Upload folder using huggingface_hub
Browse files- .gitattributes +1 -0
- chat_template.jinja +77 -0
- config.json +96 -0
- configuration_nemotron_h.py +245 -0
- generation_config.json +14 -0
- model-00001-of-00002.safetensors +3 -0
- model-00002-of-00002.safetensors +3 -0
- model.safetensors.index.json +649 -0
- modeling_nemotron_h.py +1639 -0
- recipe.yaml +10 -0
- special_tokens_map.json +1026 -0
- tokenizer.json +3 -0
- tokenizer_config.json +0 -0
.gitattributes
CHANGED
|
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
+
tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
chat_template.jinja
ADDED
|
@@ -0,0 +1,77 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{# ───── header (system message) ───── #}
|
| 2 |
+
{{- "<|im_start|>system\n" -}}
|
| 3 |
+
{# ───── get the custom instructions + optional thinking override ───── #}
|
| 4 |
+
{%- if messages[0].role == "system" -%}
|
| 5 |
+
{%- set custom_instructions = messages[0].content.rstrip() -%}
|
| 6 |
+
{%- endif -%}
|
| 7 |
+
{# ───── set the system prompt ───── #}
|
| 8 |
+
{%- if custom_instructions -%}
|
| 9 |
+
{{- custom_instructions -}}
|
| 10 |
+
{%- else -%}
|
| 11 |
+
{{- "You are a helpful AI assistant named Dicta-LM 3.0, Trained by Dicta, the Israel Center for Text Analysis. Your role is to provide accurate, helpful, and well-structured responses to user questions and requests.\nProvide clear, logical, and precise answers that thoroughly address what the user is asking for. Structure your responses in a way that is easy to understand and follow." -}}
|
| 12 |
+
{%- endif -%}
|
| 13 |
+
{%- set ns = namespace(multi_step_tool=true, last_query_index=messages | length - 1) -%}
|
| 14 |
+
{%- for message in messages[::-1] -%}
|
| 15 |
+
{%- set index = messages | length - 1 - loop.index0 -%}
|
| 16 |
+
{%- if ns.multi_step_tool and message.role == "user" and message.content is string and not (message.content.startswith("<tool_response>") and message.content.endswith("</tool_response>")) -%}
|
| 17 |
+
{%- set ns.multi_step_tool = false -%}
|
| 18 |
+
{%- set ns.last_query_index = index -%}
|
| 19 |
+
{%- endif -%}
|
| 20 |
+
{%- endfor -%}
|
| 21 |
+
{%- if tools -%}
|
| 22 |
+
{{- "\n\n## Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" -}}
|
| 23 |
+
{%- for tool in tools -%}
|
| 24 |
+
{{- "\n" -}}
|
| 25 |
+
{{- tool | tojson -}}
|
| 26 |
+
{%- endfor -%}
|
| 27 |
+
{{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call>" -}}
|
| 28 |
+
{%- endif -%}
|
| 29 |
+
{{- "<|im_end|>" -}}
|
| 30 |
+
{%- for message in messages -%}
|
| 31 |
+
{%- if message.content is string -%}
|
| 32 |
+
{%- set content = message.content -%}
|
| 33 |
+
{%- else -%}
|
| 34 |
+
{%- set content = "" -%}
|
| 35 |
+
{%- endif -%}
|
| 36 |
+
{%- if message.role == "user" or message.role == "system" and not loop.first -%}
|
| 37 |
+
{{- "\n<|im_start|>" + message.role + "\n" + content + "<|im_end|>" -}}
|
| 38 |
+
{%- elif message.role == "assistant" -%}
|
| 39 |
+
{% generation %}
|
| 40 |
+
{{- "\n<|im_start|>" + message.role + "\n" + content -}}
|
| 41 |
+
{%- if message.tool_calls -%}
|
| 42 |
+
{%- for tool_call in message.tool_calls -%}
|
| 43 |
+
{%- if loop.first and content or not loop.first -%}
|
| 44 |
+
{{- "\n" -}}
|
| 45 |
+
{%- endif -%}
|
| 46 |
+
{%- if tool_call.function -%}
|
| 47 |
+
{%- set tool_call = tool_call.function -%}
|
| 48 |
+
{%- endif -%}
|
| 49 |
+
{{- "<tool_call>\n{\"name\": \"" -}}
|
| 50 |
+
{{- tool_call.name -}}
|
| 51 |
+
{{- "\", \"arguments\": " -}}
|
| 52 |
+
{%- if tool_call.arguments is string -%}
|
| 53 |
+
{{- tool_call.arguments -}}
|
| 54 |
+
{%- else -%}
|
| 55 |
+
{{- tool_call.arguments | tojson -}}
|
| 56 |
+
{%- endif -%}
|
| 57 |
+
{{- "}\n</tool_call>" -}}
|
| 58 |
+
{%- endfor -%}
|
| 59 |
+
{%- endif -%}
|
| 60 |
+
{{- "<|im_end|>" -}}
|
| 61 |
+
{% endgeneration %}
|
| 62 |
+
{%- elif message.role == "tool" -%}
|
| 63 |
+
{%- if loop.first or messages[loop.index0 - 1].role != "tool" -%}
|
| 64 |
+
{{- "\n<|im_start|>user" -}}
|
| 65 |
+
{%- endif -%}
|
| 66 |
+
{{- "\n<tool_response>\n" -}}
|
| 67 |
+
{{- content -}}
|
| 68 |
+
{{- "\n</tool_response>" -}}
|
| 69 |
+
{%- if loop.last or messages[loop.index0 + 1].role != "tool" -%}
|
| 70 |
+
{{- "<|im_end|>" -}}
|
| 71 |
+
{%- endif -%}
|
| 72 |
+
{%- endif -%}
|
| 73 |
+
{%- endfor -%}
|
| 74 |
+
{{- "\n" -}}
|
| 75 |
+
{%- if add_generation_prompt -%}
|
| 76 |
+
{{- "<|im_start|>assistant\n" -}}
|
| 77 |
+
{%- endif -%}
|
config.json
ADDED
|
@@ -0,0 +1,96 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"NemotronHForCausalLM"
|
| 4 |
+
],
|
| 5 |
+
"attention_bias": false,
|
| 6 |
+
"attention_dropout": 0.0,
|
| 7 |
+
"auto_map": {
|
| 8 |
+
"AutoConfig": "configuration_nemotron_h.NemotronHConfig",
|
| 9 |
+
"AutoModelForCausalLM": "modeling_nemotron_h.NemotronHForCausalLM"
|
| 10 |
+
},
|
| 11 |
+
"bos_token_id": 1,
|
| 12 |
+
"chunk_size": 128,
|
| 13 |
+
"conv_kernel": 4,
|
| 14 |
+
"dtype": "bfloat16",
|
| 15 |
+
"eos_token_id": 2,
|
| 16 |
+
"expand": 2,
|
| 17 |
+
"head_dim": 128,
|
| 18 |
+
"hidden_dropout": 0.0,
|
| 19 |
+
"hidden_size": 5120,
|
| 20 |
+
"hybrid_override_pattern": "M-M-M-M*-M-M-M-M*-M-M-M-M*-M-M-M-M*-M-M-M-M*-M-M-M-M*-M-M-M-M-",
|
| 21 |
+
"initializer_range": 0.02,
|
| 22 |
+
"intermediate_size": 20480,
|
| 23 |
+
"layer_norm_epsilon": 1e-05,
|
| 24 |
+
"mamba_head_dim": 80,
|
| 25 |
+
"mamba_hidden_act": "silu",
|
| 26 |
+
"mamba_num_groups": 8,
|
| 27 |
+
"mamba_num_heads": 128,
|
| 28 |
+
"mamba_proj_bias": false,
|
| 29 |
+
"mamba_state_dim": 128,
|
| 30 |
+
"max_position_embeddings": 131072,
|
| 31 |
+
"mlp_bias": false,
|
| 32 |
+
"mlp_hidden_act": "relu2",
|
| 33 |
+
"model_type": "nemotron_h",
|
| 34 |
+
"n_groups": 8,
|
| 35 |
+
"num_attention_heads": 40,
|
| 36 |
+
"num_hidden_layers": 62,
|
| 37 |
+
"num_key_value_heads": 8,
|
| 38 |
+
"num_logits_to_keep": 1,
|
| 39 |
+
"num_query_groups": 8,
|
| 40 |
+
"pad_token_id": 0,
|
| 41 |
+
"quantization_config": {
|
| 42 |
+
"config_groups": {
|
| 43 |
+
"group_0": {
|
| 44 |
+
"format": "pack-quantized",
|
| 45 |
+
"input_activations": null,
|
| 46 |
+
"output_activations": null,
|
| 47 |
+
"targets": [
|
| 48 |
+
"Linear"
|
| 49 |
+
],
|
| 50 |
+
"weights": {
|
| 51 |
+
"actorder": "static",
|
| 52 |
+
"block_structure": null,
|
| 53 |
+
"dynamic": false,
|
| 54 |
+
"group_size": 128,
|
| 55 |
+
"num_bits": 4,
|
| 56 |
+
"observer": "minmax",
|
| 57 |
+
"observer_kwargs": {},
|
| 58 |
+
"strategy": "group",
|
| 59 |
+
"symmetric": true,
|
| 60 |
+
"type": "int"
|
| 61 |
+
}
|
| 62 |
+
}
|
| 63 |
+
},
|
| 64 |
+
"format": "pack-quantized",
|
| 65 |
+
"global_compression_ratio": null,
|
| 66 |
+
"ignore": [
|
| 67 |
+
"lm_head"
|
| 68 |
+
],
|
| 69 |
+
"kv_cache_scheme": null,
|
| 70 |
+
"quant_method": "compressed-tensors",
|
| 71 |
+
"quantization_status": "compressed",
|
| 72 |
+
"sparsity_config": {},
|
| 73 |
+
"transform_config": {},
|
| 74 |
+
"version": "0.12.2"
|
| 75 |
+
},
|
| 76 |
+
"rescale_prenorm_residual": true,
|
| 77 |
+
"residual_in_fp32": false,
|
| 78 |
+
"rms_norm_eps": 1e-05,
|
| 79 |
+
"sliding_window": null,
|
| 80 |
+
"ssm_state_size": 128,
|
| 81 |
+
"tie_word_embeddings": false,
|
| 82 |
+
"time_step_floor": 0.0001,
|
| 83 |
+
"time_step_limit": [
|
| 84 |
+
0.0,
|
| 85 |
+
Infinity
|
| 86 |
+
],
|
| 87 |
+
"time_step_max": 0.1,
|
| 88 |
+
"time_step_min": 0.001,
|
| 89 |
+
"time_step_rank": 256,
|
| 90 |
+
"transformers_version": "4.57.3",
|
| 91 |
+
"use_bias": false,
|
| 92 |
+
"use_cache": true,
|
| 93 |
+
"use_conv_bias": true,
|
| 94 |
+
"use_mamba_kernels": true,
|
| 95 |
+
"vocab_size": 131072
|
| 96 |
+
}
|
configuration_nemotron_h.py
ADDED
|
@@ -0,0 +1,245 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2024 AI21 Labs Ltd. and the HuggingFace Inc. team. All rights reserved.
|
| 3 |
+
# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
|
| 4 |
+
#
|
| 5 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 6 |
+
# you may not use this file except in compliance with the License.
|
| 7 |
+
# You may obtain a copy of the License at
|
| 8 |
+
#
|
| 9 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 10 |
+
#
|
| 11 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 12 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 13 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 14 |
+
# See the License for the specific language governing permissions and
|
| 15 |
+
# limitations under the License.
|
| 16 |
+
"""NemotronH model configuration"""
|
| 17 |
+
|
| 18 |
+
import re
|
| 19 |
+
|
| 20 |
+
from transformers.configuration_utils import PretrainedConfig
|
| 21 |
+
from transformers.utils import logging
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
logger = logging.get_logger(__name__)
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
class NemotronHConfig(PretrainedConfig):
|
| 28 |
+
r"""
|
| 29 |
+
This is the configuration class to store the configuration of a [`NemotronHModel`]. It is used to instantiate a
|
| 30 |
+
NemotronH model according to the specified arguments, defining the model architecture. Instantiating a configuration
|
| 31 |
+
with the defaults will yield a similar configuration to that of the NemotronH-v0.1 model.
|
| 32 |
+
|
| 33 |
+
[todo](todo)
|
| 34 |
+
|
| 35 |
+
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
| 36 |
+
documentation from [`PretrainedConfig`] for more information.
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
Args:
|
| 40 |
+
vocab_size (`int`, *optional*, defaults to 131072):
|
| 41 |
+
Vocabulary size of the NemotronH model. Defines the number of different tokens that can be represented by the
|
| 42 |
+
`inputs_ids` passed when calling [`NemotronHModel`]
|
| 43 |
+
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
|
| 44 |
+
Whether the model's input and output word embeddings should be tied. Note that this is only relevant if the
|
| 45 |
+
model has a output word embedding layer.
|
| 46 |
+
hidden_size (`int`, *optional*, defaults to 4096):
|
| 47 |
+
Dimension of the hidden representations.
|
| 48 |
+
intermediate_size (`int`, *optional*, defaults to 21504):
|
| 49 |
+
Dimension of the MLP representations.
|
| 50 |
+
num_hidden_layers (`int`, *optional*, defaults to 52):
|
| 51 |
+
Number of hidden layers in the Transformer encoder.
|
| 52 |
+
hybrid_override_pattern (`str`, *optional*, defaults to `"M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M-"`):
|
| 53 |
+
The pattern of the hybrid model. The pattern is a string of characters where each character represents M: Mamba2, *: Attention, -: MLP
|
| 54 |
+
num_attention_heads (`int`, *optional*, defaults to 32):
|
| 55 |
+
Number of attention heads for each attention layer in the Transformer encoder.
|
| 56 |
+
attention_head_dim (`int`, *optional*, defaults to 128):
|
| 57 |
+
Dimension of each attention head.
|
| 58 |
+
num_key_value_heads (`int`, *optional*, defaults to 8):
|
| 59 |
+
This is the number of key_value heads that should be used to implement Grouped Query Attention. If
|
| 60 |
+
`num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
|
| 61 |
+
`num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used.
|
| 62 |
+
mlp_hidden_act (`str`, *optional*, defaults to "relu2"):
|
| 63 |
+
The non-linear activation function in the MLP layers.
|
| 64 |
+
attention_bias (`bool`, *optional*, defaults to `False`):
|
| 65 |
+
Whether to use bias in attention layers.
|
| 66 |
+
mlp_bias (`bool`, *optional*, defaults to `False`):
|
| 67 |
+
Whether to use bias in MLP layers.
|
| 68 |
+
use_bias (`bool`, *optional*, defaults to `False`):
|
| 69 |
+
Whether to use bias in the model.
|
| 70 |
+
initializer_range (`float`, *optional*, defaults to 0.02):
|
| 71 |
+
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
|
| 72 |
+
layer_norm_epsilon (`float`, *optional*, defaults to 1e-5):
|
| 73 |
+
The epsilon used by the layer normalization layers.
|
| 74 |
+
residual_in_fp32 (`bool`, *optional*, defaults to `False`):
|
| 75 |
+
Whether or not residuals should be in `float32`. If set to `False` residuals will keep the same `dtype` as the rest of the model.
|
| 76 |
+
use_cache (`bool`, *optional*, defaults to `True`):
|
| 77 |
+
Whether or not the model should return the last key/values attentions (not used by all models). Only
|
| 78 |
+
relevant if `config.is_decoder=True`.
|
| 79 |
+
num_logits_to_keep (`int` or `None`, *optional*, defaults to 1):
|
| 80 |
+
Number of prompt logits to calculate during generation. If `None`, all logits will be calculated. If an
|
| 81 |
+
integer value, only last `num_logits_to_keep` logits will be calculated.
|
| 82 |
+
pad_token_id (`int`, *optional*, defaults to 0):
|
| 83 |
+
The id of the padding token.
|
| 84 |
+
bos_token_id (`int`, *optional*, defaults to 1):
|
| 85 |
+
The id of the "beginning-of-sequence" token.
|
| 86 |
+
eos_token_id (`int`, *optional*, defaults to 2):
|
| 87 |
+
The id of the "end-of-sequence" token.
|
| 88 |
+
sliding_window (`int`, *optional*, defaults to None):
|
| 89 |
+
Sliding window attention window size.
|
| 90 |
+
max_position_embeddings (`int`, *optional*, defaults to 4096):
|
| 91 |
+
The maximum sequence length that this model might ever be used with.
|
| 92 |
+
attention_dropout (`float`, *optional*, defaults to 0.0):
|
| 93 |
+
The dropout ratio for the attention probabilities.
|
| 94 |
+
hidden_dropout (`float`, *optional*, defaults to 0.0):
|
| 95 |
+
The dropout ratio for the hidden states.
|
| 96 |
+
use_mamba_kernels (`bool`, *optional*, defaults to `True`):
|
| 97 |
+
Flag indicating whether or not to use the fast mamba kernels. These are available only if `mamba-ssm` and
|
| 98 |
+
`causal-conv1d` are installed, and the mamba modules are running on a CUDA device.
|
| 99 |
+
ssm_state_size (`int`, *optional*, defaults to 128):
|
| 100 |
+
The dimension of the mamba state space latents.
|
| 101 |
+
mamba_num_heads (`int`, *optional*, defaults to 128):
|
| 102 |
+
Number of heads in Mamba layers.
|
| 103 |
+
mamba_n_groups (`int`, *optional*, defaults to 8):
|
| 104 |
+
Number of groups in Mamba layers.
|
| 105 |
+
mamba_head_dim (`int`, *optional*, defaults to 64):
|
| 106 |
+
Dimension of each Mamba head.
|
| 107 |
+
mamba_d_conv (`int`, *optional*, defaults to 4):
|
| 108 |
+
The size of the mamba convolution kernel.
|
| 109 |
+
mamba_expand (`int`, *optional*, defaults to 2):
|
| 110 |
+
Expanding factor used to determine the mamba intermediate size.
|
| 111 |
+
mamba_hidden_act (`str`, *optional*, defaults to "silu"):
|
| 112 |
+
The non-linear activation function in the Mamba layers.
|
| 113 |
+
mamba_dt_min (`float`, *optional*, defaults to 0.001):
|
| 114 |
+
Minimum value for the time step in Mamba.
|
| 115 |
+
mamba_dt_max (`float`, *optional*, defaults to 0.1):
|
| 116 |
+
Maximum value for the time step in Mamba.
|
| 117 |
+
mamba_dt_limit (`tuple`, *optional*, defaults to (0.0, float("inf"))):
|
| 118 |
+
Limits for the time step in Mamba.
|
| 119 |
+
mamba_dt_init_floor (`float`, *optional*, defaults to 1e-4):
|
| 120 |
+
Floor value for time step initialization in Mamba.
|
| 121 |
+
mamba_conv_bias (`bool`, *optional*, defaults to `True`):
|
| 122 |
+
Whether to use bias in the convolution layer of the mamba mixer block.
|
| 123 |
+
mamba_proj_bias (`bool`, *optional*, defaults to `False`):
|
| 124 |
+
Whether to use bias in the input and output projections of the mamba mixer block.
|
| 125 |
+
mamba_chunk_size (`int`, *optional*, defaults to 256):
|
| 126 |
+
Size of chunks for Mamba processing.
|
| 127 |
+
rescale_prenorm_residual (`bool`, *optional*, defaults to `True`):
|
| 128 |
+
Whether to rescale the pre-normalization residual connections.
|
| 129 |
+
"""
|
| 130 |
+
|
| 131 |
+
model_type = "nemotron_h"
|
| 132 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
| 133 |
+
|
| 134 |
+
def __init__(
|
| 135 |
+
self,
|
| 136 |
+
vocab_size=131072,
|
| 137 |
+
tie_word_embeddings=False,
|
| 138 |
+
hidden_size=4096,
|
| 139 |
+
intermediate_size=21504,
|
| 140 |
+
num_hidden_layers=52,
|
| 141 |
+
hybrid_override_pattern="M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M*-M-M-M-M-M-",
|
| 142 |
+
num_attention_heads=32,
|
| 143 |
+
#attention_head_dim=128,
|
| 144 |
+
head_dim=128,
|
| 145 |
+
num_key_value_heads=8, # nemo: num_query_groups
|
| 146 |
+
mlp_hidden_act="relu2",
|
| 147 |
+
attention_bias=False,
|
| 148 |
+
mlp_bias=False,
|
| 149 |
+
use_bias=False,
|
| 150 |
+
initializer_range=0.02, # nemo: init_method_std
|
| 151 |
+
layer_norm_epsilon=1e-5, # nemo: layernorm_epsilon
|
| 152 |
+
residual_in_fp32=False, # Megatron Core default value
|
| 153 |
+
use_cache=True,
|
| 154 |
+
num_logits_to_keep=1,
|
| 155 |
+
pad_token_id=0,
|
| 156 |
+
bos_token_id=1,
|
| 157 |
+
eos_token_id=2,
|
| 158 |
+
sliding_window=None,
|
| 159 |
+
max_position_embeddings=4096,
|
| 160 |
+
attention_dropout=0.0,
|
| 161 |
+
hidden_dropout=0.0, # * ADDED
|
| 162 |
+
use_mamba_kernels=True,
|
| 163 |
+
ssm_state_size=128, # mamba_state_size
|
| 164 |
+
mamba_num_heads=128,
|
| 165 |
+
mamba_n_groups=8, # nemo: mamba_ssm_ngroups = num_heads
|
| 166 |
+
mamba_head_dim=64,
|
| 167 |
+
mamba_d_conv=4,
|
| 168 |
+
mamba_expand=2,
|
| 169 |
+
mamba_hidden_act="silu",
|
| 170 |
+
mamba_dt_min=0.001,
|
| 171 |
+
mamba_dt_max=0.1,
|
| 172 |
+
mamba_dt_limit=(0.0, float("inf")),
|
| 173 |
+
mamba_dt_init_floor=1e-4,
|
| 174 |
+
mamba_conv_bias=True,
|
| 175 |
+
mamba_proj_bias=False,
|
| 176 |
+
mamba_chunk_size=256,
|
| 177 |
+
rescale_prenorm_residual=True,
|
| 178 |
+
**kwargs,
|
| 179 |
+
):
|
| 180 |
+
self.vocab_size = vocab_size
|
| 181 |
+
self.tie_word_embeddings = tie_word_embeddings
|
| 182 |
+
self.hidden_size = hidden_size
|
| 183 |
+
self.intermediate_size = intermediate_size
|
| 184 |
+
self.num_hidden_layers = num_hidden_layers
|
| 185 |
+
self.hybrid_override_pattern = hybrid_override_pattern
|
| 186 |
+
self.num_attention_heads = num_attention_heads
|
| 187 |
+
#self.attention_head_dim = attention_head_dim
|
| 188 |
+
self.head_dim = head_dim
|
| 189 |
+
self.sliding_window = sliding_window
|
| 190 |
+
self.max_position_embeddings = max_position_embeddings
|
| 191 |
+
self.attention_dropout = attention_dropout
|
| 192 |
+
self.hidden_dropout = hidden_dropout
|
| 193 |
+
|
| 194 |
+
# Validate hybrid_override_pattern
|
| 195 |
+
# M: Mamba2, *: Attention, -: MLP
|
| 196 |
+
assert len(self.hybrid_override_pattern) == self.num_hidden_layers, "hybrid_override_pattern must have the same length as num_hidden_layers"
|
| 197 |
+
assert re.match(r"^[*-M]+$", self.hybrid_override_pattern), "hybrid_override_pattern must only contain characters 'M', '*', or '-'"
|
| 198 |
+
|
| 199 |
+
# for backward compatibility
|
| 200 |
+
if num_key_value_heads is None:
|
| 201 |
+
num_key_value_heads = num_attention_heads
|
| 202 |
+
|
| 203 |
+
self.num_key_value_heads = num_key_value_heads
|
| 204 |
+
self.mlp_hidden_act = mlp_hidden_act
|
| 205 |
+
self.attention_bias = attention_bias
|
| 206 |
+
self.mlp_bias = mlp_bias
|
| 207 |
+
self.use_bias = use_bias
|
| 208 |
+
self.initializer_range = initializer_range
|
| 209 |
+
self.layer_norm_epsilon = layer_norm_epsilon
|
| 210 |
+
self.residual_in_fp32 = residual_in_fp32
|
| 211 |
+
|
| 212 |
+
self.use_cache = use_cache
|
| 213 |
+
self.num_logits_to_keep = num_logits_to_keep
|
| 214 |
+
|
| 215 |
+
self.use_mamba_kernels = use_mamba_kernels
|
| 216 |
+
self.n_groups = mamba_n_groups
|
| 217 |
+
self.mamba_head_dim = mamba_head_dim
|
| 218 |
+
self.ssm_state_size = ssm_state_size
|
| 219 |
+
self.mamba_num_heads = mamba_num_heads
|
| 220 |
+
self.conv_kernel = mamba_d_conv
|
| 221 |
+
self.expand = mamba_expand
|
| 222 |
+
self.mamba_hidden_act = mamba_hidden_act
|
| 223 |
+
self.time_step_min = mamba_dt_min
|
| 224 |
+
self.time_step_max = mamba_dt_max
|
| 225 |
+
self.time_step_limit = mamba_dt_limit
|
| 226 |
+
self.time_step_floor = mamba_dt_init_floor
|
| 227 |
+
self.use_conv_bias = mamba_conv_bias
|
| 228 |
+
self.mamba_proj_bias = mamba_proj_bias
|
| 229 |
+
self.chunk_size = mamba_chunk_size
|
| 230 |
+
self.rescale_prenorm_residual = rescale_prenorm_residual
|
| 231 |
+
|
| 232 |
+
super().__init__(
|
| 233 |
+
pad_token_id=pad_token_id,
|
| 234 |
+
bos_token_id=bos_token_id,
|
| 235 |
+
eos_token_id=eos_token_id,
|
| 236 |
+
tie_word_embeddings=tie_word_embeddings,
|
| 237 |
+
**kwargs,
|
| 238 |
+
)
|
| 239 |
+
|
| 240 |
+
@property
|
| 241 |
+
def layers_block_type(self):
|
| 242 |
+
return [
|
| 243 |
+
"mamba" if self.hybrid_override_pattern[i] == "M" else
|
| 244 |
+
"attention" if self.hybrid_override_pattern[i] == "*" else "mlp"
|
| 245 |
+
for i in range(self.num_hidden_layers)]
|
generation_config.json
ADDED
|
@@ -0,0 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_from_model_config": true,
|
| 3 |
+
"bos_token_id": 1,
|
| 4 |
+
"do_sample": true,
|
| 5 |
+
"eos_token_id": [
|
| 6 |
+
2,
|
| 7 |
+
21
|
| 8 |
+
],
|
| 9 |
+
"pad_token_id": 0,
|
| 10 |
+
"temperature": 0.6,
|
| 11 |
+
"top_k": 20,
|
| 12 |
+
"top_p": 0.95,
|
| 13 |
+
"transformers_version": "4.57.3"
|
| 14 |
+
}
|
model-00001-of-00002.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:764d1dd8d943254b4da6d9038626057874c71e242020fb75de3f6f596fb9ab7d
|
| 3 |
+
size 4984762136
|
model-00002-of-00002.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c33ac220721182f4fc5fa54b80ece9180e43be793dbe59d02e6efa6817059658
|
| 3 |
+
size 3358424896
|
model.safetensors.index.json
ADDED
|
@@ -0,0 +1,649 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"metadata": {
|
| 3 |
+
"total_parameters": 2800871184,
|
| 4 |
+
"total_size": 8343114880
|
| 5 |
+
},
|
| 6 |
+
"weight_map": {
|
| 7 |
+
"backbone.embeddings.weight": "model-00001-of-00002.safetensors",
|
| 8 |
+
"backbone.layers.0.mixer.A_log": "model-00001-of-00002.safetensors",
|
| 9 |
+
"backbone.layers.0.mixer.D": "model-00001-of-00002.safetensors",
|
| 10 |
+
"backbone.layers.0.mixer.conv1d.bias": "model-00001-of-00002.safetensors",
|
| 11 |
+
"backbone.layers.0.mixer.conv1d.weight": "model-00001-of-00002.safetensors",
|
| 12 |
+
"backbone.layers.0.mixer.dt_bias": "model-00001-of-00002.safetensors",
|
| 13 |
+
"backbone.layers.0.mixer.in_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 14 |
+
"backbone.layers.0.mixer.in_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 15 |
+
"backbone.layers.0.mixer.in_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 16 |
+
"backbone.layers.0.mixer.norm.weight": "model-00001-of-00002.safetensors",
|
| 17 |
+
"backbone.layers.0.mixer.out_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 18 |
+
"backbone.layers.0.mixer.out_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 19 |
+
"backbone.layers.0.mixer.out_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 20 |
+
"backbone.layers.0.norm.weight": "model-00001-of-00002.safetensors",
|
| 21 |
+
"backbone.layers.1.mixer.down_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 22 |
+
"backbone.layers.1.mixer.down_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 23 |
+
"backbone.layers.1.mixer.down_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 24 |
+
"backbone.layers.1.mixer.up_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 25 |
+
"backbone.layers.1.mixer.up_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 26 |
+
"backbone.layers.1.mixer.up_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 27 |
+
"backbone.layers.1.norm.weight": "model-00001-of-00002.safetensors",
|
| 28 |
+
"backbone.layers.10.mixer.down_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 29 |
+
"backbone.layers.10.mixer.down_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 30 |
+
"backbone.layers.10.mixer.down_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 31 |
+
"backbone.layers.10.mixer.up_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 32 |
+
"backbone.layers.10.mixer.up_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 33 |
+
"backbone.layers.10.mixer.up_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 34 |
+
"backbone.layers.10.norm.weight": "model-00001-of-00002.safetensors",
|
| 35 |
+
"backbone.layers.11.mixer.A_log": "model-00001-of-00002.safetensors",
|
| 36 |
+
"backbone.layers.11.mixer.D": "model-00001-of-00002.safetensors",
|
| 37 |
+
"backbone.layers.11.mixer.conv1d.bias": "model-00001-of-00002.safetensors",
|
| 38 |
+
"backbone.layers.11.mixer.conv1d.weight": "model-00001-of-00002.safetensors",
|
| 39 |
+
"backbone.layers.11.mixer.dt_bias": "model-00001-of-00002.safetensors",
|
| 40 |
+
"backbone.layers.11.mixer.in_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 41 |
+
"backbone.layers.11.mixer.in_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 42 |
+
"backbone.layers.11.mixer.in_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 43 |
+
"backbone.layers.11.mixer.norm.weight": "model-00001-of-00002.safetensors",
|
| 44 |
+
"backbone.layers.11.mixer.out_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 45 |
+
"backbone.layers.11.mixer.out_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 46 |
+
"backbone.layers.11.mixer.out_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 47 |
+
"backbone.layers.11.norm.weight": "model-00001-of-00002.safetensors",
|
| 48 |
+
"backbone.layers.12.mixer.down_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 49 |
+
"backbone.layers.12.mixer.down_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 50 |
+
"backbone.layers.12.mixer.down_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 51 |
+
"backbone.layers.12.mixer.up_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 52 |
+
"backbone.layers.12.mixer.up_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 53 |
+
"backbone.layers.12.mixer.up_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 54 |
+
"backbone.layers.12.norm.weight": "model-00001-of-00002.safetensors",
|
| 55 |
+
"backbone.layers.13.mixer.A_log": "model-00001-of-00002.safetensors",
|
| 56 |
+
"backbone.layers.13.mixer.D": "model-00001-of-00002.safetensors",
|
| 57 |
+
"backbone.layers.13.mixer.conv1d.bias": "model-00001-of-00002.safetensors",
|
| 58 |
+
"backbone.layers.13.mixer.conv1d.weight": "model-00001-of-00002.safetensors",
|
| 59 |
+
"backbone.layers.13.mixer.dt_bias": "model-00001-of-00002.safetensors",
|
| 60 |
+
"backbone.layers.13.mixer.in_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 61 |
+
"backbone.layers.13.mixer.in_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 62 |
+
"backbone.layers.13.mixer.in_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 63 |
+
"backbone.layers.13.mixer.norm.weight": "model-00001-of-00002.safetensors",
|
| 64 |
+
"backbone.layers.13.mixer.out_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 65 |
+
"backbone.layers.13.mixer.out_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 66 |
+
"backbone.layers.13.mixer.out_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 67 |
+
"backbone.layers.13.norm.weight": "model-00001-of-00002.safetensors",
|
| 68 |
+
"backbone.layers.14.mixer.down_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 69 |
+
"backbone.layers.14.mixer.down_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 70 |
+
"backbone.layers.14.mixer.down_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 71 |
+
"backbone.layers.14.mixer.up_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 72 |
+
"backbone.layers.14.mixer.up_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 73 |
+
"backbone.layers.14.mixer.up_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 74 |
+
"backbone.layers.14.norm.weight": "model-00001-of-00002.safetensors",
|
| 75 |
+
"backbone.layers.15.mixer.A_log": "model-00001-of-00002.safetensors",
|
| 76 |
+
"backbone.layers.15.mixer.D": "model-00001-of-00002.safetensors",
|
| 77 |
+
"backbone.layers.15.mixer.conv1d.bias": "model-00001-of-00002.safetensors",
|
| 78 |
+
"backbone.layers.15.mixer.conv1d.weight": "model-00001-of-00002.safetensors",
|
| 79 |
+
"backbone.layers.15.mixer.dt_bias": "model-00001-of-00002.safetensors",
|
| 80 |
+
"backbone.layers.15.mixer.in_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 81 |
+
"backbone.layers.15.mixer.in_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 82 |
+
"backbone.layers.15.mixer.in_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 83 |
+
"backbone.layers.15.mixer.norm.weight": "model-00001-of-00002.safetensors",
|
| 84 |
+
"backbone.layers.15.mixer.out_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 85 |
+
"backbone.layers.15.mixer.out_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 86 |
+
"backbone.layers.15.mixer.out_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 87 |
+
"backbone.layers.15.norm.weight": "model-00001-of-00002.safetensors",
|
| 88 |
+
"backbone.layers.16.mixer.k_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 89 |
+
