Skip to content

[CI] Stabilize Hybrid SSM disagg PD gsm8k accuracy test (#43301) #43570

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

Closed
haosdent wants to merge 1 commit into from
+17 −7
Closed

[CI] Stabilize Hybrid SSM disagg PD gsm8k accuracy test (#43301) #43570

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
24 changes: 17 additions & 7 deletions tests/v1/kv_connector/nixl_integration/test_accuracy.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,10 +9,11 @@
NUM_CONCURRENT = 100
TASK = "gsm8k"
FILTER = "exact_match,strict-match"
# TODO(#43186): Widened from 0.03 to absorb chunk_scan/SSU numeric jitter
# on granite-4.0-h-tiny under NIXL PD; tighten when the kernel divergence
# is fixed.
RTOL = 0.05
RTOL = 0.03
# Larger sample averages out the 1-2 prompt-level argmax flips that
# bf16 cuBLAS ULP-tie produces under different TP shapes; with 200
# prompts the per-config variance drops to ~0.02 (within RTOL).
GSM8K_LIMIT = 200

# Model-specific expected values
EXPECTED_VALUES = {
Expand All @@ -22,8 +23,8 @@
"deepseek-ai/DeepSeek-V2-Lite-Chat": 0.65,
"google/gemma-3-4b-it": 0.74,
"nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8": 0.84,
"ibm-granite/granite-4.0-h-tiny": 0.77,
"Qwen/Qwen3.5-0.8B": 0.33,
"ibm-granite/granite-4.0-h-tiny": 0.80,
"Qwen/Qwen3.5-0.8B": 0.36,
}

SIMPLE_PROMPT = (
Expand Down Expand Up @@ -59,17 +60,26 @@ def test_accuracy():
model="local-completions",
model_args=model_args,
tasks=TASK,
limit=GSM8K_LIMIT,
# Force greedy so chosen-token is argmax of logits; under random
# sampling, gsm8k has ~0.05 variance from temperature noise that
# masks measurement of small numerical differences across runs.
gen_kwargs="temperature=0,do_sample=False,top_k=1",
)

measured_value = results["results"][TASK][FILTER]
expected_value = EXPECTED_VALUES.get(MODEL_NAME)

# Always print so the per-config measure is in CI logs even when
# the assertion passes; lets us trend-track drift instead of only
# seeing a number when something breaks.
print(f"Measured value: {measured_value} (expected {expected_value})")

if expected_value is None:
print(
f"Warning: No expected value found for {MODEL_NAME}. "
"Skipping accuracy check."
)
print(f"Measured value: {measured_value}")
return

assert (
Expand Down
Loading