mindspore 2.4.10__cp310-cp310-manylinux1_x86_64.whl → 2.5.0__cp310-cp310-manylinux1_x86_64.whl

mindspore/device_manager.py

@@ -0,0 +1,134 @@
+# Copyright 2024 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+
+"""Device manager interfaces."""
+
+import os
+from mindspore import log as logger
+from mindspore._c_expression import DeviceManagerConf, DeviceContextManager, MSContext, CollectiveManager
+from mindspore._checkparam import args_type_check
+from mindspore.parallel._ps_context import _need_reset_device_target_for_ps
+
+__all__ = ['set_device', 'set_deterministic']
+
+
+@args_type_check(device_target=str, device_id=int)
+def set_device(device_target, device_id=None):
+    """
+    Set device target and device id for running environment.
+
+    Note:
+        - The `device_target` must be set in the ["CPU", "GPU", "Ascend"], there is no default value.
+        - Suggest setting `device_target` and `device_id` before calling :func:`mindspore.communication.init`.
+
+    Args:
+        device_target (str): The target device to run, only support "Ascend", "GPU", and "CPU".
+        device_id (int): ID of the target device, the value must be in [0, device_num_per_host-1].
+            The frame will set different default behaviours according to the scenario:
+            if it is a single-card scenario, the frame will be set to 0.
+            In a distributed scenario where msrun is started, the framework will
+            automatically negotiate the available device_id values.
+            In a distributed scenario with other startup methods, the frame is set to 0.
+            "device_num_per_host" refers to the total number of devices on the host.
+
+    Examples:
+        >>> import mindspore as ms
+        >>> ms.set_device("Ascend", 1)
+    """
+    valid_targets = ["CPU", "GPU", "Ascend"]
+    if device_target not in valid_targets:
+        raise ValueError(f"The argument 'device_target' must be one of {valid_targets}, but got {device_target}.")
+    # If in Parameter Server mode, Ascend card should not be used by server and scheduler.
+    if _need_reset_device_target_for_ps(device_target):
+        logger.info("Reset device target to CPU when set_device.")
+        device_target = "CPU"
+
+    is_default = False
+    if device_id is None:
+        device_id = 0
+        is_default = True
+    if device_id < 0:
+        raise ValueError("The device id must bigger than or equal to 0.")
+
+    MSContext.get_instance().set_device_target_inner(device_target)
+
+    if DeviceManagerConf.get_instance().is_device_enable():
+        old_device_target = DeviceManagerConf.get_instance().get_device_target()
+        old_device_id = DeviceManagerConf.get_instance().get_device_id()
+        if old_device_target != device_target or old_device_id != device_id:
+            raise RuntimeError("The 'mindspore.set_device' can not be modified.")
+        return
+
+    device_context = DeviceContextManager.get_instance().get_device_context(device_target)
+    if device_context is not None and device_context.initialized():
+        raise RuntimeError("The runtime has been initialized, please set it before the kernel is executed, "
+                           "or before calling 'mindspore.communication.init()'. "
+                           "Suggest setting it as early as possible.")
+    DeviceManagerConf.get_instance().set_device(device_target, device_id, is_default)
+
+
+@args_type_check(deterministic=bool)
+def set_deterministic(deterministic):
+    """
+    Enables or disables deterministic computing.
+
+    When deterministic computing is enabled, the same output is generated if an operator is executed
+    for multiple times with the same hardware and input.This often slows down operator execution.
+    In distributed scenario, we suggest user to set deterministic mode before
+    calling :func:`mindspore.communication.init` to enable deterministic operation for
+    communication operators in the global communication group.
+
+    The framework not enabled deterministic computation by default.
+
+    Args:
+        deterministic (bool): Whether to enable deterministic computing.
+
+    Examples:
+        >>> import mindspore as ms
+        >>> ms.set_deterministic(True)
+    """
+    # Check the configuration environment whether valid.
+    if DeviceManagerConf.get_instance().is_deterministic_configured():
+        raise RuntimeError("The 'mindspore.set_deterministic' can not be set repeatedly.")
+
+    # Must wait for all async created groups to be initialized so that
+    # deterministic feature could be consistent between all processes.
+    CollectiveManager.get_instance().wait_all_comm_init()
+
+    # Check the hccl_deterministic and te_parallel_compiler.
+    hccl_deterministic = os.getenv("HCCL_DETERMINISTIC")
+    te_parallel_compiler = os.getenv("TE_PARALLEL_COMPILER")
+    if deterministic:
+        if hccl_deterministic and hccl_deterministic != "true":
+            logger.warning(f"Environment 'HCCL_DETERMINISTIC' should be 'true' when set deterministic='True', but "
+                           f"got '{hccl_deterministic}'. 'HCCL_DETERMINISTIC' will be set to 'true'.")
+        if te_parallel_compiler and te_parallel_compiler != "1":
+            logger.warning(f"Environment 'TE_PARALLEL_COMPILER' should be '1' when set deterministic='True', but "
+                           f"got '{te_parallel_compiler}'. 'TE_PARALLEL_COMPILER' will be set to '1'.")
+        os.environ["HCCL_DETERMINISTIC"] = "true"
+        os.environ["TE_PARALLEL_COMPILER"] = "1"
+    else:
+        if hccl_deterministic and hccl_deterministic != "false":
+            logger.warning(f"Environment 'HCCL_DETERMINISTIC' should not be set or be 'false' when set "
+                           f"deterministic='False', but got '{hccl_deterministic}'. 'HCCL_DETERMINISTIC' "
+                           f"will be unset.")
+            del os.environ["HCCL_DETERMINISTIC"]
+        if te_parallel_compiler and te_parallel_compiler != "0":
+            logger.warning(f"Environment 'TE_PARALLEL_COMPILER' should not be set or be '0' when set "
+                           f"deterministic='False', but got '{te_parallel_compiler}'. 'TE_PARALLEL_COMPILER' "
+                           f"will be unset.")
+            del os.environ["TE_PARALLEL_COMPILER"]
+
+    DeviceManagerConf.get_instance().set_deterministic(deterministic)

