 xiaolibai-sys
|
164896ed78
|
bump v1.1.4: nvfp4 support, GPU LoRA fold, loader I/O pipeline, async D2H, unified logging
- nvfp4: _quantization_metadata -> comfy_quant synthesis, SlotEntry extra params (block_scale), dual-scale LoRA fold, orig_dtype fixup
- LoRA: plain-mode block LoRAs no longer silently dropped; folded on GPU after weights land
- loader: parallel positioned reads, pipelined group load, no per-tensor clone, header-only dtype probes
- block swap: read_block intra-block parallel + disk_workers node option; async D2H write-back on transfer stream; lazy evict_all skips wasted write-back
- arch: split models/loader.py out of wan_model.py; single utils/keys.py; LoRA folding consolidated in utils/lora.py; global utils/log.py ([BerniniR][Tag] format, BERNINI_LOG_LEVEL)
|
2026-07-21 14:07:05 +08:00 |
|
|
xiaolibai-sys
|
ad19628149
|
v1.1.1
Unify block-swap loading paths: SlotEntry + ring buffer + PinStage
- SlotEntry unifies regular Tensor and QuantizedTensor (fp8) handling,
replacing p.data = d (a no-op on __torch_wrapper_subclass__) with
module._parameters[key] = nn.Parameter(...)
- GPU ring buffer (cursor-based) replaces free-list — no explicit release
- PinStage @dataclass: pinned CPU staging ring, size = prefetch_count
- _detect_model_config merges duplicated config detection
- _block_swap config flag removed — model reads from self._block_swap_config
- StreamingBlockPipeline removed (dead code)
- BlockSwapManager always used when block_to_swap > 0; block_to_swap = 0
falls back to standard GPU-only loading (1x memory, zero overhead)
- Fix: TeaCache pre-load bypassed block_swap_args (sampler.py:603)
- Fix: cache/guard paths now restore _block_swap_config if missing
- Fix: CUDA OOM with pin_memory on Windows — pin pool = prefetch_count
- Fix: black/noise output with pin path — sync transfer stream before assign_to
- Fix: shutdown size mismatch — offload_block copies slot-to-slot, not via freed params
- Remove: pinned_pool.py, block_ortho.py (dead code)
- Docs: README.md, README_DEVELOPER.md updated for v1.1.x architecture
|
2026-07-17 13:23:32 +08:00 |
|
|
Wen Li
|
91ce7daa7b
|
Initial commit: ComfyUI-BerniniR Wrapper v1.0.0
Seven-mode guidance (CFG/APG/RAAG/S2/Z2/STG_A/STG_R), context-window sampling, segment prompt-travel, lazy streaming load, layered VRAM management.
|
2026-07-15 00:54:57 +08:00 |
|