Skip to content

[Tests][LM Eval] Fix test seeding for consistent results #2395

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
dsikka merged 7 commits into from
Feb 25, 2026
Merged

[Tests][LM Eval] Fix test seeding for consistent results #2395

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
370bcdf
add seeding
dsikka Feb 23, 2026
22b29bc
try again
dsikka Feb 24, 2026
43fe84b
update
dsikka Feb 24, 2026
e461b88
Update flexible-extract metric value in YAML config
dsikka Feb 24, 2026
d278235
Merge branch 'main' into make_lm_eval_deterministic
dsikka Feb 24, 2026
8030d1b
update seeds
dsikka Feb 24, 2026
e49d94c
increase threshold
dsikka Feb 24, 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
5 changes: 3 additions & 2 deletions tests/lmeval/configs/w4a16_awq_sym.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,10 +4,11 @@ scheme: W4A16_awq_sym
recipe: tests/e2e/vLLM/recipes/WNA16/recipe_w4a16_awq_sym.yaml
dataset_id: HuggingFaceH4/ultrachat_200k
dataset_split: train_sft
seed: 100
lmeval:
recovery_threshold:
exact_match,strict-match: 0.91
exact_match,flexible-extract: 0.91
exact_match,strict-match: 0.92
exact_match,flexible-extract: 0.92
metrics:
exact_match,flexible-extract: 0.70
exact_match,strict-match: 0.70
1 change: 1 addition & 0 deletions tests/lmeval/configs/w4a4_nvfp4.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,6 +3,7 @@ model: meta-llama/Llama-3.1-8B-Instruct
scheme: NVFP4
dataset_id: HuggingFaceH4/ultrachat_200k
dataset_split: train_sft
seed: 100
num_calibration_samples: 20
lmeval:
# NVFP4 (4-bit weights + 4-bit activations) has lower recovery than FP8/INT8
Expand Down
21 changes: 16 additions & 5 deletions tests/lmeval/test_lmeval.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -100,12 +100,18 @@ def set_up(self, test_data_file: str):
self.recipe = eval_config.get("recipe")
self.quant_type = eval_config.get("quant_type")
self.save_dir = eval_config.get("save_dir")
self.seed = eval_config.get("seed", None)
self.seed = eval_config.get("seed", 42)

if self.seed is not None:
random.seed(self.seed)
numpy.random.seed(self.seed)
torch.manual_seed(self.seed)
random.seed(self.seed)
numpy.random.seed(self.seed)
torch.manual_seed(self.seed)

if torch.cuda.is_available():
torch.cuda.manual_seed_all(self.seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False

logger.info(f"Seed set to {self.seed} with deterministic mode enabled")

logger.info("========== RUNNING ==============")
logger.info(self.scheme)
Expand Down Expand Up @@ -187,6 +193,11 @@ def _eval_model(self, model: str) -> dict:
limit=self.lmeval.limit,
apply_chat_template=self.lmeval.apply_chat_template,
batch_size=self.lmeval.batch_size,
# Pass seeds to lm_eval for deterministic evaluation
random_seed=self.seed,
numpy_random_seed=self.seed,
torch_random_seed=self.seed,
fewshot_random_seed=self.seed,
)

return results
Expand Down
Loading