Revert "[DSpark] Support pipeline-parallel targets in aggregated serving (#56956)" (#58484)

This commit is contained in:
Nick Hill
2026-09-24 08:48:36 -07:00
committed by GitHub
parent d636456b1f
commit 09fe178dba
11 changed files with 28 additions and 629 deletions
@@ -41,7 +41,6 @@ if current_platform.is_cuda():
def _dsv4_topk_kernel(
gating_output_ptr,
correction_bias_ptr,
is_padding_ptr,
topk_weights_ptr,
topk_ids_ptr,
routed_scaling_factor,
@@ -50,7 +49,6 @@ if current_platform.is_cuda():
image_sentinel_lo,
NUM_EXPERTS: tl.constexpr,
BLOCK_N: tl.constexpr,
HAS_PADDING: tl.constexpr,
HAS_VL: tl.constexpr,
launch_pdl: tl.constexpr,
):
@@ -108,11 +106,6 @@ if current_platform.is_cuda():
output_mask = topk_offsets < 6
output_offsets = row * 6 + topk_offsets
if HAS_PADDING:
is_padding = tl.load(is_padding_ptr + row)
selected_weights = tl.where(is_padding, 0.0, selected_weights)
selected_ids = tl.where(is_padding, -1, selected_ids)
if launch_pdl:
tl.extra.cuda.gdc_launch_dependents()
@@ -125,20 +118,11 @@ def dsv4_topk(
correction_bias: torch.Tensor,
indices_dtype: torch.dtype,
routed_scaling_factor: float,
is_padding: torch.Tensor | None = None,
input_ids: torch.Tensor | None = None,
bias_vl: torch.Tensor | None = None,
image_sentinel_lo: int = 0,
) -> tuple[torch.Tensor, torch.Tensor]:
num_tokens, num_experts = gating_output.shape
if is_padding is not None:
assert is_padding.dtype == torch.bool
assert is_padding.shape == (num_tokens,)
assert is_padding.device == gating_output.device
assert is_padding.is_contiguous()
assert indices_dtype in (torch.int32, torch.int64), (
"Padding requires a signed indices dtype for the -1 sentinel."
)
has_vl = bias_vl is not None and image_sentinel_lo > 0
if bias_vl is not None:
assert input_ids is not None, "bias_vl routing requires input_ids"
@@ -152,7 +136,6 @@ def dsv4_topk(
_dsv4_topk_kernel[(num_tokens,)](
gating_output,
correction_bias,
is_padding,
topk_weights,
topk_ids,
routed_scaling_factor,
@@ -161,7 +144,6 @@ def dsv4_topk(
image_sentinel_lo,
NUM_EXPERTS=num_experts,
BLOCK_N=triton.next_power_of_2(num_experts),
HAS_PADDING=is_padding is not None,
HAS_VL=has_vl,
num_warps=1,
launch_pdl=current_platform.is_arch_support_pdl(),
@@ -148,13 +148,9 @@ def fused_topk_bias(
)
output_indices_dtype = torch.int32 if indices_type is None else indices_type
padding_mask = _get_padding_mask(gating_output.shape[0])
if (
scoring_func == "sqrtsoftplus"
and hash_indices_table is None
# dsv4_topk marks padded rows with a -1 sentinel, which unsigned
# indices cannot represent; keep the generic path for that pair.
and (padding_mask is None or output_indices_dtype != torch.uint32)
and can_use_dsv4_topk(
gating_output,
e_score_correction_bias,
@@ -169,7 +165,6 @@ def fused_topk_bias(
e_score_correction_bias,
output_indices_dtype,
routed_scaling_factor,
is_padding=padding_mask,
input_ids=input_tokens,
bias_vl=bias_vl,
image_sentinel_lo=image_sentinel_lo,