20.feed_forward.w1.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.20.feed_forward.w1.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.20.feed_forward.w3.qdata": "model-00002-of-00003.safetensors",
+    "layers.20.feed_forward.w3.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.20.feed_forward.w3.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.20.feed_forward.w2.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.20.feed_forward.w2.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.20.ffn_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.20.attention_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.21.attention.wqkv.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.21.attention.wqkv.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.21.attention.wo.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.21.attention.wo.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.21.attention.q_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.21.attention.k_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.21.feed_forward.w1.qdata": "model-00002-of-00003.safetensors",
+    "layers.21.feed_forward.w1.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.21.feed_forward.w1.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.21.feed_forward.w3.qdata": "model-00002-of-00003.safetensors",
+    "layers.21.feed_forward.w3.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.21.feed_forward.w3.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.21.feed_forward.w2.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.21.feed_forward.w2.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.21.ffn_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.21.attention_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.22.attention.wqkv.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.22.attention.wqkv.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.22.attention.wo.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.22.attention.wo.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.22.attention.q_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.22.attention.k_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.22.feed_forward.w1.qdata": "model-00002-of-00003.safetensors",
+    "layers.22.feed_forward.w1.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.22.feed_forward.w1.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.22.feed_forward.w3.qdata": "model-00002-of-00003.safetensors",
+    "layers.22.feed_forward.w3.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.22.feed_forward.w3.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.22.feed_forward.w2.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.22.feed_forward.w2.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.22.ffn_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.22.attention_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.23.attention.wqkv.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.23.attention.wqkv.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.23.attention.wo.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.23.attention.wo.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.23.attention.q_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.23.attention.k_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.23.feed_forward.w1.qdata": "model-00002-of-00003.safetensors",
+    "layers.23.feed_forward.w1.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.23.feed_forward.w1.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.23.feed_forward.w3.qdata": "model-00002-of-00003.safetensors",
+    "layers.23.feed_forward.w3.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.23.feed_forward.w3.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.23.feed_forward.w2.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.23.feed_forward.w2.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.23.ffn_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.23.attention_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.24.attention.wqkv.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.24.attention.wqkv.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.24.attention.wo.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.24.attention.wo.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.24.attention.q_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.24.attention.k_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.24.feed_forward.w1.qdata": "model-00002-of-00003.safetensors",
+    "layers.24.feed_forward.w1.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.24.feed_forward.w1.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.24.feed_forward.w3.qdata": "model-00002-of-00003.safetensors",
+    "layers.24.feed_forward.w3.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.24.feed_forward.w3.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.24.feed_forward.w2.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.24.feed_forward.w2.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.24.ffn_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.24.attention_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.25.attention.wqkv.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.25.attention.wqkv.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.25.attention.wo.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.25.attention.wo.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.25.attention.q_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.25.attention.k_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.25.feed_forward.w1.qdata": "model-00002-of-00003.safetensors",
+    "layers.25.feed_forward.w1.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.25.feed_forward.w1.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.25.feed_forward.w3.qdata": "model-00002-of-00003.safetensors",
+    "layers.25.feed_forward.w3.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.25.feed_forward.w3.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.25.feed_forward.w2.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.25.feed_forward.w2.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.25.ffn_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.25.attention_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.26.attention.wqkv.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.26.attention.wqkv.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.26.attention.wo.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.26.attention.wo.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.26.attention.q_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.26.attention.k_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.26.feed_forward.w1.qdata": "model-00002-of-00003.safetensors",
+    "layers.26.feed_forward.w1.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.26.feed_forward.w1.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.26.feed_forward.w3.qdata": "model-00002-of-00003.safetensors",
+    "layers.26.feed_forward.w3.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.26.feed_forward.w3.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.26.feed_forward.w2.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.26.feed_forward.w2.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.26.ffn_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.26.attention_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.27.attention.wqkv.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.27.attention.wqkv.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.27.attention.wo.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.27.attention.wo.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.27.attention.q_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.27.attention.k_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.27.feed_forward.w1.qdata": "model-00002-of-00003.safetensors",
+    "layers.27.feed_forward.w1.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.27.feed_forward.w1.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.27.feed_forward.w3.qdata": "model-00002-of-00003.safetensors",
+    "layers.27.feed_forward.w3.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.27.feed_forward.w3.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.27.feed_forward.w2.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.27.feed_forward.w2.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.27.ffn_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.27.attention_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.28.attention.wqkv.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.28.attention.wqkv.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.28.attention.wo.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.28.attention.wo.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.28.attention.q_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.28.attention.k_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.28.feed_forward.w1.qdata": "model-00002-of-00003.safetensors",
+    "layers.28.feed_forward.w1.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.28.feed_forward.w1.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.28.feed_forward.w3.qdata": "model-00002-of-00003.safetensors",
+    "layers.28.feed_forward.w3.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.28.feed_forward.w3.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.28.feed_forward.w2.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.28.feed_forward.w2.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.28.ffn_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.28.attention_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.29.attention.wqkv.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.29.attention.wqkv.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.29.attention.wo.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.29.attention.wo.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.29.attention.q_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.29.attention.k_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.29.feed_forward.w1.qdata": "model-00002-of-00003.safetensors",
+    "layers.29.feed_forward.w1.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.29.feed_forward.w1.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.29.feed_forward.w3.qdata": "model-00002-of-00003.safetensors",
+    "layers.29.feed_forward.w3.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.29.feed_forward.w3.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.29.feed_forward.w2.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.29.feed_forward.w2.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.29.ffn_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.29.attention_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.30.attention.wqkv.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.30.attention.wqkv.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.30.attention.wo.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.30.attention.wo.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.30.attention.q_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.30.attention.k_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.30.feed_forward.w1.qdata": "model-00002-of-00003.safetensors",
+    "layers.30.feed_forward.w1.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.30.feed_forward.w1.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.30.feed_forward.w3.qdata": "model-00002-of-00003.safetensors",
+    "layers.30.feed_forward.w3.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.30.feed_forward.w3.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.30.feed_forward.w2.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.30.feed_forward.w2.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.30.ffn_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.30.attention_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.31.attention.wqkv.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.31.attention.wqkv.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.31.attention.wo.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.31.attention.wo.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.31.attention.q_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.31.attention.k_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.31.feed_forward.w1.qdata": "model-00002-of-00003.safetensors",
+    "layers.31.feed_forward.w1.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.31.feed_forward.w1.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.31.feed_forward.w3.qdata": "model-00002-of-00003.safetensors",
+    "layers.31.feed_forward.w3.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.31.feed_forward.w3.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.31.feed_forward.w2.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.31.feed_forward.w2.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.31.ffn_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.31.attention_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.32.attention.wqkv.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.32.attention.wqkv.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.32.attention.wo.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.32.attention.wo.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.32.attention.q_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.32.attention.k_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.32.feed_forward.w1.qdata": "model-00002-of-00003.safetensors",
+    "layers.32.feed_forward.w1.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.32.feed_forward.w1.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.32.feed_forward.w3.qdata": "model-00002-of-00003.safetensors",
+    "layers.32.feed_forward.w3.weight_block_scale": "model-00002-of-00003.safetensors",
+    "layers.32.feed_forward.w3.weight_scale": "model-00002-of-00003.safetensors",
+    "layers.32.feed_forward.w2.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.32.feed_forward.w2.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.32.ffn_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.32.attention_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.33.attention.wqkv.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.33.attention.wqkv.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.33.attention.wo.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.33.attention.wo.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.33.attention.q_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.33.attention.k_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.33.feed_forward.w1.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.33.feed_forward.w1.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.33.feed_forward.w3.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.33.feed_forward.w3.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.33.feed_forward.w2.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.33.feed_forward.w2.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.33.ffn_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.33.attention_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.34.attention.wqkv.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.34.attention.wqkv.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.34.attention.wo.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.34.attention.wo.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.34.attention.q_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.34.attention.k_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.34.feed_forward.w1.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.34.feed_forward.w1.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.34.feed_forward.w3.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.34.feed_forward.w3.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.34.feed_forward.w2.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.34.feed_forward.w2.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.34.ffn_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.34.attention_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.35.attention.wqkv.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.35.attention.wqkv.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.35.attention.wo.weight_fp8": "model-00002-of-00003.safetensors",
+    "layers.35.attention.wo.weight_scale_storage": "model-00002-of-00003.safetensors",
+    "layers.35.attention.q_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.35.attention.k_norm.weight": "model-00002-of-00003.safetensors",
+    "layers.35.feed_forward.w1.weight_fp8": "model-00003-of-00003.safetensors",
+    "layers.35.feed_forward.w1.weight_scale_storage": "model-00003-of-00003.safetensors",
+    "layers.35.feed_forward.w3.weight_fp8": "model-00003-of-00003.safetensors",
+    "layers.35.feed_forward.w3.weight_scale_storage": "model-00003-of-00003.safetensors",
+    "layers.35.feed_forward.w2.weight_fp8": "model-00003-of-00003.safetensors",
+    "layers.35.feed_forward.w2.weight_scale_storage": "model-00003-of-00003.safetensors",
+    "layers.35.ffn_norm.weight": "model-00003-of-00003.safetensors",
+    "layers.35.attention_norm.weight": "model-00003-of-00003.safetensors",
+    "norm.weight": "model-00003-of-00003.safetensors",
+    "fast_embeddings.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.0.attention.wqkv.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.0.attention.wo.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.0.feed_forward.w1.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.0.feed_forward.w3.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.0.feed_forward.w2.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.0.ffn_norm.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.0.attention_norm.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.1.attention.wqkv.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.1.attention.wo.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.1.feed_forward.w1.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.1.feed_forward.w3.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.1.feed_forward.w2.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.1.ffn_norm.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.1.attention_norm.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.2.attention.wqkv.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.2.attention.wo.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.2.feed_forward.w1.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.2.feed_forward.w3.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.2.feed_forward.w2.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.2.ffn_norm.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.2.attention_norm.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.3.attention.wqkv.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.3.attention.wo.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.3.feed_forward.w1.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.3.feed_forward.w3.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.3.feed_forward.w2.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.3.ffn_norm.weight": "model-00003-of-00003.safetensors",
+    "fast_layers.3.attention_norm.weight": "model-00003-of-00003.safetensors",
+    "fast_norm.weight": "model-00003-of-00003.safetensors",
+    "fast_output.weight": "model-00003-of-00003.safetensors"
+  }
+}

quantization.json

@@ -0,0 +1,2197 @@
+{
+  "schema_version": 1,
+  "format": "fish-s2-pro-project-local-nvfp4-mixed",
+  "status": "release-v1-english-qualified-single-reference-clone-pass",
+  "created_utc": "2026-08-27T04:51:14Z",
+  "base_model": "fishaudio/s2-pro@1de9996b6be38b745688de084d87a5633f714e4e",
+  "fish_speech_revision": "e5e292632cb11e7a27b2b7487f58f612bc101e13",
+  "fish_scales_ops_revision": "a04c1b63b1a7a670840fb3e97a82c0dbe2a35ded",
+  "xpo3_source_revision": "7a03467b90d6feff6bd196928dfe156bd173f36e",
+  "implementation_files": {
+    "runtime/experimental/nvfp4/checkpoint.py": {
+      "sha256": "41870d3effbcf4e5647ba62e107bbacf0db172e963b0cf14a9503349858b7114",
+      "bytes": 7342
+    },
+    "runtime/experimental/nvfp4/modules.py": {
+      "sha256": "275be371fed819985805cf496735c32e679779fe1ccf0096f26e948bd2a668f7",
+      "bytes": 70345
+    },
+    "runtime/experimental/fp8/modules.py": {
+      "sha256": "3c4b2f71d0c3f078ad5f295c6ada96a5eb4c9c63ea518177985bdd6112766c5a",
+      "bytes": 10873
+    },
+    "runtime/experimental/codec.py": {
+      "sha256": "5bf0c02babbfefe1a736f3e844f6a46aac18d9653ba7444fd324d19abaf9dbfe",
+      "bytes": 11733
+    },
+    "runtime/server.py": {
+      "sha256": "2975cb70ebc20828e39719306d1e36075141531ea457aa2b0b12f700dfea68d1",
+      "bytes": 5276
+    }
+  },
+  "torch": "2.11.0+cu130",
+  "compiled_cuda": "13.0",
+  "hardware_family": "sm_120",
+  "runtime_variant": "nvfp4-w4a16-gate-up-middle30-mxfp8-rest",
+  "policy": "w4a16_gate_up_middle30_mxfp8_rest",
+  "activation_scaling": {
+    "checkpoint_file": "activation-scaling-report.json",
+    "source_path": "outputs/kernels/s2-pro-nvfp4-ste-channel-scales-all36-english-profile-rtx5080-20260826.json",
+    "sha256": "20024a68428ee44e4aa1b8101cc6357310d7670ebdd2ab87e91eb4f8f54d43ef",
+    "bytes": 2657835,
+    "deployment": "folded into selected FFN norms and packed gate/up weights"
+  },
+  "qualified_sampling": {
+    "language": "English",
+    "temperature": 1.0,
+    "top_p": 0.85,
+    "top_k": 30
+  },
+  "coverage": {
+    "nvfp4_modules": 60,
+    "nvfp4_layers": [
+      3,
+      4,
+      5,
+      6,
+      7,
+      8,
+      9,
+      10,
+      11,
+      12,
+      13,
+      14,
+      15,
+      16,
+      17,
+      18,
+      19,
+      20,
+      21,
+      22,
+      23,
+      24,
+      25,
+      26,
+      27,
+      28,
+      29,
+      30,
+      31,
+      32
+    ],
+    "nvfp4_roles": [
+      "w1",
+      "w3"
+    ],
+    "mxfp8_modules": 120,
+    "slow_projection_modules": 180
+  },
+  "weight_execution": {
+    "nvfp4": "packed E2M1 W4A16 small-M kernel through M=1; native NVFP4 W4A4 above M=1",
+    "mxfp8": "native dynamic MXFP8 W8A8 GEMM",
+    "output_dtype": "BF16"
+  },
+  "bf16_exclusions": [
+    "text/codebook/fast embeddings and tied text output",
+    "fast transformer and fast output head",
+    "normalization (including folded channel scales)",
+    "RoPE, KV cache, and sampling",
+    "DAC codec"
+  ],
+  "conversion_seconds": 26.938446992018726,
+  "conversion": {
+    "policy": "w4a16_gate_up_middle30_mxfp8_rest",
+    "backend": "selective_nvfp4_weight_bf16_activation+mxfp8_rest",
+    "w4a16_max_m": 1,
+    "nvfp4_mlp_roles": [
+      "w1",
+      "w3"
+    ],
+    "nvfp4_layers": [
+      3,
+      4,
+      5,
+      6,
+      7,
+      8,
+      9,
+      10,
+      11,
+      12,
+      13,
+      14,
+      15,
+      16,
+      17,
+      18,
+      19,
+      20,
+      21,
+      22,
+      23,
+      24,
+      25,
+      26,
+      27,
+      28,
+      29,
+      30,
+      31,
+      32
+    ],
+    "activation_scaling": {
+      "sweep_report": "/workspace/outputs/kernels/s2-pro-nvfp4-ste-channel-scales-all36-english-profile-rtx5080-20260826.json",
+      "calibration_reports": [
+        "/workspace/outputs/calibration/english-profile-api-multiseed-20260826/en-short-01-seed17-bf16-history/slow-mlp-inputs.json",
+        "/workspace/outputs/calibration/english-profile-api-multiseed-20260826/en-medium-01-seed17-bf16-history/slow-mlp-inputs.json",
+        "/workspace/outputs/calibration/english-profile-controls-multiseed-20260826/en-control-01-seed7-bf16-history/slow-mlp-inputs.json"
+      ],
+      "calibration_history_sha256": [
+        "ee3b0d712b0f2c7ad84a0dc149c3aa21df7511d4c8c459a6def5e07d8920881f",
+        "3d50440208a7c8198d197deb91888139adf91582a1389157d845ea7be9346dca",
+        "4381e395febcf831e90190b48105ce98301f64d46e76a7d4372ff4b17879e486"
+      ],
+      "scale_clamp": 16.0,
+      "activation_statistic": "equal_corpus_geometric_mean_absmax",
+      "correction_tensors": null,
+      "layers": [
+        {
+          "layer": 3,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.34115383428238044,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.4666644036769867,
+          "scale_p50": 0.9958776235580444,
+          "scale_max": 5.142515182495117
+        },
+        {
+          "layer": 4,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.3441897165672957,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.212601438164711,
+          "scale_p50": 0.9981337785720825,
+          "scale_max": 4.799542427062988
+        },
+        {
+          "layer": 5,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.30929929335239237,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.27531394362449646,
+          "scale_p50": 0.975422739982605,
+          "scale_max": 10.06588363647461
+        },
+        {
+          "layer": 6,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.18018434671767158,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.1165572926402092,
+          "scale_p50": 0.9514962434768677,
+          "scale_max": 7.606848239898682
+        },
+        {
+          "layer": 7,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.06065399728329279,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.3172764480113983,
+          "scale_p50": 1.0104628801345825,
+          "scale_max": 3.054215431213379
+        },
+        {
+          "layer": 8,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.011250608751427471,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.1061466634273529,
+          "scale_p50": 1.0128974914550781,
+          "scale_max": 2.625473976135254
+        },
+        {
+          "layer": 9,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.0,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 1.0,
+          "scale_p50": 1.0,
+          "scale_max": 1.0
+        },
+        {
+          "layer": 10,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.007783796885644456,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.4347914755344391,
+          "scale_p50": 1.00908625125885,
+          "scale_max": 3.0566306114196777
+        },
+        {
+          "layer": 11,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.02843783886713991,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.39092525839805603,
+          "scale_p50": 1.0116372108459473,
+          "scale_max": 3.0836970806121826
+        },
+        {
+          "layer": 12,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.012714668301119225,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.4127214550971985,
+          "scale_p50": 1.0075856447219849,
+          "scale_max": 3.343895196914673
+        },
+        {
+          "layer": 13,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.013615598069565804,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.2628592550754547,
+          "scale_p50": 1.0010664463043213,
+          "scale_max": 3.645423173904419
+        },
+        {
+          "layer": 14,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.022818199391143934,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.2197420746088028,
+          "scale_p50": 0.9946508407592773,
+          "scale_max": 4.616183757781982
+        },
+        {
+          "layer": 15,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.010961721473621622,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.39072415232658386,
+          "scale_p50": 0.9946622848510742,
+          "scale_max": 2.7863168716430664
+        },
+        {
+          "layer": 16,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.08529829675949463,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.2035234123468399,
+          "scale_p50": 0.9768345355987549,
+          "scale_max": 2.824779987335205
+        },
+        {
+          "layer": 17,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.05795215068343107,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.37192386388778687,
+          "scale_p50": 0.9935908317565918,
+          "scale_max": 3.944956064224243
+        },
+        {
+          "layer": 18,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.1330481612253437,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.4043632745742798,
+          "scale_p50": 0.9705244302749634,
+          "scale_max": 5.344815731048584
+        },
+        {
+          "layer": 19,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.11661456631179568,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.2677452862262726,
+          "scale_p50": 0.9670252799987793,
+          "scale_max": 4.1772003173828125
+        },
+        {
+          "layer": 20,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.15586651036886956,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.3513035774230957,
+          "scale_p50": 1.0072706937789917,
+          "scale_max": 2.7189722061157227
+        },
+        {
+          "layer": 21,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.13147452299013052,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.3314710557460785,
+          "scale_p50": 1.0048860311508179,
+          "scale_max": 2.3973195552825928
+        },
+        {
+          "layer": 22,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.17074986711459594,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.30894899368286133,
+          "scale_p50": 1.019906997680664,
+          "scale_max": 2.589536428451538
+        },
+        {
+          "layer": 23,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.17914724568745244,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.08279751241207123,
+          "scale_p50": 1.0246715545654297,
+          "scale_max": 2.500920057296753
+        },
+        {
+          "layer": 24,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.2385892277628,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.30888521671295166,
+          "scale_p50": 1.0113917589187622,
+          "scale_max": 2.790740489959717
+        },
+        {
+          "layer": 25,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.18576108916475464,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.30071744322776794,
+          "scale_p50": 1.0063565969467163,
+          "scale_max": 2.702918291091919
+        },
+        {
+          "layer": 26,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.17304804845125255,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.2675870954990387,
+          "scale_p50": 1.0074219703674316,
+          "scale_max": 3.0248024463653564
+        },
+        {
+          "layer": 27,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.07856216820359907,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.4005119204521179,
+          "scale_p50": 1.0053906440734863,
+          "scale_max": 2.7383763790130615
+        },
+        {
+          "layer": 28,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.06775815208747848,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.42719921469688416,
+          "scale_p50": 1.0111608505249023,
+          "scale_max": 2.56790828704834
+        },
+        {
+          "layer": 29,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.07100470156359495,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.1727953404188156,
+          "scale_p50": 1.011107087135315,
+          "scale_max": 2.124830484390259
+        },
+        {
+          "layer": 30,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.07547034941098363,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.4036831259727478,
+          "scale_p50": 1.018500566482544,
+          "scale_max": 2.685163974761963
+        },
+        {
+          "layer": 31,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.08205942511115716,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.3752962648868561,
+          "scale_p50": 1.0237677097320557,
+          "scale_max": 2.359099864959717
+        },
+        {
+          "layer": 32,
+          "alpha": null,
+          "calibration_improvement_fraction": 0.10552935903626048,
+          "clip_ratio": 1.0,
+          "scale_mode": "explicit_channel",
+          "correction_rank": null,
+          "sparse_correction_channels": null,
+          "input_hadamard_block_size": 0,
+          "scale_min": 0.39887359738349915,
+          "scale_p50": 1.0268824100494385,
+          "scale_max": 2.298870801925659
+        }
+      ]
+    },
+    "modules": 180,
+    "projections": 180,
+    "parameters": 3633315840,
+    "nvfp4_parameters": 1494220800,
+    "mxfp8_parameters": 2139095040,
+    "low_rank_correction_parameters": 0,
+    "low_rank_correction_bytes": 0,
+    "sparse_correction_parameters": 0,
+    "sparse_correction_bytes": 0,
+    "correction_parameters": 0,
+    "correction_bytes": 0,
+    "theoretical_bf16_source_bytes": 7266631680,
+    "packed_weight_bytes": 3046441200,
+    "nvfp4_packed_weight_bytes": 840499440,
+    "mxfp8_packed_weight_bytes": 2205941760,
+    "probe_cosine_min": 0.9942060112953186,
+    "probe_cosine_mean": 0.9979391674200694,
+    "probe_cosine_max": 0.9994385242462158,
+    "records": [
+      {
+        "name": "layers.0.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9994018077850342,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.0.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9994044303894043,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.0.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9992644190788269,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.0.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9991974830627441,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.0.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9992890357971191,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.1.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9992879033088684,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.1.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9993317723274231,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.1.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9992480874061584,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.1.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9993327856063843,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.1.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9993032217025757,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.2.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.999262809753418,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.2.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9992883801460266,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.2.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9992047548294067,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.2.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9993103742599487,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.2.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9992554187774658,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.3.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9994385242462158,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.3.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9992822408676147,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.3.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9948533773422241,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.3.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9952349662780762,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.3.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9992232918739319,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.4.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9992364645004272,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.4.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9992576837539673,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.4.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9952384233474731,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.4.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9954122304916382,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.4.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9993067979812622,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.5.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9992440938949585,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.5.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9992841482162476,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.5.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9942060112953186,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.5.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9951051473617554,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.5.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9992941617965698,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.6.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9992526769638062,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.6.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9993124008178711,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.6.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9946960210800171,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.6.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9953465461730957,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.6.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9993586540222168,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.7.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9992846250534058,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.7.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9992576837539673,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.7.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9948426485061646,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.7.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9953813552856445,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.7.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9992576241493225,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.8.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9992561340332031,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.8.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9993311166763306,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.8.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9954076409339905,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.8.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9952717423439026,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.8.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9992897510528564,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.9.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9992669820785522,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.9.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9993153214454651,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.9.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.995424747467041,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.9.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9953783750534058,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.9.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.999322772026062,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.10.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9993197321891785,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.10.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9993413090705872,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.10.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9951403737068176,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.10.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9953807592391968,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.10.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9992862939834595,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.11.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9992955923080444,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.11.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.999306857585907,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.11.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9950716495513916,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.11.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9955061078071594,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.11.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9993455410003662,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.12.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.999247670173645,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.12.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9992870092391968,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.12.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9951099753379822,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.12.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9955266714096069,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.12.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.999378502368927,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.13.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9992713332176208,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.13.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9993178248405457,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.13.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9951012134552002,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.13.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9953491687774658,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.13.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9993041753768921,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.14.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9992854595184326,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.14.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9992244243621826,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.14.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9949911832809448,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.14.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9952331781387329,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.14.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9993098378181458,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.15.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9993874430656433,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.15.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9993578791618347,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.15.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9955209493637085,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.15.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9953669309616089,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.15.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9992756247520447,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.16.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9992870092391968,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.16.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9993147253990173,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.16.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9951826333999634,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.16.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9952670335769653,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.16.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9992488622665405,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.17.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9993287324905396,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.17.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9993382096290588,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.17.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9949835538864136,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.17.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9951303005218506,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.17.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9992471933364868,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.18.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9993221759796143,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.18.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9993113279342651,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.18.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9948270916938782,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.18.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9951784610748291,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.18.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9993153810501099,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.19.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9992598295211792,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.19.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9992833137512207,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.19.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9952045679092407,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.19.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9950501918792725,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.19.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9993191957473755,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.20.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9993005990982056,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.20.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.999289870262146,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.20.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9953472018241882,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.20.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9954727292060852,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.20.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9993296265602112,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.21.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9992555975914001,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.21.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9992495775222778,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.21.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9953234195709229,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.21.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9952938556671143,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.21.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9992718696594238,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.22.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9992914199829102,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.22.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9992764592170715,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.22.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9949873089790344,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.22.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9951849579811096,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.22.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9992700815200806,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.23.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.99925696849823,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.23.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9992924928665161,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.23.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9949401617050171,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.23.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.995291531085968,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.23.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9993242025375366,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.24.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9992597699165344,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.24.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.999344527721405,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.24.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9952665567398071,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.24.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9955209493637085,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.24.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9993101954460144,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.25.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9993680715560913,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.25.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9992563724517822,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.25.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9952845573425293,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.25.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9953001141548157,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.25.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9993116855621338,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.26.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.999308705329895,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.26.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9992237091064453,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.26.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9952062368392944,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.26.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9953439831733704,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.26.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9993381500244141,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.27.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9992940425872803,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.27.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9992112517356873,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.27.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.995395302772522,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.27.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9952636361122131,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.27.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9993325471878052,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.28.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9993436336517334,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.28.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9992949962615967,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.28.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9951224327087402,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.28.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9954068064689636,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.28.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9992546439170837,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.29.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9993007183074951,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.29.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9992722272872925,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.29.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9955875873565674,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.29.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9953714609146118,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.29.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9993128776550293,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.30.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9992680549621582,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.30.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9993020296096802,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.30.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9952709078788757,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.30.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9954016208648682,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.30.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.99930739402771,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.31.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9993473291397095,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.31.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9992945194244385,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.31.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.995145320892334,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.31.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9953712224960327,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.31.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9993300437927246,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.32.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.999280571937561,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.32.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9992582201957703,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.32.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9953033924102783,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.32.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9953725934028625,
+        "precision": "nvfp4_w4a16_through_m1_activation_scaled"
+      },
+      {
+        "name": "layers.32.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.999283492565155,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.33.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9993834495544434,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.33.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9992942810058594,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.33.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9992245435714722,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.33.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.999274492263794,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.33.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9992848038673401,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.34.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9992939233779907,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.34.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9992879629135132,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.34.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9993096590042114,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.34.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9992994070053101,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.34.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9992780685424805,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.35.attention.wqkv",
+        "in_features": 2560,
+        "out_features": 6144,
+        "parameters": 15728640,
+        "probe_cosine": 0.9993366599082947,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.35.attention.wo",
+        "in_features": 4096,
+        "out_features": 2560,
+        "parameters": 10485760,
+        "probe_cosine": 0.9992329478263855,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.35.feed_forward.w1",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9992811679840088,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.35.feed_forward.w3",
+        "in_features": 2560,
+        "out_features": 9728,
+        "parameters": 24903680,
+        "probe_cosine": 0.9993214011192322,
+        "precision": "mxfp8_w8a8"
+      },
+      {
+        "name": "layers.35.feed_forward.w2",
+        "in_features": 9728,
+        "out_features": 2560,
+        "parameters": 24903680,
+        "probe_cosine": 0.9992574453353882,
+        "precision": "mxfp8_w8a8"
+      }
+    ]
+  },
+  "packed_tensor_counts": {
+    "qdata": 60,
+    "weight_block_scale": 60,
+    "weight_scale": 60,
+    "weight_fp8": 120,
+    "weight_scale_storage": 120
+  },
+  "state_tensor_count": 598,
+  "state_payload_bytes": 4903514352,
+  "measured_runtime": {
+    "loaded_allocation_bytes": 5744721408,
+    "loaded_allocation_gib": 5.35,
+    "note": "Measured with compact codec/reference offload and a 3072-token cache; not checkpoint file size.",
+    "short_zero_shot_peak_gib_min": 5.607,
+    "short_zero_shot_peak_gib_max": 5.769,
+    "held_out_long_peak_gib": 7.561,
+    "rtx5080_short_control_rtf": "approximately 1.05-1.06",
+    "rtx5080_semantic_frames_per_second": "approximately 20.4-20.6",
+    "rtx5060ti_focused_zero_shot_rtf": "approximately 2.06",
+    "ordinary_segment_streaming": false
+  },
+  "qualification_evidence": {
+    "outputs/quality-suite/audio-signal-gate-nvfp4-middle30-gate-up-english-profile-validation-20260826.json": {
+      "sha256": "dead831243f3bbc8aa2f059158683433e317038ff2356459d186b39d2cd71a66",
+      "bytes": 67309
+    },
+    "outputs/quality-suite/audio-signal-gate-nvfp4-middle30-gate-up-english-profile-long-20260826.json": {
+      "sha256": "7604f81336f81ec43d41e0b6a168af6453b3a0fb504772e457c55c351ca7bdac",
+      "bytes": 9054
+    },
+    "outputs/quality-suite/audio-signal-gate-nvfp4-middle30-gate-up-english-profile-api-default-3seed-20260826.json": {
+      "sha256": "4ca0333e560ac8ae86bfefbf3e20fa8932a4e9e5a53755dd36ff80d9e0ebcb31",
+      "bytes": 74872
+    },
+    "outputs/quality-suite/audio-signal-gate-nvfp4-middle30-gate-up-english-profile-controls-long-3seed-20260826.json": {
+      "sha256": "f806dfe4294263f9f1d7418415fa610732541c1aaa37d010311e68e58855b188",
+      "bytes": 75055
+    },
+    "outputs/quality-suite/asr-nvfp4-middle30-english-profile-validation-long-20260826.json": {
+      "sha256": "4a914e3067249bb53527646a13f2f5b8671e8115945f1d639ec74f5aced18531",
+      "bytes": 7532
+    },
+    "outputs/zero-shot/nvfp4-middle30-english-focused-fixed-reference-long-5seed-20260826/objective-speaker-prosody.json": {
+      "sha256": "f706fb5d8e1f07abcecaf1b9a3756a3ce2fc0c4b5764170ec31a19173d2cf7d8",
+      "bytes": 25395
+    },
+    "outputs/zero-shot/nvfp4-middle30-english-focused-fixed-reference-long-5seed-20260826/objective-asr.json": {
+      "sha256": "94cf914803808154f5d5f23b82f50b726df97fff27ba3e4cce5cd1dfa2814e8d",
+      "bytes": 15655
+    }
+  },
+  "checksums": {
+    ".source": {
+      "sha256": "e5d3907e41ff60c4e50108a4fcc4b64de346d5a02aa45819211f0f364bfd8799",
+      "bytes": 240
+    },
+    "LICENSE.md": {
+      "sha256": "aa7d9206e9d710590987a3636934f643529c00cd490323594e6206aaa0c32d80",
+      "bytes": 10360
+    },
+    "README.md": {
+      "sha256": "c010e235b428b3a45f118df4e2eb9053278fbc85a5543deb542a60f5e239f18e",
+      "bytes": 10945
+    },
+    "activation-scaling-report.json": {
+      "sha256": "20024a68428ee44e4aa1b8101cc6357310d7670ebdd2ab87e91eb4f8f54d43ef",
+      "bytes": 2657835
+    },
+    "chat_template.jinja": {
+      "sha256": "87a2728cb8dc9fe424d624542f6060ec05a1d285ebbec578bb078900e33396b5",
+      "bytes": 4116
+    },
+    "codec.pth": {
+      "sha256": "74fc41c5a7151c6f350af8bd7e5d6e3accfcc7f3dfbfac23afd35af07052bb2f",
+      "bytes": 1871099728
+    },
+    "config.json": {
+      "sha256": "bd1b81bf140b1eac0a15fe75319f5552902ad3b9903d557ddada358aef743125",
+      "bytes": 2139
+    },
+    "model-00001-of-00003.safetensors": {
+      "sha256": "b304711a5a7a102dc27a49616048e72aca0c695ef4648e860a4b9b5f2c39d694",
+      "bytes": 1992555752
+    },
+    "model-00002-of-00003.safetensors": {
+      "sha256": "ab75ab15859d88276b88d9c4672f196489a71336fb062335f7833da8e35e277b",
+      "bytes": 1984565944
+    },
+    "model-00003-of-00003.safetensors": {
+      "sha256": "0187275d6a87c4879b24b54462d8b3f487318ab402213f27bc9cca017b7d3816",
+      "bytes": 926458160
+    },
+    "model.safetensors.index.json": {
+      "sha256": "25229032b7d5a846fb2f946c46a35229f79babc569dad1dbc577705d892b2fba",
+      "bytes": 48027
+    },
+    "special_tokens_map.json": {
+      "sha256": "c2ff18fde6e43b7408435bc8ed079af74531befba549358a97cbc59ce606bc6b",
+      "bytes": 101864
+    },
+    "tokenizer.json": {
+      "sha256": "f24e08099d45a8adf3f52f5f0b03276e433bb9d689bb15fcbcc48ce58744588b",
+      "bytes": 12217872
+    },
+    "tokenizer_config.json": {
+      "sha256": "b8d149343ae425b0da67e6708686aceb51be7815d9792f265fc12ff04d5e9856",
+      "bytes": 860832
+    }
+  },
+  "fresh_load_verification": {
+    "status": "passed",
+    "created_utc": "2026-08-27T04:52:43Z",
+    "report": "outputs/nvfp4-checkpoint-verification/20260827/fresh-load/verification.json",
+    "report_sha256": "b70642be85b6e908d8be45dd7cd59673d54f2213a100232b5190d9ba4568bd32",
+    "all_14_artifact_checksums_verified": true,
+    "bf16_source_projection_materialized": false,
+    "nvfp4_modules": 60,
+    "mxfp8_modules": 120,
+    "meta_tensors": 0,
+    "fixed_code_exact_match": true,
+    "fixed_code_sha256": "64799db24ec3c4564bd901b06f868c3968297974b591fe46aa32cc930d8ed466"
+  },
+  "human_listening_correction": {
+    "initial_unblinded_finding": "The standalone NVFP4 output sounded bland, monotone, robotic, flat, and insufficiently like the reference; explicit tags did not repair perceived expressiveness.",
+    "bf16_control_finding": "The pinned BF16 base shared the broad expressiveness and speaker-likeness weakness versus VoxCPM2, so flat delivery alone is not NVFP4-specific damage.",
+    "blind_gate_status": "single_listener_single_reference_english_clone_identity_pass",
+    "blind_gate_response": "20260827T082148Z-5e83e8bb",
+    "speaker_likeness": "all four BF16 and all four NVFP4 samples scored 5/5",
+    "preferences": {
+      "nvfp4": 2,
+      "bf16": 1,
+      "tie": 1
+    },
+    "mean_reference_style": {
+      "nvfp4": 4.5,
+      "bf16": 4.5
+    },
+    "mean_naturalness": {
+      "nvfp4": 3.75,
+      "bf16": 3.75
+    },
+    "limitation": "This does not qualify cloning across other speakers, languages, or recording conditions."
+  },
+  "scope_and_limitations": [
+    "English automated signal, spectral, ASR, speaker-embedding, and held-out long-termination gates passed.",
+    "A four-pair blind test with one listener and one consented English reference found no precision-specific speaker-identity loss versus BF16.",
+    "S2-Pro BF16 and this V1 derivative both sounded flatter and less expressive than VoxCPM2 in the evaluated reference comparison; V1 does not claim to fix that inherited limitation.",
+    "The objective speaker embedding saturated and did not predict human preference, so new voices still require listening evaluation.",
+    "This policy is not qualified for other languages; use MXFP8 or BF16 when multilingual coverage matters.",
+    "The runtime requires NVIDIA Blackwell SM120, CUDA 13.0, and the bundled/pinned software path.",
+    "Loaded allocation is 5.350 GiB, but held-out long-form peak reached 7.561 GiB.",
+    "The ordinary single-speaker API is not true low-latency semantic/audio streaming."
+  ],
+  "packaged_utc": "2026-08-27T08:41:54Z",
+  "release": {
+    "version": "1.0",
+    "channel": "v1",
+    "xpo3_release": false,
+    "scope": "English-only research and non-commercial release",
+    "repository_name": "ajh-code/Fish-Audio-S2-Pro-NVFP4-Balanced",
+    "self_contained_weights": true,
+    "display_name": "V1 · Fish Audio S2-Pro · NVFP4 Balanced",
+    "profile": "balanced"
+  },
+  "conversion_implementation_files": {
+    "experimental/nvfp4/checkpoint.py": {
+      "sha256": "41870d3effbcf4e5647ba62e107bbacf0db172e963b0cf14a9503349858b7114",
+      "bytes": 7342
+    },
+    "experimental/nvfp4/modules.py": {
+      "sha256": "c12801c376ea3c74027e0ff6672c078858249c7c19b14bcc434b3af895a4ebf7",
+      "bytes": 70176
+    },
+    "experimental/fp8/modules.py": {
+      "sha256": "3c4b2f71d0c3f078ad5f295c6ada96a5eb4c9c63ea518177985bdd6112766c5a",
+      "bytes": 10873
+    }
+  },
+  "license": {
+    "name": "Fish Audio Research License",
+    "commercial_use": "requires a separate written license from Fish Audio",
+    "notice_file": "Notice",
+    "built_with_fish_audio": true
+  }
+}

