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

Inline comments:
In `@bionemo-recipes/recipes/evo2_megatron/README.md`:
- Around line 157-162: Add a language identifier (bash) to the fenced code
blocks containing the shell commands (e.g., the blocks that show
evo2_export_mbridge_to_vortex and evo2_convert_savanna_to_mbridge) so
markdownlint stops warning; locate the backtick fences around those command
examples and change the opening fence from ``` to ```bash for each occurrence
(including the second block around the evo2_convert_savanna_to_mbridge example).

In
`@bionemo-recipes/recipes/evo2_megatron/src/bionemo/evo2/models/eden_provider.py`:
- Around line 130-139: The patch_eden_tokenizer function is defined but never
used at runtime; either remove this dead function and its export (and update the
unit test to use the runtime implementation) or integrate it into the recipes
tokenizer flow by importing and calling patch_eden_tokenizer immediately after
the tokenizer is constructed in predict.py (so the tokenizer uses BOS=1, EOS=2,
SEP=3, PAD=0); also ensure any exported symbol lists are updated to remove the
orphaned function if you delete it and avoid duplicating functionality already
present in the other package-level patch implementation.

In
`@bionemo-recipes/recipes/evo2_megatron/src/bionemo/evo2/models/evo2_provider.py`:
- Around line 1115-1132: The bug: MODEL_OPTIONS is built as
{**HYENA_MODEL_OPTIONS, **EDEN_MODEL_OPTIONS} which lets EDEN override HYENA on
key collision, but infer_model_type checks HYENA first causing inconsistent
behavior; fix by adding a runtime collision check after constructing
MODEL_OPTIONS that computes collisions = set(HYENA_MODEL_OPTIONS) &
set(EDEN_MODEL_OPTIONS) and either raise a clear ValueError (or log and resolve
to a chosen precedence) if collisions is non-empty, and update infer_model_type
to rely on MODEL_OPTIONS (or document the chosen precedence) so behavior is
consistent; also update infer_model_type's docstring to include an Args section
describing the model_size parameter.
- Around line 656-699: Hyena20bModelProvider defines an incorrect/unused
attribute short_conv_len and an orphan hyena_out_proj_bias; remove
short_conv_len (it duplicates/typoed counterpart hyena_short_conv_len) and
delete hyena_out_proj_bias unless you add a corresponding field in HyenaConfig
and wire it into the model code; update Hyena20bModelProvider by removing the
short_conv_len and hyena_out_proj_bias declarations (or rename short_conv_len to
hyena_short_conv_len only if HyenaConfig lacks that field and you also add it
there).

In `@bionemo-recipes/recipes/evo2_megatron/src/bionemo/evo2/run/infer.py`:
- Around line 334-336: You're directly mutating the private attribute
model_provider._pg_collection with
ProcessGroupCollection.use_mpu_process_groups(), which is fragile; instead
expose and use a public API or setter on the provider (e.g., add or call a
method like set_process_group_collection or a constructor/init parameter on the
ModelProvider class) so non-Hyena models can be configured without touching
internals—update the provider implementation to accept and store the
ProcessGroupCollection via that public method and replace the direct assignment
at the call site with the new setter or init call, keeping Hyena models'
internal behavior unchanged.

In
`@bionemo-recipes/recipes/evo2_megatron/src/bionemo/evo2/utils/checkpoint/mbridge_to_vortex.py`:
- Around line 133-167: The export currently silently omits required tensors
(e.g., embedding.word_embeddings.weight, decoder.final_norm.weight and per-layer
weights produced by _convert_hyena_layer, _convert_attention_layer, and
_convert_mlp) so add a validation pass after the loop that defines the mandatory
target keys (embedding_layer.weight, unembed.weight, norm.scale and all expected
per-layer keys derived from prefix/block_prefix and the pattern) and check their
presence in the resulting vortex_sd (or mbridge_state_dict if conversions expect
source keys); collect missing keys into a list and raise a descriptive exception
listing layer and key names (including references to the layer index and symbol)
before writing the .pt/config.json to fail fast on bad --model-size or --no-te
choices.
- Around line 48-56: The current logic assumes mbridge_ckpt_dir is the
checkpoint root and fails if the user passes an iter_* directory; modify the
resolver around latest_file/iter_dir so that if mbridge_ckpt_dir.name matches
the iter_* pattern (e.g., startswith "iter_" or matches r"^iter_\d+$") you treat
mbridge_ckpt_dir itself as the iter_dir; otherwise keep the existing flow (check
for latest_checkpointed_iteration.txt, parse iteration into iter_{:07d}, or
fallback to glob("iter_*")). Update uses of iteration/iter_dirs to reflect this
early-path selection so valid direct iter_* paths are accepted.

