TensorSharp

English | 中文

A C# inference engine for running large language models (LLMs) locally using GGUF model files. TensorSharp provides a console application, a web-based chatbot interface, and Ollama/OpenAI-compatible HTTP APIs for programmatic access.

Documentation Map

Start here	Use this when you want to...
Quick build and usage	Build the solution, compile the native GGML bridge, and run the CLI or server
Supported model architectures	Check which GGUF architecture keys, modalities, thinking mode, and tool calling paths are implemented
Compute backends	Choose between pure C# CPU, direct CUDA/cuBLAS, MLX Metal, GGML CPU, GGML Metal, and GGML CUDA
HTTP APIs	Use the Ollama-compatible, OpenAI-compatible, or Web UI SSE endpoints
Per-model architecture cards	Read end-to-end documentation of one architecture (origin, forward graph, components, parameters, and how TensorSharp implements / optimizes prefill and decode)
Paged attention & continuous batching	Understand the vLLM-style paged KV cache, prefix sharing, and iteration-level scheduler
Inference benchmark matrix	Compare TensorSharp against llama.cpp and Ollama across text / multimodal workloads, with KV-cache dtype sweeps
Server API examples	Copy complete curl and Python examples for the server surface
Server integration tests	Exercise the public API contract against a running server

Current Status

Area	Status
Model families	Gemma 3/4, Qwen 3, Qwen 3.5/3.6-family GGUFs (qwen35, qwen35moe, qwen3next), GPT OSS, Nemotron-H (incl. Nemotron 3 Nano Omni), and Mistral 3
Inference hosts	CLI, interactive REPL, ASP.NET Core web UI, Ollama-style API, OpenAI Chat Completions-style API
Backends	Pure C# CPU, direct CUDA/cuBLAS (cuda), MLX Metal (mlx), GGML CPU, GGML Metal, GGML CUDA
Multimodal	Gemma 4 image/video/audio; Gemma 3, Qwen 3.5-family, Mistral 3, and Nemotron-H Omni image input
Continuous batching	vLLM-style paged KV cache, block-hash prefix sharing across requests, iteration-level scheduler (enabled by default; opt-out via --no-continuous-batching)
Server model scope	One explicitly hosted GGUF via --model; optional explicit projector via --mmproj; no directory scanning
Observability	Structured per-turn logs, queue status, and KV-cache reuse metrics across Web UI, Ollama, and OpenAI response shapes

Features

• Multi-architecture support -- Gemma 4, Gemma 3, Qwen 3, Qwen 3.5/3.6-family, GPT OSS, Nemotron-H, Mistral 3
• Multimodal inference -- image, video, and audio inputs (Gemma 4); images for Gemma 3 / Qwen 3.5-family / Mistral 3 / Nemotron-H Omni
• Thinking / reasoning mode -- structured chain-of-thought output with <think> / <|channel>thought / <|channel>analysis tags (Qwen 3, Qwen 3.5/3.6-family, Gemma 4, GPT OSS, Nemotron-H)
• Tool calling / function calling -- models can invoke user-defined tools; multi-turn tool-call conversations supported across all three API styles
• Quantized model support -- loads GGUF files with Q4_K_M, Q8_0, F16, MXFP4, and other quantization formats; performs native quantized matmul without dequantizing to FP32, including memory-efficient pure C# CPU loading for large GGUFs
• GPU-accelerated -- GGML Metal on macOS, GGML CUDA on Windows/Linux with NVIDIA GPUs, a direct CUDA/cuBLAS backend with PTX kernels, and an MLX backend for Apple Silicon (mlx-c / Metal), all with CPU fallbacks for unsupported ops
• Optimized pure C# CPU backend -- managed GEMM fast paths plus fused SIMD kernels for RMSNorm, RoPE, softmax, fused activations, and other inference hot paths
• Continuous batching & paged KV cache -- vLLM-style block-paged KV pool with block-hash prefix sharing across requests, iteration-level scheduler that admits / preempts sequences mid-batch, optional SSD-backed tier for very large KV working sets, and a native fused paged-attention kernel (TSGgml_PagedAttentionForward) that drives ggml_flash_attn_ext on Metal/CUDA. Enabled by default in TensorSharp.Server; opt-out with --no-continuous-batching. See docs/PAGED_ATTENTION_AND_CONTINUOUS_BATCHING.md.
• Batched / parallel inference -- IBatchedPagedModel.ForwardBatch implementations for Mistral 3, Gemma 4, GPT OSS, Qwen 3, Qwen 3.5/3.6-family, and Nemotron-H all run by default and pack N sequences into a single forward pass with paged K/V scatter and per-sequence attention via the native kernel. Each model exposes a TS_<FAMILY>_BATCHED=0 escape hatch (e.g. TS_GEMMA4_BATCHED=0, TS_QWEN35_BATCHED=0, TS_GPTOSS_BATCHED=0, TS_NEMOTRON_BATCHED=0) to fall back to the per-sequence KV-swap path for A/B comparison or regression isolation.
• Ollama & OpenAI API compatibility -- drop-in replacement endpoints for existing tooling
• Configurable sampling -- temperature, top-k, top-p, min-p, repetition/presence/frequency penalties, seed, stop sequences
• Chat templates -- auto-loaded from GGUF metadata (Jinja2), with hardcoded fallbacks per architecture
• Inference engine -- the new InferenceEngine (worker-thread scheduler + paged block pool) replaces the legacy single-request FIFO queue inside TensorSharp.Server. The HTTP adapters still emit queue-position chunks for backward compatibility but the engine itself handles concurrency.
• Batch processing -- JSONL input support in the console application, plus a built-in inference benchmark for prefill/decode throughput
• Streaming -- token-by-token output via SSE (web) or stdout (console), with abort/stop support for in-flight generations
• Hybrid SSM-Transformer -- Nemotron-H mixes Mamba2 SSM layers, attention-only layers, and MoE FFN layers in a single model. The Mamba2 step has both a per-sequence native kernel and a batched native kernel (TSGgml_NemotronMamba2BatchedStepF32, NEON SIMD + GCD parallelism) used by the batched path.
• Hybrid Attention-Recurrent -- Qwen 3.5/3.6-family models mix full-attention layers with GatedDeltaNet recurrent layers; the batched path keeps recurrent running state in a per-slot recurrent-state pool
• Mixture of Experts -- Gemma 4 MoE variants (e.g. gemma-4-26B-A4B), GPT OSS MoE (e.g. gpt-oss-20b), Qwen 3.5/3.6-family MoE (qwen35moe / qwen3next variants such as Qwen3.5-35B-A3B), and Nemotron-H MoE FFN layers
• Batched GPU MoE -- a single fused GGML graph dispatch handles all selected experts (plus the optional shared expert and residual add) for Qwen 3.5/3.6-family and Nemotron-H decode, eliminating per-expert round-trips
• KV cache codecs -- pluggable codec interface (IKvBlockCodec) with a built-in TurboQuant (Q4 / Q8) compressed codec for paged blocks, configurable via --paged-kv-quant-bits
• Message editing -- edit or delete previous messages in the web chat UI and regenerate from that point
• Text/Image/Audio/Video uploads -- the web UI accepts file uploads up to 500 MB, with automatic token-budget-aware truncation for large text files
• Per-turn observability -- structured logs capture the full user input and the full raw assistant output (both <think> reasoning and the final result) plus the KV cache hit ratio. The same cache-hit stats are surfaced through every API: prompt_cache_hit_tokens / prompt_cache_hit_ratio (Ollama), usage.prompt_tokens_details.cached_tokens (OpenAI), and promptTokens / kvReusedTokens / kvReusePercent in the Web UI SSE done event

Supported Model Architectures

Architecture	GGUF arch keys	Example Models	Multimodal	Thinking	Tool Calling	Card
Gemma 4	gemma4	gemma-4-E4B, gemma-4-31B, gemma-4-26B-A4B (MoE)	Image, Video, Audio	Yes	Yes	gemma4.md
Gemma 3	gemma3	gemma-3-4b	Image	No	No	gemma3.md
Qwen 3	qwen3	Qwen3-4B	Text only	Yes	Yes	qwen3.md
Qwen 3.5 / 3.6 family	qwen35, qwen35moe, qwen3next	Qwen3.5-9B (hybrid Attn+Recurrent), Qwen3.5-35B-A3B / Qwen3.6-35B-A3B (MoE)	Image	Yes	Yes	qwen35.md
GPT OSS	gptoss, gpt-oss	gpt-oss-20b (MoE)	Text only	Yes (always)	No	gptoss.md
Nemotron-H	nemotron_h, nemotron_h_moe	Nemotron-H-8B, Nemotron-H-47B (Hybrid SSM-Transformer, MoE), Nemotron 3 Nano Omni (image)	Image (Omni)	Yes	Yes	nemotron.md
Mistral 3	mistral3	Mistral-Small-3.1-24B-Instruct	Image	No	No	mistral3.md

See the per-model architecture cards for end-to-end documentation of each architecture (origin, forward graph, components, parameters, weight naming, and how TensorSharp implements / optimizes prefill and decode).

Model Downloads (GGUF)

TensorSharp loads models in GGUF format. Below are Hugging Face links where you can download GGUF files for each supported architecture. Pick a quantization that fits your hardware (Q4_K_M for low memory, Q8_0 for higher quality, etc.).

