DFlash speculative decoding for MiniMax-M2.7 (FSDP2): auto mask-token, FSDP2 resume fixes, per-checkpoint draft export#1621

Merged

yeyu-nvidia merged 32 commits into

mainfrom

yeyu/dflash-auto-mask-token

Jun 15, 2026

Commits on Jun 10, 2026

fix: auto-add mask token for DFlash when tokenizer lacks one

yeyu-nvidia
and
claude
committed
feat: add requeue support to build_slurm_executor

yeyu-nvidia
and
claude
committed
feat: FSDP2 efficient loading + requeue retries fix

yeyu-nvidia
and
claude
committed
fix: restore FSDP2 buffer patch and remove buffer-to-CUDA block for checkpoint resume

yeyu-nvidia
and
claude
committed
fix: use broadcast dtype instead of local param dtype in FSDP2 buffer patch

yeyu-nvidia
and
claude
committed
feat: restore DFlashExportCallback for per-checkpoint draft export

yeyu-nvidia
and
claude
committed
feat: MiniMax-M2.7-DFlash launcher example (online + offline + specdec_bench)

yeyu-nvidia
and
claude
committed
feat: loss_mask + chat-template support in compute_hidden_states_vllm.py

yeyu-nvidia
and
claude
committed
fix: tolerate missing dist metadata in modelopt.__version__

yeyu-nvidia
and
claude
committed
fix: standalone aux-layer resolution in vLLM dump + --block_size in DFLASH bench

yeyu-nvidia
and
claude
committed
feat: DFlash export injects long-context RoPE (target rope_theta + YaRN)

yeyu-nvidia
and
claude
committed
fix: vLLM hidden-state dump accepts 'messages' key (not just 'conversations')

yeyu-nvidia
and
claude
committed
fix: vLLM hidden-state dump stages on /dev/shm with sync lock

yeyu-nvidia
and
claude
committed
review: address CodeRabbit nitpicks + revert modelopt version guard

yeyu-nvidia
and
claude
committed
style: ruff format fsdp2_buffer_patch.py (license header + line wrapping)

yeyu-nvidia
and
claude
committed
review: converge DFlash export RoPE to a config field; gate export callback on FSDP2

yeyu-nvidia
and
claude
committed
review: avoid embedding resize for DFlash mask token; reuse an existing reserved row

yeyu-nvidia
and
claude
committed
review: document applicability/scope of fsdp2_buffer_patch

yeyu-nvidia
and
claude
committed
test: cover DFlash export rope-scaling field + mask-token resolution

yeyu-nvidia
and
claude
committed
review: rename PATCH_FSDP2_BUFFERS -> PATCH_FSDP2_BUFFERS_TF457; trim dflash doc

yeyu-nvidia
and
claude
committed
fix: CI failures — drop bogus num_error, ruff nits, remove scratch YAMLs

yeyu-nvidia
and
claude
committed

Commits on Jun 11, 2026

Commits on Jun 12, 2026

Commits on Jun 13, 2026

review: gate export callback by fsdp_state_dict_type, rename it, move dtype diag

yeyu-nvidia
and
claude
committed

Commits on Jun 15, 2026