unsloth/studio
Daniel Han 21bdc8fa8c
Studio: treat data-center Blackwell (sm_100/sm_103) as Blackwell in llama.cpp prebuilt selection (#6584)
* Studio: treat data-center Blackwell (sm_100/sm_103) as Blackwell in llama.cpp prebuilt selection

_host_is_blackwell gated on _BLACKWELL_MIN_SM = 120, but data-center Blackwell
parts report a lower compute capability than consumer Blackwell: B100/B200 are
sm_100 and B300/GB300 are sm_103, while RTX 50 is sm_120 and DGX Spark is
sm_121. Because 100 and 103 are both < 120, every data-center Blackwell host was
classified as non-Blackwell, so two GPU-targeting paths never fired for a
B200/B300:

  - the Linux blackwell_runtime_override that prefers the highest CUDA-major
    runtime line shipping a bundle covering the host SMs (so a cu12x torch could
    pin a cuda12 bundle over a native cuda13 one), and
  - _drop_blackwell_incapable_windows_cuda, which removes cuda-12.4 builds that
    load and validate but run Blackwell on a slow PTX-JIT path.

The result is a B200/B300 being handed a prebuilt that does not natively offload
its SM, i.e. the llama.cpp prebuilt is not really for the GPU. The Blackwell
floor is sm_100, so set _BLACKWELL_MIN_SM = 100. The toolkit floor (12.8) is
unchanged and already correct for sm_100/sm_103.

Surfaced loading unsloth/GLM-5.2-GGUF UD-IQ1_S on 8x B200.

Adds tests covering the sm_100/sm_103 classification, the Linux cuda13
preference for a data-center host, and the Windows cuda-12.4 drop.

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Trim comments to be succinct (no behavior change)

* studio: require CUDA 12.9 for sm_103/sm_121 Blackwell prebuilts

sm_103 (B300/GB300) and sm_121 (DGX Spark) have no native compiler
target before CUDA 12.9; the family floor of 12.8 only covers
sm_100/101/120. Make the Windows-CUDA Blackwell filter SM-aware so a
legacy win-cuda-12.8 bundle is dropped on an sm_103/sm_121 host while
sm_100/sm_120 hosts keep the 12.8 floor.

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
2026-06-23 05:15:09 -07:00
..
backend Studio: treat data-center Blackwell (sm_100/sm_103) as Blackwell in llama.cpp prebuilt selection (#6584) 2026-06-23 05:15:09 -07:00
frontend Studio: refresh chat guided tour for the redesigned model picker (#6597) 2026-06-22 22:53:49 -07:00
src-tauri Windows installer: fix DiskPart UAC mid-install, drive-root cache, and spurious unsloth.exe rename warning (#6296) 2026-06-22 03:09:08 -07:00
__init__.py Final cleanup 2026-03-12 18:28:04 +00:00
install_llama_prebuilt.py Studio: treat data-center Blackwell (sm_100/sm_103) as Blackwell in llama.cpp prebuilt selection (#6584) 2026-06-23 05:15:09 -07:00
install_node_prebuilt.py Studio: use an isolated Node.js for the frontend build instead of replacing the system Node/npm (#6533) 2026-06-21 21:17:29 -07:00
install_python_stack.py Windows installer: fix DiskPart UAC mid-install, drive-root cache, and spurious unsloth.exe rename warning (#6296) 2026-06-22 03:09:08 -07:00
LICENSE.AGPL-3.0 Add AGPL-3.0 license to studio folder 2026-03-09 19:36:25 +00:00
package-lock.json ci: advisory lockfile supply-chain audit (no install-script changes) (#5604) 2026-05-19 05:56:56 -07:00
package.json ci: advisory lockfile supply-chain audit (no install-script changes) (#5604) 2026-05-19 05:56:56 -07:00
setup.bat Final cleanup 2026-03-12 18:28:04 +00:00
setup.ps1 studio/setup.sh: guard empty CUDA arch detection in the source build (#5854) (#6481) 2026-06-23 01:26:43 -07:00
setup.sh studio/setup.sh: guard empty CUDA arch detection in the source build (#5854) (#6481) 2026-06-23 01:26:43 -07:00
Unsloth_Studio_Colab.ipynb Fix/studio colab proxy and iframe - Unsloth Studio not loading in Colab (iframe "refused to connect" and wrong URL) (#5844) 2026-05-28 23:54:48 -07:00