dotllm 0.3.0-dev.2973

This is a prerelease version of dotllm.

There is a newer prerelease version of this package available.
See the version list below for details.

dotnet tool install --global dotllm --version 0.3.0-dev.2973

This package contains a .NET tool you can call from the shell/command line.

dotnet new tool-manifest
                    

                            if you are setting up this repo

dotnet tool install --local dotllm --version 0.3.0-dev.2973

This package contains a .NET tool you can call from the shell/command line.

#tool dotnet:?package=dotllm&version=0.3.0-dev.2973&prerelease

The NuGet Team does not provide support for this client. Please contact its maintainers for support.

nuke :add-package dotllm --version 0.3.0-dev.2973

The NuGet Team does not provide support for this client. Please contact its maintainers for support.

dotLLM

High-performance LLM inference engine written natively in C#/.NET

Website · Documentation · Roadmap · Discussions

</div>

About

dotLLM is a ground-up LLM inference engine for .NET — not a wrapper around llama.cpp or Python libraries. All orchestration, model loading, tokenization, sampling, and CPU compute are implemented in pure C#, with CUDA GPU acceleration via PTX kernels loaded through the CUDA Driver API (no native shared library). It targets transformer and SSM/MoE model families including Llama, Mistral, Phi, Qwen, DeepSeek-V2/V3, Mixtral, Granite-MoE, Nemotron-H, and Mamba-3 with SIMD-optimized CPU plus CUDA and Vulkan GPU backends.

Status: Active post-Mamba-3 development — speculative decoding, paged KV-cache allocation, Native AOT (experimental), startup warm-up, constrained decoding (JSON/schema/regex/grammar), tool calling, prompt caching, logprobs, LoRA adapter loading, CUDA CPU/GPU hybrid offload, and native Vulkan real-weight parity for multiple architectures. See docs/SUPPORTED_MODELS.md for the current support matrix.

Key Features

Performance

Zero-GC inference — unmanaged memory (NativeMemory.AlignedAlloc, 64-byte aligned) for all tensor data; no managed heap allocations on the hot path
SIMD vectorization — TensorPrimitives + hand-tuned System.Runtime.Intrinsics for quantized matmul, RMSNorm, RoPE, softmax
Memory-mapped model loading — GGUF files loaded via MemoryMappedFile; OS demand-paging means multi-GB models load in milliseconds
Quantized inference — FP16/BF16, Q8_0, K-quants, and other GGUF quantization formats; fused scale×int dot-product kernels operating directly on quantized blocks where implemented

Architecture Support

Transformer/SSM models — dense transformers (incl. ternary BitNet b1.58), MLA/MoE DeepSeek variants, MoE families, hybrid Nemotron-H, and pure-SSM Mamba-3. See docs/SUPPORTED_MODELS.md for the full matrix.
Attention mechanisms — MHA, MQA, GQA, sliding-window masks, and DeepSeek MLA via parameterized ModelConfig, with IAttentionStrategy for kernel selection
Position encoding — RoPE variants and Mamba-3 data-dependent RoPE plumbing, with a pluggable IPositionEncoding surface for additional encodings
Composable sampling — ISamplerStep chain: repetition penalty → temperature → top-k → top-p → min-p → categorical sample

Serving

OpenAI-compatible API — /v1/chat/completions, /v1/completions, tool calling, streaming via ASP.NET
Paged KV-cache — PagedAttention with block-level allocation, prefix caching, and copy-on-write
Speculative decoding — draft-verify-accept with KV-cache rollback (greedy mode today; non-greedy planned — see issue #121)
Structured output — FSM/PDA-based constrained decoding guaranteeing valid JSON, JSON Schema, regex, and grammar
(Planned) Continuous batching — iteration-level scheduling with preemption and priority queuing — Phase 9, see Roadmap

Extensibility

Pluggable backends — IBackend interface with separate packages per backend (CPU, CUDA, Vulkan, ROCm/HIP foundation)
Diagnostic hooks — zero-cost IInferenceHook points for activation capture, logit lens, SAE integration
LoRA adapters — runtime loading, no weight merging, CPU/Vulkan application paths, and concurrent multi-adapter serving
OpenTelemetry observability — System.Diagnostics.Metrics + Activity for throughput, latency, and per-request tracing, OTLP exporter wired on the server side

Architecture Overview

dotLLM is organized as a layered architecture where each layer depends only on the layers below it:

┌─────────────────────────────────────────┐
│            DotLLM.Server                │  ASP.NET OpenAI-compatible API
├─────────────────────────────────────────┤
│            DotLLM.Engine                │  KV-cache, scheduler, samplers,
│                                         │  constraints, speculative decoding
├──────────┬──────────┬───────────────────┤
│ DotLLM.  │ DotLLM.  │ DotLLM.Cpu/Cuda   │  GGUF/SafeTensors, BPE/SPM,
│ Models   │Tokenizers│ (backends)        │  SIMD kernels / CUDA kernels
├──────────┴──────────┴───────────────────┤
│            DotLLM.Core                  │  Interfaces, tensor types, config
└─────────────────────────────────────────┘

Each layer is its own project/package (the .nupkgs are attached to every GitHub release; see NuGet Packages for what is on nuget.org). DotLLM.Core defines all abstractions (ITensor, IBackend, IModel, ISamplerStep, etc.) while concrete implementations live in their respective projects.

Getting Started

There are two paths: grab a pre-built release and run it, or clone the repo and build from source.

Use a pre-built release

Pick one of three install options. Builds are published from the dev branch as GitHub prereleases (0.3.0-dev.<n>); there is no stable release yet, so the NuGet and dnx options need --prerelease. See docs/PACKAGING.md for the release channels.

Option A — run or install as a .NET tool (needs the .NET 10 SDK, not just the runtime: dnx and dotnet tool are SDK commands):

# Run without installing (downloads and caches the NuGet package `dotllm`)
dnx dotllm --prerelease -- run QuantFactory/SmolLM-135M-GGUF -p "The capital of France is" -n 64
dnx dotllm --prerelease -- serve QuantFactory/SmolLM-135M-GGUF   # OpenAI-compatible API + chat UI

# ...or install the `dotllm` command globally
dotnet tool install -g dotllm --prerelease
dotllm run QuantFactory/SmolLM-135M-GGUF -p "The capital of France is" -n 64
dotllm serve QuantFactory/SmolLM-135M-GGUF

Notes on dnx (checked against the .NET 10 SDK):

Arguments after the package id are forwarded to the tool, but dnx claims --help/-h and --version for itself: dnx dotllm --prerelease serve --help prints dnx's help, while dnx dotllm --prerelease -- serve --help prints dotllm serve's. Put -- before the tool's arguments whenever they include those flags (it is harmless otherwise).
--yes (-y) is accepted and skips the confirmation dnx may ask for before running a package for the first time; use it in scripts. No prompt appeared when run from a non-interactive shell.
Pin a version with dnx dotllm@0.3.0-dev.2470 -- ... for reproducibility.
The package id is dotllm (lower case; the tool command is also dotllm). Do not install DotLLM.Cli: that id belongs to the upstream project and is frozen at 0.1.0-preview.3.

Option B — download a self-contained binary (no .NET install needed — the runtime is bundled):

Grab the archive for your platform from the fork's releases page (pick the newest dotLLM 0.3.0-dev.<n>; prereleases are never marked "latest", so releases/latest will not find them):

Windows x64: dotllm-<version>-win-x64.zip (also contains dotllm-tray.exe, the system tray)
Linux x64: dotllm-<version>-linux-x64.tar.gz
macOS (Apple Silicon): dotllm-<version>-osx-arm64.tar.gz

Every release also carries a SHA256SUMS file and a build-provenance attestation. The builds are not Authenticode-signed, so Windows SmartScreen will warn; verify the download instead:

sha256sum -c SHA256SUMS --ignore-missing
gh attestation verify dotllm-<version>-linux-x64.tar.gz --repo jamesburton/dotLLM

Unpack and run:

# Linux / macOS
tar -xzf dotllm-<version>-linux-x64.tar.gz
cd dotllm-<version>-linux-x64
./dotllm run QuantFactory/SmolLM-135M-GGUF -p "The capital of France is" -n 64
./dotllm serve QuantFactory/SmolLM-135M-GGUF             # OpenAI-compatible API + chat UI

# Windows
Expand-Archive dotllm-<version>-win-x64.zip -DestinationPath .
cd dotllm-<version>-win-x64
.\dotllm.exe run QuantFactory/SmolLM-135M-GGUF -p "The capital of France is" -n 64

Tagged (v*) releases additionally attach experimental Native AOT builds (dotllm-<version>-aot-<rid>.{zip,tar.gz}) — smaller and faster to start, but please file an issue if you hit a crash.

Option C — reference the libraries from your .NET app — see NuGet Packages below.

Build from source

Clone the repository and build with the .NET 10 SDK.

Prerequisites:

.NET 10 SDK
Python 3.10+ with pip install rich InquirerPy — only needed for the benchmark scripts under scripts/
Optional: llama.cpp for comparison benchmarks (see llama.cpp setup)

git clone https://github.com/kkokosa/dotLLM.git
cd dotLLM
dotnet build -c Release

When built from source, replace dotllm <subcommand> in the Usage examples below with dotnet run --project src/DotLLM.Cli -c Release -- <subcommand>.

Usage

The dotllm CLI has these commands:

dotllm model — pull, add, list, search, inspect and delete GGUF models; create and manage named model profiles; import ollama models
dotllm run — single-shot text generation with a performance summary
dotllm chat — interactive multi-turn REPL with chat template formatting
dotllm serve — OpenAI-compatible HTTP API (plus Anthropic- and ollama-compatible routes) with a built-in web chat UI
dotllm ps / dotllm stop — list / unload the models of a running server
dotllm bench / dotllm perplexity — throughput and quality measurement

Naming and fetching models

One resolver turns a model name into a GGUF in run, chat, serve, bench, and in a request's model field:

you write	resolves to
`C:\models\x.gguf`	that file
`owner/repo`	the repo's best local GGUF (Q4_K_M first, not the largest file)
`owner/repo:Q8_0`	the file whose name contains the tag (same as `--quant Q8_0`)
`llama3.2:3b`	a model in your local ollama store, loaded in place (see ollama)
`my-profile`	a model profile

Missing models are downloaded for you by dotllm run|chat|serve owner/repo[:tag] (resumable, with progress, into the Hugging Face hub cache). The server only downloads on demand if started with --auto-pull — otherwise a remote client could trigger multi-gigabyte downloads just by naming a repo. dotllm model pull fetches ahead of time. Models already in ~/.dotllm/models/ or the Hugging Face hub cache (HF_HUB_CACHE, ~/.cache/huggingface/hub) are found without re-downloading. Details: docs/SERVER.md.

Manage models

# Search HuggingFace for GGUF repos
dotllm model search llama --limit 5

# Download ahead of time (a tag picks the file: owner/repo:Q4_K_M)
dotllm model pull QuantFactory/SmolLM-135M-GGUF

# Multi-GGUF repo: name the file. Without --file (or a :tag) pull lists the files
# and needs an interactive terminal to choose one, so scripts/CI must pass it.
dotllm model pull TheBloke/Llama-2-7B-GGUF --file llama-2-7b.Q4_K_M.gguf

# List everything available locally
dotllm model list

# Show architecture, quantizations, and tokenizer info for a repo
dotllm model info QuantFactory/SmolLM-135M-GGUF

# Remove a downloaded repo (--quant narrows to one file) or a profile
dotllm model delete QuantFactory/SmolLM-135M-GGUF

Custom models and profiles

# Register your own GGUF (hard-linked, not copied) or download one from a URL; listed as local/<name>
dotllm model add C:\models\my-finetune.gguf --name my-finetune

# A profile (Modelfile equivalent): base model + system prompt + sampling defaults + device + keep-alive
dotllm model create terse --from bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M \
    --system "Answer in one sentence." --temperature 0.2 --device vulkan --keep-alive 600

dotllm model show terse            # merged settings and where the base model resolves
dotllm model cp terse terse-hot    # copy a profile, or alias a model under a new name

Profiles live in ~/.dotllm/profiles/ (DOTLLM_PROFILES_DIR overrides). A profile's name is a model id for the server ("model": "terse"), which applies its system prompt and defaults to requests that omit them. dotllm run|chat resolve a profile to its base model but do not yet apply its system prompt or sampling. model create options: --from/-f, --system, --system-file, --temperature, --top-p, --top-k, --min-p, --repeat-penalty, --max-tokens, --seed, --stop (repeatable), --device/-d, --gpu-layers, --keep-alive, --description, --force.

Use your ollama models

dotllm run llama3.2:3b -p "Hello"            # loaded in place from your ollama store, nothing copied
dotllm model import-ollama                   # turn the whole store into profiles (system prompt + parameters)
dotllm model pull ollama:llama3.2:3b         # or pull from registry.ollama.ai (sha256-verified)

dotLLM only reads the ollama store, never writes to it; ollama's Go prompt templates are not carried over (the chat template embedded in the GGUF is used). See docs/SERVER.md.

