feat: Transition to FullyContiguous Host and Disk layouts #3090
Conversation
WalkthroughIntroduces per-edge layout types (device/host/disk) across workers and Python bindings; replaces a boolean layout flag with explicit LayoutType. Adds auto-detection for device layout during KV-cache registration. Implements inter-layout CUDA copy paths (FullyContiguous ↔ LayerSeparate). Expands layout utilities with verification APIs and worker-side verifiers. Adds extensive tests, minor Dockerfile newline. Changes
Sequence Diagram(s)sequenceDiagram
autonumber
actor Py as Python Caller
participant PyW as PyKvConnectorWorker
participant Conn as KvConnectorWorker
participant BM as KvbmWorkerConfig Builder
participant L as LayoutType
Py->>PyW: register_kv_caches(..., device_layout_type?, host_layout_type?, disk_layout_type?)
Note right of PyW: Map PyLayoutType -> LayoutType (optional)
PyW->>Conn: register_kv_caches(..., dev?, host?, disk?)
alt device_layout_type provided
Conn->>BM: device_layout_type(dev)
else auto-detect
Conn->>Conn: capture first tensor shape
Conn->>L: layer_separate_auto(shape, num_device_blocks) or default
Conn->>BM: device_layout_type(detected or default)
end
alt host_layout_type provided
Conn->>BM: host_layout_type(host)
else
Conn->>BM: host_layout_type(FullyContiguous)
end
alt disk_layout_type provided
Conn->>BM: disk_layout_type(disk)
else
Conn->>BM: disk_layout_type(FullyContiguous)
end
BM-->>Py: configured worker ready
sequenceDiagram
autonumber
participant Src as Source Layout
participant Cpy as copy_block (CUDA)
participant Dst as Destination Layout
Src->>Cpy: request copy (S_layout, D_layout)
alt FC -> FC
Cpy->>Dst: single memcpy
else FC -> LayerSeparate
loop layers x outer
Cpy->>Dst: scatter copy chunks into per-layer regions
end
Note right of Cpy: assert src_offset == src_size
else LayerSeparate -> FC
loop layers x outer
Cpy->>Dst: gather copy chunks into contiguous buffer
end
Note right of Cpy: assert dst_offset == dst_size
end
Estimated code review effort🎯 4 (Complex) | ⏱️ ~60 minutes Possibly related PRs
Poem
Pre-merge checks❌ Failed checks (1 warning)
✅ Passed checks (2 passed)
📜 Recent review detailsConfiguration used: Path: .coderabbit.yaml Review profile: CHILL Plan: Pro 📒 Files selected for processing (5)
✅ Files skipped from review due to trivial changes (1)
🔇 Additional comments (14)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 17
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (7)
lib/llm/src/block_manager/block/data/local.rs (1)
123-130: Guard against overflow when computing fully-contiguous block sizeMultiplying mr.size() by num_layers() and num_outer_dims() can overflow usize and won’t be caught. Use checked_mul and surface a BlockError instead of silently wrapping.
Apply this diff:
- let size = mr.size() * self.num_layers() * self.num_outer_dims(); + let per_region = mr.size(); + let size = per_region + .checked_mul(self.num_layers()) + .and_then(|v| v.checked_mul(self.num_outer_dims())) + .ok_or_else(|| BlockError::InvalidState("block size overflow".to_string()))?;lib/bindings/python/rust/llm/block_manager/vllm/connector/worker.rs (2)
353-356: Don’t panic in runtime hot path when a slot is missingPanic will take down the process. Log and continue (or convert to an error return higher up) to keep the worker resilient.
Apply this diff:
- } else { - // made this condition more strict slot existence checks were added as a prerequesite - // to be added to the maybe_finished_offloading set. - panic!("request slot missing for {request_id}; however, it was present when added to the maybe finished offloading set"); - } + } else { + // stricter existence checks are in place, but avoid crashing the process + tracing::error!( + request_id, + "request slot missing; was present when added to maybe_finished_offloading" + ); + continue; + }
382-383: Avoid panic in onboarding check as wellMirror the non-panicking handling here.
