TensorFold

TensorFold serves language models on Apple Silicon and NVIDIA GPUs through an OpenAI-compatible API. Each model family supplies its own kernels and draft verification.

python -m pip install git+https://github.com/ashhart/TensorFold.git
tensorfold serve Vontra/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit

On a Mac, Homebrew installs it too: brew install ashhart/tensorfold/tensorfold.

Use http://127.0.0.1:8080/v1 as the client base URL and the model ID from /v1/models. Both backends serve chat completions, completions and OpenAI's Responses API (/v1/responses); see the API reference. Python 3.11 or newer is required, and MLX 0.32.2 or newer on a Mac (pip installs it). See the runbook for installation and a first request. On NVIDIA GPUs the CUDA kernels need compute capability 8.9 or newer: Ada (RTX 40 series), Hopper and Blackwell, including the DGX Spark's GB10 and the RTX 50 series. NVFP4 and FP8 checkpoints run from 8.9: their own math where the GPU has each mma (FP4 on 12.x, FP8 from 8.9), W4A16 elsewhere; the RTX 40, Hopper and B200 builds are compiled and bit-checked on Blackwell but not yet run on those cards. RTX 30 cards (8.6) aren't supported, and the server refuses a GPU below 8.9 at startup.

Image input

Install the vision extra, python -m pip install 'tensorfold[vision] @ git+https://github.com/ashhart/TensorFold.git', and start a supported GLM-5.3-Flash or Qwen3.5/3.8 dense checkpoint with --vision to accept image and text content parts through the same lane engine; Flash Next CUDA also accepts images with --vision --parallel 2 or more. GLM-5.3-Flash images run on MLX; dense Qwen's run on MLX and CUDA. See image input for the API, checkpoint requirements, cache behavior and qualification status.

Models

tensorfold models lists families and checkpoints. tensorfold info MODEL checks configuration without fetching weights. serve downloads a missing checkpoint; pull downloads it ahead of time.

tensorfold pull Vontra/Qwen3.8-27B-MLX-4bit z-lab/Qwen3.8-27B-DFlash2
tensorfold serve Vontra/Qwen3.8-27B-MLX-4bit

Qwen3.8-27B reads MLX affine 2-, 3-, 4-, 5-, 6- and 8-bit checkpoints, including mixed layer formats. It reads packed rows in groups of 32, 64 and 128 on Apple Silicon and CUDA, with hardware qualification still pending for the newer paths. M5 keeps its native tensor-unit kernels for compatible formats, and other formats use the row decoder. See quantized checkpoints for the exact scope. On CUDA, pull DFlash2 before serving; without it, explicitly choose --no-drafts for the serial reference.

Nemotron uses TensorFold projections and routed-expert kernels. Its load-time row check controls drafting; keep the installed MLX version within the package requirements. The named checkpoint includes mtp-4bit.safetensors, which pull and serve check for.

Flash Next requires 4-bit/group-32 weights. Without an MTP head it can run without MTP drafting on MLX; on CUDA, explicitly pass --no-drafts. On one CUDA GPU it also reads the two NVFP4 exports in the table as they ship, and block-scaled FP8 (ModelOpt FP8_PB_WO) linears in such exports; see the recipe for their formats, the checks they passed and what is not supported. Nemotron CUDA requires 4-bit/group-64 weights and an MTP head unless --no-drafts is set. GLM on MLX reads 4-bit/group-64 weights and mlx-lm's mixed-bit conversions, whose 5-, 6- and 8-bit tensors take their own row kernels; it needs MLX 0.32.2 or later. GLM CUDA reads MLX 4-bit/group-64 weights and the experimental Mia-AiLab/GLM-5.3-Flash-EXL3-TR3-4bpw conversion. GLM's optional incoai/GLM-5.3-Flash-DFlash2 checkpoint has non-commercial license terms, described in third-party notices.

Gemma 4 has no draft head. It drafts copies of its context, and chains from z-lab's DFlash model when served with --drafter z-lab/gemma-4-26B-A4B-it-DFlash (pulled once). Its kernels read 4-bit weights in groups of 32 or 64 with an 8-bit router, as the mlx-community conversion stores them; serve refuses other Gemma 4 layouts before downloading.

DeepSeek-V4-Flash reads the mlx-community conversion (affine 4-bit/group-64 weights, mxfp4 routed experts) and needs MLX 0.32.2 or later. Its draft heads are DeepSeek's DSpark blocks and MTP layer (MIT), converted: tensorfold pull Vontra/DeepSeek-V4-Flash-DSpark-MLX once and serve drafts with it; see its recipe.