Run — single-shot generation

Encodes a prompt, streams tokens to stdout, and prints a performance + memory summary.

# Greedy generation (default: temperature=0, max-tokens=128)
dotllm run QuantFactory/SmolLM-135M-GGUF -p "The capital of France is" -n 64

# Sampled generation
dotllm run QuantFactory/SmolLM-135M-GGUF -p "Once upon a time" -n 128 -t 0.7 --top-k 40 --top-p 0.95

# JSON output (for scripting / piping)
dotllm run QuantFactory/SmolLM-135M-GGUF -p "Hello" --json

# Long prompts: read from a file instead of the command line
dotllm run QuantFactory/SmolLM-135M-GGUF --prompt-file prompt.txt -n 256

# Select a specific quantization when a repo has multiple .gguf files
dotllm run QuantFactory/SmolLM-135M-GGUF -p "Test" -q Q8_0

# GPU inference: CUDA (NVIDIA GPU) ...
dotllm run QuantFactory/SmolLM-135M-GGUF -p "The capital of France is" --device gpu

# ... or Vulkan (AMD / Intel / NVIDIA; the GPU path on AMD iGPUs such as Strix Halo)
dotllm run QuantFactory/SmolLM-135M-GGUF -p "The capital of France is" --device vulkan

# NUMA / P-core aware CPU threading
dotllm run QuantFactory/SmolLM-135M-GGUF -p "Test" --threads 8 --decode-threads 4 --numa-pin

# KV-cache quantization (Q8_0 / Q4_0) to fit longer contexts in memory
dotllm run QuantFactory/SmolLM-135M-GGUF -p "..." --cache-type-k q8_0 --cache-type-v q8_0

# Constrained JSON output
dotllm run QuantFactory/SmolLM-135M-GGUF -p "List 3 colors as JSON." --response-format json_object

# Speculative decoding — target + draft must share the same vocabulary.
# Example pair (validated by scripts/test_models_speculative.py):
#   target: Llama-3.2-3B-Instruct Q8_0     draft: Llama-3.2-1B-Instruct Q4_K_M
dotllm model pull bartowski/Llama-3.2-3B-Instruct-GGUF
dotllm model pull bartowski/Llama-3.2-1B-Instruct-GGUF
dotllm run bartowski/Llama-3.2-3B-Instruct-GGUF -q Q8_0 \
    -p "Explain what a CPU cache is in one sentence." -n 48 \
    --speculative-model bartowski/Llama-3.2-1B-Instruct-GGUF --speculative-k 5

Sample output:

── dotllm | Llama 30L/576H | Q8_0 | 16 threads | greedy ──────────────────
The capital of France is Paris. Paris is a city of romance and culture,

╭──────────────────────────────────────────────────────────────────────────╮
│                                                                          │
│  Generation Complete                                      163.27 tok/s   │
│                                                                          │
│  Performance                                                             │
│  Prefill            12.3 ms       6 tokens       487.80 tok/s            │
│  Decode             91.8 ms      15 tokens       163.40 tok/s            │
│  Sampling            0.1 ms      15 tokens                               │
│  ──────────────────────────────────────────────────────────              │
│  Total             104.2 ms      21 tokens       201.54 tok/s            │
│  Load              456.7 ms                                              │
│                                                                          │
│  Memory                                                                  │
│  Weights         136.73 MiB      (memory-mapped)                         │
│  Compute           2.25 MiB                                              │
│  KV Cache        158.20 MiB      (192 slots)                             │
│  ──────────────────────────────────────────────────────────              │
│  Total           297.18 MiB                                              │
│                                                                          │
│  length | 6 prompt, 15 generated                                         │
╰──────────────────────────────────────────────────────────────────────────╯

Chat — interactive REPL

Multi-turn chat with persistent history, using the model's built-in chat template (falling back to ChatML). Prompt caching reuses KV-cache state across turns so subsequent turns skip redundant prefill.

# Basic chat
dotllm chat QuantFactory/SmolLM-135M-GGUF

# With a system prompt and sampling
dotllm chat QuantFactory/SmolLM-135M-GGUF --system "You are a helpful assistant." -t 0.8 --top-p 0.95

# GPU + KV-cache quantization for long contexts
dotllm chat bartowski/Llama-3.2-3B-Instruct-GGUF --device gpu --cache-type-k q8_0 --cache-type-v q8_0

chat takes the same --device values as run (auto default, cpu, vulkan, gpu/cuda); on Vulkan it uses the model's own device KV-cache (paged/quantized KV and the prompt prefix cache are off there, as in serve).

In-session commands: /exit or /quit to leave, /clear to reset history (keeps the system prompt), /system <text> to change the system prompt.

Sample session:

── dotllm chat | Llama 30L/576H | Q8_0 | 16 threads | greedy ─────────────
Type /exit to quit, /clear to reset history, /system <text> to set system prompt.

>>> Hello, how are you?
I'm doing well, thank you for asking! How can I help you today?
[42 prompt tokens, 18 generated tokens, 28 ms TTFT, 487.8 prefill tok/s, 163.4 decode tok/s]

>>> What is 2+2?
2 + 2 = 4.
[78 prompt tokens, 12 generated tokens, 45 ms TTFT, 312.5 prefill tok/s, 155.2 decode tok/s]

>>> /clear
History cleared.
>>> /exit

Serve — OpenAI-compatible API + chat UI

Starts a local HTTP server exposing an OpenAI-compatible API (/v1/chat/completions, /v1/completions, /v1/models, /v1/tokenize, streaming SSE, tool calling) plus a built-in single-page web chat UI. Paged KV-cache, prompt caching, and startup warm-up are on by default. The browser opens automatically unless --no-browser is set.

--device defaults to auto for serve: for each model it tries CUDA (if a GPU is present and the model fits), then Vulkan, then the CPU, and a load that fails on one device falls through to the next. Pass cpu, gpu:N/cuda:N (CUDA) or vulkan to force one (an explicit GPU that cannot load the model is an error, not a CPU server). dotllm run/chat/bench use the same auto default and the same loader. Idle models unload after --keep-alive seconds (default 300).

# Start the server with a loaded model and open the chat UI
dotllm serve QuantFactory/SmolLM-135M-GGUF

# Bind a public interface, custom port, API-only, no auto-browser
dotllm serve QuantFactory/SmolLM-135M-GGUF --host 0.0.0.0 --port 9000 --no-ui --no-browser

# Start without a model — pick one from the chat UI
dotllm serve

# Force Vulkan, and let clients name a not-yet-downloaded model ("owner/repo:Q4_K_M") to fetch it
dotllm serve --device vulkan --auto-pull

# CUDA with partial hybrid offload and more warm-up iterations
dotllm serve bartowski/Llama-3.2-3B-Instruct-GGUF --device gpu --gpu-layers 24 --warmup-iterations 5

# Speculative decoding — draft must share the target's vocabulary
dotllm serve bartowski/Llama-3.2-3B-Instruct-GGUF -q Q8_0 \
    --speculative-model bartowski/Llama-3.2-1B-Instruct-GGUF --speculative-k 5

Thinking models (Qwen3.x, Bonsai, DeepSeek-R1 style) are handled for you: the <think> span is split into message.reasoning_content (OpenAI), a thinking content block (/v1/messages) or message.thinking (/api/chat), and enable_thinking, chat_template_kwargs and reasoning_effort (Anthropic thinking, ollama think) reach the chat template. --reasoning-format none|auto|deepseek (default auto) picks the behaviour — none returns the raw output as before; details in docs/SERVER.md.

Any OpenAI-compatible client works against the running server:

curl -N http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "SmolLM-135M",
    "messages": [{"role":"user","content":"Say hi in one word."}],
    "stream": true
  }'

Other API surfaces

Route(s)	What
`POST /v1/messages`, `/v1/messages/count_tokens`	Anthropic Messages API (docs/ANTHROPIC_API.md)
`POST /v1/embeddings`	embeddings (docs/SERVER.md)
`/api/chat`, `/api/generate`, `/api/tags`, `/api/ps`, `/api/show`, `/api/version`, `/api/pull`, `/api/delete`	ollama-compatible API — point Open WebUI, Continue, ollama-python/js etc. at the server (`OLLAMA_HOST=http://localhost:8080`). `/api/embed`, `/api/embeddings`, `/api/create`, `/api/copy`, `/api/push` answer `501` with a pointer to the dotLLM equivalent; `tools` in `/api/chat` is `501` (use `/v1/chat/completions`)
`POST /v1/systemone`	Jev-compatible decision endpoint: typed questions in, per-option probabilities out, one prefill and no decode loop (built for Tev1-style decision models; tune with `--decision-temperature` / `--decision-orderings`)
`POST /v1/models/{unload,pull,enable,disable}`, `PUT /v1/settings`, `POST /v1/admin/shutdown`	management API, off by default; enabled by `--allow-model-admin`. `dotllm ps`, `dotllm stop [model\\|--all\\|--server]` and the Windows tray are clients of it. `--allow-lora-admin` separately gates runtime LoRA load/unload

Reference: docs/SERVER.md. The Windows system tray app (starts/attaches to a server, autostart, update check) is documented in docs/TRAY.md.

To embed the same endpoints inside your own ASP.NET Core app, see Host the OpenAI API in your ASP.NET app below.

CLI option reference

Common options (shared by run, chat, and serve):

Option	Short	Default	Description
`--device`	`-d`	`auto` (`run`, `chat`, `bench`, `serve`); `cpu` (`perplexity`, so scores stay comparable)	`auto` (CUDA if the model fits, else Vulkan, else CPU - a CPU result always prints a warning); `cpu`; `vulkan`; `gpu`/`cuda`, `gpu:1`/`cuda:1` (CUDA). An explicit GPU that cannot be honoured is an error (what/why/model vs device memory, and `--device cpu` to opt in) - never a silent CPU run. Every command prints `device: requested -> resolved`.
`--gpu-layers`		(all if `gpu`, 0 if `cpu`)	Transformer layers on GPU (hybrid offload)
`--rope-scaling`, `--rope-freq-base`, `--rope-scale`, `--yarn-*`		(from GGUF)	RoPE / YaRN overrides (`run`, `serve`)
`--prefill-chunk-size`		0	Max prompt tokens per prefill forward pass (0 = whole prompt)
`--threads`		0 (auto)	CPU threads for inference
`--decode-threads`		0 (auto)	Decode threads (capped at memory channels)
`--numa-pin`		false	Pin workers to NUMA-local cores (multi-socket)
`--pcore-only`		false	Pin workers to P-cores only (Intel hybrid)
`--quant`	`-q`	(auto)	Quant filter when a repo has multiple `.gguf` files (e.g., `Q4_K_M`)
`--cache-type-k`		`f32`	KV-cache key quant: `f32`, `q8_0`, `q4_0`
`--cache-type-v`		`f32`	KV-cache value quant: `f32`, `q8_0`, `q4_0`
`--kv-quant-k` / `--kv-quant-v`		(unset)	Synonyms of `--cache-type-k/-v` (`f16`, `q8_0`, `q4_0`); on `run`, `bench` and `serve`. The only way to select quantized KV for the CUDA hybrid-dense model (Bonsai-2 / qwen35), where they override `DOTLLM_HYBRID_KV`. See docs/KV_CACHE.md
`--speculative-model`		(none)	Draft model for speculative decoding (must share vocab)
`--speculative-k`		3	Draft tokens per speculative step
`--mtp-head`		(auto)	`serve`: path of a separate MTP head GGUF (qwen4exp `mtp-*.gguf`); an auto-detected sibling is attached when it fits the device. See docs/SERVER.md
`--no-mtp`		off	`serve`: opt out of default-on MTP self-speculation (models with an embedded MTP head, expected concurrency < 5, no `--speculative-model`). `--mtp` is accepted for back-compat. See docs/SERVER.md

Sampling & constraints (shared by run and chat):

Option	Short	Default (`run`)	Default (`chat`)	Description
`--max-tokens`	`-n`	128	512	Max tokens per generation
`--temp`	`-t`	0 (greedy)	0 (greedy)	Sampling temperature
`--top-k`		0 (off)	0 (off)	Top-K sampling
`--top-p`		1.0	1.0	Nucleus threshold
`--min-p`		0 (off)	0 (off)	Min-P threshold
`--repeat-penalty`		1.0	1.0	Repetition penalty
`--repeat-last-n`		0 (full)	0 (full)	Penalty lookback window
`--seed`	`-s` (run only)	(random)	(random)	Random seed for reproducibility
`--cache-window`		0	0	Full-precision tail window for KV quant
`--paged`		off	off	Use paged (block-based) KV-cache
`--response-format`		`text`	`text`	`text`, `json_object`, `json_schema`, `regex`, `grammar`
`--schema`		—	—	JSON Schema (or `@file.json`) for `json_schema`
`--pattern`		—	—	Regex pattern for `regex`
`--grammar`		—	—	GBNF grammar (or `@file.gbnf`) for `grammar`
`--tools`		—	—	Tool definitions JSON (or `@file.json`)

run-only:

