Repository navigation
0.70.2 P0: Qwen3.5-35B-A3B-UD-IQ4_XS.gguf capability_match: no CUDA forward for qwen35moe (known failure in 0.70.1) #4665
Description
Known failure shipped in 0.70.1 (aprender 0.70.1, candidate cc4463f) under C272 5b; P0 for 0.70.2. Not a regression: it fails the same way on v0.69.5-rc.1 (5660b08).
- Model: Qwen3.5-35B-A3B-UD-IQ4_XS.gguf
- Host(s): lambda
- Check: capability_match
- 0.70.1 result: FAIL: this build has no CUDA forward for architecture 'qwen35moe' (named refusal)
- v0.69.5-rc.1 result: FAIL: identical refusal
- Logs: 0.70.1: /mnt/nvme-raid0/tmp/evid-0701b/lambda-ladder.log; rc.1: /mnt/nvme-raid0/tmp/screen-0701/logs/inv_Qwen3.5-35B-A3B-UD-IQ4_XS.gguf.log
- Reproduce:
scripts/model_ladder.sh --host <lambda|gx10> --only inv:Qwen3.5-35B-A3B-UD-IQ4_XS.gguf
Done when: the ladder row is OK on lambda and gx10.
Activity
feat(cuda): qwen35moe CUDA forward — routed experts + gated shared ex…
contract(ladder): drop the qwen35moe red_unsupported key — the CUDA f…
docs(roadmap): GH-4665 acceptance criteria + measured ladder legs
evidence(GH-4665): commit both ladder receipts; the sign-off holds th…
audit(GH-4665): quorum round 2 AGREED — sonnet-5, gemini-3.1-pro-high…
noahgift commentedon Oct 5, 2026
Re-measure on current main (aprender-w4578): this is a named refusal for a missing architecture, not a GPU fallback. Proposed re-rule: refused-by-name, out of 0.70.2.
apr 0.70.1 (6192aa9b88), which ismainat 6192aa9. Release build with--features cuda, sha256 prefixd8208cdfc6d3686a, run under the GPU lock.apr run Qwen3.5-35B-A3B-UD-IQ4_XS.gguf --prompt "What is 2+2? Answer with one number." --max-tokens 16 --tracegave rc=8:--no-gpu. The load is refused before any backend is chosen, so there is no CUDA path that falls back here.Proposal: re-rule.
The refusal names the file's architecture, gives a clear error, and is deterministic. That is the "refuse the file by name" outcome #4664's ruling allows. Making the row OK needs a new qwen35moe forward, meaning MoE routing plus the Gated Delta Net layers on both CPU and GPU. That is an architecture feature, not a 0.70.2 P0 bug fix.
The proposed re-rule:
red_model) rather than a P0 failure.Agent: aprender-w4578
noahgift commentedon Oct 5, 2026
Re-ruled (release cop, 2026-10-05): refused-by-name for 0.70.2.
On main,
apr run --gpuon this file exits 12 with the by-name refusal: "this build has no CUDA forward for architecture 'qwen35moe' ... This is a refusal, not a fallback". That is exactly what the existingred_unsupported["Qwen3.5-35B-A3B*.gguf"]key requires, so the ladder row reads RED-UNSUPPORTED rather than FAIL.red_unsupportedcitation moves from 0.70.0: qwen35moe (Qwen3.5-35B-A3B hybrid SSM MoE) CUDA forward #3977 to qwen35moe (Qwen3.5-35B-A3B, hybrid Gated DeltaNet + MoE): CPU and CUDA forward #4801 in the next B6 fold.P0is removed. This issue stays open.Agent: aprender-w4578
noahgift commentedon Oct 7, 2026
0.70.2 ships on a CRUX-smoke emergency scope (operator ruling C322, A1 and A2). This issue is refused for 0.70.2 and moves to milestone 0.71.0, with no per-issue contract edit. The 0.70.2 release notes name it as a known failure.
noahgift commentedon Oct 14, 2026
Folded into #5311 as a checklist row (TRIAGE-100, 2026-10-14): the open-ticket cap is 100, so this ticket is now tracked on that master and closed as not planned. Nothing is lost: reopen it when its work resumes, and an open PR for it reopens it. Pushed branch:
fix/4665-qwen35moe-forward.Agent: aprender-clerk
noahgift commentedon Oct 14, 2026
Row moved: #5311 was split into named masters (TRIAGE-100 master naming), so this ticket's checklist row is now on #5327.
Agent: aprender-clerk