fix/issue 656 default block size factor by miroslavln · Pull Request #1 · miroslavln/llm-d-kv-cache

miroslavln · 2026-06-10T17:12:13Z

fs_backend: add get_offload_block_hash to storage_events test stub (fs_backend: add get_offload_block_hash to storage_events test stub llm-d/llm-d-kv-cache#613)
The test stubs vllm.v1.kv_offload.base to load the real manager
module in isolation. Commit 8cf550e added get_offload_block_hash
to manager.py's imports, but the stub wasn't updated, so collection
of tests/test_storage_events.py failed with "cannot import name
'get_offload_block_hash' from 'vllm.v1.kv_offload.base' (unknown
location)". Add an identity stub so the import resolves and the
existing assertions (which pass plain ints as keys) still hold.

Signed-off-by: Kfir Toledo kfir.toledo@ibm.com
fix(pvc_evictor): walk current FileMapper layout in crawler (fix(pvc_evictor): walk current FileMapper layout in crawler llm-d/llm-d-kv-cache#611)
- pvc_evictor: walk current FileMapper layout in crawler
After fs_backend: collapse FileMapper path + add config.json audit llm-d/llm-d-kv-cache#585 collapsed the FileMapper layout to
<root>/<safe_model_name>_<sha256-12>_r<rank>/<hhh>/<hh>_g<group_idx>/*.bin,
the crawler still walked the pre-fs_backend: collapse FileMapper path + add config.json audit llm-d/llm-d-kv-cache#585 deep tree
(block_size_*/tp_*_pp_size_*/rank_*/<dtype>/...). Those paths no
longer exist on disk, so the crawler discovered zero files and the
evictor never freed any blocks.

This rewrites stream_cache_files_with_mapper to:
- find rank directories by their _r<digits> suffix anywhere under
  the cache root, instead of pattern-matching the deprecated deep tree;
- iterate the first-level hex bucket ({hhh}) and apply the existing
  hex_modulo_range sharding;
- yield .bin files from any second-level bucket underneath, leaving
  the _g<group_idx> encoding opaque so the walker doesn't need to
  understand kv-cache groups.
The new walker no longer instantiates FileMapper (it only inspects
on-disk names), so the FILEMAPPER_AVAILABLE import guard and the
matching early-exit in crawler_process are removed. This also
unblocks running the evictor container without vllm installed, since
the walker was the only consumer of the from llmd_fs_backend.file_mapper import FileMapper import — which transitively required vllm.

parse_filemapper_params is dropped along with its sole caller.

Signed-off-by: Miro mironikolov@google.com
- feat: parameterize first-level hex bucket directory length in pvc crawler
Signed-off-by: Miro mironikolov@google.com

Signed-off-by: Miro mironikolov@google.com
fix(pvc_evictor): remove unused re import after fix(pvc_evictor): walk current FileMapper layout in crawler llm-d/llm-d-kv-cache#611 (fix(pvc_evictor): remove unused re import after #611 llm-d/llm-d-kv-cache#615)
Signed-off-by: Guy Girmonsky guygir@gmail.com
feat(kvevents): parse HMA KV event metadata (feat(kvevents): parse HMA KV event metadata llm-d/llm-d-kv-cache#612)
- parse hma kv event metadata
Signed-off-by: Sage Ahrac sagiahrak@gmail.com

Co-authored-by: Kapil Jain 16477749+kapiljain1989@users.noreply.github.com
- reduce noise
Signed-off-by: Sage Ahrac sagiahrak@gmail.com

Signed-off-by: Sage Ahrac sagiahrak@gmail.com
Co-authored-by: Kapil Jain 16477749+kapiljain1989@users.noreply.github.com
fix Dockerfile.dev breakage (fix Dockerfile.dev breakage for llmd_fs_backend llm-d/llm-d-kv-cache#620)
fix CUDA version mismatch and dev headers symlink
- Update default CUDA_TOOLKIT_PKG to cuda-toolkit-13-0 to
  match the CUDA 13.0 base image and prevent PyTorch compilation
  version mismatch.
- Explicitly parse and update the standard /usr/local/cuda symlink
  after GKE package installation to resolve missing dev headers
  (cusparse.h) during compilation
Signed-off-by: Saikat Roychowdhury saikat.royc85@gmail.com
feat: Add PVC evictor BlockRemoved events (feat: Add PVC evictor BlockRemoved events llm-d/llm-d-kv-cache#605)
- feat: Emit BlockRemoved events in PVC evictor
Signed-off-by: Alberto Perdomo aperdomo@redhat.com
- test: Add deleter and BlockRemoved tests
Signed-off-by: Alberto Perdomo aperdomo@redhat.com
- chore: Refactor after FileMapper changes
Signed-off-by: Alberto Perdomo aperdomo@redhat.com
- chore: Minor changes
Signed-off-by: Alberto Perdomo aperdomo@redhat.com
- chore: Fix lint issues
Signed-off-by: Alberto Perdomo aperdomo@redhat.com
- fix: Use only valid paths for events
Signed-off-by: Alberto Perdomo aperdomo@redhat.com
- fix: Emit events on shutdown
Signed-off-by: Alberto Perdomo aperdomo@redhat.com

Signed-off-by: Alberto Perdomo aperdomo@redhat.com
ci: Wire fs_backend Python tests into CI (ci: Wire fs_backend Python tests into CI llm-d/llm-d-kv-cache#578)
- ci: Wire fs_backend Python tests into CI
Signed-off-by: Alberto Perdomo aperdomo@redhat.com
- chore: Clean up branch
Signed-off-by: Alberto Perdomo aperdomo@redhat.com
- fix: Use venv for llmd_fs_backend test Python deps
Signed-off-by: Alberto Perdomo aperdomo@redhat.com
- chore: Support PVC evictor events
Signed-off-by: Alberto Perdomo aperdomo@redhat.com

Signed-off-by: Alberto Perdomo aperdomo@redhat.com

deps(go): bump the go-dependencies group across 1 directory with 16 updates (deps(go): bump the go-dependencies group across 1 directory with 16 updates llm-d/llm-d-kv-cache#630)
Bumps the go-dependencies group with 10 updates in the / directory:

