Skip to content

TurboQuant KV cache (4/4): Python reference impl + last_token_logits patcher#28563

Draft
TimPietrusky wants to merge 2 commits into
microsoft:mainfrom
TimPietrusky:tim/turboquant/04-python-tooling
Draft

TurboQuant KV cache (4/4): Python reference impl + last_token_logits patcher#28563
TimPietrusky wants to merge 2 commits into
microsoft:mainfrom
TimPietrusky:tim/turboquant/04-python-tooling

Conversation

@TimPietrusky
Copy link
Copy Markdown

Description

Reference NumPy implementation, offline graph rewriter, paper-validation tests, and a one-time model patcher that unlocks long-context inference on stock HuggingFace q4f16 ONNX exports. Standalone — depends on the schema in #28560 but not on any kernel PR.

What this PR contains

Under onnxruntime/python/tools/quantization/turboquant_kv/:

  • centroids.py — Lloyd-Max solver for the K codebook. Computes the optimal scalar quantiser for N(0, 1/d) (the distribution of components of (k / ||k||) @ H_norm where k is fp16 and H_norm is the normalised Walsh-Hadamard). Deterministic given (d, bits); identical to what the C++ graph transformer in TurboQuant KV cache (1/4): graph rewrite + schema (foundation) #28560 injects.
  • hadamard.py — Sylvester-construction Walsh-Hadamard, scaled to H @ H^T = I. Same scaling as the kernels.
  • packing.py — uint8 / uint4 / uint3 bit-pack and unpack. Bit layouts match the C++ kernels in TurboQuant KV cache (2/4): CUDA kernels #28561 (CUDA) and TurboQuant KV cache (3/4): WebGPU kernels + Safari/Firefox fallback #28562 (WebGPU).
  • quantizer.pyencode_keys / decode_keys / encode_values / decode_values. Pure NumPy reference; used by both the offline rewriter and the validation harness.
  • onnx_rewriter.py — Python equivalent of the C++ graph transformer in TurboQuant KV cache (1/4): graph rewrite + schema (foundation) #28560. Useful when users want to ship a pre-rewritten .onnx instead of relying on session-create-time rewriting (e.g. so a model registry can stamp a hash).
  • validate.py — paper-replication tests. 23 / 23 pass against the TurboQuant paper's published numbers. Tests are cross-validated bit-exact against vLLM's reference implementation where overlap exists.
  • benchmark.py — standalone perf bench. Used to generate the numbers in TurboQuant KV cache (2/4): CUDA kernels #28561 and TurboQuant KV cache (3/4): WebGPU kernels + Safari/Firefox fallback #28562.
  • last_token_logits.py — standalone model patcher. HuggingFace causal-LM ONNX exports compute logits for every prompt position by default. At long contexts (S × vocab > 2³¹) this trips an int32 overflow in ORT's CUDA Cast kernel — see #28385. This patcher inserts a Slice op before the LM-head MatMul so logits are computed only for the last position (the standard logits_to_keep=1 pattern in HF transformers). One-time, ~30s, idempotent.

Why ship this even though #28560 makes online rewriting work

Even with the C++ graph transformer landed:

Depends on

Does not intersect with

Adds a `TurboQuantKVFusion` graph transformer that rewrites every
GroupQueryAttention node at session-create time to use a TurboQuant
4-bit packed KV cache, plus the schema, session-option keys, and CPU
helpers required for that rewrite.  No kernels in this PR — they
land in follow-ups for CUDA and WebGPU.

What this PR includes:

* `core/optimizer/turboquant_kv_fusion.{cc,h}` — the L2 transformer.
  Enabled by setting `optimization.turboquant_kv_method` to one of
  `turboquant_4bit_nc`, `turboquant_k3v4_nc`, `turboquant_3bit_nc`.
  Runs on CUDA + WebGPU EPs.  Computes Lloyd-Max centroids for the
  given (head_dim, key_bits) and a normalised Walsh–Hadamard matrix,
  injects both as graph initializers, and mutates each GQA node's
  attributes + past/present tensor types to (uint8, slot_bytes).

* `core/graph/contrib_ops/bert_defs.cc` — extends GroupQueryAttention
  with the new attributes (`kv_quant_method`, `key_quant_bits`,
  `value_quant_bits`, `norm_correction`) and two new optional inputs
  at slots 14 / 15 for the shared k_codebook + hadamard initializers.

* `include/onnxruntime/core/session/onnxruntime_session_options_config_keys.h`
  — public option keys `optimization.turboquant_kv_method` and
  `optimization.turboquant_kv_boundary`.

* `contrib_ops/cpu/bert/attention_common.h` + `attention_parameters.h`
  + `group_query_attention_helper.h` — `KVQuantMethod` enum, parameter
  struct extensions, and `CheckInputs` updates so the fp16 codepath
  passes through unchanged when TurboQuant isn't requested.

* `include/onnxruntime/core/framework/int3.h` — new packed `UInt3x8`
  type for 3-bit cache slots.  Used by the (forthcoming) 3-bit
  variants.

