Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@tensorrt_llm/_torch/model_config.py`:
- Around line 662-666: The CI formatting hooks failed for
tensorrt_llm/_torch/model_config.py; run the project's formatters locally and
commit the fixes: run the pre-commit formatter (pre-commit run --files
tensorrt_llm/_torch/model_config.py) or run yapf/ruff-format directly to
reformat the occurrences around ceil_div usage for num_heads and hidden_size
(references: ceil_div, num_heads, hidden_size, self.pretrained_config,
attn_tp_size, attn_cp_size) so the file matches the project's style and passes
CI.

In `@tests/unittest/_torch/test_model_config.py`:
- Around line 73-77: CI failed due to formatting in
tests/unittest/_torch/test_model_config.py; run the configured formatters and
apply their fixes: execute the pre-commit hook (pre-commit run --files
tests/unittest/_torch/test_model_config.py) or run yapf/ruff-format manually,
then ensure the block around the attn_tp_size/attn_cp_size assignments and the
ceil_div function follows project style (correct spacing, line breaks, and
indentation) so that function ceil_div and the surrounding lines satisfy yapf
and ruff-format checks before committing.

---

Outside diff comments:
In `@tensorrt_llm/_torch/model_config.py`:
- Around line 740-748: The assertion hidden_size % num_heads == 0 can fail when
attn_cp_size > 1 because num_heads and hidden_size were divided by different
parallelism factors; replace the assert with a safe computation that derives
head_size = ceil_div(hidden_size, num_heads) (or equivalent integer ceiling
division already used elsewhere) and assign that to head_size while logging a
warning about rounding instead of crashing; update the block that sets head_size
(referencing head_size, hidden_size, num_heads and the existing ceil_div helper)
to compute and use the rounded head size rather than asserting exact
divisibility.

---

Nitpick comments:
In `@tensorrt_llm/_torch/model_config.py`:
- Around line 659-660: Remove the local duplicate ceil_div function in
model_config and import the existing typed utility from
tensorrt_llm._torch.pyexecutor._util; replace the local definition with "from
tensorrt_llm._torch.pyexecutor._util import ceil_div" and ensure any callers in
model_config continue to use ceil_div (preserving type compatibility), so the
code reuses the single canonical implementation instead of duplicating it.

In `@tests/unittest/_torch/test_model_config.py`:
- Around line 76-77: The test defines a duplicate helper ceil_div which
duplicates the implementation in production (model_config.py and
pyexecutor/_util.py); remove the local ceil_div in
tests/unittest/_torch/test_model_config.py and import the canonical function
from the production utility (e.g., import ceil_div from pyexecutor._util or the
public function in model_config) so the test reuses the single source of truth
and stays in sync with production code.
- Around line 56-67: The current test
test_get_bindings_model_config_attention_dp_attn_tp_override only uses
num_attention_heads that divide evenly by the product of parallelism factors, so
it doesn't catch ceiling-division bugs; update the test to parametrically
exercise non-evenly-divisible head counts by adding cases (e.g.,
num_attention_heads=10 and 9) alongside the existing 16 and compute the expected
per-shard head count (expected_sharded_heads = ceil(num_attention_heads /
(tp_size * cp_size))) and assert the resulting Mapping/make_pretrained_config
behavior produces that expected_sharded_heads (referencing Mapping,
make_pretrained_config and
test_get_bindings_model_config_attention_dp_attn_tp_override to locate where to
add the new parametrized inputs and assertions).