dkeviv/Qwen3.5-0.8B-Text-ONNX-WebGPU-QDQ-bucket
Document WebGPU QDQ graph repair and attribution
Decompose causal depthwise convolution into standard elementwise WebGPU ops
Preserve Qwen GDN SiLU activation in standard causal conv rewrite
Replace custom CausalConv with standard grouped Conv1d
Use fused causal convolution with valid tokenizer inputs
Standardize causal convolution after tokenizer fix
Keep causal state kernel; fix tokenizer sequence path
Interleave packed q4 nibbles and zero points correctly
Use standard 2D depthwise Conv for causal state
Fix Slice opset attributes
Rewrite causal convolution with standard slices
Rewrite causal convolution with standard Conv and state slice
Use direct dynamic embedding shapes
Use dynamic Flatten for embedding rows
Enable floating Mod for standard WebGPU dequant
Fix scale cast for WebGPU type consistency
Fix dynamic embedding dequantization shapes
Add tokenizer and model metadata
Add tokenizer and model metadata
Add tokenizer and model metadata
Add tokenizer and model metadata
Add WebGPU standard-op embedding rewrite
Add WebGPU standard-op embedding rewrite
initial commit