Package	From	To
github.com/alicebob/miniredis/v2	`2.35.0`	`2.38.0`
github.com/dgraph-io/ristretto/v2	`2.3.0`	`2.4.0`
github.com/docker/docker	`28.5.1+incompatible`	`28.5.2+incompatible`
github.com/fxamacker/cbor/v2	`2.7.0`	`2.9.2`
github.com/prometheus/client_golang	`1.22.0`	`1.23.2`
github.com/redis/go-redis/v9	`9.7.3`	`9.20.0`
github.com/testcontainers/testcontainers-go	`0.40.0`	`0.42.0`
go.opentelemetry.io/contrib/instrumentation/google.golang.org/grpc/otelgrpc	`0.63.0`	`0.69.0`
go.opentelemetry.io/otel/exporters/otlp/otlptrace/otlptracegrpc	`1.39.0`	`1.44.0`
go.uber.org/zap	`1.27.0`	`1.28.0`

Updates github.com/alicebob/miniredis/v2 from 2.35.0 to 2.38.0

Updates github.com/dgraph-io/ristretto/v2 from 2.3.0 to 2.4.0

Updates github.com/docker/docker from 28.5.1+incompatible to 28.5.2+incompatible

Updates github.com/fxamacker/cbor/v2 from 2.7.0 to 2.9.2

Updates github.com/prometheus/client_golang from 1.22.0 to 1.23.2

Updates github.com/prometheus/client_model from 0.6.1 to 0.6.2

Updates github.com/redis/go-redis/v9 from 9.7.3 to 9.20.0

Updates github.com/testcontainers/testcontainers-go from 0.40.0 to 0.42.0

Updates go.opentelemetry.io/contrib/instrumentation/google.golang.org/grpc/otelgrpc from 0.63.0 to 0.69.0

Updates go.opentelemetry.io/otel from 1.43.0 to 1.44.0

Updates go.opentelemetry.io/otel/exporters/otlp/otlptrace/otlptracegrpc from 1.39.0 to 1.44.0

Updates go.opentelemetry.io/otel/sdk from 1.43.0 to 1.44.0

Updates go.opentelemetry.io/otel/trace from 1.43.0 to 1.44.0

Updates go.uber.org/zap from 1.27.0 to 1.28.0

Updates google.golang.org/grpc from 1.77.0 to 1.81.1

Updates google.golang.org/protobuf from 1.36.10 to 1.36.11

updated-dependencies:

dependency-name: github.com/alicebob/miniredis/v2
dependency-version: 2.38.0
dependency-type: direct:production
update-type: version-update:semver-minor
dependency-group: go-dependencies
dependency-name: github.com/dgraph-io/ristretto/v2
dependency-version: 2.4.0
dependency-type: direct:production
update-type: version-update:semver-minor
dependency-group: go-dependencies
dependency-name: github.com/docker/docker
dependency-version: 28.5.2+incompatible
dependency-type: direct:production
update-type: version-update:semver-patch
dependency-group: go-dependencies
dependency-name: github.com/fxamacker/cbor/v2
dependency-version: 2.9.2
dependency-type: direct:production
update-type: version-update:semver-minor
dependency-group: go-dependencies
dependency-name: github.com/prometheus/client_golang
dependency-version: 1.23.2
dependency-type: direct:production
update-type: version-update:semver-minor
dependency-group: go-dependencies
dependency-name: github.com/prometheus/client_model
dependency-version: 0.6.2
dependency-type: direct:production
update-type: version-update:semver-patch
dependency-group: go-dependencies
dependency-name: github.com/redis/go-redis/v9
dependency-version: 9.20.0
dependency-type: direct:production
update-type: version-update:semver-minor
dependency-group: go-dependencies
dependency-name: github.com/testcontainers/testcontainers-go
dependency-version: 0.42.0
dependency-type: direct:production
update-type: version-update:semver-minor
dependency-group: go-dependencies
dependency-name: go.opentelemetry.io/contrib/instrumentation/google.golang.org/grpc/otelgrpc
dependency-version: 0.69.0
dependency-type: direct:production
update-type: version-update:semver-minor
dependency-group: go-dependencies
dependency-name: go.opentelemetry.io/otel
dependency-version: 1.44.0
dependency-type: direct:production
update-type: version-update:semver-minor
dependency-group: go-dependencies
dependency-name: go.opentelemetry.io/otel/exporters/otlp/otlptrace/otlptracegrpc
dependency-version: 1.44.0
dependency-type: direct:production
update-type: version-update:semver-minor
dependency-group: go-dependencies
dependency-name: go.opentelemetry.io/otel/sdk
dependency-version: 1.44.0
dependency-type: direct:production
update-type: version-update:semver-minor
dependency-group: go-dependencies
dependency-name: go.opentelemetry.io/otel/trace
dependency-version: 1.44.0
dependency-type: direct:production
update-type: version-update:semver-minor
dependency-group: go-dependencies
dependency-name: go.uber.org/zap
dependency-version: 1.28.0
dependency-type: direct:production
update-type: version-update:semver-minor
dependency-group: go-dependencies
dependency-name: google.golang.org/grpc
dependency-version: 1.81.1
dependency-type: direct:production
update-type: version-update:semver-minor
dependency-group: go-dependencies
dependency-name: google.golang.org/protobuf
dependency-version: 1.36.11
dependency-type: direct:production
update-type: version-update:semver-patch
dependency-group: go-dependencies
...

Signed-off-by: dependabot[bot] support@github.com
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

fix: remove MIN_STAGING_BUFFER_SIZE that inflates KV cache on disk (fix: remove MIN_STAGING_BUFFER_SIZE that inflates KV cache on disk llm-d/llm-d-kv-cache#589)
The 16 MB floor on staging buffers caused every per-file write to pad
the buffer via cudaHostAlloc, then write the full padded size to disk.
For small-block models (e.g. Llama 3.1 8B where block size is ~7 MB)
this more than doubles the on-disk footprint (7 MB -> 16 MB per file)
and degrades TTFT from 2.8 s to 6.5 s at 40k-token prefix length.

Remove the constant so allocate_staging_buffer uses exactly the size
that calc_staging_bytes computes.

Fixes Remove MIN_STAGING_BUFFER_SIZE, leads to way more data stored than required on smaller models llm-d/llm-d-kv-cache#454

Signed-off-by: Jonathan Wrede wrede.jonathan00@gmail.com
style(fs-backend): collapse double blank line in thread_pool.cpp (style(fs-backend): fix clang-format break in thread_pool.cpp llm-d/llm-d-kv-cache#634)
clang-format (v21) flags the double blank line left after fix: remove MIN_STAGING_BUFFER_SIZE that inflates KV cache on disk llm-d/llm-d-kv-cache#589 removed
MIN_STAGING_BUFFER_SIZE. Fixes the repo-wide pre-commit lint gate that
currently fails on all PRs.