See the recipes for supported formats and backend limits.

Exact decoding

A draft is accepted only when it equals the token the same engine would produce serially. Sampling depends on the prompt or explicit seed, absolute position and token ID. Verify kernels keep each row's arithmetic independent of the other rows in the call. Compare a request with the same request using "draft": false to check drafted versus serial output.

The MLX engine can share a round across requests. Each stream keeps its own state and sampling key, with concurrent output required to match its solo output. Load-time checks restrict window width and shared forwards where a family cannot reproduce its serial arithmetic. On CUDA, --parallel N with N greater than one enables shared rounds for Qwen3.8-27B on one or two ranks and for Flash Next and Qwen3.6-35B-A3B on one rank. Flash Next rejects concurrent two-rank execution. GLM and Nemotron CUDA serve one request at a time; CUDA --parallel auto also means one request at a time.

Exactness is against the same engine, weights, runtime and settings. It does not imply identical output between MLX and CUDA, different quantizations, or different tensor-parallel rank counts.

Serve options

The default sampling settings come from generation_config.json. Requests can override sampling and reply length. CUDA does not implement the MLX-only options above. See API fields for request scope.

In a terminal, tensorfold serve keeps one live throughput line under its log; it is off when output is redirected, and TENSORFOLD_NO_LIVE=1 turns it off.

Context and memory

On MLX, omitted --context targets the model's metadata window and reduces it to the startup memory estimate when needed, allowing room to retain a prompt for the next turn. An explicit positive value that cannot fit one request is refused at startup. A larger explicit window may fit a request without leaving room to retain its prompt, so a later turn can need a full prefill. --context 0 removes the metadata cap; finite engine capacity and memory admission still apply. Use the reported context when configuring client compaction.

On CUDA, Qwen defaults to the affordable native capacity. GLM targets a dense 2,051-token window, and Nemotron targets 16,384 tokens; the capacity estimate can lower these defaults. Flash Next's --kv-dtype int8 or int4 counts its smaller cache, so the same memory admits a longer window. Explicit --context 0 targets the affordable native capacity for every CUDA family. A positive CUDA value must fit both the native window and the capacity estimate on every rank; otherwise startup refuses it with fitting guidance. Increasing GLM beyond its dense window enables its sparse-attention path. The startup report distinguishes native and allocated capacity.

MLX defaults to a process budget of 70% of RAM. A family can state a larger share: GLM-5.3-Flash takes 85% on a Mac with 256 GB or less, with nothing else loaded. TENSORFOLD_MEMORY_LIMIT_GB replaces that default in GiB, raising or lowering it; physical RAM and the GPU's recommended working set still cap the result. The server reserves 3 GiB for the rest of the process before setting the MLX allocator limit, so a 110 GiB process budget allows 107 GiB of MLX buffers. Concurrent admission honors the raised budget while accounting for memory held elsewhere on the machine. Admission accounts for weights, cache growth, reply tokens and prefill workspace. Retained prefixes and reusable MLX buffers have separate limits. Admission can evict retained prefixes or queue another stream; fitting weights alone does not establish a usable context size. A larger budget leaves less RAM for other applications and cached file pages. The startup line says how far the variable can raise the budget on this Mac, and a startup refusal says how far it must go. On a 32 GB Mac, Qwen3.8-27B needs more than the default 22.4 GiB, with or without its draft model.

Flash Next's startup weight check excludes n-gram tensors when the loader keeps them in host file mappings. The startup report shows resident and file-backed bytes separately. Cached file pages still consume RAM and can be reclaimed by the OS; see Flash Next memory.

An explicit reply limit is reserved before prefill. A request that exceeds context or memory is refused with fitting guidance; an omitted reply limit is capped by the remaining context. MLX reports a context refusal as HTTP 400 for a non-streamed request or as an error event after opening a stream. CUDA checks context before opening a stream.

On MLX, streams that share rounds take memory as they grow. A stream beside others holds its next 2,048 tokens of growth, not its whole reply, so --parallel streams are admitted while their real contexts fit. Before each round the server checks that the live streams' next growth fits. If it doesn't, it first frees MLX's cached buffers and retained prefixes, but only when that makes room. Then the newest streams wait a round, keeping their state, so their tokens don't change. If even the oldest stream can't grow, the newest one ends with an error that says so; the tokens it already sent stay valid.