mindspore/experimental/llm_boost/__init__.py

@@ -16,6 +16,7 @@
 from __future__ import absolute_import
 
 from mindspore.experimental.llm_boost.atb import LlamaBoost, QwenBoost
+from mindspore.experimental.llm_boost.ascend_native import *
 from mindspore.experimental.llm_boost.register import LlmBoostRegister
 
 __all__ = ["LlmBoostRegister"]

mindspore/experimental/llm_boost/ascend_native/__init__.py

@@ -0,0 +1,22 @@
+# Copyright 2024 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+"""
+Provide llm boost for inference, such as LlamaBoost.
+"""
+from __future__ import absolute_import
+
+from mindspore.experimental.llm_boost.ascend_native.llama_boost_ascend_native import LlamaBoostAscendNative
+
+__all__ = ['LlamaBoostAscendNative']

mindspore/experimental/llm_boost/ascend_native/llama_boost_ascend_native.py

@@ -0,0 +1,211 @@
+# Copyright 2024 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+"""AscendNative Llama Boost APIs."""
+
+import os
+import numpy as np
+from mindspore.common import Tensor, dtype
+from mindspore.experimental.llm_boost.ascend_native.llm_boost import LLMBoost
+from mindspore.experimental.llm_boost.register import LlmBoostRegister, LlmBoostType
+
+
+def RoundUp(val: int, align: int) -> int:
+    if align == 0:
+        return 0
+    return -(val // -align) * align
+
+
+def ConvertTensor(nd_mat: np.ndarray, transpose: bool = True, nd2nz: bool = True) -> np.ndarray:
+    """ Transforms tensor format from Nd to Nz """
+    if transpose:
+        nd_mat = np.transpose(nd_mat)
+    if not nd2nz:
+        return nd_mat
+    block_size = (16, 16)
+    r = RoundUp(nd_mat.shape[0], block_size[0])
+    c = RoundUp(nd_mat.shape[1], block_size[1])
+    r_pad = r - nd_mat.shape[0]
+    c_pad = c - nd_mat.shape[1]
+    nd_mat = np.pad(nd_mat, ((0, r_pad), (0, c_pad)))
+    nz_mat = np.transpose(np.reshape(
+        nd_mat, (r, c // block_size[1], block_size[1])), (1, 0, 2))
+    nz_mat = nz_mat.reshape(r, c)
+    return nz_mat
+
+
+@LlmBoostRegister.register(LlmBoostType.ASCEND_NATIVE, "Llama")
+class LlamaBoostAscendNative(LLMBoost):
+    r"""
+    Implements an Llama model in a single kernel.
+    it forwards the python functions to the C++ binded object
+    """
+    def _get_from_dict(self, dictionary, name):
+        """ internal function to get a specific tensor from the dictionary """
+        all_relevant_layers = [value for key, value in dictionary.items() if name in key]
+        if all_relevant_layers:
+            return all_relevant_layers[0].asnumpy()
+        return None
+
+    def _get_quant_triplet_from_dict(self, dictionary, name):
+        """ internal function to get a weight triple tensor from the dictionary """
+        weights = self._get_from_dict(dictionary, name + "._handler.weight")
+        scale = self._get_from_dict(dictionary, name + "._weight_quantizer.scale")
+        offset = self._get_from_dict(dictionary, name + "._weight_quantizer.zp_neg")
+        return weights, scale, offset
+
+    def _prepare_single_layer(self, ckpt, config, id):
+        """ prepares the dictionary of weights of a single layer """
