Move to in-library checkpoint (#43)
Browse files- Move to in-library checkpoint (886484abd64073e42c2145031d4b1108411b5538)
- Move to in-library checkpoint (e00e26a40027807e70ef1e0d10ad85dcc2ba5bbe)
- Move to in-library checkpoint (5da3542292b50719512b7c22f6f9f8cff69c8bb7)
- Move to in-library checkpoint (5a0154d86a53a3ea6c4f115ab66be55ef5d344e8)
- README.md +6 -9
- config.json +3 -4
- configuration_jamba.py +27 -17
- generation_config.json +1 -1
- model-00001-of-00021.safetensors +2 -2
- model-00002-of-00021.safetensors +2 -2
- model-00003-of-00021.safetensors +2 -2
- model-00004-of-00021.safetensors +2 -2
- model-00005-of-00021.safetensors +2 -2
- model-00006-of-00021.safetensors +2 -2
- model-00007-of-00021.safetensors +2 -2
- model-00008-of-00021.safetensors +2 -2
- model-00009-of-00021.safetensors +2 -2
- model-00010-of-00021.safetensors +2 -2
- model-00011-of-00021.safetensors +2 -2
- model-00012-of-00021.safetensors +2 -2
- model-00013-of-00021.safetensors +2 -2
- model-00014-of-00021.safetensors +2 -2
- model-00015-of-00021.safetensors +2 -2
- model-00016-of-00021.safetensors +2 -2
- model-00017-of-00021.safetensors +2 -2
- model-00018-of-00021.safetensors +2 -2
- model-00019-of-00021.safetensors +2 -2
- model-00020-of-00021.safetensors +2 -2
- model-00021-of-00021.safetensors +2 -2
- model.safetensors.index.json +0 -0
- modeling_jamba.py +397 -632
- special_tokens_map.json +28 -4
README.md
CHANGED
|
@@ -27,9 +27,9 @@ For full details of this model please read the [white paper](https://arxiv.org/a
|
|
| 27 |
|
| 28 |
## Usage
|
| 29 |
### Presequities
|
| 30 |
-
Jamba
|
| 31 |
```bash
|
| 32 |
-
pip install transformers>=4.
|
| 33 |
```
|
| 34 |
|
| 35 |
In order to run optimized Mamba implementations, you first need to install `mamba-ssm` and `causal-conv1d`:
|
|
@@ -41,12 +41,10 @@ You also have to have the model on a CUDA device.
|
|
| 41 |
You can run the model not using the optimized Mamba kernels, but it is **not** recommended as it will result in significantly lower latencies. In order to do that, you'll need to specify `use_mamba_kernels=False` when loading the model.
|
| 42 |
|
| 43 |
### Run the model
|
| 44 |
-
Please note that, at the moment, `trust_remote_code=True` is required for running the new Jamba architecture.
|
| 45 |
```python
|
| 46 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 47 |
|
| 48 |
-
model = AutoModelForCausalLM.from_pretrained("ai21labs/Jamba-v0.1"
|
| 49 |
-
trust_remote_code=True)
|
| 50 |
tokenizer = AutoTokenizer.from_pretrained("ai21labs/Jamba-v0.1")
|
| 51 |
|
| 52 |
input_ids = tokenizer("In the recent Super Bowl LVIII,", return_tensors='pt').to(model.device)["input_ids"]
|
|
@@ -57,6 +55,8 @@ print(tokenizer.batch_decode(outputs))
|
|
| 57 |
# ["<|startoftext|>In the recent Super Bowl LVIII, the Kansas City Chiefs emerged victorious, defeating the San Francisco 49ers in a thrilling overtime showdown. The game was a nail-biter, with both teams showcasing their skills and determination.\n\nThe Chiefs, led by their star quarterback Patrick Mahomes, displayed their offensive prowess, while the 49ers, led by their strong defense, put up a tough fight. The game went into overtime, with the Chiefs ultimately securing the win with a touchdown.\n\nThe victory marked the Chiefs' second Super Bowl win in four years, solidifying their status as one of the top teams in the NFL. The game was a testament to the skill and talent of both teams, and a thrilling end to the NFL season.\n\nThe Super Bowl is not just about the game itself, but also about the halftime show and the commercials. This year's halftime show featured a star-studded lineup, including Usher, Alicia Keys, and Lil Jon. The show was a spectacle of music and dance, with the performers delivering an energetic and entertaining performance.\n"]
|
| 58 |
```
|
| 59 |
|
|
|
|
|
|
|
| 60 |
<details>
|
| 61 |
<summary><strong>Loading the model in half precision</strong></summary>
|
