[Refactor] Remove dead cuda kernels (#42767)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye
2026-05-18 11:14:21 -07:00
committed by GitHub
parent 9758a6e5c5
commit 00e20e76f7
6 changed files with 1 additions and 605 deletions
+1 -143
View File
@@ -280,123 +280,6 @@ def merge_attn_states(
)
def convert_vertical_slash_indexes(
q_seqlens: torch.Tensor, # [BATCH, ]
kv_seqlens: torch.Tensor, # [BATCH, ]
vertical_indexes: torch.Tensor, # [BATCH, N_HEADS, NNZ_V]
slash_indexes: torch.Tensor, # [BATCH, N_HEADS, NNZ_S]
context_size: int,
block_size_M: int,
block_size_N: int,
causal: bool = True,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
batch_size = slash_indexes.size(0)
num_heads = slash_indexes.size(1)
nnz_slash = slash_indexes.size(2)
nnz_vertical = vertical_indexes.size(2)
num_rows = (context_size + block_size_M - 1) // block_size_M
block_count = torch.zeros(
batch_size, num_heads, num_rows, dtype=q_seqlens.dtype, device=q_seqlens.device
)
block_offset = torch.zeros(
batch_size,
num_heads,
num_rows,
nnz_slash,
dtype=q_seqlens.dtype,
device=q_seqlens.device,
)
column_count = torch.zeros(
batch_size, num_heads, num_rows, dtype=q_seqlens.dtype, device=q_seqlens.device
)
column_index = torch.zeros(
batch_size,
num_heads,
num_rows,
nnz_vertical,
dtype=q_seqlens.dtype,
device=q_seqlens.device,
)
torch.ops._C.convert_vertical_slash_indexes(
block_count,
block_offset,
column_count,
column_index,
q_seqlens,
kv_seqlens,
vertical_indexes,
slash_indexes,
context_size,
block_size_M,
block_size_N,
causal,
)
return block_count, block_offset, column_count, column_index
def convert_vertical_slash_indexes_mergehead(
q_seqlens: torch.Tensor, # [BATCH, ]
kv_seqlens: torch.Tensor, # [BATCH, ]
vertical_indexes: torch.Tensor, # [BATCH, N_HEADS, NNZ_V]
slash_indexes: torch.Tensor, # [BATCH, N_HEADS, NNZ_S]
# [N_HEADS] : different head use different number of indices
vertical_indices_count: torch.Tensor,
slash_indices_count: torch.Tensor,
context_size: int,
block_size_M: int,
block_size_N: int,
causal: bool = True,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
batch_size = slash_indexes.size(0)
num_heads = slash_indexes.size(1)
nnz_slash = slash_indexes.size(2)
nnz_vertical = vertical_indexes.size(2)
num_rows = (context_size + block_size_M - 1) // block_size_M
block_count = torch.empty(
batch_size, num_heads, num_rows, dtype=q_seqlens.dtype, device=q_seqlens.device
)
block_offset = torch.empty(
batch_size,
num_heads,
num_rows,
nnz_slash,
dtype=q_seqlens.dtype,
device=q_seqlens.device,
)
column_count = torch.empty(
batch_size, num_heads, num_rows, dtype=q_seqlens.dtype, device=q_seqlens.device
)
column_index = torch.empty(
batch_size,
num_heads,
num_rows,
nnz_vertical,
dtype=q_seqlens.dtype,
device=q_seqlens.device,
)
torch.ops._C.convert_vertical_slash_indexes_mergehead(
block_count,
block_offset,
column_count,
column_index,
q_seqlens,
kv_seqlens,
vertical_indexes,
slash_indexes,
vertical_indices_count,
slash_indices_count,
context_size,
block_size_M,
block_size_N,
causal,
)
return block_count, block_offset, column_count, column_index
# pos encoding ops
def rotary_embedding(
positions: torch.Tensor,
@@ -2658,32 +2541,7 @@ def moe_wna16_marlin_gemm(
)
if hasattr(torch.ops, "_moe_C") and hasattr(torch.ops._moe_C, "marlin_gemm_moe"):
@register_fake("_moe_C::marlin_gemm_moe")
def marlin_gemm_moe_fake(
a: torch.Tensor,
b_q_weights: torch.Tensor,
sorted_ids: torch.Tensor,
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
b_scales: torch.Tensor,
b_zero_points: torch.Tensor,
g_idx: torch.Tensor,
perm: torch.Tensor,
workspace: torch.Tensor,
b_q_type: ScalarType,
size_m: torch.SymInt,
size_n: torch.SymInt,
size_k: torch.SymInt,
is_k_full: bool,
num_experts: int,
topk: int,
moe_block_size: int,
replicate_input: bool,
apply_weights: bool,
) -> torch.Tensor:
return torch.empty((size_m, topk, size_n), dtype=a.dtype, device=a.device)
if hasattr(torch.ops, "_moe_C") and hasattr(torch.ops._moe_C, "moe_wna16_marlin_gemm"):
@register_fake("_moe_C::moe_wna16_marlin_gemm")
def moe_wna16_marlin_gemm_fake(