+        prefix = 'model.layers.' + str(id)
+        is_last = (id == config.num_layers-1)
+        layer = 'layers.' + str(id) + '.'
+        l_dict = {key: value for key, value in ckpt.items() if layer in key}
+        if config.n_kv_heads is None:
+            config.n_kv_heads = config.num_heads
+        start = 0
+        end = config.hidden_size
+        kv_start = 0
+        kv_end = int(config.hidden_size*config.n_kv_heads/config.num_heads)
+        ffn_hid = [value for key, value in l_dict.items() if "w3" in key][0].shape[0]
+        ffn_start = 0
+        ffn_end = ffn_hid
+        rank_size = int(os.getenv('RANK_SIZE', '1'))
+        #Emir if (config.parallel_mode != 2): # 2 - AUTO_PARALLEL
+        hid_size = end
+        kv_hid_size = kv_end
+        embed_size = config.vocab_size
+        rank_id = int(os.getenv('RANK_ID', '0'))
+        if (hid_size % rank_size == 0) and (ffn_hid % rank_size == 0) and (embed_size % rank_size == 0):
+            start = int(rank_id * hid_size / rank_size)
+            end = int((rank_id + 1) * hid_size / rank_size)
+            kv_start = int(rank_id * kv_hid_size / rank_size)
+            kv_end = int((rank_id + 1) * kv_hid_size / rank_size)
+            ffn_start = int(rank_id * ffn_hid / rank_size)
+            ffn_end = int((rank_id + 1) * ffn_hid / rank_size)
+        else:
+            raise RuntimeError("hidden size and ffn hidden size must be divided by rank size without remainder.  \
+                                hidden_size: ", hid_size, " ffn_hidden_size: ", ffn_hid, " rank_size: ", rank_size)
+        quant = (self._get_from_dict(l_dict, "_weight_quantizer") is not None)
+        unite_qkv = (config.num_heads == config.n_kv_heads)
+        self.dictionary[prefix + ".attention_norm.weight"] = \
+            Tensor(self._get_from_dict(l_dict, "attention_norm"), dtype=dtype.float16)
+        self.dictionary[prefix + ".ffn_norm.weight"] = \
+            Tensor(self._get_from_dict(l_dict, "ffn_norm"), dtype=dtype.float16)
+        if is_last:
+            self.dictionary['lm_head.weight'] = Tensor(ConvertTensor(ckpt['lm_head.weight'].asnumpy()[:, start:end]))
+
+        if not quant:
+            self._pack_attn_weights(l_dict, prefix, start, end, kv_start, kv_end, unite_qkv)
+            self._pack_ffn_weights(l_dict, prefix, ffn_start, ffn_end)
+        else:
+            self._pack_attn_quant_weights(l_dict, prefix, start, end, kv_start, kv_end, unite_qkv)
+            self._pack_ffn_quant_weights(l_dict, prefix, ffn_start, ffn_end)
+
+    def _pack_attn_weights(self, l_dict, prefix, start, end, kv_start, kv_end, unite_qkv):
+        """ prepares the dictionary of weights of an attention block """
+        wq = self._get_from_dict(l_dict, "wq")[start:end, :]
+        wk = self._get_from_dict(l_dict, "wk")[kv_start:kv_end, :]
+        wv = self._get_from_dict(l_dict, "wv")[kv_start:kv_end, :]
+        self.dictionary[prefix + ".attention.wo.weight"] = \
+            Tensor(ConvertTensor(self._get_from_dict(l_dict, "wo")[:, start:end]))
