R1 1.5B: Separate Evaluation and Historical Inputs

For current setup instructions, use the five-model RR / KV-Exact manual. It covers vLLM 0.31.0, SGLang 0.5.19 and TRT rc29/PyTorch. This page retains separate R1 1.5B evidence; its results do not apply to the manual's R1 7B. The older SGLang/TRT launch procedures have been replaced by the current manual. Their archived inputs remain available for evidence comparison.

Before you start

Use this historical routing evaluation only with its own source, image and model identities.

Focus of this page: Compare its original routing arms without substituting newer launch settings.

Topology and execution locations

flowchart TD
    C["Client workload"] --> G["Selected Gateway / routing mode"]
    G --> W["Selected engine worker pool<br/>count and roles from its runbook"]
    O["Operator"] -.->|"configuration and readback"| G
    O -.->|"direct engine / receipt evidence"| W
    style G fill:#e1f5fe,stroke:#0288d1,color:#1f2937
    style W fill:#e8f5e9,stroke:#43a047,color:#1f2937

This reference diagram locates the client, Gateway and engine evidence. It does not prescribe a replica count or promote historical results to a new topology. Use the exact worker count, roles, ports and artifact identities from the selected runbook or retained result; do not mix that tuple with another guide.

Where	What happens here
Operator terminal	Select the version tuple, private run directory, model snapshot and profiles.
Gateway host / namespace	Start the dedicated Gateway; apply and read back only the selected service.
Each worker host	Launch the chosen engine and prove direct inference before Gateway traffic.
Client	Send the same request through the Gateway and retain its HTTP/SSE observations.
Independent evidence	Correlate worker requests/receipts with the selected rule and request identifiers.

Reading path and reference

Identity and scope

Input	Separate evaluation identity
Model	deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
Revision	ad9f0ae0864d7fbcd1cd905e3c6c5b069cc8b562
Profile	r1-distill-qwen-15b-v1 (15b means 1.5B here)
Gateway	Source 1d51393773e4974b5e356377496c9d677dba9f9c, isolated proxy-only evaluation
Evaluated engine	Stock vLLM 0.31.0, three L4 24 GB workers
Archived inputs	SGLang 0.5.18 and TRT 1.3.0rc24; no new runtime PASS inferred

What was verified for vLLM 0.31

On 2026-10-20 a source-built Gateway at the commit above, in proxy-only mode on an isolated container network, was evaluated with three L4 24 GB workers and this stock engine/model revision. No product or engine code was changed. The sanitized QA summary records the bounded results; private host addresses and raw receipts are excluded.

Check	Observed result
Engine identity and direct completion	All three report 0.31.0; six chat/SSE requests completed; all five pinned weight files verified per worker
RR	12/12 completed; four requests per worker, request/response hashes matched backend receipts
Exact evaluation registry	READY, empty reasons, unfenced; unchanged public profile/tokenizer/template/fixtures passed the chat-surface attestation
Controlled shared-prefix routing	12/12 completed and received by the prewarmed worker; cache-hit counter delta 12
Worker loss and recovery	Six fresh requests completed on survivors; restarted worker became healthy and Exact returned to READY
Invalid mutation and mode lifecycle	Invalid mode returned 400 with unchanged rule; RR→Exact→RR completed six requests; scoped deletion proved absence and Gateway readiness

These are instrumented functional observations: a QA receipt proxy forwarded backend requests and responses byte-for-byte. They do not establish latency, throughput, in-flight SSE continuation, lossless fail-over, P/D, host eBPF/HA, or a release-image acceptance. Uncached-prefix controls, post-recovery routing redistribution and cancellation/resource return remain open gates. The original five cards and SGLang/TRT are not qualified by this R1 1.5B result.

The retained old-manifest negative test is significant: declaring 0.28 while running 0.31 produced PROFILE_VALIDATED with identity_mismatch, but goFenced:false and a chat HTTP 200. That response was ordinary/fallback serving, not Exact proof. Do not assume every non-READY condition is globally fail-closed; inspect effective status, reasons and enforcement and require the selected mode's traffic evidence.

Retained files and current setup

The configuration ZIP retains the routing-choice/ R1 1.5B bodies, its vLLM 0.31 evaluation manifest and bounded QA summary. The SGLang/TRT files in that directory are archived inputs with their own unclosed qualification tasks. They are not the official five-model engine launchers and must not be relabeled 0.5.19/rc29 without fresh validation.

Use matrix-configurations/ and evaluation-profiles/ with the current manual for the five standard models. The profile catalog and catalog/installer alignment explain source membership, candidate opt-in and pending integration. See Verification To-do for remaining release, capacity, failure/recovery and archived-example gates.

Check your result

Keep the historical workload and evidence boundaries; it does not qualify a newer engine or Gateway revision.

Select a supported model/engine tuple before interpreting results. Keep direct-worker success separate from Gateway delivery, routing evidence, engine cache reuse, performance and release qualification. Historical PASS, FAIL and BLOCKED results retain their original artifact identities.

Next steps and cleanup

Use the canonical deployment sequence rather than combining fragments from different engine/version pages. Drain before switching or deleting owned resources; retained result pages are evidence, not another deployment recipe. Read the lab starting point if required values are missing. Use cleanup and recovery when ending the run.