Commit Graph
21 Commits
Author SHA1 Message Date
tzhouam bb8735effc Add native Windows CPU Qwen TTS Omni stage and profile evidence 2026-09-23 21:18:21 +08:00
tzhouam 8569995591 Add bounded Omni Qwen3-TTS CPU Radeon whole-session stage 2026-09-23 20:52:43 +08:00
tzhouam 34beb7f01e Add bounded Omni llama.cpp Spark text stage for Windows CPU and Radeon 2026-09-23 20:30:28 +08:00
tzhouam 4fee1dbe39 Profile Qwen3-TTS hybrid Radeon and CPU route 2026-09-23 19:59:04 +08:00
tzhouam 54c086acae Record native Windows CPU Qwen3-TTS scoped E2E evidence 2026-09-23 19:19:13 +08:00
tzhouam 03d6cb9a93 Qualify standalone Spark text on Windows CPU and Radeon 890M 2026-09-23 18:57:41 +08:00
tzhouam 34295f3eb3 Qualify Spark output-head component on AMD NPU and update E2E matrix 2026-09-23 18:32:13 +08:00
tzhouam 10221f0665 Qualify scoped MiniCPM-o image and audio paths on WSL CPU 2026-09-23 17:44:02 +08:00
tzhouam b29a6ca99f Document Windows MiniCPM-o image and audio end-to-end runs 2026-09-23 17:29:13 +08:00
tzhouam a8fdb1babf Validate MiniCPM-o audio input to speech on WSL GPU 2026-09-23 17:09:37 +08:00
tzhouam 0d00ba4701 Validate MiniCPM-o image to speech on WSL GPU 2026-09-23 17:03:31 +08:00
tzhouam 8a4559c46c Validate MiniCPM-o speech pipeline on native Windows GPU 2026-09-23 16:54:17 +08:00
tzhouam 2f5d0edcf7 Qualify scoped MiniCPM-o CPU speech and Windows InternVLA policy paths 2026-09-23 16:21:28 +08:00
tzhouam 3f2cd52939 Update E2E matrix with scoped recovery evidence 2026-09-23 15:24:43 +08:00
tzhouam c0deaa76ba Repair MiniCPM-o and InternVLA scoped E2E paths 2026-09-23 15:15:07 +08:00
tzhouam 2d8b1d4583 Resume and hash-verify pending E2E checkpoints 2026-09-23 10:34:41 +08:00
tzhouam e0142926b6 Document MiniCPM-o checkpoint preflight and current deploy layout 2026-09-23 10:19:10 +08:00
tzhouam b039b6a5bd Enable scoped Qwen3.8 Omni inference and CPU TTS compatibility 2026-09-23 10:01:21 +08:00
tzhouam 75b65f859b Add local E2E profiling evidence and support roadmap 2026-09-22 19:25:02 +08:00
tzhouam ab2f2a3d65 Audit all mobile and PC model pairings with archived evidence 2026-09-22 13:54:34 +08:00
tzhouamandClaude Opus 5 9264ff81ab Include the baselines, comparisons and study documents, not just the fork
The packaging commit shipped the code and a summary of the numbers. It left out
everything the numbers are measured *against*, which is most of what makes them
checkable.

Added:

  baselines/                 llama.cpp runs (thread sweeps, NUMA sweeps, the
                             interleaved A/B passes behind the headline decode
                             figure) and the quality references: each build
                             against its own bf16 over 12 greedy prompts.
  benchmarks/.../results/    102 artifacts backing the published figures, each
                             with a provenance block naming the model,
                             quantization, thread set and build SHAs.
  docs/engines/              the engine analyses -- llama.cpp, ExecuTorch, MNN,
                             ncnn, vLLM, vLLM-Omni -- and the cross-engine and
                             device-simulation comparisons.
  docs/analysis/             the full study documents the summaries distil.
  docs/edge/comparisons.md   where this fork sits against llama.cpp and why:
                             decode is llama.cpp's, prefill is ours by ~2x, and
                             since neither engine exceeds 39% of the measured
                             222 GB/s ceiling the decode gap is the execution
                             model rather than the kernels.
  docs/edge/environment.md   pins and machine, and the correction below.

On "the modified vllm": there isn't one, and the doc says so plainly because it
is the obvious assumption. The vLLM checkout is clean upstream (0 commits
ahead, 0 uncommitted) and the installed wheel is stock 0.28.0+cpu. This fork
changes vLLM's CPU *behaviour* through vLLM's own extension points at import
time -- three CPU patches in vllm_omni/patch.py, the quantization-config
registry, the IR op registry, platform hooks -- specifically so the changes
survive a wheel upgrade rather than being lost on it. One of those patches is
why "KV sizing is not a memory lever" looked true for months: CpuPlatform
overwrote kv_cache_memory_bytes unconditionally, so the setting had no effect.
There is no patch series to apply; pin the version and let patch.py do the rest.

.gitignore gains explicit exceptions: upstream's blanket `*.json` and `results/`
rules were silently eating the measurement artifacts, which in this fork are the
substance rather than build output. Binary captures stay ignored.

No vLLM or llama.cpp source is redistributed -- only output produced with them.
NOTICE records both.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L1ax5wYLA6XbqEL1i8L9gZ
2026-09-14 14:01:12 +08:00