[Mypy Fix] Mypy fix for "vllm/model_executor/models/[tT]"

[ZHIHANCHEN03]

Purpose

Mypy fix for vllm/model_executor/models/[tT], continuing the split in #48490 after
[aA][bB] (#48977), [cC][dD] (#52003) and [eE][fF] (#53381).

10 errors across telechat2.py and terratorch.py:

• terratorch.py — annotate loaded_buffers as list[str].
• telechat2.py — narrow layer.self_attn / layer.mlp to LlamaAttention and
  LlamaMLP before touching their submodules. Module.__getattr__ is typed
  Tensor | Module, so the attribute chain could not be checked (8 of the 10 errors).
  The layer types are fixed here: TeleChat2Model does not pass layer_type, so
  LlamaModel uses LlamaDecoderLayer, whose attn_layer_type defaults to
  LlamaAttention and whose mlp is always LlamaMLP.
• telechat2.py — drop the bias with register_parameter("bias", None) instead of
  assigning None to a Parameter-typed attribute. This is the idiom linear.py
  itself uses for bias-less layers (e.g. linear.py L373/L518/L1611), and it is
  exactly what nn.Module.__setattr__ already did for the previous assignment.
• telechat2.py — attribute_map is a ClassVar on PretrainedConfig; the
  instance-level shadow is intentional, so it is marked type: ignore[misc] rather
  than restructured.

No behavior change.

Note on the EXCLUDE entry

EXCLUDE entries are prefix patterns (re.compile(f"^{'|'.join(EXCLUDE)}.*")), so
models/[tT] also covered the models/transformers package, which still has errors.
Simply dropping the entry fails the full-repo run. The entry is therefore narrowed to
models/transformers, which enforces the three [tT] files and leaves that package
excluded. models/transformers is the only subdirectory under models/, so earlier
letter groups did not hit this.

Duplicate-work check

Searched open PRs for mypy in the title and for model_executor/models mypy work.
The only open mypy PRs are #45296 (kimi / [kK]), #51043 (tests/ directories) and
#30960 (mypy → ty migration); none touch [tT].

AI assistance disclosure: Claude Code assisted with the audit, implementation and
review. I reviewed and understand all changed lines and validated the behavior locally.

Test Plan

# the four versions CI runs, on the whole repo
pre-commit run mypy-3.10 --all-files --hook-stage manual
pre-commit run mypy-3.11 --all-files --hook-stage manual
pre-commit run mypy-3.12 --all-files --hook-stage manual
pre-commit run mypy-3.13 --all-files --hook-stage manual

pre-commit run --files tools/pre_commit/mypy.py \
    vllm/model_executor/models/telechat2.py \
    vllm/model_executor/models/teleflm.py \
    vllm/model_executor/models/terratorch.py

Test Result

Run mypy for Python 3.10.................................................Passed
Run mypy for Python 3.11.................................................Passed
Run mypy for Python 3.12.................................................Passed
Run mypy for Python 3.13.................................................Passed
All applicable pre-commit hooks passed.

Before the change the same three files reported 10 errors; the naive version of this
patch (dropping the [tT] entry outright) fails the full-repo run with
Found 145 errors in 12 files, which is what motivated narrowing the entry instead.

I also verified that register_parameter("bias", None) is exactly equivalent to the
previous bias = None assignment — same bias attribute, same _parameters entry,
same state_dict() and named_parameters() — since nn.Module.__setattr__ routes a
None assigned over a registered parameter through register_parameter.

Model-level tests for TeleChat2/TeleFLM/Terratorch were not run locally: they need real
weights, and this was developed on an Apple Silicon CPU build. terratorch.py also
imports the optional terratorch package, which is not installed locally; mypy checks it
under --follow-imports silent, unaffected.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[ZHIHANCHEN03]

@yewentao256 continuing the [aA][bB] / [cC][dD] / [eE][fF] split from #48490 with [tT]. Could you take a look when you get a chance?

One thing worth flagging: EXCLUDE entries are prefix patterns (re.compile(f"^{'|'.join(EXCLUDE)}.*")), so models/[tT] also covered the models/transformers package, which still has errors — dropping the entry outright fails the full-repo run with Found 145 errors in 12 files. I narrowed the entry to models/transformers instead, which enforces the three [tT] files and leaves that package excluded. Happy to take a different approach if you would prefer one.

Verified locally with pre-commit run mypy-3.10/3.11/3.12/3.13 --all-files --hook-stage manual, all Passed.

[yewentao256]

Thanks for the work!

[yewentao256]

Suggested change

	# `[tT]` is otherwise clean; the `transformers` package is still excluded.
	r"vllm/model_executor/models/transformers",
	r"vllm/model_executor/models/transformers",

Comment not needed

[yewentao256]

Suggested change

	# `attribute_map` is a ClassVar on `PretrainedConfig`; shadowing it on the
	# instance is intentional here so LlamaModel can read TeleChat2's fields.
	vllm_config.model_config.hf_config.attribute_map = { # type: ignore[misc]
	vllm_config.model_config.hf_config.attribute_map = { # type: ignore[misc]

comment not needed

[ZHIHANCHEN03]

Thanks for the review! Both comments dropped in 8e3e891.

Re-verified after the change: pre-commit run mypy-3.10/3.11/3.12/3.13 --all-files --hook-stage manual all Passed, and the full hook set passes on the changed files.

[github-actions]

✅ @ZHIHANCHEN03, CI is now available for this PR.

• /ci run starts upstream CI; /amd-ci run starts AMD CI only.
• /ci retry retries failed jobs in the CI build for the current PR head. If the current head has no CI build, it starts a new CI build for the current head containing only jobs that failed in the latest earlier CI build for this PR.
• /amd-ci retry retries failed jobs in AMD CI for the current PR head. Use /amd-ci run when the current head has no AMD CI build.
• /ci cancel cancels scheduled or running CI builds for this PR branch; /amd-ci cancel does the same for AMD CI only.

[ZHIHANCHEN03]

/ci retry

[github-actions]

✅ No earlier CI build exists for this PR. Use /ci run first.

[ZHIHANCHEN03]

/ci run

[github-actions]

✅ Triggered Buildkite CI #85436 for commit 8e3e891f7334.

[hmellor]

/ci run

[github-actions]

✅ Triggered Buildkite CI #85649 for commit fa00c3962464.