feat: resolve oci:// model paths via llmman serve - #4922
Open
ericcurtin wants to merge 1 commit into
Open
Conversation
Contributor
There was a problem hiding this comment.
Pull request overview
This PR adds first-class support for resolving oci:// model references (CNCF ModelPack artifacts in OCI registries) into a local directory path, integrating resolution into the existing get_model() dispatch so it works across LMDeploy entry points. It delegates registry/auth/download behavior to the external llmman CLI and adjusts CLI argument handling to avoid misclassifying OCI tags as host:port server addresses.
Changes:
- Introduce
lmdeploy/oci.pyto detectoci://references and resolve them viallmman resolve, parsing a JSON stdout contract. - Route
oci://...throughget_model()inlmdeploy/utils.pyso all callers gain OCI support. - Update CLI entrypoint parsing to treat
oci://...:tagas a model reference (not a server), and add unit tests covering scheme detection and resolver behavior.
Reviewed changes
Copilot reviewed 4 out of 4 changed files in this pull request and generated 3 comments.
| File | Description |
|---|---|
tests/test_lmdeploy/test_oci.py |
Adds unit tests for oci:// scheme detection, llmman output parsing, and get_model() dispatch behavior. |
lmdeploy/utils.py |
Adds an OCI branch to get_model() to resolve oci:// references via lmdeploy.oci. |
lmdeploy/oci.py |
New module implementing oci:// detection, scheme stripping, llmman invocation, and stdout JSON parsing. |
lmdeploy/cli/entrypoint.py |
Adjusts model_path_or_server handling so oci://...:tag isn’t treated as host:port. |
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
Comment on lines
+45
to
+48
| text = str(model_path) | ||
| if is_oci_ref(text): | ||
| return text[len(SCHEME) :] | ||
| return text |
Comment on lines
+74
to
+81
| path = payload.get("path") | ||
| if not isinstance(path, str) or not path.strip(): | ||
| raise RuntimeError(f"llmman resolve {reference!r}: returned an empty path") | ||
|
|
||
| if not os.path.exists(path): | ||
| raise RuntimeError(f"llmman resolve {reference!r}: reported path {path!r} does not exist") | ||
|
|
||
| return path |
Comment on lines
+91
to
+103
| binary = _llmman_bin() | ||
| if shutil.which(binary) is None and not os.path.isfile(binary): | ||
| raise RuntimeError( | ||
| f"{binary!r} not found. Install llmman " | ||
| "(https://github.com/llmmanorg/llmman) and put it on PATH, or set " | ||
| f"{_BIN_ENV} to its location, to use oci:// model paths." | ||
| ) | ||
|
|
||
| # stderr is inherited so pull progress reaches the terminal; only stdout | ||
| # carries the contract. | ||
| completed = subprocess.run( | ||
| [binary, "resolve", reference], stdout=subprocess.PIPE, stdin=subprocess.DEVNULL, text=True, check=False | ||
| ) |
Lets a model path point at a model published as a CNCF ModelPack OCI
artifact, anywhere a HuggingFace repo id works today:
lmdeploy serve api_server oci://ghcr.io/org/model:tag
Model distribution is increasingly moving to OCI registries, which lets
a deployment reuse the registry, credentials, mirroring and air-gap
tooling it already has for container images.
Acquisition is delegated to a running `llmman serve`, which already
implements the ModelPack media types, registry auth, resumable blob
download and a content-addressed store. The daemon does the pull (POST
/api/pull, streamed as NDJSON so a multi-gigabyte fetch is not silent)
but deliberately exposes no local path, so `llmman resolve --no-pull`
reports where the bytes landed; --no-pull guarantees it only reports on
what /api/pull already fetched. The client is stdlib-only, so no new
dependency.
get_model() is the single dispatch point, so every caller (CLI, vl
builder, turbomind) gets this at once. download_dir/revision/token are
hub concepts with no ModelPack counterpart and are not forwarded.
entrypoint.py also needed a small fix: model_path_or_server treats any
':' as marking a host:port server address, which an oci:// tag would trip
over, so the scheme is matched before that check.
An explicit oci:// scheme is required rather than sniffing a bare
registry/name:tag: that shape is indistinguishable from a HuggingFace
repo id, so guessing would silently hijack existing deployments.
