[zhenshanx-nv]

Background

Per the 2026-10-29 team discussion on Community GPU CI reliability: 30+ failures in the last two weeks were "no available L40" catalog errors -- Brev's catalog search for the standard g6.4xlarge (AWS L4) qualification type comes back empty or duplicated. #1895 already retries that search (PREFLIGHT_ATTEMPTS = 4, up to ~6 minutes), but gives up if the type never regains capacity. This PR adds one same-provider fallback: g6e.4xlarge, which is not a new/unverified SKU -- it's already the manual GPU_PROFILE: large qualification type used today (workflow_dispatch only).

Exit Criteria

• After the primary type's preflight retries are exhausted with a capacity-shaped failure (empty/duplicated catalog), retry once against g6e.4xlarge before giving up.
• A deterministic rejection (bad disk, wrong provider, credential failure) is never retried -- only "no capacity" shapes fall back.
• The two-phase --prepare-only / --from-intent allocation flow stays consistent: whichever type the prepare phase actually selects is the type the reserve phase requests.
• Out of scope: Nebius fallback, OOM mitigation for large models (separate issue raised in the same discussion).

Implementation

• tools/brev_exec.py: INSTANCE_TYPE_FALLBACKS = {"aws": {DEFAULT_INSTANCE_TYPE: "g6e.4xlarge"}}. New _select_instance_with_fallback() wraps the existing _select_instance/_preflight call; on PreflightUnavailable/InventoryError it retries once against the fallback type with host_ram_gib=None (matching how the existing manual large-profile lane already bypasses the single-GPU host-RAM check). provision() now reassigns instance_type to whichever type won and persists it as lease["sku"].
• .github/workflows/community-ci.yml: the prepare step reads back the lease's resolved sku via jq after --prepare-only and sets qualification_type=g6e.4xlarge when a fallback occurred, reusing the existing reserve-step case logic that already handles that exact type.
• tools/tests/test_community_gpu_ci.py: 3 new tests -- fallback succeeds after persistent primary-type outage, fallback is not attempted for a non-AWS provider, and a deterministic disk-mismatch rejection is never retried.

Validation

Commands and Results

• python -m pytest tools/tests/test_community_gpu_ci.py tools/tests/test_community_ci.py tools/tests/test_community_gpu_lease.py: 1248 passed, 1 skipped.
• ruff check / ruff format --check on the changed Python files: clean.
• git diff --check: clean.
• Workflow YAML re-parsed with PyYAML: valid.

Hardware, Environment, and Revisions

Base ci/developer at a4405a42. Python 3.12, CPU only, no GPU used.

Not Run / Remaining Gaps

• No live Brev run against the real catalog.
• test_architecture.py, test_public_source_hygiene.py show unrelated failures in this environment (git grep: failed to create thread: Resource temporarily unavailable) -- confirmed present on the unmodified branch too via a baseline run, so not caused by this change.

Contributor Self-Review

• I have completed a self-review of this change.

Change categories

• Model or runtime behavior
• Public API
• ABI
• Bundle or artifact format
• Dependencies
• Documentation only
• CI or developer tooling

Risk level

• Low
• Medium
• High

Risk rationale: the fallback only fires after the primary type's existing retries are exhausted on a capacity-shaped error, reuses an already-validated instance type, and never retries a deterministic rejection.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

Run configuration

• Source mode: Auto
• Repository: NVIDIA/TensorRT-Model-Connect
• Branch: ci/developer
• File: .coderabbit.yaml
• Commit: a4405a4
• Layers: Repository YAML
• Review profile: CHILL
• Plan: Enterprise
• Run ID: 3156bab8-efc0-45b0-96b8-b3c0b617f5e3

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

_{Comment @coderabbitai help to get the list of available commands.}