MUVERA: compute the encoding on the GPU, not on the inference thread's CPU

Ouverte
#443 0 commentaires 0 réactions 0 personnes assignées Voir sur GitHub

Les mainteneurs répondent en général sous 1 jour

Personne n'a encore pris cette issue.

Évaluation

Difficulté
4/5
Temps estimé
3-5 jours
Accessibilité débutants
25/100
Type d'issue
Fonctionnalité
Clarté
Clairement spécifiée
Activité
Calme
Stack technique
python, pytorch

Piste de recherche

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.

Rédigé par le modèle d'indexation à partir du texte de l'issue.

Description

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:

  1. The adapter copies each item's token vectors from the GPU to the host as float32 numpy arrays, for example in colqwen2.
  2. ModelWorker runs inference on a single thread, and EncodeHandler.run_inference applies postprocessors on that same thread through PostprocessorRegistry.transform_sync.
  3. MuveraPostprocessor.transform works 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

  1. 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 token t'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. The projection_dim: 8 and center_tokens variants are not checked yet. On the GPU, float32 sums in a different order should differ only by rounding.

  2. Run it before the host copy. Give MuveraPostprocessor a device method, for example encode_device(tokens, offsets, *, is_query) -> Tensor [B, D]. The multi-vector adapters call it on their device output when the request's options carry muvera, return dense directly, and skip the float32 multivector host copy. The registry skips the numpy transform when the adapter has already filled dense. Adapters: colbert, colbert_modernbert_flash, colbert_rotary_flash, colpali, colqwen2, colqwen3, colsmol, nemo_colembed and topk_embed, the wide ones (nemo_colembed, topk_embed) first.

  3. 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 calls transform_sync on the inference thread.

  4. 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: 8 without a sketch; with and without center_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 muvera profile on a wide model, the GPU no longer waits on the CPU between batches, and muvera requests 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 muvera profile'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.

Langage dominant
Python
Étoiles
3.5k
Forks
328
Merge moyen
16 h 25 min
PR mergées (30 j)
121

Préparer son environnement

Ce projet ne fournit ni conteneur de développement, ni Dockerfile, ni guide de contribution : l'installation est à votre charge. Commencez par son README, et consultez notre guide de la première contribution pour les étapes générales.

Par où commencer

  1. Lisez l'issue en entier, puis le guide de contribution du projet.
  2. Signalez en commentaire que vous la prenez — cela évite que deux personnes fassent le même travail.
  3. Forkez le dépôt et travaillez sur une branche.
  4. Ouvrez une pull request qui référence le numéro de l'issue.

Autres issues de superlinked/sie

Toutes les issues de superlinked/sie

Issues similaires

Plus d'issues Python

Recevez les nouvelles issues par e-mail

Un résumé court des issues GitHub adaptées aux débutants.