The memory-class table below keeps the model combinations under qualification. Its GiB budget ceilings emulate the listed RAM classes before the 3 GiB process reserve. The actual default budget uses OS-reported physical memory and the GPU working set; an explicit memory budget can lower or raise it. Context and peak-memory results remain TBD until a public prompt fixture, checkpoint revision, runtime, command and measurement output accompany each result.

Each model cell needs the fitted context and peak physical process footprint. An emulated budget on a larger host is not a measurement on hardware with that RAM size. These are qualification slots, not minimum-memory promises. Weights that exceed the MLX budget are refused before loading.

The 64 GB row comes from @benwilson's run on a 64 GB M5 Pro with TensorFold 0.3.5.1 and MLX 0.31.2 (#70), not from this release. Each cell is the fitted context at the default budget and the process's lifetime peak footprint over cold and resumed prompts at that context. The Qwen3.8-27B recipe lists the checkpoint revisions, fixture, commands and results. On 0.3.5.1 with DFlash2, prompts above about 100,000 tokens were served but not kept for the next turn (#71).

Prompt caching

On MLX, chunk starts come from the rendered token sequence. Resume points are assistant-message starts and the second message start, using markers discovered from the chat template. The planner skips points less than 256 tokens from the previous chunk start and otherwise cuts at the first eligible point or after the model family's chunk. Qwen3.8 Flash Next measures one chunk at startup and takes the largest of 8,192, 4,096 and 2,048 tokens (4,096 and 2,048 on GPUs without tensor units) that still leaves room for 128K tokens of context; other families use 2,048. Without recognized markers it uses that grid. There is no configurable --prefill-grid option.

Fresh and resumed requests use the same chunk plan. Reuse stops at a matching token prefix and a valid chunk boundary; the previous reply is prefilled again under the current prompt. A template that rewrites an earlier turn can reduce reuse. A follow-up therefore need not reprocess a full grid cell, but short messages or changed earlier text can make it reprocess more than the latest reply and new messages.

Snapshots include the model, runtime, kernel and chunk-plan identity. System prefixes and retained conversations can survive restarts. CUDA engines keep their own prompt/reply states and do not use the MLX disk-snapshot or retained-prefix options.

NVIDIA GPUs

Use NVIDIA's PyTorch container for CUDA, PyTorch, Triton and the extension compiler; the package has no cuda installation extra. Install TensorFold inside the container without replacing that toolchain.

docker run -it --gpus all --ipc=host --network host nvcr.io/nvidia/pytorch:26.07-py3
python -m pip install git+https://github.com/ashhart/TensorFold.git
tensorfold pull Vontra/Qwen3.8-27B-MLX-4bit z-lab/Qwen3.8-27B-DFlash2
tensorfold serve Vontra/Qwen3.8-27B-MLX-4bit --host 0.0.0.0

Qwen3.8-27B, Flash Next and Nemotron support one or two CUDA ranks; GLM requires two. For two ranks, see the CUDA runbook. Each rank needs its checkpoint and any optional drafter. Rank 0 serves HTTP. Unified GPU/host memory also holds runtime buffers and file-backed model data; the startup estimate is not a measured maximum capacity.

Measurements

Each release's notes give its measured decode, prompt and concurrency numbers against the previous release and the standard servers, on the machines they name: see CHANGELOG.md and the GitHub releases. The recipe book gives the public prompts and benchmark command, and each family's recipe keeps its own tables.

Updating

tensorfold update --check checks for a release; tensorfold update installs it (--force reinstalls the newest release even when it is current), then the server must restart. A normal installation uses the same interpreter's pip. An editable clone must be clean and able to fast-forward to the release tag; afterwards run python -m pip install -e . in the checkout to refresh metadata and dependencies. A Homebrew install upgrades with brew upgrade tensorfold instead. --no-update-check or TENSORFOLD_NO_UPDATE_CHECK=1 disables startup checks.

When the update finishes it prints what changed since your version, from CHANGELOG.md, which lists every release. The first time a new version serves, it prints one line linking to its notes.

Development and license

Family interfaces, kernel layout and verification requirements are in the recipe book, family map and kernel map. Apache-2.0 from 0.6.0; see LICENSE, NOTICE and third-party notices. Releases up to 0.5.0 were MIT, and code written before 0.6.0 keeps its MIT notice. Model weights keep their own licenses.