feat: add EAGLE3 draft training for Kimi K2.7 Code

[maocheng23]

Summary

Adds an EAGLE3 draft-training recipe for the Kimi K2.7 Code target, stacked on the DFlash support in sgl-project#593. Includes the sglang / transformers compatibility fixes required to run train_eagle3.py (online capture) against a current sglang and a multimodal (KimiK25ForConditionalGeneration) target.

Companion to the DFlash draft from sgl-project#593 — same target, data, and tokenize cache — so EAGLE3 vs DFlash can be compared head-to-head (accept-length).

What's included

• configs/kimi-k2.7-code-eagle3.json — EAGLE3 draft config: 1-layer Llama draft, hidden 7168, eagle_aux_hidden_state_layer_ids [1,29,57] (valid for the 61-layer K2.7 target), vocab_size 163840, draft_vocab_size 32000, K2.7 token ids.
• examples/run_kimi_k2.7_code_eagle3.sh — multi-node launch. Uses pure TP for the capture and max_length 2048 so the 1T target's per-rank capture activations fit on 184 GB GPUs (rationale in comments).
• Compatibility fixes (needed for eagle3-online on a newer sglang + a multimodal target):
  • sglang_backend/utils.py: wrap_eagle3_logits_processors_in_module now snapshots named_modules() before mutating and resolves dotted (nested) names, so a multimodal target's nested language_model.logits_processor is actually replaced — previously it raised dictionary changed size during iteration, and even past that would have silently failed to wrap the nested processor. The replaced forward reads multi_item_delimiter_indices off the ForwardBatch (newer sglang dropped LogitsProcessor.multi_item_delimiter).
  • eagle3_target_model.py: logprob_start_len = -1 + set_extend_input_len(...) in extend()/extend_vlm() so prepare_mlp_sync_batch_raw's per-request logprob assertion holds under DP attention; plus a defensive import of prepare_mlp_sync_batch_raw.
  • auto.py: guard the transformers-4.57-only custom_backend import so AutoEagle3DraftModel loads on transformers ≥ 5.

Base

Stacked on sgl-project#593 (reproduction) so the diff is only the EAGLE3 additions. Once sgl-project#593 lands, this can be retargeted to main.

Status

Training in progress on 4×GB200 (WORLD=16); draft loss is converging.