Skip to content

TurboQuant KV cache (2/4): CUDA kernels#28561

Draft
TimPietrusky wants to merge 2 commits into
microsoft:mainfrom
TimPietrusky:tim/turboquant/02-cuda-kernels
Draft

TurboQuant KV cache (2/4): CUDA kernels#28561
TimPietrusky wants to merge 2 commits into
microsoft:mainfrom
TimPietrusky:tim/turboquant/02-cuda-kernels

Conversation

@TimPietrusky
Copy link
Copy Markdown

Description

CUDA kernels for the TurboQuant 4-bit KV cache path defined by #28560. When the graph transformer in #28560 has rewritten a GroupQueryAttention node's KV-cache schema to the packed uint8 layout, the CUDA GQA kernel now dispatches into the launchers added here.

Real-model end-to-end speedups (RTX A40)

Model ctx fp16 reply TQ reply speedup
LFM2.5-1.2B 4 K 6.2 s 6.0 s tied
LFM2.5-1.2B 32 K 26.0 s 24.1 s 7%
LFM2.5-1.2B 64 K 63.0 s 41.1 s 53%
LFM2.5-1.2B 128 K fp16 OOM 65 s TQ-only
Qwen3-0.6B 4 K 26.6 s 17.6 s 51%
Qwen3-0.6B 16 K 93 s 58 s 60%
Qwen3-0.6B 32 K 187 s 96 s 94%

KV cache 3.56× smaller (LFM2.5, head_dim=64) / 3.76× (Qwen3-0.6B, head_dim=128). Per-layer cosine sim vs fp16 = 0.99526 across-context. Top-1 token match 7-of-9 on a 9-step decode chain.

What this PR contains

  • contrib_ops/cuda/bert/group_query_attention_turboquant.cuh — templated kernels:
    • TQEncodeKernel packs fresh K/V (Walsh-Hadamard + Lloyd-Max → 4-bit indices).
    • TQFlashAttentionKernel (v4-lite) fused-FA on packed K with online softmax for the decode step.
    • TQFlashAttentionQTiledKernel (v5) for continuation-prefill — ~2× prompt speedup over v4-lite at 4K.
    • TQFlashAttentionWmmaKernel (v6) tensor-core Q*K^T at head_dim=64.
  • contrib_ops/cuda/bert/group_query_attention_turboquant_impl.{cu,h} — launch shims, vectorised uint4 K/V smem loads (8 fp16 per HBM transaction), Option-ε dispatch that routes the prompt step to the stock FlashAttention kernel when past_seq == 0 so model-load output is byte-stable vs fp16.
  • contrib_ops/cuda/bert/group_query_attention.cc/.h — TQ-aware dispatch. When kv_quant_method != KVQuantMethod::None we read the new attributes from TurboQuant KV cache (1/4): graph rewrite + schema (foundation) #28560 and route to the kernels above; otherwise the op is byte-identical to today.
  • contrib_ops/cuda/bert/attention_data.h — three new fields in GroupQueryAttentionData for the codebook / Hadamard pointers and the quant-method enum.

Things I'd appreciate a sharp eye on

  1. WMMA tile schedule in v6 — kBlockK is coupled to kHeadDim today. Decoupling unlocks hd=128 (Qwen3-0.6B uses this) at v6 speed. Follow-up patch ready, kept out to minimise diff size.
  2. cp.async — tried double-buffered loads, reverted because per-tile commit_group/wait_group overhead exceeded overlap savings at our tile size (74 ms → 95 ms decode at 32K). Worth revisiting if tile sizing changes.
  3. Option ε prompt delegation in group_query_attention.cc — by design, prompt step (past_sequence_length == 0) routes to the stock FA kernel for bit-equivalent output. Cost: v6 wmma prefill never fires on the first call. Step 2+ still uses it.

Validation

Host-side bit-layout tests are in #28560. CUDA kernel correctness is validated via real-model e2e runs because onnxruntime_provider_test doesn't link libonnxruntime_providers_cuda.so directly — the EP is loaded via dlopen at test time and our launcher symbols aren't exposed there today. Per-layer cosine-sim numbers above come from a Python harness running both sessions on the same prompt and comparing logits. That harness ships in the Python tooling PR.

Depends on

Does not intersect with

  • WebGPU kernel PR (#TBD).
  • Python tooling PR (#TBD).

