Problem
The dry-run estimation path reads model-shape fields only from the top-level config object.
Qwen3.5-MoE keeps the relevant fields under text_config, so the estimator cannot count parameters correctly for that model family.
Affected code
auto_round/estimation.py:48-56auto_round/estimation.py:216-217
Repro
This reproduces with the Qwen3.5 MoE config shape:
python - <<'PY'
from transformers import AutoConfig
cfg = AutoConfig.from_pretrained("Qwen/Qwen3.5-35B-A3B", trust_remote_code=True)
print("top-level hidden_size:", getattr(cfg, "hidden_size", None))
print("nested hidden_size:", getattr(cfg.text_config, "hidden_size", None))
PYExpected result:
- the estimator should fall back to
cfg.text_config for fields such as hidden_size, num_hidden_layers, intermediate_size, and vocab_size
Actual result:
- those top-level fields are
None for Qwen3.5-MoE-style configs, so parameter counting and the estimate result break or undercount
If you run the dry-run path against Qwen/Qwen3.5-35B-A3B, it cannot size the model correctly for this reason.
Why this matters
Dry-run estimation is most useful on large MoE models where memory headroom is tight.
Qwen3.5-MoE is one of those models.
Suggested fix
Add a small helper that reads each required field from the top-level config first and then falls back to config.text_config when present.
Use that helper in both _count_parameters() and _build_estimate_result().
Problem
The dry-run estimation path reads model-shape fields only from the top-level config object.
Qwen3.5-MoE keeps the relevant fields under
text_config, so the estimator cannot count parameters correctly for that model family.Affected code
auto_round/estimation.py:48-56auto_round/estimation.py:216-217Repro
This reproduces with the Qwen3.5 MoE config shape:
Expected result:
cfg.text_configfor fields such ashidden_size,num_hidden_layers,intermediate_size, andvocab_sizeActual result:
Nonefor Qwen3.5-MoE-style configs, so parameter counting and the estimate result break or undercountIf you run the dry-run path against
Qwen/Qwen3.5-35B-A3B, it cannot size the model correctly for this reason.Why this matters
Dry-run estimation is most useful on large MoE models where memory headroom is tight.
Qwen3.5-MoE is one of those models.
Suggested fix
Add a small helper that reads each required field from the top-level config first and then falls back to
config.text_configwhen present.Use that helper in both
_count_parameters()and_build_estimate_result().