Signed-off-by: Kfir Toledo kfir.toledo@ibm.com
fs_backend: migrate kv_offload API to vllm 0.22.0 (fs_backend: migrate kv_offload API to vllm 0.22.0 llm-d/llm-d-kv-cache#628)
The vllm 0.22 KV-offload Python API (base, worker, OffloadPromMetrics) is
unchanged from 0.21, so no import/signature changes are needed.

The one functional change is in spec.py: adopt the OffloadingSpec base
class's $self.hash_block_size$ (resolved via resolve_kv_cache_block_sizes)
instead of $vllm_config.cache_config.block_size$ when computing
gpu_blocks_per_file. The two are equal for standard single-group models,
but cache_config.block_size can be larger on hybrid models, so the base
value is the correct hash granularity. Single-group only (no HMA).

Bump pins: vllm 0.21.0 -> 0.22.0, package version 0.21 -> 0.22. README,
Dockerfile.dev base image, and vllm-storage.yaml deployment example
updated accordingly.

Signed-off-by: Kfir Toledo kfir.toledo@ibm.com
feat: batch KV block copies via cudaMemcpyBatchAsync in fs connector (feat: batch KV block copies via cudaMemcpyBatchAsync in fs connector llm-d/llm-d-kv-cache#607)
- csrc: batch KV block copies via cudaMemcpyBatchAsync
Submit all per-(block, layer) copies in one driver call instead of N
cudaMemcpyAsync calls. Enabled by default; toggle off with
USE_BATCH_MEMCPY_READ / USE_BATCH_MEMCPY_WRITE=0. Requires CUDA 12.8+.

Speeds up KV-cache offload writes/reads when per-layer DMA sizes are
small enough that driver dispatch dominates.

Signed-off-by: Kfir Toledo kfir.toledo@ibm.com
- csrc: fall back to per-call DMA on CUDA < 12.8
cudaMemcpyBatchAsync was introduced in CUDA 12.8 — guard the batch
path with #if CUDA_VERSION >= 12080 and route to the per-call
cudaMemcpyAsync loop below that. Default USE_BATCH_MEMCPY_* off on
older toolchains so the env knob still makes sense.

Also drop thread_local on the attrs/attrs_idx inputs (never mutated,
no per-thread duplication needed) and move the copy_blocks dispatcher
below the helpers it dispatches to.

Signed-off-by: Kfir Toledo kfir.toledo@ibm.com

Signed-off-by: Kfir Toledo kfir.toledo@ibm.com
docs(fs-backend): note upstreaming into vLLM multi-tier offloading (docs(fs-backend): note upstreaming into vLLM multi-tier offloading llm-d/llm-d-kv-cache#632)
llmd-fs-connector==0.22 (llm-d v0.8 / vLLM v0.22) is the final release of
the standalone llm-d FS connector. The filesystem offloading logic is now
upstreamed into vLLM as the FS tier of the multi-tier offloading connector
(TieringOffloadingSpec); all new features and support continue there.

Add an [!IMPORTANT] banner to the connector README and a short note in the
root README's Connectors & Utilities list, linking the vLLM KV offloading
guide ([Docs] Add KV offloading usage guide (single- and multi-tier) vllm-project/vllm#44415).

Signed-off-by: Kfir Toledo kfir.toledo@ibm.com
feat(kvblock): trace index add and evict (feat(kvblock): trace index add and evict llm-d/llm-d-kv-cache#637)
Signed-off-by: Dong Ma winterma.dong@gmail.com
ci: refresh :latest tag only on stable GitHub Release (ci: refresh ghcr :latest tag on stable GitHub Release llm-d/llm-d-kv-cache#624)
Add a latest input (default false) to the shared docker-build-and-push
action and wire it up from ci-release.yaml and
ci-release-uds-tokenizer.yaml. The callers set it to true only when the
triggering event is a non-prerelease GitHub Release.

Previously the release workflows only pushed the immutable vX.Y.Z tag (and
vllm-v* for the UDS tokenizer), so the floating :latest tag on ghcr.io
was never refreshed after the initial manual push. That left
ghcr.io/llm-d/llm-d-uds-tokenizer:latest pointing at a 29-day-old build
while v0.8.0 had been published 9 days earlier, causing version skew with
sibling components whose latest tag stayed current.

Dev / PR / pre-release / workflow_dispatch builds intentionally keep the
default (false) so the floating tag is never bumped by a non-release
artifact.

Signed-off-by: Kay Yan kay.yan@daocloud.io
feat(evictor): implement background empty directory cleanup process (feat(evictor): implement background empty directory cleanup process llm-d/llm-d-kv-cache#619)
- feat(evictor): background empty directory cleanup
Empty cache directories accumulate as files are evicted. This adds a
background folder-cleaner process (P(N+3), gated by ENABLE_DIR_CLEANUP)
that removes them.

How it works:
- The crawler detects empty rank/{hhh}/{hh} directories during its sweep
  and the deleter offers each freshly-emptied parent directory after a
  batch delete. Both feed a shared folder_queue.
- The folder cleaner pulls paths off the queue and removes them with
  os.rmdir, which is inherently safe: it is a no-op if a file has landed
  in the directory in the meantime.
Safety:
- queue_folder skips directories modified within DIR_CLEANUP_TTL_SECONDS
  (default 120s) so we don't race a writer that just created a bucket and
  is about to populate it. This is defense-in-depth on top of rmdir's
  empty-only semantics.
Config / Helm:
- New ENABLE_DIR_CLEANUP (default true) and DIR_CLEANUP_TTL_SECONDS
  (default 120) env vars, wired through config.py, the Helm values and
  Deployment template, and documented in CONFIGURATION.md.
Reporting:
- The folder cleaner reports folders_purged via a new folder_cleaner_stats
  channel surfaced in the aggregated log. The crawler's per-sweep counter
  is named empty_folders_queued to reflect that it counts directories
  handed to the cleaner, not directories it deleted itself. The deleter's
  progress/done result-queue protocol is left unchanged.
Signed-off-by: Miro mironikolov@google.com
- Fix pvc_evictor unit tests due to delete_file_batch signature update
Signed-off-by: Miro mironikolov@google.com
- test: merge empty directory cleanup unit tests from PR fix(pvc_evictor): clean up empty directories after batch file deletion llm-d/llm-d-kv-cache#625
Signed-off-by: Miro mironikolov@google.com
- fix(test): use moby container package instead of docker docker package to fix testcontainers-go build error
Signed-off-by: Miro mironikolov@google.com
- fix(ci): update golangci-lint configuration format
Signed-off-by: Miro mironikolov@google.com
- fix(kvblock): resolve ZRevRange deprecation and lll warnings
Signed-off-by: Miro mironikolov@google.com
- fix(test): resolve lll and unused gocritic lint warnings in uds_e2e_suite_test.go
Signed-off-by: Miro mironikolov@google.com
- Fix the linting errors
Signed-off-by: Miro mironikolov@google.com

