[XPU]fix: add XPU platform guards to DeepSeek-V4 ops (#42950)

Signed-off-by: Ma Jian <jian1.ma@intel.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
This commit is contained in:
Ma Jian
2026-05-23 06:29:45 +08:00
committed by GitHub
co-authored by Kunshang Ji
parent 4e2eba28be
commit 8de5cabeb7
6 changed files with 31 additions and 18 deletions
+3 -3
View File
@@ -169,9 +169,9 @@ class SiluAndMulWithClamp(CustomOp):
def __init__(self, swiglu_limit: float, *, compile_native: bool = True): def __init__(self, swiglu_limit: float, *, compile_native: bool = True):
super().__init__(compile_native=compile_native) super().__init__(compile_native=compile_native)
self.swiglu_limit = float(swiglu_limit) self.swiglu_limit = float(swiglu_limit)
if current_platform.is_rocm(): if current_platform.is_rocm() or current_platform.is_xpu():
self._forward_method = self.forward_native self._forward_method = self.forward_native
elif current_platform.is_cuda_alike() or current_platform.is_xpu(): elif current_platform.is_cuda_alike():
self.op = torch.ops._C.silu_and_mul_with_clamp self.op = torch.ops._C.silu_and_mul_with_clamp
elif current_platform.is_cpu(): elif current_platform.is_cpu():
self._forward_method = self.forward_native self._forward_method = self.forward_native
@@ -190,7 +190,7 @@ class SiluAndMulWithClamp(CustomOp):
return out return out
def forward_xpu(self, x: torch.Tensor) -> torch.Tensor: def forward_xpu(self, x: torch.Tensor) -> torch.Tensor:
return self.forward_cuda(x) return self.forward_native(x)
# --8<-- [start:mul_and_silu] # --8<-- [start:mul_and_silu]
@@ -243,7 +243,11 @@ def _fused_inv_rope_fp8_quant_kernel_impl(
(scale_inner * tma_aligned_T, 1, tma_aligned_T), (scale_inner * tma_aligned_T, 1, tma_aligned_T),
) )
grid = (tma_aligned_T, n_groups * heads_per_group) grid = (tma_aligned_T, n_groups * heads_per_group)
pdl_kwargs = {} if current_platform.is_rocm() else {"launch_pdl": False} pdl_kwargs = (
{}
if current_platform.is_rocm() or current_platform.is_xpu()
else {"launch_pdl": False}
)
_fused_inv_rope_fp8_quant_per_head[grid]( _fused_inv_rope_fp8_quant_per_head[grid](
o, o,
positions, positions,
+5 -1
View File
@@ -296,7 +296,11 @@ class DeepseekCompressor(nn.Module):
state_cache = self.state_cache.kv_cache state_cache = self.state_cache.kv_cache
# kv_state stored in first half, score_state stored in second half # kv_state stored in first half, score_state stored in second half
state_width = state_cache.shape[-1] // 2 state_width = state_cache.shape[-1] // 2
pdl_kwargs = {} if current_platform.is_rocm() else {"launch_pdl": False} pdl_kwargs = (
{}
if current_platform.is_rocm() or current_platform.is_xpu()
else {"launch_pdl": False}
)
# Store the KV and score (with fused APE addition) in the state. # Store the KV and score (with fused APE addition) in the state.
# NOTE: PDL is disabled — both this kernel and _fused_kernel below # NOTE: PDL is disabled — both this kernel and _fused_kernel below
+5 -5
View File
@@ -1153,7 +1153,7 @@ class DeepseekV4DecoderLayer(nn.Module):
x = self.ffn(x, input_ids) x = self.ffn(x, input_ids)
return x, residual, post_mix, res_mix return x, residual, post_mix, res_mix
def _forward_rocm( def _forward_native(
self, self,
x: torch.Tensor, x: torch.Tensor,
positions: torch.Tensor, positions: torch.Tensor,
@@ -1193,8 +1193,8 @@ class DeepseekV4DecoderLayer(nn.Module):
) -> tuple[ ) -> tuple[
torch.Tensor, torch.Tensor | None, torch.Tensor | None, torch.Tensor | None torch.Tensor, torch.Tensor | None, torch.Tensor | None, torch.Tensor | None
]: ]:
if current_platform.is_rocm(): if current_platform.is_rocm() or current_platform.is_xpu():
return self._forward_rocm( return self._forward_native(
x, positions, input_ids, post_mix, res_mix, residual x, positions, input_ids, post_mix, res_mix, residual
) )
@@ -1228,10 +1228,10 @@ class DeepseekV4Model(nn.Module):
# DeepseekV4MultiHeadLatentAttentionWrapper.attn_gemm_parallel_execute # DeepseekV4MultiHeadLatentAttentionWrapper.attn_gemm_parallel_execute
# (compressor kv_score, indexer.weights_proj, indexer.compressor # (compressor kv_score, indexer.weights_proj, indexer.compressor