Apply this diff:
- } else { - panic!("request slot missing for {request_id}; however, it was present when added to the maybe finished onboarding set"); - } + } else { + tracing::error!( + request_id, + "request slot missing; was present when added to maybe_finished_onboarding" + ); + continue; + }lib/bindings/python/rust/llm/block_manager/vllm/connector/trtllm_worker.rs (2)
311-323: Avoid panic on offloading slot missSame resilience concern as the vLLM connector; don’t panic in production paths.
Apply this diff:
- } else { - // made this condition more strict slot existence checks were added as a prerequesite - // to be added to the maybe_finished_offloading set. - panic!("request slot missing for {request_id}; however, it was present when added to the maybe finished offloading set"); - } + } else { + tracing::error!( + request_id, + "request slot missing; was present when added to maybe_finished_offloading" + ); + continue; + }
349-350: Avoid panic on onboarding slot missMirror non-panicking handling here.
Apply this diff:
- } else { - panic!("request slot missing for {request_id}; however, it was present when added to the maybe finished onboarding set"); - } + } else { + tracing::error!( + request_id, + "request slot missing; was present when added to maybe_finished_onboarding" + ); + continue; + }lib/llm/src/block_manager/distributed/transfer.rs (1)
104-115: Bounds-check indices from the request before indexing into pools.
source_pool_list[idx]andtarget_pool_list[idx]may panic if a request carries out-of-range indices. Validate and error out.+ // Validate indices before indexing + let src_len = source_pool_list.len(); + let tgt_len = target_pool_list.len(); + for (from, to) in request.blocks() { + if *from >= src_len || *to >= tgt_len { + return Err(anyhow::anyhow!( + "Transfer index out of bounds: from={}, to={}, src_len={}, tgt_len={}", + from, to, src_len, tgt_len + )); + } + }lib/llm/src/block_manager/distributed/worker.rs (1)
581-589: Avoid unwraps in production path (stream creation).Propagate errors instead of panicking:
- let transfer_context = Arc::new(TransferContext::new( - Arc::new(Some(agent)), - DeviceAllocator::new(config.device_id) - .unwrap() - .ctx() - .new_stream() - .unwrap(), + let stream = DeviceAllocator::new(config.device_id)? + .ctx() + .new_stream()?; + let transfer_context = Arc::new(TransferContext::new( + Arc::new(Some(agent)), + stream, Handle::current(), ));
🧹 Nitpick comments (17)
lib/llm/src/block_manager/block/transfer/nixl.rs (1)
55-56: Early return is fine, but keep assertions in debugThe explicit Ok(()) is fine; ensure any preconditions remain covered by debug assertions elsewhere.
lib/llm/src/block_manager/block/transfer/cuda.rs (3)
246-253: Incorrect debug message for overlap checkH2D path message says “D2D copy”; adjust to avoid confusion or remove overlap check entirely for H2D.
Apply this diff:
- "Source and destination device memory regions must not overlap for D2D copy" + "Source and destination memory regions must not overlap"
270-276: Same nit: overlap message mentions D2D in D2H pathMirror the wording fix here.
Apply this diff:
- "Source and destination device memory regions must not overlap for D2D copy" + "Source and destination memory regions must not overlap"
309-739: Large CUDA tests: keep but ensure they don’t run by defaultThe tests are cfg(test, feature = "testing-cuda"); that’s good. Consider marking the perf test with #[ignore] to avoid accidental long runs in CI tiers without the feature.
lib/llm/src/block_manager/distributed/transfer.rs (1)
21-23: Remove unused import.
