Skip to content

[#9316][feat] AutoDeploy: Add the accuracy test for Nemotron MOE models #9317

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

Merged
merged 1 commit into from
Nov 20, 2025
Merged

[#9316][feat] AutoDeploy: Add the accuracy test for Nemotron MOE models #9317

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
5 changes: 5 additions & 0 deletions tests/integration/defs/accuracy/references/gsm8k.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -161,6 +161,11 @@ nvidia/Nemotron-H-56B-Base-8K:
- quant_algo: FP8
kv_cache_quant_algo: FP8
accuracy: 89.27
nvidia/Nemotron-MOE:
- accuracy: 88.249
- quant_algo: FP8
kv_cache_quant_algo: FP8
accuracy: 86.884
nvidia/Llama-3.1-Nemotron-Nano-8B-v1:
- accuracy: 37.15
- quant_algo: FP8
Expand Down
5 changes: 5 additions & 0 deletions tests/integration/defs/accuracy/references/mmlu.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -270,6 +270,11 @@ nvidia/Nemotron-H-56B-Base-8K:
- quant_algo: FP8
kv_cache_quant_algo: FP8
accuracy: 83.82
nvidia/Nemotron-MOE:
- accuracy: 77.802
- quant_algo: FP8
kv_cache_quant_algo: FP8
accuracy: 73.879
microsoft/Phi-4-mini-instruct:
- accuracy: 68.98
- quant_algo: FP8
Expand Down
27 changes: 22 additions & 5 deletions tests/integration/defs/accuracy/test_llm_api_autodeploy.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,6 +18,7 @@
import pytest

from tensorrt_llm._torch.auto_deploy import LLM as AutoDeployLLM
from tensorrt_llm.quantization import QuantAlgo
from tensorrt_llm.sampling_params import SamplingParams

from ..conftest import llm_models_root
Expand Down Expand Up @@ -153,7 +154,8 @@ def test_auto_dtype(self, enable_chunked_prefill):

class TestNemotronMOE(LlmapiAccuracyTestHarness):
MODEL_NAME = "nvidia/Nemotron-MOE"
MODEL_PATH = f"{llm_models_root()}/Nemotron-MOE/"
MODEL_PATH_BF16 = f"{llm_models_root()}/Nemotron-Nano-3-30B-A3.5B-dev-1024"
MODEL_PATH_FP8 = f"{llm_models_root()}/Nemotron-Nano-3-30B-A3.5B-FP8-KVFP8-dev"

def get_default_kwargs(self):
return {
Expand Down Expand Up @@ -196,13 +198,28 @@ def get_default_sampling_params(self):
use_beam_search=beam_width > 1)

@pytest.mark.skip_less_device_memory(32000)
def test_auto_dtype(self):
pytest.skip("Nemotron-MOE is not in CI yet")
def test_bf16(self):
kwargs = self.get_default_kwargs()
sampling_params = self.get_default_sampling_params()
with AutoDeployLLM(model=self.MODEL_PATH,
tokenizer=self.MODEL_PATH,
with AutoDeployLLM(model=self.MODEL_PATH_BF16,
tokenizer=self.MODEL_PATH_BF16,
**kwargs) as llm:
task = MMLU(self.MODEL_NAME)
task.evaluate(llm, sampling_params=sampling_params)
task = GSM8K(self.MODEL_NAME)
task.evaluate(llm)

@pytest.mark.skip_less_device_memory(32000)
def test_fp8(self):
kwargs = self.get_default_kwargs()
sampling_params = self.get_default_sampling_params()
with AutoDeployLLM(model=self.MODEL_PATH_FP8,
tokenizer=self.MODEL_PATH_FP8,
**kwargs) as llm:
# Manually set quant_config for FP8 model to get the accuracy threshold
llm.args.quant_config.quant_algo = QuantAlgo.FP8
llm.args.quant_config.kv_cache_quant_algo = QuantAlgo.FP8

task = MMLU(self.MODEL_NAME)
task.evaluate(llm, sampling_params=sampling_params)
task = GSM8K(self.MODEL_NAME)
Expand Down