Signed-off-by: Miro mironikolov@google.com
feat(kvevents): Track HMA group identity in kv cache index (feat(kvevents): Track HMA group identity in kv cache index llm-d/llm-d-kv-cache#627)
- add hma group identity to kvblock entries
Signed-off-by: Sage Ahrac sagiahrak@gmail.com
- learn hma groups from kv events
Signed-off-by: Sage Ahrac sagiahrak@gmail.com
- expose hma group catalog
Expose the learned group catalog so scorer follow-up work can use the event-derived metadata.

Co-authored-by: Kapil Jain kapiljain1989@gmail.com
Signed-off-by: Sage Ahrac sagiahrak@gmail.com
- fix redis pod entry encoding
Handle JSON encoding errors and store PodEntry directly for runtime Redis index state.

Signed-off-by: Sage Ahrac sagiahrak@gmail.com
- test namings
Signed-off-by: Sage Ahrac sagiahrak@gmail.com
- use only redis field
Signed-off-by: Sage Ahrac sagiahrak@gmail.com
- encode decode func namings for redis
Signed-off-by: Sage Ahrac sagiahrak@gmail.com
- remove redundant TestPodEntryString after redis keys change
Signed-off-by: Sage Ahrac sagiahrak@gmail.com
- remove noise from git diff
Signed-off-by: Sage Ahrac sagiahrak@gmail.com

Signed-off-by: Sage Ahrac sagiahrak@gmail.com
Co-authored-by: Kapil Jain kapiljain1989@gmail.com
feat: Add HMA support to FS connector (feat: Add HMA support to FS connector llm-d/llm-d-kv-cache#476)
feat: Add HMA support to the fs connector

Signed-off-by: Kfir Toledo kfir.toledo@ibm.com
PVC Evictor: add crawler tests, CI, and docs after layout changes (PVC Evictor: add crawler tests, CI, and docs after layout changes llm-d/llm-d-kv-cache#618)
- test(pvc_evictor): add crawler tests, CI, and docs after fix(pvc_evictor): walk current FileMapper layout in crawler llm-d/llm-d-kv-cache#611
  Add pytest for stream_cache_files_with_mapper, pvc-evictor CI workflow,
  dev Makefile/requirements, Dockerfile comments for feat: Add PVC evictor BlockRemoved events llm-d/llm-d-kv-cache#605 storage events,
  and docs for the flat fs_backend layout.
  Keeps llmd_fs_backend in the image for upcoming feat: Add PVC evictor BlockRemoved events llm-d/llm-d-kv-cache#605; crawler stays
  path-only per fix(pvc_evictor): walk current FileMapper layout in crawler llm-d/llm-d-kv-cache#611. Follow-up to pvc_evictor: support updated fs_backend layout (FileMapper v0.20+) llm-d/llm-d-kv-cache#601 / fix(pvc_evictor): walk current FileMapper layout in crawler llm-d/llm-d-kv-cache#611.
Signed-off-by: Guy Girmonsky guygir@gmail.com
- fix(pvc_evictor): ruff import order in test conftest
Signed-off-by: Guy Girmonsky guygir@gmail.com

Signed-off-by: Guy Girmonsky guygir@gmail.com
fix (fix(make): use linux platform in Docker build targets llm-d/llm-d-kv-cache#631)
Signed-off-by: Alex alex.tech.lab@outlook.com
fs_backend: sync block_size_factor with default offloaded block size
When block_size is absent from kv_connector_extra_config, the fs
backend defaults offloaded_block_size to 256 tokens, but vLLM's
OffloadingSpec base class only derives block_size_factor when
block_size is explicitly present, leaving it at 1. The scheduler then
emits one offload key per GPU block while the worker consumes one key
per file (gpu_blocks_per_file blocks), so on hybrid models (e.g. Gemma
with sliding-window + full-attention KV cache groups) the second
group's key slice lands inside the first group's keys and every
transfer fails with:
```
AssertionError: Expected group_idx=1 but key encodes 0
```
Set block_size_factor = gpu_blocks_per_file in
SharedStorageOffloadingSpec so the scheduler and worker always agree on
the per-file key granularity, regardless of whether block_size is
configured explicitly. On single-group models the old mismatch did not
assert but silently named files after the wrong block hashes, crippling
the offload hit rate.

Fixes [fs_backend] assertion with gemma4 when doing video inference llm-d/llm-d-kv-cache#656

Co-Authored-By: Claude Fable 5 noreply@anthropic.com
Signed-off-by: Miro mironikolov@google.com
fs_backend: accept explicit block_size on hybrid models
vLLM's OffloadingSpec base class derives block_size_factor from
extra_config["block_size"] and asserts that all KV cache groups share
one GPU block size to do so. Hybrid models like Gemma 4 have groups
with different block sizes, so explicitly configuring "block_size"
crashed at startup with:
```
AssertionError: If 'block_size' is specified in
kv_connector_extra_config, there must be at least one KV cache
group, and all groups must have the same block size.
```
The fs backend does not need that uniformity: it sizes files in
hash_block_size (GCD of group block sizes) granularity and already
derives block_size_factor itself. Hide "block_size" from the base
class during super().init() (restoring it afterwards) so the
uniformity assert is never reached and explicit and default block_size
configurations behave identically.