| 62 |
|
|
@@ -66,7 +66,6 @@ print(tokenizer.batch_decode(outputs))
|
|
| 66 |
from transformers import AutoModelForCausalLM
|
| 67 |
import torch
|
| 68 |
model = AutoModelForCausalLM.from_pretrained("ai21labs/Jamba-v0.1",
|
| 69 |
-
trust_remote_code=True,
|
| 70 |
torch_dtype=torch.bfloat16) # you can also use torch_dtype=torch.float16
|
| 71 |
```
|
| 72 |
|
|
@@ -75,7 +74,6 @@ When using half precision, you can enable the [FlashAttention2](https://github.c
|
|
| 75 |
from transformers import AutoModelForCausalLM
|
| 76 |
import torch
|
| 77 |
model = AutoModelForCausalLM.from_pretrained("ai21labs/Jamba-v0.1",
|
| 78 |
-
trust_remote_code=True,
|
| 79 |
torch_dtype=torch.bfloat16,
|
| 80 |
attn_implementation="flash_attention_2",
|
| 81 |
device_map="auto")
|
|
@@ -91,7 +89,6 @@ from transformers import AutoModelForCausalLM, BitsAndBytesConfig
|
|
| 91 |
quantization_config = BitsAndBytesConfig(load_in_8bit=True,
|
| 92 |
llm_int8_skip_modules=["mamba"])
|
| 93 |
model = AutoModelForCausalLM.from_pretrained("ai21labs/Jamba-v0.1",
|
| 94 |
-
trust_remote_code=True,
|
| 95 |
torch_dtype=torch.bfloat16,
|
| 96 |
attn_implementation="flash_attention_2",
|
| 97 |
quantization_config=quantization_config)
|
|
@@ -108,7 +105,7 @@ from peft import LoraConfig
|
|
| 108 |
from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments
|
| 109 |
|
| 110 |
tokenizer = AutoTokenizer.from_pretrained("ai21labs/Jamba-v0.1")
|
| 111 |
-
model = AutoModelForCausalLM.from_pretrained("ai21labs/Jamba-v0.1",
|
| 112 |
|
| 113 |
dataset = load_dataset("Abirate/english_quotes", split="train")
|
| 114 |
training_args = TrainingArguments(
|
|
|
|
| 27 |
|
| 28 |
## Usage
|
| 29 |
### Presequities
|
| 30 |
+
In order to use Jamba, it is recommended you use `transformers` version 4.40.0 or higher (version 4.39.0 or higher is required):
|
| 31 |
```bash
|
| 32 |
+
pip install transformers>=4.40.0
|
| 33 |
```
|
| 34 |
|
| 35 |
In order to run optimized Mamba implementations, you first need to install `mamba-ssm` and `causal-conv1d`:
|
|
|
|
| 41 |
You can run the model not using the optimized Mamba kernels, but it is **not** recommended as it will result in significantly lower latencies. In order to do that, you'll need to specify `use_mamba_kernels=False` when loading the model.
|
| 42 |
|
| 43 |
### Run the model
|
|
|
|
| 44 |
```python
|
| 45 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 46 |
|
| 47 |
+
model = AutoModelForCausalLM.from_pretrained("ai21labs/Jamba-v0.1")
|
|
|
|
| 48 |
tokenizer = AutoTokenizer.from_pretrained("ai21labs/Jamba-v0.1")
|
| 49 |
|
| 50 |
input_ids = tokenizer("In the recent Super Bowl LVIII,", return_tensors='pt').to(model.device)["input_ids"]
|
|
|
|
| 55 |
# ["<|startoftext|>In the recent Super Bowl LVIII, the Kansas City Chiefs emerged victorious, defeating the San Francisco 49ers in a thrilling overtime showdown. The game was a nail-biter, with both teams showcasing their skills and determination.\n\nThe Chiefs, led by their star quarterback Patrick Mahomes, displayed their offensive prowess, while the 49ers, led by their strong defense, put up a tough fight. The game went into overtime, with the Chiefs ultimately securing the win with a touchdown.\n\nThe victory marked the Chiefs' second Super Bowl win in four years, solidifying their status as one of the top teams in the NFL. The game was a testament to the skill and talent of both teams, and a thrilling end to the NFL season.\n\nThe Super Bowl is not just about the game itself, but also about the halftime show and the commercials. This year's halftime show featured a star-studded lineup, including Usher, Alicia Keys, and Lil Jon. The show was a spectacle of music and dance, with the performers delivering an energetic and entertaining performance.\n"]
|
| 56 |
```
|
| 57 |
|
| 58 |
+
Please note that if you're using `transformers<4.40.0`, `trust_remote_code=True` is required for running the new Jamba architecture.
