Commit History

Delete old fp16mixed names + retired medium dit_bf16.trt (fp16 rename shipped)
48e34e2
verified

cortexelus commited on

Rename fp16mixed->fp16 DiT/T5Gemma engines+onnx (copies; deletes follow after code merge)
3f29675
verified

cortexelus commited on

Add quantized encoder+decoder tier ONNX (fp8 / w8_bf16, padded encoders)
4168236
verified

cortexelus commited on

Add fp8 tier for sm-music / sm-sfx DiTs (clean fp8-linear graft on fp16mixed)
308abaf
verified

cortexelus commited on

Add fp8 tier for sm-music / sm-sfx DiTs (clean fp8-linear graft on fp16mixed)
ac3a57e
verified

cortexelus commited on

Add fp8 tier for sm-music / sm-sfx DiTs (clean fp8-linear graft on fp16mixed)
5bcf0c7
verified

cortexelus commited on

Add fp8 tier for sm-music / sm-sfx DiTs (clean fp8-linear graft on fp16mixed)
6a96f1b
verified

cortexelus commited on

sa3-m fp8: calibrated dit_fp8.onnx (real-conditioning activation + per-channel weight scales, @ryanontheinside #47; refs unchanged dit_fp8lin.onnx.data)
8383b7d
verified

cortexelus commited on

Upload onnx/sa3-m/dit_fp8lin.onnx.data with huggingface_hub
9efa545
verified

cortexelus commited on

Upload onnx/sa3-m/dit_fp8.onnx with huggingface_hub
21b6506
verified

cortexelus commited on

Fix the bf16 medium DiT: bake RoPE's tables into the graph (clipping 3.112% -> 0.014%) (#5)
6152b1d

cortexelus commited on

Fix medium DiT FP16-mixed ONNX: bound the RoPE island so attention fuses (4.3x faster, cos 1.0000) (#3)
fb17708

cortexelus commited on

Fix missing 2π factor in seconds_freqs across all dit.onnx (FP32 source ONNX)
ecc70ac

Cortexelus Claude Opus 4.7 (1M context) commited on

Fix sa3-m DiT ONNX external_data ref (was pointing at wrong filename)
3dbdcfa

Cortexelus Claude Opus 4.7 (1M context) commited on

Decoder ONNX: clip+scale in FP32 before int32 cast
a429ffa

Cortexelus Claude Opus 4.7 (1M context) commited on

Add pre-processed FP16-mixed ONNX for SA3 DiTs
e11ce65

Cortexelus Claude Opus 4.7 (1M context) commited on

T5Gemma: FP16-mixed (FP32 attention island) — fixes BF16 numerical bug
23f5c64

Cortexelus commited on

Decoder ONNX: bake PCM postprocessing into the graph
459fd8d

Cortexelus commited on

Add ONNX files for per-arch TRT compilation
62164ab

Cortexelus commited on