extract_7z_flat replaces extract_zip_flat; NPU + DirectML now default to
git.dou.bet raw/branch/main/runtime/{openvino,directml}.7z with pinned SHA-256s
(DirectML no longer requires WA_DIRECTML_RUNTIME_URL). Adds an extraction test
against the real bundle (proves LZMA2+BCJ decode + flatten).
True only when staging DirectML would light up a present GPU this build can't
otherwise accelerate: npu compiled, no Vulkan baked in, and an AMD/Intel GPU
present (or NVIDIA without CUDA).
load_transcriber dispatches OnnxOpenVino/OnnxDirectML through one OnnxTranscriber
block. download_and_extract_runtime is parameterized (sha + readiness) so the new
stage_directml_runtime reuses it; hardware_status reports a directml package field.
load() now selects OpenVINO (NPU) or DirectML (AMD/Intel GPU) by BackendId over
the same ONNX artifacts; ensure_runtime_env takes the target runtime DLL. Adds
an opt-in directml_transcribes_speech spike for on-GPU validation.
load_transcriber now resolves the accel path instead of assuming a GPU
BackendId has a working offload. Only requests whisper GPU offload for
WhisperCuda/WhisperVulkan; NPU goes to the ONNX engine; everything else decodes
on CPU and reports the backend actually used.
whisper.cpp bakes in one GPU backend at compile time, so a BackendId alone
can't say how a vendor is served. Add AccelPath + resolve_accel(_with) as the
single source of truth over (compiled features x runtime readiness), and derive
BackendInfo.available from it in the DXGI + NPU paths. Closes the gap where a
present-but-unsupported GPU was marked available and best() routed to a no-op
GPU. Pure resolve_accel_with is unit-tested across all vendor/feature combos.
whisper.cpp's encoder always runs over a full, padded 30s mel window
(1500 encoder positions) unless audio_ctx is explicitly reduced. The
~4s live-transcription streaming windows were never setting it, so
every window paid the full 30s-equivalent encode cost on one serial
worker thread -- the actual cause of the reported 30-65s lag between
speaking and a segment appearing, not "CPU is just slow".
Scales audio_ctx proportionally to the window's real sample count
instead (audio_ctx_for_window), floored and capped to whisper.cpp's
own valid range. Verified with a controlled A/B against the real
base.en-q5_1 model: ~30% faster even under heavy concurrent CPU
contention on this machine, which is masking most of the improvement
--the fix targets the encoder's O(n^2)-ish attention cost specifically
(1500 -> ~201 positions for a 4s window), so the gain on an otherwise-
idle machine should be substantially larger.