Signed-off-by: Eric Curtin <eric.curtin@docker.com>
ericcurtin
force-pushed
the
feat/oci-modelpack-model-path
branch
from
August 30, 2026 21:35
5db75a1 to
891aa7d
Compare
ericcurtin
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Motivation
Model distribution is increasingly moving to OCI registries -- the same registries, credentials, mirroring and air-gap tooling a deployment already uses for container images. CNCF ModelPack is the spec for that.
This adds an
oci://scheme so such a model works anywhere a HuggingFace repo id does:Modification
lmdeploy/llmman.py(new): a client for a runningllmman servedaemon. Acquisition is delegated rather than hand-rolled -- llmman already implements the ModelPack media types, registry auth, resumable blob download and a content-addressed store. Stdlib-only (urllib), no new dependency:GET /api/versionprobes reachability and identity; a server answering without aversionfield is reported as "not an llmman daemon", worth distinguishing from nothing listening.POST /api/pullstreams NDJSON so a multi-gigabyte fetch is not silent. An error arrives in-band at HTTP 200, and a stream that ends withoutsuccessis also a failure -- both are errors, not a completed pull.llmman resolve --no-pullreports where the bytes landed. The daemon deliberately exposes no local path, so the CLI is the documented interface;--no-pullguarantees it only reports on what/api/pullalready fetched.LLMMAN_HOSTis honoured with llmman's own parsing, including rewriting a wildcard bind (0.0.0.0,[::]) to loopback.lmdeploy/oci.py(new): scheme handling. Kept free of heavy imports and using plainloggingrather thanlmdeploy.utils.get_logger, sinceutilsimports this module fromget_model().lmdeploy/utils.py:get_model()is the single dispatch point -- the CLI,vl/model/builder.pyandturbomind.pyall route through it, so one branch covers every entry point.download_dir/revision/tokenare hub concepts with no ModelPack counterpart and are deliberately not forwarded.lmdeploy/cli/entrypoint.py:model_path_or_servertreats any:as marking ahost:portserver address. Anoci://ghcr.io/org/model:tagcarries one in its tag and would be misread as a server, so the scheme is matched before that check.model_pathneeded no change -- anoci://reference never satisfiesos.path.exists, so it already falls through toget_model.A pull needs both the daemon reachable and the binary on
PATH(orLMDEPLOY_LLMMAN_BIN); each missing piece has its own actionable error. Neither is required unless anoci://reference is actually used.BC-breaking
No. Purely additive; the ModelScope / openmind_hub / HuggingFace branches are unchanged.
Use cases
Pulling model weights in an air-gapped or registry-mirrored cluster without needing HuggingFace reachable, using the credentials already configured for container images. One daemon can serve many lmdeploy processes from a single content-addressed store.
Checklist
ruff formatandruff checkclean on all six touched files.Testing
Two new files.
tests/test_lmdeploy/test_llmman.py(8 cases) runs against a real HTTP server on a loopback port, not mocks, so the NDJSON streaming contract is genuinely exercised:/api/versionaccepted, a non-llmman server rejected, nothing-listening reported actionably; pull success with forwarded byte progress and the exact request body asserted; in-band error at HTTP 200; a stream ending withoutsuccess; non-OK status; a non-JSON diagnostic tolerated.tests/test_lmdeploy/test_oci.py(14 cases): scheme detection incl. case-insensitivity; that a HF repo id, a local path,s3://and anhttp://host:portserver address are not claimed;strip_schemeround-trips; the resolve contract plus eight malformed-output cases; everyLLMMAN_HOSTform incl. wildcard-to-loopback; binary default/override and the missing-binary error; empty reference rejected without touching the daemon; the scheme stripped before hand-off with progress wired; and two dispatch tests assertingoci://never reachessnapshot_downloadwhile a HF repo id still does.Verified honestly:
/api/versionaccept/reject, pull success + progress + request body, in-band error, resolve contract) -- all passlmdeploy.utilsneeds torch and the full stack, unavailable here. The twoTestGetModelDispatchcases in particular are unexercised. Flagging rather than implying coverage I do not have.llmman servebacked by a real registry.