feat: add DSV4 Hygon ops

include/infinicore/ops.hpp

@@ -25,6 +25,17 @@
 #include "ops/conv2d.hpp"
 #include "ops/cross_entropy.hpp"
 #include "ops/deepseek_moe.hpp"
+#include "ops/dsv4_act_quant_fp8.hpp"
+#include "ops/dsv4_add_rmsnorm_quant.hpp"
+#include "ops/dsv4_fused_rope.hpp"
+#include "ops/dsv4_linear_bf16_fp32.hpp"
+#include "ops/dsv4_mask_topk_ids.hpp"
+#include "ops/dsv4_per_token_quant_int8.hpp"
+#include "ops/dsv4_rmsnorm_self.hpp"
+#include "ops/dsv4_silu_and_mul.hpp"
+#include "ops/dsv4_silu_mul_quant.hpp"
+#include "ops/dsv4_swa_prefill_indices.hpp"
+#include "ops/dsv4_topk_transform.hpp"
 #include "ops/embedding.hpp"
 #include "ops/flash_attention.hpp"
 #include "ops/fmin.hpp"

include/infinicore/ops/dsv4_act_quant_fp8.hpp

@@ -0,0 +1,9 @@
+#pragma once
+#include "../device.hpp"
+#include "../graph/graph.hpp"
+#include "../tensor.hpp"
+#include "common/op.hpp"
+namespace infinicore::op {
+INFINICORE_GRAPH_OP_CLASS(Dsv4ActQuantFp8, Tensor, Tensor, const Tensor &, float);
+void dsv4_act_quant_fp8_(Tensor xq, Tensor scale, const Tensor &x, float fp8_max);
+} // namespace infinicore::op

include/infinicore/ops/dsv4_add_rmsnorm_quant.hpp

@@ -0,0 +1,9 @@
+#pragma once
+#include "../device.hpp"
+#include "../graph/graph.hpp"
+#include "../tensor.hpp"
+#include "common/op.hpp"
+namespace infinicore::op {
+INFINICORE_GRAPH_OP_CLASS(Dsv4AddRMSNormQuant, Tensor, Tensor, Tensor, const Tensor &, const Tensor &, float);
+void dsv4_add_rmsnorm_quant_(Tensor res, Tensor q, Tensor scale, const Tensor &x, const Tensor &weight, float epsilon);
+} // namespace infinicore::op

include/infinicore/ops/dsv4_fused_rope.hpp

@@ -0,0 +1,15 @@
+#pragma once
+
+#include "../device.hpp"
+#include "../graph/graph.hpp"
+#include "../tensor.hpp"
+#include "common/op.hpp"
+
+namespace infinicore::op {
+
+INFINICORE_GRAPH_OP_CLASS(Dsv4FusedRope, Tensor, Tensor, const Tensor &, const Tensor &, bool);
+
+void dsv4_fused_rope_(Tensor q, Tensor k, const Tensor &freq_real, const Tensor &freq_imag, bool has_k = false);
+Tensor dsv4_fused_rope(const Tensor &q, const Tensor &freq_real, const Tensor &freq_imag);
+
+} // namespace infinicore::op

include/infinicore/ops/dsv4_linear_bf16_fp32.hpp

@@ -0,0 +1,15 @@
+#pragma once
+
+#include "../device.hpp"
+#include "../graph/graph.hpp"
+#include "../tensor.hpp"
+#include "common/op.hpp"
+
+namespace infinicore::op {
+
+INFINICORE_GRAPH_OP_CLASS(Dsv4LinearBf16Fp32, Tensor, const Tensor &, const Tensor &);
+
+Tensor dsv4_linear_bf16_fp32(const Tensor &x, const Tensor &w);
+void dsv4_linear_bf16_fp32_(Tensor y, const Tensor &x, const Tensor &w);
+
+} // namespace infinicore::op

include/infinicore/ops/dsv4_mask_topk_ids.hpp

@@ -0,0 +1,14 @@
+#pragma once
+
+#include "../device.hpp"
+#include "../graph/graph.hpp"
+#include "../tensor.hpp"
+#include "common/op.hpp"
+
+namespace infinicore::op {
+
+INFINICORE_GRAPH_OP_CLASS(Dsv4MaskTopkIds, Tensor, const Tensor &);
+
+void dsv4_mask_topk_ids_(Tensor topk_ids, const Tensor &num_token_non_padded);
+
+} // namespace infinicore::op

