Repository navigation
fix(recipe): restore DeepSeek V4 Flash GB200/GB300 perf layout verified on the model card - #6463
pruprakash wants to merge 1 commit into
Conversation
…ed on the model card The GB200 and GB300 128-GPU MXFP8 recipes moved to EP32 (GB300 also MBS 2) with 32 HybridEP ranks per NVLink domain, while the verified model card records EP64, MBS 1 and 64 HybridEP ranks. Restore the card-verified layout. VRNVL72, which inherits the GB300 recipe, keeps its own EP32 / MBS 2 layout. Signed-off-by: Pruthviraj Prakash <pruprakash@nvidia.com>
pruprakash
added
bug
|
/review |
|
/ok to test b331dd4 |
|
Automatic Claude reviews have been retired. To request a pull-request review, post a comment containing: Add |
svcnvidia-nemo-ci
left a comment
svcnvidia-nemo-ci
left a comment
There was a problem hiding this comment.
No verified findings. The GB200 recipe and GB300 recipe restore the card-recorded EP64/MBS1 layout. VRNVL72 preserves EP32/MBS2, 32 HybridEP ranks, and its clamp workaround. The B200/B300 descendants retain their explicit overrides.
Static review only; tests and 128-GPU performance runs were not executed.
Suggested test cases
Existing cases in the recipe regression tests:
test_deepseek_v4_flash_128gpu_gb200_fp8mx_configtest_deepseek_v4_flash_128gpu_gb300_fp8mx_configtest_deepseek_v4_flash_128gpu_vrnvl72_fp8mx_disables_clamp_for_cutedsltest_deepseek_v4_flash_128gpu_b300_fp8mx_configtest_deepseek_v4_flash_128gpu_b200_fp8mx_config
Note: formal approval is not available for this request, so this review is posted as a comment only.
Review generated by codex
cuichenx
left a comment
cuichenx
left a comment
There was a problem hiding this comment.
Sorry i probably misunderstood the ask earlier. We should not change perf recipes. the verification card needs to be updated instead to reflect the new changes
|
need to rerun and update https://github.com/NVIDIA-NeMo/Megatron-Bridge/blob/main/examples/model_verification_cards/deepseek-v4-flash/card.yaml#L473-L525, sorry for the confusion |
This branch was successfully deployed
3 participants
What does this PR do ?
Restores the DeepSeek V4 Flash GB200/GB300 128-GPU MXFP8 perf recipes to the layout on the verified model card. Per Chen's comment, card evidence takes precedence; this restores the card-verified layout.
Summary
The card's
pretrain_performance.GB200/.GB300commands verified TP1/PP1/CP1/EP64/ETP1, GBS/MBS 2048/1, 64 HybridEP ranks per NVLink domain. #5845 later moved both recipes to EP32 (GB300 also MBS 2) with 32 HybridEP ranks, so the verified commands now build a different layout on main.Changelog
perf_recipes/deepseek/gb200/deepseek_v4.py: EP 64, HybridEP ranks 64.perf_recipes/deepseek/gb300/deepseek_v4.py: drop EP32/MBS2 overrides (inherits EP64/MBS1), HybridEP ranks 64.perf_recipes/deepseek/vrnvl72/deepseek_v4.py: pin its own EP32/MBS2 so VRNVL72 keeps its current layout.tests/unit_tests/recipes/test_deepseek_v4_flash_perf_recipes.py: GB200/GB300 back to the card layout; VRNVL72 asserts its pin.Verification
assert 32 == 64); the branch passes all 5 tests.Nemotron 3 Super GB300 NVFP4 (same NVBug) is out of scope: its recipe moved to the NCCL EP dispatcher, so EP64 alone would not match the card.
Reviewer: @cuichenx