specdec: config_overrides for nested text_config checkpoints + load VLM-capable bases in merge_lora - #2289

Open

yeyu-nvidia wants to merge 10 commits into

NVIDIA:mainfrom

yeyu-nvidia:yeyu/specdec-config-overrides

Commits on Aug 31, 2026

feat: config_overrides + cosmos3_omni loading in load_vlm_or_llm

yeyu-nvidia
and
claude
committed
d2943e0
fix: import transformers_cosmos3 before AutoConfig so cosmos3_omni is registered

yeyu-nvidia
and
claude
committed
ce56493
ar_validate: add --config_overrides passthrough to load_vlm_or_llm
yeyu-nvidia
committed
4ff57d5
export_hf_checkpoint: add --config_overrides passthrough
yeyu-nvidia
committed
b655c47
merge_lora: load the base via load_vlm_or_llm instead of AutoModelForCausalLM
yeyu-nvidia
committed
3308ffc
use contextlib.suppress for the optional transformers_cosmos3 import
yeyu-nvidia
committed
76860b1
address review: fix VLM class regression, config/kwargs interaction, silent drops
yeyu-nvidia
committed
8eed412
address review: share one config_overrides parser across the three scripts
yeyu-nvidia
committed
efb2ca2
address second review round: llm_config targeting, fake-base warn, parse once
yeyu-nvidia
committed
4322e5e
address review: reject NaN/Infinity in --config_overrides
yeyu-nvidia
committed
91402fa