A tokenizer library for sequential recommendation systems. RecTokens converts continuous item embeddings into discrete multi-level token sequences using Residual Quantization (RQ), enabling large item catalogs to be represented as compact token IDs suitable for Language model finetuning and/or classic Transformer-based recommendation models.
Modern sequential recommendation models treat item retrieval as a language modeling problem: given a user's interaction history, generate the next item's token sequence autoregressively. RecTokens provides three components for this pipeline:
- Tokenizers — Convert item feature vectors into discrete token sequences (item IDs).
- Constrained Decoding — At inference time, efficiently restrict the model's generation to only valid item token sequences.
- HuggingFace Integration — Plug item tokens directly into any pretrained LLM from HuggingFace.
ItemAwareTokenizerextends an HF text tokenizer with item-level tokens,ItemAwareCausalLM.from_causal_lmadapts the model's embedding table, andInterleavedSequenceCollator/PrecomputedSequenceCollatorprepare mixed text+item batches for standard HFTrainerfine-tuning. This lets you fine-tune models like Qwen on recommendation sequences with minimal boilerplate.
RecTokens builds on several lines of research in generative recommendation. The idea of treating item retrieval as a language modeling problem over discrete semantic IDs originates from Rajput et al., "Recommender Systems with Generative Retrieval" (NeurIPS 2023), which introduced the TIGER system and demonstrated that items can be represented as compact hierarchical token sequences produced by Residual Quantization. He et al., "PLUM: Adapting Pre-trained Language Models for Industrial-scale Generative Recommendations" (Google, 2025) and Zhou et al., "OneRec Technical Report" (2025) extend this paradigm to production-scale LLM fine-tuning, validating the approach at the scale of real recommendation systems.
The constrained decoding components — the CSR trie structure and the fused Triton kernels for masked linear projection — are directly inspired by Su et al., "Vectorizing the Trie: Efficient Constrained Decoding for LLM-based Generative Retrieval on Accelerators" (2026), which introduces the VTNK kernel and shows how to reformulate trie-constrained beam search as a GPU-friendly sparse matrix operation, eliminating the bottleneck of sequential CPU trie traversal at large batch sizes. The fused sampling kernel (fused_linear_constrained_node_transition_sampling) additionally incorporates the Gumbel-max trick as described in Ruiz et al., "FlashSampling: Fast and Memory-Efficient Exact Sampling" (2026), fusing categorical sampling directly into the linear projection pass to avoid materializing the full logit tensor.
RQ encodes a D-dimensional embedding as L discrete codes, one per level:
r_0 = x
code_l, q_l = quantizer_l.encode(r_{l-1})
r_l = r_{l-1} - q_l (residual)
With K codes per level and L levels, the scheme supports K^L unique item IDs. For example, K=256, L=3 yields 16.7 million possible item IDs.
pip install rectokensFor development or editable installs:
pip install -e .
# With development dependencies
pip install -e ".[dev]"Requirements: Python ≥ 3.10, PyTorch ≥ 2.0, NumPy ≥ 1.24. CUDA is required for constrained decoding with GPU kernels.
This section walks through the full pipeline for training a generative recommendation model on the Amazon Reviews dataset.
AmazonReviews (in examples/data/amazon.py) auto-downloads and processes the dataset on first use. Point --root to an empty directory and the data will be downloaded there.
Supported --split values: beauty, sports, toys.
Choose RQKMeans (fast, no GPU required) or RQVAE (better reconstruction, requires GPU).
RQKMeans:
python -m examples.scripts.training.train_rqkmeans examples/configs/pretraining/train_rqkmeans_beauty.ginRQVAE:
python -m examples.scripts.training.train_rqvae examples/configs/pretraining/train_rqvae_beauty.ginEncode all item embeddings once with the fitted tokenizer and assemble the flat token-ID sequences for every user (text tokens + semantic ID tokens). This avoids repeated neural-network inference during training.
