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

Studio: backfill the DiffusionGemma visual-server on a tag-matching update (#6267)

Browse source 
* Studio: backfill the DiffusionGemma visual-server on a tag-matching update

An install made before the visual-server entered the copy allowlist (#6254)
matches on tag yet lacks the binary, so the update skip never backfills it and
DiffusionGemma fails with "runner not found". Re-extract the bundle when a
matching install is missing it, gated to the fork bundles that ship it so
upstream installs never thrash and so it self-limits once the binary lands.

* Studio: make the visual-server backfill bypass the existing-install short-circuit

The tag-matching skip is checked twice: once in install_prebuilt and again
inside validate_prebuilt_attempts via existing_install_matches_choice, which
only compares metadata and primary executables, not the DiffusionGemma
visual-server. So a stale install missing that binary still raised
ExistingInstallSatisfied and skipped the reinstall. Pass existing_install_dir
as None when a backfill is needed so the bundle actually re-extracts.
This commit is contained in:
Daniel Han 2026-06-12 10:56:09 -07:00 • committed by GitHub
commit d7dd5556bb
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 51 additions and 11 deletions
Download patch file Download diff file Expand all files Collapse all files

62
studio/install_llama_prebuilt.py
View file

@ -6490,6 +6490,31 @@ def validate_prebuilt_attempts(
raise PrebuiltFallback("no prebuilt bundle passed validation")
def diffusion_visual_server_backfill_needed(
install_dir: Path, host: HostInfo, choice: AssetChoice
) -> bool:
"""True when an existing install matches the tag but lacks the DiffusionGemma
visual-server the chosen bundle ships. An install made before the visual-server
entered the copy allowlist matches on tag yet is missing the binary, so the
tag-match skip never backfills it (DiffusionGemma then fails with "runner not
found"). Gated to the fork ("published") bundles that actually carry it, so
upstream installs -- which never ship it -- can't thrash on repeated updates.
Once a re-extract lands the binary this returns False, so it self-limits."""
if choice.source_label != "published":
return False
name = "llama-diffusion-gemma-visual-server" + (".exe" if host.is_windows else "")
if name not in runtime_patterns_for_choice(choice):
return False
for cand in (
install_dir / name,
install_dir / "build" / "bin" / name,
install_dir / "build" / "bin" / "Release" / name,
):
if cand.is_file():
return False
return True
def install_prebuilt(
install_dir: Path,
llama_tag: str,
@ -6528,11 +6553,17 @@ def install_prebuilt(
)
if release_plans and existing_install_matches_plan(install_dir, host, release_plans[0]):
current = release_plans[0]
log(
"existing llama.cpp install already matches selected release "
f"{current.release_tag} upstream_tag={current.llama_tag}; skipping download and install"
)
return
if diffusion_visual_server_backfill_needed(install_dir, host, current.attempts[0]):
log(
f"existing install matches {current.release_tag} but is missing the "
"DiffusionGemma visual-server; re-extracting the bundle to backfill it"
)
else:
log(
"existing llama.cpp install already matches selected release "
f"{current.release_tag} upstream_tag={current.llama_tag}; skipping download and install"
)
return
with tempfile.TemporaryDirectory(prefix = "unsloth-llama-prebuilt-") as tmp:
work_dir = Path(tmp)
probe_path = work_dir / "stories260K.gguf"
@ -6540,12 +6571,19 @@ def install_prebuilt(
release_count = len(release_plans)
for release_index, plan in enumerate(release_plans):
choice = plan.attempts[0]
backfill = diffusion_visual_server_backfill_needed(install_dir, host, choice)
if existing_install_matches_plan(install_dir, host, plan):
log(
"existing llama.cpp install already matches fallback release "
f"{plan.release_tag} upstream_tag={plan.llama_tag}; skipping reinstall"
)
return
if backfill:
log(
f"existing install matches fallback {plan.release_tag} but is missing "
"the DiffusionGemma visual-server; re-extracting to backfill it"
)
else:
log(
"existing llama.cpp install already matches fallback release "
f"{plan.release_tag} upstream_tag={plan.llama_tag}; skipping reinstall"
)
return
log(
"selected "
f"{choice.name} ({choice.source_label}) from published release "
@ -6563,7 +6601,9 @@ def install_prebuilt(
release_tag = plan.release_tag,
approved_checksums = plan.approved_checksums,
initial_fallback_used = release_index > 0,
existing_install_dir = install_dir,
# a backfill must reinstall, so do not let the inner
# existing-install match short-circuit the re-extract
existing_install_dir = None if backfill else install_dir,
)
except ExistingInstallSatisfied:
return