|
| 59 |
+
|
| 60 |
<details>
|
| 61 |
<summary><strong>Loading the model in half precision</strong></summary>
|
| 62 |
|
|
|
|
| 66 |
from transformers import AutoModelForCausalLM
|
| 67 |
import torch
|
| 68 |
model = AutoModelForCausalLM.from_pretrained("ai21labs/Jamba-v0.1",
|
|
|
|
| 69 |
torch_dtype=torch.bfloat16) # you can also use torch_dtype=torch.float16
|
| 70 |
```
|
| 71 |
|
|
|
|
| 74 |
from transformers import AutoModelForCausalLM
|
| 75 |
import torch
|
| 76 |
model = AutoModelForCausalLM.from_pretrained("ai21labs/Jamba-v0.1",
|
|
|
|
| 77 |
torch_dtype=torch.bfloat16,
|
| 78 |
attn_implementation="flash_attention_2",
|
| 79 |
device_map="auto")
|
|
|
|
| 89 |
quantization_config = BitsAndBytesConfig(load_in_8bit=True,
|
| 90 |
llm_int8_skip_modules=["mamba"])
|
| 91 |
model = AutoModelForCausalLM.from_pretrained("ai21labs/Jamba-v0.1",
|
|
|
|
| 92 |
torch_dtype=torch.bfloat16,
|
| 93 |
attn_implementation="flash_attention_2",
|
| 94 |
quantization_config=quantization_config)
|
|
|
|
| 105 |
from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments
|
| 106 |
|
| 107 |
tokenizer = AutoTokenizer.from_pretrained("ai21labs/Jamba-v0.1")
|
| 108 |
+
model = AutoModelForCausalLM.from_pretrained("ai21labs/Jamba-v0.1", device_map='auto')
|
| 109 |
|
| 110 |
dataset = load_dataset("Abirate/english_quotes", split="train")
|
| 111 |
training_args = TrainingArguments(
|
config.json
CHANGED
|
@@ -12,7 +12,6 @@
|
|
| 12 |
"AutoModelForSequenceClassification": "model.JambaForSequenceClassification"
|
| 13 |
},
|
| 14 |
"bos_token_id": 1,
|
| 15 |
-
"calc_logits_for_entire_prompt": false,
|
| 16 |
"eos_token_id": 2,
|
| 17 |
"expert_layer_offset": 1,
|
| 18 |
"expert_layer_period": 2,
|
|
@@ -25,15 +24,15 @@
|
|
| 25 |
"mamba_d_state": 16,
|
| 26 |
"mamba_dt_rank": 256,
|
| 27 |
"mamba_expand": 2,
|
| 28 |
-
"mamba_inner_layernorms": true,
|
| 29 |
"mamba_proj_bias": false,
|
|
|
|
| 30 |
"model_type": "jamba",
|
| 31 |
-
"n_ctx": 262144,
|
| 32 |
"num_attention_heads": 32,
|
| 33 |
"num_experts": 16,
|
| 34 |
"num_experts_per_tok": 2,
|
| 35 |
"num_hidden_layers": 32,
|
| 36 |
"num_key_value_heads": 8,
|
|
|
|
| 37 |
"output_router_logits": false,
|
| 38 |
"pad_token_id": 0,
|
| 39 |
"rms_norm_eps": 1e-06,
|
|
@@ -41,7 +40,7 @@
|
|
| 41 |
"sliding_window": null,
|
| 42 |
"tie_word_embeddings": false,
|
| 43 |
"torch_dtype": "bfloat16",
|
| 44 |
-
"transformers_version": "4.40.