+        if unite_qkv:
+            self.dictionary[prefix + ".attention.wqkv.weight"] = Tensor(ConvertTensor(np.concatenate((wq, wk, wv))))
+        else:
+            self.dictionary[prefix + ".attention.wq.weight"] = Tensor(ConvertTensor(wq))
+            self.dictionary[prefix + ".attention.wkv.weight"] = Tensor(ConvertTensor(np.concatenate((wk, wv))))
+
+    def _pack_ffn_weights(self, l_dict, prefix, ffn_start, ffn_end):
+        """ prepares the dictionary of weights of an ffn block """
+        self.dictionary[prefix + ".feed_forward.w2.weight"] = \
+            Tensor(ConvertTensor(self._get_from_dict(l_dict, "w2")[:, ffn_start:ffn_end]))
+        w1 = self._get_from_dict(l_dict, "w1")[ffn_start:ffn_end, :]
+        w3 = self._get_from_dict(l_dict, "w3")[ffn_start:ffn_end, :]
+        self.dictionary[prefix + ".feed_forward.w13.weight"] = Tensor(ConvertTensor(np.concatenate((w1, w3))))
+
+    def _pack_attn_quant_weights(self, l_dict, prefix, start, end, kv_start, kv_end, unite_qkv):
+        """ prepares the dictionary of weights of a quantized attention block """
+        wq, wq_scale, wq_offset = self._get_quant_triplet_from_dict(l_dict, "wq")
+        wk, wk_scale, wk_offset = self._get_quant_triplet_from_dict(l_dict, "wk")
+        wv, wv_scale, wv_offset = self._get_quant_triplet_from_dict(l_dict, "wv")
+        wo, wo_scale, wo_offset = self._get_quant_triplet_from_dict(l_dict, "wo")
+        self.dictionary[prefix + ".attention.wo.weight"] = Tensor(ConvertTensor(wo[:, start:end], nd2nz=False))
+        self.dictionary[prefix + ".attention.wo.weight.scale"] = Tensor(wo_scale[start:end])
+        self.dictionary[prefix + ".attention.wo.weight.offset"] = Tensor(wo_offset[start:end])
+
+        if unite_qkv:
+            self.dictionary[prefix + ".attention.wqkv.weight"] = \
+             Tensor(ConvertTensor(np.concatenate((wq[start:end, :], wk[kv_start:kv_end, :], wv[kv_start:kv_end, :])),
+                                  nd2nz=False))
+            self.dictionary[prefix + ".attention.wqkv.weight.scale"] = \
+                Tensor(np.concatenate((wq_scale[start:end], wk_scale[kv_start:kv_end], wv_scale[kv_start:kv_end])))
+            self.dictionary[prefix + ".attention.wqkv.weight.offset"] = \
+                Tensor(np.concatenate((wq_offset[start:end], wk_offset[kv_start:kv_end], wv_offset[kv_start:kv_end])))
+        else:
+            self.dictionary[prefix + ".attention.wq.weight"] = Tensor(ConvertTensor(wq[start:end, :], nd2nz=False))
+            self.dictionary[prefix + ".attention.wq.weight.scale"] = Tensor(wq_scale[start:end])
+            self.dictionary[prefix + ".attention.wq.weight.offset"] = Tensor(wq_offset[start:end])
+            self.dictionary[prefix + ".attention.wkv.weight"] = \
+                Tensor(ConvertTensor(np.concatenate((wk[kv_start:kv_end, :], wv[kv_start:kv_end, :])), nd2nz=False))
+            self.dictionary[prefix + ".attention.wkv.weight.scale"] = \
+                Tensor(np.concatenate((wk_scale[kv_start:kv_end], wv_scale[kv_start:kv_end])))
+            self.dictionary[prefix + ".attention.wkv.weight.offset"] = \
