MUVERA: compute the encoding on the GPU, not on the inference thread's CPU
メンテナーはふだん 1 日以内に返信
まだ誰も着手していません。
評価
- 難易度
- 4/5
- 見積もり時間
- 3〜5日
- 初心者へのやさしさ
- 25/100
- issue の種類
- 機能追加
- 明瞭さ
- 明確に書かれている
- 活発さ
- 静か
調査の方向性
Read packages/sie_server/src/sie_server/core/postprocessor.py: MuveraPostprocessor.transform and _build_random_cache, then trace EncodeHandler.run_inference through PostprocessorRegistry.transform_sync and an adapter such as colqwen2. First reproduce the numpy output on CPU for each MUVERA variant and compare it with the proposed batched torch formulation. Done means the device path matches numpy within float32 rounding for documents and queries, including empty and one-token items, while the numpy fallback remains.
索引モデルが issue の本文から書いたものです。
説明
Problem
The 16 model configs with a muvera profile (the ColBERT, ColPali, ColQwen, ColSmol and NVIDIA colembed models, and TopK-Embed-V1 from #417) compute MUVERA on the CPU, after the forward pass, on the model's inference thread:
- The adapter copies each item's token vectors from the GPU to the host as float32 numpy arrays, for example in colqwen2.
ModelWorkerruns inference on a single thread, andEncodeHandler.run_inferenceapplies postprocessors on that same thread throughPostprocessorRegistry.transform_sync.MuveraPostprocessor.transformworks in numpy, one item at a time and, per item, one repetition at a time: a SimHash projection, a scatter-add of every token into its partition (np.add.at), then a scatter-add into the Count Sketch.
This suited the 128-number models it was written for (the module docstring says "for ColBERT/ColPali"). The work grows with tokens × token width × repetitions, and the models added since are wide: NVIDIA's colembed models use 2,560 and 3,072 numbers per token, TopK-Embed-V1 uses 1,024 and 2,048. On TopK-Embed-V1's 2B model the CPU step takes longer per page than the forward pass itself. Because it runs on the inference thread:
- the GPU idles while MUVERA runs;
- every other request to the model, default-profile requests included, waits behind it;
- the adapter copies the full float32 token matrix to the host only to reduce it to one fixed-size vector.
Nothing in the algorithm needs the CPU. Every step is a matrix multiply or a sum into fixed slots, and the token vectors are already on the GPU when the forward pass ends.
Proposal
-
Batched formulation. Compute every item and every repetition of a batch at once, in torch, from the packed token matrix
X [N, d]and per-item offsets (item[t]is tokent's item):# Once per model, on its device: the structures _build_random_cache() makes today, # from the same seeds and numpy RNG calls, so encodings stay compatible with existing indexes. H = cat(simhash_matrices, dim=1) # [d, R*k], R = 40 repetitions, k = 6 bits, P = 2**k partitions A = cat(ams_projection_matrices, dim=1) # [d, R*p], only when projection_dim is set (text ColBERT: p = 8) cs_idx, cs_sign # [R*P*d'], only when final_projection_dim is set (page models: 10240) # Per batch. if center_tokens: X = X - per_item_mean(X)[item] bits = (X @ H > 0).view(N, R, k) part = gray_code(bits) # [N, R], same recurrence as _sketches_to_gray_partitions seg = (item[:, None] * R + arange(R)) * P + part # one segment per (item, repetition, partition) V = X if identity else (X @ A).view(N, R, p) # d' = d or p sums = zeros(B * R * P, d_prime) for r in range(R): # avoids an [N, R, d] copy of the tokens sums.index_add_(0, seg[:, r], V if identity else V[:, r]) if not is_query: # documents average, queries sum sums /= bincount(seg.flatten(), minlength=B * R * P).clamp_min(1)[:, None] fde = sums.view(B, -1) if final_dim: fde = zeros(B, final_dim).index_add_(1, cs_idx, fde * cs_sign) if normalize: fde = fde / fde.norm(dim=1, keepdim=True)A CPU check of this formulation against
MuveraPostprocessor(identity projection with the 10,240 Count Sketch, 128 to 3,072 numbers per token, documents and queries) gave bit-identical outputs. Theprojection_dim: 8andcenter_tokensvariants are not checked yet. On the GPU, float32 sums in a different order should differ only by rounding. -
Run it before the host copy. Give
MuveraPostprocessora device method, for exampleencode_device(tokens, offsets, *, is_query) -> Tensor [B, D]. The multi-vector adapters call it on their device output when the request's options carrymuvera, returndensedirectly, and skip the float32 multivector host copy. The registry skips the numpy transform when the adapter has already filleddense. Adapters:colbert,colbert_modernbert_flash,colbert_rotary_flash,colpali,colqwen2,colqwen3,colsmol,nemo_colembedandtopk_embed, the wide ones (nemo_colembed,topk_embed) first. -
Keep the numpy path as the fallback for CPU-only deployments and adapters that have not opted in. Separately, that path could leave the inference thread: the registry's async
transform()already runs postprocessors on a CPU pool, but the server callstransform_syncon the inference thread. -
Bound the memory. The per-partition sums are
[B·R·P, d']float32, about 170 MB for a batch of eight 2,048-number pages. Process the repetitions in chunks if that competes with the model's activations.