layout::BlockLayoutConfigisn’t used here. Drop it to keep warnings clean.lib/llm/src/block_manager/distributed/worker.rs (1)
217-219: Use checked math for bytes_per_block to avoid usize overflow.Guard against large shapes:
- let bytes_per_block = - num_layers * outer_dim * config.page_size * inner_dim * config.dtype_width_bytes; + let bytes_per_block = num_layers + .checked_mul(outer_dim).and_then(|v| v.checked_mul(config.page_size)) + .and_then(|v| v.checked_mul(inner_dim)).and_then(|v| v.checked_mul(config.dtype_width_bytes)) + .ok_or_else(|| anyhow::anyhow!("bytes_per_block overflow"))?;lib/bindings/python/rust/llm/block_manager/distributed/worker.rs (1)
171-181: Minor: simplify Option mapping and let rustfmt reflow.Use
map(Into::into)to reduce verbosity.- .device_layout_type(device_layout_type.map(|py_layout| py_layout.into()).unwrap_or(LayoutType::FullyContiguous)) + .device_layout_type(device_layout_type.map(Into::into).unwrap_or(LayoutType::FullyContiguous)) .host_layout_type( - host_layout_type - .map(|py_layout| py_layout.into()) - .unwrap_or(LayoutType::FullyContiguous) + host_layout_type.map(Into::into).unwrap_or(LayoutType::FullyContiguous) ) .disk_layout_type( - disk_layout_type - .map(|py_layout| py_layout.into()) - .unwrap_or(LayoutType::FullyContiguous) + disk_layout_type.map(Into::into).unwrap_or(LayoutType::FullyContiguous) )lib/llm/src/block_manager/layout.rs (3)
338-345: Confirm semantics: memory_region_size now equals one page.Setting
memory_region_size = outer_dim_stride_in_bytesalignsLocalMemoryRegion::sizeto a single [page_size × inner_dim × dtype] region. This is consistent with region addressing used elsewhere. Ensure all callers (e.g., transfer validators) expect per-page size, not per-layer size.
861-943: LayerSeparate verification and alignment checks are solid.Good coverage: region math, alignment, and capacity checks per-layer. Minor: consider early-returning on first mismatch to speed failing tests.
1621-1904: Comprehensive tests are valuable; minor nits.The no-overlap/alignment/stride tests add strong regression guarantees. Consider parameterizing constants to reduce duplication across FC/LS tests.
lib/llm/src/block_manager/layout/utils.rs (7)
19-25: align_up: guard against overflow and enforce precondition in release builds
value + alignment - 1can overflow;debug_assert!disappears in release. Use a checked add and a hard assert (or return Result) to make it safe.#[inline(always)] pub fn align_up(value: usize, alignment: usize) -> usize { - debug_assert!(alignment.is_power_of_two(), "Alignment must be a power of 2"); - (value + alignment - 1) & !(alignment - 1) + assert!(alignment.is_power_of_two() && alignment > 0, "Alignment must be a power of 2"); + let add = alignment - 1; + let sum = value.checked_add(add).expect("overflow in align_up"); + sum & !add }
27-36: Don’t leak validator::ValidationError from low-level utilsThis module otherwise uses LayoutError. Exposing
validator::ValidationErrorhere couples layers unnecessarily.-/// Validates that the given value is a power of 2. -pub fn validate_power_of_2(alignment: usize) -> Result<(), validator::ValidationError> { - if alignment.is_power_of_two() { - Ok(()) - } else { - Err(validator::ValidationError::new( - "Alignment must be a power of 2", - )) - } -} +/// Validates that the given value is a power of 2. +pub fn validate_power_of_2(alignment: usize) -> Result<(), LayoutError> { + if alignment.is_power_of_two() { + Ok(()) + } else { + Err(LayoutError::InvalidConfig("Alignment must be a power of 2".into())) + } +}If external callers truly need
validator::ValidationError, add a thin adapter outside this module.
167-190: Pre‑allocate results to avoid reallocationsWe know total regions up front.
- let mut results = Vec::new(); + let mut results = Vec::with_capacity(layout.num_blocks() * layout.num_layers() * layout.outer_dim());
266-289: Stats should ignore “unknown” address state if you adopt tri‑state addr_matchesIf you switch to
Option<bool>, update counters to only includeSome(false)as mismatches andSome(true)in successes.- if !result.addr_matches { + if matches!(result.addr_matches, Some(false)) { self.stats.addr_mismatches += 1; } ... - if result.addr_matches && result.size_matches { + if result.size_matches && matches!(result.addr_matches, Some(true) | None) { self.stats.successful_verifications += 1; }
297-300: Critical mismatches should include checksum when verify_data=trueCurrently only size mismatches are considered critical. If data verification is opted in, checksum mismatches should elevate to critical as well.
- pub fn has_critical_mismatches(&self) -> bool { - // Only check size mismatches since address verification is layout-specific - self.stats.size_mismatches > 0 - } + pub fn has_critical_mismatches(&self) -> bool { + self.stats.size_mismatches > 0 || self.stats.checksum_mismatches > 0 + }
303-328: Avoid intermediate allocations in report buildingUse string literals with push_str.
