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

Inline comments:
In `@src/megatron/bridge/models/deepseek/common.py`:
- Around line 76-80: The code assumes hf_config.architectures exists; change the
check to safely handle missing or None architectures by reading architectures =
getattr(hf_config, "architectures", None) (or using
hf_config.__dict__.get("architectures")) and only perform the
"DeepseekV32ForCausalLM" membership test if architectures is a non-empty
iterable (e.g., isinstance(architectures, (list, tuple)) or
bool(architectures)); then set configs["experimental_attention_variant"],
configs["dsa_indexer_head_dim"], configs["dsa_indexer_n_heads"], and
configs["dsa_indexer_topk"] exactly as before when the membership test passes.
Use the existing symbols hf_config and configs to locate where to apply this
guard.

In `@src/megatron/bridge/models/deepseek/deepseek_v32_bridge.py`:
- Around line 60-64: Replace the hardcoded DSA indexer constants by reading them
from the loaded HuggingFace model config object and assigning those values to
provider.dsa_indexer_head_dim, provider.dsa_indexer_n_heads and
provider.dsa_indexer_topk (and keep provider.dsa_indexer_loss_coeff if it is not
configurable); e.g., pull the values from the HF config variable used when
loading the model (e.g., config or hf_config) and fall back to sensible defaults
only if the config fields are missing, so the conversion uses the actual model
settings rather than fixed numbers.
- Around line 145-159: The code builds inv_freq (stored on
self._deepseek_inv_freq) and injects it into
converted_weights_dict[inv_freq_key] without verifying it matches the expected
tensor shape from the source; update the block that computes and assigns
inv_freq to first obtain a reference tensor (use
next(iter(converted_weights_dict.values())) or
converted_weights_dict[inv_freq_key] if present) and validate that
inv_freq.shape equals the reference tensor.shape (or the expected shape derived
from hf_config.qk_rope_head_dim: torch.arange(0, rotary_dim, 2).shape). If
shapes mismatch, either reshape/expand/convert inv_freq to match the reference
shape or raise/log a clear error and do not write the incompatible tensor; keep
all references to _deepseek_inv_freq, inv_freq, inv_freq_key,
converted_weights_dict, and hf_config.qk_rope_head_dim so the check is applied
before assigning converted_weights_dict[inv_freq_key] = inv_freq.

In `@tests/functional_tests/models/deepseek/test_deepseek_v32_conversion.py`:
- Around line 79-80: The test contains a duplicate call to
model.save_pretrained(model_dir, safe_serialization=True) (two identical
invocations); remove the redundant second call so the model is serialized only
once, keeping the remaining call to model.save_pretrained with
safe_serialization=True (reference: the save_pretrained calls on the model in
test_deepseek_v32_conversion.py).
- Line 171: The test asserts the wrong config key name: update the assertion in
test_deepseek_v32_conversion (the saved dict check) to use "index_topk" ->
"index_topk" is incorrect; change saved["index_top_k"] to saved["index_topk"] so
the test matches the conversion/config pathway that uses index_topk.
- Line 63: Guard the deletion of the fixture attribute so the fixture setup
doesn't raise when the attribute is missing: before removing
config.quantization_config in test_deepseek_v32_conversion.py, check whether the
attribute exists (e.g., with hasattr(config, "quantization_config")) or perform
the deletion inside a try/except catching AttributeError, and only remove it
when present to avoid breaking fixture creation.
- Around line 75-76: Replace the broad "except Exception: pass" in the tokenizer
setup block with targeted error handling: either catch specific expected
exceptions (e.g., OSError, FileNotFoundError, ValueError) or log the exception
and call pytest.skip(reason) or re-raise to fail the test; specifically update
the try/except surrounding the tokenizer initialization (the block that
currently uses "except Exception: pass") so failures are not silently swallowed
but reported or skipped with a clear message.

---

Nitpick comments:
In `@src/megatron/bridge/models/deepseek/deepseek_v32_bridge.py`:
- Around line 102-109: The mapping_registry method is appending a duplicate
AutoMapping already provided by get_common_mapping_list(); remove the redundant
mapping append in mapping_registry (the AutoMapping with megatron_param
"decoder.layers.*.mlp.router.expert_bias" and hf_param
"model.layers.*.mlp.gate.e_score_correction_bias") so that mapping_list only
relies on get_common_mapping_list() and does not re-add the same AutoMapping.

In `@tests/functional_tests/models/deepseek/test_deepseek_v32_conversion.py`:
- Around line 91-101: Add a test category mark (e.g., `@pytest.mark.integration`
or `@pytest.mark.system`) to the test_deepseek_conversion_parallelism test and
guard multi-process cases by checking available GPUs: compute world_size = tp *
pp * ep inside test_deepseek_conversion_parallelism and if world_size > 1 and
torch.cuda.device_count() < world_size call pytest.skip(...) with a clear
message; keep the existing `@pytest.mark.run_only_on`("GPU") and parametrization
but add the new mark and the runtime GPU-count check to avoid running parallel
variants on under-provisioned runners.