requirements.txt

@@ -0,0 +1,5 @@
+# Installed by install.sh after the pinned CUDA PyTorch wheels.
+comfy-kitchen==0.2.22
+huggingface-hub[cli]==0.34.4
+requests==2.34.2
+soundfile==0.14.0

runtime/experimental/fp8/__init__.py

@@ -0,0 +1,11 @@
+"""Project-local S2-Pro FP8 experiments."""
+
+from .modules import MXFP8Linear, convert_s2_pro_mxfp8, load_mxfp8_checkpoint
+from .server import launch_mxfp8_thread_safe_queue
+
+__all__ = [
+    "MXFP8Linear",
+    "convert_s2_pro_mxfp8",
+    "launch_mxfp8_thread_safe_queue",
+    "load_mxfp8_checkpoint",
+]

runtime/experimental/fp8/modules.py

@@ -0,0 +1,307 @@
+"""Model-local MXFP8 projection modules for S2-Pro inference research."""
+
+from __future__ import annotations
+
+import hashlib
+import json
+from dataclasses import asdict, dataclass
+from pathlib import Path
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+from safetensors.torch import load_file
+
+import fish_scales_ops as fso
+from fish_speech.models.text2semantic.llama import (
+    BaseModelArgs,
+    DualARTransformer,
+    precompute_freqs_cis,
+)
+from fish_speech.tokenizer import FishTokenizer
+
+
+@dataclass
+class ConversionRecord:
+    name: str
+    in_features: int
+    out_features: int
+    parameters: int
+    probe_cosine: float
+
+
+class MXFP8Linear(nn.Module):
+    """BF16-input linear using native 1x32 MXFP8 activation/weight GEMM."""
+
+    def __init__(
+        self,
+        weight_fp8: torch.Tensor,
+        weight_scale_storage: torch.Tensor,
+        *,
+        in_features: int,
+        out_features: int,
+    ) -> None:
+        super().__init__()
+        self.in_features = in_features
+        self.out_features = out_features
+        self.register_buffer("weight_fp8", weight_fp8)
+        # The SM120 kernel consumes K-major scales. Store their physical
+        # layout as a contiguous [K-block, N] tensor so safetensors can
+        # serialize it canonically; transpose restores the required view.
+        self.register_buffer("weight_scale_storage", weight_scale_storage)
+
+    @classmethod
+    @torch.inference_mode()
+    def from_linear(cls, linear: nn.Linear) -> "MXFP8Linear":
+        if linear.bias is not None:
+            raise ValueError("The initial S2-Pro MXFP8 path supports bias-free linears")
+        if linear.weight.device.type != "cuda":
+            raise ValueError("Quantize S2-Pro linears after moving them to CUDA")
+        if linear.weight.dtype != torch.bfloat16:
+            raise ValueError(f"Expected BF16 source weight, got {linear.weight.dtype}")
+        weight_fp8, weight_scale = fso.gemm.quantize_1x32_fp8(linear.weight)
+        return cls(
+            weight_fp8,
+            weight_scale.t().contiguous(),
+            in_features=linear.in_features,
+            out_features=linear.out_features,
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if x.shape[-1] != self.in_features:
+            raise ValueError(
+                f"Expected input width {self.in_features}, got {x.shape[-1]}"
+            )
+        prefix = x.shape[:-1]
+        x_2d = x.reshape(-1, self.in_features).contiguous()
+        x_fp8, x_scale = fso.gemm.quantize_1x32_fp8(x_2d)
+        output = fso.gemm.linear_mxfp8(
+            x_fp8,
+            self.weight_fp8,
+            x_scale,
+            self.weight_scale_storage.t(),
+        )
+        return output.reshape(*prefix, self.out_features)
+
+    def extra_repr(self) -> str:
+        return (
+            f"in_features={self.in_features}, out_features={self.out_features}, "
+            "weight=MXFP8_1x32, activation=dynamic_MXFP8_1x32, output=BF16"
+        )
+
+
+def _selected_slow_mlp(name: str, module: nn.Module) -> bool:
+    return (
+        isinstance(module, nn.Linear)
+        and name.startswith("layers.")
+        and ".feed_forward." in name
+        and name.rsplit(".", 1)[-1] in {"w1", "w2", "w3"}
+    )
+
+
+def _selected_slow_transformer(name: str, module: nn.Module) -> bool:
+    return isinstance(module, nn.Linear) and name.startswith("layers.")
+
+
+def _selected_fast_transformer(name: str, module: nn.Module) -> bool:
+    return isinstance(module, nn.Linear) and name.startswith("fast_layers.")
+
+
+def _selected_all_transformers(name: str, module: nn.Module) -> bool:
+    return _selected_slow_transformer(name, module) or _selected_fast_transformer(
+        name, module
+    )
+
+
+@torch.inference_mode()
+def convert_s2_pro_mxfp8(
+    model: nn.Module,
+    *,
+    policy: str = "slow_mlp",
+    probe_seed: int = 20260817,
+) -> dict:
+    """Replace selected S2-Pro projections without patching global linear APIs."""
+    selectors = {
+        "slow_mlp": (_selected_slow_mlp, 108),
+        "slow_transformer": (_selected_slow_transformer, 180),
+        "fast_transformer": (_selected_fast_transformer, 20),
+        "all_transformers": (_selected_all_transformers, 200),
+    }
+    if policy not in selectors:
+        raise ValueError(f"Unsupported initial MXFP8 policy: {policy}")
+
+    selector, expected_modules = selectors[policy]
+
+    candidates = [
+        (name, module)
+        for name, module in model.named_modules()
+        if selector(name, module)
+    ]
+    if len(candidates) != expected_modules:
+        raise RuntimeError(
+            f"Expected {expected_modules} {policy} projections, found {len(candidates)}"
+        )
+
+    records = []
+    generator = torch.Generator(device=candidates[0][1].weight.device)
+    generator.manual_seed(probe_seed)
+    for name, linear in candidates:
+        parent_name, attribute = name.rsplit(".", 1)
+        parent = model.get_submodule(parent_name)
+        replacement = MXFP8Linear.from_linear(linear)
+
+        probe = torch.randn(
+            1,
+            linear.in_features,
+            dtype=torch.bfloat16,
+            device=linear.weight.device,
+            generator=generator,
+        ) * 0.1
+        reference = F.linear(probe, linear.weight)
+        actual = replacement(probe)
+        probe_cosine = float(
+            F.cosine_similarity(
+                actual.float().flatten(), reference.float().flatten(), dim=0
+            ).item()
+        )
+        records.append(
+            ConversionRecord(
+                name=name,
+                in_features=linear.in_features,
+                out_features=linear.out_features,
+                parameters=linear.weight.numel(),
+                probe_cosine=probe_cosine,
+            )
+        )
+        setattr(parent, attribute, replacement)
+
+    torch.cuda.synchronize(candidates[0][1].weight.device)
+    serialized = [asdict(record) for record in records]
+    cosines = [record.probe_cosine for record in records]
+    return {
+        "policy": policy,
+        "modules": len(records),
+        "parameters": sum(record.parameters for record in records),
+        "theoretical_bf16_source_bytes": sum(
+            record.parameters * 2 for record in records
+        ),
+        "probe_cosine_min": min(cosines),
+        "probe_cosine_mean": sum(cosines) / len(cosines),
+        "probe_cosine_max": max(cosines),
+        "records": serialized,
+    }
+
+
+def _sha256(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as handle:
+        while chunk := handle.read(8 * 1024 * 1024):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def _install_empty_mxfp8_modules(model: nn.Module, policy: str) -> int:
+    selectors = {
+        "slow_transformer": (_selected_slow_transformer, 180),
+    }
+    if policy not in selectors:
+        raise ValueError(f"Unsupported artifact policy: {policy}")
+    selector, expected = selectors[policy]
+    names = [name for name, module in model.named_modules() if selector(name, module)]
+    if len(names) != expected:
+        raise RuntimeError(f"Expected {expected} artifact modules, found {len(names)}")
+    for name in names:
+        linear = model.get_submodule(name)
+        parent_name, attribute = name.rsplit(".", 1)
+        parent = model.get_submodule(parent_name)
+        replacement = MXFP8Linear(
+            torch.empty(
+                linear.out_features,
+                linear.in_features,
+                dtype=torch.float8_e4m3fn,
+                device="meta",
+            ),
+            torch.empty(
+                linear.in_features // 128,
+                linear.out_features,
+                dtype=torch.int32,
+                device="meta",
+            ),
+            in_features=linear.in_features,
+            out_features=linear.out_features,
+        )
+        setattr(parent, attribute, replacement)
+    return len(names)
+
+
+@torch.inference_mode()
+def load_mxfp8_checkpoint(
+    path: str | Path,
+    *,
+    device: str | torch.device = "cuda:0",
+    max_length: int = 4096,
+    verify_checksums: bool = False,
+) -> DualARTransformer:
+    """Load the canonical checkpoint without materializing BF16 FP8 sources."""
+    path = Path(path)
+    metadata = json.loads((path / "quantization.json").read_text())
+    if metadata["format"] != "fish-s2-pro-project-local-mxfp8":
+        raise ValueError(f"Unsupported checkpoint format: {metadata['format']}")
+    if torch.cuda.get_device_capability(device)[0] != 12:
+        raise RuntimeError("This MXFP8 artifact currently requires sm_120")
+    if verify_checksums:
+        for filename, record in metadata["checksums"].items():
+            file_path = path / filename
+            if file_path.stat().st_size != record["bytes"]:
+                raise RuntimeError(f"Size mismatch for {filename}")
+            if _sha256(file_path) != record["sha256"]:
+                raise RuntimeError(f"SHA256 mismatch for {filename}")
+
+    config = BaseModelArgs.from_pretrained(str(path))
+    config.max_seq_len = max_length
+    with torch.device("meta"):
+        model = DualARTransformer(config)
+    model.tokenizer = FishTokenizer.from_pretrained(path)
+    _install_empty_mxfp8_modules(model, metadata["policy"])
+
+    index_path = path / "model.safetensors.index.json"
+    if index_path.is_file():
+        index = json.loads(index_path.read_text())
+        shard_names = sorted(set(index["weight_map"].values()))
+    else:
+        shard_names = ["model.safetensors"]
+    expected_keys = set(model.state_dict())
+    loaded_keys: set[str] = set()
+    for shard_name in shard_names:
+        shard = load_file(path / shard_name, device="cpu")
+        unexpected = set(shard) - expected_keys
+        if unexpected:
+            raise RuntimeError(
+                f"Unexpected checkpoint tensors in {shard_name}: {sorted(unexpected)[:5]}"
+            )
+        model.load_state_dict(shard, strict=False, assign=True)
+        loaded_keys.update(shard)
+    missing = expected_keys - loaded_keys
+    if missing:
+        raise RuntimeError(f"Missing checkpoint tensors: {sorted(missing)[:5]}")
+
+    # These buffers are intentionally non-persistent and were meta tensors.
+    model.freqs_cis = precompute_freqs_cis(
+        config.max_seq_len,
+        config.head_dim,
+        config.rope_base,
+    )
+    model.causal_mask = torch.tril(
+        torch.ones(config.max_seq_len, config.max_seq_len, dtype=torch.bool)
+    )
+    model.fast_freqs_cis = precompute_freqs_cis(
+        config.num_codebooks,
+        config.fast_head_dim,
+        config.rope_base,
+    )
+    model = model.to(device=device).eval()
+    model.fixed_temperature = torch.tensor(0.7, device=device, dtype=torch.float)
+    model.fixed_top_p = torch.tensor(0.7, device=device, dtype=torch.float)
+    model.fixed_repetition_penalty = torch.tensor(1.5, device=device, dtype=torch.float)
+    model._cache_setup_done = False
+    return model

