 WildAi
|
8404100f31
|
perf: stream sharded/quant loads, console progress, dep cleanup
- streaming per-tensor assign (tensor.clone) replaces merged-dict
load_state_dict_sharded — mmap views were keeping all shard files
resident after partial offload (+15GB ghost RSS, Pin-error flood);
private RAM copies pin reliably and release shard mappings early
- console progress bar during TTS/ASR generation (progress_utils)
- GGUF read-only mmap no longer warns (copy before from_numpy)
- demote internal load-narration logs to DEBUG (success-only INFO)
- drop dead deps (einops, tokenizers, s3tokenizer, conformer),
add accelerate to pyproject (4-bit path needs it), requirements
trimmed to non-ComfyUI-bundled packages; parity guard tests
|
2026-09-01 21:39:58 +03:00 |
|