mindspore 2.4.1__cp39-cp39-manylinux1_x86_64.whl → 2.4.10__cp39-cp39-manylinux1_x86_64.whl

mindspore/lib/plugin/ascend/custom_ascendc_910b/op_impl/ai_core/tbe/custom_ascendc_910b_impl/dynamic/prompt_kv_cache.py

@@ -0,0 +1,215 @@
+#!/usr/bin/env python
+# -*- coding: UTF-8 -*-
+"""
+Copyright (c) Huawei Technologies Co., Ltd. 2023-2024. All rights reserved.
+"""
+
+import os, sys
+import ctypes
+import json
+import shutil
+from tbe.common.platform import get_soc_spec
+from tbe.common.utils import para_check
+from tbe.tikcpp import compile_op, replay_op, check_op_cap, generalize_op_params, get_code_channel, OpInfo
+from tbe.tikcpp.compile_op import CommonUtility, AscendCLogLevel
+from tbe.common.buildcfg import get_default_build_config
+from impl.util.platform_adapter import tbe_register
+from tbe.common.buildcfg import get_current_build_config
+PYF_PATH = os.path.dirname(os.path.realpath(__file__))
+
+DTYPE_MAP = {"float32": ["DT_FLOAT", "float"],
+    "float16": ["DT_FLOAT16", "half"],
+    "int8": ["DT_INT8", "int8_t"],
+    "int16": ["DT_INT16", "int16_t"],
+    "int32": ["DT_INT32", "int32_t"],
+    "int64": ["DT_INT64", "int64_t"],
+    "uint1": ["DT_UINT1", "uint8_t"],
+    "uint8": ["DT_UINT8", "uint8_t"],
+    "uint16": ["DT_UINT16", "uint16_t"],
+    "uint32": ["DT_UINT32", "uint32_t"],
+    "uint64": ["DT_UINT64", "uint64_t"],
+    "bool": ["DT_BOOL", "bool"],
+    "double": ["DT_DOUBLE", "double"],
+    "dual": ["DT_DUAL", "unknown"],
+    "dual_sub_int8": ["DT_DUAL_SUB_INT8", "unknown"],
+    "dual_sub_uint8": ["DT_DUAL_SUB_UINT8", "unknown"],
+    "string": ["DT_STRING", "unknown"],
+    "complex32": ["DT_COMPLEX32", "unknown"],
+    "complex64": ["DT_COMPLEX64", "unknown"],
+    "complex128": ["DT_COMPLEX128", "unknown"],
+    "qint8": ["DT_QINT8", "unknown"],
+    "qint16": ["DT_QINT16", "unknown"],
+    "qint32": ["DT_QINT32", "unknown"],
+    "quint8": ["DT_QUINT8", "unknown"],
+    "quint16": ["DT_QUINT16", "unknown"],
+    "resource": ["DT_RESOURCE", "unknown"],
+    "string_ref": ["DT_STRING_REF", "unknown"],
+    "int4": ["DT_INT4", "int4b_t"],
+    "bfloat16": ["DT_BF16", "bfloat16_t"]}
+
+def add_dtype_fmt_option_single(x, x_n, is_ref: bool = False):
+    options = []
+    x_fmt = x.get("format")
+    x_dtype = x.get("dtype")
+    x_n_in_kernel = x_n + '_REF' if is_ref else x_n
+    options.append("-DDTYPE_{n}={t}".format(n=x_n_in_kernel, t=DTYPE_MAP.get(x_dtype)[1]))
+    options.append("-DORIG_DTYPE_{n}={ot}".format(n=x_n_in_kernel, ot=DTYPE_MAP.get(x_dtype)[0]))
+    options.append("-DFORMAT_{n}=FORMAT_{f}".format(n=x_n_in_kernel, f=x_fmt))
+    return options
+
+def get_dtype_fmt_options(__inputs__, __outputs__):
+    options = []
+    input_names = ['cache', 'update', 'valid_seq_len', 'batch_index', 'seq_len_axis', 'new_max_seq_len', 'cur_max_seq_len']
+    output_names = ['out']
+    unique_param_name_set = set()
+    for idx, x in enumerate(__inputs__):