runtime/experimental/fp8/server.py

@@ -0,0 +1,109 @@
+"""Threaded Fish Speech inference queue for the canonical MXFP8 checkpoint."""
+
+from __future__ import annotations
+
+import queue
+import threading
+import traceback
+from pathlib import Path
+
+import torch
+from loguru import logger
+
+from fish_speech.models.text2semantic.inference import (
+    GenerateRequest,
+    WrappedGenerateResponse,
+    decode_one_token_ar,
+    generate_long,
+)
+
+from .modules import MXFP8Linear, load_mxfp8_checkpoint
+
+
+def launch_mxfp8_thread_safe_queue(
+    checkpoint_path: str | Path,
+    device: str | torch.device,
+    precision: torch.dtype,
+    compile: bool = False,
+    *,
+    max_length: int = 4096,
+    verify_checksums: bool = False,
+) -> queue.Queue:
+    """Load MXFP8 once and serialize Fish Speech generation through a queue."""
+
+    if precision is not torch.bfloat16:
+        raise ValueError("The qualified MXFP8 service requires BF16 exclusions")
+    if compile:
+        raise ValueError("torch.compile is not qualified for the MXFP8 API service")
+    if max_length < 1024:
+        raise ValueError("MXFP8 API cache length must be at least 1024")
+
+    input_queue: queue.Queue = queue.Queue()
+    init_event = threading.Event()
+    init_error: list[BaseException] = []
+
+    def worker() -> None:
+        try:
+            model = load_mxfp8_checkpoint(
+                checkpoint_path,
+                device=device,
+                max_length=max_length,
+                verify_checksums=verify_checksums,
+            )
+            with torch.device(device):
+                model.setup_caches(
+                    max_batch_size=1,
+                    max_seq_len=model.config.max_seq_len,
+                    dtype=torch.bfloat16,
+                )
+            model._cache_setup_done = True
+            module_count = sum(
+                isinstance(module, MXFP8Linear) for module in model.modules()
+            )
+            if module_count != 180:
+                raise RuntimeError(
+                    f"Expected 180 native MXFP8 projections, found {module_count}"
+                )
+            logger.info(
+                "Loaded qualified MXFP8 slow-transformer checkpoint with "
+                f"{module_count} native projections on {device}"
+            )
+        except BaseException as error:
+            init_error.append(error)
+            logger.error(traceback.format_exc())
+            init_event.set()
+            return
+
+        init_event.set()
+        while True:
+            item: GenerateRequest | None = input_queue.get()
+            if item is None:
+                break
+            response_queue = item.response_queue
+            try:
+                for chunk in generate_long(
+                    model=model,
+                    decode_one_token=decode_one_token_ar,
+                    **item.request,
+                ):
+                    response_queue.put(
+                        WrappedGenerateResponse(status="success", response=chunk)
+                    )
+            except Exception as error:
+                logger.error(traceback.format_exc())
+                response_queue.put(
+                    WrappedGenerateResponse(status="error", response=error)
+                )
+            finally:
+                if torch.cuda.is_available():
+                    torch.cuda.empty_cache()
+
+    threading.Thread(
+        target=worker,
+        name="fish-s2-pro-mxfp8-worker",
+        daemon=True,
+    ).start()
+    init_event.wait()
+    if init_error:
+        raise RuntimeError("MXFP8 model worker failed to initialize") from init_error[0]
+    return input_queue

runtime/experimental/nvfp4/__init__.py

@@ -0,0 +1,32 @@
+"""Project-local native NVFP4 experiments for Fish Audio S2-Pro."""
+
+from .checkpoint import load_mixed_nvfp4_checkpoint
+from .server import launch_mixed_nvfp4_thread_safe_queue
+
+from .modules import (
+    NVFP4FeedForward,
+    NVFP4Linear,
+    NVFP4TransformerBlock,
+    apply_nvfp4_rounding_checkpoint,
+    convert_s2_pro_nvfp4,
+    load_direct_w4a4_m1_extension,
+    load_rmsnorm_nvfp4_m1_extension,
+    load_smallm_nvfp4_extension,
+    load_silu_product_nvfp4_m1_extension,
+    quantize_nvfp4,
+)
+
+__all__ = [
+    "NVFP4FeedForward",
+    "NVFP4Linear",
+    "NVFP4TransformerBlock",
+    "apply_nvfp4_rounding_checkpoint",
+    "convert_s2_pro_nvfp4",
+    "load_mixed_nvfp4_checkpoint",
+    "launch_mixed_nvfp4_thread_safe_queue",
+    "load_direct_w4a4_m1_extension",
+    "load_rmsnorm_nvfp4_m1_extension",
+    "load_smallm_nvfp4_extension",
+    "load_silu_product_nvfp4_m1_extension",
+    "quantize_nvfp4",
+]

runtime/experimental/nvfp4/checkpoint.py

@@ -0,0 +1,203 @@
+"""Standalone loader for project-local mixed NVFP4/MXFP8 S2-Pro checkpoints."""
+
+from __future__ import annotations
+
+import hashlib
+import json
+from pathlib import Path
+from typing import Any
+
+import torch
+from safetensors.torch import load_file
+
+from fish_speech.models.text2semantic.llama import (
+    BaseModelArgs,
+    DualARTransformer,
+    precompute_freqs_cis,
+)
+from fish_speech.tokenizer import FishTokenizer
+
+from experimental.fp8 import MXFP8Linear
+from .modules import NVFP4Linear
+
+
+CHECKPOINT_FORMAT = "fish-s2-pro-project-local-nvfp4-mixed"
+
+
+def _sha256(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as handle:
+        while chunk := handle.read(8 * 1024 * 1024):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def _install_empty_projection(
+    model: DualARTransformer,
+    record: dict[str, Any],
+    *,
+    w4a16_max_m: int,
+) -> None:
+    name = str(record["name"])
+    in_features = int(record["in_features"])
+    out_features = int(record["out_features"])
+    precision = str(record["precision"])
+    original = model.get_submodule(name)
+    if not isinstance(original, torch.nn.Linear):
+        raise TypeError(f"Expected an unmodified Linear at {name}, got {type(original)}")
+    if (original.in_features, original.out_features) != (in_features, out_features):
+        raise ValueError(f"Checkpoint shape metadata does not match config at {name}")
+
+    if precision.startswith("nvfp4_w4a16_through_m"):
+        unsupported = (
+            "low_rank_corrected",
+            "sparse_channel_corrected",
+            "hadamard",
+        )
+        if any(marker in precision for marker in unsupported):
+            raise ValueError(
+                f"This standalone loader does not support corrected/rotated NVFP4: {name}"
+            )
+        replacement: torch.nn.Module = NVFP4Linear(
+            torch.empty(
+                out_features,
+                in_features // 2,
+                dtype=torch.uint8,
+                device="meta",
+            ),
+            torch.empty(
+                out_features,
+                in_features // 16,
+                dtype=torch.float8_e4m3fn,
+                device="meta",
+            ),
+            torch.empty((), dtype=torch.float32, device="meta"),
+            in_features=in_features,
+            out_features=out_features,
+            w4a16_max_m=w4a16_max_m,
+        )
+    elif precision == "mxfp8_w8a8":
+        replacement = MXFP8Linear(
+            torch.empty(
+                out_features,
+                in_features,
+                dtype=torch.float8_e4m3fn,
+                device="meta",
+            ),
+            torch.empty(
+                in_features // 128,
+                out_features,
+                dtype=torch.int32,
+                device="meta",
+            ),
+            in_features=in_features,
+            out_features=out_features,
+        )
+    else:
+        raise ValueError(f"Unsupported precision record for {name}: {precision}")
+
+    parent_name, attribute = name.rsplit(".", 1)
+    setattr(model.get_submodule(parent_name), attribute, replacement)
+
+
+@torch.inference_mode()
+def load_mixed_nvfp4_checkpoint(
+    path: str | Path,
+    *,
+    device: str | torch.device = "cuda:0",
+    max_length: int = 3072,
+    verify_checksums: bool = False,
+) -> DualARTransformer:
+    """Load a mixed checkpoint without materializing its BF16 source projections."""
+    path = Path(path)
+    metadata = json.loads((path / "quantization.json").read_text())
+    if metadata.get("format") != CHECKPOINT_FORMAT:
+        raise ValueError(f"Unsupported checkpoint format: {metadata.get('format')}")
+    device = torch.device(device)
+    if device.type != "cuda" or torch.cuda.get_device_capability(device)[0] != 12:
+        raise RuntimeError("This mixed NVFP4/MXFP8 artifact currently requires sm_120")
+    if verify_checksums:
+        for filename, record in metadata["checksums"].items():
+            file_path = path / filename
+            if file_path.stat().st_size != int(record["bytes"]):
+                raise RuntimeError(f"Size mismatch for {filename}")
+            if _sha256(file_path) != record["sha256"]:
+                raise RuntimeError(f"SHA256 mismatch for {filename}")
+
+    conversion = metadata["conversion"]
+    records = conversion["records"]
+    if len(records) != 180:
+        raise ValueError(f"Expected 180 mixed projection records, found {len(records)}")
+    if int(conversion["correction_parameters"]) != 0:
+        raise ValueError("This loader intentionally rejects correction-bearing artifacts")
+
+    config = BaseModelArgs.from_pretrained(str(path))
+    config.max_seq_len = max_length
+    with torch.device("meta"):
+        model = DualARTransformer(config)
+    model.tokenizer = FishTokenizer.from_pretrained(path)
+    for record in records:
+        _install_empty_projection(
+            model,
+            record,
+            w4a16_max_m=int(conversion["w4a16_max_m"]),
+        )
+
+    index_path = path / "model.safetensors.index.json"
+    if index_path.is_file():
+        index = json.loads(index_path.read_text())
+        shard_names = sorted(set(index["weight_map"].values()))
+    else:
+        shard_names = ["model.safetensors"]
+    expected_keys = set(model.state_dict())
+    loaded_keys: set[str] = set()
+    for shard_name in shard_names:
+        shard = load_file(path / shard_name, device="cpu")
+        unexpected = set(shard) - expected_keys
+        if unexpected:
+            raise RuntimeError(
+                f"Unexpected checkpoint tensors in {shard_name}: {sorted(unexpected)[:5]}"
+            )
+        model.load_state_dict(shard, strict=False, assign=True)
+        loaded_keys.update(shard)
+    missing = expected_keys - loaded_keys
+    if missing:
+        raise RuntimeError(f"Missing checkpoint tensors: {sorted(missing)[:5]}")
+
+    # These buffers are non-persistent, so reconstruct them after the meta load.
+    model.freqs_cis = precompute_freqs_cis(
+        config.max_seq_len,
+        config.head_dim,
+        config.rope_base,
+    )
+    model.causal_mask = torch.tril(
+        torch.ones(config.max_seq_len, config.max_seq_len, dtype=torch.bool)
+    )
+    model.fast_freqs_cis = precompute_freqs_cis(
+        config.num_codebooks,
+        config.fast_head_dim,
+        config.rope_base,
+    )
+    model = model.to(device=device).eval()
+    sampling = metadata.get("qualified_sampling", {})
+    model.fixed_temperature = torch.tensor(
+        sampling.get("temperature", 1.0), device=device, dtype=torch.float
+    )
+    model.fixed_top_p = torch.tensor(
+        sampling.get("top_p", 0.85), device=device, dtype=torch.float
+    )
+    model.fixed_repetition_penalty = torch.tensor(1.5, device=device, dtype=torch.float)
+    model._cache_setup_done = False
+
+    nvfp4_count = sum(isinstance(module, NVFP4Linear) for module in model.modules())
+    mxfp8_count = sum(isinstance(module, MXFP8Linear) for module in model.modules())
+    if (nvfp4_count, mxfp8_count) != (60, 120):
+        raise RuntimeError(
+            f"Expected 60 NVFP4 and 120 MXFP8 modules, got {nvfp4_count}/{mxfp8_count}"
+        )
+    meta_tensors = [
+        name for name, tensor in model.state_dict().items() if tensor.device.type == "meta"
+    ]
+    if meta_tensors:
+        raise RuntimeError(f"Checkpoint left meta tensors: {meta_tensors[:5]}")
+    return model