include/infinicore/ops/dsv4_per_token_quant_int8.hpp

@@ -0,0 +1,17 @@
+#pragma once
+
+#include "../device.hpp"
+#include "../graph/graph.hpp"
+#include "../tensor.hpp"
+#include "common/op.hpp"
+
+#include <utility>
+
+namespace infinicore::op {
+
+INFINICORE_GRAPH_OP_CLASS(Dsv4PerTokenQuantInt8, Tensor, Tensor, const Tensor &);
+
+std::pair<Tensor, Tensor> dsv4_per_token_quant_int8(const Tensor &x);
+void dsv4_per_token_quant_int8_(Tensor q, Tensor scale, const Tensor &x);
+
+} // namespace infinicore::op

include/infinicore/ops/dsv4_rmsnorm_self.hpp

@@ -0,0 +1,15 @@
+#pragma once
+
+#include "../device.hpp"
+#include "../graph/graph.hpp"
+#include "../tensor.hpp"
+#include "common/op.hpp"
+
+namespace infinicore::op {
+
+INFINICORE_GRAPH_OP_CLASS(Dsv4RMSNormSelf, Tensor, const Tensor &, float);
+
+Tensor dsv4_rmsnorm_self(const Tensor &x, float epsilon = 1e-6f);
+void dsv4_rmsnorm_self_(Tensor y, const Tensor &x, float epsilon = 1e-6f);
+
+} // namespace infinicore::op

include/infinicore/ops/dsv4_silu_and_mul.hpp

@@ -0,0 +1,15 @@
+#pragma once
+
+#include "../device.hpp"
+#include "../graph/graph.hpp"
+#include "../tensor.hpp"
+#include "common/op.hpp"
+
+namespace infinicore::op {
+
+INFINICORE_GRAPH_OP_CLASS(Dsv4SiluAndMul, Tensor, const Tensor &, const Tensor &);
+
+Tensor dsv4_silu_and_mul(const Tensor &gate, const Tensor &up);
+void dsv4_silu_and_mul_(Tensor y, const Tensor &gate, const Tensor &up);
+
+} // namespace infinicore::op

include/infinicore/ops/dsv4_silu_mul_quant.hpp

@@ -0,0 +1,9 @@
+#pragma once
+#include "../device.hpp"
+#include "../graph/graph.hpp"
+#include "../tensor.hpp"
+#include "common/op.hpp"
+namespace infinicore::op {
+INFINICORE_GRAPH_OP_CLASS(Dsv4SiluMulQuant, Tensor, Tensor, const Tensor &, const Tensor &);
+void dsv4_silu_mul_quant_(Tensor q, Tensor scale, const Tensor &gate, const Tensor &up);
+} // namespace infinicore::op

include/infinicore/ops/dsv4_swa_prefill_indices.hpp

@@ -0,0 +1,9 @@
+#pragma once
+#include "../device.hpp"
+#include "../graph/graph.hpp"
+#include "../tensor.hpp"
+#include "common/op.hpp"
+namespace infinicore::op {
+INFINICORE_GRAPH_OP_CLASS(Dsv4SwaPrefillIndices, Tensor, int);
+void dsv4_swa_prefill_indices_(Tensor indices, int window_size);
+} // namespace infinicore::op

include/infinicore/ops/dsv4_topk_transform.hpp

@@ -0,0 +1,15 @@
+#pragma once
+
+#include "../device.hpp"
+#include "../graph/graph.hpp"
+#include "../tensor.hpp"
+#include "common/op.hpp"
+
+namespace infinicore::op {
+
+INFINICORE_GRAPH_OP_CLASS(Dsv4TopkTransform, Tensor, const Tensor &, const Tensor &, const Tensor &, int);
+
+Tensor dsv4_topk_transform(const Tensor &scores, const Tensor &seq_lens, const Tensor &page_tables, int page_size = 64);
+void dsv4_topk_transform_(Tensor out, const Tensor &scores, const Tensor &seq_lens, const Tensor &page_tables, int page_size = 64);
+
+} // namespace infinicore::op

