semblend 0.1.0__tar.gz

semblend-0.1.0/LICENSE

@@ -0,0 +1,61 @@
+Business Source License 1.1
+
+Parameters
+
+Licensor:             WorldFlow AI, Inc.
+Licensed Work:        SemBlend 0.1.0 and later
+                      The Licensed Work is (c) 2026 WorldFlow AI, Inc.
+Additional Use Grant: You may use the Licensed Work for non-production
+                      purposes, including testing, development, evaluation,
+                      and academic research. Production use requires a
+                      commercial license from WorldFlow AI.
+Change Date:          2030-03-16
+Change License:       Apache License, Version 2.0
+
+For information about alternative licensing arrangements for the Licensed Work,
+please contact licensing@worldflow.ai.
+
+Notice
+
+Business Source License 1.1
+
+Terms
+
+The Licensor hereby grants you the right to copy, modify, create derivative
+works, redistribute, and make non-production use of the Licensed Work. The
+Licensor may make an Additional Use Grant, above, permitting limited production
+use.
+
+Effective on the Change Date, or the fourth anniversary of the first publicly
+available distribution of a specific version of the Licensed Work under this
+License, whichever comes first, the Licensor hereby grants you rights under the
+terms of the Change License, and the rights granted in the paragraph above
+terminate.
+
+If your use of the Licensed Work does not comply with the requirements currently
+in effect as described in this License, you must purchase a commercial license
+from the Licensor, its affiliated entities, or authorized resellers, or you must
+refrain from using the Licensed Work.
+
+All copies of the original and modified Licensed Work, and derivative works of
+the Licensed Work, are subject to this License. This License applies separately
+for each version of the Licensed Work and the Change Date may vary for each
+version of the Licensed Work released by Licensor.
+
+You must conspicuously display this License on each original or modified copy of
+the Licensed Work. If you receive the Licensed Work in original or modified form
+from a third party, the terms and conditions set forth in this License apply to
+your use of that work.
+
+Any use of the Licensed Work in violation of this License will automatically
+terminate your rights under this License for the current and all other versions
+of the Licensed Work.
+
+This License does not grant you any right in any trademark or logo of Licensor
+or its affiliates (provided that you may use a trademark or logo of Licensor as
+expressly required by this License).
+
+TO THE EXTENT PERMITTED BY APPLICABLE LAW, THE LICENSED WORK IS PROVIDED ON AN
+"AS IS" BASIS. LICENSOR HEREBY DISCLAIMS ALL WARRANTIES AND CONDITIONS, EXPRESS
+OR IMPLIED, INCLUDING (WITHOUT LIMITATION) WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE, NON-INFRINGEMENT, AND TITLE.

semblend-0.1.0/PKG-INFO

@@ -0,0 +1,285 @@
+Metadata-Version: 2.4
+Name: semblend
+Version: 0.1.0
+Summary: Semantic KV cache reuse for LLM inference engines (vLLM, SGLang)
+Author: WorldFlow AI
+License: Business Source License 1.1
+        
+        Parameters
+        
+        Licensor:             WorldFlow AI, Inc.
+        Licensed Work:        SemBlend 0.1.0 and later
+                              The Licensed Work is (c) 2026 WorldFlow AI, Inc.
+        Additional Use Grant: You may use the Licensed Work for non-production
+                              purposes, including testing, development, evaluation,
+                              and academic research. Production use requires a
+                              commercial license from WorldFlow AI.
+        Change Date:          2030-03-16
+        Change License:       Apache License, Version 2.0
+        
+        For information about alternative licensing arrangements for the Licensed Work,
+        please contact licensing@worldflow.ai.
+        
+        Notice
+        
+        Business Source License 1.1
+        
+        Terms
+        
+        The Licensor hereby grants you the right to copy, modify, create derivative
+        works, redistribute, and make non-production use of the Licensed Work. The
+        Licensor may make an Additional Use Grant, above, permitting limited production
+        use.
+        
+        Effective on the Change Date, or the fourth anniversary of the first publicly
+        available distribution of a specific version of the Licensed Work under this
+        License, whichever comes first, the Licensor hereby grants you rights under the
+        terms of the Change License, and the rights granted in the paragraph above
+        terminate.
+        
+        If your use of the Licensed Work does not comply with the requirements currently
+        in effect as described in this License, you must purchase a commercial license
+        from the Licensor, its affiliated entities, or authorized resellers, or you must
+        refrain from using the Licensed Work.