- report.push_str(&format!("Layout Verification Report\n")); - report.push_str(&format!("========================\n")); + report.push_str("Layout Verification Report\n"); + report.push_str("========================\n");(Apply similarly to the other lines.)
331-396: Compatibility should also check alignment; consider returning a rich report
- Missing alignment check can green‑light incompatible GDS paths.
- Returning
boolhides the reason; a structured report improves UX.Add alignment check:
if source_config.dtype_width_bytes != dest_config.dtype_width_bytes { tracing::error!("Data type width mismatch: {} vs {}", source_config.dtype_width_bytes, dest_config.dtype_width_bytes); return Ok(false); } + + if source_config.alignment != dest_config.alignment { + tracing::error!("Alignment mismatch: {} vs {}", + source_config.alignment, dest_config.alignment); + return Ok(false); + }Follow‑up (non‑blocking): return
Result<(), Incompatibility>collecting all mismatches instead of the first-fail boolean.
📜 Review details
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (11)
lib/bindings/python/rust/llm/block_manager/distributed/worker.rs(4 hunks)lib/bindings/python/rust/llm/block_manager/vllm/connector/trtllm_worker.rs(2 hunks)lib/bindings/python/rust/llm/block_manager/vllm/connector/worker.rs(2 hunks)lib/llm/src/block_manager/block/data/local.rs(2 hunks)lib/llm/src/block_manager/block/transfer/cuda.rs(2 hunks)lib/llm/src/block_manager/block/transfer/nixl.rs(2 hunks)lib/llm/src/block_manager/distributed/transfer.rs(3 hunks)lib/llm/src/block_manager/distributed/worker.rs(4 hunks)lib/llm/src/block_manager/layout.rs(6 hunks)lib/llm/src/block_manager/layout/utils.rs(3 hunks)lib/llm/src/block_manager/offload.rs(1 hunks)
🧰 Additional context used
🧬 Code graph analysis (8)
lib/llm/src/block_manager/offload.rs (4)
lib/llm/src/block_manager/layout/utils.rs (1)
verify_layout_compatibility(332-395)lib/llm/src/block_manager/storage.rs (8)
storage_type(171-171)storage_type(389-391)storage_type(469-471)storage_type(510-512)size(177-177)size(397-399)size(477-479)size(518-520)lib/llm/src/block_manager/block/data.rs (4)
storage_type(26-26)layer_view(49-58)num_layers(32-32)block_view(73-78)lib/llm/src/block_manager/storage/disk.rs (3)
storage_type(123-125)fd(90-92)size(131-133)
lib/llm/src/block_manager/block/transfer/nixl.rs (3)
lib/llm/src/block_manager/layout.rs (1)
num_layers(235-237)lib/llm/src/block_manager/block.rs (2)
num_layers(377-379)size(1049-1051)lib/llm/src/block_manager/storage/nixl.rs (7)
size(114-116)size(281-283)size(299-301)size(323-325)size(348-350)size(373-375)size(412-414)
lib/llm/src/block_manager/block/transfer/cuda.rs (4)
lib/llm/src/block_manager/block/data/local.rs (1)
num_layers(74-76)lib/llm/src/block_manager/layout.rs (9)
num_layers(235-237)storage(194-194)storage(505-507)storage(842-844)config(206-206)config(519-521)config(804-806)allocate(470-495)allocate(764-796)lib/llm/src/block_manager/block/data.rs (1)
num_layers(32-32)lib/llm/src/block_manager/storage/cuda.rs (9)
size(226-228)size(388-390)default(284-288)default(445-449)ctx(460-462)allocate(301-303)allocate(466-468)addr(222-224)addr(384-386)
lib/bindings/python/rust/llm/block_manager/vllm/connector/trtllm_worker.rs (1)
lib/llm/src/block_manager/block.rs (1)
layout(1242-1244)
lib/llm/src/block_manager/distributed/transfer.rs (3)
lib/llm/src/block_manager/block.rs (3)
layout(1242-1244)num_layers(377-379)size(1049-1051)lib/llm/src/block_manager/block/data/local.rs (1)
num_layers(74-76)lib/llm/src/block_manager/block/data.rs (1)
num_layers(32-32)
lib/llm/src/block_manager/distributed/worker.rs (1)
lib/llm/src/block_manager/layout.rs (11)
builder(296-298)layout_type(191-191)layout_type(501-503)layout_type(836-840)num_layers(235-237)outer_dim(243-245)inner_dim(253-255)config(206-206)config(519-521)config(804-806)page_size(248-250)
lib/llm/src/block_manager/layout.rs (2)
lib/llm/src/block_manager/storage.rs (8)
addr(174-174)addr(393-395)addr(473-475)addr(514-516)size(177-177)size(397-399)size(477-479)size(518-520)lib/llm/src/block_manager/storage/nixl.rs (8)
addr(277-279)size(114-116)size(281-283)size(299-301)size(323-325)size(348-350)size(373-375)size(412-414)
lib/llm/src/block_manager/layout/utils.rs (2)
lib/llm/src/block_manager/block.rs (1)
layout(1242-1244)lib/llm/src/block_manager/layout.rs (1)
outer_dim(243-245)
🪛 GitHub Actions: Rust pre-merge checks
lib/bindings/python/rust/llm/block_manager/distributed/worker.rs
[error] 10-10: Rustfmt: formatting issue detected by cargo fmt -- --check. Run 'cargo fmt' to fix.
