Model-Optimizer/modelopt_recipes/general/speculative_decoding/lilicorr.yaml

# LiLiCorr speculative-decoding training recipe (arXiv:2608.20530).
#
# LiLiCorr reuses the DFlash mode/pipeline and adds a reranker over the candidate
# lattice the parallel backbone already produces, selected via
# dflash_architecture_config.projector_type=lilicorr. The backbone keeps its
# top-k candidates per slot; a small transformer scores transitions between them
# and serving commits a path greedily, left to right. Trained jointly with the
# backbone, so the drafter learns to propose candidates that correlate into longer
# accepted sequences.
#
# The objective is the DFlash block loss plus three weighted terms:
#   loss = dflash_loss + w_ce*CE + w_margin*hinge + w_pen*penalty
# No outer multiplier, so `loss == origin_loss + lilicorr_loss` holds exactly.
# Online training is required (the penalty reads the target model's logits).
#
# This file is the published `base` variant. The `margin` variant differs only in
# the composition of the cross-entropy weight — see the two-line override below.
# Override fields via an OmegaConf dotlist.

# modelopt-schema: modelopt.recipe.config.ModelOptDFlashRecipe
metadata:
  description: LiLiCorr training recipe (DFlash backbone + candidate-lattice reranker).

# maps to ModelArguments (main.py)
model:
  model_name_or_path:
  trust_remote_code: false
  use_fake_base_for_offline: false

# maps to DataArguments (main.py)
data:
  # Online only: the distractor penalty weights each candidate by the target
  # model's own logit gap, so there has to be a target model in the process.
  mode: online
  data_path:
  # Jinja chat template with {% generation %} tags for answer_only_loss.
  chat_template:

# maps to TrainingArguments (main.py)
training:
  # --- commonly modified ---
  output_dir:
  num_train_epochs: 6
  per_device_train_batch_size: 1
  gradient_accumulation_steps: 1
  learning_rate: 6.0e-4
  warmup_ratio: 0.04
  training_seq_len: 3072
  logging_steps: 50
  save_steps: 1000
  seed: 42
  cp_size: 1
  dp_shard_size: 1
  disable_tqdm: true
  # Keep off: eval runs the DFlash backbone only (the reranker is not applied in
  # pseudo_speculative_generate), so AR here would report the backbone alone and
  # understate the trained model. Compare via export + a serving benchmark.
  estimate_ar: false
  ar_validate_steps: 0
  answer_only_loss: true

  # --- rarely modified ---
  do_eval: false
  # Cosine, unlike the sibling recipes' linear: the published variants were trained
  # under a linear warmup into a cosine decay (specforge's CosineAnnealingWarmupLR),
  # and the schedule is part of the recipe those numbers came from.
  lr_scheduler_type: cosine
  save_strategy: steps
  weight_decay: 0.0
  max_grad_norm: 1.0
  dataloader_drop_last: true
  bf16: true
  tf32: true
  remove_unused_columns: false
  ddp_timeout: 1800
  report_to: tensorboard

# maps to DFlashConfig (modelopt/torch/speculative/config.py).
dflash:
  dflash_block_size: 16
  dflash_num_anchors: 512
  dflash_use_torch_compile: false
  # The reranker's terms are added to the plain weighted cross-entropy. Turning KD
  # on would replace that base term and change the objective the published
  # checkpoints were trained under.
  dflash_self_logit_distillation: false
  # Static exponential position decay, gamma=7 for block_size=16: early in-block
  # slots gate acceptance, so they carry more weight. Not 'dpace' — the published
  # variants were trained on the static decay.
  dflash_loss_objective: decay
  dflash_loss_decay_factor: 7.0
  # Qwen3 has no native mask token; 151669 is an unused id used by the reference.
  dflash_mask_token_id: 151669

  # Objective composition. Absolute weights, validated all-or-nothing.
  # `base`   (this file): w_ce 0.25,  w_margin 0.0
  # `margin`            : w_ce 0.125, w_margin 0.125
  # Both keep w_pen 0.25, so the head's total weight is 0.50 either way and the
  # variants differ only in how the cross-entropy block is split.
  # fp32 master weights for the draft. Compute stays bf16 under autocast; this keeps the
  # master copy -- and therefore AdamW's moments -- in fp32, which a bf16 second moment
  # cannot represent the updates of. The published results for this recipe were trained
  # with this on; turning it off changes the optimizer's arithmetic, not just its memory.
  dflash_fp32_master_weights: true

  dflash_lilicorr_w_ce: 0.25
  dflash_lilicorr_w_margin: 0.0
  dflash_lilicorr_w_pen: 0.25
  # Hinge width, in units of the log-potential. Unused while w_margin is 0.
  dflash_lilicorr_margin: 2.0

  dflash_architecture_config:
    num_hidden_layers: 5
    # Draft attention/MLP dims — set explicitly (the draft is an independent
    # Qwen3 model and does NOT inherit these from the base). GQA: 8 KV heads.
    num_attention_heads: 32
    num_key_value_heads: 8
    head_dim: 128
    intermediate_size: 12288
    projector_type: lilicorr

    # Reranker geometry. Every field is required, never defaulted:
    # candidate_topk sets the lattice width and the shape of rank_embedding, and
    # logit_scale/vector_eps change the score without changing any tensor shape —
    # so a guessed value builds a head that loads cleanly and scores a different
    # function. K is otherwise free; the method works at any k.
    # dflash_init_checkpoint restores weights only and reads geometry from here, so
    # warm-starting reproduces a head only if every field matches the one the
    # checkpoint was trained with. lilicorr_logit_scale and lilicorr_vector_eps are
    # the two a mistake would not be caught on, having no effect on any shape.
    lilicorr_candidate_topk: 8
    lilicorr_hidden_size: 1024
    lilicorr_factor_dim: 1024
    lilicorr_num_layers: 2
    lilicorr_num_heads: 8
    lilicorr_mlp_ratio: 2.0
    # The factors are cosines, so this temperature sets their usable range.
    lilicorr_logit_scale: 8.0
    lilicorr_vector_eps: 1.0e-4

# `data.chat_template` is supplied per run; use the shared
# tools/launcher/examples/Qwen/Qwen3-8B/chat_template_train.jinja, as the other
# speculative recipes do, so variants stay comparable to each other.
#
# It differs slightly from the mask the published checkpoints were trained under: the
# reference leaves the empty `<think>` preamble out of the assistant span and supervises
# `<|im_end|>`, where the shared template does the opposite. The token ids are identical
# either way, so this is 6 tokens of supervision per record (4 preamble, 2 end-of-turn)
# and nothing else. Noted because it is invisible in the data, not because it is
# expected to matter at this scale.