feat: expose v2 CUDA constraints (gpu.allowedCudaVersions / gpu.minCudaVersion) - #85
Open
brodykellish wants to merge 2 commits into
Open
feat: expose v2 CUDA constraints (gpu.allowedCudaVersions / gpu.minCudaVersion)#85brodykellish wants to merge 2 commits into
brodykellish wants to merge 2 commits into
Conversation
…daVersion) rphttp2 2.9.0 nests the CUDA host constraints under the gpu block (they were top-level pre-release and are unrepresentable on CPU workloads), gives serverless a minCudaVersion floor, makes gpu.pools optional on PATCH so a CUDA-only update need not resend the pool list, and makes a pools update replace the GPU selection wholesale (clearing excludedTypes). Adopt all of it: CUDA params on create-pod / create-endpoint / update-endpoint mapped under gpu.*, client-side validation (major.minor grammar, set-XOR-floor) so callers get named errors instead of raw 422s, pools-less gpu PATCH support (also unlocks a count-only patch, previously silently dropped), a _warning when a gpuPoolIds update clears exclusions pinned via console/set-endpoint-gpus, the minCudaVersion availability filter on list-gpu-types, and the refreshed vendored spec (new unmapped ops — ssh-keys, clusters, public template catalog — allowlisted as to-dos in the parity gate).
|
The latest updates on your projects. Learn more about Vercel for GitHub. 1 Skipped Deployment
|
There was a problem hiding this comment.
Pull request overview
This pull request updates the Runpod MCP server’s v2 REST integration to adopt rphttp2 2.9.0’s schema changes for CUDA host constraints, moving allowedCudaVersions / minCudaVersion under gpu.* and extending validation + update behavior accordingly.
Changes:
- Add v2 CUDA constraint params to
create-pod,create-endpoint, andupdate-endpoint, including client-side validation and v1-mode 501s. - Allow pools-less GPU PATCHes for Serverless endpoint updates (CUDA-only / count-only), and emit a warning when a pools update clears existing
excludedTypes. - Extend
list-gpu-typeswith the v2minCudaVersionavailability filter, and refresh the vendored v2 OpenAPI fixture + parity allowlist.
Reviewed changes
Copilot reviewed 9 out of 10 changed files in this pull request and generated 3 comments.
Show a summary per file
| File | Description |
|---|---|
src/_shared/mappers.ts |
Maps new CUDA constraint params into v2 request bodies under gpu.* and adds shared validation helper. |
src/tools/pods.ts |
Exposes CUDA constraint params for pod create and validates v2-only / GPU-only behavior. |
src/tools/endpoints.ts |
Exposes CUDA constraint params for endpoint create/update, supports pools-less GPU PATCH, and adds exclusion-wipe warning behavior. |
src/tools/catalog.ts |
Adds v2 minCudaVersion availability filter support to list-gpu-types with validation. |
tests/mappers.test.ts |
Adds mapper-level tests for CUDA nesting and validation behavior. |
tests/handlers.test.ts |
Adds handler-level tests for v2 wire shape, validation failures, warnings, and v1 rejections. |
tests/spec-parity.test.ts |
Allowlists newly introduced spec operations that are intentionally not yet exposed as MCP tools. |
tests/fixtures/v2-openapi.yaml |
Updates vendored v2 spec snapshot (including new CUDA fields, clusters/account/template catalog additions, etc.). |
tests/fixtures/v2-create-pod.json |
Updates create-pod fixture to include CUDA constraints and expected v2 mapping. |
.changeset/v2-cuda-constraints.md |
Adds a changeset documenting the new user-facing behavior and the rphttp2 2.9.0 dependency. |
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
- update-endpoint 400s an explicit gpuPoolIds: [] instead of silently dropping it (gpu.pools is minItems 1 upstream and not clearable). - allowedCudaVersions description no longer references a get-capacity product parameter that does not exist. - Note on V1PodParams that the CUDA fields are v2-only.
There was a problem hiding this comment.
Pull request overview
Copilot reviewed 9 out of 10 changed files in this pull request and generated no new comments.
Suppressed comments (2)
src/tools/endpoints.ts:368
- The update-endpoint tool description’s v2 mutable-fields list is now incomplete: it omits the newly supported gpuPoolIds/gpuCount/allowedCudaVersions/minCudaVersion (and args). This can mislead tool users about what updates are possible; update the description to include these fields.
