add: DFlash block diffusion speculative decoding#1128

Open

ChenhanYu wants to merge 40 commits intomainfrom

chenhany/dflash

Commits on Apr 1, 2026

add: DFlash block diffusion speculative decoding
ChenhanYu
committed
fix: rewrite DFlash to match SpecForge reference
ChenhanYu
committed
fix: correct mask_token_id and base model forward dispatch
ChenhanYu
committed
add: auto-detect mask_token_id for DFlash across model families
ChenhanYu
committed
fix: prevent DDP deadlock during AR validation
ChenhanYu
committed
fix: avoid DynamicModule dispatch loop in forward/training paths
ChenhanYu
committed
fix: revert training/eval to super().forward() matching EAGLE pattern
ChenhanYu
committed
fix: DDP deadlock when no valid loss positions on a rank
ChenhanYu
committed
add: logit distillation option for DFlash training
ChenhanYu
committed
fix: print training accuracy to console at each log step
ChenhanYu
committed
fix: use response-only loss mask for DFlash training
ChenhanYu
committed
fix: apply assistant_masks to labels in LanguageDataCollator
ChenhanYu
committed
fix: robust response-only loss mask via regex assistant span detection
ChenhanYu
committed
docs: add DFlash section to speculative decoding README

ChenhanYu
and
claude
committed
fix: resolve DFlash components from base model architecture

ChenhanYu
and
claude
committed
fix: enable response-only loss mask for DFlash training

ChenhanYu
and
claude
committed
add: DFlash launcher example for Qwen3-8B

ChenhanYu
and
claude
committed
fix: inline values in DFlash launcher YAML for --yaml compatibility

ChenhanYu
and
claude
committed
add: unit tests for DFlash speculative decoding

ChenhanYu
and
claude
committed
fix: add docstrings to DFlash classes for coverage check

ChenhanYu
and
claude
committed
add: AR validation step to DFlash launcher pipeline

ChenhanYu
and
claude
committed
fix: split DFlash tests into CPU (unit) and GPU tests

ChenhanYu
and
claude
committed
fix: correct DFlash attention mask test for reverse-causal pattern

ChenhanYu
and
claude
committed
fix: remove __init__.py from GPU test dirs to avoid conftest conflict

ChenhanYu
and
claude
committed
fix: match dtype in DFlash GPU tests to model dtype

ChenhanYu
and
claude
committed
fix: use Optional types for nullable DFlash arguments

ChenhanYu
and
claude
committed
fix: merge AR validation into DFlash training script

ChenhanYu
and
claude
committed

Commits on Apr 2, 2026

Commits on Apr 3, 2026

feat: update DFlash training to match SpecForge latest (post-PR #473)

ChenhanYu
and
claude
committed