Skip to content

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.

install-kernels: skip mamba_ssm below sm80 (#12921)

Daniel Han8ba9e67f7d3e
diff --git a/studio/backend/tests/test_kernel_install.py b/studio/backend/tests/test_kernel_install.py
index 1e45f1e4fdb..5a6416d7c3f 100644
--- a/studio/backend/tests/test_kernel_install.py
+++ b/studio/backend/tests/test_kernel_install.py
@@ -336,14 +336,30 @@ def test_flash_attn_resolution(env, expected):
assert kernel_install.resolve_wheel_url("flash_attn", env) == expected
[email protected]("name", ["flash_attn", "mamba_ssm"])
@pytest.mark.parametrize("capability", ["7 5", "None"])
-def test_flash_attn_is_skipped_below_sm80(capability, capsys):
+def test_sm80_kernels_are_skipped_below_sm80(name, capability, capsys):
run = _Runner([], capability = capability)
+ assert kernel_install.install_kernel(name, _COLAB, run = run, exists = lambda url: True) == 0
+ assert run.installer_calls == []
+ assert f"skipping {name}, which needs sm80 or newer" in capsys.readouterr().out
+
+
+def test_causal_conv1d_still_installs_below_sm80(uv):
+ uv(False)
+ run = _Runner([False, True], capability = "7 5")
assert (
- kernel_install.install_kernel("flash_attn", _COLAB, run = run, exists = lambda url: True) == 0
+ kernel_install.install_kernel("causal_conv1d", _COLAB, run = run, exists = lambda url: True)
+ == 0
)
- assert run.installer_calls == []
- assert "skipping flash_attn" in capsys.readouterr().out
+ assert run.installer_calls[0][-1].startswith(f"{_CC1D}/causal_conv1d-1.6.1+cu13torch2.10")
+
+
+def test_the_gpu_is_probed_once_for_every_sm80_kernel(capsys):
+ run = _Runner([], capability = "7 5")
+ for name in ("flash_attn", "mamba_ssm"):
+ kernel_install.install_kernel(name, _COLAB, run = run, exists = lambda url: True)
+ assert sum("get_device_capability" in " ".join(c) for c in run.calls) == 1
def test_flash_attn_installs_on_sm80_and_newer(uv):
diff --git a/studio/backend/utils/kernel_install.py b/studio/backend/utils/kernel_install.py
index d401e2ad731..1bb88dae202 100644
--- a/studio/backend/utils/kernel_install.py
+++ b/studio/backend/utils/kernel_install.py
@@ -217,8 +217,22 @@ def resolve_wheel_url(name: str, env: dict[str, str] | None) -> str | None:
return _PINNED[name].wheel_url(env)
+# FlashAttention 2 needs sm80, and unsloth_zoo turns mamba_ssm's fast path off below sm80
+# (patch_mamba_ssm_pre_ampere_fallback), so either wheel would go unused there.
+_NEEDS_SM80 = ("flash_attn", "mamba_ssm")
+_CAPABILITY: dict = {}
+
+
def _gpu_capability(run: Callable[..., subprocess.CompletedProcess]) -> tuple[int, int] | None:
- """The best compute capability across the visible GPUs, or None."""
+ """The best compute capability across the visible GPUs, or None. Probed once per runner."""
+ if run not in _CAPABILITY:
+ _CAPABILITY[run] = _probe_gpu_capability(run)
+ return _CAPABILITY[run]
+
+
+def _probe_gpu_capability(
+ run: Callable[..., subprocess.CompletedProcess],
+) -> tuple[int, int] | None:
check = (
"import torch; n = torch.cuda.device_count() if torch.cuda.is_available() else 0; "
"print(*max(torch.cuda.get_device_capability(i) for i in range(n))) if n else None"
@@ -271,8 +285,7 @@ def install_kernel(
) -> int:
distribution, check = KERNELS[name]
url = resolve_wheel_url(name, env)
- if name == "flash_attn" and url is not None:
- # FlashAttention 2 needs Ampere or newer, and Unsloth only enables it there.
+ if name in _NEEDS_SM80 and url is not None:
capability = _gpu_capability(run)
if capability is None or capability < (8, 0):
gpu = (
@@ -280,7 +293,7 @@ def install_kernel(
if capability is None
else "the best GPU is sm%d%d" % capability
)
- print(f"Unsloth: skipping flash_attn, which needs sm80 or newer ({gpu}).")
+ print(f"Unsloth: skipping {name}, which needs sm80 or newer ({gpu}).")
return 0
torch_desc = f"torch {env.get('torch_version')}" if env else "this environment"
# Only a 404 proves nothing is published; an unreachable check falls through to the install.