Using custom models.ini in llama.app

[CYPHER0041]

I am using llama.cpp (b10282-a035a8887) compiled from source on my Mac. I have a custom models.ini file defined below.

[unsloth/Qwen3.6-27B-MTP-GGUF:UD_Q6_K_XL]
model = Models/GGUF/Qwen3.6-27B-MTP-GGUF/Qwen3.6-27B-UD-Q6_K_XL.gguf
mmproj = Models/GGUF/Qwen3.6-27B-MTP-GGUF/mmproj-F16.gguf

; CORE TUNING FOR M4 PRO
n-gpu-layers = all            ; Forces full Metal GPU offload (critical for speed)
threads = 8                   ; Matches your 8 Performance cores (avoids E-core contention)
threads-batch = 8             ; Aligns with P-cores for prompt processing
flash-attn = on               ; Reduces KV cache recomputation overhead on Metal
kv-unified = true
; MEMORY BANDWIDTH OPTIMIZATION (273 GB/s limit)
cache-type-k = q4_1           ; Compresses K-cache: ~60% less bandwidth than f16, near-lossless
cache-type-v = q4_1           ; Compresses V-cache: same benefit
kv-offload = true             ; Keeps KV cache on GPU memory (default, but explicit)

; MTP SPECULATIVE DECODING TUNING
spec-type = draft-mtp
spec-draft-n-max = 2          ; Reduce to 1 for long context
spec-draft-p-split = 0.16     ; Higher threshold trusts MTP drafts more aggressively on Metal
spec-draft-p-min = 0.08       ; Prevents fallback to slow greedy decoding too early

temp = 0.7
top-p = 0.8
top-k = 20
presence-penalty = 1.5
min-p = 0.00
chat-template-kwargs = {"preserve_thinking":true}

[bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF:Q6_K]
model = /Models/GGUF/models--bartowski--Kwaipilot_KAT-Coder-V2.5-Dev-GGUF/snapshots/d8f684f08d2950ea9d2db6a35ef7dada0707858b/Kwaipilot_KAT-Coder-V2.5-Dev-Q6_K.gguf
ctx-size = 65536
ubatch-size = 2048

I wanted to try out llama.app on my system and downloaded the binary for mac os. The ini file for llama.app is located at Library/Application Support/Llama/model.ini which had the following content.

[bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF:Q6_K]
model = Models/GGUF/models--bartowski--Kwaipilot_KAT-Coder-V2.5-Dev-GGUF/snapshots/d8f684f08d2950ea9d2db6a35ef7dada0707858b/Kwaipilot_KAT-Coder-V2.5-Dev-Q6_K.gguf
ctx-size = 262144
ubatch-size = 2048

I tried replacing it with my custom file but it keeps reverting to the original file during runtime. I even tried deleting the app, the Llama folder in Application Support and the .llama-app folder in home and then reinstalling the dmg but the same issue persists.

Is there a way for me to use my models.ini file with llama.app instead of the default one?

[Yigtwxx]

That file isn't a config file — it's a build artifact. llama.app regenerates it from its own state, so any hand edit is overwritten the next time the app touches it. Deleting Application Support and reinstalling doesn't help because nothing is being restored; the file is being recreated.

The app writes it in ModelManager.updateModelsFile():

@discardableResult
func updateModelsFile() -> Bool {
  let content = generateModelsFileContent()
  let destinationURL = UserSettings.appSupportDir.appendingPathComponent("models.ini")

  // Skip write if content is identical
  if let existingData = try? Data(contentsOf: destinationURL),
    let existingContent = String(data: existingData, encoding: .utf8),
    existingContent == content
  { return false }

  try content.write(to: destinationURL, atomically: true, encoding: .utf8)
  ...

generateModelsFileContent() builds the whole file from the scanned model cache, and it only ever emits these keys:

content += "[\(model.id)]\n"
content += "model = \(modelPath)\n"
content += "ctx-size = \(tier.rawValue)\n"
if let mmprojPath { content += "mmproj = \(mmprojPath)\n" }
// spec-type / spec-draft-model / spec-draft-n-max when an MTP head is detected
if useLargeBatch { content += "ubatch-size = 2048\n" }

There's no merge step and no hook for extra keys, so everything else in your file — threads, flash-attn, cache-type-k/v, spec-draft-p-split, temp, top-p, chat-template-kwargs — has nowhere to survive. The comparison is against the full generated string, so the moment your edited copy differs (which is immediately), the next syncServer() rewrites it. That runs on model-list refresh, context-tier changes and settings changes, which is why it reverts "during runtime" rather than only at launch.

For reference, the app then hands that path to the server as a preset (LlamaServer.swift):

let presetsPath = UserSettings.appSupportDir.appendingPathComponent("models.ini").path
...
"--models-preset", presetsPath,

So the answer to "is there a way to use my models.ini with llama.app" is no — not through the app. But --models-preset is a normal server flag, and you already have a source build, so run the server yourself and point it at your own file:

llama-server --models-preset ~/models.ini

That gets you the full preset surface, including everything the app's generator doesn't emit. The app's own CLI setup installs the same llama binary if you'd rather use that than your b10282 build.

Two things about your file while you're at it:

• spec-type and spec-draft-n-max are each defined twice in the first section. Worth collapsing so it's unambiguous which value you meant.
• The documented preset format carries a version = 1 header and supports a [*] global section for the settings you're repeating per model — see docs/preset.md and the server README section on model presets. threads, flash-attn, cache-type-k/v and the sampling params are all good candidates for [*].

If you want the app itself to respect a user-supplied preset, that'd be a feature request on ggml-org/Llama-macOS — the generator would need an "append user overrides" step, since today it owns the file end to end.

[CYPHER0041]

Thanks for the response. I am aware of the llama-server flag as I am using it currently. I removed the extra spec declarations (didn't realize that they were there).

I took a look at the macOS repo and I saw an issue regarding adding more cli arguments #71. My problem seems to be another angle to this so I will open a discussion there.