Repository navigation
Conversation
Second member of the Vela 2.0 family: a pinned vela2-decoder executor (hybrid.py) with a 10-file gate (shard index and the side-loaded broad head hashed; the 2.02 GB shard by exact size), member-aware response finishing (span_heads with spans, keyed by span question ids), worker member detection from config.json model_type, and the per-member /v1/models entry. compare_reference.py serves either member; the hybrid oracle fixtures export under oracle-hybrid-. Evidence (author-run, x86 WSL2, CPU FP32, no fla — the runtime's CPU path never dispatches to it): nine-case parity vs the pinned upstream server exact-match (four 200s incl. key order, five agreed rejections incl. the 16,384-window 413); 0.3B nine-case regression; live worker smoke with span_heads and per-member thresholds; byte-identical Rust frontend forwarding; 133 CPU contract tests.
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
…reference-serving # Conflicts: # README.md # docs/supported-models.md # recipe/README.md
Contributor
|
fix conflicts |
…ember docs Keep the Clef and pplx-decider rows from main alongside the Vela 0.8B hybrid additions; the package-gate descriptions now carry the directory-symlink refusal from main.
Collaborator
Author
|
@hsliuustc0106 Conflict fixed in |
4 tasks done
This branch has not been deployed
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.
Stacked on #151 — review the last commit only (the diff shows #151's changes too; this PR adds only the 0.8B hybrid member).
Required human contributor answers
What outcome are you trying to achieve?
Serve the second Vela 2.0 member (the 0.8B hybrid,
vllm-sr/Vela-2.0-0.8B@326b01d8) through the Python reference worker, sharing the wire layer the 0.3B established in Vela: serve Vela 2.0 0.3B through a Python reference worker #151. (Source: issue [New Model]: Support the Vela 2.0 series #137.)What is the highest-leverage variable for that outcome?
A second pinned member on the same wire layer validates that the Vela: serve Vela 2.0 0.3B through a Python reference worker #151 contract generalizes across the family (different heads, different input window, extra
span_headsfield) without a per-model fork of the transport code.How will this work move that variable rather than create busywork?
One new file (
hybrid.py) plus member-awareness inpostprocess.pyand the worker; the wire layer, transport discipline, compare tool, and recipe are reused unchanged from Vela: serve Vela 2.0 0.3B through a Python reference worker #151.Who is the customer or user, and what do they actually need?
System1-Omni users routing safety/PII/hallucination decisions through
/v1/systemonewho need the 0.8B multilingual span/set capabilities at the 16,384-token window; the community contributors tracking model support in [Tracking]: Model support status #83.Is a PR needed at all? Yes. (Source: issue #137; user-approved.)
For a bug fix: N/A.
Purpose
Add the 0.8B hybrid member to the Vela 2.0 reference worker family.
hybrid.pypins the vela2-decoder checkpoint with a 10-file SHA-256 gate. Response finishing becomes member-aware (span_headsemitted withspans). The worker detects the member fromconfig.jsonmodel_type.compare_reference.pyserves either member.Test Plan
System1-Omni Version / Commit:
model/vela2-hybrid-reference-serving@78fe143pytest tests/vela -qcompare_reference.py --model vela-2.0-0.8bcompare_reference.py --model vela-2.0-0.3bmkdocs build --strictTest Result
All passed. GPU bf16 serving unverified.
Demo / evidence
Author-run, x86 WSL2, CPU FP32, no fla:
span_heads: {pii: router, halu: router}, per-member thresholds{pii: 0.2914, halu: 0.35},usage: {input_tokens: 1041}.cmpbyte-identical.Not measured: latency, throughput, GPU parity.
Self-review