Record Qwen3-Omni two-GPU campaign and source-grounded tuning knowledge by amy-why-3459 · Pull Request #113 · InferMatrixLab/InferRSI

amy-why-3459 · 2026-10-23T04:28:42Z

Records the two-GPU Qwen3-Omni streaming-speech campaign for #112 and source-pinned MRv2 tuning knowledge. It fixes text readback semantics, freezes raw-audio and paired-WER gates, and retains the baseline matrix, separate quality outputs, screening attempts, runtime provenance and per-optimization sources.

The full Thinker → Talker → Code2Wav campaign remains INCONCLUSIVE. The complete CUDA bundle screen reports C1/C64 RTF 0.1050/0.4840 and mean WER 0.07514 versus historical baseline 0.03833: neither RTF target is reached, and the historical-baseline quality screen fails. This bundle is not recommended. Screening warmup/hardware histories differ; no comparative speedup or universal optimum is claimed. Original private model/audio payloads retain integrity hashes. Nested metadata byte checks do not establish payload availability or campaign acceptance.

The eager-MTP adapter screen reports C1/C64 RTF 0.1097/0.6117 and WER 0.04209, also above the historical quality reference. Its eight actual CUDA kernel-parity tests pass; this does not establish full-model quality. Placement and prefix/cuDNN experiments are separately registered, and retained evidence distinguishes screening from paired qualification.

The governing formal protocol is v5, including complete conditioning, three alternating pairs, observed-JIT exclusion and deterministic paired retries. Older protocols/results remain retained. Audio-input quality requires separate qualification. Each existing optimization is credited in evidence/qwen3-omni-2gpu-20261023/plan/source-attribution.md.

Validation: raw-audio checks pass on complete retained screens; the latest focused inventory suite passes. The prior isolated full suite reported 592 passed / 8 skipped plus one subsequently corrected inventory-count assertion. Earlier GitHub full-check and installed-package jobs passed on Python3.10/3.11; subsequent-head CI is reported by checks, not assumed here. Engine fixes are submitted separately as vllm-project/vllm-omni#9278 (audio timing/evidence), #9279 (peak concurrency), #9281 (JIT monitor forwarding), and #9283 (Qwen eager-MTP capability adapter).

The report-only knowledge-model-policy status previously failed separately from passing tests. Keep this PR draft while candidate validation and campaign review remain incomplete.

2026-10-24 iteration update

Protocol-v5 has ended INCONCLUSIVE under its unchanged retry rule. Pair 1 was execution-eligible but failed the quality and endpoint RTF gates. Both baseline attempts in pair 2 observed two measured-interval JIT events; pair 3 was not attempted. The complete retained evidence includes all attempted and unattempted slots, raw request rows and losslessly compressed logs. No three-pair aggregate or qualified service gain exists.

The user requested continued cross-repository search and conditional performance-ceiling analysis. New screening is separate from v5: first compare the same MRv2 deployment with the eager-MTP adapter disabled/enabled, using all 256 quality rows before any performance promotion. Native MoE tuning and predictor work-reduction diagnostics are preliminary component experiments only. No new optimization PR will be presented as validated without isolated quality and performance evidence. Hardware analytical bounds, empirical operation envelopes and service outcomes will be reported separately; no global-optimum claim is justified.

Agent-assisted-by: Codex (implementation) Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com>

Preserve upstream optimization provenance and distinguish source guidance from two-GPU qualification. Document native audio timing and throughput units without claiming measured gains. Agent-assisted-by: Codex / unknown [effort=unknown] (knowledge author) Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com>

Agent-assisted-by: Codex (implementation) Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com>