TabFM constructor accepts zero/negative dimensions; col_num_blocks=0 builds a model with no attention blocks

Start at the TabFM constructor (init) and reproduce the reported zero and negative dimension cases in a clean container. Check how the dimensional parameters reach PyTorch and ModuleList construction; done means non-positive inputs produce one clear tabfm validation error rather than building empty attention blocks or failing with an unrelated tensor-shape message.

Separate from #91 (which is about config.json keys being merged in unvalidated) — this is about the constructor itself, and it affects anyone calling TabFM(...) directly in Python.

I tested every dimensional parameter at 0 and -1 on b15593e4c1111ddb5f4f30dd2957df2edbaa04ca, in a clean container. Ten of thirteen accept 0 and build a model with no error. Three accept -1 as well.

param                 value  outcome
------------------------------------------------------------------------
embed_dim                 0  BUILT OK  (no validation)
embed_dim                -1  RuntimeError: Trying to create tensor with negative di
max_classes               0  BUILT OK  (no validation)
max_classes              -1  RuntimeError: Trying to create tensor with negative di
col_num_blocks            0  BUILT OK  (no validation)
col_num_blocks           -1  BUILT OK  (no validation)
col_nhead                 0  ZeroDivisionError
col_nhead                -1  RuntimeError: Trying to create tensor with negative di
col_num_inds              0  BUILT OK  (no validation)
col_num_inds             -1  RuntimeError: zeros: Dimension size must be non-negati
row_num_blocks            0  BUILT OK  (no validation)
row_num_blocks           -1  BUILT OK  (no validation)
row_nhead                 0  ZeroDivisionError
row_nhead                -1  RuntimeError: upper bound and lower bound inconsistent
row_num_cls               0  BUILT OK  (no validation)
row_num_cls              -1  RuntimeError: zeros: Dimension size must be non-negati
icl_num_blocks            0  BUILT OK  (no validation)
icl_num_blocks           -1  BUILT OK  (no validation)
icl_nhead                 0  ZeroDivisionError
icl_nhead                -1  RuntimeError: Trying to create tensor with negative di
ff_factor                 0  BUILT OK  (no validation)
ff_factor                -1  RuntimeError: Trying to create tensor with negative di
feature_group_size        0  BUILT OK  (no validation)
feature_group_size       -1  RuntimeError: zeros: Dimension size must be non-negati
num_freq                  0  BUILT OK  (no validation)
num_freq                 -1  RuntimeError: zeros: Dimension size must be non-negati

The one I would draw your attention to is *_num_blocks. Passing 0 or -1 makes range(...) empty, so the ModuleList is empty and the model is built with no attention blocks at all. It still runs and still returns predictions. Nothing warns. That is a worse outcome than the crashes elsewhere in the table, because a crash is at least honest about what happened.

Where errors do occur they come from PyTorch rather than from tabfm, so the message describes a tensor shape rather than the parameter the caller actually got wrong — Trying to create tensor with negative dimension does not tell someone that ff_factor was the problem.

A small guard at the top of __init__ rejecting non-positive values for these parameters would turn all twenty-six rows into one clear message. I am happy to send that PR if you would like it — I did not want to presume which parameters you consider legitimately zero-able (decoder_hidden is already None-able, so there may be others by design).

Disclosure: I used an AI assistant to help find this. I ran the matrix myself.