"backbone.layers.16.mixer.k_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 90 |
+
"backbone.layers.16.mixer.k_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 91 |
+
"backbone.layers.16.mixer.o_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 92 |
+
"backbone.layers.16.mixer.o_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 93 |
+
"backbone.layers.16.mixer.o_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 94 |
+
"backbone.layers.16.mixer.q_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 95 |
+
"backbone.layers.16.mixer.q_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 96 |
+
"backbone.layers.16.mixer.q_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 97 |
+
"backbone.layers.16.mixer.v_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 98 |
+
"backbone.layers.16.mixer.v_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 99 |
+
"backbone.layers.16.mixer.v_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 100 |
+
"backbone.layers.16.norm.weight": "model-00001-of-00002.safetensors",
|
| 101 |
+
"backbone.layers.17.mixer.down_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 102 |
+
"backbone.layers.17.mixer.down_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 103 |
+
"backbone.layers.17.mixer.down_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 104 |
+
"backbone.layers.17.mixer.up_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 105 |
+
"backbone.layers.17.mixer.up_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 106 |
+
"backbone.layers.17.mixer.up_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 107 |
+
"backbone.layers.17.norm.weight": "model-00001-of-00002.safetensors",
|
| 108 |
+
"backbone.layers.18.mixer.A_log": "model-00001-of-00002.safetensors",
|
| 109 |
+
"backbone.layers.18.mixer.D": "model-00001-of-00002.safetensors",
|
| 110 |
+
"backbone.layers.18.mixer.conv1d.bias": "model-00001-of-00002.safetensors",
|
| 111 |
+
"backbone.layers.18.mixer.conv1d.weight": "model-00001-of-00002.safetensors",
|
| 112 |
+
"backbone.layers.18.mixer.dt_bias": "model-00001-of-00002.safetensors",
|
| 113 |
+
"backbone.layers.18.mixer.in_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 114 |
+
"backbone.layers.18.mixer.in_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 115 |
+
"backbone.layers.18.mixer.in_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 116 |
+
"backbone.layers.18.mixer.norm.weight": "model-00001-of-00002.safetensors",
|
| 117 |
+
"backbone.layers.18.mixer.out_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 118 |
+
"backbone.layers.18.mixer.out_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 119 |
+
"backbone.layers.18.mixer.out_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 120 |
+
"backbone.layers.18.norm.weight": "model-00001-of-00002.safetensors",
|
| 121 |
+
"backbone.layers.19.mixer.down_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 122 |
+
"backbone.layers.19.mixer.down_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 123 |
+
"backbone.layers.19.mixer.down_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 124 |
+
"backbone.layers.19.mixer.up_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 125 |
+
"backbone.layers.19.mixer.up_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 126 |
+
"backbone.layers.19.mixer.up_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 127 |
+
"backbone.layers.19.norm.weight": "model-00001-of-00002.safetensors",
|
| 128 |
+
"backbone.layers.2.mixer.A_log": "model-00001-of-00002.safetensors",
|
| 129 |
+
"backbone.layers.2.mixer.D": "model-00001-of-00002.safetensors",
|
| 130 |
+
"backbone.layers.2.mixer.conv1d.bias": "model-00001-of-00002.safetensors",
|
| 131 |
+
"backbone.layers.2.mixer.conv1d.weight": "model-00001-of-00002.safetensors",
|
| 132 |
+
"backbone.layers.2.mixer.dt_bias": "model-00001-of-00002.safetensors",
|
| 133 |
+
"backbone.layers.2.mixer.in_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 134 |
+
"backbone.layers.2.mixer.in_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 135 |
+
"backbone.layers.2.mixer.in_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 136 |
+
"backbone.layers.2.mixer.norm.weight": "model-00001-of-00002.safetensors",
|
| 137 |
+
"backbone.layers.2.mixer.out_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 138 |
+
"backbone.layers.2.mixer.out_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 139 |
+
"backbone.layers.2.mixer.out_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 140 |
+
"backbone.layers.2.norm.weight": "model-00001-of-00002.safetensors",
|
| 141 |
+
"backbone.layers.20.mixer.A_log": "model-00001-of-00002.safetensors",
|
| 142 |
+
"backbone.layers.20.mixer.D": "model-00001-of-00002.safetensors",
|
| 143 |
+
"backbone.layers.20.mixer.conv1d.bias": "model-00001-of-00002.safetensors",
|
| 144 |
+
"backbone.layers.20.mixer.conv1d.weight": "model-00001-of-00002.safetensors",
|
| 145 |
+
"backbone.layers.20.mixer.dt_bias": "model-00001-of-00002.safetensors",
|
| 146 |
+
"backbone.layers.20.mixer.in_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 147 |
+
"backbone.layers.20.mixer.in_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 148 |
+
"backbone.layers.20.mixer.in_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 149 |
+
"backbone.layers.20.mixer.norm.weight": "model-00001-of-00002.safetensors",
|
| 150 |
+
"backbone.layers.20.mixer.out_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 151 |
+
"backbone.layers.20.mixer.out_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 152 |
+
"backbone.layers.20.mixer.out_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 153 |
+
"backbone.layers.20.norm.weight": "model-00001-of-00002.safetensors",
|
| 154 |
+
"backbone.layers.21.mixer.down_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 155 |
+
"backbone.layers.21.mixer.down_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 156 |
+
"backbone.layers.21.mixer.down_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 157 |
+
"backbone.layers.21.mixer.up_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 158 |
+
"backbone.layers.21.mixer.up_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 159 |
+
"backbone.layers.21.mixer.up_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 160 |
+
"backbone.layers.21.norm.weight": "model-00001-of-00002.safetensors",
|
| 161 |
+
"backbone.layers.22.mixer.A_log": "model-00001-of-00002.safetensors",
|
| 162 |
+
"backbone.layers.22.mixer.D": "model-00001-of-00002.safetensors",
|
| 163 |
+
"backbone.layers.22.mixer.conv1d.bias": "model-00001-of-00002.safetensors",
|
| 164 |
+
"backbone.layers.22.mixer.conv1d.weight": "model-00001-of-00002.safetensors",
|
| 165 |
+
"backbone.layers.22.mixer.dt_bias": "model-00001-of-00002.safetensors",
|
| 166 |
+
"backbone.layers.22.mixer.in_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 167 |
+
"backbone.layers.22.mixer.in_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 168 |
+
"backbone.layers.22.mixer.in_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 169 |
+
"backbone.layers.22.mixer.norm.weight": "model-00001-of-00002.safetensors",
|
| 170 |
+
"backbone.layers.22.mixer.out_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 171 |
+
"backbone.layers.22.mixer.out_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 172 |
+
"backbone.layers.22.mixer.out_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 173 |
+
"backbone.layers.22.norm.weight": "model-00001-of-00002.safetensors",
|
| 174 |
+
"backbone.layers.23.mixer.down_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 175 |
+
"backbone.layers.23.mixer.down_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 176 |
+
"backbone.layers.23.mixer.down_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 177 |
+
"backbone.layers.23.mixer.up_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 178 |
+
"backbone.layers.23.mixer.up_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 179 |
+
"backbone.layers.23.mixer.up_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 180 |
+
"backbone.layers.23.norm.weight": "model-00001-of-00002.safetensors",
|
| 181 |
+
"backbone.layers.24.mixer.A_log": "model-00001-of-00002.safetensors",
|
| 182 |
+
"backbone.layers.24.mixer.D": "model-00001-of-00002.safetensors",
|
| 183 |
+
"backbone.layers.24.mixer.conv1d.bias": "model-00001-of-00002.safetensors",
|
| 184 |
+
"backbone.layers.24.mixer.conv1d.weight": "model-00001-of-00002.safetensors",
|
| 185 |
+
"backbone.layers.24.mixer.dt_bias": "model-00001-of-00002.safetensors",
|
| 186 |
+
"backbone.layers.24.mixer.in_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 187 |
+
"backbone.layers.24.mixer.in_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 188 |
+
"backbone.layers.24.mixer.in_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 189 |
+
"backbone.layers.24.mixer.norm.weight": "model-00001-of-00002.safetensors",
|
| 190 |
+
"backbone.layers.24.mixer.out_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 191 |
+
"backbone.layers.24.mixer.out_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 192 |
+
"backbone.layers.24.mixer.out_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 193 |
+
"backbone.layers.24.norm.weight": "model-00001-of-00002.safetensors",
|
| 194 |
+
"backbone.layers.25.mixer.k_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 195 |
+
"backbone.layers.25.mixer.k_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 196 |
+
"backbone.layers.25.mixer.k_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 197 |
+
"backbone.layers.25.mixer.o_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 198 |
+
"backbone.layers.25.mixer.o_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 199 |
+
"backbone.layers.25.mixer.o_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 200 |
+
"backbone.layers.25.mixer.q_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 201 |
+
"backbone.layers.25.mixer.q_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 202 |
+
"backbone.layers.25.mixer.q_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 203 |
+
"backbone.layers.25.mixer.v_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 204 |
+
"backbone.layers.25.mixer.v_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 205 |
+
"backbone.layers.25.mixer.v_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 206 |
+
"backbone.layers.25.norm.weight": "model-00001-of-00002.safetensors",
|
| 207 |
+
"backbone.layers.26.mixer.down_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 208 |
+
"backbone.layers.26.mixer.down_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 209 |
+
"backbone.layers.26.mixer.down_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 210 |
+
"backbone.layers.26.mixer.up_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 211 |
+
"backbone.layers.26.mixer.up_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 212 |
+
"backbone.layers.26.mixer.up_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 213 |
+
"backbone.layers.26.norm.weight": "model-00001-of-00002.safetensors",
|
| 214 |
+
"backbone.layers.27.mixer.A_log": "model-00001-of-00002.safetensors",
|
| 215 |
+
"backbone.layers.27.mixer.D": "model-00001-of-00002.safetensors",
|
| 216 |
+
"backbone.layers.27.mixer.conv1d.bias": "model-00001-of-00002.safetensors",
|
| 217 |
+
"backbone.layers.27.mixer.conv1d.weight": "model-00001-of-00002.safetensors",
|
| 218 |
+
"backbone.layers.27.mixer.dt_bias": "model-00001-of-00002.safetensors",
|
| 219 |
+
"backbone.layers.27.mixer.in_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 220 |
+
"backbone.layers.27.mixer.in_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 221 |
+
"backbone.layers.27.mixer.in_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 222 |
+
"backbone.layers.27.mixer.norm.weight": "model-00001-of-00002.safetensors",
|
| 223 |
+
"backbone.layers.27.mixer.out_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 224 |
+
"backbone.layers.27.mixer.out_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 225 |
+
"backbone.layers.27.mixer.out_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 226 |
+
"backbone.layers.27.norm.weight": "model-00001-of-00002.safetensors",
|
| 227 |
+
"backbone.layers.28.mixer.down_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 228 |
+
"backbone.layers.28.mixer.down_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 229 |
+
"backbone.layers.28.mixer.down_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 230 |
+
"backbone.layers.28.mixer.up_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 231 |
+
"backbone.layers.28.mixer.up_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 232 |
+
"backbone.layers.28.mixer.up_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 233 |
+
"backbone.layers.28.norm.weight": "model-00001-of-00002.safetensors",
|
| 234 |
+
"backbone.layers.29.mixer.A_log": "model-00001-of-00002.safetensors",
|
| 235 |
+
"backbone.layers.29.mixer.D": "model-00001-of-00002.safetensors",
|
| 236 |
+
"backbone.layers.29.mixer.conv1d.bias": "model-00001-of-00002.safetensors",
|
| 237 |
+
"backbone.layers.29.mixer.conv1d.weight": "model-00001-of-00002.safetensors",
|
| 238 |
+
"backbone.layers.29.mixer.dt_bias": "model-00001-of-00002.safetensors",
|
| 239 |
+
"backbone.layers.29.mixer.in_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 240 |
+
"backbone.layers.29.mixer.in_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 241 |
+
"backbone.layers.29.mixer.in_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 242 |
+
"backbone.layers.29.mixer.norm.weight": "model-00001-of-00002.safetensors",
|
| 243 |
+
"backbone.layers.29.mixer.out_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 244 |
+
"backbone.layers.29.mixer.out_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 245 |
+
"backbone.layers.29.mixer.out_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 246 |
+
"backbone.layers.29.norm.weight": "model-00001-of-00002.safetensors",
|
| 247 |
+
"backbone.layers.3.mixer.down_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 248 |
+
"backbone.layers.3.mixer.down_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 249 |
+
"backbone.layers.3.mixer.down_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 250 |
+
"backbone.layers.3.mixer.up_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 251 |
+
"backbone.layers.3.mixer.up_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 252 |
+
"backbone.layers.3.mixer.up_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 253 |
+
"backbone.layers.3.norm.weight": "model-00001-of-00002.safetensors",
|
| 254 |
+
"backbone.layers.30.mixer.down_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 255 |
+
"backbone.layers.30.mixer.down_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 256 |
+
"backbone.layers.30.mixer.down_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 257 |
+
"backbone.layers.30.mixer.up_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 258 |
+
"backbone.layers.30.mixer.up_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 259 |
+
"backbone.layers.30.mixer.up_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 260 |
+
"backbone.layers.30.norm.weight": "model-00001-of-00002.safetensors",
|
| 261 |
+
"backbone.layers.31.mixer.A_log": "model-00001-of-00002.safetensors",
|
| 262 |
+
"backbone.layers.31.mixer.D": "model-00001-of-00002.safetensors",
|
| 263 |
+
"backbone.layers.31.mixer.conv1d.bias": "model-00001-of-00002.safetensors",
|
| 264 |
+
"backbone.layers.31.mixer.conv1d.weight": "model-00001-of-00002.safetensors",
|
| 265 |
+
"backbone.layers.31.mixer.dt_bias": "model-00001-of-00002.safetensors",
|
| 266 |
+
"backbone.layers.31.mixer.in_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 267 |
+
"backbone.layers.31.mixer.in_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 268 |
+
"backbone.layers.31.mixer.in_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 269 |
+
"backbone.layers.31.mixer.norm.weight": "model-00001-of-00002.safetensors",
|
| 270 |
+
"backbone.layers.31.mixer.out_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 271 |
+
"backbone.layers.31.mixer.out_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 272 |
+
"backbone.layers.31.mixer.out_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 273 |
+
"backbone.layers.31.norm.weight": "model-00001-of-00002.safetensors",
|
| 274 |
+
"backbone.layers.32.mixer.down_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 275 |
+
"backbone.layers.32.mixer.down_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 276 |
+
"backbone.layers.32.mixer.down_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 277 |
+
"backbone.layers.32.mixer.up_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 278 |
+
"backbone.layers.32.mixer.up_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 279 |
+
"backbone.layers.32.mixer.up_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 280 |
+
"backbone.layers.32.norm.weight": "model-00001-of-00002.safetensors",
|
| 281 |
+
"backbone.layers.33.mixer.A_log": "model-00001-of-00002.safetensors",
|
| 282 |
+
"backbone.layers.33.mixer.D": "model-00001-of-00002.safetensors",
|
| 283 |
+
"backbone.layers.33.mixer.conv1d.bias": "model-00001-of-00002.safetensors",
|
| 284 |
+
"backbone.layers.33.mixer.conv1d.weight": "model-00001-of-00002.safetensors",
|
| 285 |
+
"backbone.layers.33.mixer.dt_bias": "model-00001-of-00002.safetensors",
|
| 286 |
+
"backbone.layers.33.mixer.in_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 287 |
+
"backbone.layers.33.mixer.in_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 288 |
+
"backbone.layers.33.mixer.in_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 289 |
+
"backbone.layers.33.mixer.norm.weight": "model-00001-of-00002.safetensors",
|
| 290 |
+
"backbone.layers.33.mixer.out_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 291 |
+
"backbone.layers.33.mixer.out_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 292 |
+
"backbone.layers.33.mixer.out_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 293 |
+
"backbone.layers.33.norm.weight": "model-00001-of-00002.safetensors",
|
| 294 |
+
"backbone.layers.34.mixer.k_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 295 |
+
"backbone.layers.34.mixer.k_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 296 |
+
"backbone.layers.34.mixer.k_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 297 |
+
"backbone.layers.34.mixer.o_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 298 |
+
"backbone.layers.34.mixer.o_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 299 |
+
"backbone.layers.34.mixer.o_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 300 |
+
"backbone.layers.34.mixer.q_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 301 |
+
"backbone.layers.34.mixer.q_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 302 |
+
"backbone.layers.34.mixer.q_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 303 |
+
"backbone.layers.34.mixer.v_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 304 |
+
"backbone.layers.34.mixer.v_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 305 |
+
"backbone.layers.34.mixer.v_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 306 |
+
"backbone.layers.34.norm.weight": "model-00001-of-00002.safetensors",
|
| 307 |
+
"backbone.layers.35.mixer.down_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 308 |
+
"backbone.layers.35.mixer.down_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 309 |
+
"backbone.layers.35.mixer.down_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 310 |
+
"backbone.layers.35.mixer.up_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 311 |
+
"backbone.layers.35.mixer.up_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 312 |
+
"backbone.layers.35.mixer.up_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 313 |
+
"backbone.layers.35.norm.weight": "model-00001-of-00002.safetensors",
|
| 314 |
+
"backbone.layers.36.mixer.A_log": "model-00001-of-00002.safetensors",
|
| 315 |
+
"backbone.layers.36.mixer.D": "model-00001-of-00002.safetensors",
|
| 316 |
+
"backbone.layers.36.mixer.conv1d.bias": "model-00001-of-00002.safetensors",
|
| 317 |
+
"backbone.layers.36.mixer.conv1d.weight": "model-00001-of-00002.safetensors",
|
| 318 |
+
"backbone.layers.36.mixer.dt_bias": "model-00001-of-00002.safetensors",
|
| 319 |
+
"backbone.layers.36.mixer.in_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 320 |
+
"backbone.layers.36.mixer.in_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 321 |
+
"backbone.layers.36.mixer.in_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 322 |
+
"backbone.layers.36.mixer.norm.weight": "model-00001-of-00002.safetensors",
|
| 323 |
+
"backbone.layers.36.mixer.out_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 324 |
+
"backbone.layers.36.mixer.out_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 325 |
+
"backbone.layers.36.mixer.out_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 326 |
+
"backbone.layers.36.norm.weight": "model-00001-of-00002.safetensors",
|
| 327 |
+
"backbone.layers.37.mixer.down_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 328 |
+
"backbone.layers.37.mixer.down_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 329 |
+
"backbone.layers.37.mixer.down_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 330 |
+
"backbone.layers.37.mixer.up_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 331 |
+
"backbone.layers.37.mixer.up_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 332 |
+
"backbone.layers.37.mixer.up_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 333 |
+
"backbone.layers.37.norm.weight": "model-00001-of-00002.safetensors",
|
| 334 |
+
"backbone.layers.38.mixer.A_log": "model-00001-of-00002.safetensors",
|
| 335 |
+
"backbone.layers.38.mixer.D": "model-00001-of-00002.safetensors",
|
| 336 |
+
"backbone.layers.38.mixer.conv1d.bias": "model-00001-of-00002.safetensors",
|
| 337 |
+
"backbone.layers.38.mixer.conv1d.weight": "model-00001-of-00002.safetensors",
|
| 338 |
+
"backbone.layers.38.mixer.dt_bias": "model-00001-of-00002.safetensors",
|
| 339 |
+
"backbone.layers.38.mixer.in_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 340 |
+
"backbone.layers.38.mixer.in_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 341 |
+
"backbone.layers.38.mixer.in_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 342 |
+
"backbone.layers.38.mixer.norm.weight": "model-00001-of-00002.safetensors",
|
| 343 |
+
"backbone.layers.38.mixer.out_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 344 |
+
"backbone.layers.38.mixer.out_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 345 |
+
"backbone.layers.38.mixer.out_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 346 |
+
"backbone.layers.38.norm.weight": "model-00001-of-00002.safetensors",
|
| 347 |
+
"backbone.layers.39.mixer.down_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 348 |
+
"backbone.layers.39.mixer.down_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 349 |
+
"backbone.layers.39.mixer.down_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 350 |
+
"backbone.layers.39.mixer.up_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 351 |
+
"backbone.layers.39.mixer.up_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 352 |
+
"backbone.layers.39.mixer.up_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 353 |
+
"backbone.layers.39.norm.weight": "model-00001-of-00002.safetensors",
|
| 354 |
+
"backbone.layers.4.mixer.A_log": "model-00001-of-00002.safetensors",
|
| 355 |
+
"backbone.layers.4.mixer.D": "model-00001-of-00002.safetensors",
|
| 356 |
+
"backbone.layers.4.mixer.conv1d.bias": "model-00001-of-00002.safetensors",
|
| 357 |
+
"backbone.layers.4.mixer.conv1d.weight": "model-00001-of-00002.safetensors",
|
| 358 |
+
"backbone.layers.4.mixer.dt_bias": "model-00001-of-00002.safetensors",
|
| 359 |
+
"backbone.layers.4.mixer.in_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 360 |
+
"backbone.layers.4.mixer.in_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 361 |
+
"backbone.layers.4.mixer.in_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 362 |
+
"backbone.layers.4.mixer.norm.weight": "model-00001-of-00002.safetensors",
|
| 363 |
+
"backbone.layers.4.mixer.out_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 364 |
+
"backbone.layers.4.mixer.out_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 365 |
+
"backbone.layers.4.mixer.out_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 366 |
+
"backbone.layers.4.norm.weight": "model-00001-of-00002.safetensors",
|
| 367 |
+
"backbone.layers.40.mixer.A_log": "model-00001-of-00002.safetensors",
|
| 368 |
+
"backbone.layers.40.mixer.D": "model-00001-of-00002.safetensors",
|
| 369 |
+
"backbone.layers.40.mixer.conv1d.bias": "model-00001-of-00002.safetensors",
|
| 370 |
+
"backbone.layers.40.mixer.conv1d.weight": "model-00001-of-00002.safetensors",
|
| 371 |
+
"backbone.layers.40.mixer.dt_bias": "model-00001-of-00002.safetensors",
|
| 372 |
+
"backbone.layers.40.mixer.in_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 373 |
+
"backbone.layers.40.mixer.in_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 374 |
+
"backbone.layers.40.mixer.in_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 375 |
+
"backbone.layers.40.mixer.norm.weight": "model-00002-of-00002.safetensors",
|
| 376 |
+
"backbone.layers.40.mixer.out_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 377 |
+
"backbone.layers.40.mixer.out_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 378 |
+
"backbone.layers.40.mixer.out_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 379 |
+
"backbone.layers.40.norm.weight": "model-00001-of-00002.safetensors",
|
| 380 |
+
"backbone.layers.41.mixer.down_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 381 |
+
"backbone.layers.41.mixer.down_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 382 |
+
"backbone.layers.41.mixer.down_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 383 |
+
"backbone.layers.41.mixer.up_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 384 |
+
"backbone.layers.41.mixer.up_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 385 |
+
"backbone.layers.41.mixer.up_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 386 |
+
"backbone.layers.41.norm.weight": "model-00002-of-00002.safetensors",
|
| 387 |
+
"backbone.layers.42.mixer.A_log": "model-00002-of-00002.safetensors",
|
| 388 |
+
"backbone.layers.42.mixer.D": "model-00002-of-00002.safetensors",
|
| 389 |
+
"backbone.layers.42.mixer.conv1d.bias": "model-00002-of-00002.safetensors",
|
| 390 |
+
"backbone.layers.42.mixer.conv1d.weight": "model-00002-of-00002.safetensors",
|
| 391 |
+
"backbone.layers.42.mixer.dt_bias": "model-00002-of-00002.safetensors",
|
| 392 |
+
"backbone.layers.42.mixer.in_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 393 |
+
"backbone.layers.42.mixer.in_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 394 |
+
"backbone.layers.42.mixer.in_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 395 |
+
"backbone.layers.42.mixer.norm.weight": "model-00002-of-00002.safetensors",
|
| 396 |
+
"backbone.layers.42.mixer.out_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 397 |
+
"backbone.layers.42.mixer.out_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 398 |
+
"backbone.layers.42.mixer.out_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 399 |
+
"backbone.layers.42.norm.weight": "model-00002-of-00002.safetensors",
|
| 400 |
+
"backbone.layers.43.mixer.k_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 401 |
+
"backbone.layers.43.mixer.k_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 402 |
+
"backbone.layers.43.mixer.k_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 403 |
+
"backbone.layers.43.mixer.o_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 404 |
+
"backbone.layers.43.mixer.o_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 405 |
+
"backbone.layers.43.mixer.o_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 406 |
+
"backbone.layers.43.mixer.q_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 407 |
+
"backbone.layers.43.mixer.q_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 408 |
+
"backbone.layers.43.mixer.q_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 409 |
+
"backbone.layers.43.mixer.v_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 410 |
+
"backbone.layers.43.mixer.v_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 411 |
+
"backbone.layers.43.mixer.v_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 412 |
+
"backbone.layers.43.norm.weight": "model-00002-of-00002.safetensors",
|
| 413 |
+
"backbone.layers.44.mixer.down_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 414 |
+