+                Tensor(np.concatenate((wk_offset[kv_start:kv_end], wv_offset[kv_start:kv_end])))
+
+    def _pack_ffn_quant_weights(self, l_dict, prefix, ffn_start, ffn_end):
+        """ prepares the dictionary of weights of a quantized ffn block """
+        w1, w1_scale, w1_offset = self._get_quant_triplet_from_dict(l_dict, "w1")
+        w2, w2_scale, w2_offset = self._get_quant_triplet_from_dict(l_dict, "w2")
+        w3, w3_scale, w3_offset = self._get_quant_triplet_from_dict(l_dict, "w3")
+        self.dictionary[prefix + ".feed_forward.w2.weight"] = Tensor(ConvertTensor(w2[:, ffn_start:ffn_end],
+                                                                                   nd2nz=False))
+        self.dictionary[prefix + ".feed_forward.w2.weight.scale"] = Tensor(w2_scale[ffn_start:ffn_end])
+        self.dictionary[prefix + ".feed_forward.w2.weight.offset"] = Tensor(w2_offset[ffn_start:ffn_end])
+
+        self.dictionary[prefix + ".feed_forward.w13.weight"] = \
+                Tensor(ConvertTensor(np.concatenate((w1[ffn_start:ffn_end, :], w3[ffn_start:ffn_end, :])), nd2nz=False))
+        self.dictionary[prefix + ".feed_forward.w13.weight.scale"] = \
+                Tensor(np.concatenate((w1_scale[ffn_start:ffn_end], w3_scale[ffn_start:ffn_end])))
+        self.dictionary[prefix + ".feed_forward.w13.weight.offset"] = \
+            Tensor(np.concatenate((w1_offset[ffn_start:ffn_end], w3_offset[ffn_start:ffn_end])))
+
+    def _prepare_cos_sin_arrays(self, config, theta=10000):
+        """ prepares the cosine and sine arrays """
+        head_dim = config.hidden_size // config.num_heads
+        max_position_embedding = \
+            config.max_position_embedding if config.max_position_embedding is not None else config.seq_length
+        freqs_base = np.arange(0, head_dim, 2)[: (head_dim // 2)].astype(np.float32)
+        freqs = 1.0 / (theta ** (freqs_base / head_dim))
+        t = np.arange(0, max_position_embedding, 1).astype(np.float32)
+        freqs = np.outer(t, freqs)
+        emb = np.concatenate((freqs, freqs), axis=-1)
+        freqs_cos = Tensor(np.cos(emb), dtype=dtype.float16)
+        sin = np.sin(emb)
+
+        sin[:, :int(emb.shape[1]/2)] = -sin[:, :int(emb.shape[1]/2)]
+        self.dictionary['model.cos.weight'] = freqs_cos
+        freqs_sin = Tensor(sin, dtype=dtype.float16)
+        self.dictionary['model.sin.weight'] = freqs_sin
+
+    def set_weights(self, ckpt_dict):
+        """ load the checkpoint """
+        self.dictionary = {}
+        self.dictionary['model.tok_embeddings.embedding_weight'] = \
+            Tensor(ckpt_dict['model.tok_embeddings.embedding_weight'].asnumpy())
+        self.dictionary['model.norm_out.weight'] = \
+            Tensor(ckpt_dict['model.norm_out.weight'].asnumpy(), dtype=dtype.float16)
+        self._prepare_cos_sin_arrays(self.config)
+        for layer_id in range(self.config.num_layers):
+            self._prepare_single_layer(ckpt_dict, self.config, layer_id)
+
+        self.binder.set_weights_map(self.dictionary)