--prompt / -p — input prompt (required unless --prompt-file is given)
--prompt-file — read the prompt from a file instead of --prompt. A single trailing newline is stripped. Mutually exclusive with --prompt
--json — emit a single JSON result object (suppresses formatted output)

chat-only:

--system / -s — system prompt
--system-file — read the system prompt from a file instead of --system. A single trailing newline is stripped. Mutually exclusive with --system
--tool-choice — auto (default), none, required, or a function name
--no-prompt-cache — disable KV-cache reuse across turns
--prompt-cache-size — max cached sessions (default: 1)
--verbose / -v — debug output (finish reason, raw text, tool-call details)

serve-only:

Option	Short	Default	Description
`--host`		`localhost`	Bind address
`--port`	`-p`	`8080`	Port to listen on
`--no-ui`		false	Disable the built-in chat UI (API only)
`--no-browser`		false	Don't auto-open the browser
`--no-paged`		false	Disable paged KV-cache (paged is on by default for `serve`)
`--no-prompt-cache`		false	Disable KV-cache reuse across requests
`--prompt-cache-size`		`4`	Max cached sessions
`--no-warmup`		false	Disable startup warm-up passes
`--warmup-iterations`		`3`	Warm-up iteration count
`--auto-pull`		false	Download a missing Hugging Face / ollama model when a request names one (off by default so remote clients cannot start downloads)
`--keep-alive`		`300`	Idle-unload after N seconds (0 = unload after each request, negative = never)
`--max-resident-models`		`1`	Models resident at once (>1 keeps several loaded)
`--resident-memory-budget`		`0` (unlimited)	Total byte budget across resident models
`--allow-model-admin`		false	Enable the management API (`/v1/models/*`, `/v1/settings`, `/v1/admin/shutdown`, ollama `/api/pull` + `/api/delete`)
`--allow-lora-admin`		false	Enable `POST /v1/lora/load` and `DELETE /v1/lora/{name}`
`--mtp`		(default on)	Accepted for back-compat. `serve` auto-enables MTP self-speculation for models with an embedded MTP head when expected concurrency is < 5 and no `--speculative-model` is set; `--no-mtp` opts out
`--max-messages`		`8192`	Max messages per chat request on `/v1/chat/completions` and `/v1/messages`; `0` = unlimited; env `DOTLLM_MAX_MESSAGES` (flag wins)
`--expected-concurrency`		`0`	At 5+ concurrent requests Vulkan hybrid models serve through the continuous-batch scheduler
`--decision-temperature`, `--decision-orderings`		`0`, `1`	Calibration of `POST /v1/systemone` probabilities
`--rate-limit-rpm`, `--rate-limit-tpm`, `--rate-limit-concurrency`		`0` (off)	Per-API-key rate limits; setting any enables limiting

bench, perplexity: dotllm bench <model> [-d cpu|vulkan|cuda] [-p 512] [-n 128] [-r 5] [--depth N] [--json] times prefill/decode (llama-bench equivalent); dotllm perplexity <model> --corpus wiki.test.raw [--context 512] [--stride N] [-d ...] scores a corpus (docs/PERPLEXITY.md). bench defaults to --device auto; perplexity defaults to --device cpu (pass -d auto|vulkan|cuda to score on a GPU).

ps, stop: clients of a running server (--url, or DOTLLM_URL; default http://localhost:8080). dotllm ps lists resident models; dotllm stop <model> / --all unloads, --server stops the whole server gracefully. Both need --allow-model-admin on the server for the write actions.

Short-flag gotcha: -p is prompt under run but port under serve. -s is seed under run but system under chat. When in doubt, use the long form.

Development

Debug build

Building in Debug configuration (-c Debug) enables a debug command group with diagnostic tools for inspecting GGUF files and model internals. These commands are excluded from Release builds via #if DEBUG.

# Build in Debug mode
dotnet build src/DotLLM.Cli -c Debug

# Inspect GGUF file structure
dotnet run --project src/DotLLM.Cli -c Debug -- debug gguf-header model.gguf
dotnet run --project src/DotLLM.Cli -c Debug -- debug gguf-metadata model.gguf
dotnet run --project src/DotLLM.Cli -c Debug -- debug gguf-tensors model.gguf
dotnet run --project src/DotLLM.Cli -c Debug -- debug gguf-config model.gguf

# Tokenizer round-trip verification
dotnet run --project src/DotLLM.Cli -c Debug -- debug tokenize model.gguf --text "Hello world"

# Single forward pass with top-10 logit diagnostics
dotnet run --project src/DotLLM.Cli -c Debug -- debug forward-pass model.gguf --prompt "Hello"

# Inspect embedding vector for a token ID
dotnet run --project src/DotLLM.Cli -c Debug -- debug embed-lookup model.gguf --token-id 1

Command	Description
`debug gguf-header`	GGUF header structure (magic, version, tensor/metadata counts)
`debug gguf-metadata`	All metadata key-value pairs
`debug gguf-tensors`	Tensor descriptors (name, shape, quantization type, offset)
`debug gguf-config`	Extracted `ModelConfig` (architecture, layers, dims, RoPE params)
`debug tokenize`	Encode text → token IDs → decode, verify round-trip fidelity
`debug forward-pass`	Single forward pass, top-10 predicted tokens with softmax probabilities
`debug embed-lookup`	Raw embedding vector for a given token ID

Debug builds are significantly slower than Release (~2-10x) because JIT optimizations, inlining, and SIMD vectorization are reduced. Always use -c Release for performance measurements.

Tests

Unit and integration tests:

dotnet test

Integration tests automatically download several GGUF models (~4.5 GB total) from HuggingFace to ~/.dotllm/test-cache/ on first run. The first dotnet test will take a while; subsequent runs use the cache. To run only unit tests (no downloads): dotnet test tests/DotLLM.Tests.Unit.

Set DOTLLM_TEST_CACHE_DIR to relocate that cache — useful for pointing CI at a mounted cache volume so models aren't re-downloaded every job, or for moving it off a small system drive:

# Linux/macOS
export DOTLLM_TEST_CACHE_DIR=/mnt/ci-cache/dotllm-models

# Windows
$env:DOTLLM_TEST_CACHE_DIR = "D:\dotllm-test-cache"

The cache uses the same {owner}/{repo}/{file}.gguf layout as ~/.dotllm/models/, so you can point it at your CLI model directory to reuse models you already pulled:

$env:DOTLLM_TEST_CACHE_DIR = "$HOME\.dotllm\models"

Doing so means test models are downloaded into your dotllm model pull inventory and will appear in dotllm model list. That is intentional but opt-in — leave the variable unset to keep auto-downloaded test models separate from models you pulled explicitly. The value must be an absolute path (or a relative one, resolved against the working directory); ~ is not expanded.

GPU tests (tagged Category=GPU) require an NVIDIA GPU and run full model inference — they can take 20-30 minutes. They are skipped automatically on machines without CUDA. To exclude them explicitly: dotnet test tests/DotLLM.Tests.Unit/ --filter "Category!=GPU"

Model correctness smoke tests (scripts/test_models.py) run dotLLM CLI with greedy decoding across architectures (Llama, Mistral, Phi, Qwen) and verify expected output:

# Build CLI first
dotnet build src/DotLLM.Cli -c Release

# List available test cases and which models are cached
python scripts/test_models.py --list

# Run tests for all cached models
python scripts/test_models.py

# Download missing models and run all tests
python scripts/test_models.py --download

# Run only specific architectures
python scripts/test_models.py --filter phi,qwen

Models are downloaded from HuggingFace to ~/.dotllm/models/ on first use and cached for subsequent runs.

Sample output:

Test                                Arch       Result      Time  Details
=====================================================================================================
SmolLM-135M                         Llama      PASS        2.1s  Paris  (163.3 tok/s)
Llama-3.2-1B-Instruct-Q4            Llama      PASS        5.7s  Paris  (31.0 tok/s)
Qwen2.5-0.5B-Instruct               Qwen       PASS        3.2s  Paris  (78.5 tok/s)
Phi-3-mini-4k-instruct              Phi        PASS       12.4s  Paris  (14.2 tok/s)
=====================================================================================================

4/4 passed, 0 failed, 0 skipped

Benchmarks

Three scripts in scripts/ provide benchmarking at different levels:

bench_compare.py -- Single-point benchmark. Runs dotLLM (via BenchmarkDotNet) and optionally llama.cpp on one or more models, reports best-of-N throughput with CV (coefficient of variation):

# Benchmark dotLLM on SmolLM-135M (auto-downloads from HuggingFace)
python scripts/bench_compare.py --model QuantFactory/SmolLM-135M-GGUF --quant Q8_0

# Benchmark multiple models and quantizations
python scripts/bench_compare.py \
    --model QuantFactory/SmolLM-135M-GGUF,bartowski/Llama-3.2-1B-Instruct-GGUF \
    --quant Q4_K_M,Q8_0

# Compare dotLLM vs llama.cpp side-by-side
python scripts/bench_compare.py --model QuantFactory/SmolLM-135M-GGUF --dotllm --llamacpp

# Export results to JSON for later comparison
python scripts/bench_compare.py --model QuantFactory/SmolLM-135M-GGUF \
    --export-json benchmarks/results/baseline.json --label baseline

Sample output:

=== dotLLM Benchmark Results ===

  Model                  Prefill tok/s   Decode tok/s   Decode ms/tok   Total tok/s     CV
  SmolLM-135M.Q8_0             229.2          182.7           5.47         175.3      14.7%
  SmolLM-135M.Q4_K_M           165.0          230.1           4.35         198.2      20.5%

All values are best-of-N (max tok/s, min ms). CV is the coefficient of variation
across N iterations -- lower means more stable measurements.

bench_trend.py -- Interactive comparison of exported JSON results. Displays color-coded delta tables with noise-aware highlighting:

# Interactive mode: select runs and models to compare
python scripts/bench_trend.py

# Compare two specific result files
python scripts/bench_trend.py benchmarks/results/baseline.json benchmarks/results/optimized.json

# Show all results as a trend table
python scripts/bench_trend.py --all

Sample output (trend across three labeled runs):

                             Benchmark Trend
 Label      Date        Model               Prefill tok/s   Decode tok/s   CV
 baseline   2026-03-11  SmolLM-135M.Q4_K_M          127.6          109.5    -
 step29     2026-03-13  SmolLM-135M.Q4_K_M          142.2          127.0    -
 step30     2026-03-13  SmolLM-135M.Q4_K_M          146.0           98.6    -

bench_history.py -- Benchmark across git commits. Creates worktrees for each commit, runs bench_compare in each, and displays trend tables with per-commit deltas:

# Benchmark last 5 commits on main
python scripts/bench_history.py myrun --last 5

# Benchmark from a specific commit to HEAD
python scripts/bench_history.py myrun --from f3d3bf8

# Show results from a previous run (no benchmarking)
python scripts/bench_history.py myrun --show

# Interactively select which commits to benchmark
python scripts/bench_history.py myrun --last 10 --select

Sample output:

                     Benchmark History -- Llama-3.2-3B-Instruct-Q8_0
 Label                   Date        Prefill tok/s  %chg pf  Decode tok/s  %chg dc     CV
 test_run_0 (f3d3bf8)    2026-03-11          21.2                     7.4              3.8%
 test_run_1 (cdb5234)    2026-03-12          24.9    +17.5%           8.0    +8.1%     2.1%
 test_run_2 (a062743)    2026-03-13          24.5    ~-1.6%           7.8   ~-2.5%     4.5%
 test_run_3 (6c06fbf)    2026-03-14          24.6    ~+0.4%           7.8   ~+0.0%     3.2%
 test_run_4 (d1978d2)    2026-03-15          25.4     +3.3%           7.0   -10.3%     5.1%
 test_run_5 (572179d)    2026-03-16          25.5    ~+0.4%           7.8   +11.4%     4.2%

%chg columns show commit-to-commit deltas. ~ prefix means the change is within noise (CV threshold). CV requires multiple BenchmarkDotNet iterations (controlled by --runs in bench_compare).

Why best-of-N instead of median? On a non-isolated machine, run-to-run noise is typically 6--30%. The median includes runs degraded by OS scheduling jitter, thermal throttling, and background I/O. Best-of-N (maximum throughput) represents what the hardware can achieve and is more stable across sessions. CV is reported alongside so you can judge measurement quality -- if CV is high, the environment was noisy and even the best-of-N value should be taken with a grain of salt.

llama.cpp setup

To run comparison benchmarks against llama.cpp:

Get llama.cpp -- either download a prebuilt release or build from source:

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp && cmake -B build -DCMAKE_BUILD_TYPE=Release && cmake --build build --config Release

Point bench_compare to the binary -- either:
- Set LLAMACPP_BIN environment variable to the path of llama-completion (older builds name it llama-cli)
- Or pass --llamacpp-bin /path/to/llama-completion on each invocation
Without either, bench_compare looks for llama-completion on PATH.

Run comparison -- match the thread count:

python scripts/bench_compare.py --model QuantFactory/SmolLM-135M-GGUF --dotllm --llamacpp --threads 16

Always pass --threads when comparing engines. Their defaults are different: dotLLM auto-selects all logical cores, llama.cpp auto-selects physical cores. On a 32-logical/16-physical machine that is 32 vs 16, which makes the comparison meaningless. bench_compare warns when you omit it.

