| # Block Diffusion Base Code | |
| This is a small, shared implementation surface for comparing the reference | |
| methods without copying every paper's code structure. | |
| Current scope: | |
| - Block-level masked diffusion decoding. | |
| - Confidence-based remasking. | |
| - Multi-block scheduling hooks. | |
| - DMax/TAD-style trajectory distillation hooks. | |
| - Fast-dLLM/DFlash/PRESTO-style cache and draft/verify hooks. | |
| - Common metrics: NFE, latency, tokens per forward, exact match. | |
| Run a local smoke test: | |
| ```bash | |
| python3 -m basecode.blockdiff.cli --method confidence --steps 6 --block-size 8 | |
| python3 -m basecode.blockdiff.cli --method multiblock --steps 6 --block-size 8 --num-blocks 4 | |
| python3 -m basecode.blockdiff.cli --method speculative --steps 6 --block-size 8 | |
| ``` | |
| The toy adapter is deterministic and exists only to validate algorithmic | |
| plumbing. Real reproduction runs should add an adapter under | |
| `basecode/blockdiff/adapters/` that wraps each paper's HF/model code. | |