dylanyunlong/project_6
1
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

feat(CRITICAL): 从 GitHub 扫描搬运 ixformer SDK + xllm 完整 GDN/MoE 代码

Browse Source 
来源:
  1. Chranos/ixformer (GitHub) → ixformer_sdk/ (230 files, 70K lines)
     - inference/functions/vllm.py: vllm_moe_topk_softmax 完整实现 (2033 lines)
     - inference/functions/moe.py: MoE ops 完整实现 (1380 lines)
     - contrib/vllm_flash_attn/: FA2 Python 接口 (1018 lines)
     - contrib/tgi/fused_moe.py: TGI fused MoE (429 lines)
     - csrc/include/ixformer/: C++ kernel headers + cmake

  2. Deep-Spark/xllm (GitHub) → upstream_ref/xllm_latest/ (+15 files)
     - npu_torch/qwen3_5_decoder_layer_impl.cpp/.h
     - npu_torch/qwen3_5_gated_delta_net.cpp/.h
     - npu_torch/qwen3_next_*.cpp/.h (6 files)
     - npu_torch/attention.cpp/.h + fused_moe.cpp/.h + CMakeLists.txt
     - models/llm/qwen3_5.h + qwen3_5_mtp.h + qwen3_next.h
     - models/vlm/qwen3_5.h

调用链完整性:
  ixformer_sdk/inference/functions/vllm.py
    → ops.infer.moe_topk_softmax() (C++ 层)
    → 这就是 base 镜像 libixformer.so 里的实现

  upstream_ref/xllm_latest/core/layers/ilu/fused_moe.cpp
    → ixformer::infer::topk_softmax() (直接 C++ 调用)
    → ixformer::infer::group_gemm() → 完整 7-step MoE pipeline
This commit is contained in:
project6-dev
2026-08-11 02:31:56 +00:00
parent a8b16da5da
commit 87a19d2d00
250 changed files with 76690 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

0
ixformer_sdk/inference/models/__init__.py Normal file
View File

1
ixformer_sdk/inference/models/clip/__init__.py Normal file
View File

@@ -0,0 +1 @@
from .modeling_clip import CLIPModel

503
ixformer_sdk/inference/models/clip/configuration_clip.py Normal file
View File