mindspore/experimental/llm_boost/ascend_native/llm_boost.py

@@ -0,0 +1,52 @@
+# Copyright 2024 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+"""LLMBoost APIs."""
+
+from mindspore.common import Tensor
+
+class LLMBoost():
+    r"""
+    Implements an LLM in a single kernel.
+    it forwards the python function to the C++ binded object
+    """
+    def __init__(self, config):
+        r"""
+        initialize the parameters of the llm binder.
+        config is simply the config object of the model
+        """
+        from mindspore._c_expression import LlmBoostBinder
+        self.config = config
+        self.binder = LlmBoostBinder("AscendNative", config.model_type)
+        self.binder.init_model(config.to_dict())
+
+    def init(self):
+        """
+        Initialize the object
+        returns True if object needs input manipulation by mindformers
+        """
+        return False
+
+    def set_kvcache(self, k_caches=None, v_caches=None):
+        return
+
+    def forward(self, input_ids, batch_valid_length, position_ids=None):
+        ret = self.binder.forward([input_ids, batch_valid_length], "nothing really")
+        return Tensor(ret[0])
+
+    def set_weights(self, ckpt_dict):
+        self.binder.set_weights_map(ckpt_dict)
+
+    def add_flags(self, is_first_iteration=False):
+        self.binder.add_flags(is_first_iteration=is_first_iteration)

mindspore/experimental/llm_boost/atb/boost_base.py

@@ -112,8 +112,7 @@ class AtbBoostBase:
 
     def _convert_qkv_concat_weight(self, param_dict):
         """convert qkv concat weight"""
-        assume_num_layers = 500
-        for i in range(assume_num_layers):
+        for i in range(self.num_layers):
             # qkv weight concat
             wq_weight_name = f"model.layers.{i}.attention.wq.weight"
             wk_weight_name = f"model.layers.{i}.attention.wk.weight"
@@ -151,7 +150,7 @@ class AtbBoostBase:
             logger.info(f"transform: {qkv_concat_weight_name}")
             logger.info(f"transform: {gate_hidden_concat_weight_name}")
 
-        for i in range(assume_num_layers):
+        for i in range(self.num_layers):
             # qkv bias concat
             wq_bias_name = f"model.layers.{i}.attention.wq.bias"
             wk_bias_name = f"model.layers.{i}.attention.wk.bias"

mindspore/experimental/llm_boost/atb/llama_boost.py

@@ -43,7 +43,11 @@ class LlamaBoost(AtbBoostBase):
         )
 
     def init(self):
-        """set param"""
+        """
+        Initialize the object
+        returns True if object needs input manipulation by mindformers
+        """
+
         coder_param = {
             "normEps": self.config.rms_norm_eps,
             "normType": NormType.RMS_NORM,
@@ -93,6 +97,7 @@ class LlamaBoost(AtbBoostBase):
         }
         self.atb_encoder_operation.init(json.dumps({**encoder_param}))
         self.atb_decoder_operation.init(json.dumps({**decoder_param}))
+        return True
 
     def _prepare_inputs(
             self,

mindspore/experimental/llm_boost/register.py

@@ -23,6 +23,7 @@ class LlmBoostType:
         pass
 
     BUILDIN = 'BuildIn'
+    ASCEND_NATIVE = 'LLMBoost'
 
 
 class LlmBoostRegister:

mindspore/experimental/optim/adadelta.py

@@ -37,28 +37,32 @@ class Adadelta(Optimizer):
     Implements Adadelta algorithm.
 
