fp4local-abliterated 0.1.0__tar.gz

fp4local_abliterated-0.1.0/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

fp4local_abliterated-0.1.0/PKG-INFO

@@ -0,0 +1,97 @@
+Metadata-Version: 2.4
+Name: fp4local-abliterated
+Version: 0.1.0
+Summary: vLLM plugin for inference-time abliteration, with FP4-Local checkpoint loading
+License-Expression: Apache-2.0
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: vllm>=0.32.1rc0
+Dynamic: license-file
+
+# fp4local-abliterated
+
+A vLLM plugin that removes a refusal direction at inference time, so quantized checkpoints can be abliterated without touching (or re-quantizing) their weights. It was written for [Qwen3.8-Flash-Next-FP4-Local](https://huggingface.co/nvidia/Qwen3.8-Flash-Next-FP4-Local), whose QAD-trained NVFP4 and MXFP8 weights would lose accuracy if they were dequantized, edited and quantized again.
+
+At each residual-stream writer listed for the architecture the plugin replaces the output `x` with `x − λ (vᵀx) v`, computed in fp32. That is the same as editing each writer to `W − λ v vᵀ W`, which is what weight-editing abliteration does. Residual writes outside that list are left as they are; for Qwen4Exp these are the PLE short-convolution output and the vision tower's embeddings and deepstack inputs.
+
+The plugin also loads FP4-Local's CSF-compressed NVFP4 scales and its `tensors/` shard layout, which vLLM releases do not read yet. That part is adapted from vLLM pull request #60646, wraps two internal methods of vLLM's default loader, and switches itself off once vLLM ships the feature.
+
+## Requirements
+
+vLLM 0.32.1rc0 or newer, declared as a dependency. FP4-Local stores its PLE table in NVFP4, which vLLM reads since vllm-project/vllm#56273. vLLM 0.32.0 predates that change and allocates the table as bf16, 95 GiB instead of 27 GiB; with `--engram-config '{"cpu_offload": true}'` that is pinned host memory, which can exhaust the host. Until 0.32.1 is on PyPI, install a nightly build:
+
+```bash
+uv pip install -U vllm --pre --torch-backend=auto --extra-index-url https://wheels.vllm.ai/nightly
+```
+
+For a custom vLLM build whose version string does not satisfy the requirement, install the plugin with `--no-deps`; the build must include #56273.
+
+## Making an abliterated checkpoint
+
+The direction is an input; this package does not compute it. Given one (a 1-D tensor of `hidden_size` values in a `.pt` or `.safetensors` file):
+
+```bash
+pip install git+https://github.com/lesj0610/fp4local-abliterated
+fp4local-abliterate <base checkpoint> <output dir> --direction v.pt --lambda 1.5 \
+  --source "where the direction came from"
+```
+
+The output directory holds the base files (hard-linked when on the same filesystem) and a `config.json` with one added block, the same way quantizers add `quantization_config`:
+
+```json
+"abliteration_config": {"lambda": 1.5, "direction": [...], "source": "..."}
+```
+
+`direction` is stored normalized. λ = 1 removes the component along it; λ > 1 reverses part of it. The MTP draft model reads the same block, so target and draft stay consistent. To try another λ, write a new checkpoint; `--hf-overrides` only reaches the target model, not the draft.
+