Architecture	Model	GGUF Download
Gemma 4	gemma-4-E4B-it	ggml-org/gemma-4-E4B-it-GGUF
Gemma 4	gemma-4-31B-it	ggml-org/gemma-4-31B-it-GGUF
Gemma 4	gemma-4-26B-A4B-it (MoE)	ggml-org/gemma-4-26B-A4B-it-GGUF
Gemma 4	gemma-4-mmproj (multimodal projector)	Included in the GGUF repos above
Gemma 3	gemma-3-4b-it	google/gemma-3-4b-it-qat-q4_0-gguf
Qwen 3	Qwen3-4B	Qwen/Qwen3-4B-GGUF
Qwen 3.5 / 3.6 family	Qwen3.5-9B	unsloth/Qwen3.5-9B-GGUF
Qwen 3.5 / 3.6 family	Qwen3.5-35B-A3B	ggml-org/Qwen3.5-35B-A3B-GGUF
GPT OSS	gpt-oss-20b (MoE)	ggml-org/gpt-oss-20b-GGUF
Nemotron-H	Nemotron-H-8B-Reasoning-128K	bartowski/nvidia_Nemotron-H-8B-Reasoning-128K-GGUF
Nemotron-H	Nemotron-H-47B-Reasoning-128K	bartowski/nvidia_Nemotron-H-47B-Reasoning-128K-GGUF
Mistral 3	Mistral-Small-3.1-24B-Instruct	bartowski/Mistral-Small-3.1-24B-Instruct-2503-GGUF
Mistral 3	mistral3-mmproj (Pixtral vision projector)	bartowski/Mistral-Small-3.1-24B-Instruct-2503-GGUF

Compute Backends

Backend	Flag	Best fit	Description
Direct CUDA/cuBLAS	--backend cuda	NVIDIA inference and experimentation	Uses the CUDA Driver API, cuBLAS GEMM, PTX kernels for common float32 ops (fill, unary, binary, ternary, activations, RMSNorm, softmax, RoPE/RoPEEx, SDPA, GQA prefill/decode, causal mask, gather/concat), and native quantized matmul/get-rows for supported GGUF quant types. Unsupported ops route through CPU fallbacks while preserving tensor semantics.
MLX Metal	--backend mlx	Apple Silicon (alternative to GGML Metal)	GPU-accelerated path built on mlx-c. Implements quantized ops (Q4_K_M, Q8_0, Q5_K, Q6_K, IQ2_XXS, IQ4_XS, IQ4_NL, MXFP4, etc.) without dequantizing to FP32, fused decode/prefill Metal kernels (fused QKV preprocess, fused gate+up+SiLUMul MoE, fused multi-dim KV write), compiled-graph kernels, async worker dispatch with periodic async_eval to overlap GPU/CPU work, batched MoE decode with stacked expert weight slabs, MoE expert offload, GGUF mmap pinned in physical RAM via mlock(2), host-derived allocator caps (TS_MLX_MEMORY_LIMIT_MB / TS_MLX_CACHE_LIMIT_MB / TS_MLX_WIRED_LIMIT_MB), and a CPU fallback for ops that aren't yet wired up. Requires libmlxc (built locally by TensorSharp.Backends.MLX/build-native-macos.sh or located via TENSORSHARP_MLX_LIBRARY / TENSORSHARP_MLX_LIBRARY_DIR).
GGML Metal	--backend ggml_metal	Apple Silicon (default on macOS)	GPU-accelerated via Apple Metal. Quantized weights are mapped zero-copy from the GGUF file into Metal command buffers via host-pointer buffers, so the resident set stays close to the on-disk model size.
GGML CUDA	--backend ggml_cuda	NVIDIA inference through ggml	GPU-accelerated via GGML CUDA on Windows or Linux. Quantized weights are uploaded to device memory once at load time and the host copy is released afterwards.
GGML CPU	--backend ggml_cpu	Native CPU kernels	CPU inference using native GGML with optimized kernels. Quantized weights are mapped zero-copy from the GGUF file.
Pure C# CPU	--backend cpu	Portability and debugging	Portable CPU inference with no native dependencies.

Project Structure

TensorSharp/
├── TensorSharp.Core/            # Core tensor library (Tensor, Ops, memory, device abstraction, CPU SIMD/managed quantized kernels)
├── TensorSharp.Runtime/         # GGUF, tokenizers, templates, sampling, protocol parsing
│   ├── Paged/                   # Paged KV cache primitives (BlockPool, BlockTable, KvBlock, BlockHashIndex, PagedKvStorage, PagedKvBatchOps, ManagedPagedAttention)
│   ├── Scheduling/              # Continuous batching engine (InferenceEngine, BatchExecutor, ContinuousBatchScheduler, SequenceState, SchedulerConfig/Output, InferenceRequestHandle)
│   ├── PagedKvCacheManager.cs   # Per-session paged KV manager (block allocation, prefix reuse)
│   ├── PagedKvBlockStore.cs     # On-disk / RAM-tiered paged block storage with optional SSD spillover
│   ├── SsdKvBlockTier.cs        # SSD-backed cold tier for paged blocks
│   ├── TurboQuantKvCodec.cs     # Quantized KV block codec (Q4 / Q8) implementing IKvBlockCodec
│   ├── PrefillChunking.cs       # Chunked-prefill helper used by SWA / very long prompts
│   ├── KvBlockHash.cs           # Content-addressed block hash for prefix-cache sharing
│   └── Logging/                 # JSON-line file logger + per-turn telemetry
├── TensorSharp.Models/          # Model architectures and multimodal encoders/injectors
│   ├── Models/<Family>/         # One folder per architecture (Gemma3, Gemma4, GptOss, Mistral3, Nemotron, Qwen3, Qwen35)
│   │   ├── <Family>Model.cs                # Legacy per-sequence ModelBase implementation
│   │   └── <Family>Model.BatchedForward.cs # IBatchedPagedModel.ForwardBatch — batched/paged path (Mistral3, Gemma4, GptOss, Qwen35, Nemotron, Qwen3)
│   ├── Paged/                   # Tensor-side paged-attention helpers (TensorPagedAttention)
│   ├── KvBlockTransfer.cs       # Helpers for extract/inject of KV blocks across sequences
│   └── ModelMultimodalInjector.cs # Vision / audio / video embedding injection
├── TensorSharp.Backends.GGML/   # GGML backend bindings (Metal/CUDA/CPU via native library)
├── TensorSharp.Backends.Cuda/   # Direct CUDA backend using CUDA Driver API, cuBLAS, and PTX kernels
├── TensorSharp.Backends.MLX/    # Apple Silicon MLX backend (mlx-c / Metal). Native bridge is built via `build-native-macos.sh`.
├── TensorSharp.GGML.Native/     # Native C++ bridge to ggml (builds libGgmlOps, split into focused source files)
│   ├── ggml_ops_core.cpp                  # Element-wise, reductions, basic shape ops
│   ├── ggml_ops_elementwise.cpp           # Element-wise / activation fusions
│   ├── ggml_ops_matmul.cpp                # GEMM / quantized matmul
│   ├── ggml_ops_fused.cpp                 # Cross-cutting fused per-layer kernels
│   ├── ggml_ops_norm_attn.cpp             # Norm + attention fusions
│   ├── ggml_ops_transformer.cpp           # Full-layer fused transformer kernels (decode + prefill)
│   ├── ggml_ops_moe.cpp                   # Mixture-of-Experts forward / fused router
│   ├── ggml_ops_gated_delta_net.cpp       # Qwen 3.5/3.6 GatedDeltaNet kernels (per-seq + batched)
│   ├── ggml_ops_mamba2.cpp                # Nemotron Mamba2 kernels (per-seq + batched SIMD)
│   ├── ggml_ops_paged_attention.cpp       # Paged-attention native kernel (drives ggml_flash_attn_ext + sinks variant)
│   ├── ggml_ops_training.cpp              # Training-only kernels (unused at runtime)
│   └── tests/                              # Native unit + smoke tests
├── TensorSharp.Server/          # Web chatbot + API server (ASP.NET Core)
│   ├── Program.cs               # Slim bootstrap: DI wiring, middleware, endpoint mapping, paged-KV + continuous-batching CLI translation
│   ├── ModelService.cs          # Facade that keeps the public server inference API stable; owns the InferenceEngineHost
│   ├── ModelLifecycleService.cs # Model load/dispose and backend selection (CPU / CUDA / MLX / GGML CPU/Metal/CUDA)
│   ├── InferenceEngineHost.cs   # DI-registered per-model InferenceEngine singleton (continuous batching entry point)
│   ├── SessionKvCacheManager.cs # Active session switching, KV reuse/truncate/reset, prefill chunking (legacy per-seq path)
│   ├── ChatGenerationPipeline.cs # Prompt rendering, submits to InferenceEngine, streams tokens, stop handling
│   ├── InferenceTelemetry.cs    # Prompt/eval timing, TTFT, tokens/sec, full input/output logs
│   ├── ChatHistoryPreparer.cs   # History normalization, raw-token splice helpers, multimodal order helpers
│   ├── ChatSession.cs           # Per-conversation KV cache + tracked history
│   ├── SessionManager.cs        # Thread-safe session registry (default + per-tab sessions)
│   ├── InferenceQueue.cs        # Backward-compatible queue-status surface (engine itself handles concurrency)
│   ├── BackendCatalog.cs        # Discovery of available compute backends (CPU / CUDA / MLX / GGML*)
│   ├── TextUploadHelper.cs      # Token-budget-aware text-file truncation
│   ├── WebUiChatPolicy.cs       # Web UI chat request validation
│   ├── OpenAIResponseFormatParser.cs  # OpenAI response_format (json_object / json_schema) parsing
│   ├── Hosting/                 # Startup-time concerns: options builder (ServerOptionsBuilder), backend resolution, logging, web root, paged-KV / continuous-batching CLI translation
│   ├── RequestParsers/          # JSON request parsing (sampling, chat messages, tool functions)
│   ├── ResponseSerializers/     # Per-protocol response shape factories (Ollama, OpenAI, Web UI)
│   ├── StreamingWriters/        # SSE + NDJSON wire-format helpers
│   ├── ProtocolAdapters/        # Per-protocol request handlers (WebUiAdapter, OllamaAdapter, OpenAIChatAdapter)
│   ├── Endpoints/               # ASP.NET Core endpoint mapping (one extension method per protocol)
│   ├── Logging/                 # Request logging middleware + low-noise path support
│   ├── wwwroot/index.html       # Chat UI
│   ├── testdata/                # Integration test suites (bash + Python)
│   └── API_EXAMPLES.md          # Detailed API documentation
├── TensorSharp.Cli/             # CLI application (one-shot generation, interactive REPL, batch JSONL, benchmarks)
├── InferenceWeb.Tests/          # xUnit unit tests covering ops, KV cache, paged scheduler, batched-model correctness, web/server helpers
├── AdvUtils/                    # Utility library (logger)
├── docs/                        # Developer reference
│   ├── models/                  # Per-model architecture cards (one .md per model, EN + 中文)
│   ├── PAGED_ATTENTION_AND_CONTINUOUS_BATCHING.md  # Paged KV cache, prefix sharing, scheduler, per-model batched-forward status
│   └── inference_benchmark_matrix.md  # Cross-engine throughput matrix (TensorSharp vs llama.cpp vs Ollama)
├── benchmarks/                  # Reproducible benchmark harnesses
│   └── inference_matrix/        # Driver scripts, modelfiles, prompts, and per-cell raw JSON results
└── ExternalProjects/            # Third-party dependencies (ggml)

NuGet Packages

The repository is now split along package boundaries so consumers can depend on only the layers they actually need.