+        if x is None:
+            continue
+        x_n = input_names[idx].upper()
+        unique_param_name_set.add(x_n)
+        options += add_dtype_fmt_option_single(x, x_n)
+
+    for idx, x in enumerate(__outputs__):
+        if x is None:
+            continue
+        x_n = output_names[idx].upper()
+        if x_n in unique_param_name_set:
+            options += add_dtype_fmt_option_single(x, x_n, True)
+        else:
+            options += add_dtype_fmt_option_single(x, x_n)
+    return options
+
+def load_dso(so_path):
+    try:
+        ctypes.CDLL(so_path)
+    except OSError as error :
+        CommonUtility.print_compile_log("", error, AscendCLogLevel.LOG_ERROR)
+        raise RuntimeError("cannot open %s" %(so_path))
+    else:
+        msg = "load so succ " + so_path
+        CommonUtility.print_compile_log("", msg, AscendCLogLevel.LOG_INFO)
+
+def get_shortsoc_compile_option(compile_option_list: list, shortsoc:str):
+    compile_options = []
+    if shortsoc in compile_option_list:
+        compile_options.extend(compile_option_list[shortsoc])
+    if '__ALLSOC__' in compile_option_list:
+        compile_options.extend(compile_option_list['__ALLSOC__'])
+    return compile_options
+
+def get_kernel_source(src_file, dir_snake, dir_ex):
+    src_ex = os.path.join(PYF_PATH, "..", "ascendc", dir_ex, src_file)
+    if os.path.exists(src_ex):
+        return src_ex
+    src = os.environ.get('BUILD_KERNEL_SRC')
+    if src and os.path.exists(src):
+        return src
+    src = os.path.join(PYF_PATH, "..", "ascendc", dir_snake, src_file)
+    if os.path.exists(src):
+        return src
+    src = os.path.join(PYF_PATH, src_file)
+    if os.path.exists(src):
+        return src
+    src = os.path.join(PYF_PATH, "..", "ascendc", dir_snake, dir_snake + ".cpp")
+    if os.path.exists(src):
+        return src
+    src = os.path.join(PYF_PATH, "..", "ascendc", dir_ex, dir_ex + ".cpp")
+    if os.path.exists(src):
+        return src
+    src = os.path.join(PYF_PATH, "..", "ascendc", os.path.splitext(src_file)[0], src_file)
+    if os.path.exists(src):
+        return src
+    return src_ex
+
+def _build_args(cache_in__, update_in__, valid_seq_len_in__, batch_index_in__, seq_len_axis_in__, new_max_seq_len_in__, cur_max_seq_len_in__, out_out_):
+    __inputs__ = []
+    for arg in [cache_in__, update_in__, valid_seq_len_in__, batch_index_in__, seq_len_axis_in__, new_max_seq_len_in__, cur_max_seq_len_in__]:
+        if arg != None:
+            if isinstance(arg, (list, tuple)):
+                if len(arg) == 0:
+                    continue
+                __inputs__.append(arg[0])
+            else:
+                __inputs__.append(arg)
+        else:
+            __inputs__.append(arg)
+    __outputs__ = []
+    for arg in [out_out_]:
+        if arg != None:
+            if isinstance(arg, (list, tuple)):
+                if len(arg) == 0:
+                    continue
