Jump to content
Toggle search
Toggle menu
CMP Wiki
Toggle preferences menu
Toggle personal menu
Not logged in
Your IP address will be publicly visible if you make any edits.

Tools

From CMP Wiki
More actions

Reported 8 GB and 10 GB

Software people in the cmpunlocker Discord use to inspect, test, tune and monitor CMP 170HX cards. The unlock itself (cmpunlocker and its forks) is on Driver unlock and Unlocks; the base of the server is https://github.com/amoghmunikote/cmpunlocker. Source: @globalapathy, 2026-10-18 How to put the tools together: Acceptance checklist, Testing and burn-in, Monitoring.

Most tools on this page read the card's registers through BAR0 and need root. The ones marked Warning write to the card; read their page first.

Overview

Tool What it does Cards Author State Link
Inspect a card
gpcprobe Read-only kernel module plus gpc_probe.py (which GPCs and memory partitions are disabled or fused defective) and blackbox.py (InfoROM history: power-on time, Xid log) both @matstatman Reported, one confirmation; last commit 2026-10-20 GitHub
hbmmon nvtop-style monitor with per-HBM-stack temperature and identity, fuse map, PCIe, power 8 GB tested @matstatman Warning writes to the HBM debug bridge unless --read-only; last commit 2026-10-05 GitHub
Register probe script Older script that reports defective versus merely disabled GPCs and memory partitions both unknown, reshared by @xanthefin Reported Source: @xanthefin, 2026-10-17
infinite.im SM table Table of active SMs per NVIDIA and CMP card both @xeax Contested: one row came from an AI and was questioned site
Test a card
cmp170hx-testsuite Acceptance tests: HBM step test, tensor and SM exactness, real inference with weight read-back, Xid check, error map; saves logs per serial both @boz._. (dev-boz) Reported; single commit 2026-10-09 GitHub
club-170hx QC scripts tensor-gate.sh (burn, cross-card matmul correctness, memtest, AER snapshot), compare_vram.cu, xid-watch.sh, pcie-check.sh, set-powercap.sh both PixelML (posted by @c133, @djpandaman) Reported; last commit 2026-10-19 GitHub
memtest_vulkan Vulkan VRAM test; run 5 minutes right after a cold boot both GpuZelenograd Reported; misses some faults, see Testing and burn-in GitHub
gpu-burn CUDA burn-in; -tc uses tensor cores both wilicc Reported GitHub
mmapeak Peak tensor throughput per data type; the test to ask a seller for both ReinForce-II Reported; last commit 2025-12-09 GitHub
torch-performance-test-data PyTorch GEMM benchmark with a table that includes a CMP 170HX row both zzc0721 Reported GitHub
Tune
170tune Clock ceiling and undervolt, live HBM clock, timings, refresh, each behind hot qualification gates; receipt-gated persistence both (10 GB on a fork branch) Cachenetics (studebaker8); fork by vektorprime Warning writes registers; needs a matching unlock build GitHub, fork
LACT GPU control GUI; its curve editor accepts voltage and frequency curves 170tune refuses both LACT project Reported Source: @vcore2191, 2026-10-07
nvidia-pstated Daemon that drops idle GPUs to a low performance state both sasha0552 Theory: not tested on a 170HX GitHub
Fuses (10 GB card, irreversible)
170hx-rir-tool Dumps fuses and RIR records, volatile fuse overrides, permanent RIR burn 10 GB @lxtend (lxwx) Warning one-shot permanent write; last commit 2026-10-21 GitHub
cmp-170hx-rir-burn-util RIR burn with preflight gates plus a PCIe Gen3 driver layer on top of cmpunlocker 10 GB evanrodgers (posted by @hezz1891) Warning one-shot permanent write; last commit 2026-10-20 GitHub
Serving, benchmarks and monitoring
llm-inference-bench Prefill and decode throughput grid, scheduler stress, lil-bench, p2pmark, peer-read stall probe both local-inference-lab Reported; very active GitHub
tool-eval-bench Tool-calling quality benchmark for OpenAI-compatible servers both SeraphimSerapis Reported; run on 2x CMP by @nestaa51 Source: @nestaa51, 2026-10-14 GitHub
llm-visuals Terminal dashboard: speed, TTFT, power, HBM temperature, throttle reasons, VRAM split both DingoOz Reported; HBM readout verified on a 170HX by its author GitHub
Grafana and Prometheus Scrape vLLM's /metrics; lightweight both general Reported Source: @matstatman, 2026-10-04
gds-nvme-patch Restores GPUDirect Storage (NVMe straight to VRAM) on kernels 7.0 and 7.1 both cachenetics Reported; author-tested on one setup GitHub
localairecipes.com Recipe site and leaderboards, filterable by card type and count (hardware=170hx) both @mountains1605 Reported; much of the data is self-reported site

