Relax mamba unified cache kl threshold (#27070)
This commit is contained in:
@@ -30,7 +30,8 @@ MAMBA_TRACK_INTERVAL = 128
|
|||||||
class TestUnifiedMambaRadixCache(UnifiedRadixTreeTestMixin, CustomTestCase):
|
class TestUnifiedMambaRadixCache(UnifiedRadixTreeTestMixin, CustomTestCase):
|
||||||
"""Mamba hybrid + UnifiedRadixCache."""
|
"""Mamba hybrid + UnifiedRadixCache."""
|
||||||
|
|
||||||
kl_threshold = 0.003
|
# fp8 cache-hit KL is noisy and flaked at 0.003, match the HiCache variant
|
||||||
|
kl_threshold = 0.005
|
||||||
prefill_cache_assert = staticmethod(
|
prefill_cache_assert = staticmethod(
|
||||||
make_mamba_prefill_assert(chunk_size=MAMBA_CHUNK_SIZE)
|
make_mamba_prefill_assert(chunk_size=MAMBA_CHUNK_SIZE)
|
||||||
)
|
)
|
||||||
|
|||||||
Reference in New Issue
Block a user