"backbone.layers.44.mixer.down_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 415 |
+
"backbone.layers.44.mixer.down_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 416 |
+
"backbone.layers.44.mixer.up_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 417 |
+
"backbone.layers.44.mixer.up_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 418 |
+
"backbone.layers.44.mixer.up_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 419 |
+
"backbone.layers.44.norm.weight": "model-00002-of-00002.safetensors",
|
| 420 |
+
"backbone.layers.45.mixer.A_log": "model-00002-of-00002.safetensors",
|
| 421 |
+
"backbone.layers.45.mixer.D": "model-00002-of-00002.safetensors",
|
| 422 |
+
"backbone.layers.45.mixer.conv1d.bias": "model-00002-of-00002.safetensors",
|
| 423 |
+
"backbone.layers.45.mixer.conv1d.weight": "model-00002-of-00002.safetensors",
|
| 424 |
+
"backbone.layers.45.mixer.dt_bias": "model-00002-of-00002.safetensors",
|
| 425 |
+
"backbone.layers.45.mixer.in_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 426 |
+
"backbone.layers.45.mixer.in_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 427 |
+
"backbone.layers.45.mixer.in_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 428 |
+
"backbone.layers.45.mixer.norm.weight": "model-00002-of-00002.safetensors",
|
| 429 |
+
"backbone.layers.45.mixer.out_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 430 |
+
"backbone.layers.45.mixer.out_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 431 |
+
"backbone.layers.45.mixer.out_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 432 |
+
"backbone.layers.45.norm.weight": "model-00002-of-00002.safetensors",
|
| 433 |
+
"backbone.layers.46.mixer.down_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 434 |
+
"backbone.layers.46.mixer.down_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 435 |
+
"backbone.layers.46.mixer.down_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 436 |
+
"backbone.layers.46.mixer.up_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 437 |
+
"backbone.layers.46.mixer.up_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 438 |
+
"backbone.layers.46.mixer.up_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 439 |
+
"backbone.layers.46.norm.weight": "model-00002-of-00002.safetensors",
|
| 440 |
+
"backbone.layers.47.mixer.A_log": "model-00002-of-00002.safetensors",
|
| 441 |
+
"backbone.layers.47.mixer.D": "model-00002-of-00002.safetensors",
|
| 442 |
+
"backbone.layers.47.mixer.conv1d.bias": "model-00002-of-00002.safetensors",
|
| 443 |
+
"backbone.layers.47.mixer.conv1d.weight": "model-00002-of-00002.safetensors",
|
| 444 |
+
"backbone.layers.47.mixer.dt_bias": "model-00002-of-00002.safetensors",
|
| 445 |
+
"backbone.layers.47.mixer.in_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 446 |
+
"backbone.layers.47.mixer.in_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 447 |
+
"backbone.layers.47.mixer.in_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 448 |
+
"backbone.layers.47.mixer.norm.weight": "model-00002-of-00002.safetensors",
|
| 449 |
+
"backbone.layers.47.mixer.out_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 450 |
+
"backbone.layers.47.mixer.out_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 451 |
+
"backbone.layers.47.mixer.out_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 452 |
+
"backbone.layers.47.norm.weight": "model-00002-of-00002.safetensors",
|
| 453 |
+
"backbone.layers.48.mixer.down_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 454 |
+
"backbone.layers.48.mixer.down_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 455 |
+
"backbone.layers.48.mixer.down_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 456 |
+
"backbone.layers.48.mixer.up_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 457 |
+
"backbone.layers.48.mixer.up_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 458 |
+
"backbone.layers.48.mixer.up_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 459 |
+
"backbone.layers.48.norm.weight": "model-00002-of-00002.safetensors",
|
| 460 |
+
"backbone.layers.49.mixer.A_log": "model-00002-of-00002.safetensors",
|
| 461 |
+
"backbone.layers.49.mixer.D": "model-00002-of-00002.safetensors",
|
| 462 |
+
"backbone.layers.49.mixer.conv1d.bias": "model-00002-of-00002.safetensors",
|
| 463 |
+
"backbone.layers.49.mixer.conv1d.weight": "model-00002-of-00002.safetensors",
|
| 464 |
+
"backbone.layers.49.mixer.dt_bias": "model-00002-of-00002.safetensors",
|
| 465 |
+
"backbone.layers.49.mixer.in_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 466 |
+
"backbone.layers.49.mixer.in_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 467 |
+
"backbone.layers.49.mixer.in_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 468 |
+
"backbone.layers.49.mixer.norm.weight": "model-00002-of-00002.safetensors",
|
| 469 |
+
"backbone.layers.49.mixer.out_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 470 |
+
"backbone.layers.49.mixer.out_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 471 |
+
"backbone.layers.49.mixer.out_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 472 |
+
"backbone.layers.49.norm.weight": "model-00002-of-00002.safetensors",
|
| 473 |
+
"backbone.layers.5.mixer.down_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 474 |
+
"backbone.layers.5.mixer.down_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 475 |
+
"backbone.layers.5.mixer.down_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 476 |
+
"backbone.layers.5.mixer.up_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 477 |
+
"backbone.layers.5.mixer.up_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 478 |
+
"backbone.layers.5.mixer.up_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 479 |
+
"backbone.layers.5.norm.weight": "model-00001-of-00002.safetensors",
|
| 480 |
+
"backbone.layers.50.mixer.down_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 481 |
+
"backbone.layers.50.mixer.down_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 482 |
+
"backbone.layers.50.mixer.down_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 483 |
+
"backbone.layers.50.mixer.up_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 484 |
+
"backbone.layers.50.mixer.up_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 485 |
+
"backbone.layers.50.mixer.up_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 486 |
+
"backbone.layers.50.norm.weight": "model-00002-of-00002.safetensors",
|
| 487 |
+
"backbone.layers.51.mixer.A_log": "model-00002-of-00002.safetensors",
|
| 488 |
+
"backbone.layers.51.mixer.D": "model-00002-of-00002.safetensors",
|
| 489 |
+
"backbone.layers.51.mixer.conv1d.bias": "model-00002-of-00002.safetensors",
|
| 490 |
+
"backbone.layers.51.mixer.conv1d.weight": "model-00002-of-00002.safetensors",
|
| 491 |
+
"backbone.layers.51.mixer.dt_bias": "model-00002-of-00002.safetensors",
|
| 492 |
+
"backbone.layers.51.mixer.in_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 493 |
+
"backbone.layers.51.mixer.in_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 494 |
+
"backbone.layers.51.mixer.in_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 495 |
+
"backbone.layers.51.mixer.norm.weight": "model-00002-of-00002.safetensors",
|
| 496 |
+
"backbone.layers.51.mixer.out_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 497 |
+
"backbone.layers.51.mixer.out_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 498 |
+
"backbone.layers.51.mixer.out_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 499 |
+
"backbone.layers.51.norm.weight": "model-00002-of-00002.safetensors",
|
| 500 |
+
"backbone.layers.52.mixer.k_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 501 |
+
"backbone.layers.52.mixer.k_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 502 |
+
"backbone.layers.52.mixer.k_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 503 |
+
"backbone.layers.52.mixer.o_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 504 |
+
"backbone.layers.52.mixer.o_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 505 |
+
"backbone.layers.52.mixer.o_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 506 |
+
"backbone.layers.52.mixer.q_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 507 |
+
"backbone.layers.52.mixer.q_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 508 |
+
"backbone.layers.52.mixer.q_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 509 |
+
"backbone.layers.52.mixer.v_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 510 |
+
"backbone.layers.52.mixer.v_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 511 |
+
"backbone.layers.52.mixer.v_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 512 |
+
"backbone.layers.52.norm.weight": "model-00002-of-00002.safetensors",
|
| 513 |
+
"backbone.layers.53.mixer.down_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 514 |
+
"backbone.layers.53.mixer.down_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 515 |
+
"backbone.layers.53.mixer.down_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 516 |
+
"backbone.layers.53.mixer.up_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 517 |
+
"backbone.layers.53.mixer.up_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 518 |
+
"backbone.layers.53.mixer.up_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 519 |
+
"backbone.layers.53.norm.weight": "model-00002-of-00002.safetensors",
|
| 520 |
+
"backbone.layers.54.mixer.A_log": "model-00002-of-00002.safetensors",
|
| 521 |
+
"backbone.layers.54.mixer.D": "model-00002-of-00002.safetensors",
|
| 522 |
+
"backbone.layers.54.mixer.conv1d.bias": "model-00002-of-00002.safetensors",
|
| 523 |
+
"backbone.layers.54.mixer.conv1d.weight": "model-00002-of-00002.safetensors",
|
| 524 |
+
"backbone.layers.54.mixer.dt_bias": "model-00002-of-00002.safetensors",
|
| 525 |
+
"backbone.layers.54.mixer.in_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 526 |
+
"backbone.layers.54.mixer.in_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 527 |
+
"backbone.layers.54.mixer.in_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 528 |
+
"backbone.layers.54.mixer.norm.weight": "model-00002-of-00002.safetensors",
|
| 529 |
+
"backbone.layers.54.mixer.out_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 530 |
+
"backbone.layers.54.mixer.out_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 531 |
+
"backbone.layers.54.mixer.out_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 532 |
+
"backbone.layers.54.norm.weight": "model-00002-of-00002.safetensors",
|
| 533 |
+
"backbone.layers.55.mixer.down_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 534 |
+
"backbone.layers.55.mixer.down_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 535 |
+
"backbone.layers.55.mixer.down_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 536 |
+
"backbone.layers.55.mixer.up_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 537 |
+
"backbone.layers.55.mixer.up_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 538 |
+
"backbone.layers.55.mixer.up_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 539 |
+
"backbone.layers.55.norm.weight": "model-00002-of-00002.safetensors",
|
| 540 |
+
"backbone.layers.56.mixer.A_log": "model-00002-of-00002.safetensors",
|
| 541 |
+
"backbone.layers.56.mixer.D": "model-00002-of-00002.safetensors",
|
| 542 |
+
"backbone.layers.56.mixer.conv1d.bias": "model-00002-of-00002.safetensors",
|
| 543 |
+
"backbone.layers.56.mixer.conv1d.weight": "model-00002-of-00002.safetensors",
|
| 544 |
+
"backbone.layers.56.mixer.dt_bias": "model-00002-of-00002.safetensors",
|
| 545 |
+
"backbone.layers.56.mixer.in_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 546 |
+
"backbone.layers.56.mixer.in_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 547 |
+
"backbone.layers.56.mixer.in_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 548 |
+
"backbone.layers.56.mixer.norm.weight": "model-00002-of-00002.safetensors",
|
| 549 |
+
"backbone.layers.56.mixer.out_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 550 |
+
"backbone.layers.56.mixer.out_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 551 |
+
"backbone.layers.56.mixer.out_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 552 |
+
"backbone.layers.56.norm.weight": "model-00002-of-00002.safetensors",
|
| 553 |
+
"backbone.layers.57.mixer.down_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 554 |
+
"backbone.layers.57.mixer.down_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 555 |
+
"backbone.layers.57.mixer.down_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 556 |
+
"backbone.layers.57.mixer.up_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 557 |
+
"backbone.layers.57.mixer.up_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 558 |
+
"backbone.layers.57.mixer.up_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 559 |
+
"backbone.layers.57.norm.weight": "model-00002-of-00002.safetensors",
|
| 560 |
+
"backbone.layers.58.mixer.A_log": "model-00002-of-00002.safetensors",
|
| 561 |
+
"backbone.layers.58.mixer.D": "model-00002-of-00002.safetensors",
|
| 562 |
+
"backbone.layers.58.mixer.conv1d.bias": "model-00002-of-00002.safetensors",
|
| 563 |
+
"backbone.layers.58.mixer.conv1d.weight": "model-00002-of-00002.safetensors",
|
| 564 |
+
"backbone.layers.58.mixer.dt_bias": "model-00002-of-00002.safetensors",
|
| 565 |
+
"backbone.layers.58.mixer.in_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 566 |
+
"backbone.layers.58.mixer.in_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 567 |
+
"backbone.layers.58.mixer.in_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 568 |
+
"backbone.layers.58.mixer.norm.weight": "model-00002-of-00002.safetensors",
|
| 569 |
+
"backbone.layers.58.mixer.out_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 570 |
+
"backbone.layers.58.mixer.out_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 571 |
+
"backbone.layers.58.mixer.out_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 572 |
+
"backbone.layers.58.norm.weight": "model-00002-of-00002.safetensors",
|
| 573 |
+
"backbone.layers.59.mixer.down_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 574 |
+
"backbone.layers.59.mixer.down_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 575 |
+
"backbone.layers.59.mixer.down_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 576 |
+
"backbone.layers.59.mixer.up_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 577 |
+
"backbone.layers.59.mixer.up_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 578 |
+
"backbone.layers.59.mixer.up_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 579 |
+
"backbone.layers.59.norm.weight": "model-00002-of-00002.safetensors",
|
| 580 |
+
"backbone.layers.6.mixer.A_log": "model-00001-of-00002.safetensors",
|
| 581 |
+
"backbone.layers.6.mixer.D": "model-00001-of-00002.safetensors",
|
| 582 |
+
"backbone.layers.6.mixer.conv1d.bias": "model-00001-of-00002.safetensors",
|
| 583 |
+
"backbone.layers.6.mixer.conv1d.weight": "model-00001-of-00002.safetensors",
|
| 584 |
+
"backbone.layers.6.mixer.dt_bias": "model-00001-of-00002.safetensors",
|
| 585 |
+
"backbone.layers.6.mixer.in_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 586 |
+
"backbone.layers.6.mixer.in_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 587 |
+
"backbone.layers.6.mixer.in_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 588 |
+
"backbone.layers.6.mixer.norm.weight": "model-00001-of-00002.safetensors",
|
| 589 |
+
"backbone.layers.6.mixer.out_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 590 |
+
"backbone.layers.6.mixer.out_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 591 |
+
"backbone.layers.6.mixer.out_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 592 |
+
"backbone.layers.6.norm.weight": "model-00001-of-00002.safetensors",
|
| 593 |
+
"backbone.layers.60.mixer.A_log": "model-00002-of-00002.safetensors",
|
| 594 |
+
"backbone.layers.60.mixer.D": "model-00002-of-00002.safetensors",
|
| 595 |
+
"backbone.layers.60.mixer.conv1d.bias": "model-00002-of-00002.safetensors",
|
| 596 |
+
"backbone.layers.60.mixer.conv1d.weight": "model-00002-of-00002.safetensors",
|
| 597 |
+
"backbone.layers.60.mixer.dt_bias": "model-00002-of-00002.safetensors",
|
| 598 |
+
"backbone.layers.60.mixer.in_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 599 |
+
"backbone.layers.60.mixer.in_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 600 |
+
"backbone.layers.60.mixer.in_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 601 |
+
"backbone.layers.60.mixer.norm.weight": "model-00002-of-00002.safetensors",
|
| 602 |
+
"backbone.layers.60.mixer.out_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 603 |
+
"backbone.layers.60.mixer.out_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 604 |
+
"backbone.layers.60.mixer.out_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 605 |
+
"backbone.layers.60.norm.weight": "model-00002-of-00002.safetensors",
|
| 606 |
+
"backbone.layers.61.mixer.down_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 607 |
+
"backbone.layers.61.mixer.down_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 608 |
+
"backbone.layers.61.mixer.down_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 609 |
+
"backbone.layers.61.mixer.up_proj.weight_packed": "model-00002-of-00002.safetensors",
|
| 610 |
+
"backbone.layers.61.mixer.up_proj.weight_scale": "model-00002-of-00002.safetensors",
|
| 611 |
+
"backbone.layers.61.mixer.up_proj.weight_shape": "model-00002-of-00002.safetensors",
|
| 612 |
+
"backbone.layers.61.norm.weight": "model-00002-of-00002.safetensors",
|
| 613 |
+
"backbone.layers.7.mixer.k_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 614 |
+
"backbone.layers.7.mixer.k_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 615 |
+
"backbone.layers.7.mixer.k_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 616 |
+
"backbone.layers.7.mixer.o_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 617 |
+
"backbone.layers.7.mixer.o_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 618 |
+
"backbone.layers.7.mixer.o_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 619 |
+
"backbone.layers.7.mixer.q_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 620 |
+
"backbone.layers.7.mixer.q_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 621 |
+
"backbone.layers.7.mixer.q_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 622 |
+
"backbone.layers.7.mixer.v_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 623 |
+
"backbone.layers.7.mixer.v_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 624 |
+
"backbone.layers.7.mixer.v_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 625 |
+
"backbone.layers.7.norm.weight": "model-00001-of-00002.safetensors",
|
| 626 |
+
"backbone.layers.8.mixer.down_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 627 |
+
"backbone.layers.8.mixer.down_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 628 |
+
"backbone.layers.8.mixer.down_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 629 |
+
"backbone.layers.8.mixer.up_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 630 |
+
"backbone.layers.8.mixer.up_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 631 |
+
"backbone.layers.8.mixer.up_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 632 |
+
"backbone.layers.8.norm.weight": "model-00001-of-00002.safetensors",
|
| 633 |
+
"backbone.layers.9.mixer.A_log": "model-00001-of-00002.safetensors",
|
| 634 |
+
"backbone.layers.9.mixer.D": "model-00001-of-00002.safetensors",
|
| 635 |
+
"backbone.layers.9.mixer.conv1d.bias": "model-00001-of-00002.safetensors",
|
| 636 |
+
"backbone.layers.9.mixer.conv1d.weight": "model-00001-of-00002.safetensors",
|
| 637 |
+
"backbone.layers.9.mixer.dt_bias": "model-00001-of-00002.safetensors",
|
| 638 |
+
"backbone.layers.9.mixer.in_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 639 |
+
"backbone.layers.9.mixer.in_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 640 |
+
"backbone.layers.9.mixer.in_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 641 |
+
"backbone.layers.9.mixer.norm.weight": "model-00001-of-00002.safetensors",
|
| 642 |
+
"backbone.layers.9.mixer.out_proj.weight_packed": "model-00001-of-00002.safetensors",
|
| 643 |
+
"backbone.layers.9.mixer.out_proj.weight_scale": "model-00001-of-00002.safetensors",
|
| 644 |
+
"backbone.layers.9.mixer.out_proj.weight_shape": "model-00001-of-00002.safetensors",
|
| 645 |
+
"backbone.layers.9.norm.weight": "model-00001-of-00002.safetensors",
|
| 646 |
+
"backbone.norm_f.weight": "model-00002-of-00002.safetensors",
|
| 647 |
+
"lm_head.weight": "model-00002-of-00002.safetensors"
|
| 648 |
+
}
|
| 649 |
+
}
|
modeling_nemotron_h.py
ADDED
|
@@ -0,0 +1,1639 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2024 HuggingFace Inc. team.
|
| 3 |
+
# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
|
| 4 |
+
#
|
| 5 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 6 |
+
# you may not use this file except in compliance with the License.
|
| 7 |
+
# You may obtain a copy of the License at
|
| 8 |
+
#
|
| 9 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 10 |
+
#
|
| 11 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 12 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 13 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 14 |
+
# See the License for the specific language governing permissions and
|
| 15 |
+
# limitations under the License.
|
| 16 |
+
"""PyTorch NemotronH model."""
|
| 17 |
+
|
| 18 |
+
import math
|
| 19 |
+
from dataclasses import dataclass
|
| 20 |
+
from typing import Any, Dict, Optional, Tuple, Union
|
| 21 |
+
|
| 22 |
+
import torch
|
| 23 |
+
import torch.utils.checkpoint
|
| 24 |
+
from torch import nn
|
| 25 |
+
from torch.nn import CrossEntropyLoss
|
| 26 |
+
|
| 27 |
+
from transformers.activations import ACT2FN
|
| 28 |
+
from transformers.cache_utils import DynamicCache # we need __iter__ and __len__ of pkv
|
| 29 |
+
from transformers.generation import GenerationMixin
|
| 30 |
+
from transformers.modeling_attn_mask_utils import (
|
| 31 |
+
AttentionMaskConverter,
|
| 32 |
+
)
|
| 33 |
+
from transformers.modeling_utils import PreTrainedModel
|
| 34 |
+
from transformers.utils import (
|
| 35 |
+
ModelOutput,
|
| 36 |
+
add_code_sample_docstrings,
|
| 37 |
+
add_start_docstrings,
|
| 38 |
+
add_start_docstrings_to_model_forward,
|
| 39 |
+
logging,
|
| 40 |
+
)
|
| 41 |
+
from transformers.utils.import_utils import (
|
| 42 |
+
is_causal_conv1d_available,
|
| 43 |
+
is_flash_attn_2_available,
|
| 44 |
+
is_flash_attn_greater_or_equal_2_10,
|
| 45 |
+
is_mamba_2_ssm_available,
|
| 46 |
+
)
|
| 47 |
+
from .configuration_nemotron_h import NemotronHConfig
|
| 48 |
+
|
| 49 |
+
|
| 50 |
+
logger = logging.get_logger(__name__)
|
| 51 |
+
|
| 52 |
+
|
| 53 |
+
# Copied from transformers.models.mamba.modeling_mamba2.modeling_mamba2.py with MAMBA2->NEMOTRONH,Mamba2->NemotronH
|
| 54 |
+
# For Mamba2 components Mamba2->NemotronHMamba2
|
| 55 |
+
if is_mamba_2_ssm_available():
|
| 56 |
+
from mamba_ssm.ops.triton.selective_state_update import selective_state_update
|
| 57 |
+
from mamba_ssm.ops.triton.ssd_combined import mamba_chunk_scan_combined, mamba_split_conv1d_scan_combined
|
| 58 |
+
else:
|
| 59 |
+
mamba_chunk_scan_combined, mamba_split_conv1d_scan_combined, selective_state_update = None, None, None
|
| 60 |
+
|
| 61 |
+
try:
|
| 62 |
+
#from mamba_ssm.ops.triton.layernorm_gated import RMSNorm as RMSNormGated
|
| 63 |
+
from mamba_ssm.ops.triton.layernorm_gated import rmsnorm_fn
|
| 64 |
+
except ImportError:
|
| 65 |
+
raise ImportError("mamba-ssm is required by the Mamba model but cannot be imported")
|
| 66 |
+
|
| 67 |
+
if is_causal_conv1d_available():
|
| 68 |
+
from causal_conv1d import causal_conv1d_fn, causal_conv1d_update
|
| 69 |
+
else:
|
| 70 |
+
causal_conv1d_update, causal_conv1d_fn = None, None
|
| 71 |
+
|
| 72 |
+
if is_flash_attn_2_available():
|
| 73 |
+
from transformers.modeling_flash_attention_utils import _flash_attention_forward
|
| 74 |
+
|
| 75 |
+
is_fast_path_available = all(
|
| 76 |
+
(
|
| 77 |
+
selective_state_update,
|
| 78 |
+
mamba_chunk_scan_combined,
|
| 79 |
+
mamba_split_conv1d_scan_combined,
|
| 80 |
+
causal_conv1d_fn,
|
| 81 |
+
causal_conv1d_update,
|
| 82 |
+
)
|
| 83 |
+
)
|
| 84 |
+
|
| 85 |
+
|
| 86 |
+
_CHECKPOINT_FOR_DOC = "nvidia/Nemotron-H-56B-Base-8K"
|
| 87 |
+
_CONFIG_FOR_DOC = "NemotronHConfig"
|
| 88 |
+
|
| 89 |
+
|
| 90 |
+
# Helper methods for segment sum computation
|
| 91 |
+
|
| 92 |
+
|
| 93 |
+
def pad_tensor_by_size(input_tensor: torch.Tensor, pad_size: int):
|
| 94 |
+
"""
|
| 95 |
+
Padding x tensor with `pad_size` on the seq_len dim (dim=1)
|
| 96 |
+
|
| 97 |
+
Assumes that we only have tensors of either size 4 or 3
|
| 98 |
+
"""
|
| 99 |
+
pad_shape = (0, 0, 0, 0, 0, pad_size, 0, 0) if len(input_tensor.shape) == 4 else (0, 0, 0, pad_size, 0, 0)
|
| 100 |
+
|
| 101 |
+
return torch.nn.functional.pad(input_tensor, pad_shape, mode="constant", value=0)
|
| 102 |
+
|
| 103 |
+
|
| 104 |
+
def reshape_into_chunks(input_tensor, pad_size, chunk_size):
|
| 105 |
+
"""
|
| 106 |
+
Padding input_tensor with `pad_size` on the seq_len dim (dim=1) and
|
| 107 |
+
simultaneously splitting it into chunk sequences.
|
| 108 |
+
|
| 109 |
+
Assumes that we only have tensors of either size 4 or 3
|
| 110 |
+
"""
|
| 111 |
+
# [bsz, seq_len, ...] -> [bsz, seq_len multiple of chunk_size, ...]
|
| 112 |
+
input_tensor = pad_tensor_by_size(input_tensor, pad_size)
|
| 113 |
+
|
| 114 |
+
if len(input_tensor.shape) == 3:
|
| 115 |
+
# [bsz, seq_len multiple of chunk_size, num_heads] -> [bsz, -1, chunk_size, num_heads]
|
| 116 |
+
return input_tensor.reshape(input_tensor.shape[0], -1, chunk_size, input_tensor.shape[2])
|
| 117 |
+
else:
|
| 118 |
+
# [bsz, seq_len multiple of chunk_size, num_heads, head_dim or state_size] -> [bsz, -1, chunk_size, num_heads, head_dim or state_size]
|
| 119 |
+
return input_tensor.reshape(
|
| 120 |
+
input_tensor.shape[0], -1, chunk_size, input_tensor.shape[2], input_tensor.shape[3]
|
| 121 |
+
)
|
| 122 |
+
|
| 123 |
+
|
| 124 |
+
def segment_sum(input_tensor):
|
| 125 |
+
"""
|
| 126 |
+
More stable segment sum calculation. Uses cumulative sums and masking instead of direct subtractions.
|
| 127 |
+
"""
|
| 128 |
+
chunk_size = input_tensor.size(-1)
|
| 129 |
+
# 1. expand input tensor to have an additional dimension and repeat along that dimension
|
| 130 |
+
# [..., chunk_size] -> [..., chunk_size, chunk_size]
|
| 131 |
+
input_tensor = input_tensor[..., None].expand(*input_tensor.size(), chunk_size)
|
| 132 |
+
# 2. create a lower triangular mask with the diagonal set to 0 to 0 out elements above diag
|
| 133 |
+
mask = torch.tril(torch.ones(chunk_size, chunk_size, device=input_tensor.device, dtype=torch.bool), diagonal=-1)
|
| 134 |
+
input_tensor = input_tensor.masked_fill(~mask, 0)
|
| 135 |
+
# 3. compute actual cumsum
|
| 136 |
+
tensor_segsum = torch.cumsum(input_tensor, dim=-2)
|
| 137 |
+
|
| 138 |
+
# 4. apply mask to keep only the lower triangular part of the cumulative sum result (incl diagonal this time)
|
| 139 |
+
mask = torch.tril(torch.ones(chunk_size, chunk_size, device=input_tensor.device, dtype=torch.bool), diagonal=0)
|
| 140 |
+
tensor_segsum = tensor_segsum.masked_fill(~mask, -torch.inf)
|
| 141 |
+
return tensor_segsum
|
| 142 |
+
|
| 143 |
+
|
| 144 |
+
def apply_mask_to_padding_states(hidden_states, attention_mask):
|
| 145 |
+
"""
|
| 146 |
+
Tunes out the hidden states for padding tokens, see https://github.com/state-spaces/mamba/issues/66
|
| 147 |
+
"""
|
| 148 |
+
if attention_mask is not None and attention_mask.shape[1] > 1 and attention_mask.shape[0] > 1:
|
| 149 |
+
dtype = hidden_states.dtype
|
| 150 |
+
hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)
|
| 151 |
+
|
| 152 |
+
return hidden_states
|
| 153 |
+
|
| 154 |
+
# Copied from https://github.com/huggingface/transformers/blob/main/src/transformers/models/jamba/modeling_jamba.py
|
| 155 |
+
class HybridMambaAttentionDynamicCache(DynamicCache):
|
| 156 |
+
"""
|
| 157 |
+
A dynamic cache that can handle both the attention cache (which has a seq_len dimension) and the mamba cache
|
| 158 |
+
(which has a constant shape regardless of seq_len).
|
| 159 |
+
|
| 160 |
+
This cache has two sets of lists of tensors: `key_cache` and `value_cache` for attention cache and `conv_states`
|
| 161 |
+
and `ssm_states` for mamba cache. Each of these lists has `num_layers` tensors. The expected shape for each tensor
|
| 162 |
+
For attention layers, `key_cache` and `value_cache` have a shape of `(batch_size, num_heads, seq_len, head_dim)`,
|
| 163 |
+
while `conv_states` and `ssm_states` have a shape of `(batch_size, 0)` (empty tensors).
