From 1ce83c40aa6087445703d009d904d793b76b2c30 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Thu, 2 Apr 2026 09:45:56 -0700
Subject: [PATCH] fix(studio): build llama.cpp from master instead of latest
 release tag (#4790)

The latest ggml-org/llama.cpp release (b8635) does not include Gemma 4
support (ggml-org/llama.cpp#21309 merged after the release was cut).
This causes `llama-server` to fail with "unknown model architecture:
gemma4" when loading Gemma 4 GGUFs.

Temporarily default _DEFAULT_LLAMA_TAG to "master" so all new installs
build from the llama.cpp master branch which includes Gemma 4 support.
Once a new upstream release is cut with Gemma 4, this can be reverted
back to "latest".

Changes:
- setup.sh: add _DEFAULT_LLAMA_TAG="master" maintainer default
- setup.ps1: add $DefaultLlamaTag="master" maintainer default
- install_llama_prebuilt.py: change DEFAULT_LLAMA_TAG fallback to "master"

Users can still override via UNSLOTH_LLAMA_TAG env var.
---
 studio/install_llama_prebuilt.py | 2 +-
 studio/setup.ps1                 | 3 ++-
 studio/setup.sh                  | 5 ++++-
 3 files changed, 7 insertions(+), 3 deletions(-)

diff --git a/studio/install_llama_prebuilt.py b/studio/install_llama_prebuilt.py
index 898606514e..61e57614a8 100755
--- a/studio/install_llama_prebuilt.py
+++ b/studio/install_llama_prebuilt.py
@@ -59,7 +59,7 @@ def env_int(name: str, default: int, *, minimum: int | None = None) -> int:
     return value
 
 
-DEFAULT_LLAMA_TAG = os.environ.get("UNSLOTH_LLAMA_TAG", "latest")
+DEFAULT_LLAMA_TAG = os.environ.get("UNSLOTH_LLAMA_TAG", "master")
 # Force all installs to use mainline llama.cpp from ggml-org.
 # Previously: DEFAULT_PUBLISHED_REPO = os.environ.get("UNSLOTH_LLAMA_RELEASE_REPO", "unslothai/llama.cpp")
 DEFAULT_PUBLISHED_REPO = "ggml-org/llama.cpp"
diff --git a/studio/setup.ps1 b/studio/setup.ps1
index 130b66c7df..396a04486f 100644
--- a/studio/setup.ps1
+++ b/studio/setup.ps1
@@ -29,6 +29,7 @@ $PackageDir = Split-Path -Parent $ScriptDir
 # --------------------------------------------------------------------------
 $DefaultLlamaPrForce = ""
 $DefaultLlamaSource = "https://github.com/ggml-org/llama.cpp"
+$DefaultLlamaTag = "master"
 
 # Verbose can be enabled either by CLI flag or by UNSLOTH_VERBOSE=1.
 $script:UnslothVerbose = ($env:UNSLOTH_VERBOSE -eq '1')
@@ -1597,7 +1598,7 @@ if (-not (Test-Path $UnslothHome)) { New-Item -ItemType Directory -Force $Unslot
 $LlamaCppDir = Join-Path $UnslothHome "llama.cpp"
 $NeedLlamaSourceBuild = $false
 $SkipPrebuiltInstall = $false
-$RequestedLlamaTag = if ($env:UNSLOTH_LLAMA_TAG) { $env:UNSLOTH_LLAMA_TAG } else { "latest" }
+$RequestedLlamaTag = if ($env:UNSLOTH_LLAMA_TAG) { $env:UNSLOTH_LLAMA_TAG } else { $DefaultLlamaTag }
 # Force all installs to use mainline llama.cpp from ggml-org.
 # Previously: $HelperReleaseRepo = if ($env:UNSLOTH_LLAMA_RELEASE_REPO) { $env:UNSLOTH_LLAMA_RELEASE_REPO } else { "unslothai/llama.cpp" }
 $HelperReleaseRepo = "ggml-org/llama.cpp"
diff --git a/studio/setup.sh b/studio/setup.sh
index 9096c3bb19..116dc12e08 100755
--- a/studio/setup.sh
+++ b/studio/setup.sh
@@ -14,9 +14,12 @@ RULE=$(printf '\342\224\200%.0s' {1..52})
 #
 #   _DEFAULT_LLAMA_PR_FORCE : PR number to build by default ("" = normal path)
 #   _DEFAULT_LLAMA_SOURCE   : git clone URL for source builds
+#   _DEFAULT_LLAMA_TAG      : llama.cpp ref to build ("latest" = newest release,
+#                             "master" = bleeding-edge, "bNNNN" = specific tag)
 # ──────────────────────────────────────────────────────────────────────────
 _DEFAULT_LLAMA_PR_FORCE=""
 _DEFAULT_LLAMA_SOURCE="https://github.com/ggml-org/llama.cpp"
+_DEFAULT_LLAMA_TAG="master"
 
 # ── Colors (same palette as startup_banner / install_python_stack) ──
 if [ -n "${NO_COLOR:-}" ]; then
@@ -478,7 +481,7 @@ LLAMA_SERVER_BIN="$LLAMA_CPP_DIR/build/bin/llama-server"
 _NEED_LLAMA_SOURCE_BUILD=false
 _LLAMA_CPP_DEGRADED=false
 _LLAMA_FORCE_COMPILE="${UNSLOTH_LLAMA_FORCE_COMPILE:-0}"
-_REQUESTED_LLAMA_TAG="${UNSLOTH_LLAMA_TAG:-latest}"
+_REQUESTED_LLAMA_TAG="${UNSLOTH_LLAMA_TAG:-${_DEFAULT_LLAMA_TAG}}"
 # Force all installs to use mainline llama.cpp from ggml-org.
 # Previously: _HELPER_RELEASE_REPO="${UNSLOTH_LLAMA_RELEASE_REPO:-unslothai/llama.cpp}"
 _HELPER_RELEASE_REPO="ggml-org/llama.cpp"