init v0.23.0

Signed-off-by: Sun Ruoxi <sunruoxi@4paradigm.com>

csrc/attention/ngram_spec_decode/op_host/op_api/aclnn_ngram_spec_decode.cpp

@@ -0,0 +1,73 @@
+#include <string.h>
+#include "graph/types.h"
+#include "aclnn_ngram_spec_decode.h"
+
+enum NnopbaseHcclServerType {
+    NNOPBASE_HCCL_SERVER_TYPE_AICPU = 0,
+    NNOPBASE_HCCL_SERVER_TYPE_MTE,
+    NNOPBASE_HCCL_SERVER_TYPE_END
+};
+extern "C" void __attribute__((weak)) NnopbaseSetHcclServerType(void *executor, NnopbaseHcclServerType sType);
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+extern aclnnStatus aclnnInnerNgramSpecDecodeGetWorkspaceSize(
+    const aclTensor *tokenIds,
+    const aclTensor *numTokensNoSpec,
+    const aclTensor *sampledTokenIds,
+    const aclTensor *discardRequestMask,
+    int64_t vocabSize,
+    int64_t minN,
+    int64_t maxN,
+    int64_t k,
+    const aclTensor *nextTokenIds,
+    const aclTensor *draftTokenIds,
+    const aclTensor *numValidDraftTokens,
+    uint64_t *workspaceSize,
+    aclOpExecutor **executor);
+
+extern aclnnStatus aclnnInnerNgramSpecDecode(
+    void *workspace,
+    uint64_t workspaceSize,
+    aclOpExecutor *executor,
+    aclrtStream stream);
+
+aclnnStatus aclnnNgramSpecDecodeGetWorkspaceSize(
+    const aclTensor *tokenIds,
+    const aclTensor *numTokensNoSpec,
+    const aclTensor *sampledTokenIds,
+    const aclTensor *discardRequestMask,
+    int64_t vocabSize,
+    int64_t minN,
+    int64_t maxN,
+    int64_t k,
+    const aclTensor *nextTokenIds,
+    const aclTensor *draftTokenIds,
+    const aclTensor *numValidDraftTokens,
+    uint64_t *workspaceSize,
+    aclOpExecutor **executor)
+{
+    return aclnnInnerNgramSpecDecodeGetWorkspaceSize(
+        tokenIds, numTokensNoSpec, sampledTokenIds, discardRequestMask,
+        vocabSize, minN, maxN, k,
+        nextTokenIds, draftTokenIds, numValidDraftTokens,
+        workspaceSize, executor);
+}
+
+aclnnStatus aclnnNgramSpecDecode(
+    void *workspace,
+    uint64_t workspaceSize,
+    aclOpExecutor *executor,
+    aclrtStream stream)
+{
+    if (NnopbaseSetHcclServerType) {
+        NnopbaseSetHcclServerType(executor, NNOPBASE_HCCL_SERVER_TYPE_MTE);
+    }
+    return aclnnInnerNgramSpecDecode(workspace, workspaceSize, executor, stream);
+}
+
+#ifdef __cplusplus
+}
+#endif

csrc/attention/ngram_spec_decode/op_host/op_api/aclnn_ngram_spec_decode.h

@@ -0,0 +1,56 @@
+#ifndef ACLNN_NGRAM_SPEC_DECODE_H_
+#define ACLNN_NGRAM_SPEC_DECODE_H_
+
+#include "aclnn/acl_meta.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* function: aclnnNgramSpecDecodeGetWorkspaceSize
+ * tokenIds : required, [batch_size, max_seq_len], int32
+ * numTokensNoSpec : required, [batch_size], int32
+ * sampledTokenIds : required, [batch_size, max_new_tokens], int32
+ * discardRequestMask : required, [batch_size], int32
+ * vocabSize : required, int
+ * minN : required, int
+ * maxN : required, int
+ * k : required, int
+ * nextTokenIds : required, [batch_size], int32
+ * draftTokenIds : required, [batch_size, k], int32
+ * numValidDraftTokens : required, [batch_size], int32
+ * workspaceSize : size of workspace(output).
+ * executor : executor context(output).
+ */
+__attribute__((visibility("default"))) aclnnStatus aclnnNgramSpecDecodeGetWorkspaceSize(
+    const aclTensor *tokenIds,
+    const aclTensor *numTokensNoSpec,
+    const aclTensor *sampledTokenIds,
+    const aclTensor *discardRequestMask,
+    int64_t vocabSize,
+    int64_t minN,
+    int64_t maxN,
+    int64_t k,
+    const aclTensor *nextTokenIds,
+    const aclTensor *draftTokenIds,
+    const aclTensor *numValidDraftTokens,
+    uint64_t *workspaceSize,
+    aclOpExecutor **executor);
+
+/* function: aclnnNgramSpecDecode
+ * workspace : workspace memory addr(input).
+ * workspaceSize : size of workspace(input).
+ * executor : executor context(input).
+ * stream : acl stream.
+ */
+__attribute__((visibility("default"))) aclnnStatus aclnnNgramSpecDecode(
+    void *workspace,
+    uint64_t workspaceSize,
+    aclOpExecutor *executor,
+    aclrtStream stream);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // ACLNN_NGRAM_SPEC_DECODE_H_