[Refactor] Remove dead kernel code (#46405)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
@@ -2962,70 +2962,6 @@ def qr_max_size() -> int:
|
||||
return torch.ops._C_custom_ar.qr_max_size()
|
||||
|
||||
|
||||
def get_flash_mla_metadata(
|
||||
cache_seqlens: torch.Tensor,
|
||||
num_heads_per_head_k: int,
|
||||
num_heads_k: int,
|
||||
) -> tuple[torch.Tensor, torch.Tensor]:
|
||||
"""
|
||||
Arguments:
|
||||
cache_seqlens: (batch_size), dtype torch.int32.
|
||||
num_heads_per_head_k: Equals to seq_len_q * num_heads_q // num_heads_k.
|
||||
num_heads_k: num_heads_k.
|
||||
|
||||
Return:
|
||||
tile_scheduler_metadata: (num_sm_parts, TileSchedulerMetaDataSize), dtype torch.int32.
|
||||
num_splits: (batch_size + 1), dtype torch.int32.
|
||||
"""
|
||||
return torch.ops._C.get_flash_mla_metadata(
|
||||
cache_seqlens, num_heads_per_head_k, num_heads_k
|
||||
)
|
||||
|
||||
|
||||
def flash_mla_with_kvcache(
|
||||
q: torch.Tensor,
|
||||
k_cache: torch.Tensor,
|
||||
block_table: torch.Tensor,
|
||||
cache_seqlens: torch.Tensor,
|
||||
head_dim_v: int,
|
||||
tile_scheduler_metadata: torch.Tensor,
|
||||
num_splits: torch.Tensor,
|
||||
softmax_scale: float | None = None,
|
||||
causal: bool = False,
|
||||
) -> tuple[torch.Tensor, torch.Tensor]:
|
||||
"""
|
||||
Arguments:
|
||||
q: (batch_size, seq_len_q, num_heads_q, head_dim).
|
||||
k_cache: (num_blocks, page_block_size, num_heads_k, head_dim).
|
||||
block_table: (batch_size, max_num_blocks_per_seq), torch.int32.
|
||||
cache_seqlens: (batch_size), torch.int32.
|
||||
head_dim_v: Head_dim of v.
|
||||
tile_scheduler_metadata: (num_sm_parts, TileSchedulerMetaDataSize), torch.int32, return by get_mla_metadata.
|
||||
num_splits: (batch_size + 1), torch.int32, return by get_mla_metadata.
|
||||
softmax_scale: float. The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim).
|
||||
causal: bool. Whether to apply causal attention mask.
|
||||
|
||||
Return:
|
||||
out: (batch_size, seq_len_q, num_heads_q, head_dim_v).
|
||||
softmax_lse: (batch_size, num_heads_q, seq_len_q), torch.float32.
|
||||
"""
|
||||
if softmax_scale is None:
|
||||
softmax_scale = q.shape[-1] ** (-0.5)
|
||||
out, softmax_lse = torch.ops._C.flash_mla_fwd_kvcache(
|
||||
q,
|
||||
k_cache,
|
||||
None,
|
||||
head_dim_v,
|
||||
cache_seqlens,
|
||||
block_table,
|
||||
softmax_scale,
|
||||
causal,
|
||||
tile_scheduler_metadata,
|
||||
num_splits,
|
||||
)
|
||||
return out, softmax_lse
|
||||
|
||||
|
||||
def sm100_cutlass_mla_decode(
|
||||
out: torch.Tensor,
|
||||
lse: torch.Tensor,
|
||||
|
||||
Reference in New Issue
Block a user