+Do not add the block to a checkpoint whose weights were already abliterated; the projection would be applied twice.
+
+## Serving
+
+```bash
+vllm serve <output dir> \
+  --engram-config '{"cpu_offload": true}' \
+  --speculative-config '{"method": "mtp", "num_speculative_tokens": 3}'
+```
+
+vLLM loads the plugin through its `vllm.general_plugins` entry point; no flag is needed. Checkpoints without `abliteration_config` run unchanged.
+
+`--engram-config` is optional. It moves the PLE table from GPU memory into pinned host memory, which cannot be swapped out: plan for 36 GiB of free host RAM (the 27 GiB table, rounded up to powers of two by PyTorch's pinned-memory allocator).
+
+With `--tensor-parallel-size` above 1, add `--mm-encoder-tp-mode data`. The vision MLP's intermediate size (4304) cannot be split across GPUs into NVFP4 blocks of 16, so each GPU keeps its own copy of the vision tower instead.
+
+At startup each worker logs one line for the model and one for the MTP draft:
+
+```
+Abliteration on AbliteratedQwen4ExpForConditionalGeneration: lambda 1.5, ... residual writes in ... layers
+Abliteration on AbliteratedQwen4ExpMTP: lambda 1.5, 5 residual writes in 1 layers
+```
+
+If the line is missing, the model is being served without abliteration.
+
+## Architectures
+
+| vLLM architecture | Models | Tested with |
+|---|---|---|
+| Qwen4Exp (`Qwen4ExpForCausalLM`, `Qwen4ExpForConditionalGeneration`, `Qwen4ExpMTP`) | Qwen3.8-Flash-Next family | `nvidia/Qwen3.8-Flash-Next-FP4-Local` on vLLM 0.32.1rc1.dev467 (nightly), sm_80, tensor parallel, MTP on |
+
+Each architecture is one entry in [`adapters.py`](src/fp4local_abliterated/adapters.py): the vLLM module that defines it, the decoder layer class, and which submodules write into the residual stream. Everything else is shared. The plugin registers an `Abliterated<Arch>` subclass for every listed architecture through `ModelRegistry`, and attaches forward hooks after vLLM builds the model.
+
+To add an architecture, list its residual writers (embedding, attention output, MLP or MoE output, and anything else added to the residual) as they are named in vLLM's model code, then check that the startup log counts one write per writer per layer.
+
+## Where the direction for FP4-Local came from
+
+The direction used for [lesj0610/Qwen3.8-Flash-Next-FP4-Local-Abliterated](https://huggingface.co/lesj0610/Qwen3.8-Flash-Next-FP4-Local-Abliterated) was not captured by us. It is the direction of [windowsxp811203/Qwen3.8-Flash-Next-Abliterated](https://huggingface.co/windowsxp811203/Qwen3.8-Flash-Next-Abliterated), a bf16 weight-edited build of the same base model, recovered exactly from the difference between its edited and original weights (the difference is rank-1). Their model card describes how it was captured: difference of mean residual activations on harmful and harmless prompts, after 37 of the 48 decoder layers. The method is from Arditi et al., [Refusal in Language Models Is Mediated by a Single Direction](https://arxiv.org/abs/2406.11717) (2024).
+
+## Tests
+
+```bash
+pytest tests/
+```
+
+The CSF and CLI tests run on CPU. The projection tests need a CUDA GPU.
+
+## License
+
+Apache-2.0. `csf_scales.py` is adapted from vLLM (Apache-2.0).