Fixes [fs_backend] Assertion with gemma4 when block size is explicitly defined llm-d/llm-d-kv-cache#657

Co-Authored-By: Claude Fable 5 noreply@anthropic.com
Signed-off-by: Miro mironikolov@google.com

…lm-d#613) The test stubs vllm.v1.kv_offload.base to load the real manager module in isolation. Commit 8cf550e added get_offload_block_hash to manager.py's imports, but the stub wasn't updated, so collection of tests/test_storage_events.py failed with "cannot import name 'get_offload_block_hash' from 'vllm.v1.kv_offload.base' (unknown location)". Add an identity stub so the import resolves and the existing assertions (which pass plain ints as keys) still hold. Signed-off-by: Kfir Toledo <kfir.toledo@ibm.com>

* pvc_evictor: walk current FileMapper layout in crawler After llm-d#585 collapsed the FileMapper layout to `<root>/<safe_model_name>_<sha256-12>_r<rank>/<hhh>/<hh>_g<group_idx>/*.bin`, the crawler still walked the pre-llm-d#585 deep tree (`block_size_*/tp_*_pp_size_*/rank_*/<dtype>/...`). Those paths no longer exist on disk, so the crawler discovered zero files and the evictor never freed any blocks. This rewrites `stream_cache_files_with_mapper` to: * find rank directories by their `_r<digits>` suffix anywhere under the cache root, instead of pattern-matching the deprecated deep tree; * iterate the first-level hex bucket ({hhh}) and apply the existing hex_modulo_range sharding; * yield .bin files from any second-level bucket underneath, leaving the `_g<group_idx>` encoding opaque so the walker doesn't need to understand kv-cache groups. The new walker no longer instantiates `FileMapper` (it only inspects on-disk names), so the `FILEMAPPER_AVAILABLE` import guard and the matching early-exit in `crawler_process` are removed. This also unblocks running the evictor container without vllm installed, since the walker was the only consumer of the `from llmd_fs_backend.file_mapper import FileMapper` import — which transitively required vllm. `parse_filemapper_params` is dropped along with its sole caller. Signed-off-by: Miro <mironikolov@google.com> * feat: parameterize first-level hex bucket directory length in pvc crawler Signed-off-by: Miro <mironikolov@google.com> --------- Signed-off-by: Miro <mironikolov@google.com>

Signed-off-by: Guy Girmonsky <guygir@gmail.com>

* parse hma kv event metadata Signed-off-by: Sage Ahrac <sagiahrak@gmail.com> Co-authored-by: Kapil Jain <16477749+kapiljain1989@users.noreply.github.com> * reduce noise Signed-off-by: Sage Ahrac <sagiahrak@gmail.com> --------- Signed-off-by: Sage Ahrac <sagiahrak@gmail.com> Co-authored-by: Kapil Jain <16477749+kapiljain1989@users.noreply.github.com>

fix CUDA version mismatch and dev headers symlink - Update default CUDA_TOOLKIT_PKG to cuda-toolkit-13-0 to match the CUDA 13.0 base image and prevent PyTorch compilation version mismatch. - Explicitly parse and update the standard /usr/local/cuda symlink after GKE package installation to resolve missing dev headers (cusparse.h) during compilation Signed-off-by: Saikat Roychowdhury <saikat.royc85@gmail.com>

* feat: Emit BlockRemoved events in PVC evictor Signed-off-by: Alberto Perdomo <aperdomo@redhat.com> * test: Add deleter and BlockRemoved tests Signed-off-by: Alberto Perdomo <aperdomo@redhat.com> * chore: Refactor after FileMapper changes Signed-off-by: Alberto Perdomo <aperdomo@redhat.com> * chore: Minor changes Signed-off-by: Alberto Perdomo <aperdomo@redhat.com> * chore: Fix lint issues Signed-off-by: Alberto Perdomo <aperdomo@redhat.com> * fix: Use only valid paths for events Signed-off-by: Alberto Perdomo <aperdomo@redhat.com> * fix: Emit events on shutdown Signed-off-by: Alberto Perdomo <aperdomo@redhat.com> --------- Signed-off-by: Alberto Perdomo <aperdomo@redhat.com>

* ci: Wire fs_backend Python tests into CI Signed-off-by: Alberto Perdomo <aperdomo@redhat.com> * chore: Clean up branch Signed-off-by: Alberto Perdomo <aperdomo@redhat.com> * fix: Use venv for llmd_fs_backend test Python deps Signed-off-by: Alberto Perdomo <aperdomo@redhat.com> * chore: Support PVC evictor events Signed-off-by: Alberto Perdomo <aperdomo@redhat.com> --------- Signed-off-by: Alberto Perdomo <aperdomo@redhat.com>