Project	NuGet package	Public namespace	Responsibility
TensorSharp.Core	TensorSharp.Core	TensorSharp	Tensor primitives, ops, allocators, storage, and device abstraction
TensorSharp.Runtime	TensorSharp.Runtime	TensorSharp.Runtime	GGUF parsing, tokenizers, prompt rendering, sampling, output protocol parsing, paged KV cache, continuous-batching scheduler
TensorSharp.Models	TensorSharp.Models	TensorSharp.Models	ModelBase, architecture implementations, multimodal encoders, batched / paged forward passes, and model-side execution helpers
TensorSharp.Backends.GGML	TensorSharp.Backends.GGML	TensorSharp.GGML	GGML-backed execution and native interop
TensorSharp.Backends.Cuda	TensorSharp.Backends.Cuda	TensorSharp.Cuda	Direct CUDA allocator, storage, cuBLAS GEMM, PTX kernels, and quantized CUDA ops
TensorSharp.Backends.MLX	TensorSharp.Backends.MLX	TensorSharp.MLX	Apple Silicon MLX backend (mlx-c / Metal) with quantized / fused / compiled kernels and MoE expert offload
TensorSharp.Server	TensorSharp.Server	TensorSharp.Server	ASP.NET Core server, OpenAI/Ollama adapters, inference engine host, web UI
TensorSharp.Cli	TensorSharp.Cli	TensorSharp.Cli	Console host and debugging / batch tooling

This split keeps engine users off the web stack, keeps API-layer changes from leaking into core/runtime packages, and makes future benchmark or eval-harness projects easier to publish independently.

Validate package metadata and README dependency boundaries before publishing:

pwsh ./eng/verify-packages.ps1

The verifier runs dotnet pack for the public packages above and fails if an internal dependency such as AdvUtils leaks into the .nuspec, or if a TensorSharp package depends on a layer outside this table.

Prerequisites

• .NET 10 SDK
• macOS (Metal backend): CMake 3.20+ and Xcode command-line tools for building the native GGML library; the MLX backend additionally builds libmlxc from TensorSharp.Backends.MLX/Native/ via bash TensorSharp.Backends.MLX/build-native-macos.sh
• Windows (GGML CPU / CUDA backends): CMake 3.20+ and Visual Studio 2022 C++ build tools; for ggml_cuda or cuda, install an NVIDIA driver plus CUDA Toolkit 12.x or another compatible CUDA toolkit with cuBLAS
• Linux (GGML CPU / CUDA backends): CMake 3.20+; for ggml_cuda or cuda, install an NVIDIA driver plus CUDA Toolkit 12.x or another compatible CUDA toolkit with cuBLAS
• GGUF model files (e.g., from Hugging Face)

Building

Build the entire solution

dotnet build TensorSharp.slnx

Build individual applications

# Console application
dotnet build TensorSharp.Cli/TensorSharp.Cli.csproj

# Web application
dotnet build TensorSharp.Server/TensorSharp.Server.csproj

Build the native GGML library

The native library is built automatically during the first dotnet build if it doesn't exist. To build it manually:

cd TensorSharp.GGML.Native

macOS:

bash build-macos.sh

Linux (CPU-only):

bash build-linux.sh

Linux (GGML_CUDA enabled):

bash build-linux.sh --cuda

Windows (CPU-only):

.\build-windows.ps1 --no-cuda

Windows (GGML_CUDA enabled):

.\build-windows.ps1 --cuda

On Windows and Linux, the native build script auto-detects the visible NVIDIA GPU compute capability and passes a narrow CMAKE_CUDA_ARCHITECTURES value to ggml-cuda (for example 86-real on an RTX 3080), which cuts CUDA build time substantially. The native build also runs in parallel by default with a conservative job cap so nvcc does not overwhelm typical developer machines.

If you want to override the auto-detected architecture list or the default build parallelism, use either environment variables or explicit build flags:

TENSORSHARP_GGML_NATIVE_CUDA_ARCHITECTURES='86-real;89-real' bash build-linux.sh --cuda
bash build-linux.sh --cuda --cuda-arch='86-real;89-real'
TENSORSHARP_GGML_NATIVE_BUILD_PARALLEL_LEVEL=2 bash build-linux.sh --cuda

$env:TENSORSHARP_GGML_NATIVE_CUDA_ARCHITECTURES='86-real;89-real'; .\build-windows.ps1 --cuda
.\build-windows.ps1 --cuda --cuda-arch='86-real;89-real'
$env:TENSORSHARP_GGML_NATIVE_BUILD_PARALLEL_LEVEL=2; .\build-windows.ps1 --cuda

You can also request a CUDA-enabled native build from dotnet build:

TENSORSHARP_GGML_NATIVE_ENABLE_CUDA=ON dotnet build TensorSharp.Cli/TensorSharp.Cli.csproj -c Release

$env:TENSORSHARP_GGML_NATIVE_ENABLE_CUDA='ON'; dotnet build TensorSharp.Cli/TensorSharp.Cli.csproj -c Release

On macOS this compiles libGgmlOps.dylib with Metal GPU support. On Windows and Linux, the native scripts preserve an existing CUDA-enabled build and auto-enable GGML_CUDA when a CUDA toolchain is detected; build-windows.ps1 --cuda, build-linux.sh --cuda, and TENSORSHARP_GGML_NATIVE_ENABLE_CUDA=ON force CUDA explicitly. The build output is automatically copied to the application's output directory.

The direct cuda backend is built as managed C# plus PTX kernels. During dotnet build, TensorSharp.Backends.Cuda compiles native/kernels/*.cu to native/ptx/*.ptx when nvcc is available; if nvcc is missing, the build continues and PTX-backed ops use CPU fallbacks. cuBLAS-backed GEMM still requires the CUDA runtime libraries to be discoverable at run time.

Build the native MLX library (macOS only)

The MLX backend depends on libmlxc (the C bindings for MLX). The repository pins a known-good tag of mlx-c in TensorSharp.Backends.MLX/Native/MLX_C_VERSION and a helper script fetches and builds it:

bash TensorSharp.Backends.MLX/build-native-macos.sh

The script writes the resulting libraries (libmlxc.dylib, libmlx.dylib, and any backend deps) into TensorSharp.Backends.MLX/Native/dist/. At run time the backend probes the application directory first; you can also point it to a custom install with TENSORSHARP_MLX_LIBRARY=<path-to-libmlxc.dylib> or TENSORSHARP_MLX_LIBRARY_DIR=<dir-with-libmlxc>. If the library cannot be located the backend reports unavailable and --backend mlx is rejected at startup.

Usage

Console Application

cd TensorSharp.Cli/bin

# Text inference
./TensorSharp.Cli --model <model.gguf> --input prompt.txt --output result.txt \
    --max-tokens 200 --backend ggml_metal

# Text inference on Windows/Linux + NVIDIA GPU
./TensorSharp.Cli --model <model.gguf> --input prompt.txt --output result.txt \
    --max-tokens 200 --backend ggml_cuda

# Interactive turn-by-turn chat (REPL) with KV cache reuse and slash commands
./TensorSharp.Cli --model <model.gguf> --backend ggml_metal --interactive
./TensorSharp.Cli --model <model.gguf> --backend ggml_metal -i \
    --system "You are a terse assistant." --temperature 0.7 --top-p 0.9 --think

# Image inference (Gemma 3/4, Qwen 3.5-family)
./TensorSharp.Cli --model <model.gguf> --image photo.png --backend ggml_metal

# Video inference (Gemma 4)
./TensorSharp.Cli --model <model.gguf> --video clip.mp4 --backend ggml_metal

# Audio inference (Gemma 4)
./TensorSharp.Cli --model <model.gguf> --audio speech.wav --backend ggml_metal

# Thinking / reasoning mode
./TensorSharp.Cli --model <model.gguf> --input prompt.txt --backend ggml_metal --think

# Tool calling
./TensorSharp.Cli --model <model.gguf> --input prompt.txt --backend ggml_metal \
    --tools tools.json

# With sampling parameters
./TensorSharp.Cli --model <model.gguf> --input prompt.txt --backend ggml_metal \
    --temperature 0.7 --top-p 0.9 --top-k 40 --repeat-penalty 1.2 --seed 42

# Batch processing (JSONL)
./TensorSharp.Cli --model <model.gguf> --input-jsonl requests.jsonl \
    --output results.txt --backend ggml_metal

# Multi-turn chat simulation with KV-cache reuse (mirrors the web UI behavior)
./TensorSharp.Cli --model <model.gguf> --multi-turn-jsonl chat.jsonl \
    --backend ggml_metal --max-tokens 200

# Throughput benchmark: best-of-N prefill and decode timing
./TensorSharp.Cli --model <model.gguf> --backend ggml_metal \
    --benchmark --bench-prefill 256 --bench-decode 128 --bench-runs 3

# KV-cache reuse benchmark: measure prefill speedup across multiple chat turns
# (compares with-cache vs forced-reset prefill latency for an 8-turn conversation)
./TensorSharp.Cli --model <model.gguf> --backend ggml_metal \
    --bench-kvcache --bench-kv-turns 4 --max-tokens 64

# Inspect the rendered prompt and tokenization without running inference
./TensorSharp.Cli --model <model.gguf> --input prompt.txt --dump-prompt

# Compare hardcoded fallback templates against GGUF Jinja2 templates for every
# *.gguf file in a directory (useful when adding new architectures)
./TensorSharp.Cli --test-templates ~/models

Command-line options:

Option	Description
--model <path>	Path to a GGUF model file (required)
--input <path>	Text file containing the user prompt
--input-jsonl <path>	JSONL file with batch requests (one JSON per line)
--multi-turn-jsonl <path>	JSONL file for multi-turn chat simulation with KV cache reuse
--output <path>	Write generated text to this file
--image <path>	Image file for vision inference
--video <path>	Video file for video inference
--audio <path>	Audio file (WAV, MP3, OGG) for audio inference
--mmproj <path>	Path to the multimodal projector GGUF file
--max-tokens <N>	Maximum tokens to generate (default: 100)
--backend <type>	Compute backend: cpu, cuda, mlx, ggml_cpu, ggml_metal, or ggml_cuda
--kv-cache-dtype <type>	KV cache precision: f32 (default), f16, or q8_0. Quantized / half-precision KV caches reduce memory at the cost of small numerical drift; benchmarks live in docs/inference_benchmark_matrix.md.
--interactive / -i	Start an interactive REPL chat session (turn-by-turn input/output) with KV cache reuse, slash commands, hot-swappable model/backend/projector, file attachments (image, audio, video, text) and live sampling tuning. See the Interactive REPL commands section below for the full list.
--system <text>	System prompt to seed the interactive session (overridden inside the REPL by /system)
--system-file <path>	Read the initial system prompt from a UTF-8 text file (alternative to --system)
--think	Enable thinking/reasoning mode (chain-of-thought)
--tools <path>	JSON file with tool/function definitions
--temperature <f>	Sampling temperature (0 = greedy)
--top-k <N>	Top-K filtering (0 = disabled)
--top-p <f>	Nucleus sampling threshold (1.0 = disabled)
--min-p <f>	Minimum probability filtering (0 = disabled)
--repeat-penalty <f>	Repetition penalty (1.0 = none)
--presence-penalty <f>	Presence penalty (0 = disabled)
--frequency-penalty <f>	Frequency penalty (0 = disabled)
--seed <N>	Random seed (-1 = non-deterministic)
--stop <string>	Stop sequence (can be repeated)
--dump-prompt	Render the prompt + tokenization and exit (no generation)
--benchmark	Run a synthetic prefill/decode throughput benchmark
--bench-prefill <N>	Synthetic prefill length in tokens (default: 32)
--bench-decode <N>	Synthetic decode length in tokens (default: 64)
--bench-runs <N>	Number of benchmark runs; reports best and average (default: 1)
--bench-kvcache	Run a multi-turn KV-cache reuse benchmark (with-cache vs forced-reset prefill)
--bench-kv-turns <N>	Number of conversation turns for --bench-kvcache (default: 4, max: 8)
--bench-chunked	Run a chunked-prefill micro-benchmark (Gemma 4)
--warmup-runs <N>	Number of throw-away forward passes before timing real text / multimodal prompts (default: 0)
--test-chunked-prefill	Run the chunked-prefill correctness check (compares chunked vs non-chunked logits)
--correct-prefill <N>	Prompt length used by --test-chunked-prefill
--correct-decode <N>	Decode length used by --test-chunked-prefill
--test	Run built-in tokenizer + Qwen3 chat-template + ollama-comparison tests
--test-templates <dir>	Validate hardcoded chat templates against GGUF Jinja2 templates for every *.gguf in <dir>
--log-level <lvl>	Console + file logger level: trace, debug, info, warning, error, critical, off
--log-dir <path>	Directory for the JSON-line file logger (default: <binDir>/logs)
--log-file <0|1>	Disable (0) or enable (1) the file logger (default: enabled)
--log-console <0|1>	Disable (0) or enable (1) the console logger (default: enabled)

The multimodal projector file is auto-detected if placed alongside the model file with a recognized name (e.g., gemma-4-mmproj-F16.gguf).

JSONL input format:

Each line is a JSON object with messages, optional prompt, and optional sampling parameters:

{"id": "q1", "messages": [{"role": "user", "content": "What is 2+3?"}], "max_tokens": 50}
{"id": "q2", "messages": [{"role": "user", "content": "Write a haiku."}], "max_tokens": 100, "temperature": 0.8}

Interactive REPL commands:

Once the CLI is launched with --interactive / -i, you can drive the running session with slash commands. Type /help (or /?) inside the REPL for the same list. Anything that does not start with / is treated as a user turn.

The prompt header summarizes the current state on every turn — model, backend, architecture, context length, projector, conversation depth, and any attachments queued for the next turn (e.g. [turn 3 (2 attachments pending)]> ). Press Ctrl+C while generating to interrupt the current reply; press Ctrl+C at the prompt to exit.

Conversation:

Command	Description
/help, /?	Show all interactive commands
/exit, /quit	Leave the session
/reset, /new	Clear conversation history and KV cache
/history	Print the conversation history
/save <file>	Append the current transcript to a UTF-8 file
/system <text>	Set the system prompt (empty argument clears it). Resets KV cache.
/think on|off	Toggle thinking/reasoning mode for supported models
/multiline on|off	Toggle multi-line input (terminate the message with a single . on its own line)

Model and runtime:

Command	Description
/info, /status	Show the loaded model, backend, architecture, context/vocab size, projector, conversation depth, and pending attachments
/model <path>	Load a different .gguf model on the current backend (resets the session)
/backend <name>	Reload the current model on a different backend: cpu, cuda, mlx, ggml_cpu, ggml_metal, or ggml_cuda
/mmproj <path>	Load (or replace) the multimodal projector for the current model. Aliases: /projector

Sampling (live, persists across turns):

Command	Description
/sampling, /show	Print the current sampling configuration
/max <N>	Maximum reply length in tokens
/temp <float>	Sampling temperature (0 = greedy)
/topk <int>	Top-K filtering (0 = disabled)
/topp <float>	Top-P / nucleus threshold (1.0 = disabled)
/minp <float>	Min-P filtering (0 = disabled)
/repeat <float>	Repetition penalty (1.0 = none)
/presence <float>	Presence penalty
/frequency <float>	Frequency penalty
/seed <int>	Random seed (-1 = non-deterministic)
/stop <text>	Add a stop sequence
/clearstop	Remove all stop sequences

Uploads (queued for the next user turn, then auto-cleared after the turn):

Command	Description
/image <path>, /img <path>	Attach an image (vision-capable models only)
/audio <path>	Attach an audio file (Gemma 4)
/video <path>, /vid <path>	Attach a video; frames are extracted automatically (Gemma 4)
/text <path>, /file <path>, /txt <path>	Inline a UTF-8 text/markdown/csv/code file into the next prompt (large files are token-budget truncated)
/clearattach	Drop any pending image/audio/video/text attachments without sending a turn

Quoted paths (single or double quotes) are accepted, so drag-and-drop from a file manager works on macOS. Multimodal commands require a multimodal projector to be loaded — pass --mmproj at startup or use /mmproj <path> from the REPL.

Web Application

cd TensorSharp.Server/bin

# Start the server with the exact hosted model
./TensorSharp.Server --model ./models/model.gguf --backend ggml_metal

# Linux + NVIDIA GPU
./TensorSharp.Server --model ./models/model.gguf --backend ggml_cuda

# Multimodal models: host an explicit projector too
./TensorSharp.Server --model ./models/model.gguf --mmproj ./models/mmproj.gguf --backend ggml_cuda

# Configure server-wide default sampling parameters
# (used whenever a request does not override the value itself)
./TensorSharp.Server --model ./models/model.gguf --backend ggml_metal \
    --temperature 0.7 --top-p 0.9 --top-k 40 --repeat-penalty 1.1 \
    --presence-penalty 0.0 --frequency-penalty 0.0 --seed 42 \
    --stop "</s>" --stop "<|endoftext|>"

Open http://localhost:5000 in your browser. The web interface supports:

• Multi-turn chat conversations
• Per-tab chat sessions: each browser tab owns its own KV cache; clicking "New Chat" disposes the current session server-side so its cache is released
• A single hosted GGUF selected explicitly with --model
• An explicit hosted multimodal projector via --mmproj when needed
• Image, video, and audio uploads for multimodal inference (up to 500 MB)
• Thinking/reasoning mode toggle
• Tool calling with function definitions
• Streaming token generation via Server-Sent Events
• Request queue with real-time position feedback
• Message editing and deletion with regeneration from any point in the conversation
• Free scrolling: scroll up to read earlier replies while new tokens stream in; the chat auto-scrolls again as soon as the user scrolls back to the bottom

Use --model to choose the hosted GGUF file and --mmproj to choose the hosted projector. TensorSharp.Server no longer scans a MODEL_DIR.

Server command-line options:

Option	Description
--model <path>	GGUF file to host (required for inference; if omitted, the server starts but /api/models/load will report no hosted model)
--mmproj <path>	Multimodal projector GGUF (resolved relative to the model directory when only a filename is given; pass none to disable). Requires --model.
--backend <type>	Default compute backend: cpu, cuda, mlx, ggml_cpu, ggml_metal, or ggml_cuda
--max-tokens <N>	Default maximum tokens to generate when a request omits the limit (default: 20000)
--temperature <f>	Default sampling temperature when a request does not provide one (0 = greedy)
--top-k <N>	Default top-K filtering when a request does not provide one (0 = disabled)
--top-p <f>	Default nucleus sampling threshold when a request does not provide one (1.0 = disabled)
--min-p <f>	Default min-p filtering when a request does not provide one (0 = disabled)
--repeat-penalty <f>	Default repetition penalty when a request does not provide one (1.0 = none)
--presence-penalty <f>	Default presence penalty when a request does not provide one (0 = disabled)
--frequency-penalty <f>	Default frequency penalty when a request does not provide one (0 = disabled)
--seed <N>	Default random seed when a request does not provide one (-1 = non-deterministic)
--stop <string>	Default stop sequence (can be repeated). Per-request stop/stop_sequences fully replace the default list rather than merge with it.
--continuous-batching / --no-continuous-batching	Enable (default) or disable iteration-level paged-batching. When enabled the server admits / preempts sequences mid-batch and packs them into one forward pass on models that implement IBatchedPagedModel. --no-continuous-batching falls back to per-sequence KV-swap for every model. Alias: --paged-batching / --no-paged-batching.
--paged-kv / --no-paged-kv	Force enable or disable the vLLM-style paged KV cache for the active session. When enabled the KV blocks live in a global block pool with prefix-cache sharing. Aliases: --paged-kv-cache / --no-paged-kv-cache.
--paged-kv-block-size <N>	Tokens per paged KV block (default: 256). Smaller blocks share more aggressively but pay more bookkeeping.
--paged-kv-ram-mb <N>	Soft cap for the paged-block RAM working set in megabytes. Blocks beyond the cap spill to SSD when --paged-kv-ssd-dir is set.
--paged-kv-ssd-dir <dir>	Directory used as the SSD cold tier for paged blocks. Optional but recommended for very large multi-session workloads.
--paged-kv-ssd-mb <N>	Maximum SSD usage in megabytes for the cold tier.
--paged-kv-quant-bits <0|4|8>	Optional KV block quantization (TurboQuantKvCodec). 0 (default) keeps blocks in their native dtype; 4 / 8 halve / quarter the per-block bandwidth at small numerical cost. Recurrent-state models silently fall back to passthrough.

