mirror of
https://github.com/unslothai/unsloth.git
synced 2026-07-09 15:58:41 +00:00
Studio: remove dead direct_linux_release_plan path (#7030)
parse_direct_linux_release_bundle and direct_linux_release_plan are no longer reached by any live code path. Fork Linux installs resolve through _fork_manifest_release_plans -> _linux_published_attempts, and the upstream (ggml-org) path uses direct_upstream_release_plan. The dead parser also called _resolve_linux_bundle_profile, which no longer exists, so its CUDA branch would raise NameError if ever executed. Drop both functions and the obsolete TestDirectLinuxNvidiaCpuGate; its live equivalent TestLinuxPublishedAttemptsNvidiaCpuGate already covers the NVIDIA no-silent-CPU behaviour.
This commit is contained in:
Daniel Han
GitHub
parent
b5dca66cb1
commit
fb5dc91bb4
2 changed files with 0 additions and 218 deletions
|
|
@ -1286,162 +1286,6 @@ def synthetic_checksums_for_release(
|
|||
)
|
||||
|
||||
|
||||
def parse_direct_linux_release_bundle(
|
||||
repo: str, release: dict[str, Any]
|
||||
) -> PublishedReleaseBundle | None:
|
||||
release_tag = release.get("tag_name")
|
||||
if not isinstance(release_tag, str) or not release_tag:
|
||||
return None
|
||||
|
||||
assets = release_asset_map(release)
|
||||
artifacts: list[PublishedLlamaArtifact] = []
|
||||
inferred_labels: list[str] = []
|
||||
|
||||
linux_asset_re = re.compile(
|
||||
r"^app-(?P<label>.+)-(?P<target>linux-x64(?:-cpu)?|linux-x64-cuda\d+-(?:older|newer|portable))\.tar\.gz$"
|
||||
)
|
||||
for asset_name in sorted(assets):
|
||||
match = linux_asset_re.fullmatch(asset_name)
|
||||
if not match:
|
||||
continue
|
||||
inferred_labels.append(match.group("label"))
|
||||
target = match.group("target")
|
||||
if target in {"linux-x64", "linux-x64-cpu"}:
|
||||
artifacts.append(
|
||||
PublishedLlamaArtifact(
|
||||
asset_name = asset_name,
|
||||
install_kind = "linux-cpu",
|
||||
runtime_line = None,
|
||||
coverage_class = None,
|
||||
supported_sms = [],
|
||||
min_sm = None,
|
||||
max_sm = None,
|
||||
bundle_profile = None,
|
||||
rank = 1000,
|
||||
)
|
||||
)
|
||||
continue
|
||||
|
||||
bundle_profile = target.removeprefix("linux-x64-")
|
||||
profile = _resolve_linux_bundle_profile(bundle_profile)
|
||||
if profile is None:
|
||||
continue
|
||||
artifacts.append(
|
||||
PublishedLlamaArtifact(
|
||||
asset_name = asset_name,
|
||||
install_kind = "linux-cuda",
|
||||
runtime_line = str(profile["runtime_line"]),
|
||||
coverage_class = str(profile["coverage_class"]),
|
||||
supported_sms = [str(value) for value in profile["supported_sms"]],
|
||||
min_sm = int(profile["min_sm"]),
|
||||
max_sm = int(profile["max_sm"]),
|
||||
bundle_profile = bundle_profile,
|
||||
rank = int(profile["rank"]),
|
||||
)
|
||||
)
|
||||
|
||||
if not artifacts:
|
||||
return None
|
||||
|
||||
upstream_tag = (
|
||||
release_tag
|
||||
if is_release_tag_like(release_tag)
|
||||
else inferred_labels[0]
|
||||
if len(set(inferred_labels)) == 1 and inferred_labels
|
||||