fp4local_abliterated-0.1.0/README.md

@@ -0,0 +1,86 @@
+# fp4local-abliterated
+
+A vLLM plugin that removes a refusal direction at inference time, so quantized checkpoints can be abliterated without touching (or re-quantizing) their weights. It was written for [Qwen3.8-Flash-Next-FP4-Local](https://huggingface.co/nvidia/Qwen3.8-Flash-Next-FP4-Local), whose QAD-trained NVFP4 and MXFP8 weights would lose accuracy if they were dequantized, edited and quantized again.
+
+At each residual-stream writer listed for the architecture the plugin replaces the output `x` with `x − λ (vᵀx) v`, computed in fp32. That is the same as editing each writer to `W − λ v vᵀ W`, which is what weight-editing abliteration does. Residual writes outside that list are left as they are; for Qwen4Exp these are the PLE short-convolution output and the vision tower's embeddings and deepstack inputs.
+
+The plugin also loads FP4-Local's CSF-compressed NVFP4 scales and its `tensors/` shard layout, which vLLM releases do not read yet. That part is adapted from vLLM pull request #60646, wraps two internal methods of vLLM's default loader, and switches itself off once vLLM ships the feature.
+
+## Requirements
+
+vLLM 0.32.1rc0 or newer, declared as a dependency. FP4-Local stores its PLE table in NVFP4, which vLLM reads since vllm-project/vllm#56273. vLLM 0.32.0 predates that change and allocates the table as bf16, 95 GiB instead of 27 GiB; with `--engram-config '{"cpu_offload": true}'` that is pinned host memory, which can exhaust the host. Until 0.32.1 is on PyPI, install a nightly build:
+
+```bash
+uv pip install -U vllm --pre --torch-backend=auto --extra-index-url https://wheels.vllm.ai/nightly
+```
+
+For a custom vLLM build whose version string does not satisfy the requirement, install the plugin with `--no-deps`; the build must include #56273.
+
+## Making an abliterated checkpoint
+
+The direction is an input; this package does not compute it. Given one (a 1-D tensor of `hidden_size` values in a `.pt` or `.safetensors` file):
+
+```bash
+pip install git+https://github.com/lesj0610/fp4local-abliterated
+fp4local-abliterate <base checkpoint> <output dir> --direction v.pt --lambda 1.5 \
+  --source "where the direction came from"
+```
+
+The output directory holds the base files (hard-linked when on the same filesystem) and a `config.json` with one added block, the same way quantizers add `quantization_config`:
+
+```json
+"abliteration_config": {"lambda": 1.5, "direction": [...], "source": "..."}
+```
+
+`direction` is stored normalized. λ = 1 removes the component along it; λ > 1 reverses part of it. The MTP draft model reads the same block, so target and draft stay consistent. To try another λ, write a new checkpoint; `--hf-overrides` only reaches the target model, not the draft.
+
+Do not add the block to a checkpoint whose weights were already abliterated; the projection would be applied twice.
+
+## Serving
+
+```bash
+vllm serve <output dir> \
+  --engram-config '{"cpu_offload": true}' \
+  --speculative-config '{"method": "mtp", "num_speculative_tokens": 3}'
+```
+
+vLLM loads the plugin through its `vllm.general_plugins` entry point; no flag is needed. Checkpoints without `abliteration_config` run unchanged.
+
+`--engram-config` is optional. It moves the PLE table from GPU memory into pinned host memory, which cannot be swapped out: plan for 36 GiB of free host RAM (the 27 GiB table, rounded up to powers of two by PyTorch's pinned-memory allocator).
+
+With `--tensor-parallel-size` above 1, add `--mm-encoder-tp-mode data`. The vision MLP's intermediate size (4304) cannot be split across GPUs into NVFP4 blocks of 16, so each GPU keeps its own copy of the vision tower instead.
+
+At startup each worker logs one line for the model and one for the MTP draft:
+
+```
+Abliteration on AbliteratedQwen4ExpForConditionalGeneration: lambda 1.5, ... residual writes in ... layers
+Abliteration on AbliteratedQwen4ExpMTP: lambda 1.5, 5 residual writes in 1 layers
+```
+
+If the line is missing, the model is being served without abliteration.
+
+## Architectures
+
+| vLLM architecture | Models | Tested with |
+|---|---|---|
+| Qwen4Exp (`Qwen4ExpForCausalLM`, `Qwen4ExpForConditionalGeneration`, `Qwen4ExpMTP`) | Qwen3.8-Flash-Next family | `nvidia/Qwen3.8-Flash-Next-FP4-Local` on vLLM 0.32.1rc1.dev467 (nightly), sm_80, tensor parallel, MTP on |
+
+Each architecture is one entry in [`adapters.py`](src/fp4local_abliterated/adapters.py): the vLLM module that defines it, the decoder layer class, and which submodules write into the residual stream. Everything else is shared. The plugin registers an `Abliterated<Arch>` subclass for every listed architecture through `ModelRegistry`, and attaches forward hooks after vLLM builds the model.
+
+To add an architecture, list its residual writers (embedding, attention output, MLP or MoE output, and anything else added to the residual) as they are named in vLLM's model code, then check that the startup log counts one write per writer per layer.
+
+## Where the direction for FP4-Local came from
+
+The direction used for [lesj0610/Qwen3.8-Flash-Next-FP4-Local-Abliterated](https://huggingface.co/lesj0610/Qwen3.8-Flash-Next-FP4-Local-Abliterated) was not captured by us. It is the direction of [windowsxp811203/Qwen3.8-Flash-Next-Abliterated](https://huggingface.co/windowsxp811203/Qwen3.8-Flash-Next-Abliterated), a bf16 weight-edited build of the same base model, recovered exactly from the difference between its edited and original weights (the difference is rank-1). Their model card describes how it was captured: difference of mean residual activations on harmful and harmless prompts, after 37 of the 48 decoder layers. The method is from Arditi et al., [Refusal in Language Models Is Mediated by a Single Direction](https://arxiv.org/abs/2406.11717) (2024).
+
+## Tests
+
+```bash
+pytest tests/
+```
+
+The CSF and CLI tests run on CPU. The projection tests need a CUDA GPU.
+
+## License
+
+Apache-2.0. `csf_scales.py` is adapted from vLLM (Apache-2.0).

