vLLM Reference Audit

This audit maps the local vLLM-related reference work to turboquant. Each later vLLM phase must update the affected row before the phase is treated as done.

Status values:

• present: implemented and tested in this repository.
• partial: some behavior exists, but a listed runtime path or test group is absent.
• missing: no implementation exists here yet.
• rejected: not being adopted, with a reason.

File Map

Reference file	Feature name	Data structures	Public functions/classes	Runtime path	Tests	Our target	Status
turboquant-vllm/src/turboquant_vllm/lloyd_max.py	Lloyd-Max codebooks	codebook centroids, boundaries	solve_lloyd_max, LloydMaxCodebook	quantizer setup	test_lloyd_max.py	turboquant/core/codebook.py	present
turboquant-vllm/src/turboquant_vllm/quantizer.py	MSE and QJL codecs	indices, norms, signs	TurboQuantMSE, TurboQuantProd	HF and TQ4 diagnostics	test_quantizer.py	turboquant/core/	present
turboquant-vllm/src/turboquant_vllm/compressors.py	key/value compressors	compressed keys, compressed values	key and value compressor classes	diagnostic compression	test_compressors.py	turboquant/kv/keys.py, turboquant/kv/values.py	partial
turboquant-vllm/src/turboquant_vllm/kv_cache.py	HF dynamic cache wrapper	compressed layers, lifecycle state	dynamic cache wrapper, stats methods	HF diagnostic path	test_kv_cache_*.py	turboquant/integration/hf/dynamic_cache.py	partial
turboquant-vllm/src/turboquant_vllm/verify.py	model verification	model shape record, per-layer metrics	config detector, verification runner	CLI quality check	test_verify*.py	turboquant/integration/vllm/verify.py	partial
turboquant-vllm/src/turboquant_vllm/benchmark.py	A/B measurement	run result dicts	model load, patched run, benchmark runner	local benchmark	test_benchmark.py	scripts/qwen_ab_eval.py	partial
turboquant-vllm/src/turboquant_vllm/triton/tq4_compress.py	TQ4 pack kernel	packed index bytes, fp32 norms	tq4_compress	packed cache write	test_triton_multi_dim.py	turboquant/triton/tq4_update.py	partial
turboquant-vllm/src/turboquant_vllm/triton/tq4_decompress.py	TQ4 unpack kernel	packed index bytes, norms	tq4_decompress	diagnostic readback	test_tq4_out_parameter.py	turboquant/triton/tq4_decode.py	partial
turboquant-vllm/src/turboquant_vllm/triton/fused_qk_attention.py	packed score kernel	packed key rows	fused_qk_scores	score-only diagnostic	kernel tests	turboquant/triton/tq4_decode.py	missing
turboquant-vllm/src/turboquant_vllm/triton/fused_paged_tq4_attention.py	TQ4 paged decode	block table, packed pages	fused_paged_tq4_decode	decode attention	test_fused_paged_tq4_attention.py	turboquant/triton/tq4_decode.py	partial
turboquant-vllm/src/turboquant_vllm/triton/fused_paged_tq4_int8_prefill.py	INT8-Q prefill	int8 query scale, packed pages	prefill wrapper	prefill experiment	test_fused_paged_tq4_int8.py	turboquant/triton/int8_prefill.py	missing
turboquant-vllm/src/turboquant_vllm/vllm/tq4_backend.py	plugin-style backend	TQ4 spec, backend, impl	backend registration, page shape, dispatch	vLLM plugin path	test_vllm_*.py	turboquant/integration/vllm/	partial
vllm-turboquant/vllm/v1/attention/ops/turboquant_kv_cache.py	recipe layout and vector codec	group layout, kernel meta	layout, transforms, pack/unpack	native vLLM recipe path	test_turboquant.py	turboquant/integration/vllm/recipe.py, vector.py	partial
vllm-turboquant/vllm/v1/attention/ops/turboquant_metadata.py	metadata JSON	tensor/layer/calibration metadata	load, save, discovery, TP slicing	native vLLM recipe path	test_turboquant.py	turboquant/integration/vllm/metadata.py	partial
vllm-turboquant/vllm/v1/attention/ops/triton_turboquant_kv_update.py	recipe update kernel	packed recipe page rows	turboquant_write_packed_kv	KV cache update	test_turboquant.py	turboquant/triton/recipe_update.py	missing
vllm-turboquant/vllm/v1/attention/ops/triton_turboquant_decode.py	recipe decode kernel	block table, token metadata, packed pages	decode forward wrapper	paged attention	test_turboquant.py	turboquant/triton/recipe_decode.py	missing
vllm-turboquant/vllm/v1/attention/backends/triton_attn.py	backend integration	metadata builder, layer impl state	update, decode, fallback gates	vLLM runtime	test_turboquant.py	turboquant/integration/vllm/runtime.py	missing
vllm-turboquant/vllm/config/cache.py	cache config gate	cache dtype, gate, metadata path	config validators	vLLM config	config tests	turboquant/integration/vllm/registry.py	missing
vllm-turboquant/vllm/engine/arg_utils.py	CLI flags	engine args	enable flag, metadata path	vLLM CLI	config tests	scripts/verify_vllm_integration.py	missing
vllm-turboquant/benchmarks/generate_turboquant_metadata.py	activation calibration	activation accumulator, metadata builder	metadata generator CLI	recipe metadata build	test_turboquant.py	turboquant/integration/vllm/calibration.py, scripts/generate_vllm_metadata.py	missing

Feature Checklist

Feature group	Our target	Status	Next action
Lloyd-Max codebooks	turboquant/core/codebook.py	present	keep exact Beta tests
QR/RHT/block-Hadamard transforms	turboquant/core/rotation.py, turboquant/integration/vllm/vector.py	partial	expose query/group helpers
QJL residual	turboquant/core/qjl.py, turboquant/integration/vllm/vector.py	partial	add recipe decode contract
bit packing	turboquant/packing/bits.py	present	keep 1/2/3/4/5-bit tests
TQ4 cache wrapper	turboquant/integration/hf/dynamic_cache.py	partial	add Transformers model tests
TQ4 packed page layout	turboquant/integration/vllm/page_cache.py	present	add kernel-backed writer
asymmetric K/V bits	turboquant/integration/vllm/recipe.py, page cache	partial	add page bytes and tests
turboquant25/35 recipe layout	turboquant/integration/vllm/recipe.py	partial	add public aliases from plan
metadata JSON	turboquant/integration/vllm/metadata.py	partial	validate recipe and runtime shape
activation calibration	turboquant/integration/vllm/calibration.py	present	add model-family tests
TP metadata slicing	turboquant/integration/vllm/metadata.py	present	expand tests
vLLM config hooks	turboquant/integration/vllm/registry.py	missing	add gate checks
vLLM cache dtype registry	turboquant/integration/vllm/registry.py	missing	map dtype to uint8
vLLM attention backend selection	turboquant/integration/vllm/runtime.py	missing	add dispatch layer
packed KV update kernel	turboquant/triton/tq4_update.py, recipe_update.py	missing	add reference then Triton
packed paged decode kernel	turboquant/triton/tq4_decode.py, recipe_decode.py	missing	add reference then Triton
INT8-Q prefill	turboquant/triton/int8_prefill.py	missing	add gated experiment
CUDA graph buffers	turboquant/integration/vllm/runtime.py	missing	add buffer owner
SWA/global-layer behavior	HF and vLLM adapters	missing	add config tests
model verification	turboquant/integration/vllm/verify.py	partial	add real model smoke
Qwen A/B benchmark	scripts/qwen_ab_eval.py	partial	add vLLM variants