else release_tag
|
||||
)
|
||||
selection_log = [
|
||||
f"published_release: repo={repo}",
|
||||
f"published_release: tag={release_tag}",
|
||||
f"published_release: upstream_tag={upstream_tag}",
|
||||
"published_release: direct_asset_scan=linux",
|
||||
]
|
||||
return PublishedReleaseBundle(
|
||||
repo = repo,
|
||||
release_tag = release_tag,
|
||||
upstream_tag = upstream_tag,
|
||||
assets = assets,
|
||||
manifest_asset_name = DEFAULT_PUBLISHED_MANIFEST_ASSET,
|
||||
artifacts = artifacts,
|
||||
selection_log = selection_log,
|
||||
)
|
||||
|
||||
|
||||
def direct_linux_release_plan(
|
||||
release: dict[str, Any], host: HostInfo, repo: str, requested_tag: str
|
||||
) -> InstallReleasePlan | None:
|
||||
bundle = parse_direct_linux_release_bundle(repo, release)
|
||||
if bundle is None:
|
||||
return None
|
||||
if not direct_release_matches_request(
|
||||
release_tag = bundle.release_tag,
|
||||
llama_tag = bundle.upstream_tag,
|
||||
requested_tag = requested_tag,
|
||||
):
|
||||
return None
|
||||
|
||||
attempts: list[AssetChoice] = []
|
||||
if host.has_usable_nvidia:
|
||||
# Prefer the cudart major Studio loads at runtime (torch's bundled
|
||||
# libcudart), not the newest on disk. Otherwise a stray cuda13
|
||||
# runtime outranks the torch cuda12 the binary links against.
|
||||
torch_preference = detect_torch_cuda_runtime_preference(host)
|
||||
selection = linux_cuda_choice_from_release(
|
||||
host,
|
||||
bundle,
|
||||
preferred_runtime_line = torch_preference.runtime_line,
|
||||
selection_preamble = torch_preference.selection_log,
|
||||
)
|
||||
if selection is not None:
|
||||
attempts.extend(selection.attempts)
|
||||
elif not host.has_rocm:
|
||||
# A ROCm-only host gets no CPU asset: leaving attempts empty lets the
|
||||
# raise below trigger a HIP source build instead of shipping a CPU
|
||||
# binary on a GPU host (this ggml-org path has no per-gfx ROCm asset).
|
||||
cpu_choice = published_asset_choice_for_kind(bundle, "linux-cpu")
|
||||
if cpu_choice is not None:
|
||||
attempts.append(cpu_choice)
|
||||
# NVIDIA hosts whose CUDA selection produced nothing fall through to the
|
||||
# raise below (mirroring the ROCm policy above): the caller then walks
|
||||
# back to an older release that still ships a usable CUDA line instead of
|
||||
# silently installing a CPU binary on a GPU host. Today's walk-back only
|
||||
# works because partial releases ship no CPU bundle; this keeps it working
|
||||
# if a future partial release does.
|
||||
if not attempts:
|
||||
raise PrebuiltFallback("no compatible Linux prebuilt asset was found")
|
||||
approved_checksums = synthetic_checksums_for_release(
|
||||
repo,
|
||||
bundle.release_tag,
|
||||
bundle.upstream_tag,
|
||||
)
|
||||
resolved_upstream_tag = bundle.upstream_tag
|
||||
if DEFAULT_PUBLISHED_SHA256_ASSET in bundle.assets and not is_release_tag_like(
|
||||
bundle.upstream_tag
|
||||
):
|
||||
approved_checksums = load_approved_release_checksums(repo, bundle.release_tag)
