[Model] Sync upstream BT=chunk_size fix for GDN chunk_fwd_kernel_o, simplify warmup to single pass (#38343)

Signed-off-by: AuYang <459461160@qq.com>
Co-authored-by: Jiangyun Zhu <riverclouds.zhu@qq.com>
This commit is contained in:
Xu Jinyang
2026-03-31 23:03:24 +04:00
committed by GitHub
co-authored by Jiangyun Zhu
parent 077a9a8e37
commit b779eb3363
3 changed files with 61 additions and 68 deletions
+3 -1
View File
@@ -24,10 +24,12 @@ logger = logging.getLogger(__name__)
COMPILER_MODE = os.getenv("FLA_COMPILER_MODE") == "1"
FLA_CI_ENV = os.getenv("FLA_CI_ENV") == "1"
FLA_GDN_FIX_BT = os.getenv("FLA_GDN_FIX_BT", "0") == "1"
SUPPRESS_LEVEL = int(os.getenv("GDN_RECOMPUTE_SUPPRESS_LEVEL", "0"))
# Default chunk size used across FLA triton kernels (kda, chunk, chunk_o, etc.)
FLA_CHUNK_SIZE = 64
def tensor_cache(fn: Callable[..., torch.Tensor]) -> Callable[..., torch.Tensor]:
"""