ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Studio: extend the pinned Blackwell GPU fallback to CUDA 13.0 drivers (#5920)

Browse source 
The pinned b9360 cuda-13.1 Windows fallback for Blackwell (sm_120) only fired
when the driver advertised CUDA >= 13.1. The mainstream Blackwell branch ships
the r580 driver that reports CUDA 13.0, so those hosts missed the pin, were
gated off the in-release 13.3 build, and dropped to the CPU-only cuda-12.4
build.

b9360's binary is native sm_120a SASS (no PTX, so no JIT) and its bundled
runtime is cuda-13.1 cudart; both run on a 13.0 r580+ driver under CUDA
minor-version compatibility. Lower _PINNED_BLACKWELL_DRIVER_FLOOR to (13, 0) so
the pin covers the 13.0 branch too. The generic published-runtime gate stays
conservative (an unverified in-release 13.1 build is still gated off 13.0); only
the specific, hash-pinned, SASS-verified b9360 build is extended.

Refs #5887.
This commit is contained in:
Daniel Han 2026-06-01 06:35:28 -07:00 • committed by GitHub
commit f695fbd0fe
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 14 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

7
studio/install_llama_prebuilt.py
View file

@ -237,13 +237,16 @@ _MAX_PROBE_CUDA_MAJOR = 19
# Last ggml-org release whose Windows win-cuda-13 build is still sub-13.3
# (cuda-13.1, b9360, 2026-05-27). Upstream bumped win-cuda-13 to 13.3 at b9365
# and now ships only cuda-12.4 + cuda-13.3. cuda-12.4 predates Blackwell (ggml
# compiles sm_120 only at toolkit >= 12.8), so a Blackwell host on a 13.1/13.2
# compiles sm_120 only at toolkit >= 12.8), so a Blackwell host on a 13.0/13.1/13.2
# driver is gated off 13.3 and would drop to a CPU-only 12.4 build. b9360 is
# immutable, so we pin its cuda-13.1 build (plus paired cudart) as a GPU
# fallback for exactly those hosts. See unslothai/unsloth#5887.
_PINNED_BLACKWELL_FALLBACK_TAG = "b9360"
_PINNED_BLACKWELL_FALLBACK_RUNTIME = "13.1"
_PINNED_BLACKWELL_DRIVER_FLOOR = (13, 1)
# Floor at 13.0: b9360 ships native sm_120a SASS (no PTX/JIT) and a bundled
# cuda-13.1 cudart, both of which run on a CUDA 13.0 r580+ driver via CUDA
# minor-version compatibility, so the mainstream 13.0 Blackwell branch is covered.
_PINNED_BLACKWELL_DRIVER_FLOOR = (13, 0)
_BLACKWELL_MIN_SM = 120
# ggml compiles Blackwell sm_120 only at toolkit >= 12.8, so an in-release
# windows-cuda build at or above this already covers Blackwell and makes the

14
tests/studio/install/test_selection_logic.py
View file

@ -2021,7 +2021,7 @@ class TestWindowsCudaAttempts:
class TestPinnedBlackwellCudaFallback:
"""A Blackwell host on a 13.1/13.2 driver, gated off the in-release 13.3
"""A Blackwell host on a 13.0/13.1/13.2 driver, gated off the in-release 13.3
build, gets the pinned immutable b9360 cuda-13.1 GPU build instead of the
CPU-only cuda-12.4 drop. The pin is dormant for everyone else."""
@ -2072,15 +2072,19 @@ class TestPinnedBlackwellCudaFallback:
# Ada/Hopper run the cuda-12.4 build fine; the pin must not fire.
assert _pinned_windows_cuda_fallback(self._win_host((13, 1), [sm]), []) is None
def test_pin_not_offered_to_driver_13_0(self):
# 13.0 cannot run the 13.1 build (forward minor); residual CPU gap.
def test_pin_offered_for_driver_13_0(self):
# b9360 is native sm_120a SASS (no JIT) and ships a cuda-13.1 cudart,
# both of which run on a 13.0 r580+ driver via CUDA minor-version
# compatibility. 13.0 is the mainstream Blackwell branch, so it must fire.
assert (
_pinned_windows_cuda_fallback(self._win_host((13, 0), ["120"]), []) is None
_pinned_windows_cuda_fallback(self._win_host((13, 0), ["120"]), [])
is not None
)
def test_pin_not_offered_below_floor(self):
# 12.x predates Blackwell entirely; the pin stays dormant below 13.0.
assert (
_pinned_windows_cuda_fallback(self._win_host((12, 8), ["120"]), []) is None
_pinned_windows_cuda_fallback(self._win_host((12, 9), ["120"]), []) is None
)
def test_pin_not_offered_without_driver(self):