Sources for the rows without their own page:

The others are on each tool's page.

Notes on some of them

  • Probes for the defective-versus-disabled question. @djpandaman: units marked defective have always proven truly defective; disabled units can be enabled. Source: @djpandaman, 2026-10-17 The infinite.im table lists active SMs, so inactive ones may be reactivated if not defective Source: @xeax, 2026-10-18; @emogop questioned its 90HX row and @xeax said that row came from his AI and asked for corrections. Source: @emogop, 2026-10-18 No Windows or EFI tool shows disabled versus defective SMs on GA102 or GA104 yet. Source: @emogop, 2026-10-19
  • A probe request for 10 GB owners. @sekkit_75019 asked 10 GB owners to run gen3_sku_probe.py; no results were posted. Source: @sekkit_75019, 2026-10-04
  • localairecipes.com is run by @mountains1605. Recipes are links to git and Docker Hub with revision history; anyone can read without logging in, submissions use an email link, and bots are welcome. Source: @mountains1605, 2026-10-06
    Source: @mountains1605, 2026-10-06 As of 2026-10-27 it had strong 2, 3 and 4 card recipes but only two single-card ones. Source: @mountains1605, 2026-10-27 Its bot had not been updating existing recipes well; speeds and contexts were due for revision. Source: @mountains1605, 2026-10-27 It also has a benchmarks page, and context length was being added to the recipe cards. Source: @mountains1605, 2026-10-17
    Source: @mountains1605, 2026-10-20 Recipes on this wiki are on All recipes.
  • Fan ducts. 170hx-fanduct is a parametric OpenSCAD duct generator; @c133 used it to make a duct from one 120 mm fan to three cards. Source: @c133, 2026-10-08
    Source: @c133, 2026-10-08 More on Cooling and power.

Engines and software people asked about

Short notes; engine and recipe details belong on the recipe pages.