+                __outputs__.append(arg[0])
+            else:
+                __outputs__.append(arg)
+        else:
+            __outputs__.append(arg)
+    __attrs__ = []
+    return __inputs__, __outputs__, __attrs__
+
+@tbe_register.register_operator("PromptKvCache", trans_bool_to_s8=False)
+@para_check.check_op_params(para_check.REQUIRED_INPUT, para_check.REQUIRED_INPUT, para_check.REQUIRED_INPUT, para_check.REQUIRED_INPUT, para_check.REQUIRED_INPUT, para_check.REQUIRED_INPUT, para_check.REQUIRED_INPUT, para_check.REQUIRED_OUTPUT, para_check.KERNEL_NAME)
+def prompt_kv_cache(cache_in__, update_in__, valid_seq_len_in__, batch_index_in__, seq_len_axis_in__, new_max_seq_len_in__, cur_max_seq_len_in__, out_out_, kernel_name="prompt_kv_cache", impl_mode=""):
+    # do ascendc build step
+    if get_current_build_config("enable_op_prebuild"):
+        return
+    __inputs__, __outputs__, __attrs__ = _build_args(cache_in__, update_in__, valid_seq_len_in__, batch_index_in__, seq_len_axis_in__, new_max_seq_len_in__, cur_max_seq_len_in__, out_out_)
+    options = get_dtype_fmt_options(__inputs__, __outputs__)
+    options += ["-x", "cce"]
+    bisheng = os.environ.get('BISHENG_REAL_PATH')
+    if bisheng is None:
+        bisheng = shutil.which("bisheng")
+    if bisheng != None:
+        bisheng_path = os.path.dirname(bisheng)
+        tikcpp_path = os.path.realpath(os.path.join(bisheng_path, "..", "..", "tikcpp"))
+    else:
+        tikcpp_path = os.path.realpath("/usr/local/Ascend/latest/compiler/tikcpp")
+    options.append("-I" + tikcpp_path)
+    options.append("-I" + os.path.join(tikcpp_path, "..", "..", "include"))
+    options.append("-I" + os.path.join(tikcpp_path, "tikcfw"))
+    options.append("-I" + os.path.join(tikcpp_path, "tikcfw", "impl"))
+    options.append("-I" + os.path.join(tikcpp_path, "tikcfw", "interface"))
+    options.append("-I" + os.path.join(PYF_PATH, "..", "ascendc", "common"))
+    if impl_mode == "high_performance":
+        options.append("-DHIGH_PERFORMANCE=1")
+    elif impl_mode == "high_precision":
+        options.append("-DHIGH_PRECISION=1")
+    if get_current_build_config("enable_deterministic_mode") == 1:
+        options.append("-DDETERMINISTIC_MODE=1")
+    else:
+        options.append("-DDETERMINISTIC_MODE=0")
+
+    custom_compile_options = {},
+    custom_all_compile_options = {},
+    soc_version = get_soc_spec("SOC_VERSION")
+    soc_short = get_soc_spec("SHORT_SOC_VERSION").lower()
+    custom_compile_options_soc = get_shortsoc_compile_option(custom_compile_options[0], soc_short)
+    custom_all_compile_options_soc = get_shortsoc_compile_option(custom_all_compile_options[0], soc_short)
+    options += custom_all_compile_options_soc
+    options += custom_compile_options_soc
+
+    origin_func_name = "prompt_kv_cache"
+    ascendc_src_dir_ex = "prompt_kv_cache"
+    ascendc_src_dir = "prompt_kv_cache"
+    ascendc_src_file = "prompt_kv_cache.cpp"
+    src = get_kernel_source(ascendc_src_file, ascendc_src_dir, ascendc_src_dir_ex)
+
+    msg = "start compile Acend C Operator PromptKvCache, kernel name is " + kernel_name
+    CommonUtility.print_compile_log("", msg, AscendCLogLevel.LOG_INFO)
+    op_type = "PromptKvCache"
+    code_channel = get_code_channel(src, kernel_name, op_type, options)
+    op_info = OpInfo(kernel_name = kernel_name, op_type = op_type, inputs = __inputs__, outputs = __outputs__,\
+        attrs = __attrs__, impl_mode = impl_mode, origin_inputs=[cache_in__, update_in__, valid_seq_len_in__, batch_index_in__, seq_len_axis_in__, new_max_seq_len_in__, cur_max_seq_len_in__], origin_outputs = [out_out_],\
+                param_type_dynamic = False, mc2_ctx = [], param_type_list = ['required', 'required', 'required', 'required', 'required', 'required', 'required', 'required'], init_value_list = [None],\
+                output_shape_depend_on_compute = [])
+    compile_op(src, origin_func_name, op_info, options, code_channel, '{}')
+
+def op_select_format(cache_in__, update_in__, valid_seq_len_in__, batch_index_in__, seq_len_axis_in__, new_max_seq_len_in__, cur_max_seq_len_in__, out_out_, impl_mode=""):
+    __inputs__, __outputs__, __attrs__ = _build_args(cache_in__, update_in__, valid_seq_len_in__, batch_index_in__, seq_len_axis_in__, new_max_seq_len_in__, cur_max_seq_len_in__, out_out_)
+    result = check_op_cap("op_select_format", "PromptKvCache", __inputs__, __outputs__, __attrs__)
+    return result.decode("utf-8")
+
+def get_op_specific_info(cache_in__, update_in__, valid_seq_len_in__, batch_index_in__, seq_len_axis_in__, new_max_seq_len_in__, cur_max_seq_len_in__, out_out_, impl_mode=""):
+    __inputs__, __outputs__, __attrs__ = _build_args(cache_in__, update_in__, valid_seq_len_in__, batch_index_in__, seq_len_axis_in__, new_max_seq_len_in__, cur_max_seq_len_in__, out_out_)
+    result = check_op_cap("get_op_specific_info", "PromptKvCache", __inputs__, __outputs__, __attrs__)
+    return result.decode("utf-8")

mindspore/lib/plugin/ascend/custom_ascendc_910b/op_impl/ai_core/tbe/kernel/ascend310p/all_finite/AllFinite_52f59e2a65d9b1bb002de35c2819754a.json

@@ -0,0 +1,80 @@
+{
+ "binFileName": "AllFinite_52f59e2a65d9b1bb002de35c2819754a",
+ "binFileSuffix": ".o",
+ "blockDim": -1,
+ "coreType": "AiCore",
+ "globalworkspace_spec_workspace": {
+  "size": 32,
+  "type": 0
+ },
+ "kernelName": "AllFinite_52f59e2a65d9b1bb002de35c2819754a",
+ "magic": "RT_DEV_BINARY_MAGIC_ELF",
+ "memoryStamping": [],
+ "opParaSize": 64,
+ "parameters": [
+  null,
+  null,
+  null
+ ],
+ "sha256": "e78697ea51b702330033a1e17ce5b102df80ae6d9a29f099d3e97d36e9f5728c",
+ "workspace": {
+  "num": 1,
+  "size": [
+   -1
+  ],
+  "type": [
+   0
+  ]
+ },
+ "kernelList": [
+  {
+   "kernelName": "AllFinite_52f59e2a65d9b1bb002de35c2819754a_0"
+  }
+ ],
+ "optionalInputMode": "gen_placeholder",
+ "optionalOutputMode": "gen_placeholder",
+ "compileInfo": {},
+ "supportInfo": {
+  "implMode": "high_performance",
+  "int64Mode": false,
+  "simplifiedKeyMode": 0,
+  "simplifiedKey": [
+   "AllFinite/d=0,p=1/0,2/12,2",
+   "AllFinite/d=0,p=0/0,2/12,2",
+   "AllFinite/d=1,p=1/0,2/12,2",
+   "AllFinite/d=1,p=0/0,2/12,2"
+  ],
+  "optionalInputMode": "gen_placeholder",
+  "optionalOutputMode": "gen_placeholder",
+  "staticKey": "be3739d2ca4bc4d2e000c3359e0160d8d8f747cc5df8b0a05d03649390348d1f,d0f82552295948866b5bab1c0ef5a9cd9662ee964af71eeea40b0c8d27c08835",
+  "inputs": [
+   {
+    "name": "gradient",
+    "index": 0,
+    "dtype": "float32",
+    "format": "ND",
+    "paramType": "required",
+    "shape": [
+     -2
+    ],
+    "format_match_mode": "FormatAgnostic"
+   }
+  ],
+  "outputs": [
+   {
+    "name": "is_finite",
+    "index": 0,
+    "dtype": "bool",
+    "format": "ND",
+    "paramType": "required",
+    "shape": [
+     -2
+    ],
+    "format_match_mode": "FormatAgnostic"
+   }
+  ],
+  "opMode": "dynamic",
+  "deterministic": "ignore"
+ },
+ "filePath": "ascend310p/bin/all_finite/AllFinite_52f59e2a65d9b1bb002de35c2819754a.json"
+}

mindspore/lib/plugin/ascend/custom_ascendc_910b/op_impl/ai_core/tbe/kernel/ascend310p/all_finite/AllFinite_6b5e50e30256d85838d6ce83514df20f.json

@@ -0,0 +1,80 @@
+{
+ "binFileName": "AllFinite_6b5e50e30256d85838d6ce83514df20f",
+ "binFileSuffix": ".o",
+ "blockDim": -1,
+ "coreType": "AiCore",
+ "globalworkspace_spec_workspace": {
+  "size": 32,
+  "type": 0
+ },
+ "kernelName": "AllFinite_6b5e50e30256d85838d6ce83514df20f",
+ "magic": "RT_DEV_BINARY_MAGIC_ELF",
+ "memoryStamping": [],
+ "opParaSize": 64,
+ "parameters": [
+  null,
+  null,
+  null
+ ],
+ "sha256": "c74d3489875cc5f3cb20db6dc697d74a09d1aa355bf8d1c2913519f96ae964dd",
+ "workspace": {
+  "num": 1,
+  "size": [
+   -1
+  ],
+  "type": [
+   0
+  ]
+ },
+ "kernelList": [
+  {
+   "kernelName": "AllFinite_6b5e50e30256d85838d6ce83514df20f_0"
+  }
+ ],
+ "optionalInputMode": "gen_placeholder",
+ "optionalOutputMode": "gen_placeholder",
+ "compileInfo": {},
+ "supportInfo": {
+  "implMode": "high_performance",
+  "int64Mode": false,
+  "simplifiedKeyMode": 0,
+  "simplifiedKey": [
+   "AllFinite/d=0,p=1/27,2/12,2",
+   "AllFinite/d=0,p=0/27,2/12,2",
+   "AllFinite/d=1,p=1/27,2/12,2",
+   "AllFinite/d=1,p=0/27,2/12,2"
+  ],
+  "optionalInputMode": "gen_placeholder",
+  "optionalOutputMode": "gen_placeholder",
+  "staticKey": "9dc4423703e6fab72bb3c8372391b2e2af5d522f3de908e3e400a1b5e18e7158,7f66e3c6666676ff1823556ab378ab14ced057187dd81817f072c958578fcd6a",
+  "inputs": [
+   {
+    "name": "gradient",
+    "index": 0,
+    "dtype": "bfloat16",
+    "format": "ND",
+    "paramType": "required",
+    "shape": [
+     -2
+    ],
+    "format_match_mode": "FormatAgnostic"
+   }
+  ],
+  "outputs": [
+   {
+    "name": "is_finite",
+    "index": 0,
+    "dtype": "bool",
+    "format": "ND",
+    "paramType": "required",
+    "shape": [
+     -2
+    ],