runtime/experimental/nvfp4/modules.py

@@ -0,0 +1,1761 @@
+"""Portable native NVFP4 modules for the S2-Pro NVFP4 V1 runtime."""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import os
+import sys
+from dataclasses import asdict, dataclass
+from functools import lru_cache
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+from torch.nn.attention import SDPBackend, sdpa_kernel
+
+from fish_speech.models.text2semantic.llama import apply_rotary_emb
+
+
+ROOT = Path(
+    os.environ.get("FISH_NVFP4_ROOT", Path(__file__).resolve().parents[3])
+).resolve()
+NATIVE_SOURCE_ROOT = Path(
+    os.environ.get(
+        "FISH_NVFP4_NATIVE_ROOT",
+        ROOT / "runtime" / "native",
+    )
+).resolve()
+BUILD_ROOT = Path(
+    os.environ.get(
+        "FISH_NVFP4_BUILD_ROOT",
+        ROOT / "runtime-data" / "torch-extensions",
+    )
+).resolve()
+COMFY_KITCHEN_ROOT = os.environ.get("COMFY_KITCHEN_ROOT")
+DIRECT_SOURCE_ROOT = NATIVE_SOURCE_ROOT / "direct_w4a4_m1"
+SILU_PRODUCT_SOURCE_ROOT = NATIVE_SOURCE_ROOT / "silu_product_nvfp4_m1"
+RMSNORM_SOURCE_ROOT = NATIVE_SOURCE_ROOT / "rmsnorm_nvfp4_m1"
+SMALLM_SOURCE_ROOT = NATIVE_SOURCE_ROOT / "smallm_gemv"
+DIRECT_SOURCE_FILES = (
+    "direct_w4a4_m1.cpp",
+    "direct_w4a4_m1.cu",
+    "direct_w4a4_m1.h",
+)
+SILU_PRODUCT_SOURCE_FILES = (
+    "silu_product_nvfp4_m1.cpp",
+    "silu_product_nvfp4_m1.cu",
+    "silu_product_nvfp4_m1.h",
+)
+RMSNORM_SOURCE_FILES = (
+    "rmsnorm_nvfp4_m1.cpp",
+    "rmsnorm_nvfp4_m1.cu",
+    "rmsnorm_nvfp4_m1.h",
+)
+SMALLM_SOURCE_FILES = (
+    "smallm_gemv.cpp",
+    "smallm_gemv.cu",
+    "smallm_gemv.h",
+)
+
+if COMFY_KITCHEN_ROOT and COMFY_KITCHEN_ROOT not in sys.path:
+    sys.path.insert(0, COMFY_KITCHEN_ROOT)
+
+from comfy_kitchen.tensor import QuantizedTensor, TensorCoreNVFP4Layout
+
+
+@dataclass
+class ConversionRecord:
+    name: str
+    in_features: int
+    out_features: int
+    parameters: int
+    probe_cosine: float
+
+
+@dataclass
+class FusedMLPConversionRecord:
+    name: str
+    in_features: int
+    intermediate_features: int
+    out_features: int
+    parameters: int
+    probe_cosine: float
+
+
+@dataclass
+class FusedTransformerConversionRecord:
+    name: str
+    parameters: int
+    wqkv_probe_cosine: float
+    wo_probe_cosine: float
+    mlp_probe_cosine: float
+
+
+def quantize_nvfp4(
+    value: torch.Tensor,
+    *,
+    scale: torch.Tensor | float | None = None,
+) -> QuantizedTensor:
+    """Apply the pinned tensor-wide dynamic NVFP4 policy."""
+    return QuantizedTensor.from_float(
+        value,
+        "TensorCoreNVFP4Layout",
+        scale=scale,
+    )
+
+
+@torch.inference_mode()
+def apply_nvfp4_rounding_checkpoint(
+    model: nn.Module,
+    checkpoint: Path | str,
+) -> dict[str, Any]:
+    """Apply a packed learned-rounding delta after ordinary model conversion."""
+
+    from safetensors.torch import load_file
+
+    checkpoint_path = Path(checkpoint)
+    if checkpoint_path.is_dir():
+        candidates = sorted(checkpoint_path.glob("*.safetensors"))
+        if len(candidates) != 1:
+            raise ValueError(
+                f"Expected one safetensors file in {checkpoint_path}, found {len(candidates)}"
+            )
+        checkpoint_path = candidates[0]
+    tensors = load_file(str(checkpoint_path), device=str(next(model.parameters()).device))
+    qdata_keys = sorted(key for key in tensors if key.endswith(".qdata"))
+    if not qdata_keys:
+        raise ValueError(f"No packed qdata entries in {checkpoint_path}")
+    records = []
+    for qdata_key in qdata_keys:
+        module_name = qdata_key.removesuffix(".qdata")
+        module = model.get_submodule(module_name)
+        if not isinstance(module, NVFP4Linear):
+            raise TypeError(f"Checkpoint target is not NVFP4Linear: {module_name}")
+        block_key = module_name + ".block_scale"
+        tensor_key = module_name + ".tensor_scale"
+        if block_key not in tensors or tensor_key not in tensors:
+            raise ValueError(f"Checkpoint lacks scales for {module_name}")
+        checkpoint_qdata = tensors[qdata_key]
+        checkpoint_block_scale = tensors[block_key]
+        checkpoint_tensor_scale = tensors[tensor_key]
+        if checkpoint_qdata.shape != module.qdata.shape:
+            raise ValueError(f"qdata shape mismatch for {module_name}")
+        if not torch.equal(checkpoint_block_scale, module.weight_block_scale):
+            raise ValueError(f"Block scale mismatch for {module_name}")
+        if not torch.equal(checkpoint_tensor_scale, module.weight_scale):
+            raise ValueError(f"Tensor scale mismatch for {module_name}")
+        changed_bytes = int((checkpoint_qdata != module.qdata).sum())
+        old_codes = torch.stack(
+            (module.qdata >> 4, module.qdata & 0x0F),
+            dim=-1,
+        )
+        new_codes = torch.stack(
+            (checkpoint_qdata >> 4, checkpoint_qdata & 0x0F),
+            dim=-1,
+        )
+        changed_weights = int((old_codes != new_codes).sum())
+        module.qdata.copy_(checkpoint_qdata)
+        records.append(
+            {
+                "module": module_name,
+                "changed_packed_bytes": changed_bytes,
+                "changed_weights": changed_weights,
+                "weights": module.out_features * module.in_features,
+            }
+        )
+    return {
+        "checkpoint": str(checkpoint_path),
+        "modules": len(records),
+        "changed_packed_bytes": sum(row["changed_packed_bytes"] for row in records),
+        "changed_weights": sum(row["changed_weights"] for row in records),
+        "records": records,
+    }
+
+
+def _hadamard_blocks(value: torch.Tensor, block_size: int) -> torch.Tensor:
+    """Materialized orthonormal block Hadamard for quality prototypes."""
+    if block_size < 2 or block_size & (block_size - 1):
+        raise ValueError(f"Hadamard block size must be a power of two: {block_size}")
+    if value.shape[-1] % block_size:
+        raise ValueError(
+            f"Width {value.shape[-1]} is not divisible by block size {block_size}"
+        )
+    original_dtype = value.dtype
+    transformed = value.float().reshape(*value.shape[:-1], -1, block_size)
+    stride = 1
+    while stride < block_size:
+        pairs = transformed.reshape(
+            *transformed.shape[:-1],
+            block_size // (2 * stride),
+            2,
+            stride,
+        )
+        left = pairs[..., 0, :]
+        right = pairs[..., 1, :]
+        transformed = torch.stack((left + right, left - right), dim=-2).reshape(
+            *transformed.shape
+        )
+        stride *= 2
+    return (transformed.reshape_as(value) / block_size**0.5).to(original_dtype)
+
+
+def _direct_source_hash() -> str:
+    digest = hashlib.sha256()
+    for filename in DIRECT_SOURCE_FILES:
+        digest.update((DIRECT_SOURCE_ROOT / filename).read_bytes())
+    return digest.hexdigest()
+
+
+def _source_hash(source_root: Path, filenames: tuple[str, ...]) -> str:
+    digest = hashlib.sha256()
+    for filename in filenames:
+        digest.update((source_root / filename).read_bytes())
+    return digest.hexdigest()
+
+
+@lru_cache(maxsize=1)
+def load_direct_w4a4_m1_extension(*, verbose: bool = False) -> Any:
+    """Build the pinned direct packed-NVFP4 M=1 primitive locally."""
+    from torch.utils.cpp_extension import load
+
+    missing = [
+        filename
+        for filename in DIRECT_SOURCE_FILES
+        if not (DIRECT_SOURCE_ROOT / filename).is_file()
+    ]
+    if missing:
+        raise RuntimeError(f"Missing pinned NVFP4 native sources: {missing}")
+    build_root = BUILD_ROOT / "direct_w4a4_m1"
+    build_root.mkdir(parents=True, exist_ok=True)
+    os.environ.setdefault("TORCH_CUDA_ARCH_LIST", "12.0")
+    os.environ.setdefault("MAX_JOBS", "2")
+    return load(
+        name=f"s2_pro_direct_w4a4_m1_{_direct_source_hash()[:12]}",
+        sources=[
+            str(DIRECT_SOURCE_ROOT / "direct_w4a4_m1.cpp"),
+            str(DIRECT_SOURCE_ROOT / "direct_w4a4_m1.cu"),
+        ],
+        extra_cflags=["-O3"],
+        extra_cuda_cflags=["-O3", "--use_fast_math", "-lineinfo"],
+        extra_include_paths=[str(DIRECT_SOURCE_ROOT)],
+        build_directory=str(build_root),
+        with_cuda=True,
+        verbose=verbose,
+        is_python_module=True,
+    )
+
+
+@lru_cache(maxsize=1)
+def load_silu_product_nvfp4_m1_extension(*, verbose: bool = False) -> Any:
+    """Build the pinned fused SiLU/product-to-NVFP4 M=1 primitive."""
+    from torch.utils.cpp_extension import load
+
+    missing = [
+        filename
+        for filename in SILU_PRODUCT_SOURCE_FILES
+        if not (SILU_PRODUCT_SOURCE_ROOT / filename).is_file()
+    ]
+    if missing:
+        raise RuntimeError(f"Missing pinned NVFP4 SiLU sources: {missing}")
+    build_root = BUILD_ROOT / "silu_product_nvfp4_m1"
+    build_root.mkdir(parents=True, exist_ok=True)
+    os.environ.setdefault("TORCH_CUDA_ARCH_LIST", "12.0")
+    os.environ.setdefault("MAX_JOBS", "2")
+    return load(
+        name=(
+            "s2_pro_silu_product_nvfp4_m1_"
+            f"{_source_hash(SILU_PRODUCT_SOURCE_ROOT, SILU_PRODUCT_SOURCE_FILES)[:12]}"
+        ),
+        sources=[
+            str(SILU_PRODUCT_SOURCE_ROOT / "silu_product_nvfp4_m1.cpp"),
+            str(SILU_PRODUCT_SOURCE_ROOT / "silu_product_nvfp4_m1.cu"),
+        ],
+        extra_cflags=["-O3"],
+        extra_cuda_cflags=["-O3", "--use_fast_math", "-lineinfo"],
+        extra_include_paths=[str(SILU_PRODUCT_SOURCE_ROOT)],
+        build_directory=str(build_root),
+        with_cuda=True,
+        verbose=verbose,
+        is_python_module=True,
+    )
+
+
+@lru_cache(maxsize=1)
+def load_rmsnorm_nvfp4_m1_extension(*, verbose: bool = False) -> Any:
+    """Build the pinned fused Fish-compatible RMSNorm-to-NVFP4 M=1 primitive."""
+    from torch.utils.cpp_extension import load
+
+    missing = [
+        filename
+        for filename in RMSNORM_SOURCE_FILES
+        if not (RMSNORM_SOURCE_ROOT / filename).is_file()
+    ]
+    if missing:
+        raise RuntimeError(f"Missing pinned NVFP4 RMSNorm sources: {missing}")
+    build_root = BUILD_ROOT / "rmsnorm_nvfp4_m1"
+    build_root.mkdir(parents=True, exist_ok=True)
+    os.environ.setdefault("TORCH_CUDA_ARCH_LIST", "12.0")
+    os.environ.setdefault("MAX_JOBS", "2")
+    return load(
+        name=(
+            "s2_pro_rmsnorm_nvfp4_m1_"
+            f"{_source_hash(RMSNORM_SOURCE_ROOT, RMSNORM_SOURCE_FILES)[:12]}"
+        ),
+        sources=[
+            str(RMSNORM_SOURCE_ROOT / "rmsnorm_nvfp4_m1.cpp"),
+            str(RMSNORM_SOURCE_ROOT / "rmsnorm_nvfp4_m1.cu"),
+        ],
+        extra_cflags=["-O3"],
+        extra_cuda_cflags=["-O3", "--use_fast_math", "-lineinfo"],
+        extra_include_paths=[str(RMSNORM_SOURCE_ROOT)],
+        build_directory=str(build_root),
+        with_cuda=True,
+        verbose=verbose,
+        is_python_module=True,
+    )
+
+
+@lru_cache(maxsize=1)
+def load_smallm_nvfp4_extension(*, verbose: bool = False) -> Any:
+    """Build the pinned packed-weight W4A16 small-M GEMV primitive."""
+    from torch.utils.cpp_extension import load
+
+    missing = [
+        filename
+        for filename in SMALLM_SOURCE_FILES
+        if not (SMALLM_SOURCE_ROOT / filename).is_file()
+    ]
+    if missing:
+        raise RuntimeError(f"Missing pinned NVFP4 small-M sources: {missing}")
+    build_root = BUILD_ROOT / "smallm_gemv"
+    build_root.mkdir(parents=True, exist_ok=True)
+    os.environ.setdefault("TORCH_CUDA_ARCH_LIST", "12.0")
+    os.environ.setdefault("MAX_JOBS", "2")
+    return load(
+        name=(
+            "s2_pro_smallm_nvfp4_"
+            f"{_source_hash(SMALLM_SOURCE_ROOT, SMALLM_SOURCE_FILES)[:12]}"
+        ),
+        sources=[
+            str(SMALLM_SOURCE_ROOT / "smallm_gemv.cpp"),
+            str(SMALLM_SOURCE_ROOT / "smallm_gemv.cu"),
+        ],
+        extra_cflags=["-O3"],
+        extra_cuda_cflags=["-O3", "--use_fast_math", "-lineinfo"],
+        extra_include_paths=[str(SMALLM_SOURCE_ROOT)],
+        build_directory=str(build_root),
+        with_cuda=True,
+        verbose=verbose,
+        is_python_module=True,
+    )
+
+
+class NVFP4Linear(nn.Module):
+    """BF16-input linear using a native packed NVFP4 weight with BF16 output."""
+
+    def __init__(
+        self,
+        qdata: torch.Tensor,
+        weight_block_scale: torch.Tensor,
+        weight_scale: torch.Tensor,
+        *,
+        in_features: int,
+        out_features: int,
+        direct_m1: bool = False,
+        w4a16_max_m: int = 0,
+        correction_down: torch.Tensor | None = None,
+        correction_up: torch.Tensor | None = None,
+        sparse_correction_indices: torch.Tensor | None = None,
+        sparse_correction_weight: torch.Tensor | None = None,
+        input_hadamard_block_size: int = 0,
+    ) -> None:
+        super().__init__()
+        if w4a16_max_m < 0:
+            raise ValueError("w4a16_max_m must be nonnegative")
+        if direct_m1 and w4a16_max_m:
+            raise ValueError("Choose only one NVFP4 M=1 backend")
+        self.in_features = int(in_features)
+        self.out_features = int(out_features)
+        self.direct_m1 = bool(direct_m1)
+        self.w4a16_max_m = int(w4a16_max_m)
+        if input_hadamard_block_size and (
+            input_hadamard_block_size < 2
+            or input_hadamard_block_size & (input_hadamard_block_size - 1)
+            or self.in_features % input_hadamard_block_size
+        ):
+            raise ValueError(
+                "Input Hadamard block size must be a power of two that divides "
+                f"the input width, got {input_hadamard_block_size}"
+            )
+        self.input_hadamard_block_size = int(input_hadamard_block_size)
+        self.register_buffer("qdata", qdata)
+        self.register_buffer("weight_block_scale", weight_block_scale)
+        self.register_buffer("weight_scale", weight_scale)
+        if (correction_down is None) != (correction_up is None):
+            raise ValueError("Low-rank correction requires both down and up factors")
+        if correction_down is not None:
+            if correction_down.dim() != 2 or correction_up.dim() != 2:
+                raise ValueError("Low-rank correction factors must be 2D")
+            if correction_down.shape[1] != self.in_features:
+                raise ValueError("Low-rank down factor input width does not match")
+            if correction_up.shape != (self.out_features, correction_down.shape[0]):
+                raise ValueError("Low-rank up factor shape does not match")
+        self.register_buffer("correction_down", correction_down)
+        self.register_buffer("correction_up", correction_up)
+        if (sparse_correction_indices is None) != (sparse_correction_weight is None):
+            raise ValueError(
+                "Sparse correction requires both channel indices and a weight"
+            )
+        if sparse_correction_indices is not None:
+            if (
+                sparse_correction_indices.dim() != 1
+                or sparse_correction_indices.dtype != torch.int64
+            ):
+                raise ValueError("Sparse correction indices must be 1D int64")
+            if sparse_correction_weight.shape != (
+                self.out_features,
+                sparse_correction_indices.numel(),
+            ):
+                raise ValueError("Sparse correction weight shape does not match")
+            if sparse_correction_weight.dtype != torch.bfloat16:
+                raise ValueError("Sparse correction weight must be BF16")
+            if sparse_correction_indices.numel() and (
+                int(sparse_correction_indices.min()) < 0
+                or int(sparse_correction_indices.max()) >= self.in_features
+            ):
+                raise ValueError("Sparse correction channel index is out of range")
+        self.register_buffer("sparse_correction_indices", sparse_correction_indices)
+        self.register_buffer("sparse_correction_weight", sparse_correction_weight)
+
+    @classmethod
+    @torch.inference_mode()
+    def from_weight(
+        cls,
+        weight: torch.Tensor,
+        *,
+        direct_m1: bool = False,
+        w4a16_max_m: int = 0,
+        quantization_scale: torch.Tensor | float | None = None,
+        correction_down: torch.Tensor | None = None,
+        correction_up: torch.Tensor | None = None,
+        sparse_correction_indices: torch.Tensor | None = None,
+        sparse_correction_weight: torch.Tensor | None = None,
+        input_hadamard_block_size: int = 0,
+    ) -> "NVFP4Linear":
+        if weight.device.type != "cuda":
+            raise ValueError("Quantize S2-Pro weights after moving them to CUDA")
+        if weight.dtype != torch.bfloat16 or weight.dim() != 2:
+            raise ValueError(f"Expected a 2D BF16 source weight, got {weight.dtype} {weight.shape}")
+        packed = quantize_nvfp4(
+            weight.contiguous(),
+            scale=quantization_scale,
+        )
+        return cls(
+            packed._qdata,
+            packed._params.block_scale,
+            packed._params.scale,
+            in_features=weight.shape[1],
+            out_features=weight.shape[0],
+            direct_m1=direct_m1,
+            w4a16_max_m=w4a16_max_m,
+            correction_down=correction_down,
+            correction_up=correction_up,
+            sparse_correction_indices=sparse_correction_indices,
+            sparse_correction_weight=sparse_correction_weight,
+            input_hadamard_block_size=input_hadamard_block_size,
+        )
+
+    @classmethod
+    @torch.inference_mode()
+    def from_linear(
+        cls,
+        linear: nn.Linear,
+        *,
+        direct_m1: bool = False,
+        w4a16_max_m: int = 0,
+    ) -> "NVFP4Linear":
+        if linear.bias is not None:
+            raise ValueError("The initial S2-Pro NVFP4 path supports bias-free linears")
+        if linear.weight.device.type != "cuda":
+            raise ValueError("Quantize S2-Pro linears after moving them to CUDA")
+        if linear.weight.dtype != torch.bfloat16:
+            raise ValueError(f"Expected BF16 source weight, got {linear.weight.dtype}")
+        return cls.from_weight(
+            linear.weight,
+            direct_m1=direct_m1,
+            w4a16_max_m=w4a16_max_m,
+        )
+
+    def _weight_quantized_tensor(self) -> QuantizedTensor:
+        params = TensorCoreNVFP4Layout.Params(
+            scale=self.weight_scale,
+            orig_dtype=torch.bfloat16,
+            orig_shape=(self.out_features, self.in_features),
+            block_scale=self.weight_block_scale,
+        )
+        return QuantizedTensor(self.qdata, "TensorCoreNVFP4Layout", params)
+
+    def _direct(self, activation: QuantizedTensor) -> torch.Tensor:
+        params = activation._params
+        return load_direct_w4a4_m1_extension().linear(
+            activation._qdata,
+            params.block_scale,
+            params.scale,
+            self.qdata,
+            self.weight_block_scale,
+            self.weight_scale,
+            None,
+        )
+
+    def project_packed_m1(
+        self,
+        qdata: torch.Tensor,
+        block_scale: torch.Tensor,
+        tensor_scale: torch.Tensor,
+    ) -> torch.Tensor:
+        """Project an already packed logical M=1 activation without wrappers."""
+        return load_direct_w4a4_m1_extension().linear(
+            qdata,
+            block_scale,
+            tensor_scale,
+            self.qdata,
+            self.weight_block_scale,
+            self.weight_scale,
+            None,
+        )
+
+    def project_quantized(self, activation: QuantizedTensor) -> torch.Tensor:
+        logical_m = int(activation._params.orig_shape[0])
+        if self.direct_m1 and logical_m == 1:
+            return self._direct(activation)
+        return F.linear(activation, self._weight_quantized_tensor(), None)
+
+    def forward(self, value: torch.Tensor) -> torch.Tensor:
+        if value.shape[-1] != self.in_features:
+            raise ValueError(
+                f"Expected input width {self.in_features}, got {value.shape[-1]}"
+            )
+        input_shape = tuple(value.shape)
+        correction_input = value.reshape(-1, self.in_features).contiguous()
+        flattened = correction_input
+        if self.input_hadamard_block_size:
+            flattened = _hadamard_blocks(
+                flattened,
+                self.input_hadamard_block_size,
+            ).contiguous()
+        if 0 < flattened.shape[0] <= self.w4a16_max_m:
+            output = load_smallm_nvfp4_extension().linear(
+                flattened,
+                self.qdata,
+                self.weight_block_scale,
+                self.weight_scale,
+                None,
+            )
+        else:
+            activation = quantize_nvfp4(flattened)
+            output = self.project_quantized(activation)
+        if self.correction_down is not None:
+            correction = F.linear(
+                F.linear(correction_input, self.correction_down),
+                self.correction_up,
+            )
+            output = output + correction
+        if self.sparse_correction_indices is not None:
+            selected = correction_input.index_select(
+                1,
+                self.sparse_correction_indices,
+            )
+            output = output + F.linear(selected, self.sparse_correction_weight)
+        return output.reshape(*input_shape[:-1], self.out_features)
+
+    def extra_repr(self) -> str:
+        backend = (
+            f"w4a16_through_m{self.w4a16_max_m}+w4a4_above_threshold"
+            if self.w4a16_max_m
+            else ("direct_m1+tensorcore" if self.direct_m1 else "tensorcore")
+        )
+        return (
+            f"in_features={self.in_features}, out_features={self.out_features}, "
+            f"weight=NVFP4_E2M1, activation=dynamic_NVFP4, output=BF16, backend={backend}, "
+            f"correction_rank={0 if self.correction_down is None else self.correction_down.shape[0]}, "
+            f"sparse_correction_channels="
+            f"{0 if self.sparse_correction_indices is None else self.sparse_correction_indices.numel()}, "
+            f"input_hadamard_block_size={self.input_hadamard_block_size}"
+        )
+
+
+class NVFP4FeedForward(nn.Module):
+    """S2 SwiGLU with shared input packing and fused product packing at M=1."""
+
+    def __init__(
+        self,
+        w1: NVFP4Linear,
+        w2: NVFP4Linear,
+        w3: NVFP4Linear,
+    ) -> None:
+        super().__init__()
+        self.w1 = w1
+        self.w2 = w2
+        self.w3 = w3
+        self.decode_backend = "w4a16" if w1.w4a16_max_m else "w4a4"
+
+    @classmethod
+    @torch.inference_mode()
+    def from_module(
+        cls,
+        module: nn.Module,
+        *,
+        decode_backend: str = "w4a4",
+        w4a16_max_m: int = 1,
+    ) -> "NVFP4FeedForward":
+        if decode_backend not in {"w4a4", "w4a16"}:
+            raise ValueError(f"Unsupported NVFP4 MLP decode backend: {decode_backend}")
+        for name in ("w1", "w2", "w3"):
+            if not isinstance(getattr(module, name, None), nn.Linear):
+                raise TypeError(f"Expected BF16 FeedForward.{name} linear")
+        result = cls(
+            NVFP4Linear.from_linear(
+                module.w1,
+                direct_m1=decode_backend == "w4a4",
+                w4a16_max_m=(w4a16_max_m if decode_backend == "w4a16" else 0),
+            ),
+            NVFP4Linear.from_linear(
+                module.w2,
+                direct_m1=decode_backend == "w4a4",
+                w4a16_max_m=(w4a16_max_m if decode_backend == "w4a16" else 0),
+            ),
+            NVFP4Linear.from_linear(
+                module.w3,
+                direct_m1=decode_backend == "w4a4",
+                w4a16_max_m=(w4a16_max_m if decode_backend == "w4a16" else 0),
+            ),
+        )
+        result.decode_backend = decode_backend
+        return result
+
+    def _forward_m1(self, flattened: torch.Tensor) -> torch.Tensor:
+        if self.decode_backend == "w4a16":
+            return self.w2(F.silu(self.w1(flattened)) * self.w3(flattened))
+        packer = load_silu_product_nvfp4_m1_extension()
+        qdata, block_scale, tensor_scale = packer.quantize_input(flattened)
+        return self.forward_packed_m1(qdata, block_scale, tensor_scale)
+
+    def forward_packed_m1(
+        self,
+        qdata: torch.Tensor,
+        block_scale: torch.Tensor,
+        tensor_scale: torch.Tensor,
+    ) -> torch.Tensor:
+        """Consume an already packed normalized M=1 activation."""
+        packer = load_silu_product_nvfp4_m1_extension()
+        gate = self.w1.project_packed_m1(qdata, block_scale, tensor_scale)
+        up = self.w3.project_packed_m1(qdata, block_scale, tensor_scale)
+        product_qdata, product_block_scale, product_tensor_scale = packer.quantize(
+            gate, up
+        )
+        return self.w2.project_packed_m1(
+            product_qdata,
+            product_block_scale,
+            product_tensor_scale,
+        )
+
+    def _forward_tensorcore(self, flattened: torch.Tensor) -> torch.Tensor:
+        activation = quantize_nvfp4(flattened)
+        gate = self.w1.project_quantized(activation)
+        up = self.w3.project_quantized(activation)
+        return self.w2(F.silu(gate) * up)
+
+    def forward(self, value: torch.Tensor) -> torch.Tensor:
+        input_shape = tuple(value.shape)
+        flattened = value.reshape(-1, input_shape[-1]).contiguous()
+        if (
+            self.decode_backend == "w4a16"
+            and flattened.shape[0] <= self.w1.w4a16_max_m
+        ):
+            output = self.w2(F.silu(self.w1(flattened)) * self.w3(flattened))
+        elif flattened.shape[0] == 1:
+            output = self._forward_m1(flattened)
+        else:
+            output = self._forward_tensorcore(flattened)
+        return output.reshape(*input_shape[:-1], self.w2.out_features)
+
+    def extra_repr(self) -> str:
+        return (
+            f"decode_backend={self.decode_backend}, "
+            f"fused_silu_product_pack_m1={self.decode_backend == 'w4a4'}"
+        )
+
+
+class NVFP4TransformerBlock(nn.Module):
+    """Slow S2 block with fused norm/MLP packing for autoregressive M=1."""
+
+    def __init__(self, source: nn.Module) -> None:
+        super().__init__()
+        source.attention.wqkv = NVFP4Linear.from_linear(
+            source.attention.wqkv, direct_m1=True
+        )
+        source.attention.wo = NVFP4Linear.from_linear(
+            source.attention.wo, direct_m1=True
+        )
+        self.attention = source.attention
+        self.feed_forward = NVFP4FeedForward.from_module(source.feed_forward)
+        self.ffn_norm = source.ffn_norm
+        self.attention_norm = source.attention_norm
+        self.train(source.training)
+
+    def _packed_norm(
+        self,
+        value: torch.Tensor,
+        norm: nn.Module,
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        flattened = value.reshape(-1, value.shape[-1]).contiguous()
+        return tuple(
+            load_rmsnorm_nvfp4_m1_extension().quantize(
+                flattened,
+                norm.weight.contiguous(),
+                float(norm.eps),
+            )
+        )
+
+    def _attention_m1(
+        self,
+        x: torch.Tensor,
+        packed_norm: tuple[torch.Tensor, torch.Tensor, torch.Tensor],
+        freqs_cis: torch.Tensor,
+        mask: torch.Tensor | None,
+        input_pos: torch.Tensor | None,
+    ) -> torch.Tensor:
+        attention = self.attention
+        bsz, seqlen, _ = x.shape
+        qkv = attention.wqkv.project_packed_m1(*packed_norm)
+        q_size = attention.n_head * attention.head_dim
+        kv_size = attention.n_local_heads * attention.head_dim
+        q, k, v = qkv.split([q_size, kv_size, kv_size], dim=-1)
+        q = q.view(bsz, seqlen, attention.n_head, attention.head_dim)
+        k = k.view(bsz, seqlen, attention.n_local_heads, attention.head_dim)
+        v = v.view(bsz, seqlen, attention.n_local_heads, attention.head_dim)
+        if attention.attention_qk_norm:
+            q = attention.q_norm(q)
+            k = attention.k_norm(k)
+        q = apply_rotary_emb(q, freqs_cis)
+        k = apply_rotary_emb(k, freqs_cis)
+        q, k, v = (item.transpose(1, 2) for item in (q, k, v))
+        if attention.kv_cache is not None:
+            k, v = attention.kv_cache.update(input_pos, k, v)
+        repeat = attention.n_head // attention.n_local_heads
+        k = k.repeat_interleave(repeat, dim=1)
+        v = v.repeat_interleave(repeat, dim=1)
+        if attention.use_sdpa:
+            if mask is None:
+                with sdpa_kernel(SDPBackend.FLASH_ATTENTION):
+                    y = F.scaled_dot_product_attention(
+                        q,
+                        k,
+                        v,
+                        dropout_p=attention.dropout if attention.training else 0.0,
+                        is_causal=True,
+                    )
+            else:
+                y = F.scaled_dot_product_attention(
+                    q,
+                    k,
+                    v,
+                    attn_mask=mask,
+                    dropout_p=attention.dropout if attention.training else 0.0,
+                )
+        else:
+            y = attention.eq_scaled_dot_product_attention(q, k, v, attn_mask=mask)
+        y = y.transpose(1, 2).contiguous().view(bsz, seqlen, q_size)
+        y_flat = y.reshape(-1, q_size)
+        output_pack = load_silu_product_nvfp4_m1_extension().quantize_input(y_flat)
+        return attention.wo.project_packed_m1(*output_pack).reshape(
+            bsz, seqlen, attention.dim
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        freqs_cis: torch.Tensor,
+        mask: torch.Tensor,
+        input_pos: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        if x.reshape(-1, x.shape[-1]).shape[0] != 1:
+            h = x + self.attention(
+                self.attention_norm(x), freqs_cis, mask, input_pos
+            )
+            return h + self.feed_forward(self.ffn_norm(h))
+        attention_pack = self._packed_norm(x, self.attention_norm)
+        h = x + self._attention_m1(x, attention_pack, freqs_cis, mask, input_pos)
+        ffn_pack = self._packed_norm(h, self.ffn_norm)
+        feed_forward = self.feed_forward.forward_packed_m1(*ffn_pack)
+        return h + feed_forward.reshape_as(h)
+
+    def extra_repr(self) -> str:
+        return "fused_rmsnorm_pack_m1=True, fused_mlp_pack_m1=True"
+
+
+def _selected_slow_mlp(name: str, module: nn.Module) -> bool:
+    return (
+        isinstance(module, nn.Linear)
+        and name.startswith("layers.")
+        and ".feed_forward." in name
+        and name.rsplit(".", 1)[-1] in {"w1", "w2", "w3"}
+    )
+
+
+def _selected_slow_transformer(name: str, module: nn.Module) -> bool:
+    return isinstance(module, nn.Linear) and name.startswith("layers.")
+
+
+def _selected_slow_attention(name: str, module: nn.Module) -> bool:
+    return (
+        isinstance(module, nn.Linear)
+        and name.startswith("layers.")
+        and ".attention." in name
+        and name.rsplit(".", 1)[-1] in {"wqkv", "wo"}
+    )
+
+
+def _selected_slow_mlp_block(name: str, module: nn.Module) -> bool:
+    return (
+        name.startswith("layers.")
+        and name.endswith(".feed_forward")
+        and all(isinstance(getattr(module, role, None), nn.Linear) for role in ("w1", "w2", "w3"))
+    )
+
+
+def _selected_slow_transformer_block(name: str, module: nn.Module) -> bool:
+    parts = name.split(".")
+    return (
+        len(parts) == 2
+        and parts[0] == "layers"
+        and parts[1].isdigit()
+        and hasattr(module, "attention")
+        and hasattr(module, "feed_forward")
+        and hasattr(module, "attention_norm")
+        and hasattr(module, "ffn_norm")
+    )
+
+
+def _packed_weight_bytes(model: nn.Module) -> int:
+    return sum(
+        int(module.qdata.numel() * module.qdata.element_size())
+        + int(module.weight_block_scale.numel() * module.weight_block_scale.element_size())
+        + int(module.weight_scale.numel() * module.weight_scale.element_size())
+        for module in model.modules()
+        if isinstance(module, NVFP4Linear)
+    )
+
+
+def _resolve_workspace_path(value: str | Path, *, relative_to: Path) -> Path:
+    path = Path(value)
+    if path.is_absolute():
+        return path
+    relative_candidate = relative_to / path
+    if relative_candidate.exists():
+        return relative_candidate
+    return ROOT / path
+
+
+def _clip_weight_blocks(weight: torch.Tensor, ratio: float) -> torch.Tensor:
+    if ratio == 1.0:
+        return weight
+    if not 0 < ratio <= 1:
+        raise ValueError(f"NVFP4 clip ratio must be in (0,1], got {ratio}")
+    if weight.shape[1] % 128:
+        raise ValueError(f"Input width must be divisible by 128, got {weight.shape}")
+    blocks = weight.float().reshape(weight.shape[0], -1, 128)
+    threshold = blocks.abs().amax(dim=-1, keepdim=True) * ratio
+    return torch.clamp(blocks, min=-threshold, max=threshold).reshape_as(weight)
+
+
+@torch.inference_mode()
+def _prepare_activation_scaling(
+    model: nn.Module,
+    report_path: Path,
+    selected_layers: set[int],
+    *,
+    fold_norms: bool = True,
+) -> tuple[
+    dict[int, torch.Tensor],
+    dict[int, float],
+    dict[int, dict[str, float]],
+    dict[int, dict[str, torch.Tensor]],
+    dict[str, Any],
+]:
+    sweep = json.loads(report_path.read_text())
+    calibration_values = sweep.get("calibration_reports")
+    if calibration_values is None:
+        calibration_values = [sweep["calibration_report"]]
+    calibration_report_paths = [
+        _resolve_workspace_path(value, relative_to=report_path.parent)
+        for value in calibration_values
+    ]
+    calibrations = []
+    activation_sets = []
+    for calibration_report_path in calibration_report_paths:
+        calibration = json.loads(calibration_report_path.read_text())
+        with np.load(calibration_report_path.parent / calibration["arrays"]) as values:
+            activation_sets.append(values["activations"].copy())
+        calibrations.append(calibration)
+    activations = np.concatenate(activation_sets, axis=1)
+    rows = {int(row["layer"]): row for row in sweep["results"]}
+    missing = sorted(selected_layers - rows.keys())
+    if missing:
+        raise ValueError(f"Activation-scaling report lacks layers: {missing}")
+    clamp = float(sweep["scale_clamp"])
+    activation_statistic = sweep.get("activation_statistic", "absmax")
+    scales: dict[int, torch.Tensor] = {}
+    clip_ratios: dict[int, float] = {}
+    quantization_scales: dict[int, dict[str, float]] = {}
+    corrections: dict[int, dict[str, torch.Tensor]] = {}
+    correction_tensors = None
+    correction_tensors_value = sweep.get("correction_tensors")
+    if correction_tensors_value is not None:
+        from safetensors.torch import load_file
+
+        correction_path = _resolve_workspace_path(
+            correction_tensors_value,
+            relative_to=report_path.parent,
+        )
+        correction_tensors = load_file(
+            correction_path,
+            device=str(model.layers[0].ffn_norm.weight.device),
+        )
+    records = []
+    for layer_index in sorted(selected_layers):
+        block = model.layers[layer_index]
+        gate = block.feed_forward.w1.weight
+        up = block.feed_forward.w3.weight
+        row = rows[layer_index]
+        improvement = float(row["improvement_fraction_vs_unscaled"])
+        explicit_channel_scale = row.get("channel_scale")
+        if explicit_channel_scale is not None:
+            if len(explicit_channel_scale) != gate.shape[1]:
+                raise ValueError(