  • Reported Strata (GPU-as-cache MoE offload) ran slower than vLLM on a 170HX, and the CMP is not officially supported. Source: @_orrorin, 2026-10-11
    Source: @artistosu, 2026-10-11 Moving a card from x4 to x16 gave about 3 percent more decode but about 4 times the prefill. Source: @wakeupmbrc1, 2026-10-07
    Source: @wakeupmbrc1, 2026-10-07 About 48 GB of system RAM is the practical minimum. Source: @xtj7, 2026-10-07 Warning It pulls in further external libraries, unlike vLLM and llama.cpp built from source; one member had an agent audit it before use. Source: @kellertreppe123, 2026-10-14
    Source: @sojun80, 2026-10-14
  • Reported tokenspeed (called tensorfold in the chat) required compute capability 8.9 or newer, so it refused the 170HX; @c133's SM80 pull request was closed by a bot as "won't fix", and it had no pipeline parallel. Source: @mountains1605, 2026-10-05
    Source: @c133, 2026-10-05
    Source: @voidgiraffe, 2026-10-05 Theory On 2026-10-20 @alanoo relayed that TensorFold now ships sm_80 kernels; nobody reported a 170HX run, and the question whether it beats vLLM on these cards is open. Source: @voidgiraffe, 2026-10-04
    Source: @alanoo, 2026-10-20
  • Reported ExLlamaV3 gained DeepSeek V4.1 Flash support and per-card expert placement through @anon33anon's pull requests, without tensor parallel yet; he found vLLM's EXL3 path much slower than native ExLlamaV3. Source: @anon33anon, 2026-10-12
  • Reported Deepinfer-SM80 ports DeepSeek's DeepInfer kernels into vLLM for sm_80: +20 percent prefill, +6 percent decode, validated on 6 GPUs as TP2 x PP3. Source: @reala7096, 2026-10-20
    Source: @reala7096, 2026-10-20
    Source: @mountains1605, 2026-10-20 Only kernels are ported, so the gains are modest. Source: @reala7096, 2026-10-20
  • Reported Why EXL3 is slow on this card. EXL3 decodes its weights with integer math on the regular CUDA cores, and the 170HX has about a third of an RTX 3090's ALU throughput; a measured EXL3 GEMM used only 6 to 15 percent of the memory bandwidth, about the same for 1 or 16 tokens. Marlin INT4 is fast because its dequant is cheap and feeds the tensor cores. Source: @wakeupmbrc1, 2026-10-13
    Source: @wakeupmbrc1, 2026-10-14
  • Reported vLLM run-to-run variance. Re-running the same quantization comparison gave results about 3 percent apart; @vektor_prime traced it to upstream vLLM and fixed it with a batch-invariance change in his recipe repository. Source: @vektor_prime, 2026-10-07
    Source: @vektor_prime, 2026-10-07
  • Contested LMCache for spilling KV cache to CPU RAM: with it vLLM processed one request at a time instead of eight, and @c133 found it also wanted VRAM. Source: @nikhilr1696, 2026-10-05
    Source: @c133, 2026-10-05
  • Reported NeuralFree (EonMux/NeuralFree): a work-in-progress inference framework by @phoenixone7. GGUF support is in; safetensors support and testing are still to come. He split it out of an older trading-bot project and plans to keep supporting it because of prefix-caching problems he hit in vLLM. No CMP numbers posted. Source: repo github.com/EonMux/NeuralFree by EonMux on GitHub, posted by @phoenixone7, 2026-10-23
    Source: @phoenixone7, 2026-10-23
  • Theory EAGLE-4 speculative decoding was flagged as faster than MTP; nobody tried it on a 170HX. Source: @cibernox, 2026-10-20
  • Theory flare-local-inference (NVIDIA) is built for Blackwell; a port to these cards would need a purpose-written engine. Source: @mountains1605, 2026-10-21
    Source: @dufrelic, 2026-10-22
  • Refusal overlays. @mountains1605 uses msuiche's directional overlays (1 to 9 MB files that remove refusals at runtime without touching the weights) and finds tool use better than with uncensored fine-tunes; no measured data. Source: @mountains1605, 2026-10-09
    Source: @mountains1605, 2026-10-09 He put such an overlay on DeepSeek V4.1 Flash 1017 after it refused reverse-engineering work. Source: @cmdrwaka324, 2026-10-23
    Source: @mountains1605, 2026-10-23
  • Harnesses and front-ends. Looping on GLM 5.3 is often a harness problem; omp, DSH and Hermes have their users. Source: @sojun80, 2026-10-20
    Source: @globalapathy, 2026-10-20
    Source: @globalapathy, 2026-10-21 Others named: maki, CadetCoder, Paseo, rameness. Source: @analyticeth, 2026-10-04
    Source: @c133, 2026-10-05
    Source: @phoenixone7, 2026-10-05
    Source: @tekkenz, 2026-10-21 For chat from a phone or browser: Open WebUI, AnythingLLM Mobile (one endpoint only), a Matrix server with Element X, or Mattermost with agents. Source: @segmond, 2026-10-24
    Source: @xeax, 2026-10-25
    Source: @nathang6866, 2026-10-25
    Source: @mfj4, 2026-10-25
  • Persisting the power cap: a systemd service, a crontab entry or a sysvinit service. Source: @robbbbbb7973, 2026-10-15
    Source: @xanthefin, 2026-10-15 See Monitoring.