include/infiniop.h

@@ -44,6 +44,17 @@
 #include "infiniop/ops/digamma.h"
 #include "infiniop/ops/dist.h"
 #include "infiniop/ops/dot.h"
+#include "infiniop/ops/dsv4_act_quant_fp8.h"
+#include "infiniop/ops/dsv4_add_rmsnorm_quant.h"
+#include "infiniop/ops/dsv4_fused_rope.h"
+#include "infiniop/ops/dsv4_linear_bf16_fp32.h"
+#include "infiniop/ops/dsv4_mask_topk_ids.h"
+#include "infiniop/ops/dsv4_per_token_quant_int8.h"
+#include "infiniop/ops/dsv4_rmsnorm_self.h"
+#include "infiniop/ops/dsv4_silu_and_mul.h"
+#include "infiniop/ops/dsv4_silu_mul_quant.h"
+#include "infiniop/ops/dsv4_swa_prefill_indices.h"
+#include "infiniop/ops/dsv4_topk_transform.h"
 #include "infiniop/ops/embedding.h"
 #include "infiniop/ops/equal.h"
 #include "infiniop/ops/erf.h"

include/infiniop/ops/dsv4_act_quant_fp8.h

@@ -0,0 +1,9 @@
+#ifndef __INFINIOP_DSV4_ACT_QUANT_FP8_API_H__
+#define __INFINIOP_DSV4_ACT_QUANT_FP8_API_H__
+#include "../operator_descriptor.h"
+typedef struct InfiniopDescriptor *infiniopDsv4ActQuantFp8Descriptor_t;
+__INFINI_C __export infiniStatus_t infiniopCreateDsv4ActQuantFp8Descriptor(infiniopHandle_t, infiniopDsv4ActQuantFp8Descriptor_t *, infiniopTensorDescriptor_t, infiniopTensorDescriptor_t, infiniopTensorDescriptor_t, float);
+__INFINI_C __export infiniStatus_t infiniopGetDsv4ActQuantFp8WorkspaceSize(infiniopDsv4ActQuantFp8Descriptor_t, size_t *);
+__INFINI_C __export infiniStatus_t infiniopDsv4ActQuantFp8(infiniopDsv4ActQuantFp8Descriptor_t, void *, size_t, void *, void *, const void *, void *);
+__INFINI_C __export infiniStatus_t infiniopDestroyDsv4ActQuantFp8Descriptor(infiniopDsv4ActQuantFp8Descriptor_t);
+#endif

include/infiniop/ops/dsv4_add_rmsnorm_quant.h

@@ -0,0 +1,9 @@
+#ifndef __INFINIOP_DSV4_ADD_RMSNORM_QUANT_API_H__
+#define __INFINIOP_DSV4_ADD_RMSNORM_QUANT_API_H__
+#include "../operator_descriptor.h"
+typedef struct InfiniopDescriptor *infiniopDsv4AddRMSNormQuantDescriptor_t;
+__INFINI_C __export infiniStatus_t infiniopCreateDsv4AddRMSNormQuantDescriptor(infiniopHandle_t, infiniopDsv4AddRMSNormQuantDescriptor_t *, infiniopTensorDescriptor_t, infiniopTensorDescriptor_t, infiniopTensorDescriptor_t, infiniopTensorDescriptor_t, infiniopTensorDescriptor_t, float);
+__INFINI_C __export infiniStatus_t infiniopGetDsv4AddRMSNormQuantWorkspaceSize(infiniopDsv4AddRMSNormQuantDescriptor_t, size_t *);
+__INFINI_C __export infiniStatus_t infiniopDsv4AddRMSNormQuant(infiniopDsv4AddRMSNormQuantDescriptor_t, void *, size_t, void *, void *, void *, const void *, const void *, void *);
+__INFINI_C __export infiniStatus_t infiniopDestroyDsv4AddRMSNormQuantDescriptor(infiniopDsv4AddRMSNormQuantDescriptor_t);
+#endif

include/infiniop/ops/dsv4_fused_rope.h

@@ -0,0 +1,13 @@
+#ifndef __INFINIOP_DSV4_FUSED_ROPE_API_H__
+#define __INFINIOP_DSV4_FUSED_ROPE_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopDsv4FusedRopeDescriptor_t;
+
+__INFINI_C __export infiniStatus_t infiniopCreateDsv4FusedRopeDescriptor(infiniopHandle_t handle, infiniopDsv4FusedRopeDescriptor_t *desc_ptr, infiniopTensorDescriptor_t q_desc, infiniopTensorDescriptor_t k_desc, infiniopTensorDescriptor_t freq_real_desc, infiniopTensorDescriptor_t freq_imag_desc, int has_k);
+__INFINI_C __export infiniStatus_t infiniopGetDsv4FusedRopeWorkspaceSize(infiniopDsv4FusedRopeDescriptor_t desc, size_t *size);
+__INFINI_C __export infiniStatus_t infiniopDsv4FusedRope(infiniopDsv4FusedRopeDescriptor_t desc, void *workspace, size_t workspace_size, void *q, void *k, const void *freq_real, const void *freq_imag, void *stream);
+__INFINI_C __export infiniStatus_t infiniopDestroyDsv4FusedRopeDescriptor(infiniopDsv4FusedRopeDescriptor_t desc);
+
+#endif // __INFINIOP_DSV4_FUSED_ROPE_API_H__

