Skip to content

fix(epp): bound prefill reservations for v1.3.0 Telus backport - #1

Closed
sarma-rafay wants to merge 17 commits into
RafaySystems:backport/v1.3.0/epp-pass-full-request-bodyfrom
sarma-rafay:backport/v1.3.0/epp-pass-full-request-body-telus
Closed

fix(epp): bound prefill reservations for v1.3.0 Telus backport#1
sarma-rafay wants to merge 17 commits into
RafaySystems:backport/v1.3.0/epp-pass-full-request-bodyfrom
sarma-rafay:backport/v1.3.0/epp-pass-full-request-body-telus

Conversation

@sarma-rafay

@sarma-rafay sarma-rafay commented Aug 12, 2026

Copy link
Copy Markdown

Summary

Rebuilds backport/v1.3.0/epp-pass-full-request-body-telus from NVIDIA
tmonty12/dyn-12864-prefill-reservations-release-1.3.0 (tip
94901a158) and re-applies Rafay hack/ EPP build commits on top.

Brings the full prefill-reservation stack and the newer hardening
fixes (C-handle scoping, booking cleanup executor, ownership/reaping,
late release, raised admission defaults) onto the Rafay full-body
backport line. BuildOpenAIRequestJSON remains the request-build path
(NVIDIA #11991).

Validation

  • Cherry-picked hack/ commits (5e8a84a0b, 246af95f2, 5eea05f6b)
    onto NVIDIA tip with DCO sign-off; no conflicts.
  • Confirmed BuildOpenAIRequestJSON in disagg/shared.go and
    dynamo_kv_scorer/plugin.go.
  • Confirmed reservation surfaces present:
    booking_executor.go, prefill_scorer.go,
    lib/llm/.../reservations.rs.
  • Diff vs base is the reservation/router stack plus hack/ (expected
    with NVIDIA-as-baseline).

Source: ai-dynamo/dynamo branch
tmonty12/dyn-12864-prefill-reservations-release-1.3.0.

@sarma-rafay
sarma-rafay deployed to external_collaborator August 12, 2026 01:28 — with GitHub Actions Active
@sarma-rafay
sarma-rafay deployed to external_collaborator August 12, 2026 01:28 — with GitHub Actions Active
@github-actions

Copy link
Copy Markdown

👋 Hi sarma-rafay! Thank you for contributing to RafaySystems/dynamo.

Just a reminder: The NVIDIA Test Github Validation CI runs an essential subset of the testing framework to quickly catch errors.Your PR reviewers may elect to test the changes comprehensively before approving your changes.

🚀

avinash-rafay and others added 17 commits August 12, 2026 11:50
…orer (ai-dynamo#11991)

Signed-off-by: Avinash Varma <avinashvarmap@gmail.com>
Co-authored-by: Avinash Varma <avinashvarmap@gmail.com>
(cherry picked from commit ffae596)
Signed-off-by: Thomas Montfort <tjmontfort12@gmail.com>
Signed-off-by: Thomas Montfort <tjmontfort12@gmail.com>
Signed-off-by: Thomas Montfort <tjmontfort12@gmail.com>
Signed-off-by: Thomas Montfort <tjmontfort12@gmail.com>
Signed-off-by: Thomas Montfort <tjmontfort12@gmail.com>
(cherry picked from commit 2ead078)
Signed-off-by: Thomas Montfort <tjmontfort12@gmail.com>
(cherry picked from commit cab8cac)
Signed-off-by: Thomas Montfort <tjmontfort12@gmail.com>
(cherry picked from commit c32ce78)
Signed-off-by: Thomas Montfort <tjmontfort12@gmail.com>
(cherry picked from commit fa3e7bc)
Signed-off-by: Thomas Montfort <tjmontfort12@gmail.com>
Signed-off-by: Thomas Montfort <tjmontfort12@gmail.com>
(cherry picked from commit 3343242)
Signed-off-by: Thomas Montfort <tjmontfort12@gmail.com>
(cherry picked from commit 7b90075)
Signed-off-by: Thomas Montfort <tjmontfort12@gmail.com>
Signed-off-by: Thomas Montfort <tjmontfort12@gmail.com>
(cherry picked from commit ece553d)
Adds a self-contained way to build a drop-in dynamo-frontend image carrying the
EPP KV-router tool-call fix:

- hack/Dockerfile.epp — compiles the EPP binary from a single build context
  (repo root), so a plain `docker build` works (no buildx / named contexts).
- hack/Dockerfile.frontend-overlay — overlays the built /epp onto the upstream
  monolithic frontend image, leaving every other role byte-identical.
- hack/Makefile — orchestrates epp-image -> frontend-image -> push, with
  overridable VERSION / REGISTRY / BASE_IMAGE.
- hack/README.md — usage.

Build: `make -C hack push`
Signed-off-by: Avinash Varma <avinashvarmap@gmail.com>
(cherry picked from commit dcc1cfd)
Signed-off-by: Sarma Ayyadevara <sarma@rafay.co>
deploy/inference-gateway/epp/go.mod declares 'go 1.26.3', but the builder
stage pinned golang:1.25. The build only succeeded because Go's automatic
toolchain switching downloaded go1.26.3 mid-build, which needs network access
in the builder stage and silently floats the compiler version.

Pin the builder to the version go.mod already requires so the image build is
reproducible and does no toolchain download.

Signed-off-by: Avinash Varma <avinashvarmap@gmail.com>
Signed-off-by: Sarma Ayyadevara <sarma@rafay.co>
hack/Dockerfile.epp copied only .cargo/, the root manifests and lib/ into the
Rust builder stage. deploy/inference-gateway/ext-proc is a workspace member in
the root Cargo.toml, so cargo refuses to resolve the workspace without its
manifest on disk and the stage fails before compiling anything:

  error: failed to load manifest for workspace member `/dynamo/deploy/inference-gateway/ext-proc`
  referenced by workspace at `/dynamo/Cargo.toml`

The canonical deploy/inference-gateway/epp/Dockerfile already copies it for
exactly this reason. Mirror that here so `make -C hack push` works.

Signed-off-by: Avinash Varma <avinashvarmap@gmail.com>
Signed-off-by: Sarma Ayyadevara <sarma@rafay.co>
@sarma-rafay
sarma-rafay force-pushed the backport/v1.3.0/epp-pass-full-request-body-telus branch from 5422753 to 54d8577 Compare August 14, 2026 21:12
@sarma-rafay
sarma-rafay deployed to external_collaborator August 14, 2026 21:12 — with GitHub Actions Active
@github-actions github-actions Bot added the documentation Improvements or additions to documentation label Aug 14, 2026
@sarma-rafay sarma-rafay changed the title fix(epp): bound prefill reservations for v1.3.0 Telus backport fix(epp): merge NVIDIA prefill reservation hardening (v1.3.0) Aug 14, 2026
@sarma-rafay sarma-rafay changed the title fix(epp): merge NVIDIA prefill reservation hardening (v1.3.0) fix(epp): bound prefill reservations for v1.3.0 Telus backport Aug 14, 2026
@sarma-rafay sarma-rafay closed this by deleting the head repository Aug 14, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

documentation Improvements or additions to documentation external-contribution fix router

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants