Skip to content

[https://nvbugs/6193836][test] Use EP=8 + attention DP for minimax_m2.5 8-GPU perf #14613

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
ruodil wants to merge 2 commits into
base: main
Choose a base branch
from
Open

[https://nvbugs/6193836][test] Use EP=8 + attention DP for minimax_m2.5 8-GPU perf #14613

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c1bdd95
[https://nvbugs/6193836][test] Use EP=8 + attention DP for minimax_m2…
ruodil May 27, 2026
d2ba3a7
Merge branch 'main' into user/ruodil/fix-minimax-m2.5-tp8-divisibility
yufeiwu-nv May 27, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 17 additions & 0 deletions tests/integration/defs/perf/pytorch_model_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -251,6 +251,23 @@ def get_model_yaml_config(model_label: str,
'enable_attention_dp': True,
}
},
# MiniMax-M2.5 FP8 on 8 GPUs: intermediate_size=1536 with weight_block_size=128
# is not divisible under TP=8 (1536/8=192), so route MoE through EP=8 and use
# attention DP instead of TP.
{
'patterns': [
'minimax_m2.5_fp8-bench-pytorch-float8-input_output_len:128,128-ep:8-gpus:8',
'minimax_m2.5_fp8-bench-pytorch-float8-input_output_len:500,2000-ep:8-gpus:8',
'minimax_m2.5_fp8-bench-pytorch-float8-input_output_len:2000,500-ep:8-gpus:8',
'minimax_m2.5_fp8-bench-pytorch-float8-input_output_len:1000,1000-ep:8-gpus:8',
'minimax_m2.5_fp8-bench-pytorch-float8-input_output_len:1000,2000-ep:8-gpus:8',
'minimax_m2.5_fp8-bench-pytorch-float8-maxbs:1-input_output_len:1000,1000-reqs:10-con:1-ep:8-gpus:8',
'minimax_m2.5_fp8-bench-pytorch-float8-maxbs:512-input_output_len:1000,1000-con:512-ep:8-gpus:8',
],
'config': {
'enable_attention_dp': True,
}
},
{
'patterns': [
'qwen3_4b-bench-pytorch-streaming-bfloat16-maxbs:4-kv_frac:0.6-input_output_len:500,100-reqs:200-con:4',
Expand Down
16 changes: 8 additions & 8 deletions tests/integration/test_lists/qa/llm_perf_core.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -321,14 +321,14 @@ llm_perf_core:
#llama_v3.3_70b_instruct_fp8
#pytorch backend
- perf/test_perf.py::test_perf[llama_v3.3_70b_instruct_fp8-bench-pytorch-float8-maxbs:4096-maxnt:20000-kv_frac:0.85-input_output_len:20000,2000-reqs:1000-tp:8-gpus:8] TIMEOUT(120)
#minimax_m2.5 (FP8 216G, 8-GPU)
- perf/test_perf.py::test_perf[minimax_m2.5_fp8-bench-pytorch-float8-input_output_len:128,128-tp:8-gpus:8]
- perf/test_perf.py::test_perf[minimax_m2.5_fp8-bench-pytorch-float8-input_output_len:500,2000-tp:8-gpus:8]
- perf/test_perf.py::test_perf[minimax_m2.5_fp8-bench-pytorch-float8-input_output_len:2000,500-tp:8-gpus:8]
- perf/test_perf.py::test_perf[minimax_m2.5_fp8-bench-pytorch-float8-input_output_len:1000,1000-tp:8-gpus:8]
- perf/test_perf.py::test_perf[minimax_m2.5_fp8-bench-pytorch-float8-input_output_len:1000,2000-tp:8-gpus:8]
- perf/test_perf.py::test_perf[minimax_m2.5_fp8-bench-pytorch-float8-maxbs:1-input_output_len:1000,1000-reqs:10-con:1-tp:8-gpus:8] #min_latency
- perf/test_perf.py::test_perf[minimax_m2.5_fp8-bench-pytorch-float8-maxbs:512-input_output_len:1000,1000-con:512-tp:8-gpus:8] #max_throughput
#minimax_m2.5 (FP8 216G, 8-GPU) - use EP=8 + attention DP; TP=8 would split intermediate_size 1536/8=192 (not divisible by 128 block size)
- perf/test_perf.py::test_perf[minimax_m2.5_fp8-bench-pytorch-float8-input_output_len:128,128-ep:8-gpus:8]
- perf/test_perf.py::test_perf[minimax_m2.5_fp8-bench-pytorch-float8-input_output_len:500,2000-ep:8-gpus:8]
- perf/test_perf.py::test_perf[minimax_m2.5_fp8-bench-pytorch-float8-input_output_len:2000,500-ep:8-gpus:8]
- perf/test_perf.py::test_perf[minimax_m2.5_fp8-bench-pytorch-float8-input_output_len:1000,1000-ep:8-gpus:8]
- perf/test_perf.py::test_perf[minimax_m2.5_fp8-bench-pytorch-float8-input_output_len:1000,2000-ep:8-gpus:8]
- perf/test_perf.py::test_perf[minimax_m2.5_fp8-bench-pytorch-float8-maxbs:1-input_output_len:1000,1000-reqs:10-con:1-ep:8-gpus:8] #min_latency
- perf/test_perf.py::test_perf[minimax_m2.5_fp8-bench-pytorch-float8-maxbs:512-input_output_len:1000,1000-con:512-ep:8-gpus:8] #max_throughput
Comment thread
ruodil marked this conversation as resolved.
#llama_v3.1_nemotron_ultra_253b (nemotron-nas BF16 474G, 8-GPU)
- perf/test_perf.py::test_perf[llama_v3.1_nemotron_ultra_253b-bench-pytorch-bfloat16-input_output_len:128,128-tp:8-gpus:8]
- perf/test_perf.py::test_perf[llama_v3.1_nemotron_ultra_253b-bench-pytorch-bfloat16-input_output_len:500,2000-tp:8-gpus:8]
Expand Down
Loading