[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:
co-authored by
Jiangyun Zhu
parent
077a9a8e37
commit
b779eb3363
@@ -16,7 +16,7 @@ from vllm.triton_utils import tl, triton
|
||||
|
||||
from .index import prepare_chunk_indices
|
||||
from .op import exp
|
||||
from .utils import FLA_GDN_FIX_BT, check_shared_mem, is_nvidia_hopper
|
||||
from .utils import FLA_CHUNK_SIZE, check_shared_mem, is_nvidia_hopper
|
||||
|
||||
BKV_LIST = [64, 128] if check_shared_mem() else [32, 64]
|
||||
NUM_WARPS = [2, 4] if is_nvidia_hopper else [2, 4, 8]
|
||||
@@ -146,11 +146,11 @@ def chunk_fwd_o(
|
||||
g: torch.Tensor | None = None, # cumsum of log decay
|
||||
scale: float | None = None,
|
||||
cu_seqlens: torch.Tensor | None = None,
|
||||
chunk_size: int = 64,
|
||||
chunk_size: int = FLA_CHUNK_SIZE,
|
||||
) -> torch.Tensor:
|
||||
B, T, Hg, K, V = *q.shape, v.shape[-1]
|
||||
H = v.shape[-2]
|
||||
BT = 64 if FLA_GDN_FIX_BT else min(chunk_size, max(16, triton.next_power_of_2(T)))
|
||||
BT = chunk_size
|
||||
chunk_indices = (
|
||||
prepare_chunk_indices(cu_seqlens, BT) if cu_seqlens is not None else None
|
||||
)
|
||||
|
||||
@@ -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]:
|
||||
"""
|
||||
|
||||
Reference in New Issue
Block a user