Per-request fields in the chat / generate JSON payloads (e.g. temperature, top_p, top_k, min_p, repeat_penalty, presence_penalty, frequency_penalty, seed, stop/stop_sequences) always win over these server-wide defaults; the defaults only fill in fields the client omits.

Runtime environment variables:

Variable	Description
BACKEND	Default compute backend (cpu, cuda, mlx, ggml_cpu, ggml_metal, or ggml_cuda), used when --backend is not passed (default: ggml_metal on macOS, ggml_cpu elsewhere)
MAX_TOKENS	Default maximum generation length when neither --max-tokens nor a request-level limit is set (default: 20000)
MAX_TEXT_FILE_CHARS	Character cap used to truncate plain-text uploads when no tokenizer is available (default: 8000)
VIDEO_MAX_FRAMES	Maximum evenly spaced video frames extracted for video prompts (default: 4)
PORT / ASPNETCORE_URLS	Standard ASP.NET Core listener configuration (default port: 5000)
TENSORSHARP_TEMPERATURE	Default sampling temperature when neither --temperature nor the request body sets one
TENSORSHARP_TOP_K	Default top-K when neither --top-k nor the request body sets one
TENSORSHARP_TOP_P	Default top-P when neither --top-p nor the request body sets one
TENSORSHARP_MIN_P	Default min-P when neither --min-p nor the request body sets one
TENSORSHARP_REPEAT_PENALTY	Default repetition penalty when neither --repeat-penalty nor the request body sets one
TENSORSHARP_PRESENCE_PENALTY	Default presence penalty when neither --presence-penalty nor the request body sets one
TENSORSHARP_FREQUENCY_PENALTY	Default frequency penalty when neither --frequency-penalty nor the request body sets one
TENSORSHARP_SEED	Default random seed when neither --seed nor the request body sets one
TENSORSHARP_LOG_LEVEL	Minimum log level for both console and file loggers: Trace, Debug, Information, Warning, Error, Critical (default: Information). Also honored by TensorSharp.Cli.
TENSORSHARP_LOG_DIR	Directory the JSON-line file logger writes to (default: <binDir>/logs). Also honored by TensorSharp.Cli.
TENSORSHARP_LOG_FILE	Set to 0 to disable the file logger and keep only the console output (default: enabled). Also honored by TensorSharp.Cli.

Paged KV cache & continuous-batching tunables (read at process / model start)

These can be set with either the --paged-kv* / --continuous-batching CLI flags (which translate to the env vars below) or directly via the environment:

Variable	Description
TS_KV_PAGED_CACHE	1 / 0 to force-enable / disable the paged KV cache for the active session. The CLI shortcuts are --paged-kv / --no-paged-kv.
TS_KV_BLOCK_SIZE	Tokens per paged KV block (default: 256).
TS_KV_CACHE_MAX_RAM_MB	Soft cap for the paged-block RAM working set in megabytes.
TS_KV_CACHE_SSD_DIR	Directory used as the SSD cold tier for paged blocks.
TS_KV_CACHE_MAX_SSD_MB	Maximum SSD usage in megabytes for the cold tier.
TS_KV_PAGED_QUANT_BITS	KV block quantization bits (0 = passthrough, 4, or 8).
TS_SCHED_DISABLE_BATCHED	1 forces the per-sequence KV-swap fallback even when a model implements IBatchedPagedModel. The CLI shortcut is --no-continuous-batching.
TS_SCHED_MAX_BATCHED_TOKENS	Scheduler per-step token budget (default: 4096).
TS_SCHED_MAX_RUNNING_SEQS	Maximum in-flight sequences (default: 16).
TS_SCHED_PREFILL_CHUNK	Maximum prefill tokens per step (default: 1024).
TS_SCHED_NUM_BLOCKS	Physical blocks in the engine block pool (default: 256).
TS_SCHED_BLOCK_SIZE	Tokens per block on the engine side (default: 256).
TS_SCHED_PREFIX_CACHE	0 disables block-hash prefix sharing across requests.
TS_SCHED_DECODE_QUANTUM	Tokens before a sequence-switch is allowed (default: block size).
TS_QWEN35_BATCHED	Set to 0 to force the Qwen 3.5/3.6 family onto the legacy per-sequence KV-swap path (default: batched/paged). Also implicitly disabled by --no-continuous-batching.
TS_QWEN35_BATCHED_GDN_NATIVE	Use the native batched GatedDeltaNet kernel inside Qwen 3.5/3.6 batched path.
TS_GEMMA4_BATCHED	Set to 0 to force Gemma 4 onto the legacy per-sequence KV-swap path (default: batched/paged).
TS_GPTOSS_BATCHED	Set to 0 to force GPT OSS onto the legacy per-sequence KV-swap path (default: batched/paged).
TS_GPTOSS_PAGED_ATTN_MANAGED	Use the managed (C#) paged-attention-with-sinks kernel inside GPT OSS batched path.
TS_NEMOTRON_BATCHED	Set to 0 to force Nemotron-H onto the legacy per-sequence KV-swap path (default: batched/paged).
TS_NEMOTRON_MAMBA2_BATCHED_NATIVE	Use the native Mamba2 batched step kernel inside Nemotron-H batched path.
TS_PAGED_ATTN_KERNEL	Paged-attention dispatch kernel for Mistral3Model.BatchedForward: native (default), tensor (C# Tensor-based), or managed (pure C# scalar).
TS_MLX_PIPELINED_DECODE	Set to 1 to enable pipelined greedy decode on the MLX backend (CLI only).
TS_MLX_MLOCK_GGUF	1 (default) pins the GGUF mmap region in physical RAM via mlock(2) so model weights stay resident between forward passes. Set to 0 to skip (use if the process memlock rlimit is too low or you want the OS to manage paging). MLX backend only.
TS_MLX_FUSED_KV_WRITE	1 (default) uses a single multi-dim slice_update to write the per-token KV block. Set to 0 to revert to the per-head loop (A/B testing / regression isolation).
TS_MLX_BATCHED_MOE_DECODE	1 (default) collapses K per-expert decode dispatches to one batched dispatch per (gate/up/down) kind for Qwen 3.5/3.6 MoE. Set to 0 on memory-constrained machines (saves ~weight-doubling overhead from the stacked weight slabs).
TS_MLX_MOE_FUSED_GATE_UP_SILU	1 (default) fuses gate matmul + up matmul + SiLUMul into one Metal kernel for batched MoE decode. Set to 0 to A/B against the legacy 3-dispatch path.
TS_MLX_DEVICE_ROUTER	1 (opt-in) keeps MoE router top-K + softmax on device to skip ~60 host syncs/token on Qwen 3.6-35B-A3B. Requires greedy router + batched MoE matmul.
TS_MLX_LOG_MEMORY_POLICY	1 (default) prints once-per-load MLX memory-policy lines (wired limit, GGUF mlock status, allocator caps). Set to 0 to silence.
TS_MLX_MEMORY_LIMIT_MB / TS_MLX_CACHE_LIMIT_MB / TS_MLX_WIRED_LIMIT_MB	Override the MLX allocator hard cap / unused-buffer cache cap / wired-buffer residency cap (megabytes). Defaults are derived from the host's unified-memory capacity.
TS_MLX_EVAL_EVERY_N_LAYERS / TS_MLX_GEMMA4_EVAL_EVERY_N_LAYERS	Periodic mlx_async_eval cadence during decode to overlap GPU work with host queueing. Default 4 (sweep on E4B Q8_0 shows ~7% decode win vs. disabled). Set to 0 to disable.
TENSORSHARP_MLX_LIBRARY / TENSORSHARP_MLX_LIBRARY_DIR	Override the search path for libmlxc when using --backend mlx.

Sampling parameter precedence (highest wins):

1. Per-request JSON fields in the API call (e.g. temperature, top_p, stop).
2. Server-wide CLI flags (e.g. --temperature, --top-p, --stop).
3. TENSORSHARP_* environment variables listed above.
4. Built-in SamplingConfig defaults (temperature=1.0, top_k=0, top_p=1.0, min_p=0, repeat_penalty=1.0, presence/frequency penalties 0, seed=-1, no stop sequences).

Feature × environment variable matrix

Quick reference for which environment variables (and matching CLI flags) gate each major feature. Variables in bold are required to turn the feature on; everything else is a tunable for a feature that's already enabled by default.