python -m examples.scripts.preprocessing.precompute_sequences examples/configs/preprocessing/precompute_sequences_beauty.ginKey config parameters (examples/configs/preprocessing/precompute_sequences_beauty.gin):
main.item_tok_type—"rqvae"(default) or"rqkmeans"main.seq_splits— tuple of splits:("train", "eval", "test")main.include_future— append the held-out future item to each sequencemain.batch_size— batch size for item embedding encoding (default 512)
python -m examples.scripts.training.finetune_qwen examples/configs/finetuning/finetune_qwen_beauty.ginKey config parameters (examples/configs/finetuning/finetune_qwen_beauty.gin):
train.loss_on—"all"(default),"items", or"text"train.bf16/train.gradient_checkpointing— recommended for GPU trainingtrain.wandb_project/train.wandb_run_name— optional W&B logging (Nonefor no logging)train.precomputed_eval_path— if set, enables mid-training evaluation
Fits codebooks via mini-batch K-means. No training loop required — call fit_step on batches of embeddings, then encode.
from rectokens.tokenizers.rq_kmeans import RQKMeansTokenizer
import numpy as np
import torch
tokenizer = RQKMeansTokenizer(
num_levels=3, # number of RQ levels (token sequence length)
codebook_size=256, # codes per level
dim=128, # embedding dimension
)
data = torch.from_numpy(np.random.randn(10_000, 128).astype("float32"))
for start in range(0, len(data), 256):
tokenizer.fit_step(data[start : start + 256])
features = torch.randn(8, 128)
tokens = tokenizer.encode(features) # TokenSequence, codes shape: (8, 3)
reconstructed = tokenizer.decode(tokens) # Tensor shape: (8, 128)
tokenizer.save("tokenizer.pt")
tokenizer = RQKMeansTokenizer.load("tokenizer.pt")Learns tokenization end-to-end via an encoder–quantizer–decoder architecture. Uses a Vector Quantization (VQ) objective with EMA codebook updates and a dead-code restart mechanism.
from rectokens.tokenizers.rqvae import RQVAETokenizer
import torch
import torch.nn.functional as F
tokenizer = RQVAETokenizer(
input_dim=128,
latent_dim=64,
hidden_dim=256,
num_levels=3,
codebook_size=256,
commitment_weight=0.25,
ema_decay=0.99,
)
optimizer = torch.optim.Adam(tokenizer.parameters(), lr=1e-3)
for batch in data_loader:
out = tokenizer(batch)
loss = F.mse_loss(out.recon, batch) + out.commitment_loss
optimizer.zero_grad()
loss.backward()
optimizer.step()
tokenizer._fitted = True
tokenizer.eval()
tokens = tokenizer.encode(features)
reconstructed = tokenizer.decode(tokens)
tokenizer.save("tokenizer.pt")
tokenizer = RQVAETokenizer.load("tokenizer.pt")The forward pass returns an RQVAEOutput NamedTuple with fields recon (reconstruction), commitment_loss, codes, and p_unique_ids (fraction of distinct token tuples in the batch).
At inference time, a recommendation model must generate token sequences that correspond to actual items in the catalog. RecTokens provides a GPU-accelerated trie for this constraint.
A trie over all valid item token sequences, stored in Compressed Sparse Row (CSR) format for efficient GPU tensor operations. The first few layers use dense lookup tables for O(1) indexing; deeper layers use sparse CSR traversal.
from rectokens.schemas.compact_csr_trie import CompactCSRTrie
# Build from the codes tensor of all items
sem_ids = tokenizer.encode(all_item_features)
trie = CompactCSRTrie.from_sorted_batch(
sem_ids.codes,
vocab_size=256,
dense_lookup_layers=2,
)from rectokens.decoding.constrained_decoding import autoregressive_generate
from rectokens.schemas.config import GenerationConfig
config = GenerationConfig(
steps=3, # token sequence length
k=10, # number of items to retrieve
beam_size=50, # beam width
temperature=1.0,
)
# model: any nn.Module whose forward returns logits over the vocab
generated = autoregressive_generate(
model=model,
trie=trie,
input_ids=user_history_ids,
generation_config=config,
)
# generated shape: (B, k, steps)The core primitive for constrained decoding is a masked linear projection. Two implementations are provided:
vtnk_pytorch— Pure PyTorch; applies a validity mask to logits before sampling.fused_linear_constrained_node_transition— Custom Triton kernel that fuses the matrix multiply and constraint masking into a single GPU kernel for maximum throughput.fused_linear_constrained_node_transition_sampling— Extends the fused kernel to additionally fuse multinomial sampling into the same GPU pass, eliminating a separatetorch.multinomialcall.