…pdates (llm-d#630) Bumps the go-dependencies group with 10 updates in the / directory: | Package | From | To | | --- | --- | --- | | [github.com/alicebob/miniredis/v2](https://github.com/alicebob/miniredis) | `2.35.0` | `2.38.0` | | [github.com/dgraph-io/ristretto/v2](https://github.com/dgraph-io/ristretto) | `2.3.0` | `2.4.0` | | [github.com/docker/docker](https://github.com/docker/docker) | `28.5.1+incompatible` | `28.5.2+incompatible` | | [github.com/fxamacker/cbor/v2](https://github.com/fxamacker/cbor) | `2.7.0` | `2.9.2` | | [github.com/prometheus/client_golang](https://github.com/prometheus/client_golang) | `1.22.0` | `1.23.2` | | [github.com/redis/go-redis/v9](https://github.com/redis/go-redis) | `9.7.3` | `9.20.0` | | [github.com/testcontainers/testcontainers-go](https://github.com/testcontainers/testcontainers-go) | `0.40.0` | `0.42.0` | | [go.opentelemetry.io/contrib/instrumentation/google.golang.org/grpc/otelgrpc](https://github.com/open-telemetry/opentelemetry-go-contrib) | `0.63.0` | `0.69.0` | | [go.opentelemetry.io/otel/exporters/otlp/otlptrace/otlptracegrpc](https://github.com/open-telemetry/opentelemetry-go) | `1.39.0` | `1.44.0` | | [go.uber.org/zap](https://github.com/uber-go/zap) | `1.27.0` | `1.28.0` | Updates `github.com/alicebob/miniredis/v2` from 2.35.0 to 2.38.0 - [Release notes](https://github.com/alicebob/miniredis/releases) - [Changelog](https://github.com/alicebob/miniredis/blob/master/CHANGELOG.md) - [Commits](alicebob/miniredis@v2.35.0...v2.38.0) Updates `github.com/dgraph-io/ristretto/v2` from 2.3.0 to 2.4.0 - [Release notes](https://github.com/dgraph-io/ristretto/releases) - [Changelog](https://github.com/dgraph-io/ristretto/blob/main/CHANGELOG.md) - [Commits](dgraph-io/ristretto@v2.3.0...v2.4.0) Updates `github.com/docker/docker` from 28.5.1+incompatible to 28.5.2+incompatible - [Release notes](https://github.com/docker/docker/releases) - [Commits](moby/moby@v28.5.1...v28.5.2) Updates `github.com/fxamacker/cbor/v2` from 2.7.0 to 2.9.2 - [Release notes](https://github.com/fxamacker/cbor/releases) - [Commits](fxamacker/cbor@v2.7.0...v2.9.2) Updates `github.com/prometheus/client_golang` from 1.22.0 to 1.23.2 - [Release notes](https://github.com/prometheus/client_golang/releases) - [Changelog](https://github.com/prometheus/client_golang/blob/main/CHANGELOG.md) - [Commits](prometheus/client_golang@v1.22.0...v1.23.2) Updates `github.com/prometheus/client_model` from 0.6.1 to 0.6.2 - [Release notes](https://github.com/prometheus/client_model/releases) - [Commits](prometheus/client_model@v0.6.1...v0.6.2) Updates `github.com/redis/go-redis/v9` from 9.7.3 to 9.20.0 - [Release notes](https://github.com/redis/go-redis/releases) - [Changelog](https://github.com/redis/go-redis/blob/master/RELEASE-NOTES.md) - [Commits](redis/go-redis@v9.7.3...v9.20.0) Updates `github.com/testcontainers/testcontainers-go` from 0.40.0 to 0.42.0 - [Release notes](https://github.com/testcontainers/testcontainers-go/releases) - [Commits](testcontainers/testcontainers-go@v0.40.0...v0.42.0) Updates `go.opentelemetry.io/contrib/instrumentation/google.golang.org/grpc/otelgrpc` from 0.63.0 to 0.69.0 - [Release notes](https://github.com/open-telemetry/opentelemetry-go-contrib/releases) - [Changelog](https://github.com/open-telemetry/opentelemetry-go-contrib/blob/main/CHANGELOG.md) - [Commits](open-telemetry/opentelemetry-go-contrib@zpages/v0.63.0...zpages/v0.69.0) Updates `go.opentelemetry.io/otel` from 1.43.0 to 1.44.0 - [Release notes](https://github.com/open-telemetry/opentelemetry-go/releases) - [Changelog](https://github.com/open-telemetry/opentelemetry-go/blob/main/CHANGELOG.md) - [Commits](open-telemetry/opentelemetry-go@v1.43.0...v1.44.0) Updates `go.opentelemetry.io/otel/exporters/otlp/otlptrace/otlptracegrpc` from 1.39.0 to 1.44.0 - [Release notes](https://github.com/open-telemetry/opentelemetry-go/releases) - [Changelog](https://github.com/open-telemetry/opentelemetry-go/blob/main/CHANGELOG.md) - [Commits](open-telemetry/opentelemetry-go@v1.39.0...v1.44.0) Updates `go.opentelemetry.io/otel/sdk` from 1.43.0 to 1.44.0 - [Release notes](https://github.com/open-telemetry/opentelemetry-go/releases) - [Changelog](https://github.com/open-telemetry/opentelemetry-go/blob/main/CHANGELOG.md) - [Commits](open-telemetry/opentelemetry-go@v1.43.0...v1.44.0) Updates `go.opentelemetry.io/otel/trace` from 1.43.0 to 1.44.0 - [Release notes](https://github.com/open-telemetry/opentelemetry-go/releases) - [Changelog](https://github.com/open-telemetry/opentelemetry-go/blob/main/CHANGELOG.md) - [Commits](open-telemetry/opentelemetry-go@v1.43.0...v1.44.0) Updates `go.uber.org/zap` from 1.27.0 to 1.28.0 - [Release notes](https://github.com/uber-go/zap/releases) - [Changelog](https://github.com/uber-go/zap/blob/master/CHANGELOG.md) - [Commits](uber-go/zap@v1.27.0...v1.28.0) Updates `google.golang.org/grpc` from 1.77.0 to 1.81.1 - [Release notes](https://github.com/grpc/grpc-go/releases) - [Commits](grpc/grpc-go@v1.77.0...v1.81.1) Updates `google.golang.org/protobuf` from 1.36.10 to 1.36.11 --- updated-dependencies: - dependency-name: github.com/alicebob/miniredis/v2 dependency-version: 2.38.0 dependency-type: direct:production update-type: version-update:semver-minor dependency-group: go-dependencies - dependency-name: github.com/dgraph-io/ristretto/v2 dependency-version: 2.4.0 dependency-type: direct:production update-type: version-update:semver-minor dependency-group: go-dependencies - dependency-name: github.com/docker/docker dependency-version: 28.5.2+incompatible dependency-type: direct:production update-type: version-update:semver-patch dependency-group: go-dependencies - dependency-name: github.com/fxamacker/cbor/v2 dependency-version: 2.9.2 dependency-type: direct:production update-type: version-update:semver-minor dependency-group: go-dependencies - dependency-name: github.com/prometheus/client_golang dependency-version: 1.23.2 dependency-type: direct:production update-type: version-update:semver-minor dependency-group: go-dependencies - dependency-name: github.com/prometheus/client_model dependency-version: 0.6.2 dependency-type: direct:production update-type: version-update:semver-patch dependency-group: go-dependencies - dependency-name: github.com/redis/go-redis/v9 dependency-version: 9.20.0 dependency-type: direct:production update-type: version-update:semver-minor dependency-group: go-dependencies - dependency-name: github.com/testcontainers/testcontainers-go dependency-version: 0.42.0 dependency-type: direct:production update-type: version-update:semver-minor dependency-group: go-dependencies - dependency-name: go.opentelemetry.io/contrib/instrumentation/google.golang.org/grpc/otelgrpc dependency-version: 0.69.0 dependency-type: direct:production update-type: version-update:semver-minor dependency-group: go-dependencies - dependency-name: go.opentelemetry.io/otel dependency-version: 1.44.0 dependency-type: direct:production update-type: version-update:semver-minor dependency-group: go-dependencies - dependency-name: go.opentelemetry.io/otel/exporters/otlp/otlptrace/otlptracegrpc dependency-version: 1.44.0 dependency-type: direct:production update-type: version-update:semver-minor dependency-group: go-dependencies - dependency-name: go.opentelemetry.io/otel/sdk dependency-version: 1.44.0 dependency-type: direct:production update-type: version-update:semver-minor dependency-group: go-dependencies - dependency-name: go.opentelemetry.io/otel/trace dependency-version: 1.44.0 dependency-type: direct:production update-type: version-update:semver-minor dependency-group: go-dependencies - dependency-name: go.uber.org/zap dependency-version: 1.28.0 dependency-type: direct:production update-type: version-update:semver-minor dependency-group: go-dependencies - dependency-name: google.golang.org/grpc dependency-version: 1.81.1 dependency-type: direct:production update-type: version-update:semver-minor dependency-group: go-dependencies - dependency-name: google.golang.org/protobuf dependency-version: 1.36.11 dependency-type: direct:production update-type: version-update:semver-patch dependency-group: go-dependencies ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

