SemBlend

Semantic donor discovery for engine-native KV caches.

SemBlend helps inference engines reuse expensive long-context prefill work when exact-prefix cache matching misses. It finds semantically similar prior requests, checks whether their cached KV may be useful, and hands bounded donor evidence to engine-native connector/provider paths.

SemBlend is the engine-facing semantic KV adapter layer: semantic search is evidence, and the inference backend remains authoritative for whether donor KV is materialized, recomputed, declined, or ignored.

Exact prefix cache only:
  same meaning, different framing  ->  exact cache miss  ->  cold prefill

Engine cache + SemBlend:
  same meaning, different framing  ->  semantic donor evidence  ->  backend validates reuse

What SemBlend Is

• A Python package for semantic donor discovery, alignment, and quality-gated KV reuse planning.
• A set of engine adapters for vLLM, SGLang, TensorRT-LLM, and Dynamo-facing experiments.
• A process-local library by default: no hosted service, daemon, or network dependency is required.
• Complementary to exact-prefix caching: exact hits should remain the fastest and highest-priority path.

What SemBlend Is Not

• Not a semantic response cache.
• Not a replacement for vLLM, SGLang, TensorRT-LLM, Dynamo, llm-d, LMCache, or HiCache.
• Not a guarantee that every semantic match becomes materialized KV reuse.
• Not a license to write approximate donor KV into exact prefix-cache state without backend support.

Integration Status

Surface	Status	What it is for	Notes
Core pipeline	Stable beta	Embedding, donor search, alignment, quality gates, layer-deviation planning	Installed by default
SGLang	Active	Engine-local semantic provider and Radix/HiCache integration	Patch-based path today; upstream provider interface is in progress
vLLM compatibility	Supported	Reproduce existing vLLM benchmark path through dynamic connector loading	May use LMCache for KV transfer; direct vLLM connector work is the strategic direction
TensorRT-LLM	Experimental	Connector/provider mapping and launch surface	Validate backend semantics before claiming materialized reuse
Dynamo-facing provider	Experimental	Semantic evidence/provider surface for Dynamo-style routing experiments

Installation

pip install semblend            # CPU-only core: numpy + rapidfuzz
pip install semblend[embedder]  # + sentence-transformers + ONNX Runtime
pip install semblend[onnx-gpu]  # + ONNX Runtime GPU
pip install semblend[vllm]      # + vLLM compatibility integration
pip install semblend[sglang]    # + SGLang integration
pip install semblend[trtllm]    # + TensorRT-LLM integration surface

Quick Start: SGLang

pip install semblend[sglang] sglang

# CLI launcher: applies the current RadixCache patch path automatically.
semblend-sglang \
  --model-path Qwen/Qwen2.5-7B-Instruct \
  --host 0.0.0.0 \
  --port 8000

Or patch programmatically before SGLang initializes:

from semblend.integration.sglang.radix_patcher import patch_radix_cache

patch_radix_cache()
# Start SGLang server after patching.

A first-class SemanticPrefixProvider interface is in progress upstream: sgl-project/sglang#20806

Quick Start: vLLM Compatibility Path

The current vLLM compatibility path integrates through vLLM dynamic connector loading and may use LMCache for KV transfer. This remains the supported path for reproducing the benchmark results below.

pip install semblend[vllm] vllm lmcache

vllm serve Qwen/Qwen2.5-7B-Instruct-AWQ \
  --kv-transfer-config '{
    "kv_connector": "SemBlendConnectorV1",
    "kv_connector_module_path": "semblend.integration.vllm.connector_v1",
    "kv_role": "kv_both"
  }'

Quick Start: TensorRT-LLM Experimental Surface

pip install semblend[trtllm] tensorrt_llm

semblend-trtllm --help

TensorRT-LLM support should be treated as an integration surface, not a blanket guarantee of semantic materialization. The backend still owns connector metadata, KV block lifecycle, transfer semantics, and decline/fallback behavior.

How It Works

Request
  -> embed or fingerprint prompt
  -> search local donor store
  -> align donor and target tokens
  -> apply quality gates and layer-deviation policy
  -> return donor evidence or a backend-native load plan
  -> backend validates, materializes, recomputes, or declines

1. Embed: long prompts are segmented into overlapping windows and embedded with MiniLM or ONNX Runtime GPU paths when configured.
2. Search: SemBlend searches a process-local donor store for semantically similar prior requests.
3. Align: token and chunk alignment identify reusable spans. Optional fuzzy chunk matching can handle shifted boundaries.
4. Plan: quality gates, reuse ratio, and bathtub-style layer sensitivity decide whether donor evidence is safe enough to pass to the backend.
5. Materialize or decline: the engine-local connector/provider decides whether to load KV, recompute layers, fall back to normal prefill, or record discovery-only telemetry.

