WildAi 2aaef8228b feat: native lowvram streaming for oversized models (v2.6.0) ...

- comfy_stream: convert the transformers tree to comfy-cast streaming modules; direct-param containers (Block1D gammas) self-relocate to the compute device on attribute access
- GGUF/ConvRot quant residents stream natively, dtype-preserving
- patcher: drop force-placement workaround, pure core delegation
- vendored generate: derive stage devices from activations, not parameter residency (lies under partial offload)
- generation: place inputs on the runtime compute device
- tests: lowvram protocol, streaming parity, residents, e2e

2026-08-27 17:52:17 +03:00

..

__init__.py

major refactoring

2025-09-10 12:06:26 +03:00

asr_generation.py

feat: unload-on-change eviction + quant-resident runtime for GGUF and quantized safetensors (v2.4.0 -> v2.5.0)

2026-08-26 12:17:24 +03:00

asr_loader.py

perf: meta-init + assign loading, non-destructive offload contract (v2.3.0)

2026-08-19 18:56:35 +03:00

attention_utils.py

refactor init

2026-08-14 20:26:31 +03:00

audio_backend.py

feat: torchaudio-primary audio backend, librosa now optional (v2.1.0)

2026-08-15 21:54:53 +03:00

audio_utils.py

feat: torchaudio-primary audio backend, librosa now optional (v2.1.0)

2026-08-15 21:54:53 +03:00

base_loader.py

fix: cpu-first loading + neg-branch rope fix

2026-08-15 16:46:56 +03:00

comfy_stream.py

feat: native lowvram streaming for oversized models (v2.6.0)

2026-08-27 17:52:17 +03:00

convrot_quant.py

feat: native lowvram streaming for oversized models (v2.6.0)

2026-08-27 17:52:17 +03:00

custom_types.py

feat: external model input via VibeVoiceLoadExternalModel node (v2.2.0)

2026-08-16 23:59:04 +03:00

device_utils.py

refactor init

2026-08-14 20:26:31 +03:00

dtype_utils.py

feat: unload-on-change eviction + quant-resident runtime for GGUF and quantized safetensors (v2.4.0 -> v2.5.0)

2026-08-26 12:17:24 +03:00

external_loader.py

fix: lowvram partial-load completion, VibeVoice-7B rename follow-through, copy-free packaged tokenizer

2026-08-26 13:24:55 +03:00

generation.py

feat: native lowvram streaming for oversized models (v2.6.0)

2026-08-27 17:52:17 +03:00

gguf_quant.py

feat: native lowvram streaming for oversized models (v2.6.0)

2026-08-27 17:52:17 +03:00

loader.py

feat: native lowvram streaming for oversized models (v2.6.0)

2026-08-27 17:52:17 +03:00

model_info.py

fix: lowvram partial-load completion, VibeVoice-7B rename follow-through, copy-free packaged tokenizer

2026-08-26 13:24:55 +03:00

model_registry.py

feat: unload-on-change eviction + quant-resident runtime for GGUF and quantized safetensors (v2.4.0 -> v2.5.0)

2026-08-26 12:17:24 +03:00

patcher.py

feat: native lowvram streaming for oversized models (v2.6.0)

2026-08-27 17:52:17 +03:00

quant_common.py

feat: unload-on-change eviction + quant-resident runtime for GGUF and quantized safetensors (v2.4.0 -> v2.5.0)

2026-08-26 12:17:24 +03:00

utils.py

refactor init

2026-08-14 20:26:31 +03:00

vendor_streaming_types.py

feat: native lowvram streaming for oversized models (v2.6.0)

2026-08-27 17:52:17 +03:00