…lm-d#589) The 16 MB floor on staging buffers caused every per-file write to pad the buffer via cudaHostAlloc, then write the full padded size to disk. For small-block models (e.g. Llama 3.1 8B where block size is ~7 MB) this more than doubles the on-disk footprint (7 MB -> 16 MB per file) and degrades TTFT from 2.8 s to 6.5 s at 40k-token prefix length. Remove the constant so allocate_staging_buffer uses exactly the size that calc_staging_bytes computes. Fixes llm-d#454 Signed-off-by: Jonathan Wrede <wrede.jonathan00@gmail.com>

…-d#634) clang-format (v21) flags the double blank line left after llm-d#589 removed MIN_STAGING_BUFFER_SIZE. Fixes the repo-wide pre-commit lint gate that currently fails on all PRs. Signed-off-by: Kfir Toledo <kfir.toledo@ibm.com>

The vllm 0.22 KV-offload Python API (base, worker, OffloadPromMetrics) is unchanged from 0.21, so no import/signature changes are needed. The one functional change is in spec.py: adopt the OffloadingSpec base class's $self.hash_block_size$ (resolved via resolve_kv_cache_block_sizes) instead of $vllm_config.cache_config.block_size$ when computing gpu_blocks_per_file. The two are equal for standard single-group models, but cache_config.block_size can be larger on hybrid models, so the base value is the correct hash granularity. Single-group only (no HMA). Bump pins: vllm 0.21.0 -> 0.22.0, package version 0.21 -> 0.22. README, Dockerfile.dev base image, and vllm-storage.yaml deployment example updated accordingly. Signed-off-by: Kfir Toledo <kfir.toledo@ibm.com>

…lm-d#607) * csrc: batch KV block copies via cudaMemcpyBatchAsync Submit all per-(block, layer) copies in one driver call instead of N cudaMemcpyAsync calls. Enabled by default; toggle off with USE_BATCH_MEMCPY_READ / USE_BATCH_MEMCPY_WRITE=0. Requires CUDA 12.8+. Speeds up KV-cache offload writes/reads when per-layer DMA sizes are small enough that driver dispatch dominates. Signed-off-by: Kfir Toledo <kfir.toledo@ibm.com> * csrc: fall back to per-call DMA on CUDA < 12.8 cudaMemcpyBatchAsync was introduced in CUDA 12.8 — guard the batch path with #if CUDA_VERSION >= 12080 and route to the per-call cudaMemcpyAsync loop below that. Default USE_BATCH_MEMCPY_* off on older toolchains so the env knob still makes sense. Also drop thread_local on the attrs/attrs_idx inputs (never mutated, no per-thread duplication needed) and move the copy_blocks dispatcher below the helpers it dispatches to. Signed-off-by: Kfir Toledo <kfir.toledo@ibm.com> --------- Signed-off-by: Kfir Toledo <kfir.toledo@ibm.com>

…lm-d#632) llmd-fs-connector==0.22 (llm-d v0.8 / vLLM v0.22) is the final release of the standalone llm-d FS connector. The filesystem offloading logic is now upstreamed into vLLM as the FS tier of the multi-tier offloading connector (TieringOffloadingSpec); all new features and support continue there. Add an [!IMPORTANT] banner to the connector README and a short note in the root README's Connectors & Utilities list, linking the vLLM KV offloading guide (vllm-project/vllm#44415). Signed-off-by: Kfir Toledo <kfir.toledo@ibm.com>

Signed-off-by: Dong Ma <winterma.dong@gmail.com>

Add a `latest` input (default `false`) to the shared docker-build-and-push action and wire it up from `ci-release.yaml` and `ci-release-uds-tokenizer.yaml`. The callers set it to `true` only when the triggering event is a non-prerelease GitHub Release. Previously the release workflows only pushed the immutable `vX.Y.Z` tag (and `vllm-v*` for the UDS tokenizer), so the floating `:latest` tag on ghcr.io was never refreshed after the initial manual push. That left `ghcr.io/llm-d/llm-d-uds-tokenizer:latest` pointing at a 29-day-old build while `v0.8.0` had been published 9 days earlier, causing version skew with sibling components whose `latest` tag stayed current. Dev / PR / pre-release / workflow_dispatch builds intentionally keep the default (`false`) so the floating tag is never bumped by a non-release artifact. Signed-off-by: Kay Yan <kay.yan@daocloud.io>

…lm-d#619) * feat(evictor): background empty directory cleanup Empty cache directories accumulate as files are evicted. This adds a background folder-cleaner process (P(N+3), gated by ENABLE_DIR_CLEANUP) that removes them. How it works: - The crawler detects empty rank/{hhh}/{hh} directories during its sweep and the deleter offers each freshly-emptied parent directory after a batch delete. Both feed a shared folder_queue. - The folder cleaner pulls paths off the queue and removes them with os.rmdir, which is inherently safe: it is a no-op if a file has landed in the directory in the meantime. Safety: - queue_folder skips directories modified within DIR_CLEANUP_TTL_SECONDS (default 120s) so we don't race a writer that just created a bucket and is about to populate it. This is defense-in-depth on top of rmdir's empty-only semantics. Config / Helm: - New ENABLE_DIR_CLEANUP (default true) and DIR_CLEANUP_TTL_SECONDS (default 120) env vars, wired through config.py, the Helm values and Deployment template, and documented in CONFIGURATION.md. Reporting: - The folder cleaner reports folders_purged via a new folder_cleaner_stats channel surfaced in the aggregated log. The crawler's per-sweep counter is named empty_folders_queued to reflect that it counts directories handed to the cleaner, not directories it deleted itself. The deleter's progress/done result-queue protocol is left unchanged. Signed-off-by: Miro <mironikolov@google.com> * Fix pvc_evictor unit tests due to delete_file_batch signature update Signed-off-by: Miro <mironikolov@google.com> * test: merge empty directory cleanup unit tests from PR llm-d#625 Signed-off-by: Miro <mironikolov@google.com> * fix(test): use moby container package instead of docker docker package to fix testcontainers-go build error Signed-off-by: Miro <mironikolov@google.com> * fix(ci): update golangci-lint configuration format Signed-off-by: Miro <mironikolov@google.com> * fix(kvblock): resolve ZRevRange deprecation and lll warnings Signed-off-by: Miro <mironikolov@google.com> * fix(test): resolve lll and unused gocritic lint warnings in uds_e2e_suite_test.go Signed-off-by: Miro <mironikolov@google.com> * Fix the linting errors Signed-off-by: Miro <mironikolov@google.com> --------- Signed-off-by: Miro <mironikolov@google.com>