In
`@bionemo-recipes/recipes/evo2_megatron/tests/bionemo/evo2/test_checkpoint_roundtrip.py`:
- Around line 39-79: The fixtures download mutable HuggingFace checkpoints and
allow unsafe pickle loading; update savanna_checkpoint_path and
vortex_reference_path to pass explicit immutable commit SHAs via the revision=
parameter in their hf_hub_download(...) calls (use the specific commit SHA for
SAVANNA_1B_REPO and VORTEX_1B_REPO to pin the golden data), and modify
vortex_reference_sd to load the reference safely by calling torch.load(...,
map_location="cpu", weights_only=True) (or prefer safetensors if a .safetensors
artifact exists) so remote .pt files cannot execute pickle code.

In
`@bionemo-recipes/recipes/evo2_megatron/tests/bionemo/evo2/test_eden_llama_roundtrip.py`:
- Around line 148-186: The test currently computes original_preds via
_run_predict(eden_ckpt, ...) but then only compares original_hf and
reimported_hf logits; change the test to run _run_predict on the roundtripped HF
checkpoint (use hf_reimported_dir) to produce hf_preds and then compare
original_preds["log_probs_seqs"] to hf_preds["log_probs_seqs"] (or the
equivalent per-token log-prob key) using a numeric assert (e.g.,
torch.testing.assert_close or numpy.testing.assert_allclose) so the comparison
actually verifies the roundtrip predictions; locate and update the block that
currently creates original_hf/reimported_hf and the final
torch.testing.assert_close to instead call _run_predict for hf_reimported_dir
(or both HF and eden if you want) and perform the assertion on the
"log_probs_seqs" entries.

---

Nitpick comments:
In
`@bionemo-recipes/recipes/evo2_megatron/src/bionemo/evo2/utils/checkpoint/savanna_to_mbridge.py`:
- Line 132: Add a brief inline comment next to the torch.load call (the line
setting raw = torch.load(str(path), map_location="cpu", weights_only=False))
explaining that weights_only=False is required to load Savanna checkpoints
containing non-tensor data but is unsafe for untrusted files because it uses
pickle and can execute arbitrary code; state that callers must ensure the path
is trusted (or sanitize/validate inputs) before loading.
- Around line 95-96: The bare except in the single-file download block (the
except Exception: that logs "Single-file download failed for {repo_id}, trying
multi-part shards...") is too broad; replace it by catching specific HF and
network-related exceptions (for example
huggingface_hub.exceptions.RepositoryNotFoundError,
huggingface_hub.utils.entry_not_found_error or RevisionNotFoundError, and
requests.exceptions.HTTPError/ConnectionError) and bind the exception (except
(RepositoryNotFoundError, RevisionNotFoundError, HTTPError, ConnectionError) as
e:) so you can log the actual error (include exc_info or str(e)) while allowing
other unexpected exceptions to propagate (re-raise or don’t catch them). Locate
the except block around the single-file download attempt and update the except
clause and logger call accordingly.

In `@bionemo-recipes/recipes/evo2_megatron/tests/bionemo/evo2/run/test_infer.py`:
- Around line 714-805: The Eden tests are launching the installed module (-m
bionemo.evo2.run.infer) instead of the local infer.py; update
test_infer_eden_deterministic to use the same launch pattern as
test_infer_eden_runs: replace the "-m bionemo.evo2.run.infer" style invocation
with the script path from _infer_script_path() (use that function to build the
cmd entry) and ensure the env prepends PYTHONPATH the same way (reuse
PRETEST_ENV modification logic used in test_infer_eden_runs) so the local source
infer.py is executed; reference test_infer_eden_deterministic,
test_infer_eden_runs, _infer_script_path, and PRETEST_ENV to locate where to
apply the changes.

In
`@bionemo-recipes/recipes/evo2_megatron/tests/bionemo/evo2/test_eden_llama_roundtrip.py`:
- Line 63: The list comprehension using torch.load(pf) to build preds should
pass an explicit weights_only argument to avoid the deprecation warning and
ensure correct behavior for tensor-only prediction files; update the
comprehension that references pred_files and preds to call torch.load(pf,
weights_only=True) (or weights_only=False if non-weight objects are expected) so
loading is explicit and future-proof.