Bench tools for the cap mod and board work

How to do the x16 cap mod is on Unlocks. What members used:

  • Hot air: an Atten ST-862D (1000 W) works but is slow; @globalapathy uses the medium nozzle at 15 percent air so the capacitors do not blow away, and finds the leftmost group by the mounting hole the hardest. Source: @globalapathy, 2026-10-27
    Source: @globalapathy, 2026-10-27
    Source: @globalapathy, 2026-10-27
  • Iron: a Pinecil or any 60 W or stronger iron with SMD tips; the Sequre S99 (USB-C, C245 tips, 100 W). Source: @civiloid, 2026-10-11
    Source: @globalapathy, 2026-10-11
  • Microsoldering tweezers are essential; the capacitors are almost impossible to handle otherwise. Clean with isopropyl alcohol, foam swabs, a brush and wipes. Source: @_arcanum_, 2026-10-11
    Source: @_arcanum_, 2026-10-11
  • Microscope: @civiloid would rather have a QianLi Gorilla 6558 or Mechanic Robot 6560T with a camera than work off a screen alone. Source: @civiloid, 2026-10-11 For soldering he calls a new Chinese binocular microscope with a camera module the most cost-effective choice: small, good enough optics, and the camera plugs into any screen. QianLi and Mechanic are consumer brands that buy white-label units, so the actual manufacturer is cheaper. Source: @civiloid, 2026-10-14
    Source: @civiloid, 2026-10-14
    Source: @civiloid, 2026-10-14 @_arcanum_ paid about USD 300 on Taobao for the QianLi 6558 UV-light version; the IR module is about 140 more. Source: @_arcanum_, 2026-10-14
    Source: @_arcanum_, 2026-10-14
    Source: @_arcanum_, 2026-10-14 A used Vision Engineering TS-4 is another option; skip the TS-3, which is louder (fans on the supply and the halogen housing) and older in optics and electronics. Source: @cmdrwaka324, 2026-10-14
    Source: @cmdrwaka324, 2026-10-14 On the TS series, video output needs the camera optics (the CS adapter) fitted when bought; scopes in this class run USD 2,000 to 3,000, and eyepiece-only scopes cost about 5x less. Many cheap camera heads send one eye's light fully to the camera, so you lose binocular view while recording. Source: @endsjeremy, 2026-10-14
    Source: @cmdrwaka324, 2026-10-14
    Source: @endsjeremy, 2026-10-14
    Source: @endsjeremy, 2026-10-14
    Source: @endsjeremy, 2026-10-14
  • A solder stencil for the mod is in the channel pins, but making one needs a laser cutter. Source: @globalapathy, 2026-10-27
    Source: @globalapathy, 2026-10-27
  • A video of the mod done quickly by a Canadian GPU repair YouTuber: https://www.youtube.com/watch?v=HO2T4TgNzlQ Source: @_arcanum_, 2026-10-11
  • Replacing 0402 capacitors is home work; reballing a die this size needs a repair shop. Source: @globalapathy, 2026-10-18
    Source: @_orrorin, 2026-10-18
  • Theory Laser soldering with a 1064 nm fiber laser was discussed; without temperature feedback it would burn through boards unless tuned on something of the same thermal mass. Nobody has tried it. Source: @globalapathy, 2026-10-27
    Source: @globalapathy, 2026-10-27
  • Warning SPI flasher voltage. Flashing a VBIOS chip directly needs a programmer at the chip's voltage. Flash chips are 1.8 V or 3 V and a 3 V programmer fries a 1.8 V chip. Source: @wakeupmbrc1, 2026-10-13
    Source: @wakeupmbrc1, 2026-10-13 A CH341-based programmer sold as 3.3 V was really 5 V; a CH347-based one selects voltages in software and is faster. Source: @ichundes, 2026-10-13 The 170HX VBIOS is signed, so flashing a modified image is a separate problem; see Unlocks.

Gaps

  • No shared benchmark suite exists yet. Source: @djpandaman, 2026-10-12
    Source: @boz._., 2026-10-12
  • A community database of gpcprobe outputs was proposed, with a field for where each card was bought; GitHub submissions in a standard format were suggested. Not started. Source: @matstatman, 2026-10-09
    Source: @globalapathy, 2026-10-09 @matstatman planned a script that reads out more than gpcprobe, and @globalapathy offered reports from four 8 GB cards. Source: @matstatman, 2026-10-09
    Source: @globalapathy, 2026-10-09

