feat(performance): add --mlperf_flavor for MLPerf v6.0 apples-to-appl…#3878
Open
rsalagame-nvidia wants to merge 2 commits into
Open
feat(performance): add --mlperf_flavor for MLPerf v6.0 apples-to-appl…#3878rsalagame-nvidia wants to merge 2 commits into
rsalagame-nvidia wants to merge 2 commits into
Conversation
…es Llama3.1 runs (GB200 only) Signed-off-by: Rahul Salagame <rsalagame@login-ptyche01.ptyche.clusters.nvidia.com>
rsalagame-nvidia
requested review from
a team,
erhoo82 and
malay-nagda
as code owners
yaoyu-33
added
area:perf
jepio
reviewed
May 19, 2026
|
|
||
| # (model_recipe_name, compute_dtype, num_gpus) -> (TP, PP, VP, CP, MBS, GBS, parity_mode); shapes derived from v5.1 NVIDIA submission configs. | ||
| _MLPERF_V51_SHAPES: Dict[Tuple[str, str, int], Tuple[int, int, int, int, int, int, str]] = { | ||
| ("llama3_8b", "fp8_cs", 8): (1, 1, 1, 1, 1, 8, "F16_ATTN"), |
There was a problem hiding this comment.
Suggested change
| ("llama3_8b", "fp8_cs", 8): (1, 1, 1, 1, 1, 8, "F16_ATTN"), | |
| ("llama3_8b", "fp8_cs", 8): (1, 1, 1, 1, 2, 16, "F16_ATTN"), |
jepio
reviewed
May 19, 2026
Comment on lines
+20
to
+29
| ("llama3_8b", "fp8_cs", 16): (1, 1, 1, 2, 1, 8, "F16_ATTN"), | ||
| ("llama3_8b", "fp8_cs", 32): (1, 1, 1, 2, 1, 16, "F16_ATTN"), | ||
| ("llama3_8b", "fp8_cs", 64): (1, 1, 1, 2, 1, 32, "F16_ATTN"), | ||
| ("llama3_8b", "fp8_cs", 72): (1, 1, 1, 2, 1, 36, "F16_ATTN"), | ||
| ("llama3_8b", "fp8_cs", 128): (2, 1, 1, 4, 1, 16, "F16_ATTN"), | ||
| ("llama3_8b", "nvfp4", 8): (1, 1, 1, 1, 2, 16, "FP4_ATTN"), | ||
| ("llama31_405b","fp8_cs", 256): (4, 8, 8, 2, 1, 576, "405B"), | ||
| ("llama31_405b","fp8_cs", 512): (4, 8, 8, 2, 1, 1152, "405B"), | ||
| ("llama31_405b","nvfp4", 256): (4, 8, 8, 2, 1, 576, "405B"), | ||
| ("llama31_405b","nvfp4", 512): (4, 8, 8, 2, 1, 1152, "405B"), |
There was a problem hiding this comment.
For maintainability it would be helpful if we had a variable at the end of the datastructure that encodes the mlperf tuning source: llama31_8b_2x4, llama31_8b_18x4, llama31_8b_512x4 (or _small,_med,_large). And then in the functions where we apply the settings we'd have blocks (or function calls) that configure all overrides for that specific config. This would be instead of parity_mode.
jepio
reviewed
May 19, 2026
Comment on lines
+500
to
+505
| """Apply MLPerf v5.1 apples-to-apples recipe knobs; gated by MLPERF_PARITY_{F16_ATTN,FP4_ATTN,405B} env vars (set by perf_plugins).""" | ||
| f16_only = bool(os.environ.get("MLPERF_PARITY_F16_ATTN")) | ||
| fp4_attn = bool(os.environ.get("MLPERF_PARITY_FP4_ATTN")) | ||
| parity_405b = bool(os.environ.get("MLPERF_PARITY_405B")) | ||
| if not (f16_only or fp4_attn or parity_405b): | ||
| return recipe |
There was a problem hiding this comment.
Do we have to pass custom env variables between stages - can we have a get_mlperf_flavor_config?
…ataset optional (falls back to mock); fix 8 GPU FP8 shape; add 16/32/64 FP4 shape entries Signed-off-by: Rahul Salagame <rsalagame@login-ptyche01.ptyche.clusters.nvidia.com>
rsalagame-nvidia
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds --mlperf_flavor to scripts/performance/setup_experiment.py for MLPerf v6.0 apples-to-apples Llama3.1 runs on GB200. New utils/mlperf_flavor.py resolves the v6.0 shape per (model_recipe_name, compute_dtype, num_gpus) for Llama3 8B (8/16/32/64/72/128 GPU FP8 + 8 GPU NVFP4) and Llama3.1 405B (256/512 GPU FP8+NVFP4), wires the MLPerf preprocessed C4 dataset, appends container mounts, and triggers gated parity knobs in utils/overrides.py (up to 6 recipe knobs incl. CUDA graphs) + perf_plugins.py (parity env vars matching the v5.1 effective env). Gated to gpu=gb200 since shapes are derived from v6.0 GB200 reference configs; other GPU types not yet validated. Fully opt-in; no behavior change without the flag.