+                    f"Layer {layer_index} channel scale has "
+                    f"{len(explicit_channel_scale)} values, expected {gate.shape[1]}"
+                )
+            scale = torch.tensor(
+                explicit_channel_scale,
+                device=gate.device,
+                dtype=torch.float32,
+            )
+            if not torch.isfinite(scale).all() or (scale <= 0).any():
+                raise ValueError(
+                    f"Layer {layer_index} channel scale must be finite and positive"
+                )
+            alpha = None
+            clip_ratio = float(row.get("best_clip_ratio", 1.0))
+        elif improvement <= 0:
+            scale = torch.ones(gate.shape[1], device=gate.device, dtype=torch.float32)
+            alpha = None
+            clip_ratio = 1.0
+        else:
+            alpha = float(row["best_alpha"])
+            clip_ratio = float(row.get("best_clip_ratio", 1.0))
+            activation = torch.from_numpy(activations[layer_index]).to(
+                device=gate.device,
+                dtype=torch.float32,
+            )
+            if activation_statistic == "absmax":
+                activation_scale = activation.abs().amax(dim=0)
+            elif activation_statistic == "abs_p99":
+                activation_scale = torch.quantile(activation.abs(), 0.99, dim=0)
+            elif activation_statistic == "abs_p999":
+                activation_scale = torch.quantile(activation.abs(), 0.999, dim=0)
+            elif activation_statistic == "mean_abs":
+                activation_scale = activation.abs().mean(dim=0)
+            elif activation_statistic == "rms":
+                activation_scale = activation.square().mean(dim=0).sqrt()
+            else:
+                raise ValueError(
+                    f"Unknown activation statistic {activation_statistic!r}"
+                )
+            activation_scale = activation_scale.clamp_min(1e-6)
+            weight_max = torch.maximum(
+                gate.float().abs().amax(dim=0),
+                up.float().abs().amax(dim=0),
+            ).clamp_min(1e-6)
+            scale = activation_scale.pow(alpha) / weight_max.pow(1.0 - alpha)
+            scale = scale / torch.exp(torch.mean(torch.log(scale)))
+            scale = scale.clamp(min=1.0 / clamp, max=clamp)
+        role_quantization_scales = row.get("weight_tensor_scales")
+        if role_quantization_scales is not None:
+            parsed_role_scales = {
+                role: float(role_quantization_scales[role])
+                for role in ("w1", "w3")
+            }
+            if any(
+                not np.isfinite(value) or value <= 0
+                for value in parsed_role_scales.values()
+            ):
+                raise ValueError(
+                    f"Layer {layer_index} weight tensor scales must be finite and positive"
+                )
+            quantization_scales[layer_index] = parsed_role_scales
+        low_rank = row.get("low_rank_correction")
+        if low_rank is not None:
+            if correction_tensors is None:
+                raise ValueError(
+                    f"Layer {layer_index} has a low-rank correction without tensors"
+                )
+            down_key = low_rank["down_key"]
+            up_keys = low_rank["up_keys"]
+            try:
+                correction = {
+                    "down": correction_tensors[down_key],
+                    "w1_up": correction_tensors[up_keys["w1"]],
+                    "w3_up": correction_tensors[up_keys["w3"]],
+                }
+            except KeyError as error:
+                raise ValueError(
+                    f"Layer {layer_index} correction tensor is missing: {error}"
+                ) from error
+            if any(value.dtype != torch.bfloat16 for value in correction.values()):
+                raise ValueError(
+                    f"Layer {layer_index} correction tensors must be BF16"
+                )
+            corrections[layer_index] = correction
+        sparse = row.get("sparse_channel_correction")
+        if sparse is not None:
+            if correction_tensors is None:
+                raise ValueError(
+                    f"Layer {layer_index} has a sparse correction without tensors"
+                )
+            try:
+                sparse_correction = {
+                    "sparse_indices": correction_tensors[sparse["indices_key"]],
+                    "w1_sparse_weight": correction_tensors[
+                        sparse["weight_keys"]["w1"]
+                    ],
+                    "w3_sparse_weight": correction_tensors[
+                        sparse["weight_keys"]["w3"]
+                    ],
+                }
+            except KeyError as error:
+                raise ValueError(
+                    f"Layer {layer_index} sparse correction tensor is missing: {error}"
+                ) from error
+            if sparse_correction["sparse_indices"].dtype != torch.int64:
+                raise ValueError(
+                    f"Layer {layer_index} sparse correction indices must be int64"
+                )
+            if any(
+                sparse_correction[key].dtype != torch.bfloat16
+                for key in ("w1_sparse_weight", "w3_sparse_weight")
+            ):
+                raise ValueError(
+                    f"Layer {layer_index} sparse correction weights must be BF16"
+                )
+            corrections.setdefault(layer_index, {}).update(sparse_correction)
+        input_hadamard_block_size = int(row.get("input_hadamard_block_size", 0))
+        if input_hadamard_block_size and (
+            input_hadamard_block_size < 2
+            or input_hadamard_block_size & (input_hadamard_block_size - 1)
+            or gate.shape[1] % input_hadamard_block_size
+        ):
+            raise ValueError(
+                f"Layer {layer_index} has invalid input Hadamard block size "
+                f"{input_hadamard_block_size}"
+            )
+        if fold_norms:
+            norm = block.ffn_norm.weight
+            norm.data.copy_((norm.float() / scale).to(dtype=norm.dtype))
+        scales[layer_index] = scale
+        clip_ratios[layer_index] = clip_ratio
+        records.append(
+            {
+                "layer": layer_index,
+                "alpha": alpha,
+                "calibration_improvement_fraction": improvement,
+                "clip_ratio": clip_ratio,
+                "scale_mode": (
+                    "explicit_channel"
+                    if explicit_channel_scale is not None
+                    else "activation_formula"
+                ),
+                "correction_rank": (
+                    None if low_rank is None else int(low_rank["rank"])
+                ),
+                "sparse_correction_channels": (
+                    None if sparse is None else int(sparse["channels"])
+                ),
+                "input_hadamard_block_size": input_hadamard_block_size,
+                "scale_min": float(scale.min()),
+                "scale_p50": float(torch.quantile(scale, 0.50)),
+                "scale_max": float(scale.max()),
+            }
+        )
+    return scales, clip_ratios, quantization_scales, corrections, {
+        "sweep_report": str(report_path),
+        "calibration_reports": [str(path) for path in calibration_report_paths],
+        "calibration_history_sha256": [
+            calibration["history_sha256"] for calibration in calibrations
+        ],
+        "scale_clamp": clamp,
+        "activation_statistic": activation_statistic,
+        "correction_tensors": correction_tensors_value,
+        "layers": records,
+    }
+
+
+@torch.inference_mode()
+def convert_s2_pro_nvfp4(
+    model: nn.Module,
+    *,
+    policy: str = "slow_mlp",
+    direct_m1: bool = False,
+    w4a16_max_m: int = 1,
+    nvfp4_layers: set[int] | None = None,
+    activation_scaling_report: Path | None = None,
+    probe_seed: int = 20260818,
+) -> dict[str, Any]:
+    """Replace selected S2-Pro projections without modifying the FP8 path."""
+    if w4a16_max_m < 1:
+        raise ValueError("w4a16_max_m must be positive")
+    selectors = {
+        "slow_mlp": (_selected_slow_mlp, 108),
+        "slow_transformer": (_selected_slow_transformer, 180),
+    }
+    if policy == "w4a16_slow_transformer":
+        candidates = [
+            (name, module)
+            for name, module in model.named_modules()
+            if _selected_slow_transformer(name, module)
+        ]
+        if len(candidates) != 180:
+            raise RuntimeError(
+                f"Expected 180 w4a16_slow_transformer projections, found {len(candidates)}"
+            )
+        generator = torch.Generator(device=candidates[0][1].weight.device)
+        generator.manual_seed(probe_seed)
+        records = []
+        for name, linear in candidates:
+            parent_name, attribute = name.rsplit(".", 1)
+            parent = model.get_submodule(parent_name)
+            probe = torch.randn(
+                1,
+                linear.in_features,
+                dtype=torch.bfloat16,
+                device=linear.weight.device,
+                generator=generator,
+            ) * 0.1
+            reference = F.linear(probe, linear.weight)
+            replacement = NVFP4Linear.from_linear(
+                linear, w4a16_max_m=w4a16_max_m
+            )
+            actual = replacement(probe)
+            records.append(
+                ConversionRecord(
+                    name=name,
+                    in_features=linear.in_features,
+                    out_features=linear.out_features,
+                    parameters=linear.weight.numel(),
+                    probe_cosine=float(
+                        F.cosine_similarity(
+                            actual.float().flatten(),
+                            reference.float().flatten(),
+                            dim=0,
+                        ).item()
+                    ),
+                )
+            )
+            setattr(parent, attribute, replacement)
+        torch.cuda.synchronize(candidates[0][1].weight.device)
+        serialized = [asdict(record) for record in records]
+        cosines = [record.probe_cosine for record in records]
+        parameters = sum(record.parameters for record in records)
+        return {
+            "policy": policy,
+            "backend": (
+                f"w4a16_through_m{w4a16_max_m}+w4a4_above_threshold"
+            ),
+            "w4a16_max_m": w4a16_max_m,
+            "modules": len(records),
+            "projections": len(records),
+            "parameters": parameters,
+            "theoretical_bf16_source_bytes": parameters * 2,
+            "packed_weight_bytes": _packed_weight_bytes(model),
+            "probe_cosine_min": min(cosines),
+            "probe_cosine_mean": sum(cosines) / len(cosines),
+            "probe_cosine_max": max(cosines),
+            "records": serialized,
+        }
+    mixed_role_policies = {
+        "w4a16_mlp_mxfp8_attention": ({"w1", "w2", "w3"}, None),
+        "w4a16_gate_up_mxfp8_rest": ({"w1", "w3"}, None),
+        "w4a16_down_mxfp8_rest": ({"w2"}, None),
+        "w4a16_gate_up_middle6_mxfp8_rest": ({"w1", "w3"}, set(range(15, 21))),
+        "w4a16_gate_up_middle12_mxfp8_rest": ({"w1", "w3"}, set(range(12, 24))),
+        "w4a16_gate_up_middle18_mxfp8_rest": ({"w1", "w3"}, set(range(9, 27))),
+        "w4a16_gate_up_middle24_mxfp8_rest": ({"w1", "w3"}, set(range(6, 30))),
+        "w4a16_gate_up_middle30_mxfp8_rest": ({"w1", "w3"}, set(range(3, 33))),
+        "w4a16_gate_up_custom_mxfp8_rest": ({"w1", "w3"}, "custom"),
+    }
+    if policy in mixed_role_policies:
+        from experimental.fp8 import MXFP8Linear
+
+        candidates = [
+            (name, module)
+            for name, module in model.named_modules()
+            if _selected_slow_transformer(name, module)
+        ]
+        if len(candidates) != 180:
+            raise RuntimeError(
+                f"Expected 180 {policy} projections, "
+                f"found {len(candidates)}"
+            )
+        nvfp4_roles, policy_layers = mixed_role_policies[policy]
+        if policy_layers == "custom":
+            if not nvfp4_layers:
+                raise ValueError(f"{policy} requires at least one nvfp4 layer")
+            invalid_layers = sorted(set(nvfp4_layers) - set(range(36)))
+            if invalid_layers:
+                raise ValueError(f"Invalid slow-transformer layers: {invalid_layers}")
+            selected_nvfp4_layers = set(nvfp4_layers)
+        else:
+            if nvfp4_layers is not None:
+                raise ValueError("nvfp4_layers is only valid with the custom policy")
+            selected_nvfp4_layers = policy_layers
+        activation_scales: dict[int, torch.Tensor] = {}
+        activation_clip_ratios: dict[int, float] = {}
+        activation_quantization_scales: dict[int, dict[str, float]] = {}
+        activation_corrections: dict[int, dict[str, torch.Tensor]] = {}
+        activation_hadamard_blocks: dict[int, int] = {}
+        activation_scaling = None
+        if activation_scaling_report is not None:
+            if nvfp4_roles != {"w1", "w3"}:
+                raise ValueError(
+                    "Activation scaling requires NVFP4 gate and up projections together"
+                )
+            layers_to_scale = (
+                set(range(36))
+                if selected_nvfp4_layers is None
+                else set(selected_nvfp4_layers)
+            )
+            (
+                activation_scales,
+                activation_clip_ratios,
+                activation_quantization_scales,
+                activation_corrections,
+                activation_scaling,
+            ) = _prepare_activation_scaling(
+                model,
+                Path(activation_scaling_report),
+                layers_to_scale,
+            )
+            activation_hadamard_blocks = {
+                int(record["layer"]): int(record["input_hadamard_block_size"])
+                for record in activation_scaling["layers"]
+                if int(record.get("input_hadamard_block_size", 0))
+            }
+        generator = torch.Generator(device=candidates[0][1].weight.device)
+        generator.manual_seed(probe_seed)
+        records = []
+        nvfp4_parameters = 0
+        mxfp8_parameters = 0
+        for name, linear in candidates:
+            parent_name, attribute = name.rsplit(".", 1)
+            parent = model.get_submodule(parent_name)
+            probe = torch.randn(
+                1,
+                linear.in_features,
+                dtype=torch.bfloat16,
+                device=linear.weight.device,
+                generator=generator,
+            ) * 0.1
+            reference = F.linear(probe, linear.weight)
+            role = name.rsplit(".", 1)[-1]
+            layer = int(name.split(".", 2)[1])
+            selected_layer = (
+                selected_nvfp4_layers is None or layer in selected_nvfp4_layers
+            )
+            if ".feed_forward." in name and role in nvfp4_roles and selected_layer:
+                activation_scale = activation_scales.get(layer)
+                if activation_scale is None:
+                    replacement = NVFP4Linear.from_linear(
+                        linear, w4a16_max_m=w4a16_max_m
+                    )
+                    actual_probe = probe
+                else:
+                    scaled_weight = (
+                        linear.weight.float() * activation_scale
+                    )
+                    scaled_weight = _clip_weight_blocks(
+                        scaled_weight,
+                        activation_clip_ratios.get(layer, 1.0),
+                    )
+                    input_hadamard_block_size = activation_hadamard_blocks.get(
+                        layer,
+                        0,
+                    )
+                    if input_hadamard_block_size:
+                        scaled_weight = _hadamard_blocks(
+                            scaled_weight,
+                            input_hadamard_block_size,
+                        )
+                    scaled_weight = scaled_weight.to(dtype=torch.bfloat16)
+                    replacement = NVFP4Linear.from_weight(
+                        scaled_weight,
+                        w4a16_max_m=w4a16_max_m,
+                        quantization_scale=activation_quantization_scales.get(
+                            layer, {}
+                        ).get(role),
+                        correction_down=activation_corrections.get(
+                            layer, {}
+                        ).get("down"),
+                        correction_up=activation_corrections.get(
+                            layer, {}
+                        ).get(f"{role}_up"),
+                        sparse_correction_indices=activation_corrections.get(
+                            layer, {}
+                        ).get("sparse_indices"),
+                        sparse_correction_weight=activation_corrections.get(
+                            layer, {}
+                        ).get(f"{role}_sparse_weight"),
+                        input_hadamard_block_size=input_hadamard_block_size,
+                    )
+                    actual_probe = (
+                        probe.float() / activation_scale
+                    ).to(dtype=torch.bfloat16)
+                precision = (
+                    f"nvfp4_w4a16_through_m{w4a16_max_m}"
+                    + ("_activation_scaled" if activation_scale is not None else "")
+                    + (
+                        "_low_rank_corrected"
+                        if "down" in activation_corrections.get(layer, {})
+                        else ""
+                    )
+                    + (
+                        "_sparse_channel_corrected"
+                        if "sparse_indices" in activation_corrections.get(layer, {})
+                        else ""
+                    )
+                    + (
+                        f"_hadamard{activation_hadamard_blocks[layer]}"
+                        if layer in activation_hadamard_blocks
+                        else ""
+                    )
+                )
+                nvfp4_parameters += linear.weight.numel()
+            else:
+                replacement = MXFP8Linear.from_linear(linear)
+                actual_probe = probe
+                precision = "mxfp8_w8a8"
+                mxfp8_parameters += linear.weight.numel()
+            actual = replacement(actual_probe)
+            record = ConversionRecord(
+                name=name,
+                in_features=linear.in_features,
+                out_features=linear.out_features,
+                parameters=linear.weight.numel(),
+                probe_cosine=float(
+                    F.cosine_similarity(
+                        actual.float().flatten(),
+                        reference.float().flatten(),
+                        dim=0,
+                    ).item()
+                ),
+            )
+            records.append({**asdict(record), "precision": precision})
+            setattr(parent, attribute, replacement)
+        torch.cuda.synchronize(candidates[0][1].weight.device)
+        nvfp4_bytes = _packed_weight_bytes(model)
+        mxfp8_bytes = sum(
+            module.weight_fp8.numel() * module.weight_fp8.element_size()
+            + module.weight_scale_storage.numel()
+            * module.weight_scale_storage.element_size()
+            for module in model.modules()
+            if isinstance(module, MXFP8Linear)
+        )
+        cosines = [record["probe_cosine"] for record in records]
+        parameters = nvfp4_parameters + mxfp8_parameters
+        correction_tensors = {
+            value.data_ptr(): value
+            for correction in activation_corrections.values()
+            for value in correction.values()
+        }
+        correction_parameters = sum(
+            value.numel()
+            for value in correction_tensors.values()
+            if value.is_floating_point()
+        )
+        correction_bytes = sum(
+            value.numel() * value.element_size()
+            for value in correction_tensors.values()
+        )
+        low_rank_tensors = {
+            value.data_ptr(): value
+            for correction in activation_corrections.values()
+            for key, value in correction.items()
+            if key == "down" or key.endswith("_up")
+        }
+        sparse_correction_tensors = {
+            value.data_ptr(): value
+            for correction in activation_corrections.values()
+            for key, value in correction.items()
+            if key == "sparse_indices" or key.endswith("_sparse_weight")
+        }
+        low_rank_correction_parameters = sum(
+            value.numel() for value in low_rank_tensors.values()
+        )
+        low_rank_correction_bytes = sum(
+            value.numel() * value.element_size()
+            for value in low_rank_tensors.values()
+        )
+        sparse_correction_parameters = sum(
+            value.numel()
+            for value in sparse_correction_tensors.values()
+            if value.is_floating_point()
+        )
+        sparse_correction_bytes = sum(
+            value.numel() * value.element_size()
+            for value in sparse_correction_tensors.values()
+        )
+        return {
+            "policy": policy,
+            "backend": "selective_nvfp4_weight_bf16_activation+mxfp8_rest",
+            "w4a16_max_m": w4a16_max_m,
+            "nvfp4_mlp_roles": sorted(nvfp4_roles),
+            "nvfp4_layers": (
+                "all"
+                if selected_nvfp4_layers is None
+                else sorted(selected_nvfp4_layers)
+            ),
+            "activation_scaling": activation_scaling,
+            "modules": len(records),
+            "projections": len(records),
+            "parameters": parameters,
+            "nvfp4_parameters": nvfp4_parameters,
+            "mxfp8_parameters": mxfp8_parameters,
+            "low_rank_correction_parameters": low_rank_correction_parameters,
+            "low_rank_correction_bytes": low_rank_correction_bytes,
+            "sparse_correction_parameters": sparse_correction_parameters,
+            "sparse_correction_bytes": sparse_correction_bytes,
+            "correction_parameters": correction_parameters,
+            "correction_bytes": correction_bytes,
+            "theoretical_bf16_source_bytes": parameters * 2,
+            "packed_weight_bytes": nvfp4_bytes + mxfp8_bytes,
+            "nvfp4_packed_weight_bytes": nvfp4_bytes,
+            "mxfp8_packed_weight_bytes": mxfp8_bytes,
+            "probe_cosine_min": min(cosines),
+            "probe_cosine_mean": sum(cosines) / len(cosines),
+            "probe_cosine_max": max(cosines),
+            "records": records,
+        }
+    if policy == "hybrid_slow_transformer":
+        from experimental.fp8 import MXFP8Linear
+
+        nvfp4 = convert_s2_pro_nvfp4(
+            model,
+            policy="slow_mlp_fused",
+            probe_seed=probe_seed,
+        )
+        candidates = [
+            (name, module)
+            for name, module in model.named_modules()
+            if _selected_slow_attention(name, module)
+        ]
+        if len(candidates) != 72:
+            raise RuntimeError(
+                f"Expected 72 hybrid attention projections, found {len(candidates)}"
+            )
+        generator = torch.Generator(device=candidates[0][1].weight.device)
+        generator.manual_seed(probe_seed + 1)
+        attention_records = []
+        for name, linear in candidates:
+            parent_name, attribute = name.rsplit(".", 1)
+            parent = model.get_submodule(parent_name)
+            probe = torch.randn(
+                1,
+                linear.in_features,
+                dtype=torch.bfloat16,
+                device=linear.weight.device,
+                generator=generator,
+            ) * 0.1
+            reference = F.linear(probe, linear.weight)
+            replacement = MXFP8Linear.from_linear(linear)
+            actual = replacement(probe)
+            attention_records.append(
+                ConversionRecord(
+                    name=name,
+                    in_features=linear.in_features,
+                    out_features=linear.out_features,
+                    parameters=linear.weight.numel(),
+                    probe_cosine=float(
+                        F.cosine_similarity(
+                            actual.float().flatten(),
+                            reference.float().flatten(),
+                            dim=0,
+                        ).item()
+                    ),
+                )
+            )
+            setattr(parent, attribute, replacement)
+        torch.cuda.synchronize(candidates[0][1].weight.device)
+        attention_parameters = sum(record.parameters for record in attention_records)
+        attention_packed_bytes = sum(
+            module.weight_fp8.numel() * module.weight_fp8.element_size()
+            + module.weight_scale_storage.numel()
+            * module.weight_scale_storage.element_size()
+            for module in model.modules()
+            if isinstance(module, MXFP8Linear)
+        )
+        cosines = [
+            *[record["probe_cosine"] for record in nvfp4["records"]],
+            *[record.probe_cosine for record in attention_records],
+        ]
+        return {
+            "policy": policy,
+            "backend": "fused_nvfp4_mlp+mxfp8_attention",
+            "modules": nvfp4["modules"] + len(attention_records),
+            "projections": nvfp4["projections"] + len(attention_records),
+            "parameters": nvfp4["parameters"] + attention_parameters,
+            "theoretical_bf16_source_bytes": (
+                nvfp4["theoretical_bf16_source_bytes"]
+                + attention_parameters * 2
+            ),
+            "packed_weight_bytes": (
+                nvfp4["packed_weight_bytes"] + attention_packed_bytes
+            ),
+            "nvfp4_mlp": nvfp4,
+            "mxfp8_attention": {
+                "modules": len(attention_records),
+                "parameters": attention_parameters,
+                "packed_weight_bytes": attention_packed_bytes,
+                "records": [asdict(record) for record in attention_records],
+            },
+            "probe_cosine_min": min(cosines),
+            "probe_cosine_mean": sum(cosines) / len(cosines),
+            "probe_cosine_max": max(cosines),
+        }
+    if policy == "slow_transformer_fused":
+        candidates = [
+            (name, module)
+            for name, module in model.named_modules()
+            if _selected_slow_transformer_block(name, module)
+        ]
+        if len(candidates) != 36:
+            raise RuntimeError(
+                f"Expected 36 slow_transformer_fused blocks, found {len(candidates)}"
+            )
+        generator = torch.Generator(device=candidates[0][1].attention.wqkv.weight.device)
+        generator.manual_seed(probe_seed)
+        records = []
+        for name, module in candidates:
+            parent_name, attribute = name.rsplit(".", 1)
+            parent = model.get_submodule(parent_name)
+            dim = module.attention.wqkv.in_features
+            probe = torch.randn(
+                1,
+                dim,
+                dtype=torch.bfloat16,
+                device=module.attention.wqkv.weight.device,
+                generator=generator,
+            ) * 0.1
+            wo_probe = torch.randn(
+                1,
+                module.attention.wo.in_features,
+                dtype=torch.bfloat16,
+                device=module.attention.wo.weight.device,
+                generator=generator,
+            ) * 0.1
+            reference_wqkv = module.attention.wqkv(probe)
+            reference_wo = module.attention.wo(wo_probe)
+            reference_mlp = module.feed_forward(probe)
+            parameters = sum(
+                linear.weight.numel()
+                for linear in (
+                    module.attention.wqkv,
+                    module.attention.wo,
+                    module.feed_forward.w1,
+                    module.feed_forward.w2,
+                    module.feed_forward.w3,
+                )
+            )
+            replacement = NVFP4TransformerBlock(module)
+            actual_wqkv = replacement.attention.wqkv(probe)
+            actual_wo = replacement.attention.wo(wo_probe)
+            actual_mlp = replacement.feed_forward(probe)
+            records.append(
+                FusedTransformerConversionRecord(
+                    name=name,
+                    parameters=parameters,
+                    wqkv_probe_cosine=float(
+                        F.cosine_similarity(
+                            actual_wqkv.float().flatten(),
+                            reference_wqkv.float().flatten(),
+                            dim=0,
+                        ).item()
+                    ),
+                    wo_probe_cosine=float(
+                        F.cosine_similarity(
+                            actual_wo.float().flatten(),
+                            reference_wo.float().flatten(),
+                            dim=0,
+                        ).item()
+                    ),
+                    mlp_probe_cosine=float(
+                        F.cosine_similarity(
+                            actual_mlp.float().flatten(),
+                            reference_mlp.float().flatten(),
+                            dim=0,
+                        ).item()
+                    ),
+                )
+            )
+            setattr(parent, attribute, replacement)
+        torch.cuda.synchronize(candidates[0][1].attention.wqkv.qdata.device)
+        serialized = [asdict(record) for record in records]
+        cosines = [
+            cosine
+            for record in records
+            for cosine in (
+                record.wqkv_probe_cosine,
+                record.wo_probe_cosine,
+                record.mlp_probe_cosine,
+            )
+        ]
+        parameters = sum(record.parameters for record in records)
+        return {
+            "policy": policy,
+            "backend": "direct_fused_m1+tensorcore_prefill",
+            "modules": len(records),
+            "projections": len(records) * 5,
+            "parameters": parameters,
+            "theoretical_bf16_source_bytes": parameters * 2,
+            "packed_weight_bytes": _packed_weight_bytes(model),
+            "probe_cosine_min": min(cosines),
+            "probe_cosine_mean": sum(cosines) / len(cosines),
+            "probe_cosine_max": max(cosines),
+            "records": serialized,
+        }
+    if policy == "slow_mlp_fused":
+        candidates = [
+            (name, module)
+            for name, module in model.named_modules()
+            if _selected_slow_mlp_block(name, module)
+        ]
+        if len(candidates) != 36:
+            raise RuntimeError(
+                f"Expected 36 slow_mlp_fused blocks, found {len(candidates)}"
+            )
+        generator = torch.Generator(device=candidates[0][1].w1.weight.device)
+        generator.manual_seed(probe_seed)
+        records = []
+        for name, module in candidates:
+            parent_name, attribute = name.rsplit(".", 1)
+            parent = model.get_submodule(parent_name)
+            probe = torch.randn(
+                1,
+                module.w1.in_features,
+                dtype=torch.bfloat16,
+                device=module.w1.weight.device,
+                generator=generator,
+            ) * 0.1
+            reference = module(probe)
+            replacement = NVFP4FeedForward.from_module(module)
+            actual = replacement(probe)
+            records.append(
+                FusedMLPConversionRecord(
+                    name=name,
+                    in_features=module.w1.in_features,
+                    intermediate_features=module.w1.out_features,
+                    out_features=module.w2.out_features,
+                    parameters=(
+                        module.w1.weight.numel()
+                        + module.w2.weight.numel()
+                        + module.w3.weight.numel()
+                    ),
+                    probe_cosine=float(
+                        F.cosine_similarity(
+                            actual.float().flatten(),
+                            reference.float().flatten(),
+                            dim=0,
+                        ).item()
+                    ),
+                )
+            )
+            setattr(parent, attribute, replacement)
+        torch.cuda.synchronize(candidates[0][1].w1.weight.device)
+        serialized = [asdict(record) for record in records]
+        cosines = [record.probe_cosine for record in records]
+        parameters = sum(record.parameters for record in records)
+        return {
+            "policy": policy,
+            "backend": "direct_fused_m1+tensorcore_prefill",
+            "modules": len(records),
+            "projections": len(records) * 3,
+            "parameters": parameters,
+            "theoretical_bf16_source_bytes": parameters * 2,
+            "packed_weight_bytes": _packed_weight_bytes(model),
+            "probe_cosine_min": min(cosines),
+            "probe_cosine_mean": sum(cosines) / len(cosines),
+            "probe_cosine_max": max(cosines),
+            "records": serialized,
+        }
+    if policy not in selectors:
+        raise ValueError(f"Unsupported initial NVFP4 policy: {policy}")
+    selector, expected_modules = selectors[policy]
+    candidates = [
+        (name, module)
+        for name, module in model.named_modules()
+        if selector(name, module)
+    ]
+    if len(candidates) != expected_modules:
+        raise RuntimeError(
+            f"Expected {expected_modules} {policy} projections, found {len(candidates)}"
+        )
+
+    generator = torch.Generator(device=candidates[0][1].weight.device)
+    generator.manual_seed(probe_seed)
+    records = []
+    for name, linear in candidates:
+        parent_name, attribute = name.rsplit(".", 1)
+        parent = model.get_submodule(parent_name)
+        replacement = NVFP4Linear.from_linear(linear, direct_m1=direct_m1)
+        probe = torch.randn(
+            1,
+            linear.in_features,
+            dtype=torch.bfloat16,
+            device=linear.weight.device,
+            generator=generator,
+        ) * 0.1
+        reference = F.linear(probe, linear.weight)
+        actual = replacement(probe)
+        probe_cosine = float(
+            F.cosine_similarity(
+                actual.float().flatten(), reference.float().flatten(), dim=0
+            ).item()
+        )
+        records.append(
+            ConversionRecord(
+                name=name,
+                in_features=linear.in_features,
+                out_features=linear.out_features,
+                parameters=linear.weight.numel(),
+                probe_cosine=probe_cosine,
+            )
+        )
+        setattr(parent, attribute, replacement)
+
+    torch.cuda.synchronize(candidates[0][1].weight.device)
+    serialized = [asdict(record) for record in records]
+    cosines = [record.probe_cosine for record in records]
+    return {
+        "policy": policy,
+        "backend": "direct_m1+tensorcore" if direct_m1 else "tensorcore",
+        "modules": len(records),
+        "parameters": sum(record.parameters for record in records),
+        "theoretical_bf16_source_bytes": sum(record.parameters * 2 for record in records),
+        "packed_weight_bytes": _packed_weight_bytes(model),
+        "probe_cosine_min": min(cosines),
+        "probe_cosine_mean": sum(cosines) / len(cosines),
+        "probe_cosine_max": max(cosines),
+        "records": serialized,
+    }

