Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
"""The CUDA major of an sd.cpp build is the libcudart it links (DT_NEEDED), not the first one a string names.
The CUDA 13 prebuilt's ``libggml-cuda.so`` names "libcudart.so.12" in a dlopen fallback list too; read by string
alone it looked like a CUDA 12 build and was paired with the cu12 cuDNN, which aborts the process on a CUDA 13-only
host ("Invalid handle. Cannot load symbol cublasLtGetVersion", seen on a Colab L4)."""
from __future__ import annotations
import struct
from pathlib import Path
from core.inference import sd_cpp_cudnn as cd
def _elf_with_needed(
path: Path,
needed: list[str],
extra: bytes = b"",
) -> str:
"""A minimal ELF64 LE: a .dynstr, a .dynamic with DT_NEEDED entries, and ``extra`` bytes (rodata strings)."""
dynstr = b"\0"
offsets = []
for name in needed:
offsets.append(len(dynstr))
dynstr += name.encode() + b"\0"
dynamic = b"".join(struct.pack("<qQ", 1, off) for off in offsets) + struct.pack("<qQ", 0, 0)
body_off = 64
dynstr_off = body_off + len(extra)
dynamic_off = dynstr_off + len(dynstr)
shoff = dynamic_off + len(dynamic)
header = bytearray(64)
header[:6] = b"\x7fELF\x02\x01"
struct.pack_into("<Q", header, 0x28, shoff)
struct.pack_into("<HHH", header, 0x3A, 64, 3, 0)
null = struct.pack("<IIQQQQIIQQ", 0, 0, 0, 0, 0, 0, 0, 0, 0, 0)
sec_dynstr = struct.pack("<IIQQQQIIQQ", 0, 3, 0, 0, dynstr_off, len(dynstr), 0, 0, 1, 0)
sec_dynamic = struct.pack("<IIQQQQIIQQ", 0, 6, 0, 0, dynamic_off, len(dynamic), 1, 0, 8, 16)
path.write_bytes(bytes(header) + extra + dynstr + dynamic + null + sec_dynstr + sec_dynamic)
return str(path)
def test_cuda13_build_naming_cudart12_as_fallback_reads_as_13(tmp_path):
lib = _elf_with_needed(
tmp_path / "libggml-cuda.so.0",
["libcublasLt.so.13", "libcudart.so.13", "libcublas.so.13"],
extra = b"GGML_CUDA_CUDNN_LIB\0libcudart.so.12\0libcudart.so.13\0",
)
assert cd._elf_needed(Path(lib)) == ["libcublasLt.so.13", "libcudart.so.13", "libcublas.so.13"]
exe = tmp_path / "sd-server"
exe.write_bytes(b"\x7fELF" + b"\0" * 64)
assert cd.binary_cudnn_build(str(exe)) == (True, 13)
plan = cd.plan_cudnn_attention(
str(exe), (8, 9), allow_install = False, platform = "linux", environ = {}
)
assert plan.state == cd.STATE_UNAVAILABLE and "CUDA 13" in plan.reason
def test_cuda12_build_still_reads_as_12(tmp_path):
_elf_with_needed(
tmp_path / "libggml-cuda.so.0",
["libcudart.so.12", "libcublas.so.12"],
extra = b"GGML_CUDA_CUDNN_LIB\0libcudart.so.12\0libcudart.so.13\0",
)
exe = tmp_path / "sd-cli"
exe.write_bytes(b"\x7fELF" + b"\0" * 64)
assert cd.binary_cudnn_build(str(exe)) == (True, 12)
def test_not_an_elf_keeps_the_string_scan(tmp_path):
path = tmp_path / "sd-cli"
path.write_bytes(b"MZ GGML_CUDA_CUDNN_LIB\0libcudart.so.12\0")
assert cd._elf_needed(path) is None
assert cd.binary_cudnn_build(str(path)) == (True, 12)