rectokens.integrations.hf bridges RecTokens item tokenizers with HuggingFace models for end-to-end training of generative recommendation models. The integration has three components:
ItemAwareTokenizer— extends a HF text tokenizer with item tokens (<item_L{l}_C{c}>), one per level/code pair. Encodes mixed text+item sequences to flat token id lists, decodes them back, and builds aCompactCSRTrieover the catalog for constrained generation.InterleavedSequenceCollator— collates mixed text/item example lists into paddedinput_ids,attention_mask, andlabelstensors ready formodel(**batch). Supportsloss_on="all"|"items"|"text"to mask loss to specific token types.PrecomputedSequenceCollator— lightweight collator for training on pre-encoded integer tensors produced byprecompute_sequences.py. No neural-network calls at collation time.ItemAwareCausalLM(rectokens.integrations.hf.model) — wraps any HF causal-LM with item-aware constrained generation. UseItemAwareCausalLM.from_causal_lmto load a model, resize its embedding table andlm_headto include item tokens, and optionally initialize item embeddings from projected RQ codebook vectors.
Use this path when you want to encode sequences on-the-fly during training, or for quick prototyping. For large-scale training, prefer the precomputed workflow above.
import torch
from torch.utils.data import DataLoader
from transformers import AutoTokenizer
from rectokens.integrations.hf.tokenizer import ItemAwareTokenizer
from rectokens.integrations.hf.collator import PrecomputedSequenceCollator
from rectokens.integrations.hf.model import ItemAwareCausalLM
from rectokens.tokenizers.rqvae import RQVAETokenizer
# 1. Load a pre-fitted item tokenizer (frozen)
item_tok = RQVAETokenizer.load("item_tok.pt").eval()
# 2. Wrap HF tokenizer to register item tokens in the vocabulary
hf_tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-2B")
aware_tokenizer = ItemAwareTokenizer(
hf_tokenizer, item_tokenizer=item_tok, num_levels=3, codebook_size=256
)
# vocab now includes 3×256 + 2 = 770 new tokens: <item_L0_C0> … <item_L2_C255> + <|item_start|> + <|item_end|>
# 3. Load model and resize embeddings to the extended vocab
model = ItemAwareCausalLM.from_causal_lm(
"Qwen/Qwen3.5-2B", aware_tokenizer, torch_dtype=torch.bfloat16
).cuda()
# 4. Encode item embeddings on-the-fly to integer semantic IDs (codes)
item_embs = torch.randn(3, item_tok.input_dim) # (N, D) — replace with real embeddings
with torch.no_grad():
codes = item_tok.encode(item_embs).codes.tolist() # list of [c0, c1, c2] per item
# 5. Build flat token-id sequences — items are semantic ID special tokens:
# <|item_start|> <item_L0_Cx> <item_L1_Cy> <item_L2_Cz> <|item_end|>
def item_ids(c): return [aware_tokenizer.item_sep_token_id, *[aware_tokenizer.item_token_id(l, v) for l, v in enumerate(c)], aware_tokenizer.item_end_token_id]
def text_ids(s): return aware_tokenizer.encode(s, add_special_tokens=False)
sequences = [
torch.tensor(text_ids("User watched ") + item_ids(codes[0]) + text_ids(" then ") + item_ids(codes[1]) + text_ids(". Next: ")),
torch.tensor(text_ids("User bought ") + item_ids(codes[2]) + text_ids(". Recommend: ")),
]
# 6. Build examples: labels are -100 at text positions, item token ids otherwise
orig_vocab = aware_tokenizer.original_vocab_size
examples = [
{"input_ids": s, "labels": torch.where(s >= orig_vocab, s, torch.full_like(s, -100))}
for s in sequences
]
# 7. Collate and train
collator = PrecomputedSequenceCollator(pad_token_id=hf_tokenizer.eos_token_id, max_length=512)
loader = DataLoader(examples, batch_size=2, collate_fn=collator)
optimizer = torch.optim.AdamW(model.parameters(), lr=2e-4)
model.train()
for batch in loader:
optimizer.zero_grad()
out = model(**{k: v.cuda() for k, v in batch.items()})
out.loss.backward()
optimizer.step()Use PrecomputedSequenceCollator with PrecomputedSequenceDataset (from examples.data.amazon) when training on sequences produced by precompute_sequences.py. No item tokenizer neural net is needed at training time — omit item_tokenizer (defaults to None).
