Kingy AI
Subscribe, 2.1M subscribersWork with Kingy
Skip to content

AI Fix Finder / Model memory debugger

GGUF preflight and model-load debugger

Inspect a GGUF header on your device before loading a model. Compare its declared context and architecture with your runtime settings, then check a load error locally. For Ollama, paste a running-model status snapshot to compare reported context and CPU/GPU placement with your intended settings. Verified Ollama and LM Studio cases distinguish model lookup from context failures. A model file’s size is only part of its memory requirement. Check the error first, then compare weights, KV cache and runtime allowances for a supported configuration.

Published October 16, 2026 Pacific. Ollama and LM Studio repairs and browser-local GGUF preflight added October 17, 2026. The Ollama loaded-runtime check was added the same day. This tool reuses retained measurements; their original dates and tested limits remain unchanged.

Free browser-local debugger · llama.cpp, Ollama and LM Studio

Why won’t this model run?

Inspect a GGUF header on your device, compare it with your runtime settings, and check a model-load error, a KV-cache allocation failure, or a prompt that exceeds its context. Ollama and LM Studio checks use their own verified error shapes; their memory fit remains uncalibrated. Compare memory components using the existing Kingy fit estimator.

GGUF files, running-model status and log text stay in this tab. Diagnosis makes no API calls, runs no models or commands, and saves no inputs. Downloaded diagnoses exclude your log, version text, filename and model metadata strings and runtime identifiers.

Load a retained case summary or enter your setup. Examples replay existing evidence.

GGUF preflight · before you load a model

Select one GGUF file to inspect its header and metadata in this browser. No upload or model loading. This checks little-endian GGUF v3 within a 16 MiB prefix; other formats or larger headers remain unresolved.

The check reads at most 16 MiB, skips tokenizer contents, and stops after five seconds. It does not checksum weights, validate every tensor payload, or prove that a file is complete, authentic, compatible or safe to run. File size is not runtime RAM or VRAM.

No GGUF selected. Log diagnosis and reference comparisons work without a file.

Ollama loaded-runtime check

Paste a running-model snapshot to compare the context and placement you intended with what Ollama reports. This browser does not connect to your local server or run commands.

Use output from ollama ps, or the JSON response from curl http://localhost:11434/api/ps, copied on your device. JSON gives numeric context/memory fields. CLI formats without a CONTEXT column leave context unknown. Remove private model names if needed.

No runtime snapshot supplied.

A pasted snapshot can be stale. Reported runtime memory is not GGUF file size, free memory or a fit guarantee. JSON memory proportions do not prove layer placement; exact CPU/GPU labels come from supported CLI output. Ollama digests are model references, not a verified GGUF file checksum.

This field is a reminder; the browser cannot verify your build.

The catalog contains exact retained Q4_K_M files. A matching name alone does not verify your file.

Placement and execution settings

Fit estimates cover one CUDA GPU, full model offload, logical batch 2048 and one slot. Other settings, Apple unified memory and CPU-only setups receive log checks, without a fit verdict. Ollama and LM Studio receive runtime-specific log checks only; the llama.cpp CUDA calibration is not transferred to them. The reference machine has Linux, a Ryzen 9 7950X, an RTX 4090 and 128 GB installed RAM.

Up to 40,000 characters. For Ollama and LM Studio, paste the complete error JSON object, without an HTTP header prefix. For llama.cpp, include buffer allocation lines and actual n_ctx / n_ctx_per_seq when available. Leave out keys, private paths and prompts. The tool never estimates prompt tokens from character counts.

Loading local debugger…

Evidence, versions and limits

Loaded-runtime checker v1 reuses the retained October 17 Ollama 0.40.5 Qwen3-4B CPU snapshots: 512 allocated tokens rejected the 518-token input; the unchanged task passed at 2,048. Snapshot parsing and configuration comparisons are local checks, not new inference tests. Placement mismatches report a difference from your stated goal; they do not prove a GPU defect or provide a verified GPU repair. Ollama running-model API and context and placement documentation.

GGUF preflight v1 reads declared architecture, weight file type, training context, dimensions and shard metadata, then checks tensor-directory structure and offsets within the file. These declarations are untrusted; matching a name, size or header does not establish the retained artifact SHA-256. KV precision, allocated context and GPU placement are runtime choices, not facts established by a GGUF header. Unsupported architectures and split files receive no file-specific fit verdict. Official GGUF specification.

Ollama and LM Studio repairs were verified October 17, 2026 on one Mac using the same existing Qwen3-4B Q4_K_M file. Three baseline and corrected attempts plus a reverted control are retained per qualified repair. These are same-agent repeats, without independent human or second-machine validation. Context corrections preserve the complete original source and require all three exact facts. HTTP status alone is not task acceptance.

The estimator reuses release fit-release-2026.08.27.5 and estimator v0.2. Its retained CUDA calibration covers exact Qwen3.8-27B, Qwen3.6-27B and Gemma 4 31B-it Q4_K_M files on the Kingy RTX 4090 system, using llama.cpp b10453, commit 3cb7ffb1a1f612d5e4a46244ae5a3c77ad934a70. Measurements were retained in August 2026; publishing this tool does not refresh them. Other hardware comparisons are uncalibrated estimates.

The context repair used a separate Qwen3-4B CPU test on an M4 Pro, llama.cpp b11429, October 15, 2026. Its 2,048-token correction is specific to that fixed task. Increasing context can increase memory requirements. Successful allocation and correct task output are separate checks.

Check a GGUF before loading it

Select one local GGUF v3 file. The browser reads at most 16 MiB to inspect declared architecture, weight file type, training context, dimensions, shard metadata and the tensor directory. The file stays on your device. Unknown formats, larger headers and split-model memory fit remain unresolved.

Header inspection does not verify a full-file checksum or the tensor payloads. The model’s weight quantization differs from runtime KV-cache precision. Allocated context and GPU placement still need runtime settings or logs. A matching header does not establish the identity of a tested Kingy artifact.

Check what Ollama actually loaded

Paste running-model status from ollama ps or /api/ps, select the model and enter the context or placement you intended. The local check reports supported fields, flags observed disagreements and links applicable retained repair evidence. It does not contact your server, execute commands, verify file identity or infer that your hardware fits.

What this check can tell you

A KV-cache allocation failure means a cache buffer could not be allocated. A missing GGUF file or an unsupported architecture needs a different investigation. A context-size rejection concerns the rendered prompt and the allocation; increasing that allocation can require more memory.

The local check recognizes these conditions and links the matching retained repair. Its CUDA fit comparisons cover three exact Q4_K_M files. Your hardware, file identity, runtime version, free memory and task quality still need verification.

Keep the original task when testing a change

In the retained Gemma 4 case, 64K F16 allocation failed while 32K F16 and 64K Q8_0 passed on the stated RTX 4090 system. Choose a smaller context only if the complete input and output allowance still fit. A different KV precision can change output quality.