@@ -0,0 +1,503 @@
# coding=utf-8
# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
#
# 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.
""" CLIP model configuration"""
import copy
import os
from collections import OrderedDict
from typing import TYPE_CHECKING, Any, Mapping, Optional, Union
if TYPE_CHECKING:
from transformers.processing_utils import ProcessorMixin
from transformers.utils import TensorType
from transformers.configuration_utils import PretrainedConfig
from transformers.onnx import OnnxConfig
from transformers.utils import logging
logger = logging.get_logger(__name__)
CLIP_PRETRAINED_CONFIG_ARCHIVE_MAP = {
"openai/clip-vit-base-patch32": "https://huggingface.co/openai/clip-vit-base-patch32/resolve/main/config.json",
# See all CLIP models at https://huggingface.co/models?filter=clip
}
class CLIPTextConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`CLIPTextModel`]. It is used to instantiate a CLIP
text encoder according to the specified arguments, defining the model architecture. Instantiating a configuration
with the defaults will yield a similar configuration to that of the text encoder of the CLIP
[openai/clip-vit-base-patch32](https://huggingface.co/openai/clip-vit-base-patch32) architecture.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
vocab_size (`int`, *optional*, defaults to 49408):
Vocabulary size of the CLIP text model. Defines the number of different tokens that can be represented by
the `inputs_ids` passed when calling [`CLIPModel`].
hidden_size (`int`, *optional*, defaults to 512):
Dimensionality of the encoder layers and the pooler layer.
intermediate_size (`int`, *optional*, defaults to 2048):
Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
num_hidden_layers (`int`, *optional*, defaults to 12):
Number of hidden layers in the Transformer encoder.
num_attention_heads (`int`, *optional*, defaults to 8):
Number of attention heads for each attention layer in the Transformer encoder.
max_position_embeddings (`int`, *optional*, defaults to 77):
The maximum sequence length that this model might ever be used with. Typically set this to something large
just in case (e.g., 512 or 1024 or 2048).
hidden_act (`str` or `function`, *optional*, defaults to `"quick_gelu"`):
The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
`"relu"`, `"selu"` and `"gelu_new"` `"quick_gelu"` are supported.
layer_norm_eps (`float`, *optional*, defaults to 1e-5):
The epsilon used by the layer normalization layers.
attention_dropout (`float`, *optional*, defaults to 0.0):
The dropout ratio for the attention probabilities.
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
initializer_factor (`float`, *optional*, defaults to 1):
A factor for initializing all weight matrices (should be kept to 1, used internally for initialization
testing).
Example:
```python
>>> from transformers import CLIPTextConfig, CLIPTextModel
>>> # Initializing a CLIPTextConfig with openai/clip-vit-base-patch32 style configuration
>>> configuration = CLIPTextConfig()
>>> # Initializing a CLIPTextModel (with random weights) from the openai/clip-vit-base-patch32 style configuration
>>> model = CLIPTextModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```"""
model_type = "clip_text_model"
def __init__(
self,
vocab_size=49408,
hidden_size=512,
intermediate_size=2048,
projection_dim=512,
num_hidden_layers=12,
num_attention_heads=8,
max_position_embeddings=77,
hidden_act="quick_gelu",
layer_norm_eps=1e-5,
attention_dropout=0.0,
initializer_range=0.02,
initializer_factor=1.0,
pad_token_id=1,
bos_token_id=0,
eos_token_id=2,
**kwargs,
):
super().__init__(
pad_token_id=pad_token_id,
bos_token_id=bos_token_id,
eos_token_id=eos_token_id,
**kwargs,
)
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.intermediate_size = intermediate_size
self.projection_dim = projection_dim
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.max_position_embeddings = max_position_embeddings
self.layer_norm_eps = layer_norm_eps
self.hidden_act = hidden_act
self.initializer_range = initializer_range
self.initializer_factor = initializer_factor
self.attention_dropout = attention_dropout
@classmethod
def from_pretrained(
cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs
) -> "PretrainedConfig":
config_dict, kwargs = cls.get_config_dict(
pretrained_model_name_or_path, **kwargs
)
# get the text config dict if we are loading from CLIPConfig
if config_dict.get("model_type") == "clip":
config_dict = config_dict["text_config"]
if (
"model_type" in config_dict
and hasattr(cls, "model_type")
and config_dict["model_type"] != cls.model_type
):
logger.warning(
f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
)
return cls.from_dict(config_dict, **kwargs)
class CLIPVisionConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`CLIPVisionModel`]. It is used to instantiate a
CLIP vision encoder according to the specified arguments, defining the model architecture. Instantiating a
configuration with the defaults will yield a similar configuration to that of the vision encoder of the CLIP
[openai/clip-vit-base-patch32](https://huggingface.co/openai/clip-vit-base-patch32) architecture.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