Acceptance
- For each MUVERA variant SIE ships (identity projection with the 10,240 Count Sketch;
projection_dim: 8without a sketch; with and withoutcenter_tokens), for documents and queries, including empty and one-token items: the device path matches the numpy path within float32 rounding, and an item's encoding does not depend on the rest of its batch. - With the
muveraprofile on a wide model, the GPU no longer waits on the CPU between batches, andmuverarequests no longer hold up default-profile requests to the same model. - CPU-only deployments keep working through the numpy path.
- Performance is re-evaluated after the change: the
muveraprofile's throughput and latency on a GPU, before and after.
Out of scope: whether 10,240 numbers are enough for tokens of 1,024 to 3,072 numbers (search quality), and the paper's step that fills empty partitions for documents. Each needs its own measurement.
- 主要言語
- Python
- スター
- 3.5k
- フォーク
- 328
- 平均マージ
- 16時間 25分
- マージ済み PR(30日)
- 121
環境構築
このプロジェクトには開発コンテナ、Dockerfile、コントリビューションガイドがありません。まず README を読み、一般的な手順ははじめてのコントリビューションガイドを参照してください。
はじめの一歩
- issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
- 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
- リポジトリをフォークし、ブランチを切って変更します。
- issue 番号を参照したプルリクエストを送ります。
superlinked/sie のほかの issue
-
MODEL_NOT_FOUND gives no suggestion for a model listed only as model:profile対応中かも @ShaiShmuel が今日担当しました。 オープンenhancement rust
難易度 2/5 1〜3時間 初心者へのやさしさ 65/100
superlinked/sie#985 ·
メンテナーはふだん 1 日以内に返信
-
難易度 4/5 3〜5日 初心者へのやさしさ 42/100
superlinked/sie#969 ·
メンテナーはふだん 1 日以内に返信
-
難易度 5/5 1週間以上 初心者へのやさしさ 25/100
superlinked/sie#903 · コメント 4 件 ·
メンテナーはふだん 1 日以内に返信
-
難易度 5/5 1週間以上 初心者へのやさしさ 25/100
superlinked/sie#892 ·
メンテナーはふだん 1 日以内に返信
-
難易度 4/5 3〜5日 初心者へのやさしさ 25/100
superlinked/sie#851 · コメント 4 件 ·
メンテナーはふだん 1 日以内に返信
superlinked/sie の issue をすべて見る
似ている issue
-
難易度 1/5 1〜3時間 初心者へのやさしさ 76/100
samuelclay/NewsBlur#2214 ·
メンテナーはふだん 1 日以内に返信
-
難易度 2/5 1〜3時間 初心者へのやさしさ 68/100
メンテナーはふだん 1 日以内に返信
-
難易度 2/5 1〜3時間 初心者へのやさしさ 88/100
メンテナーはふだん 1 日以内に返信
-
難易度 2/5 1〜3時間 初心者へのやさしさ 68/100
readthedocs/readthedocs-builder#73 ·
メンテナーはふだん 1 日以内に返信
-
難易度 2/5 1〜3時間 初心者へのやさしさ 82/100
reflex-dev/reflex#7915 · コメント 1 件 ·
メンテナーはふだん 1 日以内に返信