+        
+        All copies of the original and modified Licensed Work, and derivative works of
+        the Licensed Work, are subject to this License. This License applies separately
+        for each version of the Licensed Work and the Change Date may vary for each
+        version of the Licensed Work released by Licensor.
+        
+        You must conspicuously display this License on each original or modified copy of
+        the Licensed Work. If you receive the Licensed Work in original or modified form
+        from a third party, the terms and conditions set forth in this License apply to
+        your use of that work.
+        
+        Any use of the Licensed Work in violation of this License will automatically
+        terminate your rights under this License for the current and all other versions
+        of the Licensed Work.
+        
+        This License does not grant you any right in any trademark or logo of Licensor
+        or its affiliates (provided that you may use a trademark or logo of Licensor as
+        expressly required by this License).
+        
+        TO THE EXTENT PERMITTED BY APPLICABLE LAW, THE LICENSED WORK IS PROVIDED ON AN
+        "AS IS" BASIS. LICENSOR HEREBY DISCLAIMS ALL WARRANTIES AND CONDITIONS, EXPRESS
+        OR IMPLIED, INCLUDING (WITHOUT LIMITATION) WARRANTIES OF MERCHANTABILITY,
+        FITNESS FOR A PARTICULAR PURPOSE, NON-INFRINGEMENT, AND TITLE.
+        
+Project-URL: Homepage, https://worldflow.ai/semblend
+Project-URL: Repository, https://github.com/worldflowai/semblend
+Project-URL: Documentation, https://docs.worldflow.ai/semblend
+Project-URL: Bug Tracker, https://github.com/worldflowai/semblend/issues
+Keywords: llm,inference,kv-cache,vllm,sglang,semantic-cache
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: License :: Other/Proprietary License
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: numpy>=1.24
+Requires-Dist: rapidfuzz>=3.0
+Provides-Extra: embedder
+Requires-Dist: sentence-transformers>=3.0; extra == "embedder"
+Requires-Dist: onnxruntime>=1.15; extra == "embedder"
+Provides-Extra: gpu
+Requires-Dist: torch>=2.0; extra == "gpu"
+Requires-Dist: triton>=2.0; extra == "gpu"
+Provides-Extra: vllm
+Requires-Dist: vllm>=0.8.0; extra == "vllm"
+Requires-Dist: torch>=2.0; extra == "vllm"
+Requires-Dist: sentence-transformers>=3.0; extra == "vllm"
+Provides-Extra: sglang
+Requires-Dist: sglang>=0.4.0; extra == "sglang"
+Requires-Dist: torch>=2.0; extra == "sglang"
+Requires-Dist: sentence-transformers>=3.0; extra == "sglang"
+Provides-Extra: dev
+Requires-Dist: pytest>=8.0; extra == "dev"
+Requires-Dist: pytest-asyncio>=0.23; extra == "dev"
+Requires-Dist: ruff>=0.4; extra == "dev"
+Provides-Extra: benchmarks
+Requires-Dist: aiohttp>=3.9; extra == "benchmarks"
+Requires-Dist: datasets>=2.16; extra == "benchmarks"
+Requires-Dist: tqdm>=4.66; extra == "benchmarks"
+Requires-Dist: rich>=13.7; extra == "benchmarks"
+Requires-Dist: rouge-score>=0.1.2; extra == "benchmarks"
+Requires-Dist: transformers>=4.40; extra == "benchmarks"
+Dynamic: license-file
+
+# SemBlend
+
+**Semantic KV cache reuse for LLM inference engines.**
+
+SemBlend extends exact-prefix KV caching (LMCache, vLLM prefix cache, SGLang RadixAttention) with *semantic* donor discovery. When a new prompt is semantically similar to a prior one but lexically different — different instruction wording, sentence ordering, or template fields — SemBlend finds and reuses the cached KV tensors, eliminating redundant prefill computation.
+
+```
+Without SemBlend:  vLLM + LMCache  →  0% hit  →  full prefill every request
+With SemBlend:     vLLM + LMCache  →  30–88% hit  →  reuse KV from similar past requests
+```
+
+## What It Does
+
+| System | Hit Condition | Semantically Similar Prompts |
+|--------|--------------|------------------------------|
+| vLLM prefix cache | Exact token-level prefix match | Full prefill (0% hit) |
+| LMCache alone | Exact 256-token chunk match | Full prefill (0% hit) |
+| **LMCache + SemBlend** | Semantic similarity ≥ 0.60 | **Reuse donor KV (30–88% hit)** |
+
+SemBlend runs ~8ms of in-process MiniLM embedding + cosine search on every request. On a miss it adds negligible overhead. On a hit it replaces a 2–17 second prefill with sub-second KV retrieval.