|
| 164 |
+
For mamba layers, `key_cache` and `value_cache` have a shape of `(batch_size, 0)` (empty tensors),
|
| 165 |
+
while `conv_states` represents the convolution state and has a shape of `(batch_size, d_inner, d_conv)`,
|
| 166 |
+
and `ssm_states` represents the ssm state and has a shape of `(batch_size, d_inner, d_state)`.
|
| 167 |
+
"""
|
| 168 |
+
|
| 169 |
+
def __init__(self, config, batch_size, dtype=torch.float16, device=None):
|
| 170 |
+
super().__init__()
|
| 171 |
+
self.dtype = dtype
|
| 172 |
+
self.hybrid_override_pattern = config.hybrid_override_pattern
|
| 173 |
+
self.has_previous_state = False # only used by mamba
|
| 174 |
+
#intermediate_size = config.expand * config.hidden_size
|
| 175 |
+
intermediate_size = config.mamba_num_heads * config.mamba_head_dim
|
| 176 |
+
ssm_state_size = config.ssm_state_size
|
| 177 |
+
conv_kernel_size = config.conv_kernel
|
| 178 |
+
self.conv_states = []
|
| 179 |
+
self.ssm_states = []
|
| 180 |
+
self.transformer_layers = []
|
| 181 |
+
for i in range(config.num_hidden_layers):
|
| 182 |
+
if self.hybrid_override_pattern[i] == "M":
|
| 183 |
+
# Mamba layer
|
| 184 |
+
self.conv_states += [
|
| 185 |
+
torch.zeros(batch_size, intermediate_size, conv_kernel_size, device=device, dtype=dtype)
|
| 186 |
+
]
|
| 187 |
+
self.ssm_states += [
|
| 188 |
+
torch.zeros(batch_size, intermediate_size, ssm_state_size, device=device, dtype=dtype)
|
| 189 |
+
]
|
| 190 |
+
else:
|
| 191 |
+
# Attention or MLP layer
|
| 192 |
+
self.conv_states += [torch.tensor([[]] * batch_size, device=device)]
|
| 193 |
+
self.ssm_states += [torch.tensor([[]] * batch_size, device=device)]
|
| 194 |
+
self.transformer_layers.append(i)
|
| 195 |
+
|
| 196 |
+
self.key_cache = [torch.tensor([[]] * batch_size, device=device) for _ in range(config.num_hidden_layers)]
|
| 197 |
+
self.value_cache = [torch.tensor([[]] * batch_size, device=device) for _ in range(config.num_hidden_layers)]
|
| 198 |
+
|
| 199 |
+
def update(
|
| 200 |
+
self,
|
| 201 |
+
key_states: torch.Tensor,
|
| 202 |
+
value_states: torch.Tensor,
|
| 203 |
+
layer_idx: int,
|
| 204 |
+
cache_kwargs: Optional[Dict[str, Any]] = None,
|
| 205 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
| 206 |
+
# Update the cache
|
| 207 |
+
if self.key_cache[layer_idx].shape[-1] == 0:
|
| 208 |
+
self.key_cache[layer_idx] = key_states
|
| 209 |
+
self.value_cache[layer_idx] = value_states
|
| 210 |
+
else:
|
| 211 |
+
self.key_cache[layer_idx] = torch.cat([self.key_cache[layer_idx], key_states], dim=2)
|
| 212 |
+
self.value_cache[layer_idx] = torch.cat([self.value_cache[layer_idx], value_states], dim=2)
|
| 213 |
+
|
| 214 |
+
return self.key_cache[layer_idx], self.value_cache[layer_idx]
|
| 215 |
+
|
| 216 |
+
def reorder_cache(self, beam_idx: torch.LongTensor):
|
| 217 |
+
"""Reorders the cache for beam search, given the selected beam indices."""
|
| 218 |
+
for layer_idx in range(len(self.key_cache)):
|
| 219 |
+
device = self.key_cache[layer_idx].device
|
| 220 |
+
self.key_cache[layer_idx] = self.key_cache[layer_idx].index_select(0, beam_idx.to(device))
|
| 221 |
+
device = self.value_cache[layer_idx].device
|
| 222 |
+
self.value_cache[layer_idx] = self.value_cache[layer_idx].index_select(0, beam_idx.to(device))
|
| 223 |
+
|
| 224 |
+
device = self.conv_states[layer_idx].device
|
| 225 |
+
self.conv_states[layer_idx] = self.conv_states[layer_idx].index_select(0, beam_idx.to(device))
|
| 226 |
+
device = self.ssm_states[layer_idx].device
|
| 227 |
+
self.ssm_states[layer_idx] = self.ssm_states[layer_idx].index_select(0, beam_idx.to(device))
|
| 228 |
+
|
| 229 |
+
def get_seq_length(self, layer_idx: Optional[int] = 0) -> int:
|
| 230 |
+
"""Returns the sequence length of the cached states. A layer index can be optionally passed."""
|
| 231 |
+
# take any layer that contains cache and not empty tensor
|
| 232 |
+
layer_idx = self.transformer_layers[0] if layer_idx not in self.transformer_layers else layer_idx
|
| 233 |
+
if len(self.key_cache) <= layer_idx:
|
| 234 |
+
return 0
|
| 235 |
+
return self.key_cache[layer_idx].shape[-2]
|
| 236 |
+
|
| 237 |
+
def to_legacy_cache(self) -> Tuple[Tuple[torch.Tensor], Tuple[torch.Tensor]]:
|
| 238 |
+
raise NotImplementedError("HybridMambaAttentionDynamicCache does not have a legacy cache equivalent.")
|
| 239 |
+
|
| 240 |
+
@classmethod
|
| 241 |
+
def from_legacy_cache(cls, past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None) -> "DynamicCache":
|
| 242 |
+
raise NotImplementedError("HybridMambaAttentionDynamicCache does not have a legacy cache equivalent.")
|
| 243 |
+
|
| 244 |
+
# Copied from modeling_mamba2.py
|
| 245 |
+
def update_conv_state(
|
| 246 |
+
self, layer_idx: int, new_conv_state: torch.Tensor, cache_init: bool = False
|
| 247 |
+
) -> torch.Tensor:
|
| 248 |
+
if cache_init:
|
| 249 |
+
self.conv_states[layer_idx] = new_conv_state.to(self.conv_states.device)
|
| 250 |
+
else:
|
| 251 |
+
self.conv_states[layer_idx] = self.conv_states[layer_idx].roll(shifts=-1, dims=-1)
|
| 252 |
+
self.conv_states[layer_idx][:, :, -1] = new_conv_state[:, 0, :].to(self.conv_states.device)
|
| 253 |
+
return self.conv_states[layer_idx]
|
| 254 |
+
|
| 255 |
+
def update_ssm_state(self, layer_idx: int, new_ssm_state: torch.Tensor):
|
| 256 |
+
self.ssm_states[layer_idx] = new_ssm_state.to(self.ssm_states.device)
|
| 257 |
+
return self.ssm_states[layer_idx]
|
| 258 |
+
|
| 259 |
+
def reset(self):
|
| 260 |
+
self.conv_states.zero_()
|
| 261 |
+
self.ssm_states.zero_()
|
| 262 |
+
|
| 263 |
+
class MambaRMSNormGated(torch.nn.Module):
|
| 264 |
+
def __init__(self, hidden_size, group_size, eps=1e-5):
|
| 265 |
+
super().__init__()
|
| 266 |
+
self.weight = nn.Parameter(torch.ones(hidden_size))
|
| 267 |
+
self.variance_epsilon = eps
|
| 268 |
+
self.group_size = group_size
|
| 269 |
+
|
| 270 |
+
# jan28b version
|
| 271 |
+
def forward(self, hidden_states, gate=None):
|
| 272 |
+
return rmsnorm_fn(x=hidden_states,
|
| 273 |
+
weight=self.weight,
|
| 274 |
+
bias=None, # No bias
|
| 275 |
+
z=gate,
|
| 276 |
+
eps=self.variance_epsilon,
|
| 277 |
+
group_size=self.group_size,
|
| 278 |
+
norm_before_gate=False
|
| 279 |
+
)
|
| 280 |
+
|
| 281 |
+
class NemotronHMamba2Mixer(nn.Module):
|
| 282 |
+
"""
|
| 283 |
+
Compute ∆, A, B, C, and D the state space parameters and compute the `contextualized_states`.
|
| 284 |
+
A, D are input independent (see Mamba paper [1] Section 3.5.2 "Interpretation of A" for why A isn't selective)
|
| 285 |
+
∆, B, C are input-dependent (this is a key difference between Mamba and the linear time invariant S4,
|
| 286 |
+
and is why Mamba is called **selective** state spaces)
|
| 287 |
+
"""
|
| 288 |
+
|
| 289 |
+
def __init__(self, config: NemotronHConfig, layer_idx: int):
|
| 290 |
+
super().__init__()
|
| 291 |
+
self.num_heads = config.mamba_num_heads
|
| 292 |
+
self.hidden_size = config.hidden_size
|
| 293 |
+
self.ssm_state_size = config.ssm_state_size
|
| 294 |
+
self.conv_kernel_size = config.conv_kernel
|
| 295 |
+
self.intermediate_size = config.mamba_num_heads * config.mamba_head_dim
|
| 296 |
+
self.layer_idx = layer_idx
|
| 297 |
+
self.use_conv_bias = config.use_conv_bias
|
| 298 |
+
self.activation = config.mamba_hidden_act
|
| 299 |
+
self.act = ACT2FN[config.mamba_hidden_act]
|
| 300 |
+
|
| 301 |
+
self.layer_norm_epsilon = config.layer_norm_epsilon
|
| 302 |
+
|
| 303 |
+
self.n_groups = config.n_groups
|
| 304 |
+
self.head_dim = config.mamba_head_dim
|
| 305 |
+
self.chunk_size = config.chunk_size
|
| 306 |
+
|
| 307 |
+
self.time_step_limit = config.time_step_limit
|
| 308 |
+
self.time_step_min = config.time_step_min
|
| 309 |
+
self.time_step_max = config.time_step_max
|
| 310 |
+
|
| 311 |
+
self.conv_dim = self.intermediate_size + 2 * self.n_groups * self.ssm_state_size
|
| 312 |
+
self.conv1d = nn.Conv1d(
|
| 313 |
+
in_channels=self.conv_dim,
|
| 314 |
+
out_channels=self.conv_dim,
|
| 315 |
+
bias=config.use_conv_bias,
|
| 316 |
+
kernel_size=config.conv_kernel,
|
| 317 |
+
groups=self.conv_dim,
|
| 318 |
+
padding=config.conv_kernel - 1,
|
| 319 |
+
)
|
| 320 |
+
|
| 321 |
+
# projection of the input hidden states
|
| 322 |
+
projection_size = self.intermediate_size + self.conv_dim + self.num_heads
|
| 323 |
+
self.in_proj = nn.Linear(
|
| 324 |
+
self.hidden_size,
|
| 325 |
+
projection_size,
|
| 326 |
+
bias=config.use_bias,
|
| 327 |
+
)
|
| 328 |
+
# selective projection used to make dt, B and C input dependant
|
| 329 |
+
|
| 330 |
+
# time step projection (discretization)
|
| 331 |
+
# instantiate once and copy inv_dt in init_weights of PretrainedModel
|
| 332 |
+
self.dt_bias = nn.Parameter(torch.ones(self.num_heads))
|
| 333 |
+
|
| 334 |
+
# S4D real initialization. These are not discretized!
|
| 335 |
+
# The core is to load them, compute the discrete states, then write the updated state. Keeps the memory bounded
|
| 336 |
+
A = torch.arange(1, self.num_heads + 1)
|
| 337 |
+
self.A_log = nn.Parameter(torch.log(A))
|
| 338 |
+
self.A_log._no_weight_decay = True
|
| 339 |
+
self.norm = MambaRMSNormGated(self.intermediate_size, eps=self.layer_norm_epsilon, group_size=self.intermediate_size // self.n_groups)
|
| 340 |
+
self.D = nn.Parameter(torch.ones(self.num_heads))
|
| 341 |
+
self.D._no_weight_decay = True
|
| 342 |
+
|
| 343 |
+
self.out_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.use_bias)
|
| 344 |
+
self.use_bias = config.use_bias
|
| 345 |
+
|
| 346 |
+
if not is_fast_path_available:
|
| 347 |
+
logger.warning_once(
|
| 348 |
+
"The fast path is not available because on of `(selective_state_update, causal_conv1d_fn, causal_conv1d_update)`"
|
| 349 |
+
" is None. Falling back to the naive implementation. To install follow https://github.com/state-spaces/mamba/#installation and"
|
| 350 |
+
" https://github.com/Dao-AILab/causal-conv1d"
|
| 351 |
+
)
|
| 352 |
+
|
| 353 |
+
def cuda_kernels_forward(
|
| 354 |
+
self,
|
| 355 |
+
hidden_states: torch.Tensor,
|
| 356 |
+
cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
|
| 357 |
+
cache_position: Optional[torch.LongTensor] = None,
|
| 358 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 359 |
+
):
|
| 360 |
+
# 1. Gated MLP's linear projection
|
| 361 |
+
hidden_states = apply_mask_to_padding_states(hidden_states, attention_mask)
|
| 362 |
+
projected_states = self.in_proj(hidden_states)
|
| 363 |
+
|
| 364 |
+
# Set up dimensions for reshapes later
|
| 365 |
+
batch_size, seq_len, _ = hidden_states.shape
|
| 366 |
+
groups_time_state_size = self.n_groups * self.ssm_state_size
|
| 367 |
+
d_mlp = (
|
| 368 |
+
projected_states.shape[-1]
|
| 369 |
+
- 2 * self.intermediate_size
|
| 370 |
+
- 2 * self.n_groups * self.ssm_state_size
|
| 371 |
+
- self.num_heads
|
| 372 |
+
) // 2
|
| 373 |
+
|
| 374 |
+
# Single step calculations via cache
|
| 375 |
+
if cache_params is not None and cache_position is not None and cache_position[0] > 0:
|
| 376 |
+
_, _, gate, hidden_states_B_C, dt = projected_states.squeeze(1).split(
|
| 377 |
+
[d_mlp, d_mlp, self.intermediate_size, self.conv_dim, self.num_heads], dim=-1
|
| 378 |
+
)
|
| 379 |
+
|
| 380 |
+
# 2. Convolution sequence transformation
|
| 381 |
+
hidden_states_B_C = causal_conv1d_update(
|
| 382 |
+
hidden_states_B_C,
|
| 383 |
+
cache_params.conv_states[self.layer_idx],
|
| 384 |
+
self.conv1d.weight.squeeze(1),
|
| 385 |
+
self.conv1d.bias,
|
| 386 |
+
self.activation,
|
| 387 |
+
)
|
| 388 |
+
|
| 389 |
+
hidden_states, B, C = torch.split(
|
| 390 |
+
hidden_states_B_C,
|
| 391 |
+
[self.intermediate_size, groups_time_state_size, groups_time_state_size],
|
| 392 |
+
dim=-1,
|
| 393 |
+
)
|
| 394 |
+
|
| 395 |
+
# 3. SSM transformation
|
| 396 |
+
A = -torch.exp(self.A_log.float()) # (nheads,)
|
| 397 |
+
A = A[:, None, ...][:, :, None].expand(-1, self.head_dim, self.ssm_state_size).to(dtype=torch.float32)
|
| 398 |
+
dt = dt[:, :, None].expand(-1, -1, self.head_dim)
|
| 399 |
+
dt_bias = self.dt_bias[:, None, ...].expand(-1, self.head_dim)
|
| 400 |
+
D = self.D[:, None, ...].expand(-1, self.head_dim)
|
| 401 |
+
B = B.view(batch_size, self.n_groups, B.shape[1] // self.n_groups)
|
| 402 |
+
C = C.view(batch_size, self.n_groups, C.shape[1] // self.n_groups)
|
| 403 |
+
hidden_states_reshaped = hidden_states.view(batch_size, self.num_heads, self.head_dim)
|
| 404 |
+
hidden_states = selective_state_update(
|
| 405 |
+
cache_params.ssm_states[self.layer_idx],
|
| 406 |
+
hidden_states_reshaped,
|
| 407 |
+
dt,
|
| 408 |
+
A,
|
| 409 |
+
B,
|
| 410 |
+
C,
|
| 411 |
+
D,
|
| 412 |
+
z=None,
|
| 413 |
+
dt_bias=dt_bias,
|
| 414 |
+
dt_softplus=True,
|
| 415 |
+
)
|
| 416 |
+
hidden_states = hidden_states.view(batch_size, self.num_heads * self.head_dim)
|
| 417 |
+
hidden_states = self.norm(hidden_states, gate)
|
| 418 |
+
|
| 419 |
+
# 4. Final linear projection
|
| 420 |
+
out = self.out_proj(hidden_states)[:, None, ...]
|
| 421 |
+
|
| 422 |
+
# Fused calculations or step by step if no initialized cache is found
|
| 423 |
+
else:
|
| 424 |
+
A = -torch.exp(self.A_log.float()) # (num_heads) or (intermediate_size, state_size)
|
| 425 |
+
dt_limit_kwargs = {} if self.time_step_limit == (0.0, float("inf")) else {"dt_limit": self.time_step_limit}
|
| 426 |
+
|
| 427 |
+
# 2-4. Fused kernel for conv1d, SSM, and the final projection
|
| 428 |
+
if self.training and cache_params is None:
|
| 429 |
+
out = mamba_split_conv1d_scan_combined(
|
| 430 |
+
projected_states,
|
| 431 |
+
self.conv1d.weight.squeeze(1),
|
| 432 |
+
self.conv1d.bias,
|
| 433 |
+
self.dt_bias,
|
| 434 |
+
A,
|
| 435 |
+
D=self.D,
|
| 436 |
+
chunk_size=self.chunk_size,
|
| 437 |
+
seq_idx=None, # was seq_idx
|
| 438 |
+
activation=self.activation,
|
| 439 |
+
rmsnorm_weight=self.norm.weight,
|
| 440 |
+
rmsnorm_eps=self.norm.variance_epsilon,
|
| 441 |
+
outproj_weight=self.out_proj.weight,
|
| 442 |
+
outproj_bias=self.out_proj.bias,
|
| 443 |
+
headdim=self.head_dim,
|
| 444 |
+
ngroups=self.n_groups,
|
| 445 |
+
norm_before_gate=False,
|
| 446 |
+
return_final_states=False,
|
| 447 |
+
**dt_limit_kwargs,
|
| 448 |
+
)
|
| 449 |
+
|
| 450 |
+
else:
|
| 451 |
+
_, _, gate, hidden_states_B_C, dt = projected_states.split(
|
| 452 |
+
[d_mlp, d_mlp, self.intermediate_size, self.conv_dim, self.num_heads], dim=-1
|
| 453 |
+
)
|
| 454 |
+
|
| 455 |
+
# 2. Convolution sequence transformation
|
| 456 |
+
# Init cache
|
| 457 |
+
if cache_params is not None:
|
| 458 |
+
hidden_states_B_C_transposed = hidden_states_B_C.transpose(1, 2)
|
| 459 |
+
conv_states = nn.functional.pad(
|
| 460 |
+
hidden_states_B_C_transposed,
|
| 461 |
+
(cache_params.conv_kernel_size - hidden_states_B_C_transposed.shape[-1], 0),
|
| 462 |
+
)
|
| 463 |
+
cache_params.update_conv_state(
|
| 464 |
+
layer_idx=self.layer_idx, new_conv_state=conv_states, cache_init=True
|
| 465 |
+
)
|
| 466 |
+
|
| 467 |
+
if self.activation not in ["silu", "swish"]:
|
| 468 |
+
hidden_states_B_C = self.act(
|
| 469 |
+
self.conv1d(hidden_states_B_C.transpose(1, 2))[..., :seq_len].transpose(1, 2)
|
| 470 |
+
)
|
| 471 |
+
else:
|
| 472 |
+
hidden_states_B_C = causal_conv1d_fn(
|
| 473 |
+
x=hidden_states_B_C.transpose(1, 2),
|
| 474 |
+
weight=self.conv1d.weight.squeeze(1),
|
| 475 |
+
bias=self.conv1d.bias,
|
| 476 |
+
activation=self.activation,
|
| 477 |
+
).transpose(1, 2)
|
| 478 |
+
hidden_states_B_C = apply_mask_to_padding_states(hidden_states_B_C, attention_mask)
|
| 479 |
+
hidden_states, B, C = torch.split(
|
| 480 |
+
hidden_states_B_C,
|
| 481 |
+
[self.intermediate_size, groups_time_state_size, groups_time_state_size],
|
| 482 |
+
dim=-1,
|
| 483 |
+
)
|
| 484 |
+
|
| 485 |
+
# 3. SSM transformation
|
| 486 |
+
scan_output, ssm_state = mamba_chunk_scan_combined(
|
| 487 |
+
hidden_states.view(batch_size, seq_len, -1, self.head_dim),
|
| 488 |
+
dt,
|
| 489 |
+
A,
|
| 490 |
+
B.view(batch_size, seq_len, self.n_groups, -1),
|
| 491 |
+
C.view(batch_size, seq_len, self.n_groups, -1),
|
| 492 |
+
chunk_size=self.chunk_size,
|
| 493 |
+
D=self.D,
|
| 494 |
+
z=None,
|
| 495 |
+
seq_idx=None,
|
| 496 |
+
return_final_states=True,
|
| 497 |
+
dt_bias=self.dt_bias,
|
| 498 |
+
dt_softplus=True,
|
| 499 |
+
**dt_limit_kwargs,
|
| 500 |
+
)
|
| 501 |
+
|
| 502 |
+
# Init cache
|
| 503 |
+
if ssm_state is not None and cache_params is not None:
|
| 504 |
+
cache_params.update_ssm_state(layer_idx=self.layer_idx, new_ssm_state=ssm_state)
|
| 505 |
+
|
| 506 |
+
scan_output = scan_output.view(batch_size, seq_len, -1)
|
| 507 |
+
|
| 508 |
+
# Multiply "gate" branch and apply extra normalization layer
|
| 509 |
+
scan_output = self.norm(scan_output, gate)
|
| 510 |
+
|
| 511 |
+
# 4. Final linear projection
|
| 512 |
+
out = self.out_proj(scan_output)
|
| 513 |
+
return out
|
| 514 |
+
|
| 515 |
+
# fmt: off
|
| 516 |
+
def torch_forward(self, input_states, cache_params: Optional[HybridMambaAttentionDynamicCache]=None, cache_position:Optional[torch.LongTensor]=None, attention_mask: Optional[torch.Tensor]=None):
|
| 517 |
+
batch_size, seq_len, _ = input_states.shape
|
| 518 |
+
dtype = input_states.dtype
|
| 519 |
+
|
| 520 |
+
# 1. Gated MLP's linear projection
|
| 521 |
+
input_states = apply_mask_to_padding_states(input_states, attention_mask)
|
| 522 |
+
projected_states = self.in_proj(input_states)
|
| 523 |
+
d_mlp = (projected_states.shape[-1] - 2 * self.intermediate_size - 2 * self.n_groups * self.ssm_state_size-self.num_heads) // 2
|
| 524 |
+
_, _, gate, hidden_states_B_C, dt = projected_states.split(
|
| 525 |
+
[d_mlp, d_mlp, self.intermediate_size, self.conv_dim, self.num_heads], dim=-1
|
| 526 |
+
)
|
| 527 |
+
|
| 528 |
+
# 2. Convolution sequence transformation
|
| 529 |
+
if cache_params is not None and cache_position is not None and cache_position[0] > 0:
|
| 530 |
+
cache_params.update_conv_state(layer_idx=self.layer_idx, new_conv_state=hidden_states_B_C, cache_init=False)
|
| 531 |
+
|
| 532 |
+
# We need to guarantee that anything regarding the cache is on the same device
|
| 533 |
+
conv_states = cache_params.conv_states[self.layer_idx].to(device=self.conv1d.weight.device)
|
| 534 |
+
|
| 535 |
+
hidden_states_B_C = torch.sum(
|
| 536 |
+
conv_states * self.conv1d.weight.squeeze(1), dim=-1
|
| 537 |
+
)
|
| 538 |
+
if self.use_conv_bias:
|
| 539 |
+
hidden_states_B_C = hidden_states_B_C + self.conv1d.bias
|
| 540 |
+
hidden_states_B_C = self.act(hidden_states_B_C)
|
| 541 |
+
else:
|
| 542 |
+
# Init cache
|
| 543 |
+
if cache_params is not None:
|
| 544 |
+
hidden_states_B_C_transposed = hidden_states_B_C.transpose(1, 2)
|
| 545 |
+
conv_states = nn.functional.pad(
|
| 546 |
+
hidden_states_B_C_transposed, (cache_params.conv_kernel_size - hidden_states_B_C_transposed.shape[-1], 0)
|
| 547 |
+
)
|
| 548 |
+
cache_params.update_conv_state(layer_idx=self.layer_idx, new_conv_state=conv_states, cache_init=True)
|
| 549 |
+
|
| 550 |
+
hidden_states_B_C = self.act(self.conv1d(hidden_states_B_C.transpose(1, 2))[..., :seq_len].transpose(1, 2))
|
| 551 |
+
|
| 552 |
+
hidden_states_B_C = apply_mask_to_padding_states(hidden_states_B_C, attention_mask)
|
| 553 |
+
hidden_states, B, C = torch.split(
|
| 554 |
+
hidden_states_B_C,
|
| 555 |
+
[self.intermediate_size, self.n_groups * self.ssm_state_size, self.n_groups * self.ssm_state_size],
|
| 556 |
+
dim=-1
|
| 557 |
+
)
|
| 558 |
+
|
| 559 |
+
# 3. SSM transformation
|
| 560 |
+
A = -torch.exp(self.A_log.float()) # [num_heads]
|
| 561 |
+
if cache_params is not None and cache_position is not None and cache_position[0] > 0:
|
| 562 |
+
# We need to guarantee that anything regarding the cache is on the same device
|
| 563 |
+
cache_device = cache_params.ssm_states.device
|
| 564 |
+
|
| 565 |
+
# Note: there is no need to pad parameter matrices here, as there is just one new token
|
| 566 |
+
# for batched generation
|
| 567 |
+
dt = dt[:, 0, :][:, None, ...]
|
| 568 |
+
dt = dt.transpose(1, 2).expand(batch_size, dt.shape[-1], self.head_dim)
|
| 569 |
+
# [num_heads] -> [num_heads, head_dim]
|
| 570 |
+
dt_bias = self.dt_bias[..., None].expand(self.dt_bias.shape[0], self.head_dim)
|
| 571 |
+
|
| 572 |
+
dt = torch.nn.functional.softplus(dt + dt_bias.to(dt.dtype))
|
| 573 |
+
dt = torch.clamp(dt, self.time_step_limit[0], self.time_step_limit[1])
|
| 574 |
+
A = A[..., None, None].expand(self.num_heads, self.head_dim, self.ssm_state_size).to(dtype=torch.float32)
|
| 575 |
+
# [bsz, num_heads, head_dim, state_size]
|
| 576 |
+
dA = (torch.exp(dt[..., None] * A)).to(device=cache_device)
|
| 577 |
+
|
| 578 |
+
# Discretize B
|
| 579 |
+
# [bsz, n_groups * state_size] -> [bsz, n_groups, 1, state_size] ->
|
| 580 |
+
# -> [bsz, n_groups, group to head repetition factor, state_size] -> [bsz, num_heads, state_size]
|
| 581 |
+
B = B.reshape(batch_size, self.n_groups, -1)[..., None, :]
|
| 582 |
+
B = B.expand(batch_size, self.n_groups, self.num_heads // self.n_groups, B.shape[-1]).contiguous()
|
| 583 |
+
B = B.reshape(batch_size, -1, B.shape[-1])
|
| 584 |
+
# [bsz, num_heads, head_dim, state_size]
|
| 585 |
+
dB = dt[..., None] * B[..., None, :]
|
| 586 |
+
|
| 587 |
+
# Discretize x into dB
|
| 588 |
+
# [bsz, intermediate_size] -> [bsz, num_heads, head_dim]
|
| 589 |
+
hidden_states = hidden_states.reshape(batch_size, -1, self.head_dim)
|
| 590 |
+
dBx = (dB * hidden_states[..., None]).to(device=cache_device)
|
| 591 |
+
|
| 592 |
+
# State calculation
|
| 593 |
+
cache_params.update_ssm_state(
|
| 594 |
+
layer_idx=self.layer_idx,
|
| 595 |
+
new_ssm_state=cache_params.ssm_states[self.layer_idx] * dA + dBx
|
| 596 |
+
)
|
| 597 |
+
|
| 598 |
+
# Subsequent output
|
| 599 |
+
# [bsz, n_groups * state_size] -> [bsz, num_heads, state_size]
|
| 600 |
+
C = C.reshape(batch_size, self.n_groups, -1)[..., None, :]
|
| 601 |
+
C = C.expand(batch_size, self.n_groups, self.num_heads // self.n_groups, C.shape[-1]).contiguous()
|
| 602 |
+
C = C.reshape(batch_size, -1, C.shape[-1])
|
| 603 |
+
# [bsz, num_heads, head_dim]
|
| 604 |
+
|
| 605 |
+
ssm_states = cache_params.ssm_states[self.layer_idx].to(device=C.device, dtype=C.dtype) # Shape: [b, h, d, n]
|
| 606 |
+
# Reshape ssm_states to merge the first two dimensions
|
| 607 |
+
ssm_states_reshaped = ssm_states.view(batch_size * self.num_heads, self.head_dim, self.ssm_state_size) # Shape: [b*h, d, n]
|
| 608 |
+
C_reshaped = C.view(batch_size * self.num_heads, self.ssm_state_size, 1) # Shape: [b*h, n, 1]
|
| 609 |
+
y = torch.bmm(ssm_states_reshaped, C_reshaped)
|
| 610 |
+
y = y.view(batch_size, self.num_heads, self.head_dim)
|
| 611 |
+
|
| 612 |
+
# D skip connection
|
| 613 |
+
# [num_heads] -> [num_heads, head_dim]
|
| 614 |
+
D = self.D[..., None].expand(self.D.shape[0], self.head_dim)
|
| 615 |
+
y = (y + hidden_states * D).to(y.dtype)
|
| 616 |
+
|
| 617 |
+
# [bsz, num_heads, head_dim] -> [bsz, 1, intermediate_size]
|
| 618 |
+
y = y.reshape(batch_size, -1)[:, None, ...]