include/infiniop/ops/dsv4_linear_bf16_fp32.h

@@ -0,0 +1,31 @@
+#ifndef __INFINIOP_DSV4_LINEAR_BF16_FP32_API_H__
+#define __INFINIOP_DSV4_LINEAR_BF16_FP32_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopDsv4LinearBf16Fp32Descriptor_t;
+
+__INFINI_C __export infiniStatus_t infiniopCreateDsv4LinearBf16Fp32Descriptor(
+    infiniopHandle_t handle,
+    infiniopDsv4LinearBf16Fp32Descriptor_t *desc_ptr,
+    infiniopTensorDescriptor_t y_desc,
+    infiniopTensorDescriptor_t x_desc,
+    infiniopTensorDescriptor_t w_desc);
+
+__INFINI_C __export infiniStatus_t infiniopGetDsv4LinearBf16Fp32WorkspaceSize(
+    infiniopDsv4LinearBf16Fp32Descriptor_t desc,
+    size_t *size);
+
+__INFINI_C __export infiniStatus_t infiniopDsv4LinearBf16Fp32(
+    infiniopDsv4LinearBf16Fp32Descriptor_t desc,
+    void *workspace,
+    size_t workspace_size,
+    void *y,
+    const void *x,
+    const void *w,
+    void *stream);
+
+__INFINI_C __export infiniStatus_t infiniopDestroyDsv4LinearBf16Fp32Descriptor(
+    infiniopDsv4LinearBf16Fp32Descriptor_t desc);
+
+#endif

include/infiniop/ops/dsv4_mask_topk_ids.h

@@ -0,0 +1,29 @@
+#ifndef __INFINIOP_DSV4_MASK_TOPK_IDS_API_H__
+#define __INFINIOP_DSV4_MASK_TOPK_IDS_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopDsv4MaskTopkIdsDescriptor_t;
+
+__INFINI_C __export infiniStatus_t infiniopCreateDsv4MaskTopkIdsDescriptor(
+    infiniopHandle_t handle,
+    infiniopDsv4MaskTopkIdsDescriptor_t *desc_ptr,
+    infiniopTensorDescriptor_t topk_ids_desc,
+    infiniopTensorDescriptor_t num_token_non_padded_desc);
+
+__INFINI_C __export infiniStatus_t infiniopGetDsv4MaskTopkIdsWorkspaceSize(
+    infiniopDsv4MaskTopkIdsDescriptor_t desc,
+    size_t *size);
+
+__INFINI_C __export infiniStatus_t infiniopDsv4MaskTopkIds(
+    infiniopDsv4MaskTopkIdsDescriptor_t desc,
+    void *workspace,
+    size_t workspace_size,
+    void *topk_ids,
+    const void *num_token_non_padded,
+    void *stream);
+
+__INFINI_C __export infiniStatus_t infiniopDestroyDsv4MaskTopkIdsDescriptor(
+    infiniopDsv4MaskTopkIdsDescriptor_t desc);
+
+#endif

include/infiniop/ops/dsv4_per_token_quant_int8.h

@@ -0,0 +1,31 @@
+#ifndef __INFINIOP_DSV4_PER_TOKEN_QUANT_INT8_API_H__
+#define __INFINIOP_DSV4_PER_TOKEN_QUANT_INT8_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopDsv4PerTokenQuantInt8Descriptor_t;
+
+__INFINI_C __export infiniStatus_t infiniopCreateDsv4PerTokenQuantInt8Descriptor(
+    infiniopHandle_t handle,
+    infiniopDsv4PerTokenQuantInt8Descriptor_t *desc_ptr,
+    infiniopTensorDescriptor_t q_desc,
+    infiniopTensorDescriptor_t scale_desc,
+    infiniopTensorDescriptor_t x_desc);
+
+__INFINI_C __export infiniStatus_t infiniopGetDsv4PerTokenQuantInt8WorkspaceSize(
+    infiniopDsv4PerTokenQuantInt8Descriptor_t desc,
+    size_t *size);
+
+__INFINI_C __export infiniStatus_t infiniopDsv4PerTokenQuantInt8(
+    infiniopDsv4PerTokenQuantInt8Descriptor_t desc,
+    void *workspace,
+    size_t workspace_size,
+    void *q,
+    void *scale,
+    const void *x,
+    void *stream);
+
+__INFINI_C __export infiniStatus_t infiniopDestroyDsv4PerTokenQuantInt8Descriptor(
+    infiniopDsv4PerTokenQuantInt8Descriptor_t desc);
+
+#endif // __INFINIOP_DSV4_PER_TOKEN_QUANT_INT8_API_H__