+
+## Performance
+
+Empirical results on A10G GPU, Qwen2.5-7B-AWQ + vLLM 0.14.1 + LMCache.
+
+### TTFT speedup on hits vs. cold prefill (diverse real content)
+
+| Context Length | Cold TTFT | SemBlend Miss (+overhead) | SemBlend Hit | Hit Speedup |
+|---------------|-----------|--------------------------|--------------|-------------|
+| 4K tokens | 1,859 ms | 1,864 ms (+0.3%) | 801 ms | **2.3x** |
+| 8K tokens | 3,193 ms | 3,315 ms (+3.8%) | 817 ms | **3.9x** |
+| 16K tokens | 5,852 ms | 6,064 ms (+3.6%) | 871 ms | **6.7x** |
+| 32K tokens | 15,418 ms | — | 1,288 ms | **12.0x** |
+
+*Hit TTFT is consistently ~800ms regardless of context length — it's bounded by KV retrieval, not prefill.*
+
+Miss overhead grows with context length (5–212ms at ~20 donors in store). Break-even: P_hit > 5% at 8K is net-positive.
+
+### Real-world hit rates (WildChat-1M user conversations)
+
+| Workload | Requests | Hit Rate | Hit-only TTFT speedup |
+|---------|----------|----------|----------------------|
+| Short prompts (≥4K chars) | 250 | 29.2% | **1.63x** |
+| Long prompts (≥8K chars) | 150 | 30.0% | **1.88x** |
+
+Hit rate increases with cosine similarity: 17% at 0.50–0.60 → 60% at 0.90–1.00.
+
+### Cross-dataset hit rates and speedups (8K tokens, Qwen2.5-7B)
+
+| Dataset | Hit Rate | Overall Speedup | Hit-only Speedup |
+|---------|----------|----------------|-----------------|
+| CNN/DailyMail (summarization) | 50% | 2.29x | 2.39x |
+| MultiNews (multi-doc summary) | 75% | 2.23x | 2.23x |
+| SAMSum (dialogue summary) | 88% | 2.37x | 2.37x |
+
+*Speedups measured over cold vLLM+LMCache baseline (0% hit on semantically different requests).*
+
+### Multi-turn dialogue (4K context, 3 turns)
+
+| Turn | Hit Rate | TTFT Speedup |
+|------|----------|-------------|
+| Turn 1 (cold) | — | 1.0x (baseline: 4019 ms) |
+| Turn 2 | 99.5% | **5.1x** (791 ms) |
+| Turn 3 | 99.5% | **5.1x** (787 ms) |
+
+Multi-turn conversations naturally reuse the same prefix → near-perfect hit rates without any workload tuning.
+
+### SGLang comparison (8K, cross-instruction RAG workload)
+
+| Engine | Hit Rate | p50 TTFT | Speedup |
+|--------|----------|----------|---------|
+| SGLang (RadixAttention, no SemBlend) | 0% | 3,850 ms | 0.98x |
+| vLLM + LMCache | 0% | 3,193 ms | 1.0x |
+| vLLM + LMCache + SemBlend | 50–88% | 817 ms | **3.9x** |
+| SGLang + SemBlend | 50–88% | 624 ms | **3.7x** |
+
+### Quality
+
+SemBlend injects donor KV with RoPE position correction. Output quality impact on high-hit workloads:
+
+| Dataset | PPL ratio (SemBlend / cold) |
+|---------|---------------------------|
+| CNN/DailyMail | 1.006 |
+| WikiHow | 1.012 |
+| XSum | 1.025 |
+| MultiNews (hit-only runs) | 1.007 |
+
+PPL ratio ≤ 1.025 on most datasets; elevated ratios (≥1.1) on low-hit runs are due to miss-run variance, not KV injection.