Comparing .NET runtimes

--runtimes runs dotLLM on several target frameworks in one comparison, each as a separate BenchmarkDotNet job with its own result row:

python scripts/bench_compare.py --model QuantFactory/SmolLM-135M-GGUF \
  --dotllm --llamacpp --threads 16 --runtimes net10.0,net11.0

Decode Throughput (tok/s)                               * = fastest (highest)
Model                Quant      llama.cpp  dotLLM (net10.0)  dotLLM (net11.0)
SmolLM-135M          Q8_0          304.6     360.3    417.0*

Each requested framework needs an SDK able to target it -- benchmarking net11.0 requires the .NET 11 SDK installed, otherwise the job fails with NETSDK1045. global.json uses rollForward: latestMajor with allowPrerelease, so machines with only the .NET 10 SDK keep resolving to it and the default (host runtime only) is unaffected.

Do not set DOTNET_ROLL_FORWARD to benchmark a newer runtime. BenchmarkDotNet generates its toolchain project to match the host runtime, so rolling the host forward makes it emit a project the pinned SDK cannot build. --runtimes is the supported mechanism.

llama.cpp is optional. All dotLLM benchmarks work without it. The --llamacpp flag simply adds a side-by-side comparison column.

NuGet Packages

The fork publishes one package to nuget.org:

Package	Description
`dotllm`	`dotnet tool` — the `dotllm` command (model / run / chat / serve / bench / ...). Use `dnx dotllm --prerelease` or `dotnet tool install -g dotllm --prerelease` (see Getting Started)

The library projects (DotLLM.Core, DotLLM.Cpu, DotLLM.Cuda, DotLLM.Vulkan, DotLLM.Hip, DotLLM.Models, DotLLM.Tokenizers, DotLLM.Engine, DotLLM.Server, DotLLM.HuggingFace, DotLLM.Diagnostics, DotLLM.Telemetry) are packed on every build and attached as .nupkg files to each GitHub release, but are not published to nuget.org by this fork. The DotLLM.* ids on nuget.org belong to the upstream project and hold only its frozen 0.1.0-preview.* builds, which do not contain this fork's work — do not use them. To consume the libraries, download the .nupkg files from a release into a local feed:

# after downloading the DotLLM.*.nupkg assets of one release into ./feed
dotnet nuget add source ./feed --name dotllm-local
dotnet add package DotLLM.Engine --version 0.3.0-dev.<n>
dotnet add package DotLLM.Cpu --version 0.3.0-dev.<n>
dotnet add package DotLLM.Models --version 0.3.0-dev.<n>
dotnet add package DotLLM.Tokenizers --version 0.3.0-dev.<n>

All packages in a release carry the same version. See docs/PACKAGING.md.

Host the OpenAI API in your ASP.NET app

DotLLM.Server is a library — reference it from your own ASP.NET Core host to expose dotLLM's OpenAI-compatible routes. Two patterns are supported.

Mode 1 — run a dedicated dotLLM WebApplication inside your process. Simplest; you hand off model loading and routing to dotLLM.

using DotLLM.Server;

var options = new ServerOptions
{
    Model = "llama-3.2-3b.Q4_K_M.gguf",
    Device = "gpu",
    GpuLayers = 32,
    Port = 8080,
    PromptCacheEnabled = true,
    UsePaged = true,
};

using var state = ServerStartup.LoadModel(options.Model, options);
var app = ServerStartup.BuildApp(state, args: [], serveUi: true);

await app.RunAsync($"http://localhost:{options.Port}");

Mode 2 — attach dotLLM routes to your own WebApplication. Use this when you want dotLLM's endpoints alongside your own routes, middleware, and services.

using DotLLM.Server;

var builder = WebApplication.CreateBuilder(args);

// Load the model up front so the state can be registered in DI
// before builder.Build() — dotLLM endpoints resolve ServerState from DI.
var modelPath = "llama-3.2-3b.Q4_K_M.gguf";
var state = ServerStartup.LoadModel(modelPath, new ServerOptions
{
    Model = modelPath,
    Device = "cpu",
    PromptCacheEnabled = true,
    ModelId = "llama-3.2-3b",
});

// Your own services...
builder.Services.AddAuthentication();

// Register dotLLM's state + JSON context in the same IServiceCollection:
builder.Services.AddDotLLM(state);

var app = builder.Build();

// Your own middleware and routes...
app.UseAuthentication();
app.MapGet("/hello", () => "hi");

// Mount dotLLM's OpenAI-compatible endpoints:
app.MapDotLLMEndpoints(serveUi: false);

app.Run();

Both modes transparently reuse the embedded chat UI assets if serveUi: true. The ServerState is IDisposable — dispose it on shutdown to release the model and KV-cache.