fp4local_abliterated-0.1.0/pyproject.toml

@@ -0,0 +1,23 @@
+[build-system]
+requires = ["setuptools>=64"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "fp4local-abliterated"
+version = "0.1.0"
+description = "vLLM plugin for inference-time abliteration, with FP4-Local checkpoint loading"
+readme = "README.md"
+license = "Apache-2.0"
+requires-python = ">=3.10"
+# 0.32.0 has no NVFP4 PLE tables (vllm-project/vllm#56273) and allocates
+# FP4-Local's table as bf16, four times its size.
+dependencies = ["vllm>=0.32.1rc0"]
+
+[project.scripts]
+fp4local-abliterate = "fp4local_abliterated.cli:main"
+
+[project.entry-points."vllm.general_plugins"]
+fp4local_abliterated = "fp4local_abliterated:register"
+
+[tool.setuptools.packages.find]
+where = ["src"]

fp4local_abliterated-0.1.0/setup.cfg

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

fp4local_abliterated-0.1.0/src/fp4local_abliterated/__init__.py

@@ -0,0 +1,25 @@
+# SPDX-License-Identifier: Apache-2.0
+"""vLLM plugin for inference-time abliteration.
+
+- Applies abliteration when config.json carries an ``abliteration_config``
+  block (see the ``fp4local-abliterate`` CLI). Without it the model runs
+  unchanged.
+- Loads FP4-Local's CSF-compressed NVFP4 block scales and subdirectory shards,
+  until vLLM does this itself.
+"""
+
+from .adapters import ADAPTERS
+
+
+def register() -> None:
+    from vllm import ModelRegistry
+
+    from .loader import patch_default_loader
+    from .models import PREFIX
+
+    patch_default_loader()
+    for adapter in ADAPTERS:
+        for arch in adapter.architectures:
+            ModelRegistry.register_model(
+                arch, f"fp4local_abliterated.models:{PREFIX}{arch}"
+            )