include/infiniop/ops/dsv4_rmsnorm_self.h

@@ -0,0 +1,18 @@
+#ifndef __INFINIOP_DSV4_RMSNORM_SELF_API_H__
+#define __INFINIOP_DSV4_RMSNORM_SELF_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopDsv4RMSNormSelfDescriptor_t;
+
+__INFINI_C __export infiniStatus_t infiniopCreateDsv4RMSNormSelfDescriptor(
+    infiniopHandle_t handle,
+    infiniopDsv4RMSNormSelfDescriptor_t *desc_ptr,
+    infiniopTensorDescriptor_t y_desc,
+    infiniopTensorDescriptor_t x_desc,
+    float epsilon);
+__INFINI_C __export infiniStatus_t infiniopGetDsv4RMSNormSelfWorkspaceSize(infiniopDsv4RMSNormSelfDescriptor_t desc, size_t *size);
+__INFINI_C __export infiniStatus_t infiniopDsv4RMSNormSelf(infiniopDsv4RMSNormSelfDescriptor_t desc, void *workspace, size_t workspace_size, void *y, const void *x, void *stream);
+__INFINI_C __export infiniStatus_t infiniopDestroyDsv4RMSNormSelfDescriptor(infiniopDsv4RMSNormSelfDescriptor_t desc);
+
+#endif // __INFINIOP_DSV4_RMSNORM_SELF_API_H__

include/infiniop/ops/dsv4_silu_and_mul.h

@@ -0,0 +1,13 @@
+#ifndef __INFINIOP_DSV4_SILU_AND_MUL_API_H__
+#define __INFINIOP_DSV4_SILU_AND_MUL_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopDsv4SiluAndMulDescriptor_t;
+
+__INFINI_C __export infiniStatus_t infiniopCreateDsv4SiluAndMulDescriptor(infiniopHandle_t handle, infiniopDsv4SiluAndMulDescriptor_t *desc_ptr, infiniopTensorDescriptor_t y_desc, infiniopTensorDescriptor_t gate_desc, infiniopTensorDescriptor_t up_desc);
+__INFINI_C __export infiniStatus_t infiniopGetDsv4SiluAndMulWorkspaceSize(infiniopDsv4SiluAndMulDescriptor_t desc, size_t *size);
+__INFINI_C __export infiniStatus_t infiniopDsv4SiluAndMul(infiniopDsv4SiluAndMulDescriptor_t desc, void *workspace, size_t workspace_size, void *y, const void *gate, const void *up, void *stream);
+__INFINI_C __export infiniStatus_t infiniopDestroyDsv4SiluAndMulDescriptor(infiniopDsv4SiluAndMulDescriptor_t desc);
+
+#endif // __INFINIOP_DSV4_SILU_AND_MUL_API_H__

include/infiniop/ops/dsv4_silu_mul_quant.h

@@ -0,0 +1,9 @@
+#ifndef __INFINIOP_DSV4_SILU_MUL_QUANT_API_H__
+#define __INFINIOP_DSV4_SILU_MUL_QUANT_API_H__
+#include "../operator_descriptor.h"
+typedef struct InfiniopDescriptor *infiniopDsv4SiluMulQuantDescriptor_t;
+__INFINI_C __export infiniStatus_t infiniopCreateDsv4SiluMulQuantDescriptor(infiniopHandle_t, infiniopDsv4SiluMulQuantDescriptor_t *, infiniopTensorDescriptor_t, infiniopTensorDescriptor_t, infiniopTensorDescriptor_t, infiniopTensorDescriptor_t);
+__INFINI_C __export infiniStatus_t infiniopGetDsv4SiluMulQuantWorkspaceSize(infiniopDsv4SiluMulQuantDescriptor_t, size_t *);
+__INFINI_C __export infiniStatus_t infiniopDsv4SiluMulQuant(infiniopDsv4SiluMulQuantDescriptor_t, void *, size_t, void *, void *, const void *, const void *, void *);
+__INFINI_C __export infiniStatus_t infiniopDestroyDsv4SiluMulQuantDescriptor(infiniopDsv4SiluMulQuantDescriptor_t);
+#endif