* `test/contrib_ops/turboquant_kv_test.cc` — host-side bit-layout
  tests for `UInt3x8`.  Kernel-level correctness is validated by the
  follow-up CUDA / WebGPU PRs.

When `optimization.turboquant_kv_method` is unset or set to "none" /
"off" the transformer doesn't fire and the graph is byte-identical
to today's output.

Design doc + reference NumPy implementation + paper-validation tests
are coming in the Python tooling PR.  The CUDA kernels (16-bit accum
WMMA + 4-bit packed cache) and the WebGPU kernels (WGSL encode/decode
with an ApplyAttention fallback for browsers without Subgroups) come
in separate PRs that each depend on this one.

Benches (LFM2.5-1.2B, RTX A40, all measured):

  ctx      fp16 decode    TQ decode      speedup
   4 K     6.2 s reply    6.0 s reply    tied
  32 K     26 s           24 s            7 %
  64 K     63 s           41 s           53 %
  128 K   (fp16 OOM)      65 s           TQ only
…patcher

Reference NumPy implementation, offline graph rewriter, paper-validation
tests, and a one-time model patcher that unlocks long-context inference
on stock HuggingFace q4f16 ONNX exports.

What this PR contains (all under
`onnxruntime/python/tools/quantization/turboquant_kv/`):

* `centroids.py` — Lloyd-Max solver for the K codebook.  Computes the
  optimal scalar quantiser for `N(0, 1/d)` (the distribution of
  components of `(k / ||k||) @ H_norm` where `k` is fp16 and `H_norm`
  is the normalised Walsh-Hadamard).  Deterministic given `(d, bits)`;
  identical to what the C++ graph transformer in microsoft#28560 injects.
* `hadamard.py` — Sylvester-construction Walsh-Hadamard, scaled to
  `H @ H^T = I`.  Same scaling as the kernels.
* `packing.py` — uint8 / uint4 / uint3 bit-pack and unpack.  Bit
  layouts match the C++ kernels in microsoft#28561 and microsoft#28562.
* `quantizer.py` — `encode_keys` / `decode_keys` / `encode_values` /
  `decode_values`.  Pure NumPy reference; used by both the offline
  rewriter and the validation harness.
* `onnx_rewriter.py` — Python equivalent of the C++ graph transformer
  in microsoft#28560.  Useful when users want to ship a pre-rewritten `.onnx`
  instead of relying on session-create-time rewriting (e.g. so a model
  registry can stamp a hash).
* `validate.py` — paper-replication tests.  23 / 23 pass against the
  TurboQuant paper's published numbers.  Tests are cross-validated
  bit-exact against vLLM's reference implementation where overlap
  exists.
* `benchmark.py` — standalone perf bench.  Used to generate the
  numbers in microsoft#28561 and microsoft#28562 (decode tok/s, KV cache bytes).
* `last_token_logits.py` — standalone model patcher.  HuggingFace
  causal-LM ONNX exports compute logits for *every* prompt position
  by default.  At long contexts (S × vocab > 2^31) this trips an
  int32 overflow in ORT's CUDA `Cast` kernel — see microsoft#28385.  This
  patcher inserts a `Slice` op before the LM-head MatMul so logits
  are computed only for the last position (the standard
  `logits_to_keep=1` pattern in HF transformers).  One-time, ~30s,
  idempotent.

Independent of the kernel PRs.  The C++ graph transformer in microsoft#28560
makes this Python tooling optional for online use, but the rewriter
+ validation tests are still useful for:

* Producing pre-rewritten models for environments where session
  options can't be set (`onnx_rewriter.py`).
* Reproducing the bit-exact bit layout test the kernels rely on
  (`packing.py` + the C++ tests in microsoft#28560).
* Validating new kernel changes against the published TurboQuant
  paper numbers without spinning up a full e2e benchmark
  (`validate.py`).
* Patching the long-context cliff on stock HF exports today,
  before microsoft#28385 is fixed upstream (`last_token_logits.py`).

Depends on microsoft#28560 only (for the schema the rewriter writes against).
Does not intersect with microsoft#28561 (CUDA) or microsoft#28562 (WebGPU) — those
read the schema; this one writes it.
@microsoft-github-policy-service
Copy link
Copy Markdown
Contributor

@TimPietrusky please read the following Contributor License Agreement(CLA). If you agree with the CLA, please reply with the following information.

@microsoft-github-policy-service agree [company="{your company}"]

Options:

  • (default - no company specified) I have sole ownership of intellectual property rights to my Submissions and I am not making Submissions in the course of work for my employer.
@microsoft-github-policy-service agree
  • (when company given) I am making Submissions in the course of work for my employer (or my employer has intellectual property rights in my Submissions by contract or applicable law). I have permission from my employer to make Submissions and enter into this Agreement on behalf of my employer. By signing below, the defined term “You” includes me and my employer.
@microsoft-github-policy-service agree company="Microsoft"
Contributor License Agreement

Contribution License Agreement

This Contribution License Agreement (“Agreement”) is agreed to by the party signing below (“You”),
and conveys certain license rights to Microsoft Corporation and its affiliates (“Microsoft”) for Your
contributions to Microsoft open source projects. This Agreement is effective as of the latest signature
date below.