|
| 45 |
"use_cache": true,
|
| 46 |
"use_mamba_kernels": true,
|
| 47 |
"vocab_size": 65536
|
|
|
|
| 12 |
"AutoModelForSequenceClassification": "model.JambaForSequenceClassification"
|
| 13 |
},
|
| 14 |
"bos_token_id": 1,
|
|
|
|
| 15 |
"eos_token_id": 2,
|
| 16 |
"expert_layer_offset": 1,
|
| 17 |
"expert_layer_period": 2,
|
|
|
|
| 24 |
"mamba_d_state": 16,
|
| 25 |
"mamba_dt_rank": 256,
|
| 26 |
"mamba_expand": 2,
|
|
|
|
| 27 |
"mamba_proj_bias": false,
|
| 28 |
+
"max_position_embeddings": 262144,
|
| 29 |
"model_type": "jamba",
|
|
|
|
| 30 |
"num_attention_heads": 32,
|
| 31 |
"num_experts": 16,
|
| 32 |
"num_experts_per_tok": 2,
|
| 33 |
"num_hidden_layers": 32,
|
| 34 |
"num_key_value_heads": 8,
|
| 35 |
+
"num_logits_to_keep": 1,
|
| 36 |
"output_router_logits": false,
|
| 37 |
"pad_token_id": 0,
|
| 38 |
"rms_norm_eps": 1e-06,
|
|
|
|
| 40 |
"sliding_window": null,
|
| 41 |
"tie_word_embeddings": false,
|
| 42 |
"torch_dtype": "bfloat16",
|
| 43 |
+
"transformers_version": "4.40.1",
|
| 44 |
"use_cache": true,
|
| 45 |
"use_mamba_kernels": true,
|
| 46 |
"vocab_size": 65536
|
configuration_jamba.py
CHANGED
|
@@ -26,9 +26,9 @@ class JambaConfig(PretrainedConfig):
|
|
| 26 |
r"""
|
| 27 |
This is the configuration class to store the configuration of a [`JambaModel`]. It is used to instantiate a
|
| 28 |
Jamba model according to the specified arguments, defining the model architecture. Instantiating a configuration
|
| 29 |
-
with the defaults will yield a similar configuration to that of the
|
| 30 |
|
| 31 |
-
[ai21labs/
|
| 32 |
|
| 33 |
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
| 34 |
documentation from [`PretrainedConfig`] for more information.
|
|
@@ -65,12 +65,12 @@ class JambaConfig(PretrainedConfig):
|
|
| 65 |
use_cache (`bool`, *optional*, defaults to `True`):
|
| 66 |
Whether or not the model should return the last key/values attentions (not used by all models). Only
|
| 67 |
relevant if `config.is_decoder=True`.
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
last
|
| 71 |
-
|
| 72 |
-
will reduce memory footprint
|
| 73 |
-
|
| 74 |
output_router_logits (`bool`, *optional*, defaults to `False`):
|
| 75 |
Whether or not the router logits should be returned by the model. Enabling this will also
|
| 76 |
allow the model to output the auxiliary loss. See [here]() for more details
|
|
@@ -84,7 +84,7 @@ class JambaConfig(PretrainedConfig):
|
|
| 84 |
The id of the "end-of-sequence" token.
|
| 85 |
sliding_window (`int`, *optional*):
|
| 86 |
Sliding window attention window size. If not specified, will default to `None`.
|
| 87 |
-
|
| 88 |
This value doesn't have any real effect. The maximum sequence length that this model is intended to be
|
| 89 |
used with. It can be used with longer sequences, but performance may degrade.
|
| 90 |
attention_dropout (`float`, *optional*, defaults to 0.0):
|
|
@@ -118,8 +118,6 @@ class JambaConfig(PretrainedConfig):
|
|
| 118 |
Flag indicating whether or not to use bias in the convolution layer of the mamba mixer block.