import torch
from transformers import AutoTokenizer
from rectokens.integrations.hf.tokenizer import ItemAwareTokenizer
from rectokens.integrations.hf.collator import PrecomputedSequenceCollator
from rectokens.integrations.hf.model import ItemAwareCausalLM
from examples.data.amazon import PrecomputedSequenceDataset
from torch.utils.data import DataLoader
# 1. Load precomputed sequences
dataset = PrecomputedSequenceDataset("data/precomputed/beauty/beauty_train.pt")
# 2. HF tokenizer + ItemAwareTokenizer (no item_tokenizer neural net needed)
hf_tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-2B")
aware_tokenizer = ItemAwareTokenizer(
hf_tokenizer,
num_levels=dataset.num_levels,
codebook_size=dataset.codebook_size,
)
# 3. Load model and resize embeddings
model = ItemAwareCausalLM.from_causal_lm(
"Qwen/Qwen3.5-2B", aware_tokenizer, torch_dtype=torch.bfloat16
).cuda()
# 4. Collate precomputed integer sequences
# Labels are already masked in the precomputed data (loss on assistant turn only)
collator = PrecomputedSequenceCollator(
pad_token_id=hf_tokenizer.eos_token_id,
max_length=512,
)
loader = DataLoader(dataset, batch_size=8, collate_fn=collator)
optimizer = torch.optim.AdamW(model.parameters(), lr=2e-4)
model.train()
for batch in loader:
optimizer.zero_grad()
out = model(**{k: v.cuda() for k, v in batch.items()})
out.loss.backward()
optimizer.step()# Build a trie over the item catalog (catalog_codes: (N, num_levels) int tensor)
trie = aware_tokenizer.build_item_trie(catalog_codes, dense_lookup_layers=1)
# Use autoregressive_generate with the extended vocab trie
from rectokens.decoding.constrained_decoding import autoregressive_generate
from rectokens.schemas.config import GenerationConfig
generated = autoregressive_generate(
model=model,
trie=trie,
input_ids=user_history_ids, # (B, T) in the extended HF vocab
generation_config=GenerationConfig(steps=3, k=10, beam_size=50),
)
# generated: (B, k, 3) item token sequences in the extended HF vocab space| Script | Purpose |
|---|---|
examples/scripts/training/train_rqkmeans.py |
Train RQKMeansTokenizer on Amazon item embeddings |
examples/scripts/training/train_rqvae.py |
Train RQVAETokenizer on Amazon item embeddings |
examples/scripts/training/finetune_qwen.py |
Finetune Qwen on precomputed sequences via HF Trainer |
examples/scripts/preprocessing/precompute_sequences.py |
Encode items + assemble interleaved token-ID sequences for all users |
examples/scripts/eval/eval_retrieval.py |
Evaluate next-item retrieval with constrained beam search |
examples/scripts/benchmark/benchmark_vtnk.py |
Benchmark constrained decoding implementations |
examples/scripts/benchmark/benchmark_fused_sample.py |
Benchmark fused sampling kernel vs sparse PyTorch + multinomial |
examples/scripts/benchmark/benchmark_nn_quantize.py |
Benchmark nearest-neighbor quantization kernels |
All scripts accept a single positional gin config file. Ready-made configs live in examples/configs/ organized into pretraining/, finetuning/, and preprocessing/ subdirectories.