runtime/experimental/nvfp4/server.py

@@ -0,0 +1,113 @@
+"""Threaded Fish Speech queue for standalone mixed NVFP4 checkpoints."""
+
+from __future__ import annotations
+
+import queue
+import threading
+import traceback
+from pathlib import Path
+
+import torch
+from loguru import logger
+
+from fish_speech.models.text2semantic.inference import (
+    GenerateRequest,
+    WrappedGenerateResponse,
+    decode_one_token_ar,
+    generate_long,
+)
+
+from experimental.fp8 import MXFP8Linear
+from .checkpoint import load_mixed_nvfp4_checkpoint
+from .modules import NVFP4Linear
+
+
+def launch_mixed_nvfp4_thread_safe_queue(
+    checkpoint_path: str | Path,
+    device: str | torch.device,
+    precision: torch.dtype,
+    compile: bool = False,
+    *,
+    max_length: int = 3072,
+    verify_checksums: bool = False,
+) -> queue.Queue:
+    """Load the frozen mixed checkpoint once and serialize generation."""
+    if precision is not torch.bfloat16:
+        raise ValueError("The mixed NVFP4 service requires BF16 exclusions")
+    if compile:
+        raise ValueError("torch.compile is not qualified for this checkpoint")
+    if max_length < 1024:
+        raise ValueError("NVFP4 API cache length must be at least 1024")
+
+    input_queue: queue.Queue = queue.Queue()
+    init_event = threading.Event()
+    init_error: list[BaseException] = []
+
+    def worker() -> None:
+        try:
+            model = load_mixed_nvfp4_checkpoint(
+                checkpoint_path,
+                device=device,
+                max_length=max_length,
+                verify_checksums=verify_checksums,
+            )
+            with torch.device(device):
+                model.setup_caches(
+                    max_batch_size=1,
+                    max_seq_len=model.config.max_seq_len,
+                    dtype=torch.bfloat16,
+                )
+            model._cache_setup_done = True
+            nvfp4_count = sum(
+                isinstance(module, NVFP4Linear) for module in model.modules()
+            )
+            mxfp8_count = sum(
+                isinstance(module, MXFP8Linear) for module in model.modules()
+            )
+            if (nvfp4_count, mxfp8_count) != (60, 120):
+                raise RuntimeError(
+                    f"Expected 60/120 NVFP4/MXFP8 modules, got {nvfp4_count}/{mxfp8_count}"
+                )
+            logger.info(
+                "Loaded standalone English NVFP4 checkpoint with "
+                f"{nvfp4_count} NVFP4 and {mxfp8_count} MXFP8 projections on {device}"
+            )
+        except BaseException as error:
+            init_error.append(error)
+            logger.error(traceback.format_exc())
+            init_event.set()
+            return
+
+        init_event.set()
+        while True:
+            item: GenerateRequest | None = input_queue.get()
+            if item is None:
+                break
+            response_queue = item.response_queue
+            try:
+                for chunk in generate_long(
+                    model=model,
+                    decode_one_token=decode_one_token_ar,
+                    **item.request,
+                ):
+                    response_queue.put(
+                        WrappedGenerateResponse(status="success", response=chunk)
+                    )
+            except Exception as error:
+                logger.error(traceback.format_exc())
+                response_queue.put(
+                    WrappedGenerateResponse(status="error", response=error)
+                )
+            finally:
+                if torch.cuda.is_available():
+                    torch.cuda.empty_cache()
+
+    threading.Thread(
+        target=worker,
+        name="fish-s2-pro-mixed-nvfp4-worker",
+        daemon=True,
+    ).start()
+    init_event.wait()
+    if init_error:
+        raise RuntimeError("Mixed NVFP4 model worker failed to initialize") from init_error[0]
+    return input_queue

runtime/native/direct_w4a4_m1/direct_w4a4_m1.cu

@@ -0,0 +1,152 @@
+#include "direct_w4a4_m1.h"
+
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <cuda_bf16.h>
+#include <cuda_fp8.h>
+#include <cuda_runtime.h>
+
+#include <cstdint>
+#include <limits>
+#include <optional>
+
+namespace {
+
+constexpr int kWarpSize = 32;
+constexpr int kWarpsPerBlock = 8;
+constexpr int kThreads = kWarpSize * kWarpsPerBlock;
+constexpr int kFp4BlockElements = 16;
+constexpr int kScaleTileOuter = 128;
+
+__device__ __forceinline__ size_t scale_offset(
+    int outer,
+    int inner_scale,
+    int scale_inner_dim) {
+  const int outer_tile = outer / kScaleTileOuter;
+  const int local_outer = outer % kScaleTileOuter;
+  const int local_inner = inner_scale & 3;
+  const int inner_tile_start = inner_scale - local_inner;
+  const size_t tile_base =
+      static_cast<size_t>(
+          inner_tile_start + outer_tile * scale_inner_dim) *
+      kScaleTileOuter;
+  return tile_base + static_cast<size_t>(local_outer & 31) * 16 +
+      static_cast<size_t>(local_outer >> 5) * 4 + local_inner;
+}
+
+__device__ __forceinline__ float e4m3_to_float(uint8_t raw) {
+  const __half_raw half_raw = __nv_cvt_fp8_to_halfraw(raw, __NV_E4M3);
+  return __half2float(static_cast<__half>(half_raw));
+}
+
+__device__ __forceinline__ float2 e2m1x2_to_float2(uint8_t packed) {
+  const __half2_raw raw =
+      __nv_cvt_fp4x2_to_halfraw2(packed, __NV_E2M1);
+  const __half2 converted(raw);
+  return __half22float2(converted);
+}
+
+__global__ void direct_w4a4_m1_kernel(
+    const uint8_t* __restrict__ activation_qdata,
+    const uint8_t* __restrict__ activation_block_scales,
+    const float* __restrict__ activation_tensor_scale,
+    const uint8_t* __restrict__ weight_qdata,
+    const uint8_t* __restrict__ weight_block_scales,
+    const float* __restrict__ weight_tensor_scale,
+    const __nv_bfloat16* __restrict__ bias,
+    __nv_bfloat16* __restrict__ output,
+    int n,
+    int packed_k,
+    int activation_scale_inner_dim,
+    int weight_scale_inner_dim) {
+  const int lane = threadIdx.x & (kWarpSize - 1);
+  const int warp_in_block = threadIdx.x / kWarpSize;
+  const int row_n = blockIdx.x * kWarpsPerBlock + warp_in_block;
+  if (row_n >= n) {
+    return;
+  }
+
+  const uint8_t* weight_row =
+      weight_qdata + static_cast<int64_t>(row_n) * packed_k;
+  const float activation_global_scale = activation_tensor_scale[0];
+  const float weight_global_scale = weight_tensor_scale[0];
+  float accumulator = 0.0f;
+
+  for (int pair = lane; pair < packed_k; pair += kWarpSize) {
+    const int scale_block = pair / (kFp4BlockElements / 2);
+    const float activation_scale =
+        e4m3_to_float(activation_block_scales[
+            scale_offset(0, scale_block, activation_scale_inner_dim)]) *
+        activation_global_scale;
+    const float weight_scale =
+        e4m3_to_float(weight_block_scales[
+            scale_offset(row_n, scale_block, weight_scale_inner_dim)]) *
+        weight_global_scale;
+    const float2 activations = e2m1x2_to_float2(activation_qdata[pair]);
+    const float2 weights = e2m1x2_to_float2(weight_row[pair]);
+    accumulator = fmaf(
+        activations.y * activation_scale,
+        weights.y * weight_scale,
+        accumulator);
+    accumulator = fmaf(
+        activations.x * activation_scale,
+        weights.x * weight_scale,
+        accumulator);
+  }
+
+#pragma unroll
+  for (int offset = 16; offset > 0; offset >>= 1) {
+    accumulator += __shfl_down_sync(0xFFFFFFFF, accumulator, offset);
+  }
+  if (lane == 0) {
+    if (bias != nullptr) {
+      accumulator += __bfloat162float(bias[row_n]);
+    }
+    output[row_n] = __float2bfloat16_rn(accumulator);
+  }
+}
+
+}  // namespace
+
+torch::Tensor direct_w4a4_m1_linear_cuda(
+    const torch::Tensor& activation_qdata,
+    const torch::Tensor& activation_block_scales,
+    const torch::Tensor& activation_tensor_scale,
+    const torch::Tensor& weight_qdata,
+    const torch::Tensor& weight_block_scales,
+    const torch::Tensor& weight_tensor_scale,
+    const std::optional<torch::Tensor>& bias) {
+  const auto device = activation_qdata.device();
+  c10::cuda::CUDAGuard guard(device);
+  const int64_t n64 = weight_qdata.size(0);
+  const int64_t packed_k64 = weight_qdata.size(1);
+  TORCH_CHECK(
+      n64 <= static_cast<int64_t>(std::numeric_limits<int>::max()) &&
+          packed_k64 <= static_cast<int64_t>(std::numeric_limits<int>::max()),
+      "direct W4A4 M=1 N or K is out of range");
+  const int n = static_cast<int>(n64);
+  const int packed_k = static_cast<int>(packed_k64);
+  torch::Tensor output = torch::empty(
+      {1, n64}, activation_qdata.options().dtype(at::kBFloat16));
+  const int block_count = (n + kWarpsPerBlock - 1) / kWarpsPerBlock;
+  const auto stream = at::cuda::getCurrentCUDAStream(device.index()).stream();
+  const __nv_bfloat16* bias_pointer =
+      bias.has_value()
+      ? reinterpret_cast<const __nv_bfloat16*>(bias->data_ptr())
+      : nullptr;
+  direct_w4a4_m1_kernel<<<block_count, kThreads, 0, stream>>>(
+      reinterpret_cast<const uint8_t*>(activation_qdata.data_ptr()),
+      reinterpret_cast<const uint8_t*>(activation_block_scales.data_ptr()),
+      reinterpret_cast<const float*>(activation_tensor_scale.data_ptr()),
+      reinterpret_cast<const uint8_t*>(weight_qdata.data_ptr()),
+      reinterpret_cast<const uint8_t*>(weight_block_scales.data_ptr()),
+      reinterpret_cast<const float*>(weight_tensor_scale.data_ptr()),
+      bias_pointer,
+      reinterpret_cast<__nv_bfloat16*>(output.data_ptr()),
+      n,
+      packed_k,
+      static_cast<int>(activation_block_scales.size(1)),
+      static_cast<int>(weight_block_scales.size(1)));
+  C10_CUDA_KERNEL_LAUNCH_CHECK();
+  return output;
+}

runtime/native/direct_w4a4_m1/direct_w4a4_m1.h

@@ -0,0 +1,14 @@
+#pragma once
+
+#include <torch/extension.h>
+
+#include <optional>
+
+torch::Tensor direct_w4a4_m1_linear_cuda(
+    const torch::Tensor& activation_qdata,
+    const torch::Tensor& activation_block_scales,
+    const torch::Tensor& activation_tensor_scale,
+    const torch::Tensor& weight_qdata,
+    const torch::Tensor& weight_block_scales,
+    const torch::Tensor& weight_tensor_scale,
+    const std::optional<torch::Tensor>& bias);

runtime/native/silu_product_nvfp4_m1/silu_product_nvfp4_m1.cpp

@@ -0,0 +1,97 @@
+#include "silu_product_nvfp4_m1.h"
+
+#include <torch/extension.h>
+
+#include <vector>
+
+namespace {
+
+std::vector<torch::Tensor> silu_product_nvfp4_m1(
+    const torch::Tensor& gate,
+    const torch::Tensor& up) {
+  TORCH_CHECK(
+      gate.is_cuda() && gate.scalar_type() == at::kBFloat16 &&
+          gate.dim() == 2 && gate.size(0) == 1 && gate.is_contiguous(),
+      "gate must be contiguous CUDA bfloat16 [1,K]");
+  TORCH_CHECK(
+      up.is_cuda() && up.scalar_type() == at::kBFloat16 &&
+          up.dim() == 2 && up.size(0) == 1 && up.is_contiguous(),
+      "up must be contiguous CUDA bfloat16 [1,K]");
+  TORCH_CHECK(
+      gate.device() == up.device() && gate.sizes() == up.sizes(),
+      "gate and up shapes/devices differ");
+  TORCH_CHECK(
+      gate.size(1) > 0 && gate.size(1) % 32 == 0,
+      "SiLU/product-to-NVFP4 requires K divisible by 32");
+  return silu_product_nvfp4_m1_cuda(gate, up);
+}
+
+std::vector<torch::Tensor> product_nvfp4_m1(
+    const torch::Tensor& left,
+    const torch::Tensor& right) {
+  TORCH_CHECK(
+      left.is_cuda() && left.scalar_type() == at::kBFloat16 &&
+          left.dim() == 2 && left.size(0) == 1 && left.is_contiguous(),
+      "left must be contiguous CUDA bfloat16 [1,K]");
+  TORCH_CHECK(
+      right.is_cuda() && right.scalar_type() == at::kBFloat16 &&
+          right.dim() == 2 && right.size(0) == 1 && right.is_contiguous(),
+      "right must be contiguous CUDA bfloat16 [1,K]");
+  TORCH_CHECK(
+      left.device() == right.device() && left.sizes() == right.sizes(),
+      "left and right shapes/devices differ");
+  TORCH_CHECK(
+      left.size(1) > 0 && left.size(1) % 32 == 0,
+      "product-to-NVFP4 requires K divisible by 32");
+  return product_nvfp4_m1_cuda(left, right);
+}
+
+std::vector<torch::Tensor> pack_nvfp4_m1(const torch::Tensor& input) {
+  TORCH_CHECK(
+      input.is_cuda() && input.scalar_type() == at::kBFloat16 &&
+          input.dim() == 2 && input.size(0) == 1 && input.is_contiguous(),
+      "input must be contiguous CUDA bfloat16 [1,K]");
+  TORCH_CHECK(
+      input.size(1) > 0 && input.size(1) % 32 == 0,
+      "NVFP4 pack requires K divisible by 32");
+  return pack_nvfp4_m1_cuda(input);
+}
+
+std::vector<torch::Tensor> pack_nvfp4_fixed_scale_m1(
+    const torch::Tensor& input,
+    const torch::Tensor& tensor_scale) {
+  TORCH_CHECK(
+      input.is_cuda() && input.scalar_type() == at::kBFloat16 &&
+          input.dim() == 2 && input.size(0) == 1 && input.is_contiguous(),
+      "input must be contiguous CUDA bfloat16 [1,K]");
+  TORCH_CHECK(
+      input.size(1) > 0 && input.size(1) % 32 == 0,
+      "fixed-scale NVFP4 pack requires K divisible by 32");
+  TORCH_CHECK(
+      tensor_scale.is_cuda() && tensor_scale.device() == input.device() &&
+          tensor_scale.scalar_type() == at::kFloat &&
+          tensor_scale.numel() == 1 && tensor_scale.is_contiguous(),
+      "tensor scale must be a contiguous CUDA float32 scalar");
+  return pack_nvfp4_fixed_scale_m1_cuda(input, tensor_scale);
+}
+
+}  // namespace
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, module) {
+  module.def(
+      "quantize",
+      &silu_product_nvfp4_m1,
+      "Fused BF16 SiLU/product to packed NVFP4 at logical M=1");
+  module.def(
+      "quantize_product",
+      &product_nvfp4_m1,
+      "Fused BF16 product to packed NVFP4 at logical M=1");
+  module.def(
+      "quantize_input",
+      &pack_nvfp4_m1,
+      "Pack a materialized BF16 input to NVFP4 at logical M=1");
+  module.def(
+      "quantize_input_fixed_scale",
+      &pack_nvfp4_fixed_scale_m1,
+      "Pack BF16 to NVFP4 with a supplied tensor scale at logical M=1");
+}

runtime/native/silu_product_nvfp4_m1/silu_product_nvfp4_m1.cu

@@ -0,0 +1,276 @@
+#include "silu_product_nvfp4_m1.h"
+
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <cuda_bf16.h>
+#include <cuda_fp4.h>
+#include <cuda_fp8.h>
+#include <cuda_runtime.h>
+
+#include <cstdint>
+#include <vector>
+
+namespace {
+
+constexpr int kThreads = 256;
+constexpr int kFp4BlockElements = 16;
+constexpr int kScaleTileOuter = 128;
+constexpr float kFp4Max = 6.0f;
+constexpr float kFp8Max = 448.0f;
+constexpr float kCombinedMax = kFp4Max * kFp8Max;
+
+__device__ __forceinline__ size_t scale_offset(
+    int outer,
+    int inner_scale,
+    int scale_inner_dim) {
+  const int outer_tile = outer / kScaleTileOuter;
+  const int local_outer = outer % kScaleTileOuter;
+  const int local_inner = inner_scale & 3;
+  const int inner_tile_start = inner_scale - local_inner;
+  const size_t tile_base =
+      static_cast<size_t>(inner_tile_start + outer_tile * scale_inner_dim) *
+      kScaleTileOuter;
+  return tile_base + static_cast<size_t>(local_outer & 31) * 16 +
+      static_cast<size_t>(local_outer >> 5) * 4 + local_inner;
+}
+
+__device__ __forceinline__ float e4m3_to_float(uint8_t raw) {
+  const __half_raw half_raw = __nv_cvt_fp8_to_halfraw(raw, __NV_E4M3);
+  return __half2float(static_cast<__half>(half_raw));
+}
+
+// Match PyTorch's two BF16 materialization boundaries:
+//   silu(gate_bf16) -> bfloat16, then * up_bf16 -> bfloat16.
+__device__ __forceinline__ float silu_product_value(
+    const __nv_bfloat16* gate,
+    const __nv_bfloat16* up,
+    int index) {
+  const float gate_value = __bfloat162float(gate[index]);
+  const float silu_value = gate_value / (1.0f + expf(-gate_value));
+  const __nv_bfloat16 silu_bf16 = __float2bfloat16_rn(silu_value);
+  const float product =
+      __bfloat162float(silu_bf16) * __bfloat162float(up[index]);
+  return __bfloat162float(__float2bfloat16_rn(product));
+}
+
+__device__ __forceinline__ float product_value(
+    const __nv_bfloat16* left,
+    const __nv_bfloat16* right,
+    int index) {
+  const float product =
+      __bfloat162float(left[index]) * __bfloat162float(right[index]);
+  return __bfloat162float(__float2bfloat16_rn(product));
+}
+
+template <int kMode>
+__global__ void product_nvfp4_m1_kernel(
+    const __nv_bfloat16* __restrict__ left,
+    const __nv_bfloat16* __restrict__ right,
+    uint8_t* __restrict__ qdata,
+    uint8_t* __restrict__ block_scales,
+    float* __restrict__ tensor_scale,
+    int k,
+    int scale_inner_dim) {
+  __shared__ float reduction[kThreads];
+  const int tid = threadIdx.x;
+  float local_max = 0.0f;
+  for (int index = tid; index < k; index += kThreads) {
+    const float value = kMode == 0
+        ? silu_product_value(left, right, index)
+        : (kMode == 1
+            ? product_value(left, right, index)
+            : __bfloat162float(left[index]));
+    local_max = fmaxf(
+        local_max,
+        fabsf(value));
+  }
+  reduction[tid] = local_max;
+  __syncthreads();
+  for (int offset = kThreads / 2; offset > 0; offset >>= 1) {
+    if (tid < offset) {
+      reduction[tid] = fmaxf(reduction[tid], reduction[tid + offset]);
+    }
+    __syncthreads();
+  }
+  // TensorCoreNVFP4Layout derives its default scale from a BF16 amax and
+  // rounds the division result to BF16 before storing a scalar FP32 Params.
+  const float global_scale = __bfloat162float(__float2bfloat16_rn(
+      __fdiv_rn(reduction[0], kCombinedMax)));
+  if (tid == 0) {
+    tensor_scale[0] = global_scale;
+  }
+  __syncthreads();
+
+  const int block_count = k / kFp4BlockElements;
+  for (int block = tid; block < block_count; block += kThreads) {
+    const int base = block * kFp4BlockElements;
+    float values[kFp4BlockElements];
+    float block_max = 0.0f;
+#pragma unroll
+    for (int element = 0; element < kFp4BlockElements; ++element) {
+      values[element] = kMode == 0
+          ? silu_product_value(left, right, base + element)
+          : (kMode == 1
+              ? product_value(left, right, base + element)
+              : __bfloat162float(left[base + element]));
+      block_max = fmaxf(block_max, fabsf(values[element]));
+    }
+    float scaled_block_scale = (block_max / kFp4Max) / global_scale;
+    scaled_block_scale = fminf(scaled_block_scale, kFp8Max);
+    const uint8_t raw_block_scale = static_cast<uint8_t>(
+        __nv_cvt_float_to_fp8(
+            scaled_block_scale, __NV_SATFINITE, __NV_E4M3));
+    block_scales[scale_offset(0, block, scale_inner_dim)] = raw_block_scale;
+    const float total_scale = global_scale * e4m3_to_float(raw_block_scale);
+#pragma unroll
+    for (int pair = 0; pair < kFp4BlockElements / 2; ++pair) {
+      const float even = values[pair * 2] / total_scale;
+      const float odd = values[pair * 2 + 1] / total_scale;
+      const float2 arguments = make_float2(odd, even);
+      qdata[base / 2 + pair] = static_cast<uint8_t>(
+          __nv_cvt_float2_to_fp4x2(arguments, __NV_E2M1, cudaRoundNearest));
+    }
+  }
+}
+
+__global__ void pack_nvfp4_fixed_scale_m1_kernel(
+    const __nv_bfloat16* __restrict__ input,
+    uint8_t* __restrict__ qdata,
+    uint8_t* __restrict__ block_scales,
+    const float* __restrict__ tensor_scale,
+    int k,
+    int scale_inner_dim) {
+  const int tid = threadIdx.x;
+  const float global_scale = tensor_scale[0];
+  const int block_count = k / kFp4BlockElements;
+  for (int block = tid; block < block_count; block += blockDim.x) {
+    const int base = block * kFp4BlockElements;
+    float values[kFp4BlockElements];
+    float block_max = 0.0f;
+#pragma unroll
+    for (int element = 0; element < kFp4BlockElements; ++element) {
+      values[element] = __bfloat162float(input[base + element]);
+      block_max = fmaxf(block_max, fabsf(values[element]));
+    }
+    float scaled_block_scale = (block_max / kFp4Max) / global_scale;
+    scaled_block_scale = fminf(scaled_block_scale, kFp8Max);
+    const uint8_t raw_block_scale = static_cast<uint8_t>(
+        __nv_cvt_float_to_fp8(
+            scaled_block_scale, __NV_SATFINITE, __NV_E4M3));
+    block_scales[scale_offset(0, block, scale_inner_dim)] = raw_block_scale;
+    const float total_scale = global_scale * e4m3_to_float(raw_block_scale);
+#pragma unroll
+    for (int pair = 0; pair < kFp4BlockElements / 2; ++pair) {
+      const float2 arguments = make_float2(
+          values[pair * 2 + 1] / total_scale,
+          values[pair * 2] / total_scale);
+      qdata[base / 2 + pair] = static_cast<uint8_t>(
+          __nv_cvt_float2_to_fp4x2(arguments, __NV_E2M1, cudaRoundNearest));
+    }
+  }
+}
+
+}  // namespace
+
+std::vector<torch::Tensor> silu_product_nvfp4_m1_cuda(
+    const torch::Tensor& gate,
+    const torch::Tensor& up) {
+  const auto device = gate.device();
+  c10::cuda::CUDAGuard guard(device);
+  const int64_t k = gate.size(1);
+  const int64_t scale_cols = (k / 16 + 3) / 4 * 4;
+  torch::Tensor qdata = torch::zeros(
+      {16, k / 2}, gate.options().dtype(at::kByte));
+  torch::Tensor block_scales = torch::zeros(
+      {128, scale_cols},
+      gate.options().dtype(at::ScalarType::Float8_e4m3fn));
+  torch::Tensor tensor_scale = torch::empty(
+      {}, gate.options().dtype(at::kFloat));
+  const auto stream = at::cuda::getCurrentCUDAStream(device.index()).stream();
+  product_nvfp4_m1_kernel<0><<<1, kThreads, 0, stream>>>(
+      reinterpret_cast<const __nv_bfloat16*>(gate.data_ptr()),
+      reinterpret_cast<const __nv_bfloat16*>(up.data_ptr()),
+      reinterpret_cast<uint8_t*>(qdata.data_ptr()),
+      reinterpret_cast<uint8_t*>(block_scales.data_ptr()),
+      reinterpret_cast<float*>(tensor_scale.data_ptr()),
+      static_cast<int>(k),
+      static_cast<int>(scale_cols));
+  C10_CUDA_KERNEL_LAUNCH_CHECK();
+  return {qdata, block_scales, tensor_scale};
+}
+
+std::vector<torch::Tensor> product_nvfp4_m1_cuda(
+    const torch::Tensor& left,
+    const torch::Tensor& right) {
+  const auto device = left.device();
+  c10::cuda::CUDAGuard guard(device);
+  const int64_t k = left.size(1);
+  const int64_t scale_cols = (k / 16 + 3) / 4 * 4;
+  torch::Tensor qdata = torch::zeros(
+      {16, k / 2}, left.options().dtype(at::kByte));
+  torch::Tensor block_scales = torch::zeros(
+      {128, scale_cols},
+      left.options().dtype(at::ScalarType::Float8_e4m3fn));
+  torch::Tensor tensor_scale = torch::empty(
+      {}, left.options().dtype(at::kFloat));
+  const auto stream = at::cuda::getCurrentCUDAStream(device.index()).stream();
+  product_nvfp4_m1_kernel<1><<<1, kThreads, 0, stream>>>(
+      reinterpret_cast<const __nv_bfloat16*>(left.data_ptr()),
+      reinterpret_cast<const __nv_bfloat16*>(right.data_ptr()),
+      reinterpret_cast<uint8_t*>(qdata.data_ptr()),
+      reinterpret_cast<uint8_t*>(block_scales.data_ptr()),
+      reinterpret_cast<float*>(tensor_scale.data_ptr()),
+      static_cast<int>(k),
+      static_cast<int>(scale_cols));
+  C10_CUDA_KERNEL_LAUNCH_CHECK();
+  return {qdata, block_scales, tensor_scale};
+}
+
+std::vector<torch::Tensor> pack_nvfp4_m1_cuda(const torch::Tensor& input) {
+  const auto device = input.device();
+  c10::cuda::CUDAGuard guard(device);
+  const int64_t k = input.size(1);
+  const int64_t scale_cols = (k / 16 + 3) / 4 * 4;
+  torch::Tensor qdata = torch::zeros(
+      {16, k / 2}, input.options().dtype(at::kByte));
+  torch::Tensor block_scales = torch::zeros(
+      {128, scale_cols},
+      input.options().dtype(at::ScalarType::Float8_e4m3fn));
+  torch::Tensor tensor_scale = torch::empty(
+      {}, input.options().dtype(at::kFloat));
+  const auto stream = at::cuda::getCurrentCUDAStream(device.index()).stream();
+  product_nvfp4_m1_kernel<2><<<1, kThreads, 0, stream>>>(
+      reinterpret_cast<const __nv_bfloat16*>(input.data_ptr()),
+      reinterpret_cast<const __nv_bfloat16*>(input.data_ptr()),
+      reinterpret_cast<uint8_t*>(qdata.data_ptr()),
+      reinterpret_cast<uint8_t*>(block_scales.data_ptr()),
+      reinterpret_cast<float*>(tensor_scale.data_ptr()),
+      static_cast<int>(k),
+      static_cast<int>(scale_cols));
+  C10_CUDA_KERNEL_LAUNCH_CHECK();
+  return {qdata, block_scales, tensor_scale};
+}
+
+std::vector<torch::Tensor> pack_nvfp4_fixed_scale_m1_cuda(
+    const torch::Tensor& input,
+    const torch::Tensor& tensor_scale) {
+  const auto device = input.device();
+  c10::cuda::CUDAGuard guard(device);
+  const int64_t k = input.size(1);
+  const int64_t scale_cols = (k / 16 + 3) / 4 * 4;
+  torch::Tensor qdata = torch::zeros(
+      {16, k / 2}, input.options().dtype(at::kByte));
+  torch::Tensor block_scales = torch::zeros(
+      {128, scale_cols},
+      input.options().dtype(at::ScalarType::Float8_e4m3fn));
+  const auto stream = at::cuda::getCurrentCUDAStream(device.index()).stream();
+  pack_nvfp4_fixed_scale_m1_kernel<<<1, kThreads, 0, stream>>>(
+      reinterpret_cast<const __nv_bfloat16*>(input.data_ptr()),
+      reinterpret_cast<uint8_t*>(qdata.data_ptr()),
+      reinterpret_cast<uint8_t*>(block_scales.data_ptr()),
+      reinterpret_cast<const float*>(tensor_scale.data_ptr()),
+      static_cast<int>(k),
+      static_cast<int>(scale_cols));
+  C10_CUDA_KERNEL_LAUNCH_CHECK();
+  return {qdata, block_scales};
+}