fp4local_abliterated-0.1.0/src/fp4local_abliterated/ablation.py

@@ -0,0 +1,124 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Inference-time abliteration.
+
+Every write into the residual stream has its component along the refusal
+direction ``v`` scaled by ``1 - lam``: ``x - lam * (v . x) v``. This equals
+editing those weights to ``W - lam * v v^T W`` but leaves the quantized weights
+untouched. The writes are module outputs, so forward hooks apply it without
+touching vLLM's model code; adapters.py says which modules they are.
+
+The checkpoint enables it with a top-level ``abliteration_config`` block in
+config.json (``direction``: hidden_size floats, ``lambda``), written by the
+``fp4local-abliterate`` CLI.
+"""
+
+import torch
+from torch import nn
+
+from vllm.logger import init_logger
+from vllm.triton_utils import tl, triton
+
+from .adapters import Adapter
+
+logger = init_logger(f"vllm.{__name__}")
+
+
+@triton.jit
+def _ablate_kernel(
+    x_ptr,
+    out_ptr,
+    direction_ptr,
+    x_stride,
+    out_stride,
+    lam,
+    N: tl.constexpr,
+    BLOCK: tl.constexpr,
+):
+    row = tl.program_id(0).to(tl.int64)
+    offs = tl.arange(0, BLOCK)
+    mask = offs < N
+    x = tl.load(x_ptr + row * x_stride + offs, mask=mask, other=0.0).to(tl.float32)
+    direction = tl.load(direction_ptr + offs, mask=mask, other=0.0)
+    coef = tl.sum(x * direction) * lam
+    out = x - coef * direction
+    tl.store(
+        out_ptr + row * out_stride + offs, out.to(out_ptr.dtype.element_ty), mask=mask
+    )
+
+
+def ablate(
+    x: torch.Tensor, direction: torch.Tensor, lam: float, inplace: bool = False
+) -> torch.Tensor:
+    """Project ``x`` along its last dim in fp32; rows may be strided views."""
+    if x.shape[-1] != direction.numel():
+        raise ValueError(
+            f"Abliteration needs full hidden rows ({direction.numel()}), got "
+            f"{tuple(x.shape)}; tensor-parallel shards are not supported here"
+        )
+    rows = x.reshape(-1, x.shape[-1])
+    if rows.stride(-1) != 1:
+        if inplace:
+            raise ValueError("In-place abliteration needs unit-stride rows")
+        rows = rows.contiguous()
+    out = rows if inplace else torch.empty(rows.shape, dtype=x.dtype, device=x.device)
+    _ablate_kernel[(rows.shape[0],)](
+        rows,
+        out,
+        direction,
+        rows.stride(0),
+        out.stride(0),
+        lam,
+        N=rows.shape[1],
+        BLOCK=triton.next_power_of_2(rows.shape[1]),
+    )
+    return x if inplace else out.view(x.shape)
+
+
+def _hook(owner: nn.Module, lam: float, packed: bool):
+    def hook(module, args, output):
+        out = output[0] if isinstance(output, tuple) else output
+        direction = owner.ablit_direction
+        if packed:
+            ablate(out[..., -direction.numel() :], direction, lam, inplace=True)
+            return output
+        out = ablate(out, direction, lam)
+        return (out, *output[1:]) if isinstance(output, tuple) else out
+
+    return hook
+
+
+def attach_abliteration(model: nn.Module, hf_config, adapter: Adapter) -> None:
+    config = getattr(hf_config, "abliteration_config", None)
+    if config is None:
+        return
+    lam = float(config["lambda"])
+    direction = torch.tensor(config["direction"], dtype=torch.float32)
+    model.register_buffer(
+        "ablit_direction", direction / direction.norm(), persistent=False
+    )
+
+    sites = 0
+    layers = 0
+    for module in model.modules():
+        kind = type(module).__name__
+        layers += kind == adapter.layer_class
+        for name in adapter.writers.get(kind, ()):
+            target = getattr(module, name, None)
+            if target is None or type(target).__name__ == "PPMissingLayer":
+                continue
+            packed = (kind, name) in adapter.packed_writers
+            target.register_forward_hook(_hook(model, lam, packed))
+            sites += 1
+
+    if layers == 0:
+        raise RuntimeError(
+            f"{type(model).__name__} has an abliteration_config but no "
+            f"{adapter.layer_class} was found; this vLLM version is not supported"
+        )
+    logger.info(
+        "Abliteration on %s: lambda %.4g, %d residual writes in %d layers",
+        type(model).__name__,
+        lam,
+        sites,
+        layers,
+    )

fp4local_abliterated-0.1.0/src/fp4local_abliterated/adapters.py

@@ -0,0 +1,47 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Per-architecture description of the residual-stream writers.
+
+Everything else (config parsing, the projection, hooking, registration) is
+shared, so supporting another vLLM architecture means adding one entry here.
+This module imports nothing from vLLM so the CLI can use it without a GPU.
+"""
+
+from dataclasses import dataclass, field
+
+
+@dataclass(frozen=True)
+class Adapter:
+    # vLLM module that defines the architectures below.
+    module: str
+    architectures: tuple[str, ...]
+    # Decoder layer class; finding none means vLLM's model code has moved on.
+    layer_class: str
+    # Owner class name -> submodules whose output is added to the residual.
+    writers: dict[str, tuple[str, ...]]
+    # (owner, submodule) pairs whose output only ends with the residual write.
+    packed_writers: frozenset[tuple[str, str]] = field(default_factory=frozenset)
+
+
+QWEN4EXP = Adapter(
+    module="vllm.models.qwen4_exp",
+    architectures=(
+        "Qwen4ExpForCausalLM",
+        "Qwen4ExpForConditionalGeneration",
+        "Qwen4ExpMTP",
+    ),
+    layer_class="Qwen4ExpDecoderLayer",
+    writers={
+        "Qwen4ExpDecoderLayer": ("linear_attn", "self_attn", "mlp"),
+        "Qwen4ExpModel": ("embed_tokens",),
+        "Qwen4ExpMultiTokenPredictor": ("embed_tokens", "fc_embedding", "fc_hidden"),
+        "Qwen4ExpPLELayer": ("kv_proj",),
+    },
+    # kv_proj packs [key | value]; only the value reaches the residual.
+    packed_writers=frozenset({("Qwen4ExpPLELayer", "kv_proj")}),
+)
+
+ADAPTERS = (QWEN4EXP,)
+
+
+def find_adapter(architecture: str) -> Adapter | None:
+    return next((a for a in ADAPTERS if architecture in a.architectures), None)