|
| 619 |
+
else:
|
| 620 |
+
# begin ssd naive implementation without einsums
|
| 621 |
+
dt = nn.functional.softplus(dt + self.dt_bias)
|
| 622 |
+
dt = torch.clamp(dt, self.time_step_limit[0], self.time_step_limit[1])
|
| 623 |
+
hidden_states = hidden_states.reshape(batch_size, seq_len, -1, self.head_dim).float()
|
| 624 |
+
B = B.reshape(batch_size, seq_len, -1, self.ssm_state_size).float()
|
| 625 |
+
C = C.reshape(batch_size, seq_len, -1, self.ssm_state_size).float()
|
| 626 |
+
B = B.repeat(1, 1, self.num_heads // self.n_groups, 1)
|
| 627 |
+
C = C.repeat(1, 1, self.num_heads // self.n_groups, 1)
|
| 628 |
+
pad_size = (self.chunk_size - seq_len % self.chunk_size) % self.chunk_size
|
| 629 |
+
|
| 630 |
+
D_residual = self.D[..., None] * pad_tensor_by_size(hidden_states, pad_size)
|
| 631 |
+
|
| 632 |
+
# Discretize x and A
|
| 633 |
+
hidden_states = hidden_states * dt[..., None]
|
| 634 |
+
A = A.to(hidden_states.dtype) * dt
|
| 635 |
+
|
| 636 |
+
# Rearrange into blocks/chunks
|
| 637 |
+
hidden_states, A, B, C = [reshape_into_chunks(t, pad_size, self.chunk_size) for t in (hidden_states, A, B, C)]
|
| 638 |
+
|
| 639 |
+
# [bsz, -1, chunk_size, num_heads] -> [bsz, num_heads, -1, chunk_size]
|
| 640 |
+
A = A.permute(0, 3, 1, 2)
|
| 641 |
+
A_cumsum = torch.cumsum(A, dim=-1)
|
| 642 |
+
|
| 643 |
+
# 1. Compute the output for each intra-chunk (diagonal blocks)
|
| 644 |
+
# This is the analog of a causal mask
|
| 645 |
+
L = torch.exp(segment_sum(A))
|
| 646 |
+
|
| 647 |
+
# Contraction of C and B to get G (attention-weights like)
|
| 648 |
+
G_intermediate = C[:, :, :, None, :, :] * B[:, :, None, :, :, :] # shape: (b, c, l, s, h, n)
|
| 649 |
+
G = G_intermediate.sum(dim=-1) # shape: (b, c, l, s, h)
|
| 650 |
+
|
| 651 |
+
# Compute M, equivalent to applying attention mask to weights
|
| 652 |
+
M_intermediate = G[..., None] * L.permute(0, 2, 3, 4, 1)[..., None]
|
| 653 |
+
M = M_intermediate.sum(dim=-1)
|
| 654 |
+
|
| 655 |
+
# Compute Y_diag (apply to values)
|
| 656 |
+
Y_diag = (M[..., None] * hidden_states[:, :, None]).sum(dim=3)
|
| 657 |
+
|
| 658 |
+
# 2. Compute the state for each intra-chunk
|
| 659 |
+
# (right term of low-rank factorization of off-diagonal blocks; B terms)
|
| 660 |
+
decay_states = torch.exp((A_cumsum[:, :, :, -1:] - A_cumsum))
|
| 661 |
+
B_decay = B * decay_states.permute(0, -2, -1, 1)[..., None]
|
| 662 |
+
states = (B_decay[..., None, :] * hidden_states[..., None]).sum(dim=2)
|
| 663 |
+
|
| 664 |
+
# 3. Compute the inter-chunk SSM recurrence; produces correct SSM states at chunk boundaries
|
| 665 |
+
# (middle term of factorization of off-diag blocks; A terms)
|
| 666 |
+
if cache_params is not None and cache_position is not None and cache_position[0] > 0:
|
| 667 |
+
previous_states = cache_params.ssm_states[self.layer_idx][:, None, ...].to(device=states.device)
|
| 668 |
+
else:
|
| 669 |
+
previous_states = torch.zeros_like(states[:, :1])
|
| 670 |
+
states = torch.cat([previous_states, states], dim=1)
|
| 671 |
+
decay_chunk = torch.exp(segment_sum(nn.functional.pad(A_cumsum[:, :, :, -1], (1, 0))))
|
| 672 |
+
decay_chunk = decay_chunk.transpose(1, 3)
|
| 673 |
+
new_states = (decay_chunk[..., None, None] * states[:, :, None, ...]).sum(dim=1)
|
| 674 |
+
states, ssm_state = new_states[:, :-1], new_states[:, -1]
|
| 675 |
+
|
| 676 |
+
# 4. Compute state -> output conversion per chunk
|
| 677 |
+
# (left term of low-rank factorization of off-diagonal blocks; C terms)
|
| 678 |
+
state_decay_out = torch.exp(A_cumsum)
|
| 679 |
+
C_times_states = (C[..., None, :] * states[:, :, None, ...])
|
| 680 |
+
state_decay_out_permuted = state_decay_out.permute(0, 2, 3, 1)
|
| 681 |
+
Y_off = (C_times_states.sum(-1) * state_decay_out_permuted[..., None])
|
| 682 |
+
|
| 683 |
+
# Add output of intra-chunk and inter-chunk terms (diagonal and off-diagonal blocks)
|
| 684 |
+
y = Y_diag + Y_off
|
| 685 |
+
# [bsz, -1, self.chunk_size, num_heads, head_dim] -> [bsz, (padded) seq_len, num_heads, head_dim]
|
| 686 |
+
y = y.reshape(batch_size, -1, self.num_heads, self.head_dim)
|
| 687 |
+
|
| 688 |
+
y = y + D_residual
|
| 689 |
+
# Cutting off padded chunks
|
| 690 |
+
if pad_size > 0:
|
| 691 |
+
y = y[:, :seq_len, :, :]
|
| 692 |
+
y = y.reshape(batch_size, seq_len, -1)
|
| 693 |
+
|
| 694 |
+
# Init cache
|
| 695 |
+
if ssm_state is not None and cache_params is not None:
|
| 696 |
+
cache_params.update_ssm_state(layer_idx=self.layer_idx, new_ssm_state=ssm_state)
|
| 697 |
+
|
| 698 |
+
scan_output = self.norm(y, gate)
|
| 699 |
+
|
| 700 |
+
# end ssd naive
|
| 701 |
+
|
| 702 |
+
# 4. Final linear projection
|
| 703 |
+
contextualized_states = self.out_proj(scan_output.to(dtype)) # [batch, seq_len, hidden_size]
|
| 704 |
+
return contextualized_states
|
| 705 |
+
# fmt: on
|
| 706 |
+
|
| 707 |
+
def forward(
|
| 708 |
+
self,
|
| 709 |
+
hidden_states,
|
| 710 |
+
cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
|
| 711 |
+
cache_position: Optional[torch.LongTensor] = None,
|
| 712 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 713 |
+
):
|
| 714 |
+
if is_fast_path_available and "cuda" in self.in_proj.weight.device.type:
|
| 715 |
+
return self.cuda_kernels_forward(hidden_states, cache_params, cache_position, attention_mask)
|
| 716 |
+
dtype = hidden_states.dtype
|
| 717 |
+
if attention_mask is not None and attention_mask.shape[1] > 1 and attention_mask.shape[0] > 1:
|
| 718 |
+
# tune out hidden states for pad tokens, see https://github.com/state-spaces/mamba/issues/66
|
| 719 |
+
hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)
|
| 720 |
+
|
| 721 |
+
return self.torch_forward(hidden_states, cache_params, cache_position, attention_mask)
|
| 722 |
+
|
| 723 |
+
|
| 724 |
+
class NemotronHRMSNorm(nn.Module):
|
| 725 |
+
def __init__(self, hidden_size, eps=1e-6):
|
| 726 |
+
"""
|
| 727 |
+
NemotronHRMSNorm is equivalent to T5LayerNorm and LlamaRMSNorm
|
| 728 |
+
"""
|
| 729 |
+
super().__init__()
|
| 730 |
+
self.weight = nn.Parameter(torch.ones(hidden_size))
|
| 731 |
+
self.variance_epsilon = eps
|
| 732 |
+
|
| 733 |
+
def forward(self, hidden_states):
|
| 734 |
+
input_dtype = hidden_states.dtype
|
| 735 |
+
hidden_states = hidden_states.to(torch.float32)
|
| 736 |
+
variance = hidden_states.pow(2).mean(-1, keepdim=True)
|
| 737 |
+
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
|
| 738 |
+
# Weights are in float32
|
| 739 |
+
return (self.weight.to(torch.float32) * hidden_states).to(input_dtype)
|
| 740 |
+
|
| 741 |
+
class NemotronHBlock(nn.Module):
|
| 742 |
+
def __init__(self, config, layer_idx):
|
| 743 |
+
super().__init__()
|
| 744 |
+
self.config = config
|
| 745 |
+
self.layer_idx = layer_idx
|
| 746 |
+
self.residual_in_fp32 = config.residual_in_fp32
|
| 747 |
+
self.norm = NemotronHRMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
|
| 748 |
+
|
| 749 |
+
# M: Mamba2, *: Attention, -: MLP
|
| 750 |
+
self.block_type = config.layers_block_type[layer_idx]
|
| 751 |
+
if self.block_type == "mamba":
|
| 752 |
+
self.mixer = NemotronHMamba2Mixer(config, layer_idx=layer_idx)
|
| 753 |
+
elif self.block_type == "attention":
|
| 754 |
+
self.mixer = NEMOTRONH_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx=layer_idx)
|
| 755 |
+
elif self.block_type == "mlp":
|
| 756 |
+
self.mixer = NemotronHMLP(config, layer_idx=layer_idx)
|
| 757 |
+
else:
|
| 758 |
+
raise ValueError(f"Invalid layer pattern {config.hybrid_override_pattern[layer_idx]}")
|
| 759 |
+
|
| 760 |
+
def forward(
|
| 761 |
+
self,
|
| 762 |
+
hidden_states,
|
| 763 |
+
cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
|
| 764 |
+
cache_position: Optional[torch.LongTensor] = None,
|
| 765 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 766 |
+
):
|
| 767 |
+
with torch.cuda.stream(torch.cuda.default_stream(hidden_states.device)):
|
| 768 |
+
# * Use torch.cuda.stream() to avoid NaN issues when using multiple GPUs
|
| 769 |
+
residual = hidden_states
|
| 770 |
+
hidden_states = self.norm(hidden_states.to(dtype=self.norm.weight.dtype))
|
| 771 |
+
if self.residual_in_fp32:
|
| 772 |
+
residual = residual.to(torch.float32)
|
| 773 |
+
|
| 774 |
+
if self.block_type == "mamba":
|
| 775 |
+
hidden_states = self.mixer(
|
| 776 |
+
hidden_states, cache_params=cache_params, cache_position=cache_position
|
| 777 |
+
)
|
| 778 |
+
elif self.block_type == "attention":
|
| 779 |
+
hidden_states = self.mixer(
|
| 780 |
+
hidden_states, cache_position=cache_position
|
| 781 |
+
)
|
| 782 |
+
hidden_states = hidden_states[0]
|
| 783 |
+
elif self.block_type == "mlp":
|
| 784 |
+
hidden_states = self.mixer(
|
| 785 |
+
hidden_states
|
| 786 |
+
)
|
| 787 |
+
else:
|
| 788 |
+
raise ValueError(f"Invalid block_type: {self.block_type}")
|
| 789 |
+
|
| 790 |
+
hidden_states = residual + hidden_states
|
| 791 |
+
return hidden_states
|
| 792 |
+
|
| 793 |
+
|
| 794 |
+
# Copied from transformers.models.nemotron.modeling_nemotron Nemotron->NemotronH
|
| 795 |
+
class NemotronHMLP(nn.Module):
|
| 796 |
+
def __init__(self, config, layer_idx: Optional[int] = None):
|
| 797 |
+
super().__init__()
|
| 798 |
+
self.config = config
|
| 799 |
+
self.layer_idx = layer_idx
|
| 800 |
+
if layer_idx is None:
|
| 801 |
+
logger.warning_once(
|
| 802 |
+
f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
|
| 803 |
+
"lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
|
| 804 |
+
"when creating this class."
|
| 805 |
+
)
|
| 806 |
+
self.hidden_size = config.hidden_size
|
| 807 |
+
#intermediate_size = config.expand * config.hidden_size
|
| 808 |
+
self.intermediate_size = config.intermediate_size
|
| 809 |
+
self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
|
| 810 |
+
self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.mlp_bias)
|
| 811 |
+
self.act_fn = ACT2FN[config.mlp_hidden_act]
|
| 812 |
+
|
| 813 |
+
def forward(self, x):
|
| 814 |
+
return self.down_proj(self.act_fn(self.up_proj(x)))
|
| 815 |
+
|
| 816 |
+
|
| 817 |
+
# Copied from transformers.models.llama.modeling_llama.repeat_kv
|
| 818 |
+
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
| 819 |
+
"""
|
| 820 |
+
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
|
| 821 |
+
num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
|
| 822 |
+
"""
|
| 823 |
+
batch, num_key_value_heads, slen, head_dim = hidden_states.shape
|
| 824 |
+
if n_rep == 1:
|
| 825 |
+
return hidden_states
|
| 826 |
+
hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
|
| 827 |
+
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
| 828 |
+
|
| 829 |
+
|
| 830 |
+
class NemotronHAttention(nn.Module):
|
| 831 |
+
"""Multi-headed attention from 'Attention Is All You Need' paper"""
|
| 832 |
+
|
| 833 |
+
def __init__(self, config: NemotronHConfig, layer_idx: Optional[int] = None):
|
| 834 |
+
super().__init__()
|
| 835 |
+
self.config = config
|
| 836 |
+
self.layer_idx = layer_idx
|
| 837 |
+
if layer_idx is None:
|
| 838 |
+
logger.warning_once(
|
| 839 |
+
f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
|
| 840 |
+
"lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
|
| 841 |
+
"when creating this class."
|
| 842 |
+
)
|
| 843 |
+
|
| 844 |
+
self.attention_dropout = config.attention_dropout
|
| 845 |
+
self.hidden_size = config.hidden_size
|
| 846 |
+
self.num_heads = config.num_attention_heads
|
| 847 |
+
if config.head_dim is not None:
|
| 848 |
+
self.head_dim = config.head_dim
|
| 849 |
+
else:
|
| 850 |
+
self.head_dim = config.hidden_size // config.num_attention_heads
|
| 851 |
+
self.num_key_value_heads = config.num_key_value_heads
|
| 852 |
+
self.num_key_value_groups = self.num_heads // self.num_key_value_heads
|
| 853 |
+
self.max_position_embeddings = config.max_position_embeddings
|
| 854 |
+
self.is_causal = True
|
| 855 |
+
|
| 856 |
+
self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
|
| 857 |
+
self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
|
| 858 |
+
self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
|
| 859 |
+
self.o_proj = nn.Linear(self.head_dim * self.num_heads, self.hidden_size, bias=config.attention_bias)
|
| 860 |
+
|
| 861 |
+
def forward(
|
| 862 |
+
self,
|
| 863 |
+
hidden_states: torch.Tensor,
|
| 864 |
+
# position_embeddings: Tuple[torch.Tensor, torch.Tensor], #TODO
|
| 865 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 866 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 867 |
+
past_key_value: Optional[HybridMambaAttentionDynamicCache] = None,
|
| 868 |
+
output_attentions: bool = False,
|
| 869 |
+
use_cache: bool = False,
|
| 870 |
+
cache_position: Optional[torch.LongTensor] = None,
|
| 871 |
+
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
| 872 |
+
bsz, q_len, _ = hidden_states.size()
|
| 873 |
+
|
| 874 |
+
query_states = self.q_proj(hidden_states)
|
| 875 |
+
key_states = self.k_proj(hidden_states)
|
| 876 |
+
value_states = self.v_proj(hidden_states)
|
| 877 |
+
|
| 878 |
+
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 879 |
+
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 880 |
+
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 881 |
+
|
| 882 |
+
if past_key_value is not None:
|
| 883 |
+
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)
|
| 884 |
+
|
| 885 |
+
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
| 886 |
+
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
| 887 |
+
|
| 888 |
+
causal_mask = attention_mask
|
| 889 |
+
if attention_mask is not None: # no matter the length, we just slice it
|
| 890 |
+
causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
|
| 891 |
+
|
| 892 |
+
if query_states.device.type == "cuda" and attention_mask is not None:
|
| 893 |
+
query_states = query_states.contiguous()
|
| 894 |
+
key_states = key_states.contiguous()
|
| 895 |
+
value_states = value_states.contiguous()
|
| 896 |
+
|
| 897 |
+
is_causal = True if causal_mask is None and q_len > 1 else False
|
| 898 |
+
|
| 899 |
+
attn_output = torch.nn.functional.scaled_dot_product_attention(
|
| 900 |
+
query_states,
|
| 901 |
+
key_states,
|
| 902 |
+
value_states,
|
| 903 |
+
attn_mask=causal_mask,
|
| 904 |
+
dropout_p=self.attention_dropout if self.training else 0.0,
|
| 905 |
+
is_causal=is_causal,
|
| 906 |
+
)
|
| 907 |
+
attn_output = attn_output.transpose(1, 2).contiguous()
|
| 908 |
+
#attn_output = attn_output.view(bsz, q_len, self.hidden_size)
|
| 909 |
+
attn_output = attn_output.view(bsz, q_len, self.num_heads * self.head_dim)
|
| 910 |
+
|
| 911 |
+
attn_output = self.o_proj(attn_output)
|
| 912 |
+
|
| 913 |
+
return attn_output, None, past_key_value
|
| 914 |
+
|
| 915 |
+
|
| 916 |
+
# Adapted from transformers.models.mistral.modeling_mistral.MistralFlashAttention2 with Mistral->Jamba
|
| 917 |
+
#class JambaFlashAttention2(JambaAttention):
|
| 918 |
+
class NemotronHFlashAttention2(NemotronHAttention):
|
| 919 |
+
"""
|
| 920 |
+
Jamba flash attention module. This module inherits from `JambaAttention` as the weights of the module stays
|
| 921 |
+
untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
|
| 922 |
+
flash attention and deal with padding tokens in case the input contains any of them.
|
| 923 |
+
"""
|
| 924 |
+
def __init__(self, *args, **kwargs):
|
| 925 |
+
super().__init__(*args, **kwargs)
|
| 926 |
+
|
| 927 |
+
# TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
|
| 928 |
+
# flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
|
| 929 |
+
# Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
|
| 930 |
+
self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
|
| 931 |
+
|
| 932 |
+
def forward(
|
| 933 |
+
self,
|
| 934 |
+
hidden_states: torch.Tensor,
|
| 935 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 936 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 937 |
+
past_key_value: Optional[HybridMambaAttentionDynamicCache] = None,
|
| 938 |
+
output_attentions: bool = False,
|
| 939 |
+
use_cache: bool = False,
|
| 940 |
+
cache_position: Optional[torch.LongTensor] = None,
|
| 941 |
+
**kwargs,
|
| 942 |
+
):
|
| 943 |
+
bsz, q_len, _ = hidden_states.size()
|
| 944 |
+
|
| 945 |
+
query_states = self.q_proj(hidden_states)
|
| 946 |
+
key_states = self.k_proj(hidden_states)
|
| 947 |
+
value_states = self.v_proj(hidden_states)
|
| 948 |
+
|
| 949 |
+
# Flash attention requires the input to have the shape
|
| 950 |
+
# batch_size x seq_length x head_dim x hidden_dim
|
| 951 |
+
# therefore we just need to keep the original shape
|
| 952 |
+
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim)
|
| 953 |
+
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 954 |
+
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 955 |
+
|
| 956 |
+
if past_key_value is not None:
|
| 957 |
+
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)
|
| 958 |
+
|
| 959 |
+
# repeat k/v heads if n_kv_heads < n_heads
|
| 960 |
+
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
| 961 |
+
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
| 962 |
+
dropout_rate = 0.0 if not self.training else self.attention_dropout
|
| 963 |
+
|
| 964 |
+
# In PEFT, usually we cast the layer norms in float32 for training stability reasons
|
| 965 |
+
# therefore the input hidden states gets silently casted in float32. Hence, we need
|
| 966 |
+
# cast them back in float16 just to be sure everything works as expected.
|
| 967 |
+
input_dtype = query_states.dtype
|
| 968 |
+
if input_dtype == torch.float32:
|
| 969 |
+
if torch.is_autocast_enabled():
|
| 970 |
+
target_dtype = torch.get_autocast_gpu_dtype()
|
| 971 |
+
# Handle the case where the model is quantized
|
| 972 |
+
elif hasattr(self.config, "_pre_quantization_dtype"):
|
| 973 |
+
target_dtype = self.config._pre_quantization_dtype
|
| 974 |
+
else:
|
| 975 |
+
target_dtype = self.q_proj.weight.dtype
|
| 976 |
+
|
| 977 |
+
logger.warning_once(
|
| 978 |
+
f"The input hidden states seems to be silently casted in float32, this might be related to"
|
| 979 |
+
f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
|
| 980 |
+
f" {target_dtype}."
|
| 981 |
+
)
|
| 982 |
+
|
| 983 |
+
query_states = query_states.to(target_dtype)
|
| 984 |
+
key_states = key_states.to(target_dtype)
|
| 985 |
+
value_states = value_states.to(target_dtype)
|
| 986 |
+
|
| 987 |
+
# Reashape to the expected shape for Flash Attention
|
| 988 |
+
key_states = key_states.transpose(1, 2)
|
| 989 |
+
value_states = value_states.transpose(1, 2)
|
| 990 |
+
|
| 991 |
+
attn_output = _flash_attention_forward(
|
| 992 |
+
query_states,
|
| 993 |
+
key_states,
|
| 994 |
+
value_states,
|
| 995 |
+
attention_mask,
|
| 996 |
+
q_len,
|
| 997 |
+
dropout=dropout_rate,
|
| 998 |
+
sliding_window=getattr(self.config, "sliding_window", None),
|
| 999 |
+
is_causal=self.is_causal,
|
| 1000 |
+
use_top_left_mask=self._flash_attn_uses_top_left_mask,
|
| 1001 |
+
)
|
| 1002 |
+
|
| 1003 |
+
#attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
|
| 1004 |
+
attn_output = attn_output.reshape(bsz, q_len, self.num_heads * self.head_dim).contiguous()
|
| 1005 |
+
attn_output = self.o_proj(attn_output)
|
| 1006 |
+
|
| 1007 |
+
if not output_attentions:
|
| 1008 |
+
attn_weights = None
|
| 1009 |
+
|
| 1010 |
+
return attn_output, attn_weights, past_key_value
|
| 1011 |
+
|
| 1012 |
+
|
| 1013 |
+
# Adapted from transformers.models.mistral.modeling_mistral.MistralSdpaAttention with Mistral->Jamba
|
| 1014 |
+
#class JambaSdpaAttention(JambaAttention):
|
| 1015 |
+
class NemotronHSdpaAttention(NemotronHAttention):
|
| 1016 |
+
"""
|
| 1017 |
+
Jamba attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
|
| 1018 |
+
`JambaAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
|
| 1019 |
+
SDPA API.
|
| 1020 |
+
"""
|
| 1021 |
+
|
| 1022 |
+
# Adapted from NemotronHAttention.forward
|
| 1023 |
+
def forward(
|
| 1024 |
+
self,
|
| 1025 |
+
hidden_states: torch.Tensor,
|
| 1026 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 1027 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 1028 |
+
past_key_value: Optional[HybridMambaAttentionDynamicCache] = None,
|
| 1029 |
+
output_attentions: bool = False,
|
| 1030 |
+
use_cache: bool = False,
|
| 1031 |
+
cache_position: Optional[torch.LongTensor] = None,
|
| 1032 |
+
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
| 1033 |
+
if output_attentions:
|
| 1034 |
+
# TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
|
| 1035 |
+
logger.warning_once(
|
| 1036 |
+
"NemotronHModel is using NemotronHSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
|
| 1037 |
+
'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
|
| 1038 |
+
)
|
| 1039 |
+
return super().forward(
|
| 1040 |
+
hidden_states=hidden_states,
|
| 1041 |
+
attention_mask=attention_mask,
|
| 1042 |
+
position_ids=position_ids,
|
| 1043 |
+
past_key_value=past_key_value,
|
| 1044 |
+
output_attentions=output_attentions,
|
| 1045 |
+
use_cache=use_cache,
|
| 1046 |
+
)
|
| 1047 |
+
|
| 1048 |
+
bsz, q_len, _ = hidden_states.size()
|
| 1049 |
+
|
| 1050 |
+
query_states = self.q_proj(hidden_states)
|
| 1051 |
+
key_states = self.k_proj(hidden_states)
|
| 1052 |
+
value_states = self.v_proj(hidden_states)
|
| 1053 |
+
|
| 1054 |
+
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 1055 |
+
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 1056 |
+
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 1057 |
+
|
| 1058 |
+
if past_key_value is not None:
|
| 1059 |
+
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)
|
| 1060 |
+
|
| 1061 |
+
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
| 1062 |
+
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
| 1063 |
+
|
| 1064 |
+
causal_mask = attention_mask
|
| 1065 |
+
if attention_mask is not None:
|
| 1066 |
+
causal_mask = causal_mask[:, :, :, : key_states.shape[-2]]
|
| 1067 |
+
|
| 1068 |
+
# SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
|
| 1069 |
+
# Reference: https://github.com/pytorch/pytorch/issues/112577.
|
| 1070 |
+
if query_states.device.type == "cuda" and attention_mask is not None:
|
| 1071 |
+
query_states = query_states.contiguous()
|
| 1072 |
+
key_states = key_states.contiguous()
|
| 1073 |
+
value_states = value_states.contiguous()
|
| 1074 |
+
|
| 1075 |
+
# We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
|
| 1076 |
+
# in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
|
| 1077 |
+
# The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
|
| 1078 |
+
is_causal = True if self.is_causal and causal_mask is None and q_len > 1 else False
|
| 1079 |
+
|
| 1080 |
+
attn_output = torch.nn.functional.scaled_dot_product_attention(
|
| 1081 |
+
query_states,
|
| 1082 |
+
key_states,
|
| 1083 |
+
value_states,
|
| 1084 |
+
attn_mask=causal_mask,
|
| 1085 |
+
dropout_p=self.attention_dropout if self.training else 0.0,
|
| 1086 |
+
is_causal=is_causal,
|
| 1087 |
+
)
|
| 1088 |
+
|
| 1089 |
+
attn_output = attn_output.transpose(1, 2).contiguous()
|
| 1090 |
+
attn_output = attn_output.view(bsz, q_len, self.hidden_size)
|
| 1091 |
+
|
| 1092 |
+
attn_output = self.o_proj(attn_output)
|
| 1093 |
+
|
| 1094 |
+
return attn_output, None, past_key_value
|
| 1095 |
+
|
| 1096 |
+
|
| 1097 |
+
NEMOTRONH_ATTENTION_CLASSES = {
|
| 1098 |
+
"eager": NemotronHAttention,
|
| 1099 |
+
"flash_attention_2": NemotronHFlashAttention2,
|
| 1100 |
+
"sdpa": NemotronHSdpaAttention,
|
| 1101 |
+
}
|
| 1102 |
+
|
| 1103 |
+
# Copied from transformers.models.mamba.modeling_mamba2.Mamba2PreTrainedModel
|
| 1104 |
+
class NemotronHPreTrainedModel(PreTrainedModel):
|
| 1105 |
+
"""
|
| 1106 |
+
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
|
| 1107 |
+
models.
|
| 1108 |
+
"""
|
| 1109 |
+
|
| 1110 |
+
config_class = NemotronHConfig
|
| 1111 |
+
base_model_prefix = "backbone"
|
| 1112 |
+
_no_split_modules = ["NemotronHBlock"]
|
| 1113 |
+
supports_gradient_checkpointing = True
|
| 1114 |
+
_is_stateful = True
|
| 1115 |
+
_supports_flash_attn_2 = True
|
| 1116 |
+
|
| 1117 |
+
def _init_weights(self, module):
|
| 1118 |
+
"""Initialize the weights."""
|
| 1119 |
+
if isinstance(module, NemotronHMamba2Mixer):
|
| 1120 |
+
module.A_log._no_weight_decay = True
|
| 1121 |
+
module.D._no_weight_decay = True
|
| 1122 |
+
|
| 1123 |
+
dt = torch.exp(
|
| 1124 |
+
torch.rand(self.config.mamba_num_heads)
|
| 1125 |
+
* (math.log(self.config.time_step_max) - math.log(self.config.time_step_min))
|
| 1126 |
+
+ math.log(self.config.time_step_min)
|
| 1127 |
+
).clamp(min=self.config.time_step_floor)
|
| 1128 |
+
|
| 1129 |
+
# # Inverse of softplus: https://github.com/pytorch/pytorch/issues/72759
|
| 1130 |
+
inv_dt = dt + torch.log(-torch.expm1(-dt))
|
| 1131 |
+
with torch.no_grad():
|
| 1132 |
+
module.dt_bias.copy_(inv_dt)
|
| 1133 |
+
module.dt_bias._no_reinit = True
|
| 1134 |
+
|
| 1135 |
+
if isinstance(module, nn.Linear):
|
| 1136 |
+
if module.bias is not None:
|
| 1137 |
+
if not getattr(module.bias, "_no_reinit", False):
|
| 1138 |
+
nn.init.zeros_(module.bias)
|
| 1139 |
+
elif isinstance(module, nn.Embedding):
|
| 1140 |
+
nn.init.normal_(module.weight, std=self.config.initializer_range)
|
| 1141 |
+
|
| 1142 |
+
# TODO: Check
|
| 1143 |
+
if self.config.rescale_prenorm_residual:
|
| 1144 |
+
# Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
|
| 1145 |
+
# > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
|
| 1146 |
+
# > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
|
| 1147 |
+
# > -- GPT-2 :: https://openai.com/blog/better-language-models/
|
| 1148 |
+
#
|
| 1149 |
+
# Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
|
| 1150 |
+
for name, p in module.named_parameters():
|
| 1151 |
+
if name in ["out_proj.weight"]:
|
| 1152 |
+
# Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
|
| 1153 |
+
# Following Pytorch init, except scale by 1/sqrt(2 * n_layer)
|
| 1154 |
+
# We need to reinit p since this code could be called multiple times
|
| 1155 |
+
# Having just p *= scale would repeatedly scale it down
|
| 1156 |
+
nn.init.kaiming_uniform_(p, a=math.sqrt(5))
|
| 1157 |
+
with torch.no_grad():
|
| 1158 |
+
p /= math.sqrt(self.config.num_hidden_layers)
|
| 1159 |
+
|
| 1160 |
+
|
| 1161 |
+
@dataclass
|
| 1162 |
+
# Copied from transformers.models.mamba.modeling_mamba2.Mamba2Output with MAMBA2->NemotronH,Mamba2->NemotronH
|
| 1163 |
+
class NemotronHOutput(ModelOutput):
|
| 1164 |
+
"""
|
| 1165 |
+
Class for the NemotronH model outputs.
|
| 1166 |
+
|
| 1167 |
+
Args:
|
| 1168 |
+
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
|
| 1169 |
+
Sequence of hidden-states at the output of the last layer of the model.
|
| 1170 |
+
cache_params (`HybridMambaAttentionDynamicCache`):
|
| 1171 |
+
The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to
|
| 1172 |
+
avoid providing the old `input_ids`.
|
| 1173 |
+
|
| 1174 |
+
Includes both the State space model state matrices after the selective scan, and the Convolutional states
|
| 1175 |
+
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
|
| 1176 |
+
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
|
| 1177 |
+
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
|
| 1178 |
+
|
| 1179 |
+
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
|
| 1180 |
+
"""
|
| 1181 |
+
|
| 1182 |
+
last_hidden_state: Optional[torch.FloatTensor] = None
|
| 1183 |
+
cache_params: Optional[HybridMambaAttentionDynamicCache] = None
|
| 1184 |
+
hidden_states: Optional[Tuple[torch.FloatTensor]] = None
|
| 1185 |
+
attentions: Optional[Tuple[torch.FloatTensor]] = None
|
| 1186 |
+
|
| 1187 |
+
|
| 1188 |
+
@dataclass
|
| 1189 |
+
# Copied from transformers.models.mamba2.modeling_mamba2.MambaCausalLMOutput with Mamba2->NemotronH
|
| 1190 |
+
class NemotronHCausalLMOutput(ModelOutput):
|
| 1191 |
+
"""
|
| 1192 |
+
Base class for causal language model (or autoregressive) outputs.
|
| 1193 |
+
|
| 1194 |
+
Args:
|
| 1195 |
+
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
|
| 1196 |
+
Language modeling loss (for next-token prediction).
|
| 1197 |
+
logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
|
| 1198 |
+
Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
|
| 1199 |
+
cache_params (`HybridMambaAttentionDynamicCache`):
|
| 1200 |
+
The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to
|
| 1201 |
+
avoid providing the old `input_ids`.
|
| 1202 |
+
|
| 1203 |
+
Includes both the State space model state matrices after the selective scan, and the Convolutional states
|
| 1204 |
+
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
|
| 1205 |
+
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
|
| 1206 |
+
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
|
| 1207 |
+
|
| 1208 |
+
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
|
| 1209 |
+
"""
|
| 1210 |
+
|
| 1211 |
+
loss: Optional[torch.FloatTensor] = None
|
| 1212 |
+
logits: Optional[torch.FloatTensor] = None
|
| 1213 |
+
cache_params: Optional[HybridMambaAttentionDynamicCache] = None
|
| 1214 |
+
hidden_states: Optional[Tuple[torch.FloatTensor]] = None
|
| 1215 |
+
attentions: Optional[Tuple[torch.FloatTensor]] = None
|
| 1216 |
+
|
| 1217 |
+
|
| 1218 |
+
NEMOTRONH_START_DOCSTRING = r"""
|
| 1219 |
+
|
| 1220 |
+
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
| 1221 |
+
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
| 1222 |
+
etc.)
|
| 1223 |
+
|
| 1224 |
+
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
|
| 1225 |
+
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
|
| 1226 |
+
and behavior.
|
| 1227 |
+
|
| 1228 |
+
Parameters:
|
| 1229 |
+
config ([`NemotronHConfig`]): Model configuration class with all the parameters of the model.
|
| 1230 |
+
Initializing with a config file does not load the weights associated with the model, only the
|
| 1231 |
+
configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
|
| 1232 |
+
"""
|
| 1233 |
+
|
| 1234 |
+
NEMOTRONH_INPUTS_DOCSTRING = r"""
|
| 1235 |
+
Args:
|
| 1236 |
+
input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`, *optional*):
|
| 1237 |
+
Indices of input sequence tokens in the vocabulary.
|
| 1238 |
+
|
| 1239 |
+
If `cache_params.seqlen_offset>0`, only `input_ids` that do not have their past calculated should be passed as
|
| 1240 |
+
`input_ids`.
|
| 1241 |
+
|
| 1242 |
+
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
| 1243 |
+
[`PreTrainedTokenizer.__call__`] for details.
|
| 1244 |
+
|
| 1245 |
+
[What are input IDs?](../glossary#input-ids)
|
| 1246 |
+
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
| 1247 |
+
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
|
| 1248 |
+
is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
|
| 1249 |
+
model's internal embedding lookup matrix.
|
| 1250 |
+
position_ids (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
| 1251 |
+
Indices of positions of each input sequence tokens in the position embeddings.
|
| 1252 |
+
cache_params (`HybridMambaAttentionDynamicCache`, *optional*):
|
| 1253 |
+
If passed along, the model uses the previous state in all the blocks (which will give the output for the
|
| 1254 |
+
`input_ids` provided as if the model add `state_input_ids + input_ids` as context).
|
| 1255 |
+
use_cache (`bool`, *optional*):
|
| 1256 |
+
If set to `True`, the `cache_params` is returned and can be used to quickly generate the next logits.
|
| 1257 |
+
output_attentions (`bool`, *optional*):
|
| 1258 |
+
Whether or not to return the attentions tensors of all attention layers.
|
| 1259 |
+
output_hidden_states (`bool`, *optional*):
|
| 1260 |
+
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
|
| 1261 |
+
more detail.
|
| 1262 |
+
return_dict (`bool`, *optional*):
|
| 1263 |
+
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
|
| 1264 |
+
cache_position (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
| 1265 |
+
The position of the current input in the cache. This is used to ensure that the cache is correctly updated.
|
| 1266 |
+
If `cache_params` is passed, `cache_position` should also be passed.
|
| 1267 |
+
attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 1268 |
+
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
| 1269 |
+
|
| 1270 |
+
- 1 for tokens that are **not masked**,
|
| 1271 |
+
- 0 for tokens that are **masked**.
|
| 1272 |
+
|
| 1273 |
+
[What are attention masks?](../glossary#attention-mask)
|
| 1274 |
+
"""
|
| 1275 |
+
|
| 1276 |
+
|
| 1277 |
+
@add_start_docstrings(
|
| 1278 |
+
"The bare NemotronH Model transformer outputting raw hidden-states without any specific head on top.",
|
| 1279 |
+
NEMOTRONH_START_DOCSTRING,
|
| 1280 |
+
)
|
| 1281 |
+
class NemotronHModel(NemotronHPreTrainedModel):
|
| 1282 |
+
def __init__(self, config):
|
| 1283 |
+
super().__init__(config)
|
| 1284 |
+
|
| 1285 |
+
self.embeddings = nn.Embedding(config.vocab_size, config.hidden_size)
|
| 1286 |
+
self.layers = nn.ModuleList([NemotronHBlock(config, layer_idx=idx) for idx in range(config.num_hidden_layers)])
|
| 1287 |
+
|
| 1288 |
+
self.gradient_checkpointing = False
|
| 1289 |
+
self.norm_f = NemotronHRMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
|
| 1290 |
+
# Initialize weights and apply final processing
|
| 1291 |
+
self._register_load_state_dict_pre_hook(self.load_hook)
|
| 1292 |
+
self.post_init()
|
| 1293 |
+
|
| 1294 |
+
def load_hook(self, state_dict, prefix, *args):
|
| 1295 |
+
for k in state_dict:
|
| 1296 |
+
if "embedding." in k:
|
| 1297 |
+
state_dict[k.replace("embedding.", "embeddings.")] = state_dict.pop(k)
|
| 1298 |
+
break
|
| 1299 |
+
|
| 1300 |
+
def get_input_embeddings(self):
|
| 1301 |
+
return self.embeddings
|
| 1302 |
+
|
| 1303 |
+
def set_input_embeddings(self, new_embeddings):
|
| 1304 |
+
self.embeddings = new_embeddings
|
| 1305 |
+
|
| 1306 |
+
@add_start_docstrings_to_model_forward(NEMOTRONH_INPUTS_DOCSTRING)
|
| 1307 |
+
@add_code_sample_docstrings(
|
| 1308 |
+
checkpoint=_CHECKPOINT_FOR_DOC,
|
| 1309 |
+
output_type=NemotronHOutput,
|
| 1310 |
+
config_class=_CONFIG_FOR_DOC,
|
| 1311 |
+
)
|
| 1312 |
+
def forward(
|
| 1313 |
+
self,
|
| 1314 |
+
input_ids: Optional[torch.LongTensor] = None,
|
| 1315 |
+
inputs_embeds: Optional[torch.LongTensor] = None,
|
| 1316 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 1317 |
+
cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
|
| 1318 |
+
use_cache: Optional[bool] = None,
|
| 1319 |
+
output_attentions: Optional[bool] = None,
|
| 1320 |
+
output_hidden_states: Optional[bool] = None,
|
| 1321 |
+
return_dict: Optional[bool] = None,
|
| 1322 |
+
cache_position: Optional[torch.LongTensor] = None,
|
| 1323 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 1324 |
+
**kwargs,
|
| 1325 |
+
) -> Union[Tuple, NemotronHOutput]:
|
| 1326 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 1327 |
+
output_hidden_states = (
|
| 1328 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 1329 |
+
)
|
| 1330 |
+
# use_cache = use_cache if use_cache is not None else self.config.use_cache
|
| 1331 |
+
use_cache = use_cache if use_cache is not None else (self.config.use_cache if not self.training else False)
|
| 1332 |
+
|
| 1333 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 1334 |
+
|
| 1335 |
+
if (input_ids is None) ^ (inputs_embeds is not None): # ^ is python for xor
|
| 1336 |
+
raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
|
| 1337 |
+
|
| 1338 |
+
if inputs_embeds is None:
|
| 1339 |
+
inputs_embeds = self.embeddings(input_ids)
|
| 1340 |
+
|
| 1341 |
+
if self.gradient_checkpointing and self.training and use_cache:
|
| 1342 |
+
logger.warning_once(
|
| 1343 |
+
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
|
| 1344 |
+
)
|
| 1345 |
+
use_cache = False
|
| 1346 |
+
|
| 1347 |
+
# From zamba_modeling.py
|
| 1348 |
+
if use_cache and cache_params is None:
|
| 1349 |
+
logger.warning_once(
|
| 1350 |
+
"NemotronH requires an initialized `NemotronHHybridDynamicCache` to return a cache. None was "
|
| 1351 |
+
"provided, so no cache will be returned."
|
| 1352 |
+
)
|
| 1353 |
+
|
| 1354 |
+
hidden_states = inputs_embeds
|
| 1355 |
+
|
| 1356 |
+
if cache_position is None:
|
| 1357 |
+
cache_position = torch.arange(hidden_states.shape[1], device=hidden_states.device)
|
| 1358 |
+
if position_ids is None:
|
| 1359 |
+
position_ids = cache_position.unsqueeze(0)
|
| 1360 |
+
|
| 1361 |
+
causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position)
|
| 1362 |
+
mamba_mask = self._update_mamba_mask(attention_mask, cache_position)
|
| 1363 |
+
|
| 1364 |
+
all_hidden_states = () if output_hidden_states else None
|
| 1365 |
+
all_self_attns = () if output_attentions else None
|
| 1366 |
+
# Until HERE
|
| 1367 |
+
|
| 1368 |
+
for layer_idx, mixer_block in enumerate(self.layers):
|
| 1369 |
+
# Depending on the layer type we opt for 2D base attention mask (Mamba) or 4D causal mask (Attention)
|
| 1370 |
+
if mixer_block.block_type == "mamba":
|
| 1371 |
+
layer_mask = mamba_mask
|
| 1372 |
+
elif mixer_block.block_type == "attention":
|
| 1373 |
+
layer_mask = causal_mask
|
| 1374 |
+
elif mixer_block.block_type == "mlp":
|
| 1375 |
+
layer_mask = None
|
| 1376 |
+
else:
|
| 1377 |
+
raise ValueError(f"Invalid block_type: {self.block_type}")
|
| 1378 |
+
|
| 1379 |
+
if output_hidden_states:
|
| 1380 |
+
all_hidden_states += (hidden_states,)
|
| 1381 |
+
|
| 1382 |
+
if self.gradient_checkpointing and self.training:
|
| 1383 |
+
hidden_states = self._gradient_checkpointing_func(
|
| 1384 |
+
mixer_block.__call__, hidden_states, cache_params, cache_position, layer_mask
|
| 1385 |
+
)
|
| 1386 |
+
else:
|
| 1387 |
+
hidden_states = mixer_block(
|
| 1388 |
+
hidden_states,
|
| 1389 |
+
cache_params=cache_params,
|
| 1390 |
+
cache_position=cache_position,
|
| 1391 |
+
attention_mask=layer_mask,
|
| 1392 |
+
)
|
| 1393 |
+
|
| 1394 |
+
# TODO: Store attentions
|
| 1395 |
+
# if output_attentions:
|
| 1396 |
+
# if layer_outputs[1] is not None:
|
| 1397 |
+
# # append attentions only of attention layers. Mamba layers return `None` as the attention weights
|
| 1398 |
+
# all_self_attns += (layer_outputs[1],)
|
| 1399 |
+
|
| 1400 |
+
# TODO (Check): should it happen before the forward pass?
|
| 1401 |
+
# if output_hidden_states:
|
| 1402 |
+
# all_hidden_states = all_hidden_states + (hidden_states,)
|
| 1403 |
+
|
| 1404 |
+
hidden_states = self.norm_f(hidden_states)
|
| 1405 |
+
|
| 1406 |
+
if output_hidden_states:
|
| 1407 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
| 1408 |
+
|
| 1409 |
+
if not return_dict:
|
| 1410 |
+
return tuple(v for v in [hidden_states, cache_params, all_hidden_states] if v is not None)
|
| 1411 |
+
|
| 1412 |
+
return NemotronHOutput(
|
| 1413 |
+
last_hidden_state=hidden_states,
|
| 1414 |
+
cache_params=cache_params if use_cache else None,
|
| 1415 |
+
hidden_states=all_hidden_states,
|
| 1416 |
+
attentions=all_self_attns,
|
| 1417 |
+
)
|
| 1418 |
+
|
| 1419 |
+
# Copied from transformers.models.jamba.modeling_jamba.JambaModel._update_causal_mask
|
| 1420 |
+
def _update_causal_mask(self, attention_mask, input_tensor, cache_position):
|
| 1421 |
+
if self.config._attn_implementation == "flash_attention_2":
|
| 1422 |
+
if attention_mask is not None and 0.0 in attention_mask:
|
| 1423 |
+
return attention_mask
|
| 1424 |
+
return None
|
| 1425 |
+
|
| 1426 |
+
dtype, device = input_tensor.dtype, input_tensor.device
|
| 1427 |
+
min_dtype = torch.finfo(dtype).min
|
| 1428 |
+
sequence_length = input_tensor.shape[1]
|
| 1429 |
+
target_length = cache_position[-1] + 1
|
| 1430 |
+
|
| 1431 |
+
causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
|
| 1432 |
+
if sequence_length != 1:
|
| 1433 |
+
causal_mask = torch.triu(causal_mask, diagonal=1)
|
| 1434 |
+
causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
|
| 1435 |
+
causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
|
| 1436 |
+
if attention_mask is not None:
|
| 1437 |
+
causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit
|
| 1438 |
+
if attention_mask.dim() == 2:
|
| 1439 |
+
mask_length = attention_mask.shape[-1]
|
| 1440 |
+
padding_mask = causal_mask[..., :mask_length].eq(0.0) * attention_mask[:, None, None, :].eq(0.0)
|
| 1441 |
+
causal_mask[..., :mask_length] = causal_mask[..., :mask_length].masked_fill(padding_mask, min_dtype)
|
| 1442 |
+
|
| 1443 |
+
if (
|
| 1444 |
+
self.config._attn_implementation == "sdpa"
|
| 1445 |
+
and attention_mask is not None
|
| 1446 |
+
and attention_mask.device.type == "cuda"
|
| 1447 |
+
):
|
| 1448 |
+
# Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
|
| 1449 |
+
# using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
|
| 1450 |
+
# Details: https://github.com/pytorch/pytorch/issues/110213
|
| 1451 |
+
causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
|
| 1452 |
+
|
| 1453 |
+
return causal_mask
|
| 1454 |
+
|
| 1455 |
+
def _update_mamba_mask(self, attention_mask, cache_position):
|
| 1456 |
+
"""
|
| 1457 |
+
No need for zeroing states when
|
| 1458 |
+
1. Cached forward
|
| 1459 |
+
2. Attending to all inputs
|
| 1460 |
+
"""
|
| 1461 |
+
mamba_mask = attention_mask
|
| 1462 |
+
if cache_position[0] > 0 or (attention_mask is not None and torch.all(attention_mask == 1)):
|
| 1463 |
+
mamba_mask = None
|
| 1464 |
+
return mamba_mask
|
| 1465 |
+
|
| 1466 |
+
|
| 1467 |
+
@add_start_docstrings(
|
| 1468 |
+
"""
|
| 1469 |
+
The NEMOTRONH Model transformer with a language modeling head on top (linear layer with weights not tied to the input
|
| 1470 |
+
embeddings).
|
| 1471 |
+
""",
|
| 1472 |
+
NEMOTRONH_START_DOCSTRING,
|
| 1473 |
+
)
|
| 1474 |
+
class NemotronHForCausalLM(NemotronHPreTrainedModel, GenerationMixin):
|
| 1475 |
+
_tied_weights_keys = ["lm_head.weight"]
|
| 1476 |
+
|
| 1477 |
+
def __init__(self, config):
|
| 1478 |
+
super().__init__(config)
|
| 1479 |
+
self.backbone = NemotronHModel(config)
|
| 1480 |
+
self.vocab_size = config.vocab_size
|
| 1481 |
+
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
| 1482 |
+
|
| 1483 |
+
# Initialize weights and apply final processing
|
| 1484 |
+
self.post_init()
|
| 1485 |
+
|
| 1486 |
+
def get_input_embeddings(self):
|
| 1487 |
+
return self.backbone.get_input_embeddings()
|
| 1488 |
+
|
| 1489 |
+
def set_input_embeddings(self, new_embeddings):
|
| 1490 |
+
return self.backbone.set_input_embeddings(new_embeddings)
|
| 1491 |
+
|
| 1492 |
+
def get_output_embeddings(self):
|
| 1493 |
+
return self.lm_head
|
| 1494 |
+
|
| 1495 |
+
def set_output_embeddings(self, new_embeddings):
|
| 1496 |
+
self.lm_head = new_embeddings
|
| 1497 |
+
|
| 1498 |
+
def get_decoder(self):
|
| 1499 |
+
return self.model
|
| 1500 |
+
|
| 1501 |
+
def set_decoder(self, decoder):
|
| 1502 |
+
self.model = decoder
|
| 1503 |
+
|
| 1504 |
+
def prepare_inputs_for_generation(
|
| 1505 |
+
self,
|
| 1506 |
+
input_ids,
|
| 1507 |
+
past_key_values=None,
|
| 1508 |
+
attention_mask=None,
|
| 1509 |
+
inputs_embeds=None,
|
| 1510 |
+
cache_position=None,
|
| 1511 |
+
position_ids=None,
|
| 1512 |
+
use_cache=True,
|
| 1513 |
+
**kwargs,
|
| 1514 |
+
):
|
| 1515 |
+
# Copy from https://github.com/huggingface/transformers/blob/main/src/transformers/models/jamba/modeling_jamba.py
|
| 1516 |
+
# Overwitten -- uses `cache_params` as opposed to `past_key_values`
|
| 1517 |
+
empty_past_kv = past_key_values is None
|
| 1518 |
+
|
| 1519 |
+
# If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens
|
| 1520 |
+
# Exception 1: when passing input_embeds, input_ids may be missing entries
|
| 1521 |
+
# Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here
|
| 1522 |
+
# Exception 3: with synced GPUs cache_position may go out of bounds, but we only want dummy token in that case.
|
| 1523 |
+
# (we can't check exception 3 while compiling)
|
| 1524 |
+
if not empty_past_kv:
|
| 1525 |
+
if (
|
| 1526 |
+
inputs_embeds is not None # Exception 1
|
| 1527 |
+
or cache_position[-1] >= input_ids.shape[1] # Exception 3
|
| 1528 |
+
):
|
| 1529 |
+
input_ids = input_ids[:, -cache_position.shape[0] :]
|
| 1530 |
+
elif input_ids.shape[1] != cache_position.shape[0]: # Default case (the "else", a no op, is Exception 2)
|
| 1531 |
+
input_ids = input_ids[:, cache_position]
|
| 1532 |
+
else:
|
| 1533 |
+
past_key_values = HybridMambaAttentionDynamicCache(
|
| 1534 |
+
self.config, input_ids.shape[0], self.dtype, device=self.device
|
| 1535 |
+
)
|
| 1536 |
+
|
| 1537 |
+
if attention_mask is not None and position_ids is None:
|
| 1538 |
+
# create position_ids on the fly for batch generation
|
| 1539 |
+
position_ids = attention_mask.long().cumsum(-1) - 1
|
| 1540 |
+
position_ids.masked_fill_(attention_mask == 0, 1)
|
| 1541 |
+
if not empty_past_kv:
|
| 1542 |
+
position_ids = position_ids[:, -input_ids.shape[1] :]
|
| 1543 |
+
|
| 1544 |
+
# if `inputs_embeds` are passed, we only want to use them in the 1st generation step
|
| 1545 |
+
if inputs_embeds is not None and empty_past_kv:
