[Metrics] Fix fwd_occupancy reading NaN on every decode log line; probe-free base-a (#27836)

This commit is contained in:
Liangsheng Yin
2026-06-10 15:42:19 -07:00
committed by GitHub
parent bdf3ef6421
commit 16124fc9b2
4 changed files with 20 additions and 14 deletions
+3 -2
View File
@@ -34,8 +34,9 @@ class TestBasicSanity(
):
served_model_name = DEFAULT_MODEL_NAME_FOR_TEST
# 5090 + Llama-3.1-8B single-batch decode with overlap scheduler +
# cuda graph measured ~99 median in CI; keep ~2pp headroom.
fwd_occupancy_threshold = 97.0
# cuda graph measured ~99 median in CI; async-assert probes are off in
# base-a, so the threshold can sit right under the measured median.
fwd_occupancy_threshold = 99.0
@classmethod
def setUpClass(cls):
@@ -33,10 +33,11 @@ class TestBasicSanityEagle3(
CustomTestCase,
):
served_model_name = DEFAULT_TARGET_MODEL_EAGLE3
# CUDA 5090 + Llama-3.1-8B measured ~99 median in CI. AMD EAGLE3
# currently sustains lower single-batch occupancy and needs a longer
# measurement window to avoid too few non-NaN samples.
fwd_occupancy_threshold = 80.0 if is_in_amd_ci() else 97.0
# CUDA 5090 + Llama-3.1-8B measured ~99 median in CI with async-assert
# probes off in base-a. AMD EAGLE3 currently sustains lower single-batch
# occupancy and needs a longer measurement window to avoid too few
# non-NaN samples.
fwd_occupancy_threshold = 80.0 if is_in_amd_ci() else 98.0
fwd_occupancy_max_new_tokens = 4096 if is_in_amd_ci() else 2048
fwd_occupancy_acc_length_threshold: float = 1.6