hidden_size (`int`, *optional*, defaults to 768):
Dimensionality of the encoder layers and the pooler layer.
intermediate_size (`int`, *optional*, defaults to 3072):
Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
num_hidden_layers (`int`, *optional*, defaults to 12):
Number of hidden layers in the Transformer encoder.
num_attention_heads (`int`, *optional*, defaults to 12):
Number of attention heads for each attention layer in the Transformer encoder.
image_size (`int`, *optional*, defaults to 224):
The size (resolution) of each image.
patch_size (`int`, *optional*, defaults to 32):
The size (resolution) of each patch.
hidden_act (`str` or `function`, *optional*, defaults to `"quick_gelu"`):
The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
`"relu"`, `"selu"` and `"gelu_new"` ``"quick_gelu"` are supported.
layer_norm_eps (`float`, *optional*, defaults to 1e-5):
The epsilon used by the layer normalization layers.
attention_dropout (`float`, *optional*, defaults to 0.0):
The dropout ratio for the attention probabilities.
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
initializer_factor (`float`, *optional*, defaults to 1):
A factor for initializing all weight matrices (should be kept to 1, used internally for initialization
testing).
Example:
```python
>>> from transformers import CLIPVisionConfig, CLIPVisionModel
>>> # Initializing a CLIPVisionConfig with openai/clip-vit-base-patch32 style configuration
>>> configuration = CLIPVisionConfig()
>>> # Initializing a CLIPVisionModel (with random weights) from the openai/clip-vit-base-patch32 style configuration
>>> model = CLIPVisionModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```"""
model_type = "clip_vision_model"
def __init__(
self,
hidden_size=768,
intermediate_size=3072,
projection_dim=512,
num_hidden_layers=12,
num_attention_heads=12,
num_channels=3,
image_size=224,
patch_size=32,
hidden_act="quick_gelu",
layer_norm_eps=1e-5,
attention_dropout=0.0,
initializer_range=0.02,
initializer_factor=1.0,
**kwargs,
):
super().__init__(**kwargs)
self.hidden_size = hidden_size
self.intermediate_size = intermediate_size
self.projection_dim = projection_dim
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.num_channels = num_channels
self.patch_size = patch_size
self.image_size = image_size
self.initializer_range = initializer_range
self.initializer_factor = initializer_factor
self.attention_dropout = attention_dropout
self.layer_norm_eps = layer_norm_eps
self.hidden_act = hidden_act
@classmethod
def from_pretrained(
cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs
) -> "PretrainedConfig":
config_dict, kwargs = cls.get_config_dict(
pretrained_model_name_or_path, **kwargs
)
# get the vision config dict if we are loading from CLIPConfig
if config_dict.get("model_type") == "clip":
config_dict = config_dict["vision_config"]
if (
"model_type" in config_dict
and hasattr(cls, "model_type")
and config_dict["model_type"] != cls.model_type
):
logger.warning(
f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
)
return cls.from_dict(config_dict, **kwargs)
class CLIPConfig(PretrainedConfig):
r"""
[`CLIPConfig`] is the configuration class to store the configuration of a [`CLIPModel`]. It is used to instantiate
a CLIP model according to the specified arguments, defining the text model and vision model configs. Instantiating
a configuration with the defaults will yield a similar configuration to that of the CLIP
[openai/clip-vit-base-patch32](https://huggingface.co/openai/clip-vit-base-patch32) architecture.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
text_config (`dict`, *optional*):
Dictionary of configuration options used to initialize [`CLIPTextConfig`].
vision_config (`dict`, *optional*):
Dictionary of configuration options used to initialize [`CLIPVisionConfig`].
projection_dim (`int`, *optional*, defaults to 512):
Dimentionality of text and vision projection layers.
logit_scale_init_value (`float`, *optional*, defaults to 2.6592):
The inital value of the *logit_scale* paramter. Default is used as per the original CLIP implementation.
kwargs (*optional*):
Dictionary of keyword arguments.
Example:
```python
>>> from transformers import CLIPConfig, CLIPModel
>>> # Initializing a CLIPConfig with openai/clip-vit-base-patch32 style configuration
>>> configuration = CLIPConfig()
>>> # Initializing a CLIPModel (with random weights) from the openai/clip-vit-base-patch32 style configuration
>>> model = CLIPModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
>>> # We can also initialize a CLIPConfig from a CLIPTextConfig and a CLIPVisionConfig
>>> from transformers import CLIPTextConfig, CLIPVisionConfig
>>> # Initializing a CLIPText and CLIPVision configuration
>>> config_text = CLIPTextConfig()
>>> config_vision = CLIPVisionConfig()
>>> config = CLIPConfig.from_text_vision_configs(config_text, config_vision)
```"""
model_type = "clip"
is_composition = True
def __init__(
self,
text_config=None,
vision_config=None,
projection_dim=512,
logit_scale_init_value=2.6592,
**kwargs,
):
# If `_config_dict` exist, we use them for the backward compatibility.
# We pop out these 2 attributes before calling `super().__init__` to avoid them being saved (which causes a lot
# of confusion!).
text_config_dict = kwargs.pop("text_config_dict", None)