News

2026-08 — Mamba-3 CUDA host (CudaMamba3TransformerModel) — closes the CPU/Vulkan-only gap. (#346) Four new F32 CUDA kernels (mamba3_data_rope_f32, mamba3_chunk_boundary_f32, mamba3_ssd_scan_siso_f32, mamba3_ssd_scan_mimo_f32) port the already-validated Vulkan GLSL compute shaders to CUDA C — one block per head, sequential in-kernel loop over t with __syncthreads() (not a per-token host-loop like the GDN scan kernel), compiled -fmad=false for CPU bit-parity; per-token preprocessing (softplus/sigmoid/RMSNorm+bias/qk_pre_dot/scale) stays host-side C#, mirroring Vulkan's design decision. CudaMamba3StateCache mirrors CudaGdnStateCache's allocation idiom with Mamba-3's four-buffer state (ssm_state/cum_angle/k_state/v_state); CudaModelLoader.LoadMamba3FromSafetensors is the dedicated safetensors entry point, reusing the CPU Mamba3WeightLoader for tensor resolution. No GGUF path on any backend — confirmed, not just unimplemented. SISO is now real-weight-verified against ib-ssm/mamba3-370M-10BT on an RTX 3060 (IbSsmMamba3CudaParityTests): prefill max_abs 1.502e-5, prefill+decode max_abs 2.074e-5, argmax matches — same standing as CPU/Vulkan. MIMO stays synthetic-fixture-only, same as CPU/Vulkan (no public MIMO checkpoint exists anywhere). F32 only — quantized/F16 CUDA paths are a follow-up.
2026-08 — MTP (Multi-Token Prediction) self-speculative decoding — CPU, CUDA, CLI, and server. (#253) Extends the existing two-model speculative-decode infrastructure with single-model self-speculation: a small MTP head predicts several future tokens per step, verified in one batched forward against the base model, reusing the DeepSeek-V3 "NextN" GGUF tensor convention (blk.{n}.nextn.{eh_proj,enorm,hnorm,embed_tokens}) — first seen at scale in froggeric/Qwen3.6-27B-MTP-GGUF, sharing Architecture.Qwen3HybridDense with PrismML's Bonsai-27B. CUDA (CudaMtpState) composes entirely from already-proven kernels (RmsNorm, RoPE, attention, SwiGLU, the Gemm dispatcher) rather than new kernel work. CLI wiring: --no-mtp opt-out for run/chat (auto-detects and enables by default); server wiring: --mtp opt-in for serve (disables the continuous-batch scheduler while active, mirroring --speculative-model's existing restriction), surfaced via mtp_active in /props. Along the way, fixed a correctness gap in GDN (Gated DeltaNet) recurrent architectures: unlike KV-cache, GDN's recurrent state has no position addressing, so a rejected draft token's effect couldn't be "un-reached" by truncation alone (#287) — new IModel.CheckpointRecurrentState()/RestoreRecurrentState() (CPU + CUDA) make rollback safe for both SpeculativeDecoder and MtpSpeculativeDecoder on hybrid GDN/attention models. Real end-to-end validation via dotllm run against the real froggeric/Qwen3.6-27B-MTP-GGUF (Q4_K_M) checkpoint.
2026-08 — Mach-1 additive-codec load path (Phase B) — SyzygyResearch/Mach-1-Additive-35B loads via the existing Qwen3MoeHybrid forward pass, zero forward-pass changes. (#266) This checkpoint is the same Qwen3.6-35B-A3B model as Phase 10's qwen35moe GGUF support, but compressed to ~1.7 bpw (7.53 GB) with a bespoke tail-biting-trellis + randomized-Hadamard codec that has no GGUF representation. Phase A (bit-exact codec decoder, previously shipped) is now wired into the standard loading surface: Qwen35MoeConfigExtractor parses the nested text_config HF config; Mach1PackedCheckpoint reads the packed/ HF repo layout via the existing generic safetensors reader (no new binary parser needed — its composite |-delimited keys are just ordinary safetensors headers) and dispatches container kind per-file rather than assuming a fixed layout; Qwen3MoeHybridTransformerModel.LoadFromMach1Packed decodes every tensor to dense fp32 and populates the exact same weight structures the GGUF loader builds. Every tier (routed experts, GDN spine, full-attention spine, LM head int5g64_packed, embeddings affine-int4-with-exceptions) and both layer kinds are independently validated against the real fixture — several bit-exact against the vendor's own golden tensor — plus the real chat_template.jinja's {% macro %} support (issue #273) verified end-to-end. Full 40-layer end-to-end generation and top-1 agreement against the base GGUF is gated on ~70-128 GB RAM for a fully dense-decoded model — a known hardware limitation the issue itself anticipated ("a validation vehicle, not a deliverable"), not a defect in the loader, which keeps all layers resident exactly like the GGUF/mmap path already does. Phase C (fused additive-domain GEMV keeping routed experts in their native ~6.2 GB packed form) remains pending on all three backends before the issue closes.
2026-07 — PQ2_0 ternary quant type + dense Qwen3.5 hybrid architecture — PrismML Bonsai-27B runs end-to-end. New QuantizationType.PQ2_0 (#151) for PrismML's Bonsai-27B: unlike I2_S's single per-tensor scale, PQ2_0 packs a per-128-element-group FP16 scale ahead of its 2-bit ternary codes; CPU scalar + CUDA GEMV/dequant kernels ship for both. New Architecture.Qwen3HybridDense (qwen35, #157) — a dense-FFN sibling of Qwen3MoeHybrid reusing the same GDN/full-attention token-mixing infrastructure, with a dense SwiGLU FFN in place of MoE routing. CPU and CUDA forward passes both validated end-to-end against the real 7.17 GB Ternary-Bonsai-27B-Q2_0.gguf (64 layers: 48 Gated-DeltaNet + 16 full-attention, PQ2_0 throughout). Along the way, fixed a VRAM-oversubscription bug on 12 GB consumer cards: a full-table F32 embedding pre-dequant (~5 GB at Bonsai's 248320-token vocab) was silently tripping WDDM's host-RAM paging fallback, masquerading as a hang; replaced with a per-call host-side row dequant. Baseline dotllm bench --device cuda: prefill 9.86 tok/s, decode 1.71 tok/s on an RTX 3060 (unoptimized, correctness-first PQ2_0 kernels — a target for follow-up optimization). Follow-up same day: the PQ2_0 F16 decode GEMV (pq2_0_gemv.cu) was rewritten to mirror I2_S's proven v2 scheme — shared-memory x-staging (kept half-width so Bonsai's largest K=17408 fits under sm_86's 48 KB static cap) plus warp-per-row (16 rows/block) instead of one-warp-per-row-with-global-x-reads, and the launch grid is no longer capped at 256 blocks. Decode 1.71 → 7.09 tok/s (4.15×); prefill unchanged (untouched dequant-then-cuBLAS path). Validated against the CPU float reference across real Bonsai shapes plus tail-row-clamp edge cases (CudaPQ2_0GemvTest). A follow-up fused-projection round (pq2_0_gemv2_f16in, mirroring i2_s_gemv2_f16in) applied to dense-FFN gate+up, attention K+V, and GDN alpha+beta — correctness-validated (fused-vs-separate parity test) but measured no further decode gain (stayed at 7.09–7.10 tok/s across all three): compute now dominates the per-call shared-x-staging cost it targets, so GEMV is no longer decode's bottleneck. Next lever is likely kernel-launch-count reduction (the codebase already has CudaDecodeGraph.cs for CUDA graph capture) rather than further GEMV tuning. Overnight follow-up (same PR): a lightweight category profiler found the attention Q+Gate de-interleave host loop (numHeads separate paired cuMemcpyDtoDAsync calls per layer) was ~12% of decode time on its own — replaced with a single gather-kernel launch (deinterleave_qgate_f32; same fix applied to GDN's Q/K/V split). Then found the GEMV kernel's v2 fix hadn't actually solved the coalescing problem it targeted — lanes still owned separate whole groups 34 bytes apart, just fewer of them; restructured so the warp cooperates on one group at a time (lane L reads code byte L — a genuinely coalesced 32-byte transaction). The same uncoalesced pattern turned out to also be present, untouched, in the prefill-path dequant kernel (dequant_pq2_0.cu) — fixing it there took prefill from 12.6 → 89.06 tok/s (7.07×). The identical defect was found and fixed in I2_S's (BitNet) sibling dequant kernel too. Final cumulative numbers from this session's original unoptimized baseline: prefill 9.86 → 89.06 tok/s (9.0×), decode 1.71 → 10.34 tok/s (6.0×) on an RTX 3060. A PQ2_0_ROWS_PER_WARP=4 tuning experiment was tried and measured worse (documented, reverted) — decode is no longer GEMV-bound at the margin the earlier fixes left it. Decode-gap follow-up: with GEMV no longer the bottleneck, the profiler was re-run to find what else remained. Fused the decode-time residual-copy+RmsNorm pattern (cuMemcpyDtoDAsync immediately followed by rmsnorm_f32, called 2×/layer × 64 layers — ~9% of decode time) into one kernel (copy_rmsnorm_f32): decode 10.34 → 10.51 tok/s (+1.6%), bit-exact validated against the separate-calls sequence. Then fused GDN's decay+sigmoid step (gdn_decay_f32 + sigmoid_f32, called once each per GDN layer × 48 layers) into gdn_decay_sigmoid_f32 — lives in the -fmad=false bit-perfect kernel file since it sits in the GDN recurrence's precision-critical path: decode 10.51 → 10.59 tok/s (+0.8%), again bit-exact validated. Re-profiling afterward showed the GEMM-bound categories (dense FFN gate+up/down, GDN projections — already coalescing-optimized) now dominate decode time; the next non-GEMM candidate (GDN's causal-conv1d input staging, ~6% of decode) needs a riskier restructure of a -fmad=false kernel for an estimated sub-1% gain and was left as documented future work rather than pursued for diminishing returns. Cumulative session decode: 1.71 → 10.59 tok/s (6.2×); prefill: 9.86 → 88.34 tok/s (9.0×). Occupancy follow-up (#159): a roofline check (RTX 3060's 360 GB/s bandwidth against Bonsai's 7.17 GB weights implies a ~50 tok/s decode ceiling; 10.59 tok/s was only ~21% of it) prompted a deeper ncu (Nsight Compute) pass, which found the decode GEMV was latency-bound, not bandwidth-bound: a single-lane scatter store (fixed, decode 10.59 → 10.95 tok/s), then the real headline fix — the per-layer activation-staging loop issued up to 68 sequential single-half loads before releasing compute; vectorizing to uint4 cut that ~8× (decode 10.74 → 12.72 tok/s). A second GEMV specialization split the kernel by shape: the small attention/GDN case (k≤5120) got its own tighter shared-memory buffer, lifting occupancy from 33%→100% theoretical (decode 12.72 → 14.68 tok/s); the larger FFN case (k=17408) got windowed activation staging instead, since it can't shrink the same way, reaching the same 83% occupancy ceiling (decode 14.68 → 15.79 tok/s). Two follow-on ideas (resizing the FFN kernel's launch grid to remove wave-quantization tail effects, and an explicit warp-shuffle broadcast for a redundant scalar read) were modeled, implemented, and measured — both regressed real throughput once the kernel was no longer latency-bound, and are reverted/documented in pq2_0_gemv.cu's history rather than shipped. Final cumulative numbers across both issues: decode 1.71 → 15.79 tok/s (9.2×), ~31.6% of the theoretical bandwidth roofline. Launch-count + correctness follow-up (#161): a fresh ncu pass plus an advisor review found the largest remaining PQ2_0 GEMV kernels had shifted from latency-bound to genuinely ALU-bound (78%/68% compute throughput), and — via tracing real call sites rather than only the profiled kernel names — that ~65% of all PQ2_0-related kernel launches per decode token (~704 of ~1,088) were pure convert_f32_to_f16/convert_f16_to_f32 plumbing bridging the model's F32 activation pipeline to the F16-only GEMV kernels. New F32-native GEMV kernel variants convert inline during the existing vectorized stage/store steps, eliminating 864 of those launches per decode token entirely: decode ~15.7 → ~16.9 tok/s (+8%). A follow-up algebraic identity (Σ(code−1)·x = Σ(code·x) − Σx, letting the group-level Σx be computed once per warp and reused across rows instead of once per byte) removed the ternary decode loop's per-element subtract: decode ~17.0 → ~17.3 tok/s (+1.9%). Two other candidates were tried and reverted (a SwiGLU-epilogue fusion regressed 5.4% — interleaving two independent weight tensors broke a previously-sequential access pattern; CUDA graph capture was investigated and explicitly deferred — the KV-cache/attention path for this hybrid model needs a real device-position-aware redesign, not a quick port of the existing dense-F16 pattern). Along the way, found and fixed a real correctness bug (not a perf issue): the LM head and token-embedding tensors were uploaded in the raw on-disk PQ2_0 layout while every GEMV/dequant kernel assumed the split [scales][codes] layout every other weight gets — confirmed via CPU-vs-GPU logit comparison (sign-uncorrelated, ruling out a simple scale bug) and fixed by routing both tensors through the same repack path as everything else; a previously-failing real-weight smoke test (-∞ logits) now passes clean. Final cumulative numbers across all three issues: decode 1.71 → 17.85 tok/s (10.4×), ~35.7% of the theoretical bandwidth roofline.
2026-06 — Phase 9 Production Serving complete — advanced scheduling (Step 59). The continuous-batch scheduler gains the full Step 59 feature set. Batched forward across in-flight sequences via IModel.ForwardBatch: steady-state decode (#348), multi-admit prefill (#349), and recurrent decode/prefill for Mamba-3 + Qwen3-MoE-Hybrid (#350) — the latter via a new IRecurrentSequenceState container the scheduler allocates and threads per sequence (IModel.CreateSequenceState/SupportsThreadedSequenceState), which also fixes the latent cross-sequence corruption of running >1 concurrent recurrent sequence against a shared default state; MaxRecurrentSequences bounds per-seq state memory. Prefill/decode disaggregation (#351): Step() splits into StepPrefill()/StepDecode() over the separate prefill queue + decode set — the seam a future disaggregated/multi-worker driver uses, while the in-process single-GPU path stays serial (the model forward is single-threaded). Fairness (#352): opt-in per-API-key start-time fair queuing in admission so a hammer-client can't starve light tenants (priority still dominates across tiers). Joins the already-shipped chunked prefill, priority admission, and recompute-on-resume preemption. Phase 9 (continuous batching, prefix sharing, advanced scheduling, rate limiting, metrics & tracing) is now Done (5/5).
2026-06 — BitNet b1.58 support — Microsoft BitNet b1.58 2B4T runs natively via the official I2_S ternary GGUF (weights ∈ {-1, 0, +1}, 4 codes/byte + per-tensor scale). New QuantizationType.I2_S dequant + ternary GEMV/GEMM kernels, config-driven FFN activation with squared-ReLU (ReluSquared) gating, and attention/FFN Sub-LN wired into the parameterized TransformerBlock (no separate model class). CPU correctness-first path (full-precision activations); a parallel/SIMD ternary kernel and GPU (W2A8) support are follow-ups (#77)
2026-05 — Strix Halo Vulkan hardening + per-sequence recurrent-state isolation. Post-Phase-10 work landing across the Strix Halo dev host (AMD Radeon 8060S iGPU, gfx1151). Sliding-window mask off-by-one (> vs >=) fixed across 5 GPU kernels (Vulkan attention_f32.comp + 3 FA shaders, CUDA attention.cu + attention_f32.cu); CUDA PTX regen deferred to a CUDA host (acknowledged in commit). IQ3 family Vulkan support complete: IQ3_XXS (3.0625 bpw) + IQ3_S (3.4375 bpw) GEMV/GEMM kernels via SSBO codebooks, bit-perfect vs ggml-quants.c (16 kernel parity tests), upload-path predicates + per-host matmul dispatch on all 4 Vulkan transformer hosts (dense / Mamba3 / NemotronH / Qwen3MoeHybrid), and 16 host-level IQ3 forward parity tests across all 4 hosts — audit-H3 trap-the-bug (kernel-level ✓ ≠ dispatch-level ✓) is now closed for IQ3 across the entire matrix. Vulkan Flash Attention shader (flash_attention_f32.comp) with online-softmax single-pass; Vulkan FA microbench landed in PERFORMANCE.md §6.4. ForwardBatch CPU Phase 5a/5b: lm_head GEMM fusion at seqLen = Σ N_i (commit 479c23f), intra-block Q/K/V/O/gate/up/down fusion across the simple-subgroup (commit 92c1345, ~2.09× speedup at 4× decode batch on SmolLM-135M Q8_0). Phase 5f Vulkan dense intra-block fusion (commit 1c04887) mirrors the CPU pattern. Gemma 3 wiring (CPU + Vulkan): attn_logit_softcapping, per-layer sliding window, query_pre_attn_scalar, final_logit_softcapping plumbed through both backends with synthetic discriminative parity tests. Per-sequence recurrent-state isolation for hosts with non-attention state: IGdnState interface + SequenceForwardRequest.GdnState slot lifts the multi-seq guard on VulkanQwen3MoeHybridTransformerModel.ForwardBatch; IMambaState mirror lifts the multi-seq guard on VulkanMamba3TransformerModel.ForwardBatch. Multi-seq parity tests on both hosts compare against per-seq Forwards on fresh model instances (abs 5e-3 / rel 1e-3). LoRA Q8_0 stage-1 Path B3 disconfirmed empirically (commit dd2892f, ~840 µs vs ≤470 µs bar at canonical rank=16); residual −16% LoRA Q8_0 prefill gap is now bounded by what an upstream Avx512Vnni.V512 PR would buy (passive at https://github.com/dotnet/runtime/issues/86849). Real-GGUF GPU parity tests still pending (29.6 GiB Q6_K_XL needs a CUDA host with >12 GiB VRAM; llama.cpp Vulkan FA bench needs -DGGML_VULKAN=ON build).
