Repository navigation
docs(misc): point GB200 fine-tunes at the card's own import - #6366
Conversation
pruprakash
added
docs
|
/review |
svcnvidia-nemo-ci
left a comment
svcnvidia-nemo-ci
left a comment
There was a problem hiding this comment.
Reviewed all eight changed files. The five checkpoint corrections match the cards’ published GPU import destinations, and the generated catalog, Sphinx pages, and Fern mirrors consistently contain the corrected paths. Two blocking coverage gaps remain in the new consistency guard, detailed inline.
Static review only: no tests or GPU workloads were executed.
Suggested test cases
test_published_pretrained_checkpoints_are_written_by_the_same_card— the corrected cards have no unproduced checkpoints beyond the two explicit GLM baselines.test_pretrained_checkpoint_check_rejects_an_unwritten_path— retain GB200 mismatch rejection and acceptance of the H100 import checkpoint.test_pretrained_checkpoint_check_covers_nested_leaves(proposed) — detect missing checkpoints in precision variants and weak-scaling points, while retaining valid producer relationships.test_pretrained_checkpoint_options_accept_equals_syntax(proposed) — exercise both supported option-value forms for readers and writers.test_generated_outputs_and_navigation_are_current— generated-output checking succeeds.- No perf tests impacted.
Review generated by codex
The Qwen3.6-35B-A3B GB200 SFT/PEFT items and the DeepSeek-V4-Flash GB200 SFT, long-context SFT and PEFT items read --pretrained_checkpoint directories that no item on their card writes. Point them at the path the card's GPU import writes, as the sibling items already do, and regenerate the catalog and model pages. Detected by: megatron-bridge QA (NVBug 6875021) Signed-off-by: Pruthviraj Prakash <pruprakash@nvidia.com>
3a2884c to
761a0fe
Compare
|
/review |
svcnvidia-nemo-ci
left a comment
svcnvidia-nemo-ci
left a comment
There was a problem hiding this comment.
Follow-up review complete; no eligible findings remain.
Both prior findings no longer apply: the checkpoint guard and its regression test were removed, and the published-command test module matches base. The five corrected readers match their cards’ published GPU-import destinations; the inspected catalog, Sphinx pages, and Fern mirrors contain the same corrected paths.
The previously reviewed head tree was unavailable, so a complete inter-review comparison was not possible. This review verified the prior findings and claimed removal, plus the current path corrections and generated copies. Static review only; no tests or GPU workloads were executed.
Suggested test cases
test_normalized_commands_equal_card_scalars— all five corrected catalog commands must equal their YAML command scalars.test_generated_outputs_and_navigation_are_current—generate(REPO_ROOT, check=True)must return true, with exactly one generated section per model page.- No perf tests impacted.
Review generated by codex
This comment was marked as outdated.
This comment was marked as outdated.
This comment was marked as off-topic.
This comment was marked as off-topic.
|
/ok to test 761a0fe |
|
Automatic Claude reviews have been retired. To request a pull-request review, post a comment containing: Add |
|
Retrospective QA-fix tracking issue: #6385. It records the fix authored by @pruprakash and delivered by this merged PR. |
This branch was successfully deployed
4 participants
What does this PR do ?
Fixes five verified GB200 fine-tune commands on the Qwen3.6-35B-A3B and DeepSeek-V4-Flash cards that load a
--pretrained_checkpointno item on their card writes, by pointing them at the card's own GPU import.Changelog
examples/model_verification_cards/qwen3.6-35b-a3b/card.yaml: GB200sft/peftreadimported-megatron/iter_0000000instead ofgb200-imported/iter_0000000.examples/model_verification_cards/deepseek-v4-flash/card.yaml: GB200sft/sft_long_context/peftreaddsv4-flash/import-gpu/iter_0000000instead ofwork/models/deepseek-v4-flash-megatron.docs/_static/model-verification/catalog-v1.json,docs/models/...,docs/fern/.../*.mdx: regenerated (--checkpasses).Additional Information
pretrain.GB200(DeepSeek) already read the import path. Same class as docs: correct Nemotron verification artifact paths #6286.--megatron-pathwritten by the card's ownhf_to_megatron_gpuitem.validate_card.pypasses for both cards,generate_model_verification_catalog.py --checkis clean, and the card/catalog/doc-consistency unit tests pass (70). QA's card audit (nmfw_tests) flagged exactly these 5 leaves on main.sft/sft_long_contextread a Hugging Face directory (glm5-2/hf-<rev>) that no item stages. They are not changed here, because repointing would change the checkpoint format.