vision_config_dict = kwargs.pop("vision_config_dict", None)
super().__init__(**kwargs)
# Instead of simply assigning `[text|vision]_config_dict` to `[text|vision]_config`, we use the values in
# `[text|vision]_config_dict` to update the values in `[text|vision]_config`. The values should be same in most
# cases, but we don't want to break anything regarding `_config_dict` that existed before commit `8827e1b2`.
if text_config_dict is not None:
if text_config is None:
text_config = {}
# This is the complete result when using `text_config_dict`.
_text_config_dict = CLIPTextConfig(**text_config_dict).to_dict()
# Give a warning if the values exist in both `_text_config_dict` and `text_config` but being different.
for key, value in _text_config_dict.items():
if (
key in text_config
and value != text_config[key]
and key not in ["transformers_version"]
):
# If specified in `text_config_dict`
if key in text_config_dict:
message = (
f"`{key}` is found in both `text_config_dict` and `text_config` but with different values. "
f'The value `text_config_dict["{key}"]` will be used instead.'
)
# If inferred from default argument values (just to be super careful)
else:
message = (
f"`text_config_dict` is provided which will be used to initialize `CLIPTextConfig`. The "
f'value `text_config["{key}"]` will be overriden.'
)
logger.warning(message)
# Update all values in `text_config` with the ones in `_text_config_dict`.
text_config.update(_text_config_dict)
if vision_config_dict is not None:
if vision_config is None:
vision_config = {}
# This is the complete result when using `vision_config_dict`.
_vision_config_dict = CLIPVisionConfig(**vision_config_dict).to_dict()
# convert keys to string instead of integer
if "id2label" in _vision_config_dict:
_vision_config_dict["id2label"] = {
str(key): value
for key, value in _vision_config_dict["id2label"].items()
}
# Give a warning if the values exist in both `_vision_config_dict` and `vision_config` but being different.
for key, value in _vision_config_dict.items():
if (
key in vision_config
and value != vision_config[key]
and key not in ["transformers_version"]
):
# If specified in `vision_config_dict`
if key in vision_config_dict:
message = (
f"`{key}` is found in both `vision_config_dict` and `vision_config` but with different "
f'values. The value `vision_config_dict["{key}"]` will be used instead.'
)
# If inferred from default argument values (just to be super careful)
else:
message = (
f"`vision_config_dict` is provided which will be used to initialize `CLIPVisionConfig`. "
f'The value `vision_config["{key}"]` will be overriden.'
)
logger.warning(message)
# Update all values in `vision_config` with the ones in `_vision_config_dict`.
vision_config.update(_vision_config_dict)
if text_config is None:
text_config = {}
logger.info(
"`text_config` is `None`. Initializing the `CLIPTextConfig` with default values."
)
if vision_config is None:
vision_config = {}
logger.info(
"`vision_config` is `None`. initializing the `CLIPVisionConfig` with default values."
)
self.text_config = CLIPTextConfig(**text_config)
self.vision_config = CLIPVisionConfig(**vision_config)
self.projection_dim = projection_dim
self.logit_scale_init_value = logit_scale_init_value
self.initializer_factor = 1.0
@classmethod
def from_text_vision_configs(
cls, text_config: CLIPTextConfig, vision_config: CLIPVisionConfig, **kwargs
):
r"""
Instantiate a [`CLIPConfig`] (or a derived class) from clip text model configuration and clip vision model
configuration.
Returns:
[`CLIPConfig`]: An instance of a configuration object
"""
return cls(
text_config=text_config.to_dict(),
vision_config=vision_config.to_dict(),
**kwargs,
)
def to_dict(self):
"""
Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
Returns:
`Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
"""
output = copy.deepcopy(self.__dict__)
output["text_config"] = self.text_config.to_dict()
output["vision_config"] = self.vision_config.to_dict()
output["model_type"] = self.__class__.model_type
return output
class CLIPOnnxConfig(OnnxConfig):
@property
def inputs(self) -> Mapping[str, Mapping[int, str]]:
return OrderedDict(
[
("input_ids", {0: "batch", 1: "sequence"}),
(
"pixel_values",
{0: "batch", 1: "num_channels", 2: "height", 3: "width"},
),
("attention_mask", {0: "batch", 1: "sequence"}),
]
)
@property
def outputs(self) -> Mapping[str, Mapping[int, str]]:
return OrderedDict(
[
("logits_per_image", {0: "batch"}),
("logits_per_text", {0: "batch"}),
("text_embeds", {0: "batch"}),
("image_embeds", {0: "batch"}),
]
)
@property
def atol_for_validation(self) -> float:
return 1e-4
def generate_dummy_inputs(
self,
processor: "ProcessorMixin",
batch_size: int = -1,
seq_length: int = -1,
framework: Optional["TensorType"] = None,
) -> Mapping[str, Any]:
text_input_dict = super().generate_dummy_inputs(
processor.tokenizer,
batch_size=batch_size,
seq_length=seq_length,
framework=framework,
)
image_input_dict = super().generate_dummy_inputs(
processor.feature_extractor, batch_size=batch_size, framework=framework
)
return {**text_input_dict, **image_input_dict}
@property
def default_onnx_opset(self) -> int:
return 14