Continuous batching & paged KV cache

Feature	Default	Env vars	CLI equivalent
Continuous-batching engine (InferenceEngine + scheduler)	ON in TensorSharp.Server	TS_SCHED_DISABLE_BATCHED=1 to force per-seq fallback	--no-continuous-batching / --continuous-batching
Paged KV cache for the active session	ON	TS_KV_PAGED_CACHE (0 / 1), TS_KV_BLOCK_SIZE	--paged-kv / --no-paged-kv, --paged-kv-block-size N
Paged KV SSD spillover (cold tier)	OFF	TS_KV_CACHE_MAX_RAM_MB, TS_KV_CACHE_SSD_DIR, TS_KV_CACHE_MAX_SSD_MB	--paged-kv-ram-mb, --paged-kv-ssd-dir, --paged-kv-ssd-mb
Paged KV block quantization (TurboQuantKvCodec)	OFF (0 = passthrough)	TS_KV_PAGED_QUANT_BITS (0 / 4 / 8)	--paged-kv-quant-bits
Block-hash prefix sharing across requests	ON	TS_SCHED_PREFIX_CACHE=0 to disable	—
Scheduler tunables (per-step token budget, max in-flight seqs, prefill chunk, block pool size, decode quantum)	engine defaults	TS_SCHED_MAX_BATCHED_TOKENS, TS_SCHED_MAX_RUNNING_SEQS, TS_SCHED_PREFILL_CHUNK, TS_SCHED_NUM_BLOCKS, TS_SCHED_BLOCK_SIZE, TS_SCHED_DECODE_QUANTUM	—

Per-model batched / paged forward (IBatchedPagedModel.ForwardBatch)

Model	Default state	Env var to flip default	Native-kernel sub-toggle
Mistral 3	ON	—	TS_PAGED_ATTN_KERNEL = native (default) / tensor / managed
Gemma 4	ON	TS_GEMMA4_BATCHED=0 to force legacy per-seq	—
Qwen 3	ON (reference port)	—	—
Qwen 3.5 / 3.6 family	ON	TS_QWEN35_BATCHED=0 to force legacy per-seq (or --no-continuous-batching)	TS_QWEN35_BATCHED_GDN_NATIVE=1 enables native batched GDN kernel; FUSED_ATTN_LAYER_MIN_SEQ_LEN=N overrides fused-attention engage threshold (default 4096)
GPT OSS	ON	TS_GPTOSS_BATCHED=0 to force legacy per-seq	TS_GPTOSS_PAGED_ATTN_MANAGED=1 forces the managed (C#) sinks softmax instead of the native paged-attention-with-sinks kernel
Nemotron-H	ON	TS_NEMOTRON_BATCHED=0 to force legacy per-seq	TS_NEMOTRON_MAMBA2_BATCHED_NATIVE=1 enables the native batched Mamba2 step (NEON SIMD + GCD parallelism)
Gemma 3	not implemented (per-seq fallback)	—	—

Backends

Feature	Default	Env vars	CLI equivalent
Default compute backend	ggml_metal (macOS), ggml_cpu (Windows/Linux)	BACKEND	--backend
MLX backend library lookup	probe app dir	TENSORSHARP_MLX_LIBRARY (full path to libmlxc), TENSORSHARP_MLX_LIBRARY_DIR (directory)	—
MLX pipelined greedy decode (CLI only)	OFF	TS_MLX_PIPELINED_DECODE=1	—
MLX mlock(2) of GGUF mmap so weights stay resident	ON	TS_MLX_MLOCK_GGUF=0 to disable	—
MLX fused multi-dim KV write (single slice_update per cache block)	ON	TS_MLX_FUSED_KV_WRITE=0 to revert to per-head loop	—
MLX batched MoE decode (Qwen 3.5/3.6 MoE)	ON	TS_MLX_BATCHED_MOE_DECODE=0 for legacy per-expert path	—
MLX fused MoE gate+up+SiLUMul Metal kernel	ON	TS_MLX_MOE_FUSED_GATE_UP_SILU=0 for legacy 3-dispatch	—
MLX on-device MoE router top-K + softmax	OFF	TS_MLX_DEVICE_ROUTER=1	—
MLX Gemma 4 layer-boundary async_eval cadence	every 4 layers	TS_MLX_GEMMA4_EVAL_EVERY_N_LAYERS=N (0 = disabled)	—
MLX allocator caps (memory / cache / wired buffer)	host-derived	TS_MLX_MEMORY_LIMIT_MB, TS_MLX_CACHE_LIMIT_MB, TS_MLX_WIRED_LIMIT_MB	—
MLX one-line memory-policy banners at load	ON	TS_MLX_LOG_MEMORY_POLICY=0 to silence	—

Sampling defaults (server-only)

These fill in fields the request body omits; per-request JSON always wins, CLI flags win over env vars.

Sampling field	Env var	CLI equivalent
temperature	TENSORSHARP_TEMPERATURE	--temperature
top_k	TENSORSHARP_TOP_K	--top-k
top_p	TENSORSHARP_TOP_P	--top-p
min_p	TENSORSHARP_MIN_P	--min-p
repeat_penalty	TENSORSHARP_REPEAT_PENALTY	--repeat-penalty
presence_penalty	TENSORSHARP_PRESENCE_PENALTY	--presence-penalty
frequency_penalty	TENSORSHARP_FREQUENCY_PENALTY	--frequency-penalty
seed	TENSORSHARP_SEED	--seed
max tokens	MAX_TOKENS	--max-tokens
stop sequences	— (CLI / per-request only)	--stop (repeatable)

Hosting & uploads (server-only)

Feature	Default	Env vars
ASP.NET Core listener	http://0.0.0.0:5000	PORT, ASPNETCORE_URLS
Plain-text upload character cap (when no tokenizer available)	8000 chars	MAX_TEXT_FILE_CHARS
Video-frame extraction count	4 frames	VIDEO_MAX_FRAMES

Logging (server + CLI)

Feature	Default	Env vars	CLI equivalent
Console + file log minimum level	Information	TENSORSHARP_LOG_LEVEL	--log-level
File logger output directory	<binDir>/logs	TENSORSHARP_LOG_DIR	--log-dir
File logger enabled	ON	TENSORSHARP_LOG_FILE=0 to disable	--log-file 0|1
Console logger enabled	ON	—	--log-console 0|1 (CLI only)

Native build (compile-time only)

These are read by build-linux.sh / build-windows.ps1 / the auto-build during dotnet build for TensorSharp.GGML.Native, not at run time.

Feature	Default	Env vars	Build-script flag
Enable GGML CUDA in the native build	auto-detected from toolchain	TENSORSHARP_GGML_NATIVE_ENABLE_CUDA=ON	--cuda / --no-cuda
Narrow CMAKE_CUDA_ARCHITECTURES list	auto-detected from visible GPU	TENSORSHARP_GGML_NATIVE_CUDA_ARCHITECTURES	--cuda-arch='86-real;89-real'
Native build parallelism cap	conservative auto-cap	TENSORSHARP_GGML_NATIVE_BUILD_PARALLEL_LEVEL	—

Server Logging

The server emits one structured Information-level entry at the start and end of every chat / generate turn, so a single grep over the log file reproduces the full request-response audit trail without replaying any traffic.

Event id	Emitted on	Carries
ChatStarted (1500)	chat.start, generate.start, plus per-protocol request banners	sampling config, message + attachment counts, userInput= (full latest user message), fullInput= (JSON-encoded array of EVERY message in the request: system prompts + all prior user/assistant turns + the new user message, with attachment counts), or the full prompt for /api/generate
ChatCompleted (1502)	chat.complete, generate.complete	token counts, KV cache reuse (kvReused, kvReusePercent), TTFT, elapsed, throughput, finish reason, full raw assistant output (reasoning + result)
ChatAborted (1503)	client disconnected mid-stream	partial output, KV reuse fraction at the time of abort
KvCacheReusePlan (1510)	per-prefix-reuse decision	Debug-level fine-grained breakdown (exact match / partial / full reset)
HttpRequestStarted/Completed (1100/1101)	every HTTP request	method, path, remote IP, status, duration; /api/queue/status is demoted to Debug so high-frequency UI polling does not drown out the per-turn entries

The raw assistant output captures <think>...</think>, <|channel|>analysis, and any other inline framing the model emits, so the log line for a single turn contains both reasoning and the user-visible result. Combined with the fullInput= field on chat.start, every turn is fully reproducible from the log file alone (request inputs + raw model output). Long uploads or long reasoning traces can produce multi-kilobyte log lines; raise the log level (TENSORSHARP_LOG_LEVEL=Warning) to suppress them while still keeping the start banner and error logs.

Sample fullInput payload (formatted for readability; it is emitted as a single line in the actual log):

[
  {"role":"system","content":"You are a helpful assistant."},
  {"role":"user","content":"What is the tallest mountain?"},
  {"role":"assistant","content":"Mount Everest."},
  {"role":"user","content":"How tall is it?","images":1}
]

The same per-turn KV cache reuse stats are surfaced through every API:

• Web UI SSE (POST /api/chat) - the done event carries promptTokens, kvReusedTokens, and kvReusePercent.
• Ollama NDJSON (POST /api/generate, POST /api/chat/ollama) - the final chunk and the non-streaming response carry prompt_cache_hit_tokens (int) and prompt_cache_hit_ratio (0..1).
• OpenAI (POST /v1/chat/completions) - the usage block carries prompt_tokens_details.cached_tokens, matching the OpenAI extension that existing SDKs already understand.

The Web UI footer line under each assistant message also surfaces the cache hit inline (e.g. 187 tokens · 2.1s · 87.2 tok/s · KV 420/512 (82%)).

HTTP APIs

TensorSharp.Server exposes three API styles. See API_EXAMPLES.md for full documentation with curl and Python examples.

Ollama-compatible API:

# List models
curl http://localhost:5000/api/tags

# Generate text
curl -X POST http://localhost:5000/api/generate \
  -H "Content-Type: application/json" \
  -d '{"model": "Qwen3-4B-Q8_0.gguf", "prompt": "Hello!", "stream": false}'

# Chat
curl -X POST http://localhost:5000/api/chat/ollama \
  -H "Content-Type: application/json" \
  -d '{"model": "Qwen3-4B-Q8_0.gguf", "messages": [{"role": "user", "content": "Hi"}], "stream": false}'

# Chat with thinking mode
curl -X POST http://localhost:5000/api/chat/ollama \
  -H "Content-Type: application/json" \
  -d '{"model": "Qwen3-4B-Q8_0.gguf", "messages": [{"role": "user", "content": "Solve 17*23"}], "think": true, "stream": false}'

# Chat with tool calling
curl -X POST http://localhost:5000/api/chat/ollama \
  -H "Content-Type: application/json" \
  -d '{"model": "Qwen3-4B-Q8_0.gguf", "messages": [{"role": "user", "content": "What is the weather?"}], "tools": [{"function": {"name": "get_weather", "description": "Get current weather", "parameters": {"properties": {"city": {"type": "string"}}, "required": ["city"]}}}], "stream": false}'

OpenAI-compatible API:

# Chat completions
curl -X POST http://localhost:5000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "Qwen3-4B-Q8_0.gguf", "messages": [{"role": "user", "content": "Hi"}], "max_tokens": 50}'

# Structured outputs (OpenAI response_format)
curl -X POST http://localhost:5000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen3-4B-Q8_0.gguf",
    "messages": [{"role": "user", "content": "Extract the city and country from: Paris, France."}],
    "response_format": {
      "type": "json_schema",
      "json_schema": {
        "name": "location_extraction",
        "strict": true,
        "schema": {
          "type": "object",
          "properties": {
            "city": {"type": "string"},
            "country": {"type": "string"},
            "confidence": {"type": ["string", "null"]}
          },
          "required": ["city", "country", "confidence"],
          "additionalProperties": false
        }
      }
    }
  }'

OpenAI Python SDK:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:5000/v1", api_key="not-needed")
response = client.chat.completions.create(
    model="Qwen3-4B-Q8_0.gguf",
    messages=[{"role": "user", "content": "What is 2+3?"}],
    max_tokens=50
)
print(response.choices[0].message.content)

Queue status:

curl http://localhost:5000/api/queue/status
# {"busy":false,"pending_requests":0,"total_processed":42}

Thinking / Reasoning Mode

Models that support thinking mode (Qwen 3, Qwen 3.5/3.6-family, Gemma 4, GPT OSS, Nemotron-H) can produce structured chain-of-thought reasoning before generating the final answer. The thinking content is separated from the main response and can be displayed or hidden by the client.