+
+## Installation
+
+```bash
+# Core (CPU-only: numpy + rapidfuzz)
+pip install semblend
+
+# With vLLM integration
+pip install semblend[vllm]
+
+# With SGLang integration
+pip install semblend[sglang]
+
+# With sentence-transformers embedder
+pip install semblend[embedder]
+```
+
+## Quick Start: vLLM
+
+```bash
+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"
+  }'
+```
+
+Configure via environment variables:
+
+| Variable | Default | Description |
+|----------|---------|-------------|
+| `SEMBLEND_ENABLED` | `1` | Enable semantic donor search |
+| `SEMBLEND_MIN_SIMILARITY` | `0.60` | Cosine similarity threshold |
+| `SEMBLEND_EMBEDDER` | `minilm` | Embedder type (`minilm`, `jaccard`, `onnx_gpu`) |
+| `SEMBLEND_FUZZY_CHUNKS` | `0` | Enable fuzzy chunk matching |
+
+## Quick Start: SGLang
+
+```bash
+pip install semblend[sglang] sglang
+
+# Option 1: CLI launcher (patches RadixCache automatically)
+semblend-sglang --model-path Qwen/Qwen2.5-7B-Instruct \
+  --host 0.0.0.0 --port 8000
+
+# Option 2: Programmatic patching
+from semblend.integration.sglang.radix_patcher import patch_radix_cache
+patch_radix_cache()  # Call before SGLang starts
+```
+
+## How It Works
+
+```
+Request → MiniLM Embed (5ms) → Cosine Search (1ms) → Align (1ms) → Inject KV
+              ↓                       ↓                    ↓
+         384-dim vector        Find similar donor    Match chunk boundaries
+                                in donor store       via MD5 hash alignment
+```
+
+1. **Embed**: Compute a 384-dim MiniLM-L6-v2 embedding (sliding-window for long docs)
+2. **Search**: Brute-force cosine similarity against the donor store (<1ms at 1K donors)
+3. **Align**: MD5 chunk hashing at 256-token boundaries finds reusable KV chunks
+4. **Inject**: Replace target token IDs with donor token IDs — LMCache/RadixCache finds cached KV
+
+## When SemBlend Helps
+
+SemBlend is most effective for workloads where prompts share a large common context:
+
+- **Document Q&A / RAG**: Same retrieved documents, different questions
+- **Summarization**: Same article, different instruction phrasing
+- **Multi-turn dialogue**: Conversation history grows across turns
+- **Code completion**: Shared repository context across requests
+
+SemBlend has minimal overhead on dissimilar workloads (e.g. code generation from scratch: measured 0% hit, 0.96x — 4% overhead from embedding + search).
+
+## License
+
+Business Source License 1.1 (BSL-1.1). Free for non-production use including testing, development, evaluation, and academic research. Production use requires a commercial license from WorldFlow AI. Converts to Apache License 2.0 on 2030-03-16.
+
+Contact: licensing@worldflow.ai
+

semblend-0.1.0/README.md

@@ -0,0 +1,172 @@
+# SemBlend
+
+**Semantic KV cache reuse for LLM inference engines.**
+
+SemBlend extends exact-prefix KV caching (LMCache, vLLM prefix cache, SGLang RadixAttention) with *semantic* donor discovery. When a new prompt is semantically similar to a prior one but lexically different — different instruction wording, sentence ordering, or template fields — SemBlend finds and reuses the cached KV tensors, eliminating redundant prefill computation.
+
+```
+Without SemBlend:  vLLM + LMCache  →  0% hit  →  full prefill every request
+With SemBlend:     vLLM + LMCache  →  30–88% hit  →  reuse KV from similar past requests
+```
+
+## What It Does
+
+| System | Hit Condition | Semantically Similar Prompts |
+|--------|--------------|------------------------------|
+| vLLM prefix cache | Exact token-level prefix match | Full prefill (0% hit) |
+| LMCache alone | Exact 256-token chunk match | Full prefill (0% hit) |
+| **LMCache + SemBlend** | Semantic similarity ≥ 0.60 | **Reuse donor KV (30–88% hit)** |
+
+SemBlend runs ~8ms of in-process MiniLM embedding + cosine search on every request. On a miss it adds negligible overhead. On a hit it replaces a 2–17 second prefill with sub-second KV retrieval.
+
+## Performance
+
+Empirical results on A10G GPU, Qwen2.5-7B-AWQ + vLLM 0.14.1 + LMCache.
+
+### TTFT speedup on hits vs. cold prefill (diverse real content)
+
+| Context Length | Cold TTFT | SemBlend Miss (+overhead) | SemBlend Hit | Hit Speedup |
+|---------------|-----------|--------------------------|--------------|-------------|
+| 4K tokens | 1,859 ms | 1,864 ms (+0.3%) | 801 ms | **2.3x** |
+| 8K tokens | 3,193 ms | 3,315 ms (+3.8%) | 817 ms | **3.9x** |
+| 16K tokens | 5,852 ms | 6,064 ms (+3.6%) | 871 ms | **6.7x** |
+| 32K tokens | 15,418 ms | — | 1,288 ms | **12.0x** |
+
+*Hit TTFT is consistently ~800ms regardless of context length — it's bounded by KV retrieval, not prefill.*
+
+Miss overhead grows with context length (5–212ms at ~20 donors in store). Break-even: P_hit > 5% at 8K is net-positive.
