Overview

vllm and sglang engine factories stop re-running fetch_model against the worker URI when initializing native preprocessor mode, and instead consume the already-resolved slug_dir that download_config populates on the frontend. This eliminates a redundant HF download on every model registration and is the natural follow-on to #9707 — the frontend now has every typed file plus harvested sibling (preprocessor_config.json, special_tokens_map.json, …) blake3-verified under ~/.cache/dynamo/mdc/by-slug/<slug>/<mdcsum>/, so the native preprocessor just needs to be pointed at that directory.

Why now

#9610 + #9707 closed the multimodal-completeness gap by harvesting non-typed siblings on both the HF and self-host paths. The TODOs in vllm_processor.py and sglang_processor.py flagged the redundant fetch_model as the next thing to remove once that landed. This is the smallest remaining piece in the gh-8749 stack before the DYN_SELF_HOST_METADATA default-ON flip (#8754 capstone); it's also a prerequisite for that flip to be a no-op in latency terms.

Details

• ModelDeploymentCard::local_dir() -> PathBuf — pure path computation, returns the same path that mdc_slug_dir(slug, mdcsum) produces. No fs side effect (the resolve pipeline owns directory creation).
• Python binding ModelDeploymentCard.local_dir() -> str plus .pyi stub.
• vllm and sglang factories replace the source_path = mdc.source_path(); if not exists: fetch_model(source_path, ignore_weights=True) block with source_path = mdc.local_dir(). The unused fetch_model import is dropped in both.
• TensorBased models skip download_config and would have an empty slug_dir, but both factories already gate on model_type.supports_chat() at the top, which excludes them — no extra guard needed.

Test

Rust unit test (model_card::tests::local_dir_matches_resolve_pipeline_path) asserts the path produced by local_dir() matches what resolve_metadata_files writes into. The existing tests/frontend/test_self_host_metadata.py covers the worker HTTP route end; the factory consumer path is exercised by the real vllm/sglang serve tests in CI (tests/serve/test_vllm.py, tests/serve/test_sglang.py).

Out of scope

• Default-ON flip of DYN_SELF_HOST_METADATA — still gated on auto-detach-cleanup landing.
• Auto-detach-cleanup itself — separate follow-up; the MetadataArtifactRegistry needs a secondary (endpoint, lora_slug) -> Vec<(slug, suffix)> index so the static detach_from_endpoint helper can clean up without holding a LocalModel handle.
• Multimodal positive e2e covering a Qwen-VL through the full self-host path — separate PR.

Related

• DEP (light): Worker self-hosted metadata files #8749 — self-hosted metadata stack
• feat(self-host): worker advertises non-typed metadata siblings (gh-8749) #9707 — worker advertises non-typed metadata siblings (merged 2026-05-27 as squash 53aa9b7)
• fix(self-host): harvest non-weight siblings into slug_dir (gh-8749) #9610 — frontend HF-sibling harvest
• feat(self-host): worker hosts metadata files via system_status_server (gh-8749) #8855, feat(self-host): frontend MDC verify-and-cache pipeline (gh-8749) #9057 — earlier merges in the same stack

Summary by CodeRabbit

Release Notes

• New Features
  • Added local_dir() method to access the local directory containing cached model files, enabling direct access to resolved metadata without requiring runtime fetch operations.