• Qwen 3 / Qwen 3.5/3.6-family / Nemotron-H: uses <think>...</think> tags
• Gemma 4: uses <|channel>thought\n...<channel|> tags
• GPT OSS: uses Harmony format with <|channel|>analysis for thinking and <|channel|>final for the response

Enable via --think (console), "think": true (Ollama API), or the thinking toggle in the web UI.

Tool Calling / Function Calling

Models can invoke user-defined tools and participate in multi-turn tool-call conversations. Define tools as JSON and pass them via --tools (console) or the tools parameter in the API.

Each architecture uses its own wire format for tool calls:

• Qwen 3 / Qwen 3.5/3.6-family / Nemotron-H: <tool_call>{"name": "...", "arguments": {...}}</tool_call>
• Gemma 4: <|tool_call>call:function_name{args}<tool_call|>

The output parser (OutputParser.cs) automatically extracts tool calls from the model's raw output regardless of architecture.

Multimodal Support

Gemma 4

Gemma 4 models support image, video, and audio inputs. Place the multimodal projector (gemma-4-mmproj-F16.gguf) in the same directory as the model file for automatic loading.

• Images: PNG, JPEG, HEIC/HEIF
• Video: MP4 (extracts up to 8 frames at 1 fps using OpenCV)
• Audio: WAV (16kHz mono), MP3, OGG Vorbis

Gemma 3

Gemma 3 supports PNG, JPEG, and HEIC/HEIF image inputs. Place its multimodal projector (mmproj-gemma3-4b-f16.gguf) next to the model file for automatic loading.

Qwen 3.5 / 3.6 family

All Qwen 3.5/3.6-family variants (qwen35, qwen35moe, and qwen3next) load through the same Qwen35Model implementation. Image inputs are supported via the dynamic-resolution Qwen35VisionEncoder; place the projector (Qwen3.5-mmproj-F16.gguf) next to the model GGUF for automatic loading. The MoE variants (e.g. Qwen3.5-35B-A3B and Qwen3.6-35B-A3B GGUFs that report the same architecture keys) additionally enable a fused MoEExpertsSwiGLUResidual GGML kernel during decode that runs all selected experts, the optional shared expert, and the residual add in a single GPU graph dispatch.

Mistral 3

Mistral 3 supports image inputs via the Pixtral vision encoder. Place the multimodal projector (mistral3-mmproj.gguf) in the same directory as the model file for automatic loading.

• Images: PNG, JPEG, HEIC/HEIF

Nemotron-H (Omni distribution)

The Nemotron Omni distribution adds a RADIO / v2_vl ViT image encoder. Pass the matching multimodal projector with --mmproj (e.g. nvidia_Nemotron-H-Omni-mmproj.gguf); the language-model GGUF stays the same. Image tokens are inserted at <image> placeholders and expanded into <img> + N tile tokens + </img> automatically by the multimodal injector.

• Images: PNG, JPEG, HEIC/HEIF
• Audio: the chat template emits <so_embedding> per uploaded audio file and the CLI runs the Parakeet-style log-mel preprocessor for verification, but actual audio inference requires a Parakeet audio mmproj that the public GGUFs do not currently ship.

Architecture

TensorSharp is structured as a layered system:

1. TensorSharp.Core provides the core Tensor type, storage abstraction, and the extensible operation registry (Ops). CPU implementations use System.Numerics.Vectors for SIMD acceleration.

2. TensorSharp.Runtime owns runtime-facing contracts and services: GGUF parsing, tokenization (SentencePiece / BPE), chat template rendering, configurable token sampling, output parsing, paged KV cache (Runtime/Paged/*), the continuous-batching scheduler / engine (Runtime/Scheduling/*), the IKvBlockCodec interface plus the TurboQuantKvCodec Q4/Q8 implementation, and reusable contracts such as IModelArchitecture, IBatchedPagedModel, IPromptRenderer, IOutputProtocolParser, IMultimodalInjector, IKVCachePolicy, and IBackendExecutionPlan.

3. TensorSharp.Models implements ModelBase plus the concrete architectures and multimodal helpers (Gemma 3/4, Qwen 3/3.5, GPT OSS, Nemotron-H, Mistral 3). Each architecture ships both the legacy per-sequence forward and an IBatchedPagedModel.ForwardBatch implementation (<Family>Model.BatchedForward.cs) for continuous batching. Models are loaded via ModelBase.Create() which auto-detects the architecture from GGUF metadata.

4. TensorSharp.Backends.GGML registers accelerated implementations of the same operations via a native C++ bridge (libGgmlOps / GgmlOps.dll) that links against ggml. On macOS this provides Metal GPU compute, and on Windows/Linux it can expose GGML CUDA for NVIDIA GPUs. Operations include native quantized matmul (Q4_K_M, Q8_0, etc.) without dequantizing to FP32, plus paged-attention (TSGgml_PagedAttentionForward, with and without attention sinks) and architecture-specific batched kernels (Mamba2, GatedDeltaNet).

5. TensorSharp.Backends.Cuda is the direct CUDA path. It uses the CUDA Driver API for device/context/storage management, cuBLAS for float32 GEMM, PTX kernels for hot scalar and transformer helper ops, and CPU fallbacks where native kernels are not implemented yet.

6. TensorSharp.Backends.MLX is the Apple Silicon MLX path. It wraps mlx-c (libmlxc) with allocator, storage, async worker dispatch, quantized + fused + compiled kernels, MoE expert offload, and a CPU fallback layer for ops that aren't yet wired up.

7. TensorSharp.Server is the HTTP/application layer. It provides Ollama-compatible and OpenAI-compatible REST APIs, the browser-based chat UI, upload handling, an InferenceEngineHost that owns the per-model continuous-batching engine, and a thin queue-status surface for backward compatibility.

8. TensorSharp.Cli is the console/application layer for local prompts, multimodal experiments, prompt inspection, JSONL batch workflows, the interactive REPL, and the built-in prefill / decode benchmarks.

Performance Optimizations

The list below is the cross-architecture summary; each per-model card under docs/models/ walks through the same kernels in context, with the exact GGML graph dispatched and the conditions under which the fused path engages.