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
Adds the CUDA TurboQuant attention kernels referenced by the graph
rewriter in microsoft#28560.  All speedups numbers are end-to-end against
fp16 baseline on RTX A40, real models pulled from HuggingFace:

  Model               ctx     fp16 reply   TQ reply   speedup
  LFM2.5-1.2B          4 K    6.2 s        6.0 s      tied
  LFM2.5-1.2B         32 K   26.0 s       24.1 s      7 %
  LFM2.5-1.2B         64 K   63.0 s       41.1 s      53 %
  LFM2.5-1.2B        128 K   (fp16 OOM)   65 s        TQ-only
  Qwen3-0.6B           4 K   26.6 s       17.6 s     51 %
  Qwen3-0.6B          16 K     93 s         58 s     60 %
  Qwen3-0.6B          32 K    187 s         96 s     94 %

KV cache 3.56× smaller (LFM2.5, hd=64) / 3.76× (Qwen3-0.6B, hd=128).
Per-layer cosine sim vs fp16: 0.99526 across-context, top-1 token
match 7/9 on a 9-step decode chain (matches CUDA TurboQuant kernels
in vLLM bit-exact in the encode path; the decode kernels have minor
differences in tile scheduling).

What this PR contains:

* `contrib_ops/cuda/bert/group_query_attention_turboquant.cuh` —
  templated CUDA kernels:
    - `TQEncodeKernel` for fresh-K/V pack (Walsh-Hadamard + Lloyd-Max).
    - `TQFlashAttentionKernel` (v4-lite) for decode-step fused FA on
      packed K with online softmax.
    - `TQFlashAttentionQTiledKernel` (v5) for continuation-prefill,
      ~2× prompt speedup over v4-lite at 4K context.
    - `TQFlashAttentionWmmaKernel` (v6) for tensor-core Q*K^T scoring
      at hd=64.  Falls through to v5 for other head_dims.
* `contrib_ops/cuda/bert/group_query_attention_turboquant_impl.{cu,h}`
  — launch shims, vectorised uint4 K/V smem loads (8-fp16 per HBM
  transaction), Option-ε dispatch that routes the prompt step to the
  stock FlashAttention kernel when past_seq=0 (bit-equivalent to fp16
  for the first call so model load is byte-stable).
* `contrib_ops/cuda/bert/group_query_attention.cc/.h` — TQ-aware
  dispatch in the existing GroupQueryAttention CUDA op.  When
  `kv_quant_method != KVQuantMethod::None` we read the new attributes
  from microsoft#28560 and route to the kernels above.  When it's None the op
  is byte-identical to today.
* `contrib_ops/cuda/bert/attention_data.h` — three new fields in
  `GroupQueryAttentionData` for the codebook / Hadamard pointers and
  quant-method enum.  Unused on the fp16 path.

Things I'd appreciate a sharp eye on:

1. **WMMA tile schedule** in v6 — `kBlockK` is currently coupled to
   `kHeadDim`.  Decoupling unlocks hd=128 (Qwen3-0.6B) at the same
   speed as hd=64 (LFM2.5).  I have a follow-up patch ready but kept
   it out of this PR to minimise diff size.
2. **cp.async** — I tried double-buffered K/V loads in v6.  Reverted
   because the per-tile commit_group/wait_group overhead exceeded the
   overlap savings at our tile size (74 ms → 95 ms decode at 32K).
   Worth revisiting if a future commit changes the tile sizing.
3. **Option ε prompt-step delegation** in `group_query_attention.cc`
   — the dispatcher checks `past_sequence_length == 0` and routes to
   the standard FA kernel.  This is correctness-first: prompt step
   output is bit-equivalent to fp16, which keeps top-1 token agreement
   with non-TQ runs stable across cold-load tests.  Cost is the wmma
   v6 prefill kernel never fires on step 1, but step 2+ still uses
   it because past_seq > 0 by then.

Depends on microsoft#28560 (foundation: graph rewrite + schema).
The WebGPU kernel PR and the Python tooling PR each depend on
microsoft#28560 too, and don't intersect with this file set.

### Validation

The host-side bit-layout tests in microsoft#28560 are necessary but not
sufficient for the CUDA kernels.  Kernel correctness is validated via
real-model e2e runs because `onnxruntime_provider_test` does not
link `libonnxruntime_providers_cuda.so` directly — the CUDA EP is
loaded via `dlopen` at test time, and our kernel launcher symbols
aren't exposed through that interface today.  Per-layer cosine sim
vs fp16 was measured by a Python harness that runs both sessions on
the same prompt and compares logits; that harness sits under
`tools/quantization/turboquant_kv/validate.py` in the Python tooling
PR.
@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