feat: cache NVIDIA GPU attestation collateral

[kvinwang]

Summary

Builds on #789 and removes NVIDIA's live OCSP/RIM services from the warm GPU-VM boot path.

• add dstack-gpu-attest-proxy, a persistent PCCS-like cache:
  • POST /ocsp keys responses by the DER OCSP request list, excluding the per-request nonce
  • GET /v1/rim/<id> caches version-addressed NVIDIA RIM documents
  • cache metadata/body digests survive restarts, expired entries fail closed, and per-kind entry counts are bounded
  • OCSP lifetime is capped by the signed nextUpdate (or one hour from thisUpdate when absent)
• add optional nvidia_attestation_proxy_url to VMM config and .sys-config.json
• route nvattest OCSP and RIM calls through the configured proxy
• package NVIDIA's allow_trust_outpost_ocsp.rego in the GPU image and select it only in proxy mode
• patch the pinned NVIDIA SDK to enforce thisUpdate/nextUpdate with OCSP_check_validity, so accepting a cached nonce mismatch cannot turn an old signed good response into an indefinite replay
• document deployment, cache behavior, and the sys-config trust boundary

The proxy never signs or rewrites collateral. The guest still verifies NVIDIA signatures, certificate/RIM chains, response freshness, measurements, CC state, and the independent fresh GPU evidence nonce added by #789.

Configuration

Run the proxy on a host/fleet address reachable during early guest boot:

dstack-gpu-attest-proxy \
  --listen 0.0.0.0:8090 \
  --cache-dir /var/cache/dstack/gpu-attest-proxy

Then configure the VMM:

[cvm]
nvidia_attestation_proxy_url = "http://10.0.2.2:8090"

See dstack/gpu-attest-proxy/README.md for TTLs and environment variables.

Verification

• cargo fmt --all -- --check
• cargo clippy -p dstack-gpu-attest-proxy -p dstack-util -p dstack-types -p dstack-vmm --all-targets -- -D warnings
• cargo test -p dstack-gpu-attest-proxy -p dstack-util -p dstack-types -p dstack-vmm
  • proxy: 10 tests
  • dstack-types: 30 tests
  • dstack-util: 47 tests
  • dstack-vmm: 42 tests
• python3 -m py_compile dstack/scripts/bin/dstack-cloud
• applied the Yocto patch with git apply --check against NVIDIA SDK commit 9d12801cea8a198ea0f29640dfaf8a4017c841c5
• live NVIDIA OCSP check using the repository's GH100 certificate chain:
  • first request: X-Dstack-Cache: MISS
  • a second request for the same CertID with a different nonce: X-Dstack-Cache: HIT
  • cached DER remained openssl ocsp-verifiable as good with the original signed nextUpdate

Stacked on #789; review only the final commit in this PR.

[Copilot]

Pull request overview

Adds a persistent “PCCS-like” cache service for NVIDIA GPU attestation collateral (OCSP + RIM) and wires it through VMM/sys-config so GPU guests can complete local attestation without relying on NVIDIA’s live services during warm boot.

Changes:

• Introduces dstack-gpu-attest-proxy HTTP service with bounded persistent caches for OCSP and RIM collateral.
• Adds nvidia_attestation_proxy_url plumbing (VMM config → .sys-config.json → guest nvattest invocation), including selecting NVIDIA’s Trust Outpost OCSP policy when proxy mode is enabled.
• Updates Yocto packaging and docs; applies an NVIDIA SDK patch to enforce OCSP thisUpdate/nextUpdate freshness validation.

Reviewed changes

Copilot reviewed 17 out of 19 changed files in this pull request and generated 9 comments.

Show a summary per file

File	Description
os/yocto/layers/meta-nvidia/recipes-graphics/nvattest/nvattest_2026.06.09.bb	Applies OCSP freshness patch and installs the Trust Outpost policy into the GPU image.
os/yocto/layers/meta-nvidia/recipes-graphics/nvattest/files/0001-validate-ocsp-response-freshness.patch	Patches NVIDIA SDK to enforce OCSP validity windows even when nonce can’t match.
dstack/vmm/vmm.toml	Documents the new optional VMM config knob.
dstack/vmm/src/config.rs	Adds nvidia_attestation_proxy_url to VMM config struct.
dstack/vmm/src/app.rs	Exposes nvidia_attestation_proxy_url via sys-config; adds coverage in tests.
dstack/scripts/bin/dstack-cloud	Supports emitting nvidia_attestation_proxy_url into generated .sys-config.json.
dstack/gpu-attest-proxy/src/ocsp.rs	Implements OCSP request key derivation (nonce-agnostic) and cache expiry calculation from signed timestamps.
dstack/gpu-attest-proxy/src/main.rs	CLI entrypoint / configuration for running the proxy daemon.
dstack/gpu-attest-proxy/src/lib.rs	HTTP server, upstream fetching, caching logic, and request routing.
dstack/gpu-attest-proxy/src/cache.rs	Persistent cache storage with metadata + body digests and bounded eviction.
dstack/gpu-attest-proxy/README.md	Documents deployment, cache behavior, and operational headers.
dstack/gpu-attest-proxy/Cargo.toml	New crate manifest for the proxy.
dstack/dstack-util/src/system_setup.rs	Routes nvattest OCSP/RIM calls through proxy when configured; selects Trust Outpost policy.
dstack/dstack-util/Cargo.toml	Adds url dependency for proxy URL validation.
dstack/dstack-types/src/lib.rs	Adds nvidia_attestation_proxy_url to sys-config schema.
dstack/Cargo.toml	Adds gpu-attest-proxy to the Rust workspace.
dstack/Cargo.lock	Locks new crate and dependency graph updates.
docs/tutorials/vmm-configuration.md	Documents how to enable the GPU attestation cache/proxy.
docs/security/cvm-boundaries.md	Documents the trust boundary implications of nvidia_attestation_proxy_url.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.