|
| 1546 |
+
# TODO(pjin): workaround fix for properly extending inputs_embeds;
|
| 1547 |
+
# longer term, may be better handled elsewhere in .generate().
|
| 1548 |
+
if input_ids is not None and inputs_embeds.shape[1] < input_ids.shape[1]:
|
| 1549 |
+
new_token_embeds = self.get_input_embeddings()(input_ids[:,inputs_embeds.shape[1]:])
|
| 1550 |
+
inputs_embeds = torch.cat([inputs_embeds, new_token_embeds], dim=1)
|
| 1551 |
+
model_inputs = {"inputs_embeds": inputs_embeds}
|
| 1552 |
+
else:
|
| 1553 |
+
model_inputs = {"input_ids": input_ids.contiguous()} # `contiguous()` needed for compilation use cases
|
| 1554 |
+
|
| 1555 |
+
model_inputs.update(
|
| 1556 |
+
{
|
| 1557 |
+
"position_ids": position_ids,
|
| 1558 |
+
"past_key_values": past_key_values,
|
| 1559 |
+
"use_cache": use_cache,
|
| 1560 |
+
"attention_mask": attention_mask,
|
| 1561 |
+
"logits_to_keep": self.config.num_logits_to_keep,
|
| 1562 |
+
"cache_position": cache_position,
|
| 1563 |
+
}
|
| 1564 |
+
)
|
| 1565 |
+
return model_inputs
|
| 1566 |
+
|
| 1567 |
+
@add_start_docstrings_to_model_forward(NEMOTRONH_INPUTS_DOCSTRING)
|
| 1568 |
+
@add_code_sample_docstrings(
|
| 1569 |
+
checkpoint=_CHECKPOINT_FOR_DOC,
|
| 1570 |
+
output_type=NemotronHCausalLMOutput,
|
| 1571 |
+
config_class=_CONFIG_FOR_DOC,
|
| 1572 |
+
)
|
| 1573 |
+
def forward(
|
| 1574 |
+
self,
|
| 1575 |
+
input_ids: Optional[torch.LongTensor] = None,
|
| 1576 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 1577 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 1578 |
+
cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
|
| 1579 |
+
labels: Optional[torch.LongTensor] = None,
|
| 1580 |
+
output_attentions: Optional[bool] = None,
|
| 1581 |
+
output_hidden_states: Optional[bool] = None,
|
| 1582 |
+
return_dict: Optional[bool] = None,
|
| 1583 |
+
use_cache: Optional[bool] = None,
|
| 1584 |
+
cache_position: Optional[torch.Tensor] = None,
|
| 1585 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 1586 |
+
**kwargs, # for now we need this for generation
|
| 1587 |
+
) -> Union[Tuple, NemotronHCausalLMOutput]:
|
| 1588 |
+
r"""
|
| 1589 |
+
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 1590 |
+
Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
|
| 1591 |
+
`labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
|
| 1592 |
+
are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
|
| 1593 |
+
"""
|
| 1594 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 1595 |
+
|
| 1596 |
+
output_hidden_states = (
|
| 1597 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 1598 |
+
)
|
| 1599 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 1600 |
+
|
| 1601 |
+
nemotron_h_outputs = self.backbone(
|
| 1602 |
+
input_ids,
|
| 1603 |
+
cache_params=cache_params,
|
| 1604 |
+
inputs_embeds=inputs_embeds,
|
| 1605 |
+
output_attentions=output_attentions,
|
| 1606 |
+
output_hidden_states=output_hidden_states,
|
| 1607 |
+
return_dict=return_dict,
|
| 1608 |
+
use_cache=use_cache,
|
| 1609 |
+
cache_position=cache_position,
|
| 1610 |
+
attention_mask=attention_mask,
|
| 1611 |
+
)
|
| 1612 |
+
hidden_states = nemotron_h_outputs[0]
|
| 1613 |
+
|
| 1614 |
+
# TODO: Check zamba_modeling.py: https://github.com/huggingface/transformers/blob/d7188ba600e36d3fd191b12e19f1b3bb81a8404f/src/transformers/models/zamba/modeling_zamba.py#L1284C1-L1286C2
|
| 1615 |
+
#logits = self.lm_head(hidden_states.to(self.lm_head.weight.dtype)).float()
|
| 1616 |
+
logits = self.lm_head(hidden_states.to(self.lm_head.weight.dtype)).float()
|
| 1617 |
+
|
| 1618 |
+
loss = None
|
| 1619 |
+
if labels is not None:
|
| 1620 |
+
# move labels to correct device to enable model parallelism
|
| 1621 |
+
labels = labels.to(logits.device)
|
| 1622 |
+
# Shift so that tokens < n predict n
|
| 1623 |
+
shift_logits = logits[..., :-1, :].contiguous()
|
| 1624 |
+
shift_labels = labels[..., 1:].contiguous()
|
| 1625 |
+
# Flatten the tokens
|
| 1626 |
+
loss_fct = CrossEntropyLoss()
|
| 1627 |
+
loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
|
| 1628 |
+
|
| 1629 |
+
if not return_dict:
|
| 1630 |
+
output = (logits,) + nemotron_h_outputs[1:]
|
| 1631 |
+
return ((loss,) + output) if loss is not None else output
|
| 1632 |
+
|
| 1633 |
+
return NemotronHCausalLMOutput(
|
| 1634 |
+
loss=loss,
|
| 1635 |
+
logits=logits,
|
| 1636 |
+
cache_params=nemotron_h_outputs.cache_params,
|
| 1637 |
+
hidden_states=nemotron_h_outputs.hidden_states,
|
| 1638 |
+
attentions=nemotron_h_outputs.attentions,
|
| 1639 |
+
)
|
recipe.yaml
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
default_stage:
|
| 2 |
+
default_modifiers:
|
| 3 |
+
GPTQModifier:
|
| 4 |
+
targets: [Linear]
|
| 5 |
+
ignore: [lm_head]
|
| 6 |
+
scheme: W4A16
|
| 7 |
+
block_size: 128
|
| 8 |
+
dampening_frac: 0.01
|
| 9 |
+
actorder: static
|
| 10 |
+
offload_hessians: false
|
special_tokens_map.json
ADDED
|
@@ -0,0 +1,1026 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"additional_special_tokens": [
|
| 3 |
+
"<unk>",
|
| 4 |
+
"<s>",
|
| 5 |
+
"</s>",
|
| 6 |
+
"[INST]",
|
| 7 |
+
"[/INST]",
|
| 8 |
+
"[AVAILABLE_TOOLS]",
|
| 9 |
+
"[/AVAILABLE_TOOLS]",
|
| 10 |
+
"[TOOL_RESULTS]",
|
| 11 |
+
"[/TOOL_RESULTS]",
|
| 12 |
+
"[TOOL_CALLS]",
|
| 13 |
+
"[IMG]",
|
| 14 |
+
"<pad>",
|
| 15 |
+
"[IMG_BREAK]",
|
| 16 |
+
"[IMG_END]",
|
| 17 |
+
"[PREFIX]",
|
| 18 |
+
"[MIDDLE]",
|
| 19 |
+
"[SUFFIX]",
|
| 20 |
+
"[SYSTEM_PROMPT]",
|
| 21 |
+
"[/SYSTEM_PROMPT]",
|
| 22 |
+
"[TOOL_CONTENT]",
|
| 23 |
+
"<|im_start|>",
|
| 24 |
+
"<|im_end|>",
|
| 25 |
+
"<SPECIAL_28>",
|
| 26 |
+
"<SPECIAL_29>",
|
| 27 |
+
"<SPECIAL_30>",
|
| 28 |
+
"<SPECIAL_31>",
|
| 29 |
+
"<SPECIAL_32>",
|
| 30 |
+
"<SPECIAL_33>",
|
| 31 |
+
"<SPECIAL_34>",
|
| 32 |
+
"<SPECIAL_35>",
|
| 33 |
+
"<SPECIAL_36>",
|
| 34 |
+
"<SPECIAL_37>",
|
| 35 |
+
"<SPECIAL_38>",
|
| 36 |
+
"<SPECIAL_39>",
|
| 37 |
+
"<SPECIAL_40>",
|
| 38 |
+
"<SPECIAL_41>",
|
| 39 |
+
"<SPECIAL_42>",
|
| 40 |
+
"<SPECIAL_43>",
|
| 41 |
+
"<SPECIAL_44>",
|
| 42 |
+
"<SPECIAL_45>",
|
| 43 |
+
"<SPECIAL_46>",
|
| 44 |
+
"<SPECIAL_47>",
|
| 45 |
+
"<SPECIAL_48>",
|
| 46 |
+
"<SPECIAL_49>",
|
| 47 |
+
"<SPECIAL_50>",
|
| 48 |
+
"<SPECIAL_51>",
|
| 49 |
+
"<SPECIAL_52>",
|
| 50 |
+
"<SPECIAL_53>",
|
| 51 |
+
"<SPECIAL_54>",
|
| 52 |
+
"<SPECIAL_55>",
|
| 53 |
+
"<SPECIAL_56>",
|
| 54 |
+
"<SPECIAL_57>",
|
| 55 |
+
"<SPECIAL_58>",
|
| 56 |
+
"<SPECIAL_59>",
|
| 57 |
+
"<SPECIAL_60>",
|
| 58 |
+
"<SPECIAL_61>",
|
| 59 |
+
"<SPECIAL_62>",
|
| 60 |
+
"<SPECIAL_63>",
|
| 61 |
+
"<SPECIAL_64>",
|
| 62 |
+
"<SPECIAL_65>",
|
| 63 |
+
"<SPECIAL_66>",
|
| 64 |
+
"<SPECIAL_67>",
|
| 65 |
+
"<SPECIAL_68>",
|
| 66 |
+
"<SPECIAL_69>",
|
| 67 |
+
"<SPECIAL_70>",
|
| 68 |
+
"<SPECIAL_71>",
|
| 69 |
+
"<SPECIAL_72>",
|
| 70 |
+
"<SPECIAL_73>",
|
| 71 |
+
"<SPECIAL_74>",
|
| 72 |
+
"<SPECIAL_75>",
|
| 73 |
+
"<SPECIAL_76>",
|
| 74 |
+
"<SPECIAL_77>",
|
| 75 |
+
"<SPECIAL_78>",
|
| 76 |
+
"<SPECIAL_79>",
|
| 77 |
+
"<SPECIAL_80>",
|
| 78 |
+
"<SPECIAL_81>",
|
| 79 |
+
"<SPECIAL_82>",
|
| 80 |
+
"<SPECIAL_83>",
|
| 81 |
+
"<SPECIAL_84>",
|
| 82 |
+
"<SPECIAL_85>",
|
| 83 |
+
"<SPECIAL_86>",
|
| 84 |
+
"<SPECIAL_87>",
|
| 85 |
+
"<SPECIAL_88>",
|
| 86 |
+
"<SPECIAL_89>",
|
| 87 |
+
"<SPECIAL_90>",
|
| 88 |
+
"<SPECIAL_91>",
|
| 89 |
+
"<SPECIAL_92>",
|
| 90 |
+
"<SPECIAL_93>",
|
| 91 |
+
"<SPECIAL_94>",
|
| 92 |
+
"<SPECIAL_95>",
|
| 93 |
+
"<SPECIAL_96>",
|
| 94 |
+
"<SPECIAL_97>",
|
| 95 |
+
"<SPECIAL_98>",
|
| 96 |
+
"<SPECIAL_99>",
|
| 97 |
+
"<SPECIAL_100>",
|
| 98 |
+
"<SPECIAL_101>",
|
| 99 |
+
"<SPECIAL_102>",
|
| 100 |
+
"<SPECIAL_103>",
|
| 101 |
+
"<SPECIAL_104>",
|
| 102 |
+
"<SPECIAL_105>",
|
| 103 |
+
"<SPECIAL_106>",
|
| 104 |
+
"<SPECIAL_107>",
|
| 105 |
+
"<SPECIAL_108>",
|
| 106 |
+
"<SPECIAL_109>",
|
| 107 |
+
"<SPECIAL_110>",
|
| 108 |
+
"<SPECIAL_111>",
|
| 109 |
+
"<SPECIAL_112>",
|
| 110 |
+
"<SPECIAL_113>",
|
| 111 |
+
"<SPECIAL_114>",
|
| 112 |
+
"<SPECIAL_115>",
|
| 113 |
+
"<SPECIAL_116>",
|
| 114 |
+
"<SPECIAL_117>",
|
| 115 |
+
"<SPECIAL_118>",
|
| 116 |
+
"<SPECIAL_119>",
|
| 117 |
+
"<SPECIAL_120>",
|
| 118 |
+
"<SPECIAL_121>",
|
| 119 |
+
"<SPECIAL_122>",
|
| 120 |
+
"<SPECIAL_123>",
|
| 121 |
+
"<SPECIAL_124>",
|
| 122 |
+
"<SPECIAL_125>",
|
| 123 |
+
"<SPECIAL_126>",
|
| 124 |
+
"<SPECIAL_127>",
|
| 125 |
+
"<SPECIAL_128>",
|
| 126 |
+
"<SPECIAL_129>",
|
| 127 |
+
"<SPECIAL_130>",
|
| 128 |
+
"<SPECIAL_131>",
|
| 129 |
+
"<SPECIAL_132>",
|
| 130 |
+
"<SPECIAL_133>",
|
| 131 |
+
"<SPECIAL_134>",
|
| 132 |
+
"<SPECIAL_135>",
|
| 133 |
+
"<SPECIAL_136>",
|
| 134 |
+
"<SPECIAL_137>",
|
| 135 |
+
"<SPECIAL_138>",
|
| 136 |
+
"<SPECIAL_139>",
|
| 137 |
+
"<SPECIAL_140>",
|
| 138 |
+
"<SPECIAL_141>",
|
| 139 |
+
"<SPECIAL_142>",
|
| 140 |
+
"<SPECIAL_143>",
|
| 141 |
+
"<SPECIAL_144>",
|
| 142 |
+
"<SPECIAL_145>",
|
| 143 |
+
"<SPECIAL_146>",
|
| 144 |
+
"<SPECIAL_147>",
|
| 145 |
+
"<SPECIAL_148>",
|
| 146 |
+
"<SPECIAL_149>",
|
| 147 |
+
"<SPECIAL_150>",
|
| 148 |
+
"<SPECIAL_151>",
|
| 149 |
+
"<SPECIAL_152>",
|
| 150 |
+
"<SPECIAL_153>",
|
| 151 |
+
"<SPECIAL_154>",
|
| 152 |
+
"<SPECIAL_155>",
|
| 153 |
+
"<SPECIAL_156>",
|
| 154 |
+
"<SPECIAL_157>",
|
| 155 |
+
"<SPECIAL_158>",
|
| 156 |
+
"<SPECIAL_159>",
|
| 157 |
+
"<SPECIAL_160>",
|
| 158 |
+
"<SPECIAL_161>",
|
| 159 |
+
"<SPECIAL_162>",
|
| 160 |
+
"<SPECIAL_163>",
|
| 161 |
+
"<SPECIAL_164>",
|
| 162 |
+
"<SPECIAL_165>",
|
| 163 |
+
"<SPECIAL_166>",
|
| 164 |
+
"<SPECIAL_167>",
|
| 165 |
+
"<SPECIAL_168>",
|
| 166 |
+
"<SPECIAL_169>",
|
| 167 |
+
"<SPECIAL_170>",
|
| 168 |
+
"<SPECIAL_171>",
|
| 169 |
+
"<SPECIAL_172>",
|
| 170 |
+
"<SPECIAL_173>",
|
| 171 |
+
"<SPECIAL_174>",
|
| 172 |
+
"<SPECIAL_175>",
|
| 173 |
+
"<SPECIAL_176>",
|
| 174 |
+
"<SPECIAL_177>",
|
| 175 |
+
"<SPECIAL_178>",
|
| 176 |
+
"<SPECIAL_179>",
|
| 177 |
+
"<SPECIAL_180>",
|
| 178 |
+
"<SPECIAL_181>",
|
| 179 |
+
"<SPECIAL_182>",
|
| 180 |
+
"<SPECIAL_183>",
|
| 181 |
+
"<SPECIAL_184>",
|
| 182 |
+
"<SPECIAL_185>",
|
| 183 |
+
"<SPECIAL_186>",
|
| 184 |
+
"<SPECIAL_187>",
|
| 185 |
+
"<SPECIAL_188>",
|
| 186 |
+
"<SPECIAL_189>",
|
| 187 |
+
"<SPECIAL_190>",
|
| 188 |
+
"<SPECIAL_191>",
|
| 189 |
+
"<SPECIAL_192>",
|
| 190 |
+
"<SPECIAL_193>",
|
| 191 |
+
"<SPECIAL_194>",
|
| 192 |
+
"<SPECIAL_195>",
|
| 193 |
+
"<SPECIAL_196>",
|
| 194 |
+
"<SPECIAL_197>",
|
| 195 |
+
"<SPECIAL_198>",
|
| 196 |
+
"<SPECIAL_199>",
|
| 197 |
+
"<SPECIAL_200>",
|
| 198 |
+
"<SPECIAL_201>",
|
| 199 |
+
"<SPECIAL_202>",
|
| 200 |
+
"<SPECIAL_203>",
|
| 201 |
+
"<SPECIAL_204>",
|
| 202 |
+
"<SPECIAL_205>",
|
| 203 |
+
"<SPECIAL_206>",
|
| 204 |
+
"<SPECIAL_207>",
|
| 205 |
+
"<SPECIAL_208>",
|
| 206 |
+
"<SPECIAL_209>",
|
| 207 |
+
"<SPECIAL_210>",
|
| 208 |
+
"<SPECIAL_211>",
|
| 209 |
+
"<SPECIAL_212>",
|
| 210 |
+
"<SPECIAL_213>",
|
| 211 |
+
"<SPECIAL_214>",
|
| 212 |
+
"<SPECIAL_215>",
|
| 213 |
+
"<SPECIAL_216>",
|
| 214 |
+
"<SPECIAL_217>",
|
| 215 |
+
"<SPECIAL_218>",
|
| 216 |
+
"<SPECIAL_219>",
|
| 217 |
+
"<SPECIAL_220>",
|
| 218 |
+
"<SPECIAL_221>",
|
| 219 |
+
"<SPECIAL_222>",
|
| 220 |
+
"<SPECIAL_223>",
|
| 221 |
+
"<SPECIAL_224>",
|
| 222 |
+
"<SPECIAL_225>",
|
| 223 |
+
"<SPECIAL_226>",
|
| 224 |
+
"<SPECIAL_227>",
|
| 225 |
+
"<SPECIAL_228>",
|
| 226 |
+
"<SPECIAL_229>",
|
| 227 |
+
"<SPECIAL_230>",
|
| 228 |
+
"<SPECIAL_231>",
|
| 229 |
+
"<SPECIAL_232>",
|
| 230 |
+
"<SPECIAL_233>",
|
| 231 |
+
"<SPECIAL_234>",
|
| 232 |
+
"<SPECIAL_235>",
|
| 233 |
+
"<SPECIAL_236>",
|
| 234 |
+
"<SPECIAL_237>",
|
| 235 |
+
"<SPECIAL_238>",
|
| 236 |
+
"<SPECIAL_239>",
|
| 237 |
+
"<SPECIAL_240>",
|
| 238 |
+
"<SPECIAL_241>",
|
| 239 |
+
"<SPECIAL_242>",
|
| 240 |
+
"<SPECIAL_243>",
|
| 241 |
+
"<SPECIAL_244>",
|
| 242 |
+
"<SPECIAL_245>",
|
| 243 |
+
"<SPECIAL_246>",
|
| 244 |
+
"<SPECIAL_247>",
|
| 245 |
+
"<SPECIAL_248>",
|
| 246 |
+
"<SPECIAL_249>",
|
| 247 |
+
"<SPECIAL_250>",
|
| 248 |
+
"<SPECIAL_251>",
|
| 249 |
+
"<SPECIAL_252>",
|
| 250 |
+
"<SPECIAL_253>",
|
| 251 |
+
"<SPECIAL_254>",
|
| 252 |
+
"<SPECIAL_255>",
|
| 253 |
+
"<SPECIAL_256>",
|
| 254 |
+
"<SPECIAL_257>",
|
| 255 |
+
"<SPECIAL_258>",
|
| 256 |
+
"<SPECIAL_259>",
|
| 257 |
+
"<SPECIAL_260>",
|
| 258 |
+
"<SPECIAL_261>",
|
| 259 |
+
"<SPECIAL_262>",
|
| 260 |
+
"<SPECIAL_263>",
|
| 261 |
+
"<SPECIAL_264>",
|
| 262 |
+
"<SPECIAL_265>",
|
| 263 |
+
"<SPECIAL_266>",
|
| 264 |
+
"<SPECIAL_267>",
|
| 265 |
+
"<SPECIAL_268>",
|
| 266 |
+
"<SPECIAL_269>",
|
| 267 |
+
"<SPECIAL_270>",
|
| 268 |
+
"<SPECIAL_271>",
|
| 269 |
+
"<SPECIAL_272>",
|
| 270 |
+
"<SPECIAL_273>",
|
| 271 |
+
"<SPECIAL_274>",
|
| 272 |
+
"<SPECIAL_275>",
|
| 273 |
+
"<SPECIAL_276>",
|
| 274 |
+
"<SPECIAL_277>",
|
| 275 |
+
"<SPECIAL_278>",
|
| 276 |
+
"<SPECIAL_279>",
|
| 277 |
+
"<SPECIAL_280>",
|
| 278 |
+
"<SPECIAL_281>",
|
| 279 |
+
"<SPECIAL_282>",
|
| 280 |
+
"<SPECIAL_283>",
|
| 281 |
+
"<SPECIAL_284>",
|
| 282 |
+
"<SPECIAL_285>",
|
| 283 |
+
"<SPECIAL_286>",
|
| 284 |
+
"<SPECIAL_287>",
|
| 285 |
+
"<SPECIAL_288>",
|
| 286 |
+
"<SPECIAL_289>",
|
| 287 |
+
"<SPECIAL_290>",
|
| 288 |
+
"<SPECIAL_291>",
|
| 289 |
+
"<SPECIAL_292>",
|
| 290 |
+
"<SPECIAL_293>",
|
| 291 |
+
"<SPECIAL_294>",
|
| 292 |
+
"<SPECIAL_295>",
|
| 293 |
+
"<SPECIAL_296>",
|
| 294 |
+
"<SPECIAL_297>",
|
| 295 |
+
"<SPECIAL_298>",
|
| 296 |
+
"<SPECIAL_299>",
|
| 297 |
+
"<SPECIAL_300>",
|
| 298 |
+
"<SPECIAL_301>",
|
| 299 |
+
"<SPECIAL_302>",
|
| 300 |
+
"<SPECIAL_303>",
|
| 301 |
+
"<SPECIAL_304>",
|
| 302 |
+
"<SPECIAL_305>",
|
| 303 |
+
"<SPECIAL_306>",
|
| 304 |
+
"<SPECIAL_307>",
|
| 305 |
+
"<SPECIAL_308>",
|
| 306 |
+
"<SPECIAL_309>",
|
| 307 |
+
"<SPECIAL_310>",
|
| 308 |
+
"<SPECIAL_311>",
|
| 309 |
+
"<SPECIAL_312>",
|
| 310 |
+
"<SPECIAL_313>",
|
| 311 |
+
"<SPECIAL_314>",
|
| 312 |
+
"<SPECIAL_315>",
|
| 313 |
+
"<SPECIAL_316>",
|
| 314 |
+
"<SPECIAL_317>",
|
| 315 |
+
"<SPECIAL_318>",
|
| 316 |
+
"<SPECIAL_319>",
|
| 317 |
+
"<SPECIAL_320>",
|
| 318 |
+
"<SPECIAL_321>",
|
| 319 |
+
"<SPECIAL_322>",
|
| 320 |
+
"<SPECIAL_323>",
|
| 321 |
+
"<SPECIAL_324>",
|
| 322 |
+
"<SPECIAL_325>",
|
| 323 |
+
"<SPECIAL_326>",
|
| 324 |
+
"<SPECIAL_327>",
|
| 325 |
+
"<SPECIAL_328>",
|
| 326 |
+
"<SPECIAL_329>",
|
| 327 |
+
"<SPECIAL_330>",
|
| 328 |
+
"<SPECIAL_331>",
|
| 329 |
+
"<SPECIAL_332>",
|
| 330 |
+
"<SPECIAL_333>",
|
| 331 |
+
"<SPECIAL_334>",
|
| 332 |
+
"<SPECIAL_335>",
|
| 333 |
+
"<SPECIAL_336>",
|
| 334 |
+
"<SPECIAL_337>",
|
| 335 |
+
"<SPECIAL_338>",
|
| 336 |
+
"<SPECIAL_339>",
|
| 337 |
+
"<SPECIAL_340>",
|
| 338 |
+
"<SPECIAL_341>",
|
| 339 |
+
"<SPECIAL_342>",
|
| 340 |
+
"<SPECIAL_343>",
|
| 341 |
+
"<SPECIAL_344>",
|
| 342 |
+
"<SPECIAL_345>",
|
| 343 |
+
"<SPECIAL_346>",
|
| 344 |
+
"<SPECIAL_347>",
|
| 345 |
+
"<SPECIAL_348>",
|
| 346 |
+
"<SPECIAL_349>",
|
| 347 |
+
"<SPECIAL_350>",
|
| 348 |
+
"<SPECIAL_351>",
|
| 349 |
+
"<SPECIAL_352>",
|
| 350 |
+
"<SPECIAL_353>",
|
| 351 |
+
"<SPECIAL_354>",
|
| 352 |
+
"<SPECIAL_355>",
|
| 353 |
+
"<SPECIAL_356>",
|
| 354 |
+
"<SPECIAL_357>",
|
| 355 |
+
"<SPECIAL_358>",
|
| 356 |
+
"<SPECIAL_359>",
|
| 357 |
+
"<SPECIAL_360>",
|
| 358 |
+
"<SPECIAL_361>",
|
| 359 |
+
"<SPECIAL_362>",
|
| 360 |
+
"<SPECIAL_363>",
|
| 361 |
+
"<SPECIAL_364>",
|
| 362 |
+
"<SPECIAL_365>",
|
| 363 |
+
"<SPECIAL_366>",
|
| 364 |
+
"<SPECIAL_367>",
|
| 365 |
+
"<SPECIAL_368>",
|
| 366 |
+
"<SPECIAL_369>",
|
| 367 |
+
"<SPECIAL_370>",
|
| 368 |
+
"<SPECIAL_371>",
|
| 369 |
+
"<SPECIAL_372>",
|
| 370 |
+
"<SPECIAL_373>",
|
| 371 |
+
"<SPECIAL_374>",
|
| 372 |
+
"<SPECIAL_375>",
|
| 373 |
+
"<SPECIAL_376>",
|
| 374 |
+
"<SPECIAL_377>",
|
| 375 |
+
"<SPECIAL_378>",
|
| 376 |
+
"<SPECIAL_379>",
|
| 377 |
+
"<SPECIAL_380>",
|
| 378 |
+
"<SPECIAL_381>",
|
| 379 |
+
"<SPECIAL_382>",
|
| 380 |
+
"<SPECIAL_383>",
|
| 381 |
+
"<SPECIAL_384>",
|
| 382 |
+
"<SPECIAL_385>",
|
| 383 |
+
"<SPECIAL_386>",
|
| 384 |
+
"<SPECIAL_387>",
|
| 385 |
+
"<SPECIAL_388>",
|
| 386 |
+
"<SPECIAL_389>",
|
| 387 |
+
"<SPECIAL_390>",
|
| 388 |
+
"<SPECIAL_391>",
|
| 389 |
+
"<SPECIAL_392>",
|
| 390 |
+
"<SPECIAL_393>",
|
| 391 |
+
"<SPECIAL_394>",
|
| 392 |
+
"<SPECIAL_395>",
|
| 393 |
+
"<SPECIAL_396>",
|
| 394 |
+
"<SPECIAL_397>",
|
| 395 |
+
"<SPECIAL_398>",
|
| 396 |
+
"<SPECIAL_399>",
|
| 397 |
+
"<SPECIAL_400>",
|
| 398 |
+
"<SPECIAL_401>",
|
| 399 |
+
"<SPECIAL_402>",
|
| 400 |
+
"<SPECIAL_403>",
|
| 401 |
+
"<SPECIAL_404>",
|
| 402 |
+
"<SPECIAL_405>",
|
| 403 |
+
"<SPECIAL_406>",
|
| 404 |
+
"<SPECIAL_407>",
|
| 405 |
+
"<SPECIAL_408>",
|
| 406 |
+
"<SPECIAL_409>",
|
| 407 |
+
"<SPECIAL_410>",
|
| 408 |
+
"<SPECIAL_411>",
|
| 409 |
+
"<SPECIAL_412>",
|
| 410 |
+
"<SPECIAL_413>",
|
| 411 |
+
"<SPECIAL_414>",
|
| 412 |
+
"<SPECIAL_415>",
|
| 413 |
+
"<SPECIAL_416>",
|
| 414 |
+
"<SPECIAL_417>",
|
| 415 |
+
"<SPECIAL_418>",
|
| 416 |
+
"<SPECIAL_419>",
|
| 417 |
+
"<SPECIAL_420>",
|
| 418 |
+
"<SPECIAL_421>",
|
| 419 |
+
"<SPECIAL_422>",
|
| 420 |
+
"<SPECIAL_423>",
|
| 421 |
+
"<SPECIAL_424>",
|
| 422 |
+
"<SPECIAL_425>",
|
| 423 |
+
"<SPECIAL_426>",
|
| 424 |
+
"<SPECIAL_427>",
|
| 425 |
+
"<SPECIAL_428>",
|
| 426 |
+
"<SPECIAL_429>",
|
| 427 |
+
"<SPECIAL_430>",
|
| 428 |
+
"<SPECIAL_431>",
|
| 429 |
+
"<SPECIAL_432>",
|
| 430 |
+
"<SPECIAL_433>",
|
| 431 |
+
"<SPECIAL_434>",
|
| 432 |
+
"<SPECIAL_435>",
|
| 433 |
+
"<SPECIAL_436>",
|
| 434 |
+
"<SPECIAL_437>",
|
| 435 |
+
"<SPECIAL_438>",
|
| 436 |
+
