Module llama_cpp 

llama.cpp-turboquant-hip installer — native Rust CMake source build.

Clones the private fork scooter-lacroix/llama.cpp-turboquant-hip, builds with HIP/ROCm CMake flags per channel policy, and installs binaries to ~/.mlstack/components/llama-cpp/bin/.

Detection Contract

Detection and verification both operate on the same binary contract: llama-cli --help must succeed (exit code 0). This ensures:

• Partial clone artifacts do not count as installed
• Broken executables do not count as installed
• Detection and verification use the same executable

ROCm Channel Policy

Channel	GPU Targets	WMMA FA	Notes
Legacy	gfx1030	No	RDNA2 only
Stable	gfx1030;gfx1100;gfx1101	Yes	RDNA2+RDNA3
Latest	gfx1030;gfx1100;gfx1101;gfx1200	Yes	RDNA2+RDNA3+RDNA4

Validation Assertions

• VAL-INTEGRATION-002: Native installer route accepts llama-cpp
• VAL-INTEGRATION-003: Installed detection succeeds with functional command
• VAL-INTEGRATION-004: Installed detection fails when command is absent
• VAL-INTEGRATION-005: Broken executable does not count as installed
• VAL-INTEGRATION-006: Partial filesystem artifacts do not count as success
• VAL-CROSS-001: Channel selection determines llama.cpp build flags
• VAL-CROSS-003: Install path and detection strategy are consistent
• VAL-CROSS-012: Detection and verification operate on the same binary contract

Structs

LlamaCppConfig

Configuration for the llama.cpp installer.

LlamaCppInstaller

The llama.cpp installer.

PostInstallVerification

Result of stronger-than—help post-install verification.

PrebuiltInstallPlan

ReleaseArchs

ReleaseAsset

ReleaseManifest

ShellCommand

A constructed shell command.

SourceInstallPlan

Enums

HipArchs

GPU architecture specification for HIP builds.

InstallStrategy

RocmChannel

ROCm channel for CMake flag selection.

Constants

DETECTION_BINARY

The primary detection binary.

DETECTION_SUBCOMMAND

The detection subcommand that must succeed for installed status.

Functions

ensure_default_gguf_model

Ensure the canonical verification/benchmark GGUF model exists at ~/.mlstack/models/<FILENAME>, downloading it from the canonical URL if absent. Shared by the llama.cpp installer verification AND the benchmark runner so there is one download path and one model URL. Validates with is_gguf and removes any corrupt/error file left behind. curl -s keeps it silent so it does not corrupt the TUI’s alternate screen.

has_partial_artifacts

Check if partial artifacts exist (clone/build dirs) without a functional binary.

is_llama_cli_functional

Check if the llama-cli binary is functional at the expected install path.

is_llama_cli_on_path

Check if llama-cli is on PATH and functional.

resolve_install_bin_dir

Resolve the install bin directory for llama-cpp.

verify_installed_binary