* add hma group identity to kvblock entries Signed-off-by: Sage Ahrac <sagiahrak@gmail.com> * learn hma groups from kv events Signed-off-by: Sage Ahrac <sagiahrak@gmail.com> * expose hma group catalog Expose the learned group catalog so scorer follow-up work can use the event-derived metadata. Co-authored-by: Kapil Jain <kapiljain1989@gmail.com> Signed-off-by: Sage Ahrac <sagiahrak@gmail.com> * fix redis pod entry encoding Handle JSON encoding errors and store PodEntry directly for runtime Redis index state. Signed-off-by: Sage Ahrac <sagiahrak@gmail.com> * test namings Signed-off-by: Sage Ahrac <sagiahrak@gmail.com> * use only redis field Signed-off-by: Sage Ahrac <sagiahrak@gmail.com> * encode decode func namings for redis Signed-off-by: Sage Ahrac <sagiahrak@gmail.com> * remove redundant TestPodEntryString after redis keys change Signed-off-by: Sage Ahrac <sagiahrak@gmail.com> * remove noise from git diff Signed-off-by: Sage Ahrac <sagiahrak@gmail.com> --------- Signed-off-by: Sage Ahrac <sagiahrak@gmail.com> Co-authored-by: Kapil Jain <kapiljain1989@gmail.com>

feat: Add HMA support to the fs connector --------- Signed-off-by: Kfir Toledo <kfir.toledo@ibm.com>

…m-d#618) * test(pvc_evictor): add crawler tests, CI, and docs after llm-d#611 Add pytest for stream_cache_files_with_mapper, pvc-evictor CI workflow, dev Makefile/requirements, Dockerfile comments for llm-d#605 storage events, and docs for the flat fs_backend layout. Keeps llmd_fs_backend in the image for upcoming llm-d#605; crawler stays path-only per llm-d#611. Follow-up to llm-d#601 / llm-d#611. Signed-off-by: Guy Girmonsky <guygir@gmail.com> * fix(pvc_evictor): ruff import order in test conftest Signed-off-by: Guy Girmonsky <guygir@gmail.com> --------- Signed-off-by: Guy Girmonsky <guygir@gmail.com>

Signed-off-by: Alex <alex.tech.lab@outlook.com>

When block_size is absent from kv_connector_extra_config, the fs backend defaults offloaded_block_size to 256 tokens, but vLLM's OffloadingSpec base class only derives block_size_factor when block_size is explicitly present, leaving it at 1. The scheduler then emits one offload key per GPU block while the worker consumes one key per file (gpu_blocks_per_file blocks), so on hybrid models (e.g. Gemma with sliding-window + full-attention KV cache groups) the second group's key slice lands inside the first group's keys and every transfer fails with: AssertionError: Expected group_idx=1 but key encodes 0 Set block_size_factor = gpu_blocks_per_file in SharedStorageOffloadingSpec so the scheduler and worker always agree on the per-file key granularity, regardless of whether block_size is configured explicitly. On single-group models the old mismatch did not assert but silently named files after the wrong block hashes, crippling the offload hit rate. Fixes llm-d#656 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Signed-off-by: Miro <mironikolov@google.com>

vLLM's OffloadingSpec base class derives block_size_factor from extra_config["block_size"] and asserts that all KV cache groups share one GPU block size to do so. Hybrid models like Gemma 4 have groups with different block sizes, so explicitly configuring "block_size" crashed at startup with: AssertionError: If 'block_size' is specified in kv_connector_extra_config, there must be at least one KV cache group, and all groups must have the same block size. The fs backend does not need that uniformity: it sizes files in hash_block_size (GCD of group block sizes) granularity and already derives block_size_factor itself. Hide "block_size" from the base class during super().__init__() (restoring it afterwards) so the uniformity assert is never reached and explicit and default block_size configurations behave identically. Fixes llm-d#657 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Signed-off-by: Miro <mironikolov@google.com>

github-actions · 2026-06-10T17:12:28Z

Unsigned commits detected! Please sign your commits.

For instructions on how to set up GPG/SSH signing and verify your commits, please see GitHub Documentation.

kfirtoledo and others added 22 commits May 27, 2026 11:56

fix(pvc_evictor): remove unused re import after llm-d#611 (llm-d#615)

5e9d31e

Signed-off-by: Guy Girmonsky <guygir@gmail.com>

feat(kvblock): trace index add and evict (llm-d#637)

7c285bf

Signed-off-by: Dong Ma <winterma.dong@gmail.com>

feat: Add HMA support to FS connector (llm-d#476)

9f03606

feat: Add HMA support to the fs connector --------- Signed-off-by: Kfir Toledo <kfir.toledo@ibm.com>

fix (llm-d#631)

1351200

Signed-off-by: Alex <alex.tech.lab@outlook.com>

github-actions Bot added the size/XXL label Jun 10, 2026

miroslavln closed this Jun 10, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix/issue 656 default block size factor#1

fix/issue 656 default block size factor#1
miroslavln wants to merge 22 commits into
mainfrom
fix/issue-656-default-block-size-factor

miroslavln commented Jun 10, 2026

Uh oh!

github-actions Bot commented Jun 10, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

10 participants

Conversation

miroslavln commented Jun 10, 2026

Uh oh!

github-actions Bot commented Jun 10, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

10 participants