2026-05 — Qwen3MoeHybrid (qwen35moe) end-to-end — Gated DeltaNet + sparse MoE. Phase 10 lands the CPU, CUDA, and Vulkan implementations of Alibaba's Qwen3.6-35B-A3B hybrid architecture (38 GDN linear-attention layers + 10 full GQA layers, each feeding a shared sparse MoE with 256 routed experts top-8 + a sigmoid-gated shared expert). CPU is bit-exact vs llama.cpp (top-1 token "Ta" matches the reference), and the per-forward expert dequant scratch (~30 GiB at Q6_K_XL scale) is eliminated by feeding the GGUF raw quant view (Q4_K / Q5_K / Q6_K / Q8_0 / Q5_0 / F32 / F16) directly into the GEMM path; --threads N is honoured for parallel projections. CUDA ships per-layer gdn_scan_step, gdn_decay, conv1d_causal, l2_normalize_heads, four fused elementwise kernels (sigmoid, silu, sigmoid_mul), Q6_K → F32 dequant, and an on-device MoE dispatcher via CudaMoeFfn (no more D2H/H2D fallback) with a model-private F32 KV cache for the 10 full-attn layers; csproj fix preserves -fmad=false across PTX rebuilds. Vulkan ships seven new GDN-related compute shaders including a multi-token GDN scan (one dispatch per layer instead of O(seqLen) with 6 per-token D2D copies eliminated) and an on-device gdn_decay_f32 + sigmoid_inplace_f32 pair replacing the host round-trip; resident routed-expert banks are opt-in via DOTLLM_VK_MOE_RESIDENT=1. Pre-existing CudaMoeGroupedGemv Q5_K divergence fixed along the way (Q4_K-style memory layout was being used to decode Q5_K). 1981 unit tests pass. Real-GGUF end-to-end GPU parity tests remain pending pursuant to hardware (29.6 GiB Q6_K_XL exceeds local 12 GiB CUDA card; Vulkan SDK + Strix Halo SPV build needed).
2026-04 — Vulkan MoE Strategy C grouped F16 coopmat path — routed MoE expert rows can now stay grouped by expert on-GPU for native F16 expert banks when VK_KHR_cooperative_matrix is available. The path computes expert counts/offsets on device, groups routed rows, runs grouped F16 coopmat W1/W3/W2 projections, ungroups rows back to the existing weighted-scatter contract, and falls back to the prior indexed F32/Q8_0 path whenever the model has active MoE LoRA, non-F16 expert banks, incompatible dimensions, or no cooperative-matrix support. Covered by standalone group/offset/grouped-matmul kernel tests and an end-to-end synthetic Vulkan-vs-CPU MoE forward parity test.
2026-04 — Vulkan: K-quants (Q4_K_M / Q5_K_M / Q6_K_M) + native F16 / BF16 matmul kernels, plus comprehensive real-weight Vulkan parity coverage. Vulkan's matmul library now ships GEMV (decode) + GEMM (prefill) shaders for every common GGUF quantisation: F32, F16, BF16, Q8_0, Q4_K_M, Q5_K_M, Q6_K_M — plus an opt-in F16 cooperative-matrix GEMM path on devices that enumerate F16xF16→F32 tiles (gfx1151 verified). K-quant + Q8_0 weights stay in source bytes on device and dequantise inside the shader inner loop, eliminating the 2× memory penalty of the previous F32-only upload path. F16 source weights also stay native; BF16 source preserves bytes via uintBitsToFloat(bits<<16) (no F16 extension required). The kernel half lands at afb2272/15099b9/29a1459/c9c08c5; upload-path + dispatcher routing at b1ee6bc/83e0732/39b7646/6787492 covers VulkanWeights (standard transformer), VulkanNemotronHWeights, VulkanMamba3Weights and the MoE overlay slots. Real-weight CPU↔Vulkan parity at the integration tier now covers TinyLlama-1.1B, Phi-3.5-mini, Qwen2.5-0.5B, Granite-3-MoE, Mamba-3 ib-ssm 370M, DeepSeek-V2-Lite-Q4_K_M (MLA + multi-shared MoE, 19m 27s end-to-end) — RealHfSafetensorsEndToEndVulkanTests + RealGgufVulkanParityTests. Vulkan suite 264 → 360 (+96 unit tests). DeepSeek-V2-Lite via the BF16 SafeTensors path self-skips at 0.7 s with a clear actionable message (the F32-expanded weights would need ~58 GB; use the Q4_K_M GGUF instead — Phase 1 K-quants make it native).
2026-04 — Long-context YaRN ramped-freq verification end-to-end — DeepSeekV2Lite_LongContext_LogitsMatchPyTorchReference was previously a 50 ms self-skip while its PyTorch reference JSON was missing. New --prompt-file + --last-n-rows flags on tests/scripts/compare_logits_py_reference.py plus a logits_offset field on the JSON schema make >4K-token references practical (a full 4920 × 102400 reference would be ~5 GB JSON; subset-to-last-8-rows is ~7.4 MB). The test now actually runs (36m 55s on this iGPU box's CPU forward) and validates positions 4912–4920 — past DeepSeek-V2-Lite's original_max_position_embeddings=4096 where YaRN's ramped inv_freq actively diverges from plain RoPE — within DriftTolerances.Tight against the PyTorch BF16 oracle. Validates the YaRN fix landed at 1e0a930. The C# reference comparator was extended to slice ours[offset..offset+ref_seq_len, :] for the comparison; short-prompt references with offset=0 are unaffected.
2026-04 — DeepSeek-V2/V3 default to Phase C latent KV-cache at the loader level (HF + GGUF) — the Phase A expanded cache scales as O(numLayers × maxSeqLen × numHeads × (qkNope + v) × 4 bytes) which for V2-Lite at max_position_embeddings=163840 is ~68 GB and OOMs on most hosts. Phase C's hybrid (latent persistence + Phase A-equivalent prefill expand + absorbed decode) keeps the same numerics but stores kv_lora_rank=512 per token instead of expanded heads — ~9 GB for the same model. HfConfigExtractor and GgufModelConfigExtractor both now set UseHybridMlaCache=true by default for Architecture.DeepSeekV2/Architecture.DeepSeekV3; production code paths get the win automatically, unit tests that build MlaConfig directly (bypassing the loader) still default to Phase A.
2026-04 — Vulkan: fused rmsnorm + matmul_q8_0 (sub-tile design), 282 → 295 tok/s SmolLM-135M.Q8_0 decode on AMD Radeon 8060S iGPU (RDNA3.5; +4.6%, beats v0.2.0-alpha.2 baseline of 291). Eliminates 60 dispatches + 60 pipeline barriers per decode at the attn-norm+Q and ffn-norm+Gate fusion sites. Each WG outputs ROWS_PER_WG=8 matmul rows so the rmsnorm reduction is amortised 8× across the rows it feeds (72 vs 576 WGs for SmolLM Q proj); the matmul phase reads normalised values from on-chip shared memory, and a leader-WG slice of the dispatch writes the global normOutput so K/V/Up matmuls keep working unchanged. A first per-row variant was ~7% slower (576× redundant rmsnorm overwhelmed the dispatch+barrier savings); sub-tile-4 broke even; sub-tile-8 with K_MAX=1024 (4 KB sharedNorm) tipped the balance. Default-on; opt out via DOTLLM_VULKAN_DISABLE_FUSED_RMSNORM_MATMUL=1. Falls back to the standalone (rmsnorm + matmul) pair for prefill (seqLen>1), non-Q8_0 weights, and hidden sizes >1024. Strict-argmax CPU↔Vulkan parity holds (8/9, unchanged). 8 unit parity tests use a CPU reference and exercise K=32 with M>1, sidestepping the standalone GEMV's known rowUints*4 over-stride bug (issue #1) — the fused shader uses the correct blocksPerRow*34 byte stride (#136)
2026-04 — Phi-3.5-mini-instruct + Granite-3.0-MoE end-to-end on real weights. Two loader extensions enable loading recent production checkpoints as-is from HuggingFace. Phi-3 (Phi3ForCausalLM) stores attention as a fused self_attn.qkv_proj [Q+K+V, hidden] and MLP as a fused mlp.gate_up_proj [2*intermediate, hidden]; the new SplitFusedProjection helper probes for these fused names, allocates one 64-byte-aligned F32 buffer per split at the correct row offsets, and BF16/F16/F32-decodes in place. Fusion-absent models unchanged. Granite-MoE (GraniteMoeForCausalLM / model_type=granitemoe) packs all experts per-layer into block_sparse_moe.input_linear [E, 2*I, H] + block_sparse_moe.output_linear [E, H, I] + a router.layer gate. New Architecture.GraniteMoe variant + LoadGraniteMoeLayer slices each expert's w1/w2/w3 from the fused tensors into per-expert F32 pointers; the existing MoeSwiGluMlp kernel handles top-8 routing natively. Real-weight verification: microsoft/Phi-3.5-mini-instruct (7.6 GB, 32-layer MHA) produces [3, 32064] logits 96192/96192 finite (stddev 7.00); ibm-granite/granite-3.0-3b-a800m-instruct (6.3 GB, 40 experts top-8, 24Q/8KV GQA) produces [3, 49155] logits 147465/147465 finite (stddev 16.13). Both gated by env vars (DOTLLM_PHI35_CHECKPOINT_PATH, DOTLLM_GRANITE3_CHECKPOINT_PATH) with auto-detect of C:/temp/dotllm-{phi35-mini,granite3-moe}/ (#136)
2026-04 — DeepSeek-V2/V3 end-to-end MLA integration — TransformerModel.Forward gains a per-layer MLA branch (RMSNorm → MlaAttention.Execute → residual → shared MoE/dense FFN dispatch), LoadDeepSeekMlaLayer safetensors weight loader resolves all MLA tensors (q_a_proj/q_a_layernorm/q_b_proj or monolithic q_proj when q_lora_rank=0, kv_a_proj_with_mqa/kv_a_layernorm/kv_b_proj, o_proj), and ModelLoader.LoadFromSafetensors now dispatches Architecture.DeepSeekV2/DeepSeekV3 (removed NotSupportedException). FFN routes to dense MLP (first-K prefix) or plural-shared-experts MoE (step 58c). GQA path for Llama/Mistral/Phi/Qwen unchanged. Verified against real yujiepan/deepseek-v2-tiny-random: 4-token prefill produces 409600/409600 finite logits (stddev 0.056). 3 new unit tests (LoRA-Q prefill, monolithic-Q prefill, LoRA-Q single-token); 1285 unit tests pass. Follow-ups: latent KV-cache, YaRN mscale, absorption optimisation, real DeepSeek-V2-Lite (16 GB) validation (#136)
2026-04 — DeepSeek multi-shared-expert MoE (n_shared_experts > 1) — extends the MoE shared-expert branch to run N shared experts in parallel (DeepSeek-V2/V3 convention; Qwen1.5-MoE at N=1 unchanged). MoeConfig.NumSharedExperts (default 1). MoeLayerWeights.SharedGateProj/UpProj/DownProj migrated from single nint to nint[]. MoeSwiGluMlp.ExecuteWithSharedExpert accepts per-shared-proj ReadOnlySpan<nint>; for k=0 the GEMM writes directly into the down-buffer (bit-identical with prior single-shared path), subsequent experts compute into pooled scratch then TensorPrimitives.Add. Plural HF naming mlp.shared_experts.{k}.{gate,up,down}_proj supported; singular Qwen1.5 mlp.shared_expert.* still loads as length-1 array. This path is now used by the DeepSeek-V2/V3 MLA forward integration as well as Qwen-MoE shared-expert variants (#136)
2026-04 — Mamba-3 MIMO streaming decode (kernel + block) — extends the SISO streaming SSD from step 60e to MIMO (rank > 1). Mamba3State.k_state rank-extended to [R, H, N] (SISO keeps [H, N]); ssm_state, cum_angle, v_state remain rank-free per canonical mamba3.py allocate_inference_cache(). New Mamba3CanonicalSsd.ExecuteMimoStreaming mirrors SISO streaming with rank-aware K; boundary term derived from the tilelang MIMO forward: ssm[h,p,n] += v_state[h,p] · (Σ_r k_state[r,h,n]) · DT[0,h] · (1-trap[0,h]) — the rank-sum analog of the SISO adjustment. Empty-span and single-chunk schedules are bit-identical to the existing one-shot ExecuteMimo. Synthetic MIMO drift (R=3) across all split schedules: y max_abs ≤ 4.7e-15, ssm_state max_abs ≤ 3.4e-13 — pure F32-reorder noise. 18 new unit tests (8 kernel + 9 block + 1 state shape). Model-level wiring (Mamba3TransformerModel MIMO path) awaits the weight-loader extension for [H, R, N] biases + mimo_* projections (#136)
2026-04 — MoE GroupedGEMM refactor (bit-identical) — MoeSwiGluMlp moves from per-token per-expert scalar loop to gather/batched-GEMM/scatter: bucket tokens by their top-k-assigned experts (ArrayPool-rented index lists, zero sustained allocations), run one batched SwiGLU per expert across all its assigned tokens (gather → w1/w3 GEMM → silu*up → w2 GEMM), scatter back preserving the original per-(token, slot) accumulation order for bit-identity. Public signatures unchanged; both Execute (Mixtral) and ExecuteWithSharedExpert (Qwen-MoE) route through the same grouped core. Verified bit-identical on all 8 existing MoE unit tests + 16 integration tests including the real yujiepan/qwen3-moe-tiny-random end-to-end forward. Follow-up: fused GroupedGEMM with SIMD-batched SwiGLU for ~2-3× further (#136)
2026-04 — MLA attention kernel (PoC) + DeepSeek-V2/V3 detection — new MlaAttention.Execute scalar kernel implements DeepSeek's Multi-head Latent Attention end-to-end: Q LoRA factorization (q_a_proj → q_a_layernorm → q_b_proj, or monolithic Q when q_lora_rank=0), KV compression via kv_a_proj_with_mqa → kv_a_layernorm → kv_b_proj (split into K_nope + per-head V), decoupled RoPE on qk_rope_head_dim sub-dim with an MQA-shared K_rope, per-head causal SDPA, and o_proj. New Architecture.{DeepSeekV2, DeepSeekV3} enum variants; HfConfigExtractor detects DeepseekV2ForCausalLM / DeepseekV3ForCausalLM and populates MlaConfig + DeepSeek-flavoured MoeConfig (n_routed_experts, n_shared_experts, first_k_dense_replace). Verified against yujiepan/deepseek-v2-tiny-random (config detection) and a reference implementation within 5e-4 absolute tolerance. Later work wired this into TransformerModel.Forward, safetensors/GGUF loading, latent KV-cache defaults, and real DeepSeek-V2-Lite verification (#136)
2026-04 — Mamba-3 streaming decode (chunk-boundary state) — Mamba-3 prefill+decode now reproduces one-shot prefill at chunk boundaries, closing the canonical shifted_γ[t] = DT[t+1]·(1-trap[t+1]) 1-token lookahead gap. Mamba3State gains two new persistent buffers per layer — k_state [n_head, d_state] (previous chunk's last-token post-RoPE, pre-scale K) and v_state [n_head, head_dim] (previous chunk's last-token V) — matching canonical mamba3_ssm.modules.mamba3 commit 7438488 (mamba3.py:142 / mamba3_siso_fwd.py:341-352). Mamba3Block.Forward gains kState / vState span parameters: at chunk start it folds the deferred ssm += v_state · k_state · DT[0] · (1 - trap[0]) term into the SSM state BEFORE the scan — exactly the contribution a one-shot forward would have produced via scale[T_prev-1] = γ + shifted_γ; at chunk end it persists this chunk's last-token post-RoPE K (from bHRN[T-1]) and V (from xBuf[T-1]). Mamba3TransformerModel.Forward(..., Mamba3State) threads all four buffers. User-visible win: on the real ib-ssm/mamba3-370M-10BT (48 layers, 32 heads, d_state=128) the last-token logit drift max_abs drops from ≈6.5 to ≈3.3 and — critically — the top-1 argmax now matches between one-shot and 2-prefill+1-decode (previously differed, so next-token sampling was inconsistent across modes). On tiny synthetic fixtures prefill+decode reproduces one-shot to F32-reorder noise (1e-8 on 2+2 splits, ~2e-6 on three-way splits). Opens the door to true O(N) streaming-decode generation (vs today's O(N²) growing-context prefill). Tightened assertions: PrefillThenDecode_ApproximatesOneShot_WithExpectedDrift renamed to PrefillThenDecode_BitEqualsOneShot (abs_tol 5e-5→1e-5, three schedules); DecodeMatchesPrefillOnRealCheckpoint ceiling 100→10 with a new Assert.Equal(argmaxPrefill, argmaxSplit). New Mamba3BlockStreamingTests (7 split schedules × {y, ssm_state, k_state, v_state}) covers the kernel-level property. MIMO streaming deferred (canonical's tilelang MIMO kernel has no input-state signature; the MIMO Block path is unreachable today). 65 Mamba-3 unit tests + 20 integration tests all green (#136)
2026-04 — Mamba-3 pooled forward scratch — eliminates per-call managed allocations on the Mamba-3 hot path. New Mamba3ForwardScratch : IDisposable (mirrors the existing NemotronHForwardState pattern) owns 13 named NativeMemory.AlignedAlloc-backed 64-byte-aligned scratch buffers (Proj, X, Z, Dt, Adt, Trap, Gamma, Scale, AnglesRaw, B, C, QkPreDot, YScan) grown power-of-two on demand via EnsureCapacity(seqLen) and reused across every layer of every Mamba3Block.Forward / ForwardMimo call. Mamba3TransformerModel allocates one scratch at construction and disposes it — on the ib-ssm 370M checkpoint this collapses ≈480 per-step managed float[] allocations (48 layers × ≈10 buffers) to zero after the first Forward. Numerics preserved bit-exact: every pre-existing Mamba-3 test (57 unit + 20 integration, including the canonical-Python SISO/MIMO block comparators and all 9 real-weight ib-ssm integration tests with end-to-end 5-token generation) still passes. ComputeMemoryBytes now reports the scratch footprint. Scratch handles both SISO and MIMO B/C widths in one instance (backed by max(1, mimoRank)·nHead·dState) (#136)
2026-04 — Qwen-MoE support (Qwen1.5/2/3-MoE + shared experts) — extends the Mixtral MoE plumbing to the HF Qwen-MoE naming convention (mlp.gate + mlp.experts.{j}.{gate_proj,up_proj,down_proj} instead of Mixtral's block_sparse_moe.gate + experts.{j}.w1/w2/w3), with optional shared-expert branch (Qwen1.5-MoE-A2.7B: mlp.shared_expert.* + optional mlp.shared_expert_gate.weight sigmoid scalar) and the norm_topk_prob=false raw-softmax gating used by Qwen1.5. New Architecture.QwenMoe enum variant dispatches Qwen{2,3}MoeForCausalLM / model_type=qwen{2,3}_moe. MoeConfig gains NormTopKProb, SharedExpertIntermediateSize, HasSharedExpertGate, DecoderSparseStep, and MlpOnlyLayers — the last two let Qwen3-MoE interleave dense MLP and MoE layers in the same model (decoder_sparse_step=2 → layer 0 dense, layer 1 MoE). New MoeSwiGluMlp.ExecuteWithSharedExpert overload runs a parallel dense SwiGLU on every token and (optionally) multiplies it by sigmoid(hidden . shared_expert_gate) before adding to the routed top-k sum. Existing Mixtral Execute call-sites are untouched — the kernel change is additive. Verified end-to-end against the real yujiepan/qwen3-moe-tiny-random HF checkpoint (~20 MB, 2 layers × 8 experts × top-2, decoder_sparse_step=2, no shared expert): detection → load → 3-token forward → finite logits. Synthetic-fixture coverage for the shared-expert + sigmoid-gate + raw-softmax path (Qwen1.5-MoE convention). DeepSeek-V2/V3 reuse the plural shared-expert loader path through the MLA integration (#136)