rectokens/
├── core/ # Abstract base classes (Tokenizer, Quantizer, Codebook)
├── tokenizers/ # RQKMeansTokenizer, RQVAETokenizer
├── quantizers/ # KMeansQuantizer, ResidualQuantizer
├── codebooks/ # EuclideanCodebook (vectorized L2 nearest-neighbor)
├── decoding/ # vtnk_pytorch, autoregressive_generate, Trie (CPU)
├── schemas/ # CompactCSRTrie, GenerationConfig, GenerationState
├── ops/ # Python wrappers for kernels
├── kernels/ # Triton GPU kernels
├── modules/ # SparseLinear, ConstraintEnforcer (abstract), SparseTrieConstraintEnforcer
└── integrations/
└── hf/ # ItemAwareTokenizer, InterleavedSequenceCollator,
# PrecomputedSequenceCollator, ItemAwareCausalLM
examples/
├── configs/
│ ├── pretraining/
│ │ ├── train_rqvae_beauty.gin # RQ-VAE tokenizer — Amazon Beauty
│ │ ├── train_rqvae_sports.gin # RQ-VAE tokenizer — Amazon Sports
│ │ ├── train_rqkmeans_beauty.gin # RQ-KMeans tokenizer — Amazon Beauty
│ │ └── train_rqkmeans_sports.gin # RQ-KMeans tokenizer — Amazon Sports
│ ├── finetuning/
│ │ ├── finetune_qwen_beauty.gin
│ │ └── eval_retrieval_beauty.gin
│ └── preprocessing/
│ └── precompute_sequences_beauty.gin
├── data/
│ └── amazon.py # AmazonReviews, ItemData, UserSequenceDataset,
│ # PrecomputedSequenceDataset
├── scripts/
│ ├── training/
│ │ ├── train_rqkmeans.py # Train RQKMeansTokenizer on Amazon item embeddings
│ │ ├── train_rqvae.py # Train RQVAETokenizer on Amazon item embeddings
│ │ └── finetune_qwen.py # Finetune Qwen via HF Trainer on precomputed data
│ ├── eval/
│ │ └── eval_retrieval.py # Evaluate next-item retrieval with beam search
│ ├── preprocessing/
│ │ └── precompute_sequences.py # Precompute interleaved token-ID sequences
│ └── benchmark/
│ ├── benchmark_vtnk.py # Benchmark constrained decoding
│ ├── benchmark_fused_sample.py # Benchmark fused sampling kernel
│ └── benchmark_nn_quantize.py # Benchmark nearest-neighbor quantization
└── utils.py # parse_config() gin helper
| Type | Description |
|---|---|
TokenSequence |
Output of encode(); holds .codes tensor of shape (N, num_levels) |
QuantizerOutput |
Single-level quantizer output: codes, quantized, residuals, commitment_loss |
ResidualQuantizerOutput |
Multi-level output: codes (B, L), quantized (B, D), level_outputs |
GenerationConfig |
Beam search config: steps, k, beam_size, temperature |
CompactCSRTrie |
GPU-resident CSR trie encoding valid item token sequences |
Run the full test suite with:
python -m pytest tests/To run a specific test file:
python -m pytest tests/test_constrained_decoding.py
python -m pytest tests/test_trie.py
python -m pytest tests/test_kernel.py
python -m pytest tests/test_tokenizers.py
python -m pytest tests/test_hf_integration.pyThe examples/scripts/benchmark/benchmark_vtnk.py script benchmarks constrained decoding implementations across batch sizes (B ∈ {256, 1024, 4096}) and vocabulary sizes (N ∈ {512, 1024, 8192, 150000}). All results below were obtained with 1% sparsity (i.e. the trie has at most 0.01 × N valid next tokens at each node), with hidden dim K=512 fixed.
python -m examples.scripts.benchmark.benchmark_vtnkThis section benchmarks fused_linear_constrained_node_transition — the kernel that fuses the linear projection and CSR trie constraint into a single GPU pass — against four baselines. Each heatmap reports the speedup ratio (values > 1 mean the fused kernel is faster) across batch sizes (B ∈ {256, 1024, 4096}) and vocabulary sizes (N ∈ {512, 1024, 8192, 150000}), with hidden dim K=512 and 1% sparsity.
Summary of findings. The fused kernel consistently outperforms all GPU baselines. Against the dense PyTorch baseline (compiled_linear+vtnk_pytorch) it is 4–9× faster across all grid points, with peak advantage at B ∈ {256, 1024} (8.2–8.8×) and a reduced but still clear margin at B=4096 (4.3–6.6×). The margin narrows at N=150k (5.2× for B ∈ {256, 1024}, 4.5× for B=4096) but remains substantial throughout. Against sparse PyTorch the advantage is 10.9–12.4× at N ≤ 1024 for B ∈ {256, 1024}, dropping to 4.9–7.4× at B=4096; at N=8192 it narrows to 1.7–8.5×, and at N=150k the fused kernel is 1.4–2.2× ahead since sparse PyTorch already skips most of the matmul work at high sparsity.
Against the two-kernel approach (separate matmul + constraint pass) the advantage is strongly N-driven: at N=150k the fused kernel is 81–95× faster, at N=8192 7.4–8.3×, and at N=1024 1.5–3.5×. At N=512 the fused kernel leads at all batch sizes: 1.1× at B=256, rising to 1.9–3.1× at larger batch sizes.