Related work on other NVIDIA cards

  • Reported Zagitta/gpu-unlock: a Linux kernel module that removes the half-rate tensor-math limit on RTX 3090 class cards with the register approach the CMP unlocks use, without driver patching. It must load before the NVIDIA driver; it only matches RTX 3090 IDs by default. Gain: a few percent prefill when the engine uses BF16 to FP32 FMA; no core or SM unlock. Source: @zagitta, 2026-10-18
    Source: @zagitta, 2026-10-19
    Source: @zagitta, 2026-10-20
    Source: @zagitta, 2026-10-23 @sojun80 used the approach on a 3060 Ti, 3070, 3070 mobile, 3080 and 3080 Ti, and is building a generic unlock tool from cmpunlocker for the 30 and 40 series. Source: @sojun80, 2026-10-21
    Source: @sojun80, 2026-10-21 Details belong on Unlocks.
    • @grapepepper patched his RTX 3090 Ti from the repository; after he shared his results others unlocked various RTX 30-series cards, and he has seen a report of an RTX 4070 Ti SUPER. Source: @grapepepper, 2026-10-20
    • eGPU. Zagitta: someone got it working on an eGPU (see the repo issues). Theory @da6b, whose 3090 sits behind a Thunderbolt 3 bridge, expects to need the Thunderbolt drivers in the initramfs and Thunderbolt security disabled so the tunnel is up before the module loads, possibly with some timing involved; an eGPU on M.2 is a direct PCIe link and simpler. Source: @zagitta, 2026-10-23
      Source: @da6b, 2026-10-23
      Source: @da6b, 2026-10-23
    • Theory @sojun80 thinks the tensor limit applies to the 20 to 40 series, "i think all of them"; he has no 20-series card left to test. Source: @sojun80, 2026-10-21
      Source: @sojun80, 2026-10-21
    • Zagitta on its origin: it builds on the CMP unlock work (the same basis as cmpunlocker, and the CMP 90HX and 50HX unlocks), and nothing in it is new; his only change is doing it without a driver patch, as a standalone early-boot module, an idea he took from the UEFI module the 90HX and 50HX Windows unlocks use. He has no Windows version planned. On 2026-10-24, asked whether it covers only GA102, he said GA104 and GA106 as well; other GPUs in the same host (a 50-series card, for example) are left alone. Source: @zagitta, 2026-10-23
      Source: @zagitta, 2026-10-24
      Source: repo github.com/Zagitta/gpu-unlock by Zagitta on GitHub, posted by @mrwojaketh, 2026-10-24
      Source: @zagitta, 2026-10-24
      Source: @zagitta, 2026-10-24
      Source: @zagitta, 2026-10-24
      Source: @zagitta, 2026-10-25
    • Reported @danske1's before and after numbers on an EVGA RTX 3090 FTW3, 8192 x 8192 x 8192 GEMM, median of 20 iterations x 5 samples: FP16 in with FP32 accumulate 81.95 to 141.07 TFLOPS (+72 %), BF16 / FP32 81.35 to 145.59 (+79 %), TF32 40.46 to 78.76 (+95 %); FP16 with FP16 accumulate (about 148) and plain FP32 (about 30) unchanged. Source: @danske1, 2026-10-23
      Source: @danske1, 2026-10-23
  • Reported JustAResearcher/gpu-unlock-ada (GitHub) takes the tensor unlock to RTX 40-series (Ada) cards. @petrikrohn_26497 reported an RTX 4090 unlocked and posted the repo. Its README credits Zagitta/gpu-unlock for the Linux module and packaging, TOTORONG/AD103_win_unlocker for the AD102 support and WildFlash1st/cmp90hx-unlock for the GA102 path; the module and packaging are GPL-2.0-only and the NVIDIA firmware stays under NVIDIA's own terms. @djpandaman, earlier the same day: the tensor unlock can be done on AD103 too, and he was trying AD102. @sevendaysworth unlocked an RTX 4080 FE by pointing an AI agent at another member's repo (the repo is not named in his post). Not a CMP 170HX tool. A day earlier @zagitta had posted TOTORONG/AD103_win_unlocker on Hugging Face ("Seems someone got ada unlock working maybe") and @dottoreargus replied "Can confirm, it does work". Source: @zagitta, 2026-10-23
    Source: @dottoreargus, 2026-10-23
    Source: @petrikrohn_26497, 2026-10-24
    Source: repo github.com/JustAResearcher/gpu-unlock-ada by JustAResearcher on GitHub, posted by @petrikrohn_26497, 2026-10-24
    Source: @djpandaman, 2026-10-24
    Source: @sevendaysworth, 2026-10-24
  • Open question: has anyone tried the unlock on a GA104 laptop GPU? @da6b asked on 2026-10-23; no answer in the thread. Source: @da6b, 2026-10-23
  • Theory @djpandaman asked AD102 owners (4090, RTX 5000 Ada, 6000 Ada) to run a read-only fuse dump script; @zagitta posted fuse and floorsweep register values from two RTX 3090s for reference. Source: @zagitta, 2026-10-26
    Source: @djpandaman, 2026-10-27
    Source: @djpandaman, 2026-10-27
  • Reverse-engineering links people shared: a laser fault injection rig build (fraktalcyber/lfi-rig) and an agent tool whose list lacks NVIDIA Falcon support; an agent with Ghidra and radare2 can disassemble most things. Source: @mountains1605, 2026-10-09
    Source: @analyticeth, 2026-10-09
    Source: @dottoreargus, 2026-10-14
  • Nobody has a fork that combines P2P and 170tune. Source: @gnasherrr, 2026-10-26