2026-04 — Mamba-3 end-to-end text generation verified on ib-ssm 370M — full pipeline confirmation: ModelLoader.LoadFromSafetensors + ModelLoader.LoadTokenizerFromHfDirectory + iterative Mamba3TransformerModel.Forward → argmax → decode. Test generates 5 tokens from "The capital of France is" → [263, 1407, 4100, 760, 310] → "The capital of France is a very important part of" (coherent English from a 10BT-trained base model). Uses growing-context one-shot prefill per step (O(N²) total, every forward canonical) to sidestep the shifted_γ chunk-edge lookahead drift that affects streaming decode; ~1.4 s/token on CPU, 7 s end-to-end for 5 tokens. Streaming-SSD decode (O(N) per token, needs k_state + v_state threading plus a streaming kernel) remains a follow-up. Gated integration test IbSsmMamba3GenerationTests runs on DOTLLM_IBSSM_CHECKPOINT_PATH / conventional paths (#136)
2026-04 — Multi-shard safetensors loader (model.safetensors.index.json) — HuggingFace sharded-checkpoint support for models above the default 5 GiB shard cap (e.g. Llama-3-8B-Instruct, Mistral-7B-Instruct). New SafetensorsIndex parses the weight_map + metadata.total_size sidecar; MultiShardSafetensorsFile opens every shard, resolves tensor-name → owning shard, and exposes the same lookup surface as the single-file reader via new ISafetensorsTensorSource interface. Consumers (TransformerWeightsSafetensorsLoader, Mamba3WeightLoader, TransformerModel.LoadFromSafetensors, Mamba3TransformerModel.LoadFromSafetensors) accept the interface; existing single-file callers unchanged. ModelLoader.LoadFromSafetensors auto-detects: directory path → probe for model.safetensors.index.json first (multi-shard) else single .safetensors; index.json path → multi-shard; single-file path → single-shard, but prefers sibling index.json when present. Duplicate tensor-name handling favours the index's authoritative mapping and throws on conflicts in non-index shards. Verified end-to-end via resharded hf-internal-testing/tiny-random-LlamaForCausalLM (2 shards, matching index.json, full forward pass: [3, 32000] finite logits, stddev=0.080) (#136)
2026-04 — HF tokenizer.json ingest for Mamba-3 / Llama-family — JSON-driven BPE adapter bridging HuggingFace tokenizer.json into the existing BpeTokenizer / SentencePieceEncoding infrastructure. HfTokenizerJsonParser decodes vocab, merges (both "a b" and modern ["a","b"] forms), added tokens, Metaspace pretokenizer, ByteFallback decoder. HfBpeTokenizerFactory maps merge rank → synthetic score (score = -rank) so the score-driven merge queue reproduces HF's earliest-rank-wins order. Special added tokens route through the control-token pre-split path; BOS/EOS auto-detected. ModelLoader.LoadTokenizerFromHfDirectory surfaces it next to LoadFromSafetensors. Verified end-to-end on real ib-ssm/mamba3-370M-10BT: "Hello world" → [15043, 3186] (matches Llama-2 canonical IDs), non-ASCII round-trips via byte fallback, "The quick brown fox" → forward through 48-layer 370M SSM → argmax decodes (#136)
2026-04 — Mamba-3 persistent state buffers — Mamba3State : IDisposable owns per-layer ssm_state + cum_angle (NativeMemory.AlignedAlloc, 64-byte) and threads them across Mamba3TransformerModel.Forward(..., Mamba3State) calls for stateful multi-call sequence advance. Building block for streaming decode (full streaming kernel — k_state + v_state + chunk-boundary shifted_γ continuation — is the next stage). State threading is deterministic bit-for-bit on re-run; prefill+decode tracks one-shot up to the canonical chunk-edge lookahead drift, which is documented and pinned by a regression-ceiling assertion. Verified on the real ib-ssm/mamba3-370M-10BT checkpoint (gated by DOTLLM_IBSSM_CHECKPOINT_PATH) (#136)
2026-04 — Mixtral-family MoE support — dense-routing top-k Mixture-of-Experts for Mixtral-convention models (Mixtral, Qwen*-MoE without shared experts, Phi-3.5-MoE). New MoeConfig on ModelConfig (NumExperts, NumExpertsPerTok, MoeIntermediateSize), Architecture.Mixtral enum variant, HfConfigExtractor detects num_local_experts / num_experts + num_experts_per_tok and surfaces Phi-3.5's moe_intermediate_size override. MoeSwiGluMlp kernel: full softmax over experts → top-k partial max-scan (stable tiebreak: lower index wins, matching torch.topk) → renormalise by sum (Mixtral convention, NOT a second softmax) → per-expert SwiGLU MLP via existing FusedOps.SwiGLU → weighted sum. TransformerModel.Forward branches on TransformerLayerWeights.Moe; safetensors loader resolves block_sparse_moe.gate + experts.{j}.w1/w2/w3, F16/BF16 → F32 upcast at load time. Verified against real yujiepan/mixtral-tiny-random (config + detection) and a synthetic 2-layer, 4-expert, top-2 fixture (full forward pass). Out of scope: shared experts (DeepSeek-V3), Qwen-MoE mlp.experts naming adapter, fused GroupedGEMM, expert parallelism, real Mixtral-8x7B validation (#136)
2026-04 — Mamba-3 architecture (prefill) — pure-SSM Mamba-3 end-to-end loading and prefill forward pass. Mamba3ConfigExtractor parses HF config.json (model_type=mamba3), Mamba3WeightLoader ingests F32 safetensors, and Mamba3TransformerModel : IModel composes embed → N × (RMSNorm + Mamba3Block.Forward + residual) → final RMSNorm → LM head. Canonical state-spaces/mamba semantics: 8-slice in_proj (z, x, B, C, dd_dt, dd_A, trap, angles), per-token per-head A = -softplus(dd_A) clamp, data-dependent RoPE with tanh(angle)·π·DT cumsum, attention-style SSD scan, SISO and MIMO entry points. ModelLoader.LoadFromSafetensors now dispatches Architecture.Mamba3. Prefill only — persistent decode state, tokenizer ingest, and generation loop are follow-ups. Verified end-to-end against canonical Python fixtures, a synthetic tiny checkpoint, and the real 1.55 GB ib-ssm/mamba3-370M-10BT checkpoint (48 layers × 32 heads × d_state=128 × vocab=32000, gated integration test via DOTLLM_IBSSM_CHECKPOINT_PATH) (#136)
2026-04 — Safetensors loader for dense transformers — ModelLoader.LoadFromSafetensors + TransformerModel.LoadFromSafetensors ingest HuggingFace model.safetensors + config.json for Llama/Mistral/Phi/Qwen. HfConfigExtractor mirrors the GGUF extractor pattern over HF JSON fields (hidden_size, num_hidden_layers, num_key_value_heads, rope_theta, tie_word_embeddings, …). bf16 tensors are upcast into 64-byte-aligned scratch at load time; F32 tensors are zero-copy mmap views. ModelLoader.Load(path) auto-detects .gguf vs .safetensors. Verified end-to-end on hf-internal-testing/tiny-random-LlamaForCausalLM
2026-04 — First public release (v0.1.0-preview.1) — dotLLM goes public. NuGet packages for all 10 libraries + DotLLM.Cli as a global dotnet tool. Self-contained single-file downloads for Windows / Linux / macOS (Apple Silicon) and experimental Native AOT builds for Linux / Windows attached to every GitHub Release. Companion website at dotllm.dev (#119)
2026-04 — Wave 7: CPU performance cleanup pass — TopKSampler replaces full Array.Sort with a hand-rolled size-K min-heap (O(N log K), stack-resident scratch); JsonSchemaConstraint adds first-char bucketing to skip the ~160 MB of struct clones per mask build when the tracker rejects most leading characters, plus LRU eviction instead of the previous full-flush cache overflow; Dequantize.Q5_0 gains an AVX2 path matching Q8_0's throughput (reuses MatMulQ5_0.ExtractQ5HighBits / vpshufb bit-extraction); BpeTokenizer pre-splits special tokens via the existing Trie.TryMatchLongest instead of the O(n × m) linear scan; ComputeThreadPool now pins the caller (inference) thread to the first candidate P-core on first Dispatch, eliminating the hybrid-CPU stall where pinned P-core workers idled at the barrier waiting for an E-core caller. New BenchmarkDotNet suites for TopK sampling, schema mask build, and special-token encode (#109)
2026-04 — Phase 7 begins: Logprobs — OpenAI-compatible logprobs: true + top_logprobs: N (0-20) on /v1/chat/completions and /v1/completions. Per-token log-softmax captured before sampling, returned in both streaming SSE chunks and non-streaming responses. Chat UI gains opt-in logprobs visualization: color-coded token confidence (green/lime/yellow/orange/red), hover tooltips with top-K alternatives and probabilities, diagnostic cues for low confidence, ambiguity, and sampling effect. DotLLM.Sample.Logprobs console sample with ANSI-colored output (#101)
2026-04 — Phase 6 complete: Speculative decoding — draft-verify-accept loop with modified rejection sampling. A small draft model proposes K candidate tokens; the target model verifies in one batched forward pass. Greedy fast-path for temperature=0. IKvCache.Rollback() for KV-cache truncation on rejection. IDecodingConstraint.Clone() for constraint rollback. Serve UI gains draft model selector and K slider. --speculative-model and --speculative-k CLI options. Speculative acceptance rate in response timings (#98)
2026-04 — Paged KV-cache — block-based KV-cache memory management (the allocation half of PagedAttention): shared block pool, block tables, ref counting, copy-on-write. Foundation for advanced prefix sharing (step 37, hard requirement) and speculative decoding (step 43, cheap rollback/fork). Attention kernels still operate on contiguous buffers via staging-buffer gather — true paged attention kernels are a future step. --paged (opt-in for CLI), on by default for serve. --no-ui flag for API-only hosting. See docs/KV_CACHE.md (#96)
2026-04 — Native AOT (experimental) — opt-in dotnet publish -p:PublishAot=true produces a single-file dotllm binary with ~50ms startup (vs ~500ms JIT). Source-generated JSON serialization across all projects, CreateSlimBuilder for ASP.NET, TrimmerRoots.xml for Spectre.Console.Cli type preservation. JIT remains default for best throughput (Dynamic PGO). See docs/AOT.md (#94)
2026-04 — Phase 6 begins: Warm-up — configurable dummy inference passes at server startup trigger .NET Tier-1 JIT promotion (Dynamic PGO) and exercise CUDA/cuBLAS pipelines. /ready probe gates on warm-up completion. --no-warmup to disable, --warmup-iterations N to configure (#92)
2026-04 — Phase 5 complete: simple prompt caching completes the constrained decoding & API phase — all 7 steps done (JSON mode, JSON Schema, regex/CFG, tool calling, API server, chat UI, prompt caching)
2026-04 — Simple prompt caching — PrefixCache keeps live KV-cache instances across generation calls. On each turn, element-wise prefix match finds cached tokens and skips redundant prefill, processing only new suffix tokens. LRU eviction with configurable max sessions. Enabled by default in chat and serve commands (--no-prompt-cache to disable). Cached token count reported in CLI stats, API timings.cached_tokens, and Chat UI stats bar. Near-100% cache hit rate in multi-turn chat, dramatically reducing TTFT on subsequent turns (#90)
2026-04 — Built-in web chat UI — dotllm serve model.gguf starts the API server and opens a browser to a bundled single-page chat UI (vanilla JS + TailwindCSS, embedded as resources in the DLL). Per-message inference stats (prefill/decode tok/s, TTFT), live sampling parameter control, model hot-swap from the UI, system prompt, verbose mode. New endpoints: /props, /v1/config, /v1/models/available, /v1/models/load. Streaming SSE now includes usage + timings in the final chunk (#86)
2026-04 — ASP.NET OpenAI-compatible API server — DotLLM.Server with /v1/chat/completions (streaming SSE + non-streaming), /v1/completions, /v1/models, /v1/tokenize, /v1/detokenize, health/ready probes. Chat template formatting, tool calling with IToolCallParser detection, response_format constrained decoding. Model loading at startup via --model/--device CLI args. Sequential request processing via semaphore (#84)
2026-04 — Tool calling — IToolCallParser implementations for Llama 3.1+, Hermes/Qwen, Mistral, and generic fallback. Auto-detection factory selects parser from model architecture and chat template content. ToolCallSchemaBuilder generates JSON Schema from tool definitions for constrained decoding (tool_choice=required). ToolCallDetector for post-generation detection, StreamingToolCallAccumulator for streaming. --tools and --tool-choice CLI options with multi-turn tool use in chat REPL. Parallel tool calls supported (#82)
2026-04 — Regex + CFG constrained decoding — RegexConstraint compiles patterns to minimized DFA (Thompson NFA → subset construction → Hopcroft minimization) with equivalence-class compression. GrammarConstraint parses GBNF grammars into PDA with InlineArray-based call stack. Both use zero-alloc struct simulators and dictionary-cached token masks. --response-format regex --pattern <pattern> and --response-format grammar --grammar <gbnf|@file> CLI support (#80)
2026-03 — JSON Schema constrained decoding — JsonSchemaConstraint layers schema tracking on JsonCharParser to enforce type constraints, required properties, enum values, nested structures. Schema compiled into flat node array with property-name tries. Zero-alloc Clone() via struct-copy. --response-format json_schema --schema <json|@file> CLI support (#78)
2026-03 — Phase 5 begins: JSON mode constrained decoding — JsonConstraint FSM guarantees syntactically valid JSON output via per-token vocabulary masking. Stack-based PDA (RFC 8259), AVX2-vectorized logit masking, state-keyed mask cache. --response-format json_object CLI flag (#76)
2026-03 — KV-cache quantization: Q8_0 and Q4_0 KV-cache compression on CPU and GPU (3.7–7.1× memory reduction). Separate --cache-type-k/--cache-type-v options, mixed-precision window --cache-window N keeps recent tokens in full precision. Dual-region storage with quantize-on-evict, per-tile dequantization in tiled attention (#74)
2026-03 — CPU/GPU hybrid layer offloading: --gpu-layers N to run first N layers on GPU, remainder on CPU. Automatic FP16→FP32 hidden state transfer at boundary. Split KV-cache (GPU FP16 + CPU FP32). Partial VRAM usage proportional to offloaded layers (#72)
2026-03 — CUDA GPU backend: PTX kernels via CUDA Driver API P/Invoke (no native shared library), cuBLAS HGEMM for prefill, custom quantized GEMV for decode (Q8_0, Q4_K, Q6_K), FP16 activation pipeline, on-the-fly weight dequantization, GPU KV-cache, --device gpu CLI flag, --device both benchmarking (#70)
2026-03 — NUMA-aware threading: adaptive spin-wait dispatch (generation counter with event fallback), NUMA topology detection (Windows/Linux), P-core/E-core awareness, CPU affinity pinning, auto-reduced decode thread count (#57)
2026-03 — Operator fusion: fused RMSNorm+quantize (decode-only, eliminates normOut intermediate buffer) and tiled SwiGLU (1KB L1-resident sigmoid buffer) reduce DRAM roundtrips on the decode hot path (#56)
2026-03 — Fast approximate exp/softmax: Schraudolph IEEE-754 bit-manipulation trick replaces polynomial exp (~3 SIMD ops vs ~12) in attention softmax. AVX2/AVX-512 fused shift+exp+sum pass eliminates 3 separate TensorPrimitives calls. Sampling softmax keeps full precision (#55)
2026-03 — Tiled attention with online softmax: O(N) memory flash-attention-style algorithm replaces O(N²) score matrix materialization, eliminates 64 MB/head allocations at ctx 4096, uses ~1 KB stack per head (#54)
2026-03 — Row-interleaved weight repacking: R4 layout stores 4 consecutive rows' blocks contiguously at model load time, improving cache/TLB locality for all quantized GEMV kernels (#52)
2026-03 — Q8_1 input quantization: precomputed block sums for Q5_0 kernels, 2-block loop unrolling, eliminates ~4 SIMD ops/block from Q5_0 vec_dot hot path (#51)
2026-03 — Fused decode dispatch: Q/K/V (3→1) and Gate/Up (2→1) projection fusion saves ~72 dispatches/layer, ~4% decode throughput improvement (#50)
2026-03 — Phase 2 complete: additional model architectures (Mistral, Phi, Qwen), sliding window attention, fused QKV support, IModel interface, ModelLoader helper (#34)
2026-03 — Streaming token generation: IAsyncEnumerable<GenerationToken> API with UTF-8-safe incremental text, CancellationToken support, and per-token finish reason/timings (#31)
2026-03 — Chat template engine: Jinja2-subset interpreter (lexer→parser→evaluator), IChatTemplate implementation, GgufChatTemplateFactory, dotllm chat REPL command (#30)
2026-03 — Mixed quantization + Q8_K: Q8_K input quantization (float32 scale, 256-element blocks, precomputed bsums), true 4-row fused K-quant kernels, re-enabled Q4_K×Q8_K/Q5_K×Q8_K/Q6_K×Q8_K fused GEMV/GEMM (#29)
2026-03 — Q4_K_M dequantization and vec_dot kernels: Q4_K, Q5_K, Q6_K scalar + AVX2 dequant and fused matmul kernels with full model-level dispatch (#28)
2026-03 — BDN inference benchmarks: end-to-end benchmarks with custom tok/s columns, auto model download, llama.cpp comparison script (#42)
2026-03 — Engine inference timings: InferenceTimings on InferenceResponse, onTokenGenerated callback, CLI refactored to use TextGenerator (#41)
2026-03 — Multi-threaded CPU inference: zero-alloc ComputeThreadPool with delegate* dispatch, parallel GEMV/GEMM and head-parallel attention (#36)
2026-03 — SIMD kernel tuning: FMA float accumulation, 4-row batched GEMV, AVX-512 paths, SIMD quantization (#26)
2026-03 — Phase 1 complete: sampling pipeline + stop conditions — first coherent multi-token generation (#24)
2026-03 — KV-cache: eval drops from 1091 ms/token to 227 ms/token (~4.8× speedup)
2026-03 — Llama forward pass: first token generation from embedding to logits
2026-02 — BPE Tokenizer with SentencePiece and tiktoken support (#16)