runtime/native/silu_product_nvfp4_m1/silu_product_nvfp4_m1.h

@@ -0,0 +1,20 @@
+#pragma once
+
+#include <torch/extension.h>
+
+#include <vector>
+
+std::vector<torch::Tensor> silu_product_nvfp4_m1_cuda(
+    const torch::Tensor& gate,
+    const torch::Tensor& up);
+
+std::vector<torch::Tensor> product_nvfp4_m1_cuda(
+    const torch::Tensor& left,
+    const torch::Tensor& right);
+
+std::vector<torch::Tensor> pack_nvfp4_m1_cuda(
+    const torch::Tensor& input);
+
+std::vector<torch::Tensor> pack_nvfp4_fixed_scale_m1_cuda(
+    const torch::Tensor& input,
+    const torch::Tensor& tensor_scale);

runtime/server.py

@@ -0,0 +1,145 @@
+#!/usr/bin/env python
+"""Serve the self-contained English S2-Pro mixed NVFP4 V1 checkpoint."""
+
+from __future__ import annotations
+
+import argparse
+import json
+import multiprocessing
+import os
+import sys
+from argparse import Namespace
+from pathlib import Path
+
+import torch
+import uvicorn
+from kui.asgi import JSONResponse
+
+
+ROOT = Path(
+    os.environ.get("FISH_NVFP4_ROOT", Path(__file__).resolve().parents[1])
+).resolve()
+RUNTIME_ROOT = ROOT / "runtime"
+if str(RUNTIME_ROOT) not in sys.path:
+    sys.path.insert(0, str(RUNTIME_ROOT))
+
+from experimental.codec import (
+    load_compact_codec_model,
+    load_reference_audio_soundfile,
+    warm_reference_encoder,
+)
+from experimental.nvfp4 import launch_mixed_nvfp4_thread_safe_queue
+
+
+DEFAULT_CHECKPOINT = ROOT
+DEFAULT_WEB_UI = RUNTIME_ROOT / "web" / "index.html"
+EXPECTED_FORMAT = "fish-s2-pro-project-local-nvfp4-mixed"
+EXPECTED_POLICY = "w4a16_gate_up_middle30_mxfp8_rest"
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--checkpoint", type=Path, default=DEFAULT_CHECKPOINT)
+    parser.add_argument("--host", default="0.0.0.0")
+    parser.add_argument("--port", type=int, default=8080)
+    parser.add_argument("--device", default="cuda:0")
+    parser.add_argument("--cache-length", type=int, default=3072)
+    parser.add_argument("--max-text-length", type=int, default=0)
+    parser.add_argument("--api-key")
+    parser.add_argument("--verify-checksums", action="store_true")
+    return parser.parse_args()
+
+
+def validate(args: argparse.Namespace) -> dict:
+    if not 1 <= args.port <= 65535:
+        raise SystemExit("--port must be between 1 and 65535")
+    metadata_path = args.checkpoint / "quantization.json"
+    codec_path = args.checkpoint / "codec.pth"
+    if not metadata_path.is_file() or not codec_path.is_file():
+        raise SystemExit(f"Incomplete NVFP4 checkpoint: {args.checkpoint}")
+    metadata = json.loads(metadata_path.read_text())
+    if metadata.get("format") != EXPECTED_FORMAT:
+        raise SystemExit(f"Unsupported checkpoint format: {metadata.get('format')}")
+    if metadata.get("policy") != EXPECTED_POLICY:
+        raise SystemExit(f"Unexpected NVFP4 policy: {metadata.get('policy')}")
+    if metadata.get("fresh_load_verification", {}).get("status") != "passed":
+        raise SystemExit("Checkpoint lacks a passed fresh-load verification")
+    if not torch.cuda.is_available() or torch.cuda.get_device_capability(args.device)[0] != 12:
+        raise SystemExit("The mixed NVFP4 kernels require an SM120 CUDA GPU")
+    return metadata
+
+
+def main() -> int:
+    args = parse_args()
+    metadata = validate(args)
+
+    import tools.server.model_manager as model_manager_module
+    import tools.server.views as server_views
+    from fish_speech.inference_engine.reference_loader import ReferenceLoader
+
+    def checkpoint_queue_loader(checkpoint_path, device, precision, compile=False):
+        return launch_mixed_nvfp4_thread_safe_queue(
+            checkpoint_path,
+            device,
+            precision,
+            compile,
+            max_length=args.cache_length,
+            verify_checksums=args.verify_checksums,
+        )
+
+    def compact_codec_loader(config_name, checkpoint_path, device):
+        codec = load_compact_codec_model(
+            config_name,
+            checkpoint_path,
+            device,
+            torch.bfloat16,
+            offload_reference=True,
+        )
+        codec._reference_warmup_report = warm_reference_encoder(codec, device)
+        return codec
+
+    model_manager_module.launch_thread_safe_queue = checkpoint_queue_loader
+    model_manager_module.load_decoder_model = compact_codec_loader
+    ReferenceLoader.load_audio = staticmethod(load_reference_audio_soundfile)
+    server_views._WEBUI_HTML = DEFAULT_WEB_UI
+
+    @server_views.routes.http.get("/v1/model")
+    async def model_info():
+        return JSONResponse(
+            {
+                "model": "V1 · Fish Audio S2-Pro NVFP4 Balanced",
+                "checkpoint": str(args.checkpoint),
+                "checkpoint_status": metadata["status"],
+                "quantization": "60 native NVFP4 W4A16/W4A4 + 120 native MXFP8 W8A8",
+                "release": metadata.get("release"),
+                "policy": metadata["policy"],
+                "sampling": metadata["qualified_sampling"],
+                "fresh_load_verification": metadata["fresh_load_verification"],
+                "device": args.device,
+                "cache_length": args.cache_length,
+            }
+        )
+
+    from tools.api_server import API
+
+    upstream_args = Namespace(
+        mode="tts",
+        device=args.device,
+        half=False,
+        compile=False,
+        llama_checkpoint_path=str(args.checkpoint),
+        decoder_checkpoint_path=str(args.checkpoint / "codec.pth"),
+        decoder_config_name="modded_dac_vq",
+        max_text_length=args.max_text_length,
+        listen=f"{args.host}:{args.port}",
+        workers=1,
+        api_key=args.api_key,
+    )
+    multiprocessing.set_start_method("spawn", force=True)
+    app = API(args=upstream_args).app
+    uvicorn.run(app, host=args.host, port=args.port, workers=1, log_level="info")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

runtime/torch-overrides.txt

@@ -0,0 +1,7 @@
+torch==2.11.0
+torchaudio==2.11.0
+protobuf>=3.20.0,<6.0.0
+tensorboard==2.20.0
+transformers==4.56.1
+tokenizers==0.22.0
+huggingface-hub==0.34.4

validate_release.py

@@ -0,0 +1,293 @@
+#!/usr/bin/env python3
+"""Validate the self-contained Fish S2-Pro NVFP4/MXFP8 V1 release."""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import struct
+from pathlib import Path, PurePosixPath
+
+
+ROOT = Path(__file__).resolve().parent
+DTYPE_BYTES = {
+    "BF16": 2,
+    "F32": 4,
+    "F8_E4M3": 1,
+    "I32": 4,
+    "U8": 1,
+}
+
+
+def fail(message: str) -> None:
+    raise RuntimeError(message)
+
+
+def sha256(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as handle:
+        for chunk in iter(lambda: handle.read(8 * 1024 * 1024), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def load_json(relative: str) -> dict:
+    path = ROOT / relative
+    if not path.is_file():
+        fail(f"missing required JSON file: {relative}")
+    try:
+        return json.loads(path.read_text(encoding="utf-8"))
+    except (OSError, json.JSONDecodeError) as error:
+        fail(f"invalid JSON in {relative}: {error}")
+
+
+def safetensors_header(path: Path) -> dict:
+    with path.open("rb") as handle:
+        raw_length = handle.read(8)
+        if len(raw_length) != 8:
+            fail(f"truncated safetensors length: {path.name}")
+        header_length = struct.unpack("<Q", raw_length)[0]
+        if header_length <= 2 or header_length > path.stat().st_size - 8:
+            fail(f"invalid safetensors header length: {path.name}")
+        try:
+            return json.loads(handle.read(header_length))
+        except json.JSONDecodeError as error:
+            fail(f"invalid safetensors header in {path.name}: {error}")
+
+
+def validate_metadata() -> dict:
+    config = load_json("config.json")
+    policy = config.get("fish_s2_quantization")
+    if not isinstance(policy, dict):
+        fail("config.json has no fish_s2_quantization dictionary")
+    expected = {
+        "format": "mixed_nvfp4_mxfp8",
+        "profile": "balanced",
+        "release": "v1",
+        "policy": "w4a16_gate_up_middle30_mxfp8_rest",
+        "nvfp4_modules": 60,
+        "mxfp8_modules": 120,
+        "runtime": "bundled",
+        "hardware_family": "sm_120",
+    }
+    if policy != expected:
+        fail(f"unexpected config quantization policy: {policy}")
+
+    quantization = load_json("quantization.json")
+    if quantization.get("format") != "fish-s2-pro-project-local-nvfp4-mixed":
+        fail("unexpected quantization format")
+    if quantization.get("policy") != "w4a16_gate_up_middle30_mxfp8_rest":
+        fail("unexpected quantization policy")
+    release = quantization.get("release", {})
+    if release.get("version") != "1.0" or release.get("xpo3_release") is not False:
+        fail("release must identify V1 and explicitly remain outside XPO3")
+    if release.get("repository_name") != "ajh-code/Fish-Audio-S2-Pro-NVFP4-Balanced":
+        fail("release metadata has the wrong Hugging Face repository name")
+    if release.get("self_contained_weights") is not True:
+        fail("release does not declare self-contained weights")
+    fresh_load = quantization.get("fresh_load_verification", {})
+    if fresh_load.get("status") != "passed":
+        fail("checkpoint lacks passed fresh-load verification")
+    conversion = quantization.get("conversion", {})
+    if len(conversion.get("records", [])) != 180:
+        fail("quantization metadata must contain 180 projection records")
+    if conversion.get("correction_parameters") != 0:
+        fail("V1 loader does not accept correction-bearing checkpoints")
+    if conversion.get("w4a16_max_m") != 1:
+        fail("V1 must use the qualified M=1 W4A16 decode boundary")
+
+    source = (ROOT / ".source").read_text(encoding="utf-8")
+    for revision in (
+        "1de9996b6be38b745688de084d87a5633f714e4e",
+        "e5e292632cb11e7a27b2b7487f58f612bc101e13",
+        "a04c1b63b1a7a670840fb3e97a82c0dbe2a35ded",
+        "7a03467b90d6feff6bd196928dfe156bd173f36e",
+    ):
+        if revision not in source:
+            fail(f".source is missing pinned revision {revision}")
+
+    license_text = (ROOT / "LICENSE.md").read_text(encoding="utf-8")
+    if "FISH AUDIO RESEARCH LICENSE AGREEMENT" not in license_text:
+        fail("LICENSE.md is not the Fish Audio Research License")
+    notice = (ROOT / "Notice").read_text(encoding="utf-8")
+    if "This model is licensed under the Fish Audio Research License" not in notice:
+        fail("Notice lacks the required Fish Audio attribution")
+    if "Built with Fish Audio" not in notice:
+        fail("Notice lacks the required Built with Fish Audio statement")
+    readme = (ROOT / "README.md").read_text(encoding="utf-8")
+    for required in (
+        "Built with Fish Audio",
+        "This is not yet an XPO3 release",
+        "Commercial use requires a separate",
+    ):
+        if required not in readme:
+            fail(f"README.md lacks required release statement: {required}")
+    return quantization
+
+
+def validate_checkpoint(quantization: dict) -> tuple[int, int, int]:
+    index = load_json("model.safetensors.index.json")
+    weight_map = index.get("weight_map")
+    if not isinstance(weight_map, dict) or not weight_map:
+        fail("checkpoint index has no weight map")
+    shard_names = sorted(set(weight_map.values()))
+    if shard_names != [
+        "model-00001-of-00003.safetensors",
+        "model-00002-of-00003.safetensors",
+        "model-00003-of-00003.safetensors",
+    ]:
+        fail(f"unexpected checkpoint shards: {shard_names}")
+
+    discovered: dict[str, str] = {}
+    logical_bytes = 0
+    for shard_name in shard_names:
+        shard_path = ROOT / shard_name
+        if not shard_path.is_file():
+            fail(f"missing checkpoint shard: {shard_name}")
+        header = safetensors_header(shard_path)
+        for name, record in header.items():
+            if name == "__metadata__":
+                continue
+            if name in discovered:
+                fail(f"duplicate tensor across shards: {name}")
+            dtype = record.get("dtype")
+            shape = record.get("shape")
+            offsets = record.get("data_offsets")
+            if dtype not in DTYPE_BYTES or not isinstance(shape, list):
+                fail(f"unsupported tensor metadata for {name}")
+            if (
+                not isinstance(offsets, list)
+                or len(offsets) != 2
+                or not all(isinstance(value, int) for value in offsets)
+                or offsets[0] < 0
+                or offsets[1] < offsets[0]
+            ):
+                fail(f"invalid data offsets for {name}")
+            elements = 1
+            for dimension in shape:
+                if not isinstance(dimension, int) or dimension < 0:
+                    fail(f"invalid shape for {name}")
+                elements *= dimension
+            tensor_bytes = elements * DTYPE_BYTES[dtype]
+            if offsets[1] - offsets[0] != tensor_bytes:
+                fail(f"tensor byte range mismatch for {name}")
+            logical_bytes += tensor_bytes
+            discovered[name] = shard_name
+
+    if discovered != weight_map:
+        missing = sorted(set(weight_map) - set(discovered))
+        extra = sorted(set(discovered) - set(weight_map))
+        fail(f"checkpoint index mismatch; missing={missing[:3]} extra={extra[:3]}")
+    expected_size = int(index.get("metadata", {}).get("total_size", -1))
+    if logical_bytes != expected_size:
+        fail(f"logical checkpoint size mismatch: {logical_bytes} != {expected_size}")
+    if logical_bytes != quantization.get("state_payload_bytes"):
+        fail("quantization state_payload_bytes does not match the checkpoint")
+
+    counts = {
+        "qdata": sum(name.endswith(".qdata") for name in weight_map),
+        "weight_block_scale": sum(
+            name.endswith(".weight_block_scale") for name in weight_map
+        ),
+        "weight_scale": sum(name.endswith(".weight_scale") for name in weight_map),
+        "weight_fp8": sum(name.endswith(".weight_fp8") for name in weight_map),
+        "weight_scale_storage": sum(
+            name.endswith(".weight_scale_storage") for name in weight_map
+        ),
+    }
+    if counts != {
+        "qdata": 60,
+        "weight_block_scale": 60,
+        "weight_scale": 60,
+        "weight_fp8": 120,
+        "weight_scale_storage": 120,
+    }:
+        fail(f"unexpected packed tensor counts: {counts}")
+    if counts != quantization.get("packed_tensor_counts"):
+        fail("packed tensor counts disagree with quantization.json")
+    return len(weight_map), len(shard_names), logical_bytes
+
+
+def validate_runtime() -> None:
+    required = (
+        "codec.pth",
+        "client.py",
+        "launch.sh",
+        "install.sh",
+        "Dockerfile",
+        "compose.yaml",
+        "runtime/server.py",
+        "runtime/web/index.html",
+        "runtime/experimental/codec.py",
+        "runtime/experimental/nvfp4/checkpoint.py",
+        "runtime/experimental/nvfp4/modules.py",
+        "runtime/experimental/fp8/modules.py",
+        "runtime/native/smallm_gemv/smallm_gemv.cpp",
+        "runtime/native/smallm_gemv/smallm_gemv.cu",
+        "runtime/native/smallm_gemv/smallm_gemv.h",
+        "runtime/native/LICENSE",
+        "vendor/fish-speech/.project-root",
+        "vendor/fish-speech/LICENSE",
+        "vendor/fish-speech/pyproject.toml",
+        "vendor/fish-speech/fish_speech/configs/modded_dac_vq.yaml",
+        "vendor/fish-speech/tools/api_server.py",
+    )
+    for relative in required:
+        if not (ROOT / relative).is_file():
+            fail(f"missing required release file: {relative}")
+
+
+def validate_manifest() -> int:
+    manifest = load_json("MANIFEST.json")
+    records = manifest.get("files")
+    if not isinstance(records, list) or not records:
+        fail("MANIFEST.json has no file records")
+    seen: set[str] = set()
+    for record in records:
+        relative = record.get("path")
+        if not isinstance(relative, str):
+            fail("manifest record has no path")
+        pure = PurePosixPath(relative)
+        if pure.is_absolute() or ".." in pure.parts or relative == "MANIFEST.json":
+            fail(f"unsafe or recursive manifest path: {relative}")
+        if relative in seen:
+            fail(f"duplicate manifest path: {relative}")
+        seen.add(relative)
+        path = ROOT.joinpath(*pure.parts)
+        if not path.is_file():
+            fail(f"manifest file is missing: {relative}")
+        if path.stat().st_size != record.get("size"):
+            fail(f"manifest size mismatch: {relative}")
+        if sha256(path) != record.get("sha256"):
+            fail(f"manifest hash mismatch: {relative}")
+    return len(records)
+
+
+def main() -> None:
+    quantization = validate_metadata()
+    checkpoint_tensors, checkpoint_shards, logical_bytes = validate_checkpoint(
+        quantization
+    )
+    validate_runtime()
+    manifest_files = validate_manifest()
+    print(
+        json.dumps(
+            {
+                "status": "pass",
+                "release": "v1",
+                "xpo3_release": False,
+                "checkpoint_tensors": checkpoint_tensors,
+                "checkpoint_shards": checkpoint_shards,
+                "checkpoint_logical_bytes": logical_bytes,
+                "nvfp4_projections": 60,
+                "mxfp8_projections": 120,
+                "manifest_files": manifest_files,
+                "self_contained_weights": True,
+            },
+            indent=2,
+        )
+    )
+
+
+if __name__ == "__main__":
+    main()

vendor/fish-speech/.gitignore

@@ -0,0 +1,114 @@
+# =============================================================================
+# Fish Speech - .gitignore
+# =============================================================================
+
+# Operating System Files
+# -----------------------
+.DS_Store
+.DS_Store?
+._*
+.Spotlight-V100
+.Trashes
+ehthumbs.db
+Thumbs.db
+
+# IDEs and Editors
+# ----------------
+.vscode/
+.idea/
+*.swp
+*.swo
+*~
+
+# Python
+# ------
+__pycache__/
+*.py[cod]
+*$py.class
+*.so
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# Virtual Environments
+# --------------------
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+/fishenv/
+
+# Project Dependencies
+# --------------------
+.pdm-python
+/fish_speech.egg-info
+
+# Data and Model Files
+# --------------------
+data/
+results/
+checkpoints/
+references/
+demo-audios/
+example/
+filelists/
+*.filelist
+
+# Audio Files
+# -----------
+*.wav
+*.mp3
+*.flac
+*.ogg
+*.m4a
+
+# Data Files
+# ----------
+*.npy
+*.npz
+*.pkl
+*.pickle
+*.lab
+/fish_speech/text/cmudict_cache.pickle
+
+# Cache and Temporary Files
+# --------------------------
+/.cache/
+/.gradio/
+/.locale/
+.pgx.*
+*log
+*.log
+site/
+
+# External Tools
+# --------------
+ffmpeg.exe
+ffprobe.exe
+/faster_whisper/
+
+# Server Related
+# --------------
+/data_server/target/
+
+# Test Files
+# ----------
+/*.test.sh
+asr-label*

vendor/fish-speech/.readthedocs.yaml

@@ -0,0 +1,19 @@
+# Read the Docs configuration file for MkDocs projects
+# See https://docs.readthedocs.io/en/stable/config-file/v2.html for details
+
+# Required
+version: 2
+
+# Set the version of Python and other tools you might need
+build:
+  os: ubuntu-22.04
+  tools:
+    python: "3.12"
+
+mkdocs:
+  configuration: mkdocs.yml
+
+# Optionally declare the Python requirements required to build your docs
+python:
+  install:
+  - requirements: docs/requirements.txt

vendor/fish-speech/API_FLAGS.txt

@@ -0,0 +1,6 @@
+# --infer
+--api
+--listen 0.0.0.0:8080 \
+--llama-checkpoint-path "checkpoints/openaudio-s1-mini" \
+--decoder-checkpoint-path "checkpoints/openaudio-s1-mini/codec.pth" \
+--decoder-config-name modded_dac_vq