# kv_score). fused_wqa_wkv stays on the default stream. # kv_score). fused_wqa_wkv stays on the default stream.
# Disable them on ROCm because of hang issues. # Disable them on ROCm / XPU because of hang issues / no overlap.
aux_stream_list = ( aux_stream_list = (
None None
if current_platform.is_rocm() if current_platform.is_rocm() or current_platform.is_xpu()
else [torch.cuda.Stream() for _ in range(3)] else [torch.cuda.Stream() for _ in range(3)]
) )
+5 -1
View File
@@ -367,7 +367,11 @@ class DeepseekSparseSWAMetadataBuilder(AttentionMetadataBuilder):
_LAYER_TYPE_C4A: None, _LAYER_TYPE_C4A: None,
_LAYER_TYPE_C128A: None, _LAYER_TYPE_C128A: None,
} }
if num_decode_tokens == 0 or current_platform.is_rocm(): if (
num_decode_tokens == 0
or current_platform.is_rocm()
or current_platform.is_xpu()
):
return out return out
for layer_type in self._layer_types: for layer_type in self._layer_types:
# get_mla_metadata() is the official FlashMLA entry point that # get_mla_metadata() is the official FlashMLA entry point that
@@ -468,12 +468,13 @@ def fp8_mqa_logits_torch(
seq_len_kv = k_fp8.shape[0] seq_len_kv = k_fp8.shape[0]
k = k_fp8.to(torch.bfloat16) k = k_fp8.to(torch.bfloat16)
q = q.to(torch.bfloat16) q = q.to(torch.bfloat16)
device = q.device
mask_lo = ( mask_lo = (
torch.arange(0, seq_len_kv, device="cuda")[None, :] >= cu_seqlen_ks[:, None] torch.arange(0, seq_len_kv, device=device)[None, :] >= cu_seqlen_ks[:, None]
) )
mask_hi = ( mask_hi = (
torch.arange(0, seq_len_kv, device="cuda")[None, :] < cu_seqlen_ke[:, None] torch.arange(0, seq_len_kv, device=device)[None, :] < cu_seqlen_ke[:, None]
) )
mask = mask_lo & mask_hi mask = mask_lo & mask_hi
@@ -1355,7 +1356,7 @@ def _rocm_sparse_attn_prefill_ragged_triton(
assert kv.ndim == 2, f"expected kv=[skv,d], got {kv.shape}" assert kv.ndim == 2, f"expected kv=[skv,d], got {kv.shape}"
assert indices.ndim == 1, f"expected indices=[nnz], got {indices.shape}" assert indices.ndim == 1, f"expected indices=[nnz], got {indices.shape}"
assert indptr.ndim == 1, f"expected indptr=[sq+1], got {indptr.shape}" assert indptr.ndim == 1, f"expected indptr=[sq+1], got {indptr.shape}"
assert q.is_cuda and kv.is_cuda and indices.is_cuda and indptr.is_cuda assert not q.is_cpu and not kv.is_cpu and not indices.is_cpu and not indptr.is_cpu
indices = _as_int32_contiguous_1d(indices) indices = _as_int32_contiguous_1d(indices)
indptr = _as_int32_contiguous_1d(indptr) indptr = _as_int32_contiguous_1d(indptr)
@@ -1459,10 +1460,10 @@ def _rocm_sparse_attn_decode_ragged_triton(
) )
assert main_indptr.ndim == 1, f"expected main_indptr=[b+1], got {main_indptr.shape}" assert main_indptr.ndim == 1, f"expected main_indptr=[b+1], got {main_indptr.shape}"
assert ( assert (
q.is_cuda not q.is_cpu
and main_cache.is_cuda and not main_cache.is_cpu
and main_indices.is_cuda and not main_indices.is_cpu
and main_indptr.is_cuda and not main_indptr.is_cpu
) )
main_indices = _as_int32_contiguous_1d(main_indices) main_indices = _as_int32_contiguous_1d(main_indices)