  1. Definitions.
    “Code” means the computer software code, whether in human-readable or machine-executable form,
    that is delivered by You to Microsoft under this Agreement.
    “Project” means any of the projects owned or managed by Microsoft and offered under a license
    approved by the Open Source Initiative (www.opensource.org).
    “Submit” is the act of uploading, submitting, transmitting, or distributing code or other content to any
    Project, 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 Project for the purpose of
    discussing and improving that Project, but excluding communication that is conspicuously marked or
    otherwise designated in writing by You as “Not a Submission.”
    “Submission” means the Code and any other copyrightable material Submitted by You, including any
    associated comments and documentation.
  2. Your Submission. You must agree to the terms of this Agreement before making a Submission to any
    Project. This Agreement covers any and all Submissions that You, now or in the future (except as
    described in Section 4 below), Submit to any Project.
  3. Originality of Work. You represent that each of Your Submissions is entirely Your original work.
    Should You wish to Submit materials that are not Your original work, You may Submit them separately
    to the Project if You (a) retain all copyright and license information that was in the materials as You
    received them, (b) in the description accompanying Your Submission, include the phrase “Submission
    containing materials of a third party:” followed by the names of the third party and any licenses or other
    restrictions of which You are aware, and (c) follow any other instructions in the Project’s written
    guidelines concerning Submissions.
  4. Your Employer. References to “employer” in this Agreement include Your employer or anyone else
    for whom You are acting in making Your Submission, e.g. as a contractor, vendor, or agent. If Your
    Submission is made in the course of Your work for an employer or Your employer has intellectual
    property rights in Your Submission by contract or applicable law, You must secure permission from Your
    employer to make the Submission before signing this Agreement. In that case, the term “You” in this
    Agreement will refer to You and the employer collectively. If You change employers in the future and
    desire to Submit additional Submissions for the new employer, then You agree to sign a new Agreement
    and secure permission from the new employer before Submitting those Submissions.
  5. Licenses.
  • Copyright License. You grant Microsoft, and those who receive the Submission directly or
    indirectly from Microsoft, a perpetual, worldwide, non-exclusive, royalty-free, irrevocable license in the
    Submission to reproduce, prepare derivative works of, publicly display, publicly perform, and distribute
    the Submission and such derivative works, and to sublicense any or all of the foregoing rights to third
    parties.
  • Patent License. You grant Microsoft, and those who receive the Submission directly or
    indirectly from Microsoft, a perpetual, worldwide, non-exclusive, royalty-free, irrevocable license under
    Your patent claims that are necessarily infringed by the Submission or the combination of the
    Submission with the Project to which it was Submitted to make, have made, use, offer to sell, sell and
    import or otherwise dispose of the Submission alone or with the Project.
  • Other Rights Reserved. Each party reserves all rights not expressly granted in this Agreement.
    No additional licenses or rights whatsoever (including, without limitation, any implied licenses) are
    granted by implication, exhaustion, estoppel or otherwise.
  1. Representations and Warranties. You represent that You are legally entitled to grant the above
    licenses. You represent that each of Your Submissions is entirely Your original work (except as You may
    have disclosed under Section 3). You represent that You have secured permission from Your employer to
    make the Submission in cases where Your Submission is made in the course of Your work for Your
    employer or Your employer has intellectual property rights in Your Submission by contract or applicable
    law. If You are signing this Agreement on behalf of Your employer, You represent and warrant that You
    have the necessary authority to bind the listed employer to the obligations contained in this Agreement.
    You are not expected to provide support for Your Submission, unless You choose to do so. UNLESS
    REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING, AND EXCEPT FOR THE WARRANTIES
    EXPRESSLY STATED IN SECTIONS 3, 4, AND 6, THE SUBMISSION PROVIDED UNDER THIS AGREEMENT IS
    PROVIDED WITHOUT WARRANTY OF ANY KIND, INCLUDING, BUT NOT LIMITED TO, ANY WARRANTY OF
    NONINFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
  2. Notice to Microsoft. You agree to notify Microsoft in writing of any facts or circumstances of which
    You later become aware that would make Your representations in this Agreement inaccurate in any
    respect.
  3. Information about Submissions. You agree that contributions to Projects and information about
    contributions may be maintained indefinitely and disclosed publicly, including Your name and other
    information that You submit with Your Submission.
  4. Governing Law/Jurisdiction. This Agreement is governed by the laws of the State of Washington, and
    the parties consent to exclusive jurisdiction and venue in the federal courts sitting in King County,
    Washington, unless no federal subject matter jurisdiction exists, in which case the parties consent to
    exclusive jurisdiction and venue in the Superior Court of King County, Washington. The parties waive all
    defenses of lack of personal jurisdiction and forum non-conveniens.
  5. Entire Agreement/Assignment. This Agreement is the entire agreement between the parties, and
    supersedes any and all prior agreements, understandings or communications, written or oral, between
    the parties relating to the subject matter hereof. This Agreement may be assigned by Microsoft.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants