issue/170: quantize_gptq

Author: xgqdut2016

include/infinicore.h

@@ -70,6 +70,7 @@ typedef enum {
     INFINI_DTYPE_C64 = 17,
     INFINI_DTYPE_C128 = 18,
     INFINI_DTYPE_BF16 = 19,
+    INFINI_DTYPE_I4 = 20,
 } infiniDtype_t;
 
 #endif // __INFINICORE_API_H__

include/infiniop.h

@@ -13,6 +13,7 @@
 #include "infiniop/ops/max_pool.h"
 #include "infiniop/ops/mlp.h"
 #include "infiniop/ops/mul.h"
+#include "infiniop/ops/quantize_gptq.h"
 #include "infiniop/ops/random_sample.h"
 #include "infiniop/ops/rearrange.h"
 #include "infiniop/ops/relu.h"

include/infiniop/ops/quantize_gptq.h

@@ -0,0 +1,40 @@
+#ifndef __INFINIOP_QUANTIZE_GPTQ_API_H__
+#define __INFINIOP_QUANTIZE_GPTQ_API_H__
+
+#include "../operator_descriptor.h"
+
+typedef InfiniopDescriptor *infiniopQuantizeGPTQDescriptor_t;
+
+__C __export infiniStatus_t infiniopCreateQuantizeGPTQDescriptor(infiniopHandle_t handle,
+                                                                 infiniopQuantizeGPTQDescriptor_t *desc_ptr,
+                                                                 infiniopTensorDescriptor_t c_desc,
+                                                                 infiniopTensorDescriptor_t a_desc,
+                                                                 infiniopTensorDescriptor_t packed_weights_desc,
+                                                                 infiniopTensorDescriptor_t b_scale_desc,
+                                                                 infiniopTensorDescriptor_t zero_desc);
+
+__C __export infiniStatus_t infiniopGetQuantizeGPTQWorkspaceSize(infiniopQuantizeGPTQDescriptor_t desc, size_t *size);
+
+__C __export infiniStatus_t infiniopQuantizeGPTQ(infiniopQuantizeGPTQDescriptor_t desc,
+                                                 void *workspace,
+                                                 size_t workspace_size,
+                                                 void *packed_weights,
+                                                 void *b_scale,
+                                                 void *zero,
+                                                 const void *a,
+                                                 const void *b,
+                                                 void *stream);
+
+__C __export infiniStatus_t infiniopQuantizeLinearGPTQ(infiniopQuantizeGPTQDescriptor_t desc,
+                                                       void *workspace,
+                                                       size_t workspace_size,
+                                                       void *c,
+                                                       const void *a,
+                                                       void *packed_weights,
+                                                       void *b_scale,
+                                                       void *zero,
+                                                       void *stream);
+
+__C __export infiniStatus_t infiniopDestroyQuantizeGPTQDescriptor(infiniopQuantizeGPTQDescriptor_t desc);
+
+#endif