[error] 28-28: Rustfmt: multi-line struct initialization formatting changed. Run 'cargo fmt'.
[error] 168-168: Rustfmt: break device_layout_type call into multi-line layout.
[error] 175-175: Rustfmt: adjust formatting around unwrap_or in device layout chain.
[error] 179-179: Rustfmt: adjust formatting in device layout chain (map/unwrap).
lib/bindings/python/rust/llm/block_manager/vllm/connector/worker.rs
[error] 20-20: Rustfmt: import LayoutType inserted/moved; reformat to match style.
[error] 169-169: Rustfmt: break device_layout_type call into multi-line layout.
[error] 175-175: Rustfmt: trailing comma added and unwrap formatting adjusted.
[error] 179-179: Rustfmt: unwrap/map formatting adjusted in device layout chain.
[error] 26-26: Rustfmt: import LayoutType removed or reordered to satisfy formatting.
lib/bindings/python/rust/llm/block_manager/vllm/connector/trtllm_worker.rs
[error] 19-19: Rustfmt: import LayoutType inserted/moved; reformat to match style.
[error] 25-25: Rustfmt: import LayoutType adjusted to satisfy formatting.
🪛 GitHub Actions: NVIDIA Dynamo Github Validation
lib/llm/src/block_manager/block/transfer/nixl.rs
[error] 1-999: Rust formatting check failed. Run 'cargo fmt' to format code (cargo fmt -- --check).
lib/llm/src/block_manager/block/transfer/cuda.rs
[error] 1-999: Rust formatting check failed. Run 'cargo fmt' to format code (cargo fmt -- --check).
lib/llm/src/block_manager/distributed/transfer.rs
[error] 1-999: Rust formatting check failed. Run 'cargo fmt' to format code (cargo fmt -- --check).
lib/llm/src/block_manager/layout/utils.rs
[error] 1-999: Rust formatting check failed. Run 'cargo fmt' to format code (cargo fmt -- --check).
🔇 Additional comments (15)
lib/bindings/python/rust/llm/block_manager/vllm/connector/worker.rs (2)
26-26: Fix rustfmt import orderingSandbox formatting check failed (cargo fmt error: "Could not locate working directory: no /proc/self/exe available"); run cargo fmt locally and commit the resulting changes or reorder imports in lib/bindings/python/rust/llm/block_manager/vllm/connector/worker.rs (around line 26) to satisfy rustfmt.
169-176: Format chained builder calls (rustfmt)Per-edge layout config is correct; format the chained builder calls to satisfy rustfmt/Clippy.
