Skip to the content.

Every external source this repository relies on, with its link and what it is used for. Pull requests and issues of 1Cat-vLLM are listed with their authors in CREDITS.md and are not repeated here. The research notes under research/ carry their own reference lists, the longest in the kernel survey.

PLI Labs’ own repositories and packages (the records behind the MEASURED figures)

Source Used for
Volta sPLIt: README, docs/levers.md, docs/method.md, docs/four-card.md, engine/production Every Volta sPLIt figure; the A/B method; the patches
Volta sPLIt on Hugging Face The build’s card
Volta Max uPLIft: README, docs/measurements.md, docs/history.md, docs/acceptance, the closeness test’s preregistration Every Volta Max uPLIft figure; the window method; the closeness judge
Volta Max uPLIft package on Hugging Face The build as one download
Volta Full NVFP4 Every Volta Full NVFP4 figure; the exact fold; the cost of host-RAM experts
deepseek-v4-flash-v100 The July 2026 llama.cpp build; the quantized key-cache bug; the interim EXL3 results
swift-1.5-v100 A model that fits one board
deepseek-v4.1-v100 Our fork of Phil Tompkins’ engine snapshot
pli-labs, research/v100 and its July 2026 case study The earlier V100 notes
Hugging Face deepseek-v4.1-flash-sm70, sm70-fp8-block32-kernels The record of the first twelve-card port and its FP8 kernels
DGX sPLIce, dsv41-window-harness The sister build on DGX Sparks and its fenced-window harness
Volta 12X AmPLIfy: README, docs/measurements.md, docs/history.md, the time budget, kit/ Every twelve-card figure; the three-stage engine change; where the time goes

Engines and kernels

Source Used for
1CatAI/1Cat-vLLM (Apache-2.0) The engine of every build. Pull requests: CREDITS.md
dnv2003/v100-skinny (MIT; fork patches Apache-2.0) The NVFP4 and FP8 kernels for Volta; the measured read ceiling of 879 GB/s; kernel throughput by row count; the comparison with an RTX 5090
vllm-project/vllm (Apache-2.0) The serving engine underneath; Model Runner V2
philbert440/deepseek-v4.1-v100 and the EXL3 pack The engine and weights of our first eight-card build; the design reference for the prefill skip
Elevatormusic/TensorFold #25 Coupled sampled drafts
0rand/deepseek-v41-tensorfold-spark The same idea in a DGX Spark recipe; expert top-p pruning (tested here, lost)
sudoingX/dgx-spark-deepseek-v41 The two-Spark launcher. sudoingX also wrote the p4_kld.py fidelity method behind Volta sPLIt’s next-token agreement figures
Peuqui/triton The Triton fork behind 1Cat’s tensor-core prefill routes
Leonccaa/ReVolta (Apache-2.0) Another V100 inference server; reviewed for ideas during the twelve-card work
mw00/project-maya (MIT), lxe/dsx (MIT) Other engines for the same model family on consumer cards; read as research leads
An NVIDIA developer-forum thread on GLM-5.3-Flash NVFP4 on eight V100s, and a Reddit post by the 1Cat authors on two years of making P100 and V100 cards work (September 2026; not linked: we could not confirm the addresses) Where the 1Cat-vLLM work started for us
ggml; llama.cpp issue #25382 The GGUF and IQ formats; the quantized key-cache bug report

Models and weights

Source Used for
deepseek-ai/DeepSeek-V4.1-Flash-1017 (MIT) The model, its drafter, config and tokenizer
nvidia/DeepSeek-V4.1-Flash-1017-NVFP4 (MIT) The NVFP4 experts; NVIDIA’s accuracy table (UPSTREAM)
ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF (MIT) The 2-bit experts of Volta sPLIt; the non-expert tensors and Engram tables of every build

Vendor documents

Source Used for
NVIDIA, Tesla V100 GPU architecture whitepaper 80 SMs, 900 GB/s HBM2, about 125 TFLOPS FP16 tensor throughput, the 6 MB L2 cache
NVIDIA, Volta tuning guide What sm_70 has: mma.sync.m8n8k4, shared memory and register limits
NVIDIA, Ampere, Hopper and Blackwell tuning guides What each later generation added
NVIDIA, DGX Spark hardware overview 128 GB of unified memory at 273 GB/s
NVIDIA, CUDA features archive CUDA 13.0 removed offline compilation for Volta
PyTorch, dropping Volta from the CUDA 12.8 binaries Which PyTorch builds still support sm_70

Writing about the hardware

Source Used for
The founder’s public posts and comments on Reddit, February to November 2026: the board post of 2026-03-11, the hardware updates of 2026-11-05 and 2026-11-12, and the comments linked where they are quoted Prices as paid, the power measurement, cooling, the hardware faults, the driver branch. Quoted in part 8 of the front page and in hardware.md
eBay listings of the parts and searches for them (linked in hardware.md) What each part is called on the used market; the price ranges are what such parts sold for in 2026
Tom’s Hardware, 2026-11-09: Lawyer builds private AI server with eight used Nvidia V100 GPUs Independent coverage of the eight-card build
TechCrunch, 2025-02-21: Sakana walks back claims that its AI can dramatically speed up model training Why AI-written kernels need adversarial tests

The skill

PLI Volta Builder: the method of this repository packaged for coding agents. Its optimization loop is borrowed, with credit, from the hill-climbing scripts at https://github.com/0xsero.