[XPU] Qwen3.8-flash-next enablement (#37213)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
This commit is contained in:
Xia Weiwen
2026-09-24 11:06:30 +08:00
committed by GitHub
co-authored by Copilot
parent f4d9d2e670
commit 5c44214d6a
6 changed files with 283 additions and 10 deletions
@@ -250,9 +250,9 @@ def fused_mamba_state_scatter_with_mask(
raise ValueError(
f"dst and src must be on the same device. {dst.device=} {src.device=}"
)
if not dst.is_cuda or not src.is_cuda:
if not (dst.is_cuda or dst.is_xpu) or dst.device != src.device:
raise ValueError(
"fused_mamba_state_scatter_with_mask only supports CUDA tensors."
"fused_mamba_state_scatter_with_mask only supports CUDA/XPU tensors."
)
if dst.ndim < 2 or src.ndim < 3:
raise ValueError(f"Unexpected tensor ranks: {dst.ndim=} {src.ndim=}")
@@ -410,10 +410,10 @@ def fused_conv_window_scatter_with_mask(
if total_requests == 0:
return
if not (dst.is_cuda and src.is_cuda and dst.device == src.device):
if not (dst.is_cuda or dst.is_xpu) or dst.device != src.device:
raise ValueError(
"fused_conv_window_scatter_with_mask requires dst and src to be CUDA "
f"tensors on the same device ({dst.device=}, {src.device=})."
"fused_conv_window_scatter_with_mask requires dst and src to be "
f"CUDA/XPU tensors on the same device ({dst.device=}, {src.device=})."
)
if dst.ndim != 4 or src.ndim != 5:
raise ValueError(f"Unexpected ranks: {dst.ndim=} (want 4) {src.ndim=} (want 5)")
@@ -95,6 +95,11 @@ def fused_sigmoid_gating_delta_rule_update(
o=o,
h0_source=initial_state_source,
h0_indices=initial_state_indices,
# Envelope-strided state pools (page-major / unified memory) have a
# per-slot pitch != HV*K*V; contiguous pools pass exactly HV*K*V.
stride_h0_source=(
initial_state_source.stride(0) if initial_state_source is not None else 0
),
cu_seqlens=cu_seqlens,
intermediate_states_buffer=intermediate_states_buffer,
intermediate_state_indices=intermediate_state_indices,
@@ -519,13 +519,16 @@ class QSAIndexer(MultiPlatformOp):
token_topk=self.token_topk,
)
def forward_cuda(
def _forward_impl(
self,
hidden_states: torch.Tensor,
positions: torch.Tensor,
forward_batch,
indexer_metadata,
) -> torch.Tensor:
"""Portable orchestration shared by different platforms.
Fast paths are gated per platforms inside kernel calls.
"""
forward_mode = forward_batch.forward_mode
is_target_verify = getattr(forward_mode, "is_target_verify", lambda: False)()
is_draft_extend = getattr(forward_mode, "is_draft_extend_v2", lambda: False)()
@@ -628,6 +631,28 @@ class QSAIndexer(MultiPlatformOp):
row_sequence_lengths,
)
def forward_cuda(
self,
hidden_states: torch.Tensor,
positions: torch.Tensor,
forward_batch,
indexer_metadata,
) -> torch.Tensor:
return self._forward_impl(
hidden_states, positions, forward_batch, indexer_metadata
)
def forward_xpu(
self,
hidden_states: torch.Tensor,
positions: torch.Tensor,
forward_batch,
indexer_metadata,
) -> torch.Tensor:
return self._forward_impl(
hidden_states, positions, forward_batch, indexer_metadata
)
__all__ = [
"QSAIndexer",
+3 -3
View File
@@ -139,14 +139,14 @@ class GatedResidual(HyperConnectionBase):
self.hidden_size * self.hc_count,
self.config.hc_lowrank,
bias=False,
device=torch.cuda.current_device(),
device=torch.get_device_module().current_device(),
dtype=config.params_dtype,
)
self.input_mix_weight_up = nn.Linear(
self.config.hc_lowrank,
self.hc_count * self.hidden_size,
bias=False,
device=torch.cuda.current_device(),
device=torch.get_device_module().current_device(),
dtype=config.params_dtype,
)
lowrank = self.config.hc_lowrank
@@ -166,7 +166,7 @@ class GatedResidual(HyperConnectionBase):
self.hidden_size * self.hc_count,
self.hc_count,
bias=False,
device=torch.cuda.current_device(),
device=torch.get_device_module().current_device(),
dtype=config.params_dtype,
)
# hc_combine rejects other shapes; device and dtype are checked per call.