Against CPU trie traversal (benchmarked for B ∈ {256, 1024} and N ∈ {512, 1024, 8192} — the CPU baseline was too slow at larger settings) the speedups are massive: 175–578× at N=512, 289–974× at N=1024, and 1582–2267× at N=8192. The dominant axis is batch size: the GPU processes B items in parallel while the CPU traverses them serially, driving the speedup proportionally with B.
vs PyTorch (dense) — torch.compile(nn.Linear) followed by vtnk_pytorch, which applies a validity mask to the logits in a separate GPU pass after the matmul.
vs Sparse PyTorch — torch.compile(sparse_linear_pytorch), which skips columns corresponding to invalid tokens during the matmul using a sparse weight representation, but remains within the PyTorch runtime.
vs Custom Kernel — torch.compile(nn.Linear) followed by constrained_node_transition, a standalone Triton kernel that applies the CSR trie mask to precomputed logits. The matmul and masking are still two separate kernel launches.
vs CPU Trie — Pure Python traversal of an in-memory Trie on CPU, iterating over each batch item to collect valid next tokens. Included as the reference baseline for the constrained decoding problem.
The examples/scripts/benchmark/benchmark_fused_sample.py script benchmarks fused_linear_constrained_node_transition_sampling — a Triton kernel that fuses the linear projection, CSR trie masking, and multinomial sampling into a single GPU pass — against torch.compile(sparse_linear_pytorch) followed by a separate torch.softmax + torch.multinomial call. Grid and setup are the same as above (K=512, 1% sparsity).
python -m examples.scripts.benchmark.benchmark_fused_sampleSummary of findings. The fused sample kernel is faster than sparse PyTorch + multinomial at every point in the benchmark grid (B ∈ {256, 1024, 4096}, N ∈ {512, 1024, 8192, 150000}).
At N ≤ 1024 the fused kernel is 9.7–11.1× faster at B ∈ {256, 1024} and 6.3–6.4× faster at B=4096: fusing the sampling step eliminates a softmax pass and a separate torch.multinomial kernel launch whose overhead rivals the matmul cost at small N. At N=8192 the advantage ranges from 5.0–9.3× across batch sizes. At N=150k the fused kernel remains consistently faster: 8.6–11.7× at B ∈ {256, 1024} and 5.7× at B=4096, because sparse PyTorch still requires the separate multinomial pass over ~1500 valid candidates regardless of batch size. fused_linear_constrained_node_transition_sampling is the recommended choice across all measured settings.
vs Sparse PyTorch + multinomial — torch.compile(sparse_linear_pytorch) (skips invalid-token columns during the matmul) followed by torch.softmax and torch.multinomial. This is the two-step baseline that separates projection from sampling.
The core operation in residual quantization is finding the nearest codebook entry for each embedding vector — a (B, D) × (N, D) nearest-neighbor search returning B integer indices. RecTokens provides four implementations, benchmarked across batch sizes B ∈ {32, …, 65536}, embedding dims D ∈ {64, 128, 256}, and codebook sizes N ∈ {64, 128, 256, 512}.
quantize_fwd_mmis the default choice for N ≥ 128. Its matrix-multiply tile structure maps well onto GPU SIMD for any batch size, and its lead widens as both N and D grow. It beatscdist_compiledat every point in the benchmark grid for N ≥ 128, and beats FAISS-GPU at every point regardless of N.quantize_fwdwins only at N ≤ 64. Its sequential-scan approach is efficient when the inner loop over N is short enough that tiling overhead is not yet amortized. At N=64 and small batch it is up to 1.6× faster thanquantize_fwd_mm.- The crossover between
fwdandmmis governed by N × D. At small N the sequential scan wins; as N grows the scan's serial inner loop becomes the bottleneck and the MM tile kernel pulls ahead. Higher D accelerates this crossover. cdist_compiledis a reasonable fallback — simpler to deploy than Triton kernels, and its cost grows predictably with B×N. It is however consistently slower thanquantize_fwd_mmand loses toquantize_fwdat small D.- FAISS-GPU (search only, pre-built index) is the slowest implementation across virtually all settings. Its per-call dispatch overhead dominates at small N and large B. The only region where it is marginally competitive is large N (≥ 256) combined with large D (256) and very small batch (B ≤ 256) — a regime uncommon in production rec systems. The Triton kernels are otherwise 2–21× faster.
- All implementations benefit from large B, but the Triton kernels benefit most. Their speedup over FAISS and cdist grows monotonically with batch size because they achieve near-linear GPU utilization scaling while FAISS's fixed dispatch cost remains constant.