include/infiniop/ops/dsv4_swa_prefill_indices.h

@@ -0,0 +1,9 @@
+#ifndef __INFINIOP_DSV4_SWA_PREFILL_INDICES_API_H__
+#define __INFINIOP_DSV4_SWA_PREFILL_INDICES_API_H__
+#include "../operator_descriptor.h"
+typedef struct InfiniopDescriptor *infiniopDsv4SwaPrefillIndicesDescriptor_t;
+__INFINI_C __export infiniStatus_t infiniopCreateDsv4SwaPrefillIndicesDescriptor(infiniopHandle_t, infiniopDsv4SwaPrefillIndicesDescriptor_t *, infiniopTensorDescriptor_t, int);
+__INFINI_C __export infiniStatus_t infiniopGetDsv4SwaPrefillIndicesWorkspaceSize(infiniopDsv4SwaPrefillIndicesDescriptor_t, size_t *);
+__INFINI_C __export infiniStatus_t infiniopDsv4SwaPrefillIndices(infiniopDsv4SwaPrefillIndicesDescriptor_t, void *, size_t, void *, void *);
+__INFINI_C __export infiniStatus_t infiniopDestroyDsv4SwaPrefillIndicesDescriptor(infiniopDsv4SwaPrefillIndicesDescriptor_t);
+#endif

include/infiniop/ops/dsv4_topk_transform.h

@@ -0,0 +1,13 @@
+#ifndef __INFINIOP_DSV4_TOPK_TRANSFORM_API_H__
+#define __INFINIOP_DSV4_TOPK_TRANSFORM_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef struct InfiniopDescriptor *infiniopDsv4TopkTransformDescriptor_t;
+
+__INFINI_C __export infiniStatus_t infiniopCreateDsv4TopkTransformDescriptor(infiniopHandle_t handle, infiniopDsv4TopkTransformDescriptor_t *desc_ptr, infiniopTensorDescriptor_t out_desc, infiniopTensorDescriptor_t scores_desc, infiniopTensorDescriptor_t seq_lens_desc, infiniopTensorDescriptor_t page_tables_desc, int page_size);
+__INFINI_C __export infiniStatus_t infiniopGetDsv4TopkTransformWorkspaceSize(infiniopDsv4TopkTransformDescriptor_t desc, size_t *size);
+__INFINI_C __export infiniStatus_t infiniopDsv4TopkTransform(infiniopDsv4TopkTransformDescriptor_t desc, void *workspace, size_t workspace_size, void *out, const void *scores, const void *seq_lens, const void *page_tables, void *stream);
+__INFINI_C __export infiniStatus_t infiniopDestroyDsv4TopkTransformDescriptor(infiniopDsv4TopkTransformDescriptor_t desc);
+
+#endif // __INFINIOP_DSV4_TOPK_TRANSFORM_API_H__

src/infinicore/ops/dsv4_act_quant_fp8/dsv4_act_quant_fp8.cc

@@ -0,0 +1,13 @@
+#include "infinicore/ops/dsv4_act_quant_fp8.hpp"
+#include "../../utils.hpp"
+namespace infinicore::op {
+INFINICORE_GRAPH_OP_DISPATCHERS_IMPL(Dsv4ActQuantFp8);
+Dsv4ActQuantFp8::Dsv4ActQuantFp8(Tensor xq, Tensor scale, const Tensor &x, float fp8_max) {
+    INFINICORE_ASSERT_TENSORS_SAME_DEVICE(xq, scale, x);
+    INFINICORE_GRAPH_OP_DISPATCH(xq->device().getType(), xq, scale, x, fp8_max);
+}
+void Dsv4ActQuantFp8::execute(Tensor xq, Tensor scale, const Tensor &x, float fp8_max) {
+    INFINICORE_GRAPH_OP_RECORD_OR_RUN(Dsv4ActQuantFp8, xq, scale, x, fp8_max);
+}
+void dsv4_act_quant_fp8_(Tensor xq, Tensor scale, const Tensor &x, float fp8_max) { Dsv4ActQuantFp8::execute(xq, scale, x, fp8_max); }
+} // namespace infinicore::op