1578
ixformer_sdk/inference/models/clip/modeling_clip.py Normal file
View File

File diff suppressed because it is too large Load Diff

1
ixformer_sdk/inference/models/codeshell_7b_chat/__init__.py Normal file
View File

@@ -0,0 +1 @@
from .modeling_codeshell import CodeShellForCausalLM

153
ixformer_sdk/inference/models/codeshell_7b_chat/configuration_codeshell.py Normal file
View File

@@ -0,0 +1,153 @@
# coding=utf-8
# Copyright 2023 WisdomShell Inc. All Rights Reserved.
# 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.
# This code is based on Bigcode's GPTBigCode configuration. It has been modified from
# its original forms to accommodate minor architectural differences compared to
# GPTBigCode Configuration that trained the model.
# coding=utf-8
# Copyright 2023 The BigCode team and HuggingFace Inc. team.
#
# 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.
""" CodeShell configuration"""
from transformers.configuration_utils import PretrainedConfig
from transformers.utils import logging
logger = logging.get_logger(__name__)
class CodeShellConfig(PretrainedConfig):
"""
This is the configuration class to store the configuration of a [`CodeShellModel`]. It is used to instantiate a
CodeShell model according to the specified arguments, defining the model architecture.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
vocab_size (`int`, *optional*, defaults to 50257):
Vocabulary size of the GPT-2 model. Defines the number of different tokens that can be represented by the
`inputs_ids` passed when calling [`CodeShellModel`].
n_positions (`int`, *optional*, defaults to 1024):
The maximum sequence length that this model might ever be used with. Typically set this to something large
just in case (e.g., 512 or 1024 or 2048).
n_embd (`int`, *optional*, defaults to 768):
Dimensionality of the embeddings and hidden states.
n_layer (`int`, *optional*, defaults to 12):
Number of hidden layers in the Transformer encoder.
n_head (`int`, *optional*, defaults to 12):
Number of attention heads for each attention layer in the Transformer encoder.
n_inner (`int`, *optional*, defaults to None):
Dimensionality of the inner feed-forward layers. `None` will set it to 4 times n_embd
activation_function (`str`, *optional*, defaults to `"gelu_pytorch_tanh"`):
Activation function, to be selected in the list `["relu", "silu", "gelu", "tanh", "gelu_new",
"gelu_pytorch_tanh"]`.
resid_pdrop (`float`, *optional*, defaults to 0.1):
The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
embd_pdrop (`float`, *optional*, defaults to 0.1):
The dropout ratio for the embeddings.
attn_pdrop (`float`, *optional*, defaults to 0.1):
The dropout ratio for the attention.
layer_norm_epsilon (`float`, *optional*, defaults to 1e-5):
The epsilon to use in the layer normalization layers.
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
scale_attn_weights (`bool`, *optional*, defaults to `True`):
Scale attention weights by dividing by sqrt(hidden_size)..
use_cache (`bool`, *optional*, defaults to `True`):
Whether or not the model should return the last key/values attentions (not used by all models).
attention_softmax_in_fp32 (`bool`, *optional*, defaults to `True`):
Whether to call the fused softmax in float32.
scale_attention_softmax_in_fp32 (`bool`, *optional*, defaults to `True`):
Whether to scale the attention softmax in float32.
attention_type (`bool`, *optional*, defaults to `True`):
Whether to use Multi-Query Attion (`True`) or Multi-Head Attention (`False`).
"""
model_type = "codeshell"
keys_to_ignore_at_inference = ["past_key_values"]
attribute_map = {
"hidden_size": "n_embd",
"max_position_embeddings": "n_positions",
"num_attention_heads": "n_head",
"num_hidden_layers": "n_layer",
}
def __init__(
self,
vocab_size=70144,
n_positions=8192,
n_embd=4096,
n_layer=42,
n_head=32,
n_inner=None,
activation_function="gelu_pytorch_tanh",
resid_pdrop=0.1,
embd_pdrop=0.1,
attn_pdrop=0.1,
layer_norm_epsilon=1e-5,
initializer_range=0.02,
scale_attn_weights=True,
use_cache=True,
bos_token_id=70000,
eos_token_id=70000,
attention_softmax_in_fp32=True,
scale_attention_softmax_in_fp32=True,
group_query_attention=True,
num_query_groups=1,
position_embedding_type="learned_absolute",
rope_scaling=None,
**kwargs,
):
self.vocab_size = vocab_size
self.n_positions = n_positions
self.n_embd = n_embd
self.n_layer = n_layer
self.n_head = n_head
self.n_inner = n_inner
self.activation_function = activation_function
self.resid_pdrop = resid_pdrop
self.embd_pdrop = embd_pdrop
self.attn_pdrop = attn_pdrop
self.layer_norm_epsilon = layer_norm_epsilon
self.initializer_range = initializer_range
self.scale_attn_weights = scale_attn_weights
self.use_cache = use_cache
self.attention_softmax_in_fp32 = attention_softmax_in_fp32
self.scale_attention_softmax_in_fp32 = scale_attention_softmax_in_fp32
self.group_query_attention = group_query_attention
self.num_query_groups = num_query_groups
self.position_embedding_type = position_embedding_type
self.rope_scaling = rope_scaling
assert self.position_embedding_type in [
"learned_absolute",
"rope",
], "position_embedding_type must be one of ['learned_absolute', 'rope']"
self.bos_token_id = bos_token_id
self.eos_token_id = eos_token_id
super().__init__(bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)