+    "format_match_mode": "FormatAgnostic"
+   }
+  ],
+  "opMode": "dynamic",
+  "deterministic": "ignore"
+ },
+ "filePath": "ascend310p/bin/all_finite/AllFinite_6b5e50e30256d85838d6ce83514df20f.json"
+}

mindspore/lib/plugin/ascend/custom_ascendc_910b/op_impl/ai_core/tbe/kernel/ascend310p/all_finite/AllFinite_74e4ac02880d452e3308c94af273562e.json

@@ -0,0 +1,80 @@
+{
+ "binFileName": "AllFinite_74e4ac02880d452e3308c94af273562e",
+ "binFileSuffix": ".o",
+ "blockDim": -1,
+ "coreType": "AiCore",
+ "globalworkspace_spec_workspace": {
+  "size": 32,
+  "type": 0
+ },
+ "kernelName": "AllFinite_74e4ac02880d452e3308c94af273562e",
+ "magic": "RT_DEV_BINARY_MAGIC_ELF",
+ "memoryStamping": [],
+ "opParaSize": 64,
+ "parameters": [
+  null,
+  null,
+  null
+ ],
+ "sha256": "7a4d7e05d646d8bf8d4f2b83f2eae9c6289e4836a6e71d3114fff56c58fb7965",
+ "workspace": {
+  "num": 1,
+  "size": [
+   -1
+  ],
+  "type": [
+   0
+  ]
+ },
+ "kernelList": [
+  {
+   "kernelName": "AllFinite_74e4ac02880d452e3308c94af273562e_0"
+  }
+ ],
+ "optionalInputMode": "gen_placeholder",
+ "optionalOutputMode": "gen_placeholder",
+ "compileInfo": {},
+ "supportInfo": {
+  "implMode": "high_performance",
+  "int64Mode": false,
+  "simplifiedKeyMode": 0,
+  "simplifiedKey": [
+   "AllFinite/d=0,p=1/1,2/12,2",
+   "AllFinite/d=0,p=0/1,2/12,2",
+   "AllFinite/d=1,p=1/1,2/12,2",
+   "AllFinite/d=1,p=0/1,2/12,2"
+  ],
+  "optionalInputMode": "gen_placeholder",
+  "optionalOutputMode": "gen_placeholder",
+  "staticKey": "b03c3347996b5f3b8699b916889ab8e39e9bb822875d5430a15bd4b70cc62534,aaee380ee48b9c261d816baac1f6fc0b820ce3b2255e1eff19382df469b7e6de",
+  "inputs": [
+   {
+    "name": "gradient",
+    "index": 0,
+    "dtype": "float16",
+    "format": "ND",
+    "paramType": "required",
+    "shape": [
+     -2
+    ],
+    "format_match_mode": "FormatAgnostic"
+   }
+  ],
+  "outputs": [
+   {
+    "name": "is_finite",
+    "index": 0,
+    "dtype": "bool",
+    "format": "ND",
+    "paramType": "required",
+    "shape": [
+     -2
+    ],
+    "format_match_mode": "FormatAgnostic"
+   }
+  ],
+  "opMode": "dynamic",
+  "deterministic": "ignore"
+ },
+ "filePath": "ascend310p/bin/all_finite/AllFinite_74e4ac02880d452e3308c94af273562e.json"
+}

mindspore/lib/plugin/ascend/custom_ascendc_910b/op_impl/ai_core/tbe/kernel/ascend310p/decoder_kv_cache/DecoderKvCache_0d5520cc587ad44ce634bf3fbcffc272.json

@@ -0,0 +1,158 @@
+{
+ "binFileName": "DecoderKvCache_0d5520cc587ad44ce634bf3fbcffc272",
+ "binFileSuffix": ".o",
+ "blockDim": -1,
+ "coreType": "AiCore",
+ "globalworkspace_spec_workspace": {
+  "size": 32,
+  "type": 0
+ },
+ "kernelName": "DecoderKvCache_0d5520cc587ad44ce634bf3fbcffc272",
+ "magic": "RT_DEV_BINARY_MAGIC_ELF",
+ "memoryStamping": [],
+ "opParaSize": 120,
+ "parameters": [
+  null,
+  null,
+  null,
+  null,
+  null,
+  null,
+  null,
+  null,
+  null
+ ],
+ "sha256": "6e26b8da8644f2eea819e2316e8fcb7f7ef1bc11b7ced62754f449b20d6a515c",