Each Triton block handles one or more rows of x and iterates over the full codebook of size N in sequence, accumulating the minimum L2 distance in registers.
Strengths: Very low launch overhead; minimal shared memory pressure; optimal at N ≤ 64 where the inner loop is short; memory access pattern is sequential and cache-friendly.
Weaknesses: The inner loop over N is serial within each thread block, so kernel time scales linearly with N. At large N or large D the kernel stalls waiting for memory, and the MM-style kernel's parallelism over N tiles wins decisively. At N=512, D=256, fwd is over 3× slower than mm.
Reformulates the L2 nearest-neighbor problem as a matrix multiplication over B×N tiles. Multiple thread blocks cooperate over the N dimension in parallel.
Strengths: Strong GPU utilization at any (B, N, D) combination; tiling amortizes launch overhead effectively; consistently fastest at N ≥ 128; no failure modes — it always beats FAISS.
Weaknesses: Higher launch overhead and shared memory pressure than the sequential kernel; slight disadvantage at N ≤ 64 + small batch where the tile setup cost is not yet amortized.
Computes the full pairwise distance matrix (B, N) using torch.cdist, then takes argmin along N. Compiled with torch.compile for fused kernel dispatch.
Strengths: Zero external dependencies beyond PyTorch; leverages cuBLAS for the pairwise distance core; D-insensitive performance because the matmul is fully blocked.
Weaknesses: Allocates an intermediate (B, N) distance buffer that grows with both B and N; two-pass execution (matmul then argmin) prevents full fusion. Consistently slower than both Triton kernels at large B, and loses to quantize_fwd at small D.
Uses a FAISS IndexFlatL2 GPU index built once from the codebook (make_gpu_index) and queried at inference time with index.search(x, 1).
Strengths: Battle-tested implementation; plugs into the broader FAISS ecosystem for approximate search extensions; cuBLAS-backed distance computation is highly optimized for large D.
Weaknesses: Large per-call dispatch overhead even on GPU, not amortized at small N or large B. The Triton kernels are 2–21× faster across the benchmark grid. Only marginally competitive at N ≥ 256, D=256, and B ≤ 256 — a narrow corner of the operating space.
python -m examples.scripts.benchmark.benchmark_nn_quantizeGrid: B ∈ {32, 256, 1024, 4096, 16384, 32768, 65536}, D ∈ {64, 128, 256}, N ∈ {64, 128, 256, 512}. Heatmap axes are batch size (B, rows) vs embedding dim (D, columns). Speedup values > 1 mean the left-hand kernel is faster. All FAISS timings exclude index build time (static codebook).
N=64 — fwd leads at small batch and low D, mm takes over at large D or large B
At N=64 the sequential scan is fast enough to outrun the MM tile kernel across most of the grid. quantize_fwd is up to 1.62× faster at (B=32, D=64), where the tile setup overhead of mm is not yet amortized and the codebook is small enough that the serial inner loop completes quickly. The advantage decays along two axes: increasing D multiplies the per-row scan cost, and increasing B eventually saturates the GPU such that the parallel tile structure of mm becomes the more efficient fit. At D=256 the advantage collapses entirely: fwd falls to 0.54–0.84× of mm for large B. The crossover diagonal runs from (small B, high D) to (large B, low D). For N=64 with D ≤ 128 and B ≤ 4096, fwd is the faster choice.
N=256 — mm dominates; fwd never recovers
With N=256 the sequential inner loop in quantize_fwd is 4× longer than at N=64, and the MM tile kernel's parallel decomposition over N fully materializes. fwd is slower at nearly every cell, ranging from 0.45× (B=32, D=256 — more than 2× slower) to a single near-parity cell at (B=65536, D=128: 1.03×). The D=256 column is universally dark: 0.45–0.61× regardless of batch size. The gradient along the D axis is the dominant signal — wider embeddings amplify the per-row scan cost linearly while the MM kernel's tile width absorbs the extra work with no additional penalty. At N=256 and above, quantize_fwd_mm is unambiguously the correct choice.