|
| 119 |
mamba_proj_bias (`bool`, *optional*, defaults to `False`):
|
| 120 |
Flag indicating whether or not to use bias in the input and output projections (["in_proj", "out_proj"]) of the mamba mixer block
|
| 121 |
-
mamba_inner_layernorms (`bool`, *optional*, defaults to `True`):
|
| 122 |
-
Flag indicating whether or not to apply layernorms to internal mamba activations
|
| 123 |
|
| 124 |
"""
|
| 125 |
|
|
@@ -139,14 +137,14 @@ class JambaConfig(PretrainedConfig):
|
|
| 139 |
initializer_range=0.02,
|
| 140 |
rms_norm_eps=1e-6,
|
| 141 |
use_cache=True,
|
| 142 |
-
|
| 143 |
output_router_logits=False,
|
| 144 |
router_aux_loss_coef=0.001,
|
| 145 |
pad_token_id=0,
|
| 146 |
bos_token_id=1,
|
| 147 |
eos_token_id=2,
|
| 148 |
sliding_window=None,
|
| 149 |
-
|
| 150 |
attention_dropout=0.0,
|
| 151 |
num_experts_per_tok=2,
|
| 152 |
num_experts=16,
|
|
@@ -161,7 +159,6 @@ class JambaConfig(PretrainedConfig):
|
|
| 161 |
mamba_dt_rank="auto",
|
| 162 |
mamba_conv_bias=True,
|
| 163 |
mamba_proj_bias=False,
|
| 164 |
-
mamba_inner_layernorms=True,
|
| 165 |
**kwargs,
|
| 166 |
):
|
| 167 |
self.vocab_size = vocab_size
|
|
@@ -171,7 +168,7 @@ class JambaConfig(PretrainedConfig):
|
|
| 171 |
self.num_hidden_layers = num_hidden_layers
|
| 172 |
self.num_attention_heads = num_attention_heads
|
| 173 |
self.sliding_window = sliding_window
|
| 174 |
-
self.
|
| 175 |
self.attention_dropout = attention_dropout
|
| 176 |
|
| 177 |
# for backward compatibility
|
|
@@ -184,7 +181,7 @@ class JambaConfig(PretrainedConfig):
|
|
| 184 |
self.rms_norm_eps = rms_norm_eps
|
| 185 |
|
| 186 |
self.use_cache = use_cache
|
| 187 |
-
self.
|
| 188 |
self.output_router_logits = output_router_logits
|
| 189 |
self.router_aux_loss_coef = router_aux_loss_coef
|
| 190 |
|
|
@@ -202,7 +199,6 @@ class JambaConfig(PretrainedConfig):
|
|
| 202 |
self.mamba_dt_rank = math.ceil(self.hidden_size / 16) if mamba_dt_rank == "auto" else mamba_dt_rank
|
| 203 |
self.mamba_conv_bias = mamba_conv_bias
|
| 204 |
self.mamba_proj_bias = mamba_proj_bias
|
| 205 |
-
self.mamba_inner_layernorms = mamba_inner_layernorms
|
| 206 |
|
| 207 |
super().__init__(
|
| 208 |
pad_token_id=pad_token_id,
|
|
@@ -211,3 +207,17 @@ class JambaConfig(PretrainedConfig):
|
|
| 211 |
tie_word_embeddings=tie_word_embeddings,
|
| 212 |
**kwargs,
|
| 213 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
r"""
|
| 27 |
This is the configuration class to store the configuration of a [`JambaModel`]. It is used to instantiate a
|
| 28 |
Jamba model according to the specified arguments, defining the model architecture. Instantiating a configuration
|
| 29 |
+
with the defaults will yield a similar configuration to that of the Jamba-v0.1 model.
|
| 30 |
|
| 31 |
+
[ai21labs/Jamba-v0.1](https://huggingface.co/ai21labs/Jamba-v0.1)
|
| 32 |
|
| 33 |
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
| 34 |
documentation from [`PretrainedConfig`] for more information.
|
|
|
|
| 65 |
use_cache (`bool`, *optional*, defaults to `True`):
|
| 66 |
Whether or not the model should return the last key/values attentions (not used by all models). Only
|
| 67 |
relevant if `config.is_decoder=True`.
|
| 68 |
+
num_logits_to_keep (`int` or `None`, *optional*, defaults to 1):
|
| 69 |
+
Number of prompt logits to calculate during generation. If `None`, all logits will be calculated. If an
|
| 70 |
+
integer value, only last `num_logits_to_keep` logits will be calculated. Default is 1 because only the
|
| 71 |
+
logits of the last prompt token are needed for generation. For long sequences, the logits for the entire
|
| 72 |
+
sequence may use a lot of memory so, setting `num_logits_to_keep=1` will reduce memory footprint
|
| 73 |
+
significantly.
