Status: ReleasedData ownership: Institute-ownedLicence: CC BY 4.0
Cross-chip byte-identical decode traces
Per-token traces quantifying how far greedy decode stays byte-identical across two different chips — the basis of mid-stream failover.
Purpose
Establish how reliably decode can hand off between chips without diverging.
Source
Captured live during a 2-chip run of the same quantized model under greedy decode.
Composition
Downloadable CSV — measured cross-backend consistency (backend, comparison, cosine, argmax) plus the CUDA↔ROCm head-to-head (Qwen2.5-0.5B, 31/32 token-identical). The full per-token trace set is larger and available on request.
Known biases
- Greedy decode only (not sampled).
Limitations
- A single 2-chip session.
- The disagreement residual concentrates at sub-f16-ULP ties and is irreducible at f16.
Access conditions
The measured cross-backend consistency tables are available for download below, under CC-BY-4.0. (The full per-token trace set is larger and available on request.)
| backend | comparison | cosine | argmax_match | argmax_total | sovereignty |
|---|---|---|---|---|---|
| CUDA (AWS L4) MoE engine | fp32 vs int8-KV teacher-forced | 0.999761 | 16 | 16 | ldd = libcudart only (no cuBLAS/cuDNN/cuTLASS) |
| Metal (Mac M-series) q4 resident | prod fused-MoE vs CPU-glue ref | 0.99999999999944 | 16 | 16 | runtime cBLAS count = 0 |
| Metal single-encoder decode (short ctx) | short 72-tok ctx | 0.99999999999989 | byte-identical | greedy | — |
| Metal single-encoder decode (long ctx) | long 72-tok ctx | 0.99999999999992 | byte-identical | greedy | — |
| CUDA-Silicon direct | chip-to-chip 30B q4 | 0.99999999 | — | — | both sovereign |