fix(studio): build llama.cpp from master for Gemma 4 support

[danielhanchen]

Summary

• Default llama.cpp build target from latest release tag to master branch
• The latest ggml-org/llama.cpp release (b8635) does not include Gemma 4 support -- model: support gemma 4 (vision + moe, no audio) ggml-org/llama.cpp#21309 merged after the release was cut
• This causes llama-server to fail with unknown model architecture: 'gemma4' when loading any Gemma 4 GGUF
• Temporary fix until the next upstream release includes Gemma 4

Changes

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

Users can still override via UNSLOTH_LLAMA_TAG env var. Once a new upstream release is cut with Gemma 4 support, revert these back to "latest".

Test plan

• Fresh install via install.sh --local builds llama.cpp from master HEAD
• Load Gemma 4 E2B GGUF in Studio -- verify it works (was failing with b8635)
• Load Qwen3.5-4B GGUF -- verify non-Gemma models still work
• UNSLOTH_LLAMA_TAG=b8635 override still works (builds from that tag)

[gemini-code-assist]

Code Review

This pull request updates the default llama.cpp tag from "latest" to "master" across the Python, PowerShell, and Shell setup scripts. Feedback highlights that switching to "master" while using the ggml-org repository forces source builds for all users, leading to significant performance regressions and reduced installation stability. Additionally, there is a suggestion to improve environment variable handling in the Python script to robustly handle empty strings and a concern regarding the maintenance overhead of duplicating the default tag across multiple files.

[gemini-code-assist]

Setting the default tag to "master" while using ggml-org/llama.cpp as the helper repository effectively forces a source build for all users, as ggml-org does not host the Unsloth-specific manifest and checksum files required for prebuilt resolution. This results in a significant performance regression (5-10 minutes per setup run) because the scripts lack logic to skip the build if the existing installation is already at the correct commit. Furthermore, "master" is a moving target, making installations less reproducible and potentially unstable.

[gemini-code-assist]

Setting the default tag to "master" while using ggml-org/llama.cpp as the helper repository effectively forces a source build for all users, as ggml-org does not host the Unsloth-specific dummy manifest and checksum files required for prebuilt resolution. This results in a significant performance regression (5-10 minutes per setup run) because the scripts lack logic to skip the build if the existing installation is already at the correct commit. Furthermore, "master" is a moving target, making installations less reproducible and potentially unstable.

[gemini-code-assist]

The default tag is now duplicated across three files (install_llama_prebuilt.py, setup.sh, and setup.ps1), which increases maintenance overhead. Additionally, using os.environ.get(..., "master") will return an empty string if the environment variable is set but empty, which might lead to unexpected behavior in normalized_requested_llama_tag (which defaults to 'latest' for empty strings). Using or ensures that empty strings also fall back to the intended default value.

Suggested change

	DEFAULT_LLAMA_TAG = os.environ.get("UNSLOTH_LLAMA_TAG", "master")
	DEFAULT_LLAMA_TAG = os.environ.get("UNSLOTH_LLAMA_TAG") or "master"