Roadmap

Phase	Description	Status
1 — End-to-End Generation	GGUF loading, dequantization, CPU ops, tokenizer, attention, forward pass, KV-cache, sampling	Done (9/9)
2 — Practical Local Inference	Engine metrics, benchmarks, Q4_K_M, chat templates, streaming, multi-threading, more architectures, safetensors loader	Done (12/12)
3 — CPU Performance	Decode dispatch, Q8_1 input, weight repacking, outer-product GEMM, tiled attention, fast exp, fusion, NUMA	In Progress (7/8)
4 — GPU Acceleration	CUDA backend, CPU/GPU hybrid, KV-cache quantization	Done (3/3)
5 — Constrained Decoding & API	JSON mode, JSON Schema, regex/CFG, tool calling, OpenAI API server, chat UI, prompt caching	Done (7/7)
6 — Improved Serving	Warm-up, NativeAOT, paged KV-cache, speculative decoding	Done (4/4)
7 — Diagnostics & Interpretability	Logprobs, hook system, logit lens, SAE integration, LoRA adapters	In Progress (2/5)
8 — Model Expansion	MLA attention, ALiBi, SmolLM3, Gemma 4, Mixture of Experts, Mamba-3	In Progress (4/6)
9 — Production Serving	Continuous batching, prefix sharing, advanced scheduling, rate limiting, metrics & tracing	Done (5/5)
10 — Qwen3MoeHybrid (Gated DeltaNet + MoE)	`qwen35moe` architecture: alternating GDN linear-attention / full GQA + shared sparse MoE FFN. CPU bit-exact vs llama.cpp; CUDA forward path (MoE GPU dispatcher live, F32 KV cache for full-attn layers); Vulkan compute shaders (multi-token GDN scan, on-device decay/sigmoid). Real-GGUF GPU parity tests pending.	In Progress (5/5 impl, parity pending)

See docs/ROADMAP.md for detailed steps, dependencies, and milestones.

Documentation

Contributing

Contributions are welcome! dotLLM uses an issue-driven workflow — every change starts with a GitHub issue describing the work. Pick an existing issue or open a new one, then submit a PR targeting main.

Contact

Questions, ideas, or feedback? Open a thread in GitHub Discussions.

Author

Built by Konrad Kokosa — .NET MVP, author of Pro .NET Memory Management (2nd ed.), and AI/agents engineer at Nethermind. Over 20 years of .NET performance work.

Website: dotllm.dev
Personal: kokosa.dev
GitHub: @kkokosa

License

dotLLM is licensed under the GNU General Public License v3.0.

Acknowledgments

llama.cpp — reference for GGUF format, quantization kernels, and CUDA implementations
Hugging Face — model ecosystem, transformers reference implementations, tokenizer specs
.NET team — TensorPrimitives, System.Runtime.Intrinsics, MemoryMappedFile, and the runtime that makes this possible

Product	Compatible and additional computed target framework versions.
.NET	net10.0 is compatible. net10.0-android was computed. net10.0-browser was computed. net10.0-ios was computed. net10.0-maccatalyst was computed. net10.0-macos was computed. net10.0-tvos was computed. net10.0-windows was computed.

Compatible target framework(s)

Included target framework(s) (in package)

Learn more about Target Frameworks and .NET Standard.

This package has no dependencies.

Version	Downloads	Last Updated
0.3.0-dev.2987	8	10/16/2026
0.3.0-dev.2979	11	10/16/2026
0.3.0-dev.2974	39	10/15/2026
0.3.0-dev.2973	48	10/14/2026
0.3.0-dev.2965	35	10/14/2026
0.3.0-dev.2959	38	10/14/2026
0.3.0-dev.2948	36	10/14/2026
0.3.0-dev.2945	30	10/14/2026
0.3.0-dev.2942	34	10/14/2026
0.3.0-dev.2930	33	10/14/2026
0.3.0-dev.2925	32	10/14/2026
0.3.0-dev.2922	32	10/14/2026
0.3.0-dev.2920	40	10/14/2026
0.3.0-dev.2918	33	10/14/2026
0.3.0-dev.2910	35	10/14/2026
0.3.0-dev.2908	30	10/14/2026
0.3.0-dev.2902	30	10/14/2026
0.3.0-dev.2899	37	10/14/2026
0.3.0-dev.2897	34	10/14/2026
0.3.0-dev.2895	48	10/14/2026