N=256 — fwd wins at small D with growing margin, but loses at large D + small batch
At D=64, quantize_fwd beats cdist_compiled at every batch size (0.98–3.08×), and the speedup grows monotonically with B — as the batch size grows the Triton kernel's per-row parallelism scales efficiently while cdist's two-pass (matmul + argmin) dispatch cost amortizes more slowly. At D=128 the advantage shrinks to 0.72–1.94× and is only reliable for B ≥ 1024. At D=256, fwd loses at small batch (0.47–0.75× for B ≤ 256) and barely reaches parity at the largest batch sizes (B=65536, 0.97×). This failure at high D reflects the sequential kernel's sensitivity to embedding width: cdist delegates to a cuBLAS batched matmul that remains equally efficient regardless of D, while fwd's inner loop cost grows with D. For N=256 and D=256, cdist_compiled is actually preferable to quantize_fwd at small batch.
N=256 — mm beats cdist at every point; speedup driven by B and D
Unlike quantize_fwd, the MM kernel beats cdist_compiled at every cell in this grid without exception. The minimum speedup is 1.03× (B=32, D=256) and the maximum is 3.72× (B=65536, D=64). The speedup grows primarily along the B axis — as batch size increases the tile structure of mm better saturates the GPU while cdist's two-kernel dispatch overhead becomes the dominant cost. The secondary gradient runs along D inversely: higher D slightly reduces mm's relative advantage because cdist's cuBLAS core is particularly well-optimized for wide embeddings. Even so, the D=256 column reaches 1.61× at B=65536, a meaningful and reliable margin. This heatmap confirms quantize_fwd_mm as the dominant general-purpose kernel for N ≥ 128.
N=64 — large and growing advantage for fwd across all batch sizes
At N=64, FAISS-GPU's fixed per-call dispatch overhead dwarfs the actual distance computation, making the Triton kernel overwhelmingly faster. quantize_fwd is 2.34–20.85× faster across the grid. The speedup pattern is strikingly regular: it is nearly uniform across D at small batch (B=32: 2.34–4.05×, increasing with D), and then fans out rapidly as B grows — at B=65536 the margin reaches 4.51–20.85×. This proportional growth with B is the direct signature of FAISS's fixed overhead: the Triton kernel's useful work scales with B while FAISS's dispatch cost stays constant, driving the ratio linearly. Even at the smallest batch size tested (B=32), fwd is 2–4× faster, meaning there is no operating point at N=64 where FAISS is competitive.
N=256 — fwd loses at large D + small batch; FAISS's cuBLAS core becomes visible
At N=256 the sequential scan's increasing cost at large D begins to rival FAISS's fixed overhead at small batch. The D=256 column dips below 1× for B ≤ 256 (0.73× at B=32, 0.80× at B=256): the serial inner loop over 256 codebook entries × 256 dimensions is slow enough that FAISS's cuBLAS-backed distance kernel, despite its overhead, matches or slightly beats it. At D=64 the advantage remains strong and B-driven (1.76–9.17×) — FAISS's overhead is not amortized here. The practical takeaway: if you are using N=256, D=256, and very small batch sizes (B ≤ 256), FAISS is a viable alternative to quantize_fwd specifically; for all other settings fwd wins.
N=64 — mm wins everywhere, uniform strength across D
At N=64, quantize_fwd_mm beats FAISS-GPU at every cell: 2.24–20.28×. The overall pattern mirrors the fwd vs FAISS heatmap — B-driven growth, large margins at large batch — but with one key difference: the D=256 column is competitive (2.24–8.37×) rather than being the weakest column. Where fwd has elevated cost at large D (sequential scan), mm's tiling keeps cost controlled, so the D=256 minimum (B=32, 2.24×) is only marginally lower than D=64 (B=32, 2.50×). mm provides a more uniform advantage profile than fwd against FAISS because it does not have the sequential scan's D-scaling weakness.
N=256 — mm wins at every cell; no failure mode at large D + small batch
This heatmap is the clearest illustration of quantize_fwd_mm's robustness. While quantize_fwd fell below FAISS at N=256, D=256, small B, mm maintains a positive margin everywhere: 1.60–11.06×. The D=256 column ranges from 1.60× (B=32) to 3.19× (B=65536) — a clear win even in the exact regime where fwd lost. The D=64 column reaches 11.06× at B=65536 driven by B-scaling. The bottom-right region (large B, any D) is brightest, consistent with FAISS's fixed overhead being overwhelmed by the volume of useful work. This heatmap, paired with the N=256 fwd vs faiss heatmap, is the strongest argument for preferring mm over fwd as the default kernel for standard rec-system codebook sizes.
Apache 2.0