|
||||
# Require exact source provenance for branch/pull/commit releases.
|
||||
# Mirrors validated_checksums_for_bundle so incomplete metadata fails
|
||||
# closed instead of degrading to the legacy branch-as-tag source
|
||||
# hydration path this PR eliminates.
|
||||
if (
|
||||
not approved_checksums.source_commit
|
||||
or exact_source_archive_hash(approved_checksums) is None
|
||||
or source_clone_url_from_checksums(approved_checksums) is None
|
||||
):
|
||||
raise PrebuiltFallback(
|
||||
f"approved checksum asset {DEFAULT_PUBLISHED_SHA256_ASSET} for "
|
||||
f"{repo}@{bundle.release_tag} did not contain exact source provenance"
|
||||
)
|
||||
attempts = apply_approved_hashes(attempts, approved_checksums)
|
||||
return InstallReleasePlan(
|
||||
requested_tag = requested_tag,
|
||||
llama_tag = resolved_upstream_tag,
|
||||
release_tag = bundle.release_tag,
|
||||
attempts = attempts,
|
||||
approved_checksums = approved_checksums,
|
||||
)
|
||||
|
||||
|
||||
def direct_upstream_release_plan(
|
||||
release: dict[str, Any], host: HostInfo, repo: str, requested_tag: str
|
||||
) -> InstallReleasePlan | None:
|
||||
|
|
|
|||
|
|
@ -2684,68 +2684,6 @@ class TestBlackwellCuda124Exclusion:
|
|||
assert kept == [cpu]
|
||||
|
||||
|
||||
# N.1c3. direct_linux_release_plan -- no silent CPU on NVIDIA hosts
|
||||
|
||||
|
||||
class TestDirectLinuxNvidiaCpuGate:
|
||||
"""A linux-cpu-only release on an NVIDIA host must raise (caller walks back to a usable CUDA line), not silently CPU-install. CPU-only hosts keep the CPU bundle."""
|
||||
|
||||
def _bundle_cpu_only(self):
|
||||
return make_release(
|
||||
[
|
||||
make_artifact(
|
||||
"llama-b8508-bin-ubuntu-x64.tar.gz",
|
||||
install_kind = "linux-cpu",
|
||||
runtime_line = None,
|
||||
coverage_class = None,
|
||||
supported_sms = [],
|
||||
min_sm = None,
|
||||
max_sm = None,
|
||||
bundle_profile = None,
|
||||
),
|
||||
]
|
||||
)
|
||||
|
||||
def _patch(self, monkeypatch):
|
||||
monkeypatch.setattr(
|
||||
INSTALL_LLAMA_PREBUILT,
|
||||
"parse_direct_linux_release_bundle",
|
||||
lambda repo, release: self._bundle_cpu_only(),
|
||||
)
|
||||
monkeypatch.setattr(
|
||||
INSTALL_LLAMA_PREBUILT,
|
||||
"detect_torch_cuda_runtime_preference",
|
||||
lambda host: CudaRuntimePreference(runtime_line = None, selection_log = []),
|
||||
)
|
||||
monkeypatch.setattr(
|
||||
INSTALL_LLAMA_PREBUILT,
|
||||
"detected_linux_runtime_lines",
|
||||
lambda: (["cuda13"], {"cuda13": ["/usr/local/cuda/lib64"]}),
|
||||
)
|
||||
|
||||
def test_nvidia_host_without_cuda_line_raises_for_walkback(self, monkeypatch):
|
||||
self._patch(monkeypatch)
|
||||
host = make_host(driver_cuda_version = (13, 1), compute_caps = ["100"])
|
||||
with pytest.raises(PrebuiltFallback, match = "no compatible Linux prebuilt"):
|
||||
INSTALL_LLAMA_PREBUILT.direct_linux_release_plan(
|
||||
{"tag_name": "b8508"}, host, "unslothai/llama.cpp", "latest"
|
||||
)
|
||||
|
||||
def test_cpu_host_still_gets_cpu_bundle(self, monkeypatch):
|
||||
self._patch(monkeypatch)
|
||||
host = make_host(
|
||||
nvidia_smi = None,
|
||||
driver_cuda_version = None,
|
||||
compute_caps = [],
|
||||
has_physical_nvidia = False,
|
||||
has_usable_nvidia = False,
|
||||
)
|
||||
plan = INSTALL_LLAMA_PREBUILT.direct_linux_release_plan(
|
||||
{"tag_name": "b8508"}, host, "unslothai/llama.cpp", "latest"
|
||||
)
|
||||
assert [a.install_kind for a in plan.attempts] == ["linux-cpu"]
|
||||
|
||||
|
||||
class TestLinuxPublishedAttemptsNvidiaCpuGate:
|
||||
"""Live fork-manifest path: an NVIDIA host whose CUDA selection finds nothing gets an empty attempt list (source-builds with CUDA), not the manifest CPU bundle. CPU-only hosts still get the CPU bundle."""
|
||||
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue