[aoshen02]

Summary

vllm/models/deepseek_v4/** reads six fields straight off the config object
while building layers, with no default anywhere in the stack:

self.n_groups = config.o_groups                       # attention.py:212
self.topk_tokens = config.index_topk                  # attention.py:928
self.hc_mult = config.hc_mult                         # amd/model.py:724 (+20 more)
self.swiglu_limit = config.swiglu_limit               # amd/model.py:535
is_hash_moe = ... < config.num_hash_layers            # amd/model.py:553

A checkpoint whose config.json omits one of them fails deep inside layer
construction with:

AttributeError: 'DeepseekV4Config' object has no attribute 'o_groups'

which names the attribute but not the file it belongs in. This PR raises from
the config classes instead, naming config.json and the full required set.

Why not defaults

Because a default would be a guess at the model's shape. index_topk sets the
sparse selection size, compress_ratios the per-layer cache type, o_groups
the o-proj grouping — a wrong value builds a model that does not match the
weights and fails later, or worse, silently produces bad output. An explicit
error is the correct behaviour for a checkpoint that does not say.

Cost

The hook is __getattr__, which Python only calls when normal lookup fails, so
a complete checkpoint pays nothing and behaves exactly as before. Unrelated
missing attributes keep raising a plain AttributeError(name), so copy,
to_dict and the transformers machinery are unaffected — covered by a test.

Applies to both DeepseekV4Config and DeepseekV41Config (V4.1 flattens
text_config onto the top level and has the same six fields missing when the
checkpoint omits them).

Not a duplicate

gh pr list --search "DeepseekV4Config" returns only #44031, which normalizes
compress_ratios for transformers >= 4.57 inside the model files — a different
problem (a field that is present but shaped differently). No open PR or issue
covers the missing-field failure mode.

Test plan

pytest tests/transformers_utils/test_deepseek_v4_config_fields.py -q
# 16 passed

CPU-only. Covers each required field on both config classes, unrelated
attributes still raising plainly, and a complete config surviving deepcopy
and to_dict.

pre-commit run --files vllm/transformers_utils/configs/deepseek_v4.py \
  vllm/transformers_utils/configs/deepseek_v41.py \
  tests/transformers_utils/test_deepseek_v4_config_fields.py
# passed

No model-quality impact: no change for any checkpoint that loads today.

AI assistance

Developed with AI assistance (Claude). The missing-field list came from reading
every bare config.<field> access in the DeepSeek-V4/V4.1 model packages, and
was confirmed by constructing both config classes and checking each attribute. I
reviewed every changed line, ran the tests above, and I am the accountable
submitter.

🤖 Generated with Claude Code

Review rounds

A later adversarial round, told about the deleted list and asked for new
defects, went after the __getattr__ itself: copy/deepcopy, pickle protocols
0-5, cloudpickle and ForkingPickler (these configs cross the process boundary
to workers), probes on a pre-__init__ instance, hasattr fallbacks,
attribute_map and get_text_config() delegation, and dataclasses field
introspection. It executed all of them against transformers 5.12.1: both
DeepseekV4Config and DeepseekV41Config pass, and no caller matches the old
error text. No new defects.

[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.

[aoshen02]

/ci run

[github-actions]

❌ This PR is 1 commit behind upstream main. Your branch must contain every commit currently on upstream main. No new CI build was started. Merge or rebase onto the latest main, then rerun /ci run. To test this branch at your own risk, use /ci run --allow-stale.

[aoshen02]

/ci run

[github-actions]

❌ This PR is 2 commits behind upstream main. Your branch must contain every commit currently on upstream main. No new CI build was started. Merge or rebase onto the latest main, then rerun /ci run. To test this branch at your own risk, use /ci run --allow-stale.

[aoshen02]

/ci run

[github-actions]

❌ This PR is 2 commits behind upstream main. Your branch must contain every commit currently on upstream main. No new CI build was started. Merge or rebase onto the latest main, then rerun /ci run. To test this branch at your own risk, use /ci run --allow-stale.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @aoshen02.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[hmellor]

Closing as we plan to delete this file in #60823