+
+### Real-world hit rates (WildChat-1M user conversations)
+
+| Workload | Requests | Hit Rate | Hit-only TTFT speedup |
+|---------|----------|----------|----------------------|
+| Short prompts (≥4K chars) | 250 | 29.2% | **1.63x** |
+| Long prompts (≥8K chars) | 150 | 30.0% | **1.88x** |
+
+Hit rate increases with cosine similarity: 17% at 0.50–0.60 → 60% at 0.90–1.00.
+
+### Cross-dataset hit rates and speedups (8K tokens, Qwen2.5-7B)
+
+| Dataset | Hit Rate | Overall Speedup | Hit-only Speedup |
+|---------|----------|----------------|-----------------|
+| CNN/DailyMail (summarization) | 50% | 2.29x | 2.39x |
+| MultiNews (multi-doc summary) | 75% | 2.23x | 2.23x |
+| SAMSum (dialogue summary) | 88% | 2.37x | 2.37x |
+
+*Speedups measured over cold vLLM+LMCache baseline (0% hit on semantically different requests).*
+
+### Multi-turn dialogue (4K context, 3 turns)
+
+| Turn | Hit Rate | TTFT Speedup |
+|------|----------|-------------|
+| Turn 1 (cold) | — | 1.0x (baseline: 4019 ms) |
+| Turn 2 | 99.5% | **5.1x** (791 ms) |
+| Turn 3 | 99.5% | **5.1x** (787 ms) |
+
+Multi-turn conversations naturally reuse the same prefix → near-perfect hit rates without any workload tuning.
+
+### SGLang comparison (8K, cross-instruction RAG workload)
+
+| Engine | Hit Rate | p50 TTFT | Speedup |
+|--------|----------|----------|---------|
+| SGLang (RadixAttention, no SemBlend) | 0% | 3,850 ms | 0.98x |
+| vLLM + LMCache | 0% | 3,193 ms | 1.0x |
+| vLLM + LMCache + SemBlend | 50–88% | 817 ms | **3.9x** |
+| SGLang + SemBlend | 50–88% | 624 ms | **3.7x** |
+
+### Quality
+
+SemBlend injects donor KV with RoPE position correction. Output quality impact on high-hit workloads:
+
+| Dataset | PPL ratio (SemBlend / cold) |
+|---------|---------------------------|
+| CNN/DailyMail | 1.006 |
+| WikiHow | 1.012 |
+| XSum | 1.025 |
+| MultiNews (hit-only runs) | 1.007 |
+
+PPL ratio ≤ 1.025 on most datasets; elevated ratios (≥1.1) on low-hit runs are due to miss-run variance, not KV injection.
+
+## Installation
+
+```bash
+# Core (CPU-only: numpy + rapidfuzz)
+pip install semblend
+
+# With vLLM integration
+pip install semblend[vllm]
+
+# With SGLang integration
+pip install semblend[sglang]
+
+# With sentence-transformers embedder
+pip install semblend[embedder]
+```
+
+## Quick Start: vLLM
+
+```bash
+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"
+  }'
+```
+
+Configure via environment variables:
+
+| Variable | Default | Description |
+|----------|---------|-------------|
+| `SEMBLEND_ENABLED` | `1` | Enable semantic donor search |
+| `SEMBLEND_MIN_SIMILARITY` | `0.60` | Cosine similarity threshold |
+| `SEMBLEND_EMBEDDER` | `minilm` | Embedder type (`minilm`, `jaccard`, `onnx_gpu`) |
+| `SEMBLEND_FUZZY_CHUNKS` | `0` | Enable fuzzy chunk matching |
+
+## Quick Start: SGLang
+
+```bash
+pip install semblend[sglang] sglang
+
+# Option 1: CLI launcher (patches RadixCache automatically)
+semblend-sglang --model-path Qwen/Qwen2.5-7B-Instruct \
+  --host 0.0.0.0 --port 8000
+
+# Option 2: Programmatic patching
+from semblend.integration.sglang.radix_patcher import patch_radix_cache
+patch_radix_cache()  # Call before SGLang starts
+```
+
+## How It Works
+
+```
+Request → MiniLM Embed (5ms) → Cosine Search (1ms) → Align (1ms) → Inject KV
+              ↓                       ↓                    ↓
+         384-dim vector        Find similar donor    Match chunk boundaries
+                                in donor store       via MD5 hash alignment
+```
+
+1. **Embed**: Compute a 384-dim MiniLM-L6-v2 embedding (sliding-window for long docs)
+2. **Search**: Brute-force cosine similarity against the donor store (<1ms at 1K donors)
+3. **Align**: MD5 chunk hashing at 256-token boundaries finds reusable KV chunks
+4. **Inject**: Replace target token IDs with donor token IDs — LMCache/RadixCache finds cached KV
+
+## When SemBlend Helps
+
+SemBlend is most effective for workloads where prompts share a large common context:
+
+- **Document Q&A / RAG**: Same retrieved documents, different questions
+- **Summarization**: Same article, different instruction phrasing
+- **Multi-turn dialogue**: Conversation history grows across turns
+- **Code completion**: Shared repository context across requests
+
+SemBlend has minimal overhead on dissimilar workloads (e.g. code generation from scratch: measured 0% hit, 0.96x — 4% overhead from embedding + search).