Safety and Accounting Boundaries

SemBlend distinguishes these outcomes:

Outcome	Meaning
Exact prefix route	The serving engine/router found an exact cache hit. This should normally win before semantic reuse.
Semantic discovery only	SemBlend found a likely donor but did not affect placement or materialization.
Semantic placement	A router or substrate used SemBlend evidence to choose a better endpoint, but reuse is not yet confirmed.
Backend materialized reuse	The backend confirmed donor KV was safely loaded, blended, or recomputed for the request.
Backend declined reuse	The backend rejected the donor because of stale state, namespace mismatch, capacity, unsupported mode, or validation failure.
Cold fallback	Normal prefill/decode path. This is expected behavior, not failure.

Reuse ROI should be counted only after backend-confirmed materialization. A semantic match by itself is evidence, not a cache hit.

Configuration

Variable	Default	Description
SEMBLEND_ENABLED	1	Enable semantic donor search
SEMBLEND_MIN_SIMILARITY	0.60	Cosine similarity threshold
SEMBLEND_EMBEDDER	minilm	Embedder type, such as minilm or ONNX GPU paths
SEMBLEND_FUZZY_CHUNKS	0	Enable fuzzy chunk matching for shifted prefixes
SEMBLEND_DISCOVERY_ONLY	0	Run lookup and telemetry without materializing donor KV
SEMBLEND_STRICT_PREFIX_BOUNDARY	1	Keep materialization inside prefix-shaped backend contracts
SEMBLEND_EXACT_MATERIALIZATION_ONLY	1	Materialize only token-exact spans where required by the backend

Performance

These numbers are from the current vLLM compatibility benchmark path: A10G GPU at 0.85 utilization, Qwen2.5-7B-AWQ, vLLM 0.14.1 plus the LMCache-compatible integration path. Runs used fresh pod isolation and real HuggingFace datasets with n=15 per cell.

TTFT speedup vs cold prefill

Context	Cold TTFT	SemBlend TTFT	Speedup
4K	2,102 ms	433 ms	4.9x
8K	3,816 ms	539 ms	7.1x
12K	5,655 ms	648 ms	8.7x
16K	7,635 ms	760 ms	10.0x
24K	11,977 ms	972 ms	12.3x

The speedup grows with context length because cold prefill scales with prompt length while validated KV retrieval is much less sensitive to context length. Results depend on backend, model, cache state, donor density, and workload overlap.

Multi-dataset validation

Dataset	Content type	8K speedup	16K speedup	24K speedup
XSum	News summaries	7.1x	10.0x	12.3x
CNN/DailyMail	Long-form journalism	7.1x	9.4x	12.2x
MultiNews	Multi-document news	-	9.3x	-

Quality

Quality was validated across 5 datasets, 4-5 context lengths each, with PPL ratio and LLM-as-judge faithfulness scoring.

Dataset	PPL range	Status	Judge cold	Judge SemBlend	Faithful
XSum	1.018-1.054	PASS	0.84	0.84	100%
CNN/DailyMail	1.011-1.049	PASS	0.87	0.86	97%
WikiHow	0.987-1.037	PASS	0.82	0.84	97%
MultiNews	0.958-1.064	PASS	0.79	0.78	100%
SAMSum	1.140-1.198	ELEVATED	0.78	0.86	87%

PPL was below 1.065 for 4 of 5 datasets at all tested lengths. SAMSum showed elevated PPL due to short dialogue turns, while LLM-as-judge rated SemBlend output higher than cold in that cell. Treat these as workload-specific validation results, not a universal guarantee.

When SemBlend Helps

SemBlend is most useful when many requests share large semantic context regions:

• Document Q&A / RAG: same retrieved passages, different questions.
• Summarization: same article, different instruction phrasing.
• Agent workflows: repeated tool traces, plans, or memory summaries.
• Code assistance: shared repository or file context across requests.
• Enterprise knowledge workloads: repeated policies, contracts, tickets, or manuals.

When SemBlend Is Usually Not Worth It

• Short prompts where lookup overhead can dominate.
• Fully novel traffic with low donor density.

Benchmarks and Reproduction

Install benchmark dependencies:

pip install semblend[benchmarks]

The public package contains benchmark and test scaffolding, but benchmark numbers should always be reported with model, backend, cache configuration, hardware, dataset, context length, and whether reuse was backend-confirmed or discovery-only.

Contributing

See CONTRIBUTING.md.

License

Apache License 2.0.

Built at WorldFlow AI. For support contact research@worldflowai.com.