File: lib/bindings/python/rust/llm/block_manager/vllm/connector/worker.rs (lines 169–176)
Run locally: cargo fmt && cargo clippy -q --no-depslib/llm/src/block_manager/block/transfer/nixl.rs (1)
1-999: Run cargo fmt (repo-wide)Formatter failures are blocking CI on this file; run
cargo fmt --alland commit the changes. Automated verification failed in the sandbox with: "Could not locate working directory.: no /proc/self/exe available" — run locally or in CI and re-request verification.lib/llm/src/block_manager/block/transfer/cuda.rs (1)
1-999: Run cargo fmt (repo-wide)Formatter failures are blocking CI on this file — run
cargo fmt --alland commit the resulting changes. Automated verification here failed (error: "Could not locate working directory: no /proc/self/exe available"); run locally or in CI to confirm formatting is fixed.lib/bindings/python/rust/llm/block_manager/vllm/connector/trtllm_worker.rs (1)
25-25: Fix rustfmt import ordering — run cargo fmtrustfmt is failing for the added
LayoutTypeimport in lib/bindings/python/rust/llm/block_manager/vllm/connector/trtllm_worker.rs (around line 25). Runcargo fmtlocally and commit the formatted changes. (Automated attempt failed with: "Could not locate working directory: no /proc/self/exe available".)lib/llm/src/block_manager/distributed/transfer.rs (1)
129-151: Run cargo fmt to fix formatting.The pipeline is failing on rustfmt. Please run
cargo fmtand commit the diff.lib/bindings/python/rust/llm/block_manager/distributed/worker.rs (2)
26-36: Enum mapping LGTM.
PyLayoutType→LayoutTypeconversion is correct and clear.
1-1: Run cargo fmt.Rustfmt errors are blocking. Please format this file.
lib/llm/src/block_manager/offload.rs (1)
1404-1408: Nice: index-based verification after bulk onboard.The revised assertions compare each disk→device pair by index and check contents; this is more robust than counting blocks.
lib/llm/src/block_manager/layout.rs (2)
113-118: Public utils module exposure LGTM.Making
utilspublic and reusing it across verifiers is consistent with added verification APIs.
558-610: Verification APIs for FullyContiguous look good.
verify_memory_regions,expected_memory_address, andverify_memory_regionare clean, use the shared verifier, and provide actionable logs.lib/llm/src/block_manager/layout/utils.rs (4)
83-104: Bounds checks look goodIndexes validated against config with clear, specific errors. LGTM.
403-447: Nice: cross‑layout compatibility testGood coverage of FC↔LS size compatibility without touching device memory.
448-476: Incompatible layouts test exercises the right surfaceThe differing blocks case is a good negative check.
105-477: Formatting check couldn't be executed in sandbox — run rustfmt locally and re-run formatting checkSandbox error returned: "Unable to proceed. Could not locate working directory.: no /proc/self/exe available. Is /proc mounted?"
Run locally and re-run CI:
cargo fmt && cargo fmt -- --check
oandreeva-nv
changed the title
oandreeva-nv
changed the title
oandreeva-nv
force-pushed
the
oandreeva_fc_g2_g3
branch
from
bea4304 to
bc4bc09
Compare
Signed-off-by: Olga Andreeva <[email protected]>
Signed-off-by: Olga Andreeva <[email protected]>
Signed-off-by: Olga Andreeva <[email protected]>
Signed-off-by: Olga Andreeva <[email protected]>
Signed-off-by: Olga Andreeva <[email protected]>
Signed-off-by: Olga Andreeva <[email protected]>
Signed-off-by: Olga Andreeva <[email protected]>
Signed-off-by: Olga Andreeva <[email protected]>
Signed-off-by: Olga Andreeva <[email protected]>
Signed-off-by: Olga Andreeva <[email protected]>
Signed-off-by: Olga Andreeva <[email protected]>
Signed-off-by: Olga Andreeva <[email protected]>
oandreeva-nv
force-pushed
the
oandreeva_fc_g2_g3
branch
from
bc4bc09 to
953f70b
Compare
|
@coderabbitai review |
✅ Actions performedReview triggered.
|
Signed-off-by: Olga Andreeva <[email protected]>
|
/ok to test c1afc69 |
|
/ok to test a8c411a |
Signed-off-by: Olga Andreeva <[email protected]>
|
/ok to test 7066c88 |
|
/ok to test b76ef8a |
Signed-off-by: Olga Andreeva <[email protected]> Signed-off-by: Olga Andreeva <[email protected]> Co-authored-by: oandreeva-nv <[email protected]>
Overview:
Fixes DIS-546
Summary by CodeRabbit
New Features
Tests
Details:
Where should the reviewer start?
Related Issues: (use one of the action keywords Closes / Fixes / Resolves / Relates to)
Summary by CodeRabbit