add: DFlash block diffusion speculative decoding - #1211

Merged

ChenhanYu merged 47 commits into

mainfrom

chenhany/dflash-v2

Apr 14, 2026

Commits on Apr 14, 2026

add: DFlash block diffusion speculative decoding

ChenhanYu
and
claude
committed
refactor: simplify DFlash
ChenhanYu
committed
refactor: simplify DFlash implementation and training pipeline

ChenhanYu
and
claude
committed
fix: revert DFlash-specific changes, add sliding window docs

ChenhanYu
and
claude
committed
refactor: consolidate docs, simplify eval, online GT as default

ChenhanYu
and
claude
committed
add: export, PTQ, and vLLM scripts; document open items

ChenhanYu
and
claude
committed
add: vLLM DFlash deployment script + validated 386 tok/s

ChenhanYu
and
claude
committed
fix: vllm_serve.sh - add TP_SIZE, fix prompt parsing, 1024 max tokens

ChenhanYu
and
claude
committed
remove: HTML results page (lives in nmm-sandbox, not Model-Optimizer)

ChenhanYu
and
claude
committed
update: DFlash unit tests for _apply meta fix, sliding window, mask_token_id error

ChenhanYu
and
claude
committed
update: dflash.md with AR evaluation methods and vLLM per-category results

ChenhanYu
and
claude
committed
simplify: remove fixed GT vs online GT discussion from dflash.md

ChenhanYu
and
claude
committed
address PR review: clarify KV injection, anchors, decay, offline, vLLM

ChenhanYu
and
claude
committed
update: categorize open items by implementation status in dflash.md

ChenhanYu
and
claude
committed
fix: address CodeRabbit review comments

ChenhanYu
and
claude
committed
add: illustrated anchor sampling example in dflash.md

ChenhanYu
and
claude
committed
clarify: architecture descriptions per reviewer feedback

ChenhanYu
and
claude
committed
add: token-level KV injection illustration in dflash.md

ChenhanYu
and
claude
committed
add: training vs inference with attention mask visualization

ChenhanYu
and
claude
committed
rename: AR Evaluation → HuggingFace AR Evaluation

ChenhanYu
and
claude
committed
address PR review feedback

ChenhanYu
and
claude
committed
address PR review: chat template, rope, export, vLLM improvements

ChenhanYu
and
claude
committed
address PR review: round 3

ChenhanYu
and
claude
committed
add Qwen3.5-4B chat template, vLLM smoke test, deployment notes

ChenhanYu
and
claude
committed
fix: code quality - inline short function call

ChenhanYu
and
claude
committed
fix: revert local_files_only, lazy rotary init, loss mask simplification

ChenhanYu
and
claude
committed
fix: GPU tests use dflash_mask_token_id (moved from architecture_config)

ChenhanYu
and
claude
committed
fix: support conversations field as fallback for messages

ChenhanYu
and
claude
committed
fix: convert EAGLE3 test data from conversations to messages format

ChenhanYu
and
claude
committed
fix: use static test data instead of downloading daring-anteater

ChenhanYu
and
claude
committed
standardize on OpenAI messages format throughout pipeline

ChenhanYu
and
claude
committed
deprecate conversations field with one-time warning

ChenhanYu
and
claude
committed
remove duplicate chat template from modelopt_recipes

ChenhanYu
and
claude
committed
fix: add conversations alias in test data for backward compatibility

ChenhanYu
and
claude
committed
fix: add missing json import in conftest (needed by tiny_conversations_path)

ChenhanYu
and
claude
committed
fix: rename make_eagle_supervised_data_module in offline test

ChenhanYu
and
claude
committed
fix: pass chat template explicitly in generation tag tests

ChenhanYu
and
claude
committed
fix: match generation tag variants, revert uv.lock

ChenhanYu
and
claude
committed
fix: generation tag verification and test assertions

ChenhanYu
and
claude
committed
fix: revert retries passthrough in core.py (breaks launcher unit test)

ChenhanYu
and
claude
committed
add: synthetic dataset, regression test, Qwen3-0.6B example

ChenhanYu
and
claude
committed
fix: support messages field in calibrate_draft_vocab and compute_hidden_states

ChenhanYu
and
claude
committed
fix: Qwen3.5-4B does not need trust_remote_code

ChenhanYu
and
claude
committed
trigger CI with run-tests label
ChenhanYu
committed
add: gpu_regression test suite with per-push CI

ChenhanYu
and
claude
committed
fix: export review feedback - dtype mismatch, default, validation

ChenhanYu
and
claude
committed
retrigger CI
ChenhanYu
committed