feat: turboquant encoding for vectors#7167

Draft

lwwmanning wants to merge 25 commits intodevelopfrom

claude/admiring-lichterman

Commits on Mar 27, 2026

feat[turboquant]: add TurboQuant vector quantization encoding

lwwmanning
and
claude
committed
feat[turboquant]: add TurboQuantCompressor and WriteStrategyBuilder integration

lwwmanning
and
claude
committed
refactor[turboquant]: integrate into BtrBlocks compressor directly

lwwmanning
and
claude
committed
bench[turboquant]: add compression/decompression throughput benchmarks

lwwmanning
and
claude
committed
perf[turboquant]: replace dense rotation with randomized Hadamard transform

lwwmanning
and
claude
committed
test[turboquant]: add theoretical error bound and inner product bias tests

lwwmanning
and
claude
committed
chore[turboquant]: fix review issues and generate public-api.lock

lwwmanning
and
claude
committed
chore[turboquant]: review cleanup — tighter tests, naming, validation

lwwmanning
and
claude
committed
docs[turboquant]: add crate-level docs with compression ratios and error bounds

lwwmanning
and
claude
committed
feat[turboquant]: support 1-8 bit quantization

lwwmanning
and
claude
committed
feat[turboquant]: support 9-bit Prod for tensor core int8 GEMM

lwwmanning
and
claude
committed
bench[turboquant]: add dim 1024 and 1536 benchmarks

lwwmanning
and
claude
committed
feat[turboquant]: add rotation sign export/import and hot-path inverse

lwwmanning
and
claude
committed
feat[turboquant]: define TurboQuantMSEArray and TurboQuantQJLArray

lwwmanning
and
claude
committed
feat[turboquant]: add new compression functions for cascaded arrays

lwwmanning
and
claude
committed
refactor[btrblocks]: simplify TurboQuant compressor for cascaded arrays

lwwmanning
and
claude
committed
chore[turboquant]: regenerate public-api.lock for new array types

lwwmanning
and
claude
committed
refactor[turboquant]: restructure into subdirectory modules, delete dead code

lwwmanning
and
claude
committed
test[turboquant]: improve test coverage and add explanatory comments

lwwmanning
and
claude
committed
perf[turboquant]: restore fast SIMD-friendly decode by expanding stored signs

lwwmanning
and
claude
committed
fix[turboquant]: address PR review findings

lwwmanning
and
claude
committed
fix[turboquant]: second-round review fixes and merge conflict resolution

lwwmanning
and
claude
committed
refactor[turboquant]: simplify code from review findings

lwwmanning
and
claude
committed
fix[turboquant]: address PR review comments from AdamGS

lwwmanning
and
claude
committed
chore[turboquant]: cleanup from second simplify pass

lwwmanning
and
claude
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat: turboquant encoding for vectors#7167

feat: turboquant encoding for vectors#7167
lwwmanning wants to merge 25 commits intodevelopfrom
claude/admiring-lichterman

Commits on Mar 27, 2026

feat[turboquant]: add TurboQuant vector quantization encoding

feat[turboquant]: add TurboQuantCompressor and WriteStrategyBuilder integration

refactor[turboquant]: integrate into BtrBlocks compressor directly

bench[turboquant]: add compression/decompression throughput benchmarks

perf[turboquant]: replace dense rotation with randomized Hadamard transform

test[turboquant]: add theoretical error bound and inner product bias tests

chore[turboquant]: fix review issues and generate public-api.lock

chore[turboquant]: review cleanup — tighter tests, naming, validation

docs[turboquant]: add crate-level docs with compression ratios and error bounds

feat[turboquant]: support 1-8 bit quantization

feat[turboquant]: support 9-bit Prod for tensor core int8 GEMM

bench[turboquant]: add dim 1024 and 1536 benchmarks

feat[turboquant]: add rotation sign export/import and hot-path inverse

feat[turboquant]: define TurboQuantMSEArray and TurboQuantQJLArray

feat[turboquant]: add new compression functions for cascaded arrays

refactor[btrblocks]: simplify TurboQuant compressor for cascaded arrays

chore[turboquant]: regenerate public-api.lock for new array types

refactor[turboquant]: restructure into subdirectory modules, delete dead code

test[turboquant]: improve test coverage and add explanatory comments

perf[turboquant]: restore fast SIMD-friendly decode by expanding stored signs

fix[turboquant]: address PR review findings

fix[turboquant]: second-round review fixes and merge conflict resolution

refactor[turboquant]: simplify code from review findings

fix[turboquant]: address PR review comments from AdamGS

chore[turboquant]: cleanup from second simplify pass