1280
ixformer_sdk/inference/models/codeshell_7b_chat/modeling_codeshell.py Normal file
View File

File diff suppressed because it is too large Load Diff

99
ixformer_sdk/inference/models/codeshell_7b_chat/modeling_codeshell_ixformer.py Normal file
View File

@@ -0,0 +1,99 @@
import math
import ixformer.functions as ixf_F
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers.utils import logging
logger = logging.get_logger(__name__)
def mha(query, key, value, attention_mask):
if attention_mask is None and query.shape[2] == key.shape[2]:
context_layer = ixf_F.scaled_dot_product_attention(
query.contiguous(), key.contiguous(), value.contiguous(), is_causal=True
)
else:
# if attention_mask is not None:
# # attention_mask = attention_mask
# attention_mask = (~attention_mask).cuda().float()*(-10000)
context_layer = ixf_F.scaled_dot_product_attention(
query.contiguous(), key.contiguous(), value.contiguous(), attention_mask
)
# context_layer = context_layer.transpose(1, 2).contiguous()
# res_shape = list(context_layer.shape)
# res_shape = res_shape[:2] + [-1]
# context_layer = context_layer.view(*res_shape)
return context_layer
# batch_size, head_num, seq_len, head_dim = query.shape
# src_len = query.shape[-2]
# tgt_len = key.shape[-2]
# if attention_mask is None and src_len == tgt_len:
# attention_mask = ~torch.tril(torch.ones([src_len, tgt_len])).bool()
# elif attention_mask is None:
# attention_mask = torch.zeros([src_len, tgt_len])
# attention_mask = attention_mask.cuda().int()
# attention_scores = ixf_F.act_bias_mm(
# query, key, scale=1 / math.sqrt(head_dim), trans_format="TN"
# )
# # softmax
# # if tgt_len > 2048:
# # if not (attention_mask == 0).all():
# # attention_scores.masked_fill_(attention_mask.bool(), -10000.0)
# # dtype = attention_scores.dtype
# # attention_probs = F.softmax(attention_scores.float(), dim=-1)
# # attention_probs = attention_probs.type(dtype)
# # else:
# # raise NotImplementedError()
# attention_probs = ixf_F.attention_masked_softmax(
# attention_scores, attention_mask.int()
# )
# # s * v
# # batch_size,head_num,seq_len,head_dim
# context_layer = ixf_F.act_bias_mm(
# attention_probs, value, trans_format="NN")
# context_layer = context_layer.transpose(1, 2).contiguous()
# context_layer = context_layer.view(
# batch_size, seq_len, head_num * head_dim)
def mlp(mlp_input, ff1_weight, ff1_bias, ff2_weight):
input_shape = list(mlp_input.shape)
mlp_input = mlp_input.view(-1, input_shape[-1])
mlp_output = ixf_F.act_bias_mm(
mlp_input, ff1_weight, ff1_bias, scale=1, act_type="gelu", trans_format="TN"
)
mlp_output = ixf_F.linear(mlp_output, ff2_weight, None)
input_shape[-1] = -1
mlp_output = mlp_output.view(*input_shape)
return mlp_output
def mlp_forward(self, hidden_states):
# [s, b, 4hp]
# intermediate_parallel = self.dense_h_to_4h(hidden_states)
# intermediate_parallel = self.activation_func(intermediate_parallel)
input_shape = list(hidden_states.shape)
hidden_states = hidden_states.view(-1, input_shape[-1])
mlp_output = ixf_F.act_bias_mm(
hidden_states,
self.c_fc.weight,
self.c_fc.bias,
scale=1,
act_type="gelu",
trans_format="TN",
)
if isinstance(self.c_proj, nn.Linear):
output = ixf_F.linear(
mlp_output,
self.c_proj.weight,
self.c_proj.bias,
)
else:
output = self.c_proj(mlp_output)
output = output.view(*input_shape)
return output