     .. math::
-       \begin{aligned}
-            &\rule{150mm}{0.4pt}                                                                 \\
-            &\textbf{input}      : \gamma \text{ (lr)}, \: \theta_0 \text{ (params)},
-                \: f(\theta) \text{ (objective)}, \: \rho \text{ (decay)},
-                \: \lambda \text{ (weight decay)}                                                \\
-            &\textbf{initialize} :  v_0  \leftarrow 0 \: \text{ (square avg)},
-                \: u_0 \leftarrow 0 \: \text{ (accumulate variables)}                     \\[-1.ex]
-            &\rule{110mm}{0.4pt}                                                                 \\
-            &\textbf{for} \: t=1 \: \textbf{to} \: \ldots \: \textbf{do}                         \\
-            &\hspace{5mm}g_t           \leftarrow   \nabla_{\theta} f_t (\theta_{t-1})           \\
-            &\hspace{5mm}if \: \lambda \neq 0                                                    \\
-            &\hspace{10mm} g_t \leftarrow g_t + \lambda  \theta_{t-1}                            \\
-            &\hspace{5mm} v_t      \leftarrow v_{t-1} \rho + g^2_t (1 - \rho)                    \\
-            &\hspace{5mm}\Delta x_t    \leftarrow   \frac{\sqrt{u_{t-1} +
-                \epsilon }}{ \sqrt{v_t + \epsilon}  }g_t \hspace{21mm}                           \\
-            &\hspace{5mm} u_t  \leftarrow   u_{t-1}  \rho +
-                 \Delta x^2_t  (1 - \rho)                                                        \\
-            &\hspace{5mm}\theta_t      \leftarrow   \theta_{t-1} - \gamma  \Delta x_t            \\
-            &\rule{110mm}{0.4pt}                                                          \\[-1.ex]
-            &\bf{return} \:  \theta_t                                                     \\[-1.ex]
-            &\rule{110mm}{0.4pt}                                                          \\[-1.ex]
-       \end{aligned}
+        \newcommand{\grad}[2]{\nabla_{#1} f_{#2}(#2_{#2 - 1})}
+        \newcommand{\updateVar}[3]{#1_{#2} \leftarrow #1_{#2 - 1} \rho + #3_{#2} (1 - \rho)}
+
+        \begin{align*}
+            &\rule{150mm}{0.4pt} \\
+            &\textbf{Input}:
+                \gamma \text{ (lr)}, \: \theta_0 \text{ (params)}, \: f(\theta) \text{ (objective)},
+                \: \rho \text{ (decay)}, \: \lambda \text{ (weight decay)} \\
+            &\textbf{Initialize}:
+                \begin{cases}
+                    v_0 \leftarrow 0 \text{ (square avg)} \\
+                    u_0 \leftarrow 0 \text{ (accumulate variables)}
+                \end{cases} \\
+            &\rule{110mm}{0.4pt} \\
+            &\textbf{For } t = 1 \text{ to } \ldots \text{ do}: \\
+            &\quad g_t \leftarrow \grad{\theta}{t} \\
+            &\quad \text{If } \lambda \neq 0: \\
+            &\quad\quad g_t \leftarrow g_t + \lambda \theta_{t - 1} \\
+            &\quad v_t \leftarrow \updateVar{v}{t}{g^2} \\
+            &\quad \Delta x_t \leftarrow \frac{\sqrt{u_{t - 1} + \epsilon}}{\sqrt{v_t + \epsilon}} g_t \\
+            &\quad u_t \leftarrow \updateVar{u}{t}{\Delta x^2} \\
+            &\quad \theta_t \leftarrow \theta_{t - 1} - \gamma \Delta x_t \\
+            &\rule{110mm}{0.4pt} \\
+            &\bf{Return}: \theta_t \\
+            &\rule{110mm}{0.4pt}
+        \end{align*}
 
     .. warning::
         This is an experimental optimizer API that is subject to change.

mindspore/experimental/optim/adam.py

@@ -78,6 +78,9 @@ class Adam(Optimizer):
        \end{aligned}
 
     .. warning::
+        The implementation formula of this optimizer interface is not completely consistent with that in the paper.
+        If you want to use an interface that is completely consistent, it is recommended to use
+        :class:`mindspore.mint.optim.Adam`, which currently only supports Ascend.
         This is an experimental optimizer API that is subject to change.
         This module must be used with lr scheduler module in `LRScheduler Class
         <https://www.mindspore.cn/docs/en/master/api_python/mindspore.nn.html#learningrateschedule-class>`_ .