"<SPECIAL_439>",
|
| 437 |
+
"<SPECIAL_440>",
|
| 438 |
+
"<SPECIAL_441>",
|
| 439 |
+
"<SPECIAL_442>",
|
| 440 |
+
"<SPECIAL_443>",
|
| 441 |
+
"<SPECIAL_444>",
|
| 442 |
+
"<SPECIAL_445>",
|
| 443 |
+
"<SPECIAL_446>",
|
| 444 |
+
"<SPECIAL_447>",
|
| 445 |
+
"<SPECIAL_448>",
|
| 446 |
+
"<SPECIAL_449>",
|
| 447 |
+
"<SPECIAL_450>",
|
| 448 |
+
"<SPECIAL_451>",
|
| 449 |
+
"<SPECIAL_452>",
|
| 450 |
+
"<SPECIAL_453>",
|
| 451 |
+
"<SPECIAL_454>",
|
| 452 |
+
"<SPECIAL_455>",
|
| 453 |
+
"<SPECIAL_456>",
|
| 454 |
+
"<SPECIAL_457>",
|
| 455 |
+
"<SPECIAL_458>",
|
| 456 |
+
"<SPECIAL_459>",
|
| 457 |
+
"<SPECIAL_460>",
|
| 458 |
+
"<SPECIAL_461>",
|
| 459 |
+
"<SPECIAL_462>",
|
| 460 |
+
"<SPECIAL_463>",
|
| 461 |
+
"<SPECIAL_464>",
|
| 462 |
+
"<SPECIAL_465>",
|
| 463 |
+
"<SPECIAL_466>",
|
| 464 |
+
"<SPECIAL_467>",
|
| 465 |
+
"<SPECIAL_468>",
|
| 466 |
+
"<SPECIAL_469>",
|
| 467 |
+
"<SPECIAL_470>",
|
| 468 |
+
"<SPECIAL_471>",
|
| 469 |
+
"<SPECIAL_472>",
|
| 470 |
+
"<SPECIAL_473>",
|
| 471 |
+
"<SPECIAL_474>",
|
| 472 |
+
"<SPECIAL_475>",
|
| 473 |
+
"<SPECIAL_476>",
|
| 474 |
+
"<SPECIAL_477>",
|
| 475 |
+
"<SPECIAL_478>",
|
| 476 |
+
"<SPECIAL_479>",
|
| 477 |
+
"<SPECIAL_480>",
|
| 478 |
+
"<SPECIAL_481>",
|
| 479 |
+
"<SPECIAL_482>",
|
| 480 |
+
"<SPECIAL_483>",
|
| 481 |
+
"<SPECIAL_484>",
|
| 482 |
+
"<SPECIAL_485>",
|
| 483 |
+
"<SPECIAL_486>",
|
| 484 |
+
"<SPECIAL_487>",
|
| 485 |
+
"<SPECIAL_488>",
|
| 486 |
+
"<SPECIAL_489>",
|
| 487 |
+
"<SPECIAL_490>",
|
| 488 |
+
"<SPECIAL_491>",
|
| 489 |
+
"<SPECIAL_492>",
|
| 490 |
+
"<SPECIAL_493>",
|
| 491 |
+
"<SPECIAL_494>",
|
| 492 |
+
"<SPECIAL_495>",
|
| 493 |
+
"<SPECIAL_496>",
|
| 494 |
+
"<SPECIAL_497>",
|
| 495 |
+
"<SPECIAL_498>",
|
| 496 |
+
"<SPECIAL_499>",
|
| 497 |
+
"<SPECIAL_500>",
|
| 498 |
+
"<SPECIAL_501>",
|
| 499 |
+
"<SPECIAL_502>",
|
| 500 |
+
"<SPECIAL_503>",
|
| 501 |
+
"<SPECIAL_504>",
|
| 502 |
+
"<SPECIAL_505>",
|
| 503 |
+
"<SPECIAL_506>",
|
| 504 |
+
"<SPECIAL_507>",
|
| 505 |
+
"<SPECIAL_508>",
|
| 506 |
+
"<SPECIAL_509>",
|
| 507 |
+
"<SPECIAL_510>",
|
| 508 |
+
"<SPECIAL_511>",
|
| 509 |
+
"<SPECIAL_512>",
|
| 510 |
+
"<SPECIAL_513>",
|
| 511 |
+
"<SPECIAL_514>",
|
| 512 |
+
"<SPECIAL_515>",
|
| 513 |
+
"<SPECIAL_516>",
|
| 514 |
+
"<SPECIAL_517>",
|
| 515 |
+
"<SPECIAL_518>",
|
| 516 |
+
"<SPECIAL_519>",
|
| 517 |
+
"<SPECIAL_520>",
|
| 518 |
+
"<SPECIAL_521>",
|
| 519 |
+
"<SPECIAL_522>",
|
| 520 |
+
"<SPECIAL_523>",
|
| 521 |
+
"<SPECIAL_524>",
|
| 522 |
+
"<SPECIAL_525>",
|
| 523 |
+
"<SPECIAL_526>",
|
| 524 |
+
"<SPECIAL_527>",
|
| 525 |
+
"<SPECIAL_528>",
|
| 526 |
+
"<SPECIAL_529>",
|
| 527 |
+
"<SPECIAL_530>",
|
| 528 |
+
"<SPECIAL_531>",
|
| 529 |
+
"<SPECIAL_532>",
|
| 530 |
+
"<SPECIAL_533>",
|
| 531 |
+
"<SPECIAL_534>",
|
| 532 |
+
"<SPECIAL_535>",
|
| 533 |
+
"<SPECIAL_536>",
|
| 534 |
+
"<SPECIAL_537>",
|
| 535 |
+
"<SPECIAL_538>",
|
| 536 |
+
"<SPECIAL_539>",
|
| 537 |
+
"<SPECIAL_540>",
|
| 538 |
+
"<SPECIAL_541>",
|
| 539 |
+
"<SPECIAL_542>",
|
| 540 |
+
"<SPECIAL_543>",
|
| 541 |
+
"<SPECIAL_544>",
|
| 542 |
+
"<SPECIAL_545>",
|
| 543 |
+
"<SPECIAL_546>",
|
| 544 |
+
"<SPECIAL_547>",
|
| 545 |
+
"<SPECIAL_548>",
|
| 546 |
+
"<SPECIAL_549>",
|
| 547 |
+
"<SPECIAL_550>",
|
| 548 |
+
"<SPECIAL_551>",
|
| 549 |
+
"<SPECIAL_552>",
|
| 550 |
+
"<SPECIAL_553>",
|
| 551 |
+
"<SPECIAL_554>",
|
| 552 |
+
"<SPECIAL_555>",
|
| 553 |
+
"<SPECIAL_556>",
|
| 554 |
+
"<SPECIAL_557>",
|
| 555 |
+
"<SPECIAL_558>",
|
| 556 |
+
"<SPECIAL_559>",
|
| 557 |
+
"<SPECIAL_560>",
|
| 558 |
+
"<SPECIAL_561>",
|
| 559 |
+
"<SPECIAL_562>",
|
| 560 |
+
"<SPECIAL_563>",
|
| 561 |
+
"<SPECIAL_564>",
|
| 562 |
+
"<SPECIAL_565>",
|
| 563 |
+
"<SPECIAL_566>",
|
| 564 |
+
"<SPECIAL_567>",
|
| 565 |
+
"<SPECIAL_568>",
|
| 566 |
+
"<SPECIAL_569>",
|
| 567 |
+
"<SPECIAL_570>",
|
| 568 |
+
"<SPECIAL_571>",
|
| 569 |
+
"<SPECIAL_572>",
|
| 570 |
+
"<SPECIAL_573>",
|
| 571 |
+
"<SPECIAL_574>",
|
| 572 |
+
"<SPECIAL_575>",
|
| 573 |
+
"<SPECIAL_576>",
|
| 574 |
+
"<SPECIAL_577>",
|
| 575 |
+
"<SPECIAL_578>",
|
| 576 |
+
"<SPECIAL_579>",
|
| 577 |
+
"<SPECIAL_580>",
|
| 578 |
+
"<SPECIAL_581>",
|
| 579 |
+
"<SPECIAL_582>",
|
| 580 |
+
"<SPECIAL_583>",
|
| 581 |
+
"<SPECIAL_584>",
|
| 582 |
+
"<SPECIAL_585>",
|
| 583 |
+
"<SPECIAL_586>",
|
| 584 |
+
"<SPECIAL_587>",
|
| 585 |
+
"<SPECIAL_588>",
|
| 586 |
+
"<SPECIAL_589>",
|
| 587 |
+
"<SPECIAL_590>",
|
| 588 |
+
"<SPECIAL_591>",
|
| 589 |
+
"<SPECIAL_592>",
|
| 590 |
+
"<SPECIAL_593>",
|
| 591 |
+
"<SPECIAL_594>",
|
| 592 |
+
"<SPECIAL_595>",
|
| 593 |
+
"<SPECIAL_596>",
|
| 594 |
+
"<SPECIAL_597>",
|
| 595 |
+
"<SPECIAL_598>",
|
| 596 |
+
"<SPECIAL_599>",
|
| 597 |
+
"<SPECIAL_600>",
|
| 598 |
+
"<SPECIAL_601>",
|
| 599 |
+
"<SPECIAL_602>",
|
| 600 |
+
"<SPECIAL_603>",
|
| 601 |
+
"<SPECIAL_604>",
|
| 602 |
+
"<SPECIAL_605>",
|
| 603 |
+
"<SPECIAL_606>",
|
| 604 |
+
"<SPECIAL_607>",
|
| 605 |
+
"<SPECIAL_608>",
|
| 606 |
+
"<SPECIAL_609>",
|
| 607 |
+
"<SPECIAL_610>",
|
| 608 |
+
"<SPECIAL_611>",
|
| 609 |
+
"<SPECIAL_612>",
|
| 610 |
+
"<SPECIAL_613>",
|
| 611 |
+
"<SPECIAL_614>",
|
| 612 |
+
"<SPECIAL_615>",
|
| 613 |
+
"<SPECIAL_616>",
|
| 614 |
+
"<SPECIAL_617>",
|
| 615 |
+
"<SPECIAL_618>",
|
| 616 |
+
"<SPECIAL_619>",
|
| 617 |
+
"<SPECIAL_620>",
|
| 618 |
+
"<SPECIAL_621>",
|
| 619 |
+
"<SPECIAL_622>",
|
| 620 |
+
"<SPECIAL_623>",
|
| 621 |
+
"<SPECIAL_624>",
|
| 622 |
+
"<SPECIAL_625>",
|
| 623 |
+
"<SPECIAL_626>",
|
| 624 |
+
"<SPECIAL_627>",
|
| 625 |
+
"<SPECIAL_628>",
|
| 626 |
+
"<SPECIAL_629>",
|
| 627 |
+
"<SPECIAL_630>",
|
| 628 |
+
"<SPECIAL_631>",
|
| 629 |
+
"<SPECIAL_632>",
|
| 630 |
+
"<SPECIAL_633>",
|
| 631 |
+
"<SPECIAL_634>",
|
| 632 |
+
"<SPECIAL_635>",
|
| 633 |
+
"<SPECIAL_636>",
|
| 634 |
+
"<SPECIAL_637>",
|
| 635 |
+
"<SPECIAL_638>",
|
| 636 |
+
"<SPECIAL_639>",
|
| 637 |
+
"<SPECIAL_640>",
|
| 638 |
+
"<SPECIAL_641>",
|
| 639 |
+
"<SPECIAL_642>",
|
| 640 |
+
"<SPECIAL_643>",
|
| 641 |
+
"<SPECIAL_644>",
|
| 642 |
+
"<SPECIAL_645>",
|
| 643 |
+
"<SPECIAL_646>",
|
| 644 |
+
"<SPECIAL_647>",
|
| 645 |
+
"<SPECIAL_648>",
|
| 646 |
+
"<SPECIAL_649>",
|
| 647 |
+
"<SPECIAL_650>",
|
| 648 |
+
"<SPECIAL_651>",
|
| 649 |
+
"<SPECIAL_652>",
|
| 650 |
+
"<SPECIAL_653>",
|
| 651 |
+
"<SPECIAL_654>",
|
| 652 |
+
"<SPECIAL_655>",
|
| 653 |
+
"<SPECIAL_656>",
|
| 654 |
+
"<SPECIAL_657>",
|
| 655 |
+
"<SPECIAL_658>",
|
| 656 |
+
"<SPECIAL_659>",
|
| 657 |
+
"<SPECIAL_660>",
|
| 658 |
+
"<SPECIAL_661>",
|
| 659 |
+
"<SPECIAL_662>",
|
| 660 |
+
"<SPECIAL_663>",
|
| 661 |
+
"<SPECIAL_664>",
|
| 662 |
+
"<SPECIAL_665>",
|
| 663 |
+
"<SPECIAL_666>",
|
| 664 |
+
"<SPECIAL_667>",
|
| 665 |
+
"<SPECIAL_668>",
|
| 666 |
+
"<SPECIAL_669>",
|
| 667 |
+
"<SPECIAL_670>",
|
| 668 |
+
"<SPECIAL_671>",
|
| 669 |
+
"<SPECIAL_672>",
|
| 670 |
+
"<SPECIAL_673>",
|
| 671 |
+
"<SPECIAL_674>",
|
| 672 |
+
"<SPECIAL_675>",
|
| 673 |
+
"<SPECIAL_676>",
|
| 674 |
+
"<SPECIAL_677>",
|
| 675 |
+
"<SPECIAL_678>",
|
| 676 |
+
"<SPECIAL_679>",
|
| 677 |
+
"<SPECIAL_680>",
|
| 678 |
+
"<SPECIAL_681>",
|
| 679 |
+
"<SPECIAL_682>",
|
| 680 |
+
"<SPECIAL_683>",
|
| 681 |
+
"<SPECIAL_684>",
|
| 682 |
+
"<SPECIAL_685>",
|
| 683 |
+
"<SPECIAL_686>",
|
| 684 |
+
"<SPECIAL_687>",
|
| 685 |
+
"<SPECIAL_688>",
|
| 686 |
+
"<SPECIAL_689>",
|
| 687 |
+
"<SPECIAL_690>",
|
| 688 |
+
"<SPECIAL_691>",
|
| 689 |
+
"<SPECIAL_692>",
|
| 690 |
+
"<SPECIAL_693>",
|
| 691 |
+
"<SPECIAL_694>",
|
| 692 |
+
"<SPECIAL_695>",
|
| 693 |
+
"<SPECIAL_696>",
|
| 694 |
+
"<SPECIAL_697>",
|
| 695 |
+
"<SPECIAL_698>",
|
| 696 |
+
"<SPECIAL_699>",
|
| 697 |
+
"<SPECIAL_700>",
|
| 698 |
+
"<SPECIAL_701>",
|
| 699 |
+
"<SPECIAL_702>",
|
| 700 |
+
"<SPECIAL_703>",
|
| 701 |
+
"<SPECIAL_704>",
|
| 702 |
+
"<SPECIAL_705>",
|
| 703 |
+
"<SPECIAL_706>",
|
| 704 |
+
"<SPECIAL_707>",
|
| 705 |
+
"<SPECIAL_708>",
|
| 706 |
+
"<SPECIAL_709>",
|
| 707 |
+
"<SPECIAL_710>",
|
| 708 |
+
"<SPECIAL_711>",
|
| 709 |
+
"<SPECIAL_712>",
|
| 710 |
+
"<SPECIAL_713>",
|
| 711 |
+
"<SPECIAL_714>",
|
| 712 |
+
"<SPECIAL_715>",
|
| 713 |
+
"<SPECIAL_716>",
|
| 714 |
+
"<SPECIAL_717>",
|
| 715 |
+
"<SPECIAL_718>",
|
| 716 |
+
"<SPECIAL_719>",
|
| 717 |
+
"<SPECIAL_720>",
|
| 718 |
+
"<SPECIAL_721>",
|
| 719 |
+
"<SPECIAL_722>",
|
| 720 |
+
"<SPECIAL_723>",
|
| 721 |
+
"<SPECIAL_724>",
|
| 722 |
+
"<SPECIAL_725>",
|
| 723 |
+
"<SPECIAL_726>",
|
| 724 |
+
"<SPECIAL_727>",
|
| 725 |
+
"<SPECIAL_728>",
|
| 726 |
+
"<SPECIAL_729>",
|
| 727 |
+
"<SPECIAL_730>",
|
| 728 |
+
"<SPECIAL_731>",
|
| 729 |
+
"<SPECIAL_732>",
|
| 730 |
+
"<SPECIAL_733>",
|
| 731 |
+
"<SPECIAL_734>",
|
| 732 |
+
"<SPECIAL_735>",
|
| 733 |
+
"<SPECIAL_736>",
|
| 734 |
+
"<SPECIAL_737>",
|
| 735 |
+
"<SPECIAL_738>",
|
| 736 |
+
"<SPECIAL_739>",
|
| 737 |
+
"<SPECIAL_740>",
|
| 738 |
+
"<SPECIAL_741>",
|
| 739 |
+
"<SPECIAL_742>",
|
| 740 |
+
"<SPECIAL_743>",
|
| 741 |
+
"<SPECIAL_744>",
|
| 742 |
+
"<SPECIAL_745>",
|
| 743 |
+
"<SPECIAL_746>",
|
| 744 |
+
"<SPECIAL_747>",
|
| 745 |
+
"<SPECIAL_748>",
|
| 746 |
+
"<SPECIAL_749>",
|
| 747 |
+
"<SPECIAL_750>",
|
| 748 |
+
"<SPECIAL_751>",
|
| 749 |
+
"<SPECIAL_752>",
|
| 750 |
+
"<SPECIAL_753>",
|
| 751 |
+
"<SPECIAL_754>",
|
| 752 |
+
"<SPECIAL_755>",
|
| 753 |
+
"<SPECIAL_756>",
|
| 754 |
+
"<SPECIAL_757>",
|
| 755 |
+
"<SPECIAL_758>",
|
| 756 |
+
"<SPECIAL_759>",
|
| 757 |
+
"<SPECIAL_760>",
|
| 758 |
+
"<SPECIAL_761>",
|
| 759 |
+
"<SPECIAL_762>",
|
| 760 |
+
"<SPECIAL_763>",
|
| 761 |
+
"<SPECIAL_764>",
|
| 762 |
+
"<SPECIAL_765>",
|
| 763 |
+
"<SPECIAL_766>",
|
| 764 |
+
"<SPECIAL_767>",
|
| 765 |
+
"<SPECIAL_768>",
|
| 766 |
+
"<SPECIAL_769>",
|
| 767 |
+
"<SPECIAL_770>",
|
| 768 |
+
"<SPECIAL_771>",
|
| 769 |
+
"<SPECIAL_772>",
|
| 770 |
+
"<SPECIAL_773>",
|
| 771 |
+
"<SPECIAL_774>",
|
| 772 |
+
"<SPECIAL_775>",
|
| 773 |
+
"<SPECIAL_776>",
|
| 774 |
+
"<SPECIAL_777>",
|
| 775 |
+
"<SPECIAL_778>",
|
| 776 |
+
"<SPECIAL_779>",
|
| 777 |
+
"<SPECIAL_780>",
|
| 778 |
+
"<SPECIAL_781>",
|
| 779 |
+
"<SPECIAL_782>",
|
| 780 |
+
"<SPECIAL_783>",
|
| 781 |
+
"<SPECIAL_784>",
|
| 782 |
+
"<SPECIAL_785>",
|
| 783 |
+
"<SPECIAL_786>",
|
| 784 |
+
"<SPECIAL_787>",
|
| 785 |
+
"<SPECIAL_788>",
|
| 786 |
+
"<SPECIAL_789>",
|
| 787 |
+
"<SPECIAL_790>",
|
| 788 |
+
"<SPECIAL_791>",
|
| 789 |
+
"<SPECIAL_792>",
|
| 790 |
+
"<SPECIAL_793>",
|
| 791 |
+
"<SPECIAL_794>",
|
| 792 |
+
"<SPECIAL_795>",
|
| 793 |
+
"<SPECIAL_796>",
|
| 794 |
+
"<SPECIAL_797>",
|
| 795 |
+
"<SPECIAL_798>",
|
| 796 |
+
"<SPECIAL_799>",
|
| 797 |
+
"<SPECIAL_800>",
|
| 798 |
+
"<SPECIAL_801>",
|
| 799 |
+
"<SPECIAL_802>",
|
| 800 |
+
"<SPECIAL_803>",
|
| 801 |
+
"<SPECIAL_804>",
|
| 802 |
+
"<SPECIAL_805>",
|
| 803 |
+
"<SPECIAL_806>",
|
| 804 |
+
"<SPECIAL_807>",
|
| 805 |
+
"<SPECIAL_808>",
|
| 806 |
+
"<SPECIAL_809>",
|
| 807 |
+
"<SPECIAL_810>",
|
| 808 |
+
"<SPECIAL_811>",
|
| 809 |
+
"<SPECIAL_812>",
|
| 810 |
+
"<SPECIAL_813>",
|
| 811 |
+
"<SPECIAL_814>",
|
| 812 |
+
"<SPECIAL_815>",
|
| 813 |
+
"<SPECIAL_816>",
|
| 814 |
+
"<SPECIAL_817>",
|
| 815 |
+
"<SPECIAL_818>",
|
| 816 |
+
"<SPECIAL_819>",
|
| 817 |
+
"<SPECIAL_820>",
|
| 818 |
+
"<SPECIAL_821>",
|
| 819 |
+
"<SPECIAL_822>",
|
| 820 |
+
"<SPECIAL_823>",
|
| 821 |
+
"<SPECIAL_824>",
|
| 822 |
+
"<SPECIAL_825>",
|
| 823 |
+
"<SPECIAL_826>",
|
| 824 |
+
"<SPECIAL_827>",
|
| 825 |
+
"<SPECIAL_828>",
|
| 826 |
+
"<SPECIAL_829>",
|
| 827 |
+
"<SPECIAL_830>",
|
| 828 |
+
"<SPECIAL_831>",
|
| 829 |
+
"<SPECIAL_832>",
|
| 830 |
+
"<SPECIAL_833>",
|
| 831 |
+
"<SPECIAL_834>",
|
| 832 |
+
"<SPECIAL_835>",
|
| 833 |
+
"<SPECIAL_836>",
|
| 834 |
+
"<SPECIAL_837>",
|
| 835 |
+
"<SPECIAL_838>",
|
| 836 |
+
"<SPECIAL_839>",
|
| 837 |
+
"<SPECIAL_840>",
|
| 838 |
+
"<SPECIAL_841>",
|
| 839 |
+
"<SPECIAL_842>",
|
| 840 |
+
"<SPECIAL_843>",
|
| 841 |
+
"<SPECIAL_844>",
|
| 842 |
+
"<SPECIAL_845>",
|
| 843 |
+
"<SPECIAL_846>",
|
| 844 |
+
"<SPECIAL_847>",
|
| 845 |
+
"<SPECIAL_848>",
|
| 846 |
+
"<SPECIAL_849>",
|
| 847 |
+
"<SPECIAL_850>",
|
| 848 |
+
"<SPECIAL_851>",
|
| 849 |
+
"<SPECIAL_852>",
|
| 850 |
+
"<SPECIAL_853>",
|
| 851 |
+
"<SPECIAL_854>",
|
| 852 |
+
"<SPECIAL_855>",
|
| 853 |
+
"<SPECIAL_856>",
|
| 854 |
+
"<SPECIAL_857>",
|
| 855 |
+
"<SPECIAL_858>",
|
| 856 |
+
"<SPECIAL_859>",
|
| 857 |
+
"<SPECIAL_860>",
|
| 858 |
+
"<SPECIAL_861>",
|
| 859 |
+
"<SPECIAL_862>",
|
| 860 |
+
"<SPECIAL_863>",
|
| 861 |
+
"<SPECIAL_864>",
|
| 862 |
+
"<SPECIAL_865>",
|
| 863 |
+
"<SPECIAL_866>",
|
| 864 |
+
"<SPECIAL_867>",
|
| 865 |
+
"<SPECIAL_868>",
|
| 866 |
+
"<SPECIAL_869>",
|
| 867 |
+
"<SPECIAL_870>",
|
| 868 |
+
"<SPECIAL_871>",
|
| 869 |
+
"<SPECIAL_872>",
|
| 870 |
+
"<SPECIAL_873>",
|
| 871 |
+
"<SPECIAL_874>",
|
| 872 |
+
"<SPECIAL_875>",
|
| 873 |
+
"<SPECIAL_876>",
|
| 874 |
+
"<SPECIAL_877>",
|
| 875 |
+
"<SPECIAL_878>",
|
| 876 |
+
"<SPECIAL_879>",
|
| 877 |
+
"<SPECIAL_880>",
|
| 878 |
+
"<SPECIAL_881>",
|
| 879 |
+
"<SPECIAL_882>",
|
| 880 |
+
"<SPECIAL_883>",
|
| 881 |
+
"<SPECIAL_884>",
|
| 882 |
+
"<SPECIAL_885>",
|
| 883 |
+
"<SPECIAL_886>",
|
| 884 |
+
"<SPECIAL_887>",
|
| 885 |
+
"<SPECIAL_888>",
|
| 886 |
+
"<SPECIAL_889>",
|
| 887 |
+
"<SPECIAL_890>",
|
| 888 |
+
"<SPECIAL_891>",
|
| 889 |
+
"<SPECIAL_892>",
|
| 890 |
+
"<SPECIAL_893>",
|
| 891 |
+
"<SPECIAL_894>",
|
| 892 |
+
"<SPECIAL_895>",
|
| 893 |
+
"<SPECIAL_896>",
|
| 894 |
+
"<SPECIAL_897>",
|
| 895 |
+
"<SPECIAL_898>",
|
| 896 |
+
"<SPECIAL_899>",
|
| 897 |
+
"<SPECIAL_900>",
|
| 898 |
+
"<SPECIAL_901>",
|
| 899 |
+
"<SPECIAL_902>",
|
| 900 |
+
"<SPECIAL_903>",
|
| 901 |
+
"<SPECIAL_904>",
|
| 902 |
+
"<SPECIAL_905>",
|
| 903 |
+
"<SPECIAL_906>",
|
| 904 |
+
"<SPECIAL_907>",
|
| 905 |
+
"<SPECIAL_908>",
|
| 906 |
+
"<SPECIAL_909>",
|
| 907 |
+
"<SPECIAL_910>",
|
| 908 |
+
"<SPECIAL_911>",
|
| 909 |
+
"<SPECIAL_912>",
|
| 910 |
+
"<SPECIAL_913>",
|
| 911 |
+
"<SPECIAL_914>",
|
| 912 |
+
"<SPECIAL_915>",
|
| 913 |
+
"<SPECIAL_916>",
|
| 914 |
+
"<SPECIAL_917>",
|
| 915 |
+
"<SPECIAL_918>",
|
| 916 |
+
"<SPECIAL_919>",
|
| 917 |
+
"<SPECIAL_920>",
|
| 918 |
+
"<SPECIAL_921>",
|
| 919 |
+
"<SPECIAL_922>",
|
| 920 |
+
"<SPECIAL_923>",
|
| 921 |
+
"<SPECIAL_924>",
|
| 922 |
+
"<SPECIAL_925>",
|
| 923 |
+
"<SPECIAL_926>",
|
| 924 |
+
"<SPECIAL_927>",
|
| 925 |
+
"<SPECIAL_928>",
|
| 926 |
+
"<SPECIAL_929>",
|
| 927 |
+
"<SPECIAL_930>",
|
| 928 |
+
"<SPECIAL_931>",
|
| 929 |
+
"<SPECIAL_932>",
|
| 930 |
+
"<SPECIAL_933>",
|
| 931 |
+
"<SPECIAL_934>",
|
| 932 |
+
"<SPECIAL_935>",
|
| 933 |
+
"<SPECIAL_936>",
|
| 934 |
+
"<SPECIAL_937>",
|
| 935 |
+
"<SPECIAL_938>",
|
| 936 |
+
"<SPECIAL_939>",
|
| 937 |
+
"<SPECIAL_940>",
|
| 938 |
+
"<SPECIAL_941>",
|
| 939 |
+
"<SPECIAL_942>",
|
| 940 |
+
"<SPECIAL_943>",
|
| 941 |
+
"<SPECIAL_944>",
|
| 942 |
+
"<SPECIAL_945>",
|
| 943 |
+
"<SPECIAL_946>",
|
| 944 |
+
"<SPECIAL_947>",
|
| 945 |
+
"<SPECIAL_948>",
|
| 946 |
+
"<SPECIAL_949>",
|
| 947 |
+
"<SPECIAL_950>",
|
| 948 |
+
"<SPECIAL_951>",
|
| 949 |
+
"<SPECIAL_952>",
|
| 950 |
+
"<SPECIAL_953>",
|
| 951 |
+
"<SPECIAL_954>",
|
| 952 |
+
"<SPECIAL_955>",
|
| 953 |
+
"<SPECIAL_956>",
|
| 954 |
+
"<SPECIAL_957>",
|
| 955 |
+
"<SPECIAL_958>",
|
| 956 |
+
"<SPECIAL_959>",
|
| 957 |
+
"<SPECIAL_960>",
|
| 958 |
+
"<SPECIAL_961>",
|
| 959 |
+
"<SPECIAL_962>",
|
| 960 |
+
"<SPECIAL_963>",
|
| 961 |
+
"<SPECIAL_964>",
|
| 962 |
+
"<SPECIAL_965>",
|
| 963 |
+
"<SPECIAL_966>",
|
| 964 |
+
"<SPECIAL_967>",
|
| 965 |
+
"<SPECIAL_968>",
|
| 966 |
+
"<SPECIAL_969>",
|
| 967 |
+
"<SPECIAL_970>",
|
| 968 |
+
"<SPECIAL_971>",
|
| 969 |
+
"<SPECIAL_972>",
|
| 970 |
+
"<SPECIAL_973>",
|
| 971 |
+
"<SPECIAL_974>",
|
| 972 |
+
"<SPECIAL_975>",
|
| 973 |
+
"<SPECIAL_976>",
|
| 974 |
+
"<SPECIAL_977>",
|
| 975 |
+
"<SPECIAL_978>",
|
| 976 |
+
"<SPECIAL_979>",
|
| 977 |
+
"<SPECIAL_980>",
|
| 978 |
+
"<SPECIAL_981>",
|
| 979 |
+
"<SPECIAL_982>",
|
| 980 |
+
"<SPECIAL_983>",
|
| 981 |
+
"<SPECIAL_984>",
|
| 982 |
+
"<SPECIAL_985>",
|
| 983 |
+
"<SPECIAL_986>",
|
| 984 |
+
"<SPECIAL_987>",
|
| 985 |
+
"<SPECIAL_988>",
|
| 986 |
+
"<SPECIAL_989>",
|
| 987 |
+
"<SPECIAL_990>",
|
| 988 |
+
"<SPECIAL_991>",
|
| 989 |
+
"<SPECIAL_992>",
|
| 990 |
+
"<SPECIAL_993>",
|
| 991 |
+
"<SPECIAL_994>",
|
| 992 |
+
"<SPECIAL_995>",
|
| 993 |
+
"<SPECIAL_996>",
|
| 994 |
+
"<SPECIAL_997>",
|
| 995 |
+
"<SPECIAL_998>",
|
| 996 |
+
"<SPECIAL_999>"
|
| 997 |
+
],
|
| 998 |
+
"bos_token": {
|
| 999 |
+
"content": "<s>",
|
| 1000 |
+
"lstrip": false,
|
| 1001 |
+
"normalized": false,
|
| 1002 |
+
"rstrip": false,
|
| 1003 |
+
"single_word": false
|
| 1004 |
+
},
|
| 1005 |
+
"eos_token": {
|
| 1006 |
+
"content": "</s>",
|
| 1007 |
+
"lstrip": false,
|
| 1008 |
+
"normalized": false,
|
| 1009 |
+
"rstrip": false,
|
| 1010 |
+
"single_word": false
|
| 1011 |
+
},
|
| 1012 |
+
"pad_token": {
|
| 1013 |
+
"content": "<pad>",
|
| 1014 |
+
"lstrip": false,
|
| 1015 |
+
"normalized": false,
|
| 1016 |
+
"rstrip": false,
|
| 1017 |
+
"single_word": false
|
| 1018 |
+
},
|
| 1019 |
+
"unk_token": {
|
| 1020 |
+
"content": "<unk>",
|
| 1021 |
+
"lstrip": false,
|
| 1022 |
+
"normalized": false,
|
| 1023 |
+
"rstrip": false,
|
| 1024 |
+
"single_word": false
|
| 1025 |
+
}
|
| 1026 |
+
}
|
tokenizer.json
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d54d32258647e5a4e66e7a211bd9a8794674666b4ebb049ccf851161a9e489cd
|
| 3 |
+
size 17078033
|
tokenizer_config.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|