+ "workspace": {
+  "num": 1,
+  "size": [
+   -1
+  ],
+  "type": [
+   0
+  ]
+ },
+ "kernelList": [
+  {
+   "kernelName": "DecoderKvCache_0d5520cc587ad44ce634bf3fbcffc272_1"
+  },
+  {
+   "kernelName": "DecoderKvCache_0d5520cc587ad44ce634bf3fbcffc272_2"
+  },
+  {
+   "kernelName": "DecoderKvCache_0d5520cc587ad44ce634bf3fbcffc272_4"
+  }
+ ],
+ "optionalInputMode": "gen_placeholder",
+ "optionalOutputMode": "gen_placeholder",
+ "compileInfo": {},
+ "supportInfo": {
+  "implMode": "high_performance",
+  "int64Mode": false,
+  "simplifiedKeyMode": 0,
+  "simplifiedKey": [
+   "DecoderKvCache/d=0,p=1/0,2/0,2/9,2/9,2/9,2/9,2/9,2/0,2",
+   "DecoderKvCache/d=0,p=0/0,2/0,2/9,2/9,2/9,2/9,2/9,2/0,2",
+   "DecoderKvCache/d=1,p=1/0,2/0,2/9,2/9,2/9,2/9,2/9,2/0,2",
+   "DecoderKvCache/d=1,p=0/0,2/0,2/9,2/9,2/9,2/9,2/9,2/0,2"
+  ],
+  "optionalInputMode": "gen_placeholder",
+  "optionalOutputMode": "gen_placeholder",
+  "staticKey": "d2856beaf2ce2851079a33584f85d8aef4bd9ee1145fdcf53bacd6d0035a1c4e,4a1f6ada9e32bab6f426997be4c2205d430f3e80f3e028831ec7146ba3e0f1f9",
+  "inputs": [
+   {
+    "name": "cache",
+    "index": 0,
+    "dtype": "float32",
+    "format": "ND",
+    "paramType": "required",
+    "shape": [
+     -2
+    ],
+    "format_match_mode": "FormatAgnostic"
+   },
+   {
+    "name": "update",
+    "index": 1,
+    "dtype": "float32",
+    "format": "ND",
+    "paramType": "required",
+    "shape": [
+     -2
+    ],
+    "format_match_mode": "FormatAgnostic"
+   },
+   {
+    "name": "valid_seq_len",
+    "index": 2,
+    "dtype": "int64",
+    "format": "ND",
+    "paramType": "required",
+    "shape": [
+     -2
+    ],
+    "format_match_mode": "FormatAgnostic"
+   },
+   {
+    "name": "batch_index",
+    "index": 3,
+    "dtype": "int64",
+    "format": "ND",
+    "paramType": "required",
+    "shape": [
+     -2
+    ],
+    "format_match_mode": "FormatAgnostic"
+   },
+   {
+    "name": "seq_len_axis",
+    "index": 4,
+    "dtype": "int64",
+    "format": "ND",
+    "paramType": "required",
+    "shape": [
+     -2
+    ],
+    "format_match_mode": "FormatAgnostic"
+   },
+   {
+    "name": "new_max_seq_len",
+    "index": 5,
+    "dtype": "int64",
+    "format": "ND",
+    "paramType": "required",
+    "shape": [
+     -2
+    ],
+    "format_match_mode": "FormatAgnostic"
+   },
+   {
+    "name": "cur_max_seq_len",
+    "index": 6,
+    "dtype": "int64",
+    "format": "ND",
+    "paramType": "required",
+    "shape": [
+     -2
+    ],
+    "format_match_mode": "FormatAgnostic"
+   }
+  ],
+  "outputs": [
+   {
+    "name": "out",
+    "index": 0,
+    "dtype": "float32",
+    "format": "ND",
+    "paramType": "required",
+    "shape": [
+     -2
+    ],
+    "format_match_mode": "FormatAgnostic"
+   }
+  ],
+  "opMode": "dynamic",
+  "deterministic": "ignore"
+ },
+ "filePath": "ascend310p/bin/decoder_kv_cache/DecoderKvCache_0d5520cc587ad44ce634bf3fbcffc272.json"
+}