+
+## License
+
+Business Source License 1.1 (BSL-1.1). Free for non-production use including testing, development, evaluation, and academic research. Production use requires a commercial license from WorldFlow AI. Converts to Apache License 2.0 on 2030-03-16.
+
+Contact: licensing@worldflow.ai
+

semblend-0.1.0/pyproject.toml

@@ -0,0 +1,89 @@
+[project]
+name = "semblend"
+version = "0.1.0"
+description = "Semantic KV cache reuse for LLM inference engines (vLLM, SGLang)"
+readme = "README.md"
+requires-python = ">=3.10"
+license = {file = "LICENSE"}
+authors = [
+    { name = "WorldFlow AI" },
+]
+keywords = ["llm", "inference", "kv-cache", "vllm", "sglang", "semantic-cache"]
+classifiers = [
+    "Development Status :: 4 - Beta",
+    "Intended Audience :: Developers",
+    "Intended Audience :: Science/Research",
+    "License :: Other/Proprietary License",
+    "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+    "Topic :: Scientific/Engineering :: Artificial Intelligence",
+]
+
+dependencies = [
+    "numpy>=1.24",
+    "rapidfuzz>=3.0",
+]
+
+[project.urls]
+Homepage = "https://worldflow.ai/semblend"
+Repository = "https://github.com/worldflowai/semblend"
+Documentation = "https://docs.worldflow.ai/semblend"
+"Bug Tracker" = "https://github.com/worldflowai/semblend/issues"
+
+[project.scripts]
+semblend-sglang = "semblend.integration.sglang.launch_semblend_sglang:main"
+
+[project.optional-dependencies]
+embedder = [
+    "sentence-transformers>=3.0",
+    "onnxruntime>=1.15",
+]
+gpu = [
+    "torch>=2.0",
+    "triton>=2.0",
+]
+vllm = [
+    "vllm>=0.8.0",
+    "torch>=2.0",
+    "sentence-transformers>=3.0",
+]
+sglang = [
+    "sglang>=0.4.0",
+    "torch>=2.0",
+    "sentence-transformers>=3.0",
+]
+dev = [
+    "pytest>=8.0",
+    "pytest-asyncio>=0.23",
+    "ruff>=0.4",
+]
+benchmarks = [
+    "aiohttp>=3.9",
+    "datasets>=2.16",
+    "tqdm>=4.66",
+    "rich>=13.7",
+    "rouge-score>=0.1.2",
+    "transformers>=4.40",
+]
+
+[build-system]
+requires = ["setuptools>=68"]
+build-backend = "setuptools.build_meta"
+
+[tool.setuptools.packages.find]
+include = ["semblend*", "semblend_core*", "synapse_kv_connector*"]
+exclude = ["benchmarks*", "paper*", "patent*", "infra*", "literature*", "results*", "tests*", "synapse_kv_connector.tests*", "scripts*"]
+
+[tool.pytest.ini_options]
+testpaths = ["tests", "synapse_kv_connector/tests"]
+addopts = "--ignore=synapse_kv_connector/tests/test_rope_correction.py --ignore=synapse_kv_connector/tests/test_triton_kernels.py"
+
+[tool.ruff]
+line-length = 100
+target-version = "py310"
+
+[tool.ruff.lint]
+select = ["E", "F", "I", "W"]
+ignore = ["E501"]