Write-ups on Discord

Summaries of longer write-ups members posted, with links to the originals.

170hx_sweep.py, bar0_write.py and iff_apply.py (10 GB card)

Theory @xeax posted iff_apply_README.md, 170hx_sweep_README.md, bar0_write.py and 170hx_sweep.py with one word: "untested". Source: @xeax, 2026-10-22
Source: @xeax, 2026-10-22
Source: @xeax, 2026-10-22
Source: @xeax, 2026-10-22 @scallion_9883 (Humvee55) had shared a message.txt when @feralzi asked for iff_apply.py, and confirmed the tools are "10gb only". Source: @scallion_9883, 2026-10-22
Source: @scallion_9883, 2026-10-22 What the sweep README describes:

  • Purpose. It tests up to 128 SMs and up to 24 HBM memory partitions, including the ones fused off at the factory, one unbind, reset, override, bind and test cycle at a time, and writes CARD_REPORT.md: a verdict per GPC, TPC and memory partition, the defective units, and the masks for the largest stable configuration of that die. Every test uses volatile overrides through the ga100_rir binary of 170hx-rir-tool, so a reboot or power cycle returns the card to stock; the README says no RIR write path exists in the code. Source: @xeax, 2026-10-22 The tool's software overrides only work once the gate fuse is cleared; see 170hx-rir-tool. Source: repo github.com/lixiangwuxian/170hx-rir-tool by lixiangwuxian on GitHub, posted by @lxtend, 2026-10-10.
  • Requirements. Debian or Ubuntu, root, one card per run, no display session on the card, a working NVIDIA driver. gpu-burn, cuda_memtest and torch are needed for real verdicts; without them tests are logged as skipped. Memory stages above 40 GiB need the patched cmpunlocker driver with the live-FBPA LMR fix (--assume-memory-driver); otherwise they are skipped. By default it only accepts device IDs 0x2082 and 0x2081. Source: @xeax, 2026-10-22
  • Commands. sudo python3 170hx_sweep.py discover (read-only scan of fuses and masks), run (resumable; about 35 short GPU cycles, 45 to 60 minutes), status, report. Useful flags: --quick (skip the per-TPC matrix), --skip-memory, --verbose. Watch dmesg -w and nvidia-smi -l 1 in another terminal. Source: @xeax, 2026-10-22
  • Safety model. Device identity checks before anything runs, a 10-second countdown, every override read back, the kernel log diffed per unit, every change logged before it runs, and resume after a crash or power cycle. Source: @xeax, 2026-10-22
  • bar0_write.py writes one 32-bit register through BAR0 and prints the read-back. The sweep uses it for three SM speed registers (0x823804, 0x82381C, 0x823820); without them revealed SMs stay throttled. Source: @xeax, 2026-10-22
  • iff_apply.py (optional) arms software IFF reset-replay records that the card's firmware re-applies at every reset, described as a cleaner route to Gen3 or Gen4. While a record is armed the sweep refuses to run (iff-show, iff-install, iff-undo). Source: @xeax, 2026-10-22
  • The README calls power readings that jump above about 90 SMs a telemetry artifact and suggests sudo nvidia-smi -lgc 300,1200. Source: @xeax, 2026-10-22

