Skip to content

[https://nvbugs/5440241][fix] Fix 70B GSM8K Accuracy drop #7075

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
Aug 20, 2025
Merged

[https://nvbugs/5440241][fix] Fix 70B GSM8K Accuracy drop #7075

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
6 changes: 4 additions & 2 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 @@ -13,10 +13,12 @@ meta-llama/Llama-3.3-70B-Instruct:
- accuracy: 83.78
- quant_algo: NVFP4
kv_cache_quant_algo: FP8
accuracy: 88.70
accuracy: 87.33
- quant_algo: FP8
kv_cache_quant_algo: FP8
accuracy: 84.08
accuracy: 90.30
- quant_algo: FP8
accuracy: 90.30
meta-llama/Llama-4-Maverick-17B-128E-Instruct:
- accuracy: 92.20
- quant_algo: FP8
Expand Down
6 changes: 4 additions & 2 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 @@ -64,10 +64,12 @@ meta-llama/Llama-3.3-70B-Instruct:
accuracy: 81.31
- quant_algo: NVFP4
kv_cache_quant_algo: FP8
accuracy: 79.31
accuracy: 78.78
- quant_algo: FP8
kv_cache_quant_algo: FP8
accuracy: 81.02
accuracy: 80.40
- quant_algo: FP8
accuracy: 80.40
meta-llama/Llama-4-Maverick-17B-128E-Instruct:
- accuracy: 86.40
- quant_algo: FP8
Expand Down
14 changes: 9 additions & 5 deletions tests/integration/defs/accuracy/test_llm_api_pytorch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -408,7 +408,7 @@ def test_fp8_eagle3_tp8(self, eagle3_one_model):
@pytest.mark.skip_less_device(4)
@skip_pre_hopper
def test_fp8_tp4(self):
model_path = f"{llm_models_root()}/modelopt-hf-model-hub/Llama-3.3-70B-Instruct-fp8"
model_path = f"{llm_models_root()}/llama-3.3-models/Llama-3.3-70B-Instruct-FP8"
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.5)
with LLM(model_path,
tensor_parallel_size=4,
Expand All @@ -417,6 +417,7 @@ def test_fp8_tp4(self):
kv_cache_config=kv_cache_config) as llm:
assert llm.args.quant_config.quant_algo == QuantAlgo.FP8
sampling_params = SamplingParams(
max_tokens=256,
temperature=0.0,
add_special_tokens=False,
)
Expand All @@ -426,16 +427,20 @@ def test_fp8_tp4(self):
task.evaluate(llm, sampling_params=sampling_params)
task = GPQADiamond(self.MODEL_NAME)
task.evaluate(llm,
sampling_params=sampling_params,
extra_evaluator_kwargs=dict(apply_chat_template=True))

@pytest.mark.skip_less_device(4)
@skip_pre_blackwell
def test_nvfp4_tp4(self):
model_path = f"{llm_models_root()}/modelopt-hf-model-hub/Llama-3.3-70B-Instruct-fp4"
with LLM(model_path, tensor_parallel_size=4) as llm:
model_path = f"{llm_models_root()}/llama-3.3-models/Llama-3.3-70B-Instruct-FP4"
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.5)
with LLM(model_path,
tensor_parallel_size=4,
max_batch_size=32,
kv_cache_config=kv_cache_config) as llm:
assert llm.args.quant_config.quant_algo == QuantAlgo.NVFP4
sampling_params = SamplingParams(
max_tokens=256,
temperature=0.0,
add_special_tokens=False,
)
Expand All @@ -445,7 +450,6 @@ def test_nvfp4_tp4(self):
task.evaluate(llm, sampling_params=sampling_params)
task = GPQADiamond(self.MODEL_NAME)
task.evaluate(llm,
sampling_params=sampling_params,
extra_evaluator_kwargs=dict(apply_chat_template=True))


Expand Down