"Update a Serverless endpoint's config. On v2 you can change image/disk/env/ports/registry/workers/scaling/networkVolumes/timeout/flashboot; on v1, scaling fields (worker min/max, idle timeout, scaler type/value, name). Only provided fields change. An endpoint's request routing (queue vs load balancer) is fixed at creation and cannot be changed here — recreate the endpoint instead. Note: passing gpuPoolIds replaces the GPU selection wholesale, which clears any GPU-type exclusions set elsewhere (console or set-endpoint-gpus) — the reply carries a _warning when that happens.",
tests/spec-parity.test.ts:140
- The comment above ALLOWLIST_UNMAPPED_OPS still says "Empty today" even though the allowlist is now non-empty, which makes the parity gate’s intent unclear. Update the comment to reflect that this is a temporary to-do list of intentionally unmapped spec operations.
// Spec operations deliberately NOT exposed as a tool. Empty today — every operation
// in the vendored spec maps to a registered tool. Add an entry (with the reason) if
// an operation ships that we choose not to cover.
// New in the 2.9.0-era spec refresh: whole resources the MCP has no tools for
// yet. A to-do list, not a permanent carve-out — delete an entry when its tool
Comment on lines
+115
to
+120
| ? `?include=AVAILABILITY&product=${product}${ | ||
| params.minCudaVersion !== undefined | ||
| ? `&minCudaVersion=${encodeURIComponent(params.minCudaVersion)}` | ||
| : '' | ||
| }` | ||
| : '' |
Member
There was a problem hiding this comment.
nit: can we simplify this nested ternary for readability?
Comment on lines
+510
to
+512
| current?.scaling?.type === 'REQUEST_COUNT' | ||
| ? 'REQUEST_COUNT' | ||
| : 'QUEUE_DELAY'; |
Member
There was a problem hiding this comment.
nit: minor concern around usage of magic strings. Seems like there isnt a project convention, but ideally could be simplified by usaing a const as object pattern?
| pools?: string[], | ||
| count?: number | ||
| params: V2EndpointParams, | ||
| mode: 'create' | 'update' |
Member
There was a problem hiding this comment.
note: comment around typing etc below.
| function endpointCommonToV2(params: V2EndpointParams): Record<string, unknown> { | ||
| function endpointCommonToV2( | ||
| params: V2EndpointParams, | ||
| mode: 'create' | 'update' |
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.
What rphttp2 2.9.0 changed on the wire
The CUDA host constraints (
allowedCudaVersions,minCudaVersion) moved from the top level of pod/endpoint bodies into thegpublock (pre-2.9.0 servers 422 the nested spelling as an unknown property; 2.9.0 servers 422 the old top-level spelling). Serverless gained agpu.minCudaVersionfloor. OnPATCH /v2/serverless,gpu.poolsbecame optional (a CUDA-only patch need not resend the pool list), and sendingpoolsnow replaces the GPU selection wholesale — clearingexcludedTypes.The MCP server needed nothing to survive 2.9.0 (it never sent or read the affected fields over REST); this PR adopts the new schema.
What this adds
allowedCudaVersions/minCudaVersionparams on create-pod, create-endpoint, and update-endpoint, emitted undergpu.*. Client-side validation names the failure instead of relaying a raw 422/400: major.minor grammar (a bare"12"is rejected with the fix in the message), set-XOR-floor exclusivity, GPU-pods-only. Update supports the spec's clear sentinels ([]clears the set,""clears the floor).gpuPATCH: a CUDA-only (or count-only — previously silently dropped) endpoint update no longer resendsgpuPoolIds.update-endpointsendsgpuPoolIdsand the endpoint carriesexcludedTypes(set via console orset-endpoint-gpus), the reply carries a_warningnaming the cleared exclusions.minCudaVersionavailability filter (bare major legal here — it only widens a read).RUNPOD_REST_VERSION=v2(thetemplateIdprecedent).Do not release/deploy this MCP version until prod api.runpod.io runs rphttp2 2.9.0 — pre-2.9.0 servers 422 the gpu-nested fields as unknown properties.
28 new tests (14 mapper, 14 handler); suite 733 tests, 725 pass / 8 pre-existing env-gated skips; tsc + eslint + prettier clean on touched files.