• Fused GPU decode (Gemma 4): all transformer layers are executed in a single GGML compute graph dispatch on Metal, reducing CPU-GPU round-trips from hundreds per token to one. This achieves ~2.6x speedup over per-operation dispatch.
• Fused GPU prefill (Gemma 4): for dense (non-MoE, non-shared, non-PLE/multimodal) layers, Gemma4LayerPrefill runs the entire transformer block (RMSNorm + QKV + QK-norm + RoPE + attention + output projection + post-attn norm + GeGLU FFN + post-FFN norm + residual + layer scalar) as a single GGML graph dispatch per layer during prefill, extending the fused approach from decode to multi-token prefill.
• Chunked prefill (Gemma 4): long prompts are split into bounded chunks (2x sliding window, max 2048 tokens) to avoid O(n^2) attention score tensors for SWA layers. Chunking is applied automatically when text-only (no multimodal embeddings) and keeps each chunk within the SWA window budget.
• Native whole-model decode (Qwen 3): all transformer layers run in one native call (TransformerModelDecode) with pre-resolved per-layer weight pointers cached at load time, removing managed-loop overhead from the decode hot path.
• Fused Qwen 3.5/3.6-family attention layer decode: a single GGML graph performs RMSNorm + fused QKV + Q/gate deinterleave + per-head QK norm + RoPE + KV cache append + flash attention + sigmoid-gated mix + output projection + residual add for each FullAttention layer. Replaces ~2 standalone GGML calls and ~6 small CPU/GPU sync points per attention layer. Engages once the cached sequence length exceeds 4096 tokens (override with FUSED_ATTN_LAYER_MIN_SEQ_LEN=N).
• Fused prefill attention (Qwen 3.5/3.6-family): FusedPrefillAttention combines Q*K^T, causal mask, softmax, and *V into a single GGML graph dispatch during multi-token prefill, eliminating ~5 separate C#-to-GGML round-trips per attention layer. Handles both initial prefill and continuation with existing KV cache entries.
• Fused output-projection + FFN (Qwen 3.5/3.6-family): for both FullAttention and GatedDeltaNet layers with dense FFN, FusedOutProjFFN merges the output projection, residual add, post-attention RMSNorm, and the full SwiGLU FFN (gate_up matmul + SiLU + down matmul + residual) into a single GGML graph dispatch, reducing two GPU round-trips to one per layer.
• Fused output-projection + norm + router (Qwen 3.5/3.6-family MoE): FusedOutProjNormRouter merges the GatedDeltaNet output projection, residual add, post-attention RMSNorm, and MoE router projection into one dispatch. The pre-computed router logits are then consumed directly by the batched MoE kernel, eliminating a separate router dispatch per MoE layer.
• Fused vision encoder (Qwen 3.5/3.6-family): FusedVisionAttention merges LayerNorm + QKV + bias + 2D RoPE + scaled dot-product attention + output projection + bias + residual into one GGML graph dispatch (~8 ops → 1). FusedVisionMLP merges LayerNorm + up + bias + GELU + down + bias + residual into one dispatch (7 ops → 1). Combined, these cut the per-block GPU round-trips from ~15 to 2.
• Fused weight projections: Q/K/V projections are fused into a single QKV matmul; gate and up projections are fused into a single gate_up matmul.
• Native quantized compute: quantized weights (Q4_K_M, Q6_K, Q8_0, IQ2_XXS, MXFP4, etc.) are used directly in matmul without expanding to FP32, saving memory and bandwidth. A batched AddmmQuantBatch kernel handles multiple sub-weight matmuls against a single quantized blob in one dispatch.
• Direct CUDA kernels: the cuda backend accelerates fill/copy, unary ops, activation fusions, RMSNorm, softmax, index select, causal masking, RoPE/RoPEEx, cuBLAS GEMM, and supported quantized matmul/get-rows while safely falling back for incomplete op coverage.
• Batched GPU MoE: MoEExpertsSwiGLUResidual (Qwen 3.5/3.6-family) and MoEExpertsForward (Nemotron-H) collapse all selected experts -- and, for Qwen 3.5/3.6-family, the optional shared expert and the residual add -- into a single GGML graph dispatch per MoE layer.
• GEMM-based vision patch embedding (Qwen 3.5/3.6-family): the patch embedding step is reformulated as parallel im2col + matrix multiplication, replacing a single-threaded scalar quintuple-nested loop with a GPU-accelerated matmul.
• Parallelized Q/gate deinterleave (Qwen 3.5/3.6-family): the Q + sigmoid-gate deinterleave in FullAttention prefill is parallelized across tokens, scaling linearly with CPU core count for long prompts.
• Optimized pure C# CPU path: managed GEMM fast paths and contiguous float32 kernels accelerate decode, softmax, RMSNorm, RoPE, fused activations, and other hot paths while keeping quantized GGUF weights compressed during CPU loading.
• Circular KV cache: sliding-window attention layers use a fixed-size circular buffer, bounding memory usage regardless of sequence length.
• KV-cache prefix reuse: multi-turn conversations reuse the longest matching token prefix across turns. Truncation is automatically backed off by the sliding-window size for SWA models so the suffix can rebuild the SWA context.
• Paged KV cache & block-hash prefix sharing: the continuous-batching engine partitions KV into fixed-size blocks, content-hashes each full block, and shares them across concurrent and sequential requests. Combined with a per-tier (RAM → SSD) PagedKvBlockStore, this gives vLLM-style memory efficiency without giving up the legacy per-session contiguous path.
• Native paged-attention kernel: TSGgml_PagedAttentionForward (and the WithSinks variant for GPT OSS) does a C++ gather of K/V from the paged buffer, builds a small GGML graph per sequence, and dispatches ggml_flash_attn_ext — the same fused Metal/CUDA flash-attention kernel the legacy single-sequence path uses. On Ministral-3-14B long-context (4×~800 tokens) it is ~21 % faster than the legacy per-sequence GGML path.
• Batched / paged forward passes: Mistral 3, Gemma 4, GPT OSS, Qwen 3.5/3.6 (incl. GatedDeltaNet recurrent state pool), and Nemotron-H (incl. Mamba2 recurrent state pool + native batched Mamba2 kernel) pack N sequences into a single ForwardBatch call with one batched linear-projection matmul per layer, paged K/V scatter via slotMapping, and per-sequence attention via the native kernel. Gemma 4 batched path reaches 1.5× legacy throughput at batch=8 short prompts and 1.6× at 4×800-token prompts; Nemotron-H Mamba2 batched reaches 3.95× at batch=3 on Apple M4 Pro. See docs/PAGED_ATTENTION_AND_CONTINUOUS_BATCHING.md.
• Kernel warmup: both CLI and Server run a tiny forward pass at startup to pre-compile GPU kernels (Metal pipeline states, CUDA JIT) and warm the memory pool, avoiding cold-start latency on the first real inference request.
• Prefill caching (Gemma 4, Qwen 3.5/3.6-family): per-forward-pass SWA mask cache (Gemma 4), NeoX RoPE cos/sin lookup table cache across global layers (Gemma 4), and RoPE position tensor cache across layers (Gemma 4, Qwen 3.5/3.6-family) eliminate redundant recomputation during prefill.
• In-place QK RMSNorm (Qwen 3.5/3.6-family): per-head QK normalization is performed in-place using a View, avoiding one tensor allocation and copy per Q/K per layer.

Memory Optimizations

• Zero-copy file-mapped quantized weights (direct CUDA, GGML CUDA, GGML Metal, GGML CPU): the GGUF model file is memory-mapped and quantized tensors are bound directly into native ops via host-pointer buffers. This removes the per-tensor copy from disk into a freshly-allocated native heap buffer that previously roughly doubled the resident set on Apple Silicon for large quantized models. For example, Qwen3.5-35B-A3B-IQ2_XXS (~10 GB GGUF) now runs with ~7 GB peak working memory under Metal instead of ~17 GB. The OS keeps the mapped file in its page cache and pages it out under memory pressure without any inference penalty on Apple Silicon (unified memory).
• Best-fit memory pool: the GGML host allocator uses a best-fit search across pooled blocks instead of first-fit, which avoids handing out a large scratch block to satisfy a tiny intermediate-tensor request and keeps the working-set tightly bounded across long-running inference.
• Bounded pool retention: the integrated-GPU / CPU memory pool now caps individual retained blocks at 64 MB and the total pool at 32 blocks. Combined with mmap-backed weights, this keeps short-lived intermediate tensors recycled fast while bounding the peak resident set.
• Memory-efficient model loading: large tensors are streamed directly to native memory without intermediate managed allocations. F32 weights and norms still load on demand; quantized weights are mmap-backed when supported by the backend.
• Paged KV block pool with optional SSD spillover: paged KV blocks live in a per-engine BlockPool with LRU eviction; the PagedKvBlockStore keeps a configurable RAM cap (TS_KV_CACHE_MAX_RAM_MB) and spills cold blocks into an SSD tier (TS_KV_CACHE_SSD_DIR) up to TS_KV_CACHE_MAX_SSD_MB. Block content-hashes are kept in a global index so prefix matches are reused across sessions and requests without rematerialising the K/V.
• KV block codecs: blocks can be optionally compressed in-place with TurboQuantKvCodec (Q4 or Q8) via --paged-kv-quant-bits, trading a small accuracy cost for half / quarter the per-block bandwidth and memory footprint. Recurrent-state models fall back to passthrough automatically.

Benchmarks

Internal regression baseline

Reference numbers measured on Qwen3.6-35B-A3B-UD-IQ2_XXS.gguf (~10 GB on disk, 256 routed experts of which 8 are active per token, with 12 full attention + 30 GatedDeltaNet recurrent layers) on an Apple M4 Pro with 24 GB unified memory:

Metric	Before (v1 baseline)	After (this branch)	Change
Process peak memory footprint	~17 GB	~8 GB	-52%
TensorSharp.Server resident set after load	~20 GB	~8 GB	-60%
Decode throughput (warm, 256 prefill / 64 decode, M4 Pro)	~3.8 tok/s	~10.8 tok/s	+2.85x
Decode latency (warm, 256 prefill / 64 decode, M4 Pro)	~264 ms/token	~92 ms/token	-65%

Reproduce with:

./TensorSharp.Cli --model Qwen3.6-35B-A3B-UD-IQ2_XXS.gguf --backend ggml_metal \
    --benchmark --bench-prefill 256 --bench-decode 64 --bench-runs 3

The memory reduction comes primarily from no longer copying the GGUF file into a separate native heap buffer (the file is now mmap-bound zero-copy into Metal command buffers). The decode throughput increase is largely a side effect of removing that ~10 GB duplicate working set, which was previously triggering OS-level memory pressure on machines with 24 GB or less of physical RAM.

Cross-engine inference matrix

For an apples-to-apples comparison of TensorSharp, llama.cpp, and Ollama on the same on-disk GGUF files (Gemma 4 E4B Q8_0 today, with text / synthetic prefill / image / audio / video tasks and KV-cache dtype sweeps for f32, f16, and q8_0), see docs/inference_benchmark_matrix.md. The driver scripts are in benchmarks/inference_matrix/scripts/ and the per-cell raw JSON outputs live under benchmarks/inference_matrix/results/.

Testing

Unit tests (xUnit)

InferenceWeb.Tests exercises in-process behavior that doesn't require a running server: managed quantized ops, direct CUDA backend kernels when a CUDA device is available, MLX backend kernels when MLX is available, paged KV cache scheduling (ContinuousBatchSchedulerTests, PagedKvCacheTests, PagedKvCacheCodecTests), batched executor correctness (BatchedExecutorTests), per-model batched-forward correctness against the legacy path (Qwen35BatchedCorrectnessTests, Mistral3BatchedForwardTests, Gemma4BatchedForwardTests, GptOssBatchedCorrectnessTests, NemotronBatchedCorrectnessTests), per-model batched perf microbenchmarks (*BatchedPerfBench.cs), TurboQuantKvCodec codec round-trips, prefill chunking, KV cache policies, KV-cache prompt rendering / multi-turn integration, chat-session and session-manager isolation, model service history and KV cache plumbing, request-logging middleware and file-logger provider, image preprocessing, media helpers, structured-output validation, text-upload helpers, model-service upload logging, web UI chat policy, model context length parsing, backend catalog resolution, and the server CLI options builder (ServerOptionsBuilderTests).

dotnet test InferenceWeb.Tests/InferenceWeb.Tests.csproj

Server integration tests

Integration tests for TensorSharp.Server are in TensorSharp.Server/testdata/. They cover all three API styles (Web UI SSE, Ollama, OpenAI), multi-turn conversations, thinking mode, tool calling, structured outputs, queue behavior, concurrent requests, and abort support. Architecture-specific features (thinking, tool calling) are auto-detected and skipped when the active model does not support them.

# Start TensorSharp.Server, then run:
python3 TensorSharp.Server/testdata/test_multiturn.py
# or
bash TensorSharp.Server/testdata/test_multiturn.sh

See TensorSharp.Server/testdata/README.md for the full test matrix.

Author

Zhongkai Fu

License

See LICENSE for details.