Repository navigation
effective-config reports Qwen3.5 CUDA residency as 'cpu residency' and drops the F2 parity verdict (v0.69.3) #4374
Description
GET /v1/effective-config misreports a CUDA-resident Qwen3.5 model as CPU-resident (v0.69.3, gx10).
Observed (gx10-a5b5, apr-review-serve from infra e872544c, apr v0.69.3 binary sha 2f274f47…, Qwen3.5-4B-Q4_K_M sha 00fe7986…, 2026-09-25 ~10:14Z):
- The journal says
Backend: GPU (CUDA, NVIDIA GB10 …) [qwen35 hybrid forward, #3090],32 layers resident on the GPU,gpu-layers … (backend=cuda), and an F2 guard receipt match. /healthsayscompute_mode: "gpu"./v1/effective-configsaysparity: {status: "not-run", cosine: null, basis: "no GPU model loaded (cpu residency)"},backend_loaded: [],cuda: null,compute_class: "unknown".
Why it matters: effective-config is the provenance envelope (REG-15 / #2971 wire shape). A reader who trusts it concludes the model runs on CPU. It also discards the F2 verdict, so no endpoint exposes the parity result of the qwen35 path. PROMETHEUS (#4354) REX-04 cannot admit cell C4 from the serve's own envelope.
Expected: for the qwen35 hybrid CUDA path, backend_loaded and cuda are populated. parity carries the F2 outcome (source fresh/receipt, positions_judged, and the measured cosine/argmax agreement if one exists).
Activity
noahgift commentedon Sep 25, 2026
Fix on branch
fix/4374-effective-config-qwen35@ 4ec849e (no PR: 10-PR cap, to be folded into B3). A Qwen3.5 serve now reportsbackend_loaded/compute_classfrom the session (cuda|cpu), andparitycarries the F2 guard's record: status PASS/FAIL/not-run, min real-position cosine, positions, source=fresh|receipt. A receipt hit reports PASS with cosine null, so C4's cosine receipt needs a serve started withAPR_F2_REVALIDATE=1. Tests: 6/6 pass (5 pure, plus a CPU test backed by the 0.8B GGUF), and CUDA lib clippy and the test-profile check are clean. The GPU falsifier is compiled but NOT run (this worker does no lambda 4090 work). It needs a gx10 run through the declared executor (infra#1088). Unrelated breakage on main:tests/driver_cuda_gguf.rsdoes not compile with --features cuda (missingquery_pre_attn_scalar/post_*_norm_weight).noahgift commentedon Oct 14, 2026
P0 to P2 (2026-10-14 P0 triage): P0 now means it blocks the next release or carries the operator's own words on it; this issue does neither, and its milestone is backlog.
noahgift commentedon Oct 14, 2026
merged into #5312 (master ticket for epic #3994; triage rule 7b, 2026-10-14). It is a checklist row there; reopen this issue when work on it starts or its PR opens.
noahgift commentedon Oct 14, 2026
Moved: this sub-ticket is now a row in #5325 (apr on GPU: refuse by name before OOM, never fall back to CPU silently), split out of #5312 by theme.