| # Validation evidence and interpretation |
|
|
| ## Tested hardware and software |
|
|
| | Item | Verified value | |
| |---|---| |
| | CPU/APU | AMD Ryzen AI MAX+ 395 | |
| | GPU | Radeon 8060S, `gfx1151` | |
| | RAM | 63.65 GiB visible | |
| | OS | Windows 11 Pro, build 26200 | |
| | AMD driver | 32.0.31041.1004 | |
| | TheRock | gfx1151-7.13.0 | |
| | HIP | 7.13.99004-3309c6114a | |
| | Compiler | AMD Clang 23.0.0 | |
| | llama.cpp | b10830 / 465e49b + one-file patch | |
|
|
| ## Tested inference configuration |
|
|
| | Setting | Value | |
| |---|---| |
| | Target | Qwen3.8-27B `UD-Q4_K_XL` | |
| | Draft | Qwen3.8-27B DFlash2 `Q8_0` | |
| | Target GPU layers | 99 | |
| | Draft GPU layers | 99 | |
| | Context | 32,768 | |
| | Batch / micro-batch | 2,048 / 512 | |
| | KV cache | q8_0 K and V | |
| | Parallel slots | 1 | |
| | DFlash p_min | 0.90 | |
| | Requested / effective n_max | 8 / 7 | |
| |
| Model files are not redistributed here. Names document the measured configuration, |
| not an endorsement or compatibility promise for similarly named checkpoints. |
| |
| ## Evidence hierarchy |
| |
| The investigation used progressively stronger checks: |
| |
| 1. **PE and dependency check** — distinguishes a real 86.8 MB executable from the |
| earlier 9,728-byte stub. |
| 2. **Version and command-line check** — proves which build and model paths are live. |
| 3. **Health and model endpoints** — proves the server loaded, but not correctness. |
| 4. **Metrics** — proves DFlash actually drafted and the target accepted draft tokens. |
| 5. **Perplexity invariant** — tests whether changing only micro-batch splitting |
| changes model probabilities. |
| 6. **Positional needles** — exercises early, middle, and late retrieval through the |
| OpenAI-compatible user surface. |
| 7. **Full longctx/deep-reasoning battery** — verifies the original failing consumer |
| workload. |
| |
| Only the last three establish the correctness repair. |
| |
| ## Published and retained evidence |
| |
| - `evidence/pass1-old-hip/`: exact first-pass local source shims and cache-init file |
| - `evidence/validation-summary.json`: machine-readable aggregate results for the |
| first pass, control run, and repaired build |
| - `scripts/probe_gfx1151_long_prompt.py`: public synthetic split-prefill probe |
|
|
| Raw private battery JSONL files were retained by the project but are not included. |
| They contain answer text that would contaminate future evaluation if published. |
| The aggregate scores, counts, token budgets, timings, and interpretation needed to |
| audit this release are preserved in the sanitized summary and this documentation. |
|
|
| ## What may be concluded |
|
|
| - The first-pass HIP executable corrupted the tested split-prefill long prompts. |
| - The fixed HIP executable did not reproduce that corruption. |
| - The fixed executable successfully exercised DFlash2 on full GPU offload. |
| - The repaired configuration scored 11/11 on the retained longctx slice. |
|
|
| ## What may not be concluded |
|
|
| - That every `gfx1151` machine, driver, or ROCm build behaves identically. |
| - That DFlash2 inherently improves answer quality. |
| - That the fixed binary is faster than every Vulkan or upstream ROCm build. |
| - That rung 9 is a stable mean rather than a favorable single sample. |
| - That 262K context or parallel production serving is validated. |
|
|