vendor/fish-speech/README.md

@@ -0,0 +1,197 @@
+<div align="center">
+<h1>Fish Speech</h1>
+
+**English** | [简体中文](docs/README.zh.md) | [Portuguese](docs/README.pt-BR.md) | [日本語](docs/README.ja.md) | [한국어](docs/README.ko.md) | [العربية](docs/README.ar.md) | [Español](docs/README.es.md)  <br>
+
+<a href="https://www.producthunt.com/products/fish-speech?embed=true&utm_source=badge-top-post-badge&utm_medium=badge&utm_source=badge-fish&#0045;audio&#0045;s1" target="_blank"><img src="https://api.producthunt.com/widgets/embed-image/v1/top-post-badge.svg?post_id=1023740&theme=light&period=daily&t=1761164814710" alt="Fish&#0032;Audio&#0032;S1 - Expressive&#0032;Voice&#0032;Cloning&#0032;and&#0032;Text&#0045;to&#0045;Speech | Product Hunt" style="width: 250px; height: 54px;" width="250" height="54" /></a>
+<a href="https://trendshift.io/repositories/7014" target="_blank">
+    <img src="https://trendshift.io/api/badge/repositories/7014" alt="fishaudio%2Ffish-speech | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/>
+</a>
+<br>
+</div>
+<br>
+
+<div align="center">
+    <img src="https://count.getloli.com/get/@fish-speech?theme=asoul" /><br>
+</div>
+
+<br>
+
+<div align="center">
+    <a target="_blank" href="https://discord.gg/Es5qTB9BcN">
+        <img alt="Discord" src="https://img.shields.io/discord/1214047546020728892?color=%23738ADB&label=Discord&logo=discord&logoColor=white&style=flat-square"/>
+    </a>
+    <a target="_blank" href="https://hub.docker.com/r/fishaudio/fish-speech">
+        <img alt="Docker" src="https://img.shields.io/docker/pulls/fishaudio/fish-speech?style=flat-square&logo=docker"/>
+    </a>
+    <a target="_blank" href="https://pd.qq.com/s/bwxia254o">
+      <img alt="QQ Channel" src="https://img.shields.io/badge/QQ-blue?logo=tencentqq">
+    </a>
+</div>
+
+<div align="center">
+    <a target="_blank" href="https://huggingface.co/fishaudio/s2-pro">
+        <img alt="HuggingFace Model" src="https://img.shields.io/badge/🤗%20-models-orange"/>
+    </a>
+    <a target="_blank" href="https://fish.audio/blog/fish-audio-open-sources-s2/">
+        <img alt="Fish Audio Blog" src="https://img.shields.io/badge/Blog-Fish_Audio_S2-1f7a8c?style=flat-square&logo=readme&logoColor=white"/>
+    </a>
+    <a target="_blank" href="https://arxiv.org/abs/2603.08823">
+        <img alt="Paper | Technical Report" src="https://img.shields.io/badge/Paper-Technical_Report-b31b1b?style=flat-square"/>
+    </a>
+</div>
+
+> [!IMPORTANT]
+> **License Notice**  
+> This codebase and its associated model weights are released under **[FISH AUDIO RESEARCH LICENSE](LICENSE)**. Please refer to [LICENSE](LICENSE) for more details. We will take action against any violation of the license.
+
+> [!WARNING]
+> **Legal Disclaimer**  
+> We do not hold any responsibility for any illegal usage of the codebase. Please refer to your local laws about DMCA and other related laws.
+
+## Quick Start
+
+### For Human
+
+Here are the official documents for Fish Audio S2, follow the instructions to get started easily.
+
+- [Installation](https://speech.fish.audio/install/)
+- [Command Line Inference](https://speech.fish.audio/inference/#command-line-inference)
+- [WebUI Inference](https://speech.fish.audio/inference/#webui-inference)
+- [Server Inference](https://speech.fish.audio/server/)
+- [Docker Setup](https://speech.fish.audio/install/#docker-setup)
+
+> [!IMPORTANT]
+> **For SGLang server, please read [SGLang-Omni README](https://github.com/sgl-project/sglang-omni/blob/main/sglang_omni/models/fishaudio_s2_pro/README.md).**
+>
+> **For vLLM Omni server, please read [vLLM-Omni Fish Speech S2 Pro Recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/fishaudio/Fish-Speech-S2-Pro.md) and the [User Guide](https://github.com/vllm-project/vllm-omni/blob/main/docs/user_guide/examples/online_serving/text_to_speech.md#fish-speech-s2-pro).**
+
+### For LLM Agent
+
+```
+Install and configure Fish-Audio S2 by following the instructions here: https://speech.fish.audio/install/
+```
+
+## Fish Audio S2 Pro
+**State-of-the-art multilingual text-to-speech (TTS) system, redefining the boundaries of voice generation.**
+
+Fish Audio S2 Pro is the most advanced multimodal model developed by [Fish Audio](https://fish.audio/). Trained on over **10 million hours** of audio data covering more than **80 languages**, S2 Pro combines a **Dual-Autoregressive (Dual-AR)** architecture with reinforcement learning (RL) alignment to generate speech that is exceptionally natural, realistic, and emotionally rich, leading the competition among both open-source and closed-source systems.
+
+The core strength of S2 Pro lies in its support for **sub-word level** fine-grained control of prosody and emotion using natural language tags (e.g., `[whisper]`, `[excited]`, `[angry]`), while natively supporting multi-speaker and multi-turn conversation generation.
+
+Visit the [Fish Audio website](https://fish.audio/) for a live playground, or read our [technical report](https://arxiv.org/abs/2603.08823) and [blog post](https://fish.audio/blog/fish-audio-open-sources-s2/) for more details.
+
+### Model Variants
+
+| Model | Size | Availability | Description |
+|------|------|-------------|-------------|
+| S2-Pro | 4B parameters | [HuggingFace](https://huggingface.co/fishaudio/s2-pro) | Full-featured flagship model with maximum quality and stability |
+
+More details of the model can be found in the [technical report](https://arxiv.org/abs/2411.01156).
+
+## Benchmark Results
+
+| Benchmark | Fish Audio S2 |
+|------|------|
+| Seed-TTS Eval — WER (Chinese) | **0.54%** (best overall) |
+| Seed-TTS Eval — WER (English) | **0.99%** (best overall) |
+| Audio Turing Test (with instruction) | **0.515** posterior mean |
+| EmergentTTS-Eval — Win Rate | **81.88%** (highest overall) |
+| Fish Instruction Benchmark — TAR | **93.3%** |
+| Fish Instruction Benchmark — Quality | **4.51 / 5.0** |
+| Multilingual (MiniMax Testset) — Best WER | **11 of 24** languages |
+| Multilingual (MiniMax Testset) — Best SIM | **17 of 24** languages |
+
+On Seed-TTS Eval, S2 achieves the lowest WER among all evaluated models including closed-source systems: Qwen3-TTS (0.77/1.24), MiniMax Speech-02 (0.99/1.90), Seed-TTS (1.12/2.25). On the Audio Turing Test, 0.515 surpasses Seed-TTS (0.417) by 24% and MiniMax-Speech (0.387) by 33%. On EmergentTTS-Eval, S2 achieves particularly strong results in paralinguistics (91.61% win rate), questions (84.41%), and syntactic complexity (83.39%).
+
+## Highlights
+
+<img src="./docs/assets/totalability.png" width=200%>
+
+### Fine-Grained Inline Control via Natural Language
+
+S2 Pro brings unprecedented "soul" to speech. Using simple `[tag]` syntax, you can precisely embed emotional instructions at any position in the text.
+- **15,000+ Unique Tags Supported**: Not limited to fixed presets; S2 supports **free-form text descriptions**. Try `[whisper in small voice]`, `[professional broadcast tone]`, or `[pitch up]`.
+- **Rich Emotion Library**:
+  `[pause]` `[emphasis]` `[laughing]` `[inhale]` `[chuckle]` `[tsk]` `[singing]` `[excited]` `[laughing tone]` `[interrupting]` `[chuckling]` `[excited tone]` `[volume up]` `[echo]` `[angry]` `[low volume]` `[sigh]` `[low voice]` `[whisper]` `[screaming]` `[shouting]` `[loud]` `[surprised]` `[short pause]` `[exhale]` `[delight]` `[panting]` `[audience laughter]` `[with strong accent]` `[volume down]` `[clearing throat]` `[sad]` `[moaning]` `[shocked]`
+
+### Innovative Dual-Autoregressive (Dual-AR) Architecture
+
+S2 Pro adopts a master-slave Dual-AR architecture consisting of a decoder-only transformer and an RVQ audio codec (10 codebooks, ~21 Hz):
+
+- **Slow AR (4B parameters)**: Operates along the time axis, predicting the primary semantic codebook.
+- **Fast AR (400M parameters)**: Generates the remaining 9 residual codebooks at each time step, reconstructing exquisite acoustic details.
+
+This asymmetric design achieves peak audio fidelity while significantly boosting inference speed.
+
+### Reinforcement Learning (RL) Alignment
+
+S2 Pro utilizes **Group Relative Policy Optimization (GRPO)** for post-training alignment. We use the same model suite for data cleaning and annotation directly as Reward Models, perfectly resolving the distribution mismatch between pre-training data and post-training objectives.
+- **Multi-Dimensional Reward Signals**: Comprehensively evaluates semantic accuracy, instruction adherence, acoustic preference scoring, and timbre similarity to ensure every second of generated speech feels intuitive to humans.
+
+### Extreme Streaming Performance (Powered by SGLang)
+
+As the Dual-AR architecture is structurally isomorphic to standard LLMs, S2 Pro natively supports all SGLang inference acceleration features, including Continuous Batching, Paged KV Cache, CUDA Graph, and RadixAttention-based Prefix Caching.
+
+**Performance on a single NVIDIA H200 GPU:**
+- **Real-Time Factor (RTF)**: 0.195
+- **Time-to-First-Audio (TTFA)**: ~100 ms
+- **Extreme Throughput**: 3,000+ acoustic tokens/s while maintaining RTF < 0.5
+
+### Robust Multilingual Support
+
+S2 Pro supports over 80 languages without requiring phonemes or language-specific preprocessing:
+
+- **Tier 1**: Japanese (ja), English (en), Chinese (zh)
+- **Tier 2**: Korean (ko), Spanish (es), Portuguese (pt), Arabic (ar), Russian (ru), French (fr), German (de)
+- **Global Coverage**: sv, it, tr, no, nl, cy, eu, ca, da, gl, ta, hu, fi, pl, et, hi, la, ur, th, vi, jw, bn, yo, xsl, cs, sw, nn, he, ms, uk, id, kk, bg, lv, my, tl, sk, ne, fa, af, el, bo, hr, ro, sn, mi, yi, am, be, km, is, az, sd, br, sq, ps, mn, ht, ml, sr, sa, te, ka, bs, pa, lt, kn, si, hy, mr, as, gu, fo, etc.
+
+### Native Multi-Speaker Generation
+
+<img src="./docs/assets/chattemplate.png" width=200%>
+
+Fish Audio S2 allows users to upload reference audio containing multiple speakers, and the model processes each speaker's features via the `<|speaker:i|>` token. You can then control the model's performance via speaker ID tokens, enabling a single generation to include multiple speakers. There is no longer a need to upload separate reference audio for each individual speaker.
+
+### Multi-Turn Generation
+
+Thanks to the expansion of the model context, our model can now leverage previous information to improve the expressiveness of subsequent generated content, thereby increasing the naturalness of the dialogue.
+
+### Rapid Voice Cloning
+
+Fish Audio S2 supports accurate voice cloning using short reference samples (typically 10-30 seconds). The model captures timbre, speaking style, and emotional tendencies, producing realistic and consistent cloned voices without additional fine-tuning.
+For SGLang Server usage, please refer to the [SGLang-Omni README](https://github.com/sgl-project/sglang-omni/blob/main/sglang_omni/models/fishaudio_s2_pro/README.md).
+
+---
+
+## Credits
+
+- [VITS2 (daniilrobnikov)](https://github.com/daniilrobnikov/vits2)
+- [Bert-VITS2](https://github.com/fishaudio/Bert-VITS2)
+- [GPT VITS](https://github.com/innnky/gpt-vits)
+- [MQTTS](https://github.com/b04901014/MQTTS)
+- [GPT Fast](https://github.com/pytorch-labs/gpt-fast)
+- [GPT-SoVITS](https://github.com/RVC-Boss/GPT-SoVITS)
+- [Qwen3](https://github.com/QwenLM/Qwen3)
+
+## Tech Report
+```bibtex
+@misc{fish-speech-v1.4,
+      title={Fish-Speech: Leveraging Large Language Models for Advanced Multilingual Text-to-Speech Synthesis},
+      author={Shijia Liao and Yuxuan Wang and Tianyu Li and Yifan Cheng and Ruoyi Zhang and Rongzhi Zhou and Yijin Xing},
+      year={2024},
+      eprint={2411.01156},
+      archivePrefix={arXiv},
+      primaryClass={cs.SD},
+      url={https://arxiv.org/abs/2411.01156},
+}
+
+@misc{liao2026fishaudios2technical,
+      title={Fish Audio S2 Technical Report}, 
+      author={Shijia Liao and Yuxuan Wang and Songting Liu and Yifan Cheng and Ruoyi Zhang and Tianyu Li and Shidong Li and Yisheng Zheng and Xingwei Liu and Qingzheng Wang and Zhizhuo Zhou and Jiahua Liu and Xin Chen and Dawei Han},
+      year={2026},
+      eprint={2603.08823},
+      archivePrefix={arXiv},
+      primaryClass={cs.SD},
+      url={https://arxiv.org/abs/2603.08823}, 
+}
+```

vendor/fish-speech/compose.base.yml

@@ -0,0 +1,25 @@
+services:
+  app-base:
+    build:
+      context: .
+      dockerfile: docker/Dockerfile
+      args:
+        BACKEND: ${BACKEND:-cuda}     # or cpu
+        CUDA_VER: ${CUDA_VER:-12.9.0}
+        UV_EXTRA: ${UV_EXTRA:-cu129}
+        UV_VERSION: ${UV_VERSION:-0.8.15}
+    volumes:
+      - ./checkpoints:/app/checkpoints
+      - ./references:/app/references
+    environment:
+      COMPILE: ${COMPILE:-0}
+    # GPU (remove this block if CPU-only):
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: all
+              capabilities: [gpu]
+    tty: true
+    stdin_open: true

vendor/fish-speech/compose.rocm.yml

@@ -0,0 +1,58 @@
+name: fish-speech-rocm
+
+# AMD ROCm compose for Fish Speech (RDNA3 / RDNA4).
+# Mount your checkpoints into ./checkpoints before running.
+#
+#   docker compose -f compose.rocm.yml --profile webui up --build
+#   docker compose -f compose.rocm.yml --profile server up --build
+
+services:
+  webui:
+    build:
+      context: .
+      dockerfile: docker/Dockerfile.rocm
+      target: webui
+    image: fish-speech-webui:rocm
+    profiles: ["webui"]
+    ports:
+      - "${GRADIO_PORT:-7860}:7860"
+    volumes:
+      - ./checkpoints:/app/checkpoints
+      - ./references:/app/references
+    environment:
+      - ROCBLAS_USE_HIPBLASLT=0
+      - COMPILE=${COMPILE:-1}
+    devices:
+      - /dev/kfd
+      - /dev/dri
+    group_add:
+      - video
+      - render
+    shm_size: "16g"
+    tty: true
+    stdin_open: true
+
+  server:
+    build:
+      context: .
+      dockerfile: docker/Dockerfile.rocm
+      target: server
+    image: fish-speech-server:rocm
+    profiles: ["server"]
+    ports:
+      - "${API_PORT:-8080}:8080"
+    volumes:
+      - ./checkpoints:/app/checkpoints
+      - ./references:/app/references
+    environment:
+      - ROCBLAS_USE_HIPBLASLT=0
+      - COMPILE=${COMPILE:-1}
+    devices:
+      - /dev/kfd
+      - /dev/dri
+    group_add:
+      - video
+      - render
+    shm_size: "16g"
+    tty: true
+    stdin_open: true

vendor/fish-speech/compose.yml

@@ -0,0 +1,26 @@
+name: fish-speech
+
+services:
+  webui:
+    extends:
+      file: compose.base.yml
+      service: app-base
+    build:
+      target: webui
+    environment:
+      COMPILE: ${COMPILE:-0}
+    profiles: ["webui"]
+    ports:
+      - "${GRADIO_PORT:-7860}:7860"
+
+  server:
+    extends:
+      file: compose.base.yml
+      service: app-base
+    build:
+      target: server
+    environment:
+      COMPILE: ${COMPILE:-0}
+    profiles: ["server"]
+    ports:
+      - "${API_PORT:-8080}:8080"

vendor/fish-speech/dockerfile.dev

@@ -0,0 +1,37 @@
+ARG VERSION=dev
+ARG BASE_IMAGE=ghcr.io/fishaudio/fish-speech:${VERSION}
+
+FROM ${BASE_IMAGE}
+
+ARG TOOLS="               \
+        git               \
+        curl              \
+        build-essential   \
+        ffmpeg            \
+        libsm6            \
+        libxext6          \
+        libjpeg-dev       \
+        zlib1g-dev        \
+        aria2             \
+        zsh               \
+        openssh-server    \
+        sudo              \
+        protobuf-compiler \
+        libasound-dev     \
+        portaudio19-dev   \
+        libportaudio2     \
+        libportaudiocpp0  \
+        cmake"
+
+RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
+    --mount=type=cache,target=/var/lib/apt,sharing=locked \
+    set -ex \
+    && apt-get update \
+    && apt-get -y install --no-install-recommends ${TOOLS}
+
+# Install oh-my-zsh so your terminal looks nice
+RUN sh -c "$(curl https://raw.githubusercontent.com/robbyrussell/oh-my-zsh/master/tools/install.sh)" "" --unattended
+
+# Set zsh as default shell
+RUN chsh -s /usr/bin/zsh
+ENV SHELL=/usr/bin/zsh

vendor/fish-speech/entrypoint.sh

@@ -0,0 +1,10 @@
+#!/bin/bash
+
+CUDA_ENABLED=${CUDA_ENABLED:-true}
+DEVICE=""
+
+if [ "${CUDA_ENABLED}" != "true" ]; then
+    DEVICE="--device cpu"
+fi
+
+exec python tools/run_webui.py ${DEVICE}

vendor/fish-speech/inference.ipynb

@@ -0,0 +1,214 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Fish Speech"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### For Windows User / win用户"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "vscode": {
+     "languageId": "bat"
+    }
+   },
+   "outputs": [],
+   "source": [
+    "!chcp 65001"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### For Linux User / Linux 用户"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import locale\n",
+    "locale.setlocale(locale.LC_ALL, 'en_US.UTF-8')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### Prepare Model"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# For Chinese users, you probably want to use mirror to accelerate downloading\n",
+    "# !set HF_ENDPOINT=https://hf-mirror.com\n",
+    "# !export HF_ENDPOINT=https://hf-mirror.com \n",
+    "\n",
+    "!hf download fishaudio/openaudio-s1-mini --local-dir checkpoints/openaudio-s1-mini/"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## WebUI Inference\n",
+    "\n",
+    "> You can use --compile to fuse CUDA kernels for faster inference (10x)."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "vscode": {
+     "languageId": "shellscript"
+    }
+   },
+   "outputs": [],
+   "source": [
+    "!python tools/run_webui.py \\\n",
+    "    --llama-checkpoint-path checkpoints/openaudio-s1-mini \\\n",
+    "    --decoder-checkpoint-path checkpoints/openaudio-s1-mini/codec.pth \\\n",
+    "    # --compile"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Break-down CLI Inference"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### 1. Encode reference audio: / 从语音生成 prompt: \n",
+    "\n",
+    "You should get a `fake.npy` file.\n",
+    "\n",
+    "你应该能得到一个 `fake.npy` 文件."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "vscode": {
+     "languageId": "shellscript"
+    }
+   },
+   "outputs": [],
+   "source": [
+    "## Enter the path to the audio file here\n",
+    "src_audio = r\"D:\\PythonProject\\vo_hutao_draw_appear.wav\"\n",
+    "\n",
+    "!python fish_speech/models/dac/inference.py \\\n",
+    "    -i {src_audio} \\\n",
+    "    --checkpoint-path \"checkpoints/openaudio-s1-mini/codec.pth\"\n",
+    "\n",
+    "from IPython.display import Audio, display\n",
+    "audio = Audio(filename=\"fake.wav\")\n",
+    "display(audio)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### 2. Generate semantic tokens from text: / 从文本生成语义 token:\n",
+    "\n",
+    "> This command will create a codes_N file in the working directory, where N is an integer starting from 0.\n",
+    "\n",
+    "> You may want to use `--compile` to fuse CUDA kernels for faster inference (~30 tokens/second -> ~300 tokens/second).\n",
+    "\n",
+    "> 该命令会在工作目录下创建 codes_N 文件, 其中 N 是从 0 开始的整数.\n",
+    "\n",
+    "> 您可以使用 `--compile` 来融合 cuda 内核以实现更快的推理 (~30 tokens/秒 -> ~300 tokens/秒)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "vscode": {
+     "languageId": "shellscript"
+    }
+   },
+   "outputs": [],
+   "source": [
+    "!python fish_speech/models/text2semantic/inference.py \\\n",
+    "    --text \"hello world\" \\\n",
+    "    --prompt-text \"The text corresponding to reference audio\" \\\n",
+    "    --prompt-tokens \"fake.npy\" \\\n",
+    "    --checkpoint-path \"checkpoints/openaudio-s1-mini\" \\\n",
+    "    --num-samples 2\n",
+    "    # --compile"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### 3. Generate speech from semantic tokens: / 从语义 token 生成人声:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "vscode": {
+     "languageId": "shellscript"
+    }
+   },
+   "outputs": [],
+   "source": [
+    "!python fish_speech/models/dac/inference.py \\\n",
+    "    -i \"codes_0.npy\" \\\n",
+    "    --checkpoint-path \"checkpoints/openaudio-s1-mini/codec.pth\"\n",
+    "\n",
+    "from IPython.display import Audio, display\n",
+    "audio = Audio(filename=\"fake.wav\")\n",
+    "display(audio)"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.14"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}

vendor/fish-speech/mkdocs.yml

@@ -0,0 +1,155 @@
+site_name: Fish Audio
+site_description: Targeting SOTA TTS solutions.
+site_url: https://speech.fish.audio
+
+# Repository
+repo_name: fishaudio/fish-speech
+repo_url: https://github.com/fishaudio/fish-speech
+edit_uri: blob/main/docs
+
+# Copyright
+copyright: Copyright © 2023-2025 by Fish Audio
+
+theme:
+  name: material
+  favicon: assets/logo.svg
+  language: en
+  features:
+    - content.action.edit
+    - content.action.view
+    - navigation.tracking
+    - navigation.footer
+    # - navigation.tabs
+    - search
+    - search.suggest
+    - search.highlight
+    - search.share
+    - content.code.copy
+  logo: assets/logo.svg
+
+  palette:
+    # Palette toggle for automatic mode
+    - media: "(prefers-color-scheme)"
+      toggle:
+        icon: material/brightness-auto
+        name: Switch to light mode
+
+    # Palette toggle for light mode
+    - media: "(prefers-color-scheme: light)"
+      scheme: default
+      toggle:
+        icon: material/brightness-7
+        name: Switch to dark mode
+      primary: black
+      font:
+        code: Roboto Mono
+
+    # Palette toggle for dark mode
+    - media: "(prefers-color-scheme: dark)"
+      scheme: slate
+      toggle:
+        icon: material/brightness-4
+        name: Switch to light mode
+      primary: black
+      font:
+        code: Roboto Mono
+
+nav:
+  - Introduction: en/index.md
+  - Installation: en/install.md
+  - Finetune: en/finetune.md
+  - Inference: en/inference.md
+  - Server: en/server.md
+  - Samples: en/samples.md
+
+# Plugins
+plugins:
+  - search:
+      separator: '[\s\-,:!=\[\]()"`/]+|\.(?!\d)|&[lg]t;|(?!\b)(?=[A-Z][a-z])'
+      lang:
+        - en
+        - zh
+        - ja
+        - pt
+        - ko
+        - ar
+  - i18n:
+      docs_structure: folder
+      languages:
+        - locale: en
+          name: English
+          default: true
+          build: true
+        - locale: zh
+          name: 简体中文
+          build: true
+          nav:
+            - 介绍: zh/index.md
+            - 安装: zh/install.md
+            - 微调: zh/finetune.md
+            - 推理: zh/inference.md
+            - 示例: zh/samples.md
+        - locale: ja
+          name: 日本語
+          build: true
+          nav:
+            - はじめに: ja/index.md
+            - インストール: ja/install.md
+            - ファインチューニング: ja/finetune.md
+            - 推論: ja/inference.md
+            - サンプル: ja/samples.md
+        - locale: pt
+          name: Português (Brasil)
+          build: true
+          nav:
+            - Introdução: pt/index.md
+            - Instalação: pt/install.md
+            - Ajuste Fino: pt/finetune.md
+            - Inferência: pt/inference.md
+            - Amostras: pt/samples.md
+        - locale: ko
+          name: 한국어
+          build: true
+          nav:
+            - 소개: ko/index.md
+            - 설치: ko/install.md
+            - 파인튜닝: ko/finetune.md
+            - 추론: ko/inference.md
+            - 샘플: ko/samples.md
+        - locale: ar
+          name: العربية
+          build: true
+          nav:
+            - مقدمة: ar/index.md
+            - التثبيت: ar/install.md
+            - الضبط الدقيق: ar/finetune.md
+            - الاستنتاج: ar/inference.md
+            - العينات: ar/samples.md
+
+markdown_extensions:
+  - pymdownx.highlight:
+      anchor_linenums: true
+      line_spans: __span
+      pygments_lang_class: true
+  - pymdownx.inlinehilite
+  - pymdownx.snippets
+  - pymdownx.superfences
+  - admonition
+  - pymdownx.details
+  - pymdownx.superfences
+  - attr_list
+  - md_in_html
+  - pymdownx.superfences
+
+extra_css:
+  - stylesheets/extra.css
+
+extra:
+  social:
+    - icon: fontawesome/brands/discord
+      link: https://discord.gg/Es5qTB9BcN
+    - icon: fontawesome/brands/docker
+      link: https://hub.docker.com/r/fishaudio/fish-speech
+    - icon: fontawesome/brands/qq
+      link: http://qm.qq.com/cgi-bin/qm/qr?_wv=1027&k=jCKlUP7QgSm9kh95UlBoYv6s1I-Apl1M&authKey=xI5ttVAp3do68IpEYEalwXSYZFdfxZSkah%2BctF5FIMyN2NqAa003vFtLqJyAVRfF&noverify=0&group_code=593946093
+  homepage: https://speech.fish.audio

vendor/fish-speech/pyrightconfig.json

@@ -0,0 +1,6 @@
+{
+    "exclude": [
+        "data",
+        "filelists"
+    ]
+}