@markshowa asked whether iff_apply's show command would return anything on an 8 GB card; no answer. Source: @markshowa, 2026-10-22

Building cmpunlocker's vGPU-and-others branch

Reported @arabus_74064's build notes (shell commands with comments in Hungarian, summarised in English): Source: @arabus_74064, 2026-10-27
Source: @arabus_74064, 2026-10-27
Source: @arabus_74064, 2026-10-27
Source: @arabus_74064, 2026-10-27

  1. Clone the branch: git clone --branch vGPU-and-others --single-branch https://github.com/amoghmunikote/cmpunlocker.git ~/cmpunlocker-vgpu.
  2. The vGPU patches are driver/patches/vgpu-*.patch, applied in the PATCH_ORDER list in driver/build.sh. Every unlock must also be declared under unlocks: in common/constants.yaml, or tools/read-constants.py rejects the build ("built but not declared", "does not match the patches").
  3. Back up common/constants.yaml, then add three entries before the ecc: block: vgpu_kvm_build (vgpu-kvm-build.patch), vgpu_ga100_host (vgpu-ga100-host.patch) and vgpu_umd_version (vgpu-umd-version.patch), each with registers: {}.
  4. Check with python3 tools/read-constants.py common/constants.yaml driver/patches driver/build.sh 10gb, then install with sudo ./install.sh --profile=10gb.

Whether vGPU then worked was not posted. vGPU itself is on Unlocks.

Model loading time

  • @matstatman's reboots took about 20 minutes because vLLM loaded about 180 GB from a Samsung QVO SATA SSD. Source: @matstatman, 2026-10-16
    Source: @matstatman, 2026-10-16
    Source: @matstatman, 2026-10-16
  • @impactic loads GLM 5.2 Full 4-bit from an M.2 NVMe in about 300 s, and 350 to 380 GB models in under 5 minutes; before optimising he saw 10 to 20 minutes per GLM load. Splitting model shards across several disks speeds up transfers. His guide, model-startup-loading-optimization-public-20261016.md, is attached on Discord. Source: @impactic, 2026-10-16
    Source: @impactic, 2026-10-16
    Source: @impactic, 2026-10-16
    Source: @impactic, 2026-10-16
    Source: @impactic, 2026-10-17
  • @peter643 customised vLLM to checkpoint running models and swap models fast: GLM 5.3 loads from scratch in 30 to 60 s on 4x 170HX, with four Gen4 x4 NVMe drives behind a PEX88096 switch reading about 24 GB/s, and a management layer meant to let agents manage the models they run. On an i7-9700K without enough BAR1 address space he wrote his own P2PDMA, a GDS mailbox and an nvidia-fs driver that uses a 128 MB BAR region. No repository was posted. Source: @peter643, 2026-10-07
    Source: @peter643, 2026-10-07 See also gds-nvme-patch.

@rh3dai's video series (8 GB card)

Source: @rh3dai, 2026-10-04
Source: @rh3dai, 2026-10-07
Source: @rh3dai, 2026-10-11

The settings in the videos are not transcribed here; 170tune covers the tool he tunes with.

Agents driving a PC through an IP-KVM

@con3po7016 posted two write-ups, computer-use-kvm-history.md and GLKVM-AGENT-SETUP_1.md, on letting an AI agent operate a PC through a GL-KVM, and proposed a small working group; the files are on Discord. Source: @con3po7016, 2026-10-10
Source: @con3po7016, 2026-10-10 @wise_axolotl_13963 linked the similar agentkvm project (https://github.com/magicsmokepress/agentkvm). Source: @wise_axolotl_13963, 2026-10-10 @info.at.mah7_78733 built an agent-enabled terminal (ssh, VNC, RDP, NanoKVM) that drives a PC with keyboard, mouse and a vision model. Source: @info.at.mah7_78733, 2026-10-20 A local-model use case, not specific to the cards.

See also

Imported from the cmpunlocker Discord with credit; see Credits.
Retrieved from "https://cmp.wiki/w/index.php?title=Tools&oldid=478"