|
| 74 |
output_router_logits (`bool`, *optional*, defaults to `False`):
|
| 75 |
Whether or not the router logits should be returned by the model. Enabling this will also
|
| 76 |
allow the model to output the auxiliary loss. See [here]() for more details
|
|
|
|
| 84 |
The id of the "end-of-sequence" token.
|
| 85 |
sliding_window (`int`, *optional*):
|
| 86 |
Sliding window attention window size. If not specified, will default to `None`.
|
| 87 |
+
max_position_embeddings (`int`, *optional*, defaults to 262144):
|
| 88 |
This value doesn't have any real effect. The maximum sequence length that this model is intended to be
|
| 89 |
used with. It can be used with longer sequences, but performance may degrade.
|
| 90 |
attention_dropout (`float`, *optional*, defaults to 0.0):
|
|
|
|
| 118 |
Flag indicating whether or not to use bias in the convolution layer of the mamba mixer block.
|
| 119 |
mamba_proj_bias (`bool`, *optional*, defaults to `False`):
|
| 120 |
Flag indicating whether or not to use bias in the input and output projections (["in_proj", "out_proj"]) of the mamba mixer block
|
|
|
|
|
|
|
| 121 |
|
| 122 |
"""
|
| 123 |
|
|
|
|
| 137 |
initializer_range=0.02,
|
| 138 |
rms_norm_eps=1e-6,
|
| 139 |
use_cache=True,
|
| 140 |
+
num_logits_to_keep=1,
|
| 141 |
output_router_logits=False,
|
| 142 |
router_aux_loss_coef=0.001,
|
| 143 |
pad_token_id=0,
|
| 144 |
bos_token_id=1,
|
| 145 |
eos_token_id=2,
|
| 146 |
sliding_window=None,
|
| 147 |
+
max_position_embeddings=262144,
|
| 148 |
attention_dropout=0.0,
|
| 149 |
num_experts_per_tok=2,
|
| 150 |
num_experts=16,
|
|
|
|
| 159 |
mamba_dt_rank="auto",
|
| 160 |
mamba_conv_bias=True,
|
| 161 |
mamba_proj_bias=False,
|
|
|
|
| 162 |
**kwargs,
|
| 163 |
):
|
| 164 |
self.vocab_size = vocab_size
|
|
|
|
| 168 |
self.num_hidden_layers = num_hidden_layers
|
| 169 |
self.num_attention_heads = num_attention_heads
|
| 170 |
self.sliding_window = sliding_window
|
| 171 |
+
self.max_position_embeddings = max_position_embeddings
|
| 172 |
self.attention_dropout = attention_dropout
|
| 173 |
|
| 174 |
# for backward compatibility
|
|
|
|
| 181 |
self.rms_norm_eps = rms_norm_eps
|
| 182 |
|
| 183 |
self.use_cache = use_cache
|
| 184 |
+
self.num_logits_to_keep = num_logits_to_keep
|
| 185 |
self.output_router_logits = output_router_logits
|
| 186 |
self.router_aux_loss_coef = router_aux_loss_coef
|
| 187 |
|
|
|
|
| 199 |
self.mamba_dt_rank = math.ceil(self.hidden_size / 16) if mamba_dt_rank == "auto" else mamba_dt_rank
|
| 200 |
self.mamba_conv_bias = mamba_conv_bias
|
| 201 |
self.mamba_proj_bias = mamba_proj_bias
|
|
|
|
| 202 |
|
| 203 |
super().__init__(
|
| 204 |
pad_token_id=pad_token_id,
|
|
|
|
| 207 |
tie_word_embeddings=tie_word_embeddings,
|
| 208 |
**kwargs,
|
| 209 |
)
|
| 210 |
+
|
| 211 |
+
@property
|
| 212 |
+
def layers_block_type(self):
|
| 213 |
+
return [
|
| 214 |
+
"attention" if i % self.attn_layer_period == self.attn_layer_offset else "mamba"
|
| 215 |
+
for i in range(self.num_hidden_layers)
|
| 216 |
+
]
|
| 217 |
+
|
| 218 |
+
@property
|
| 219 |
+
def layers_num_experts(self):
|
| 220 |
+
return [
|
| 221 |
+
self.num_experts if i % self.expert_layer_period == self.expert_layer_offset else 1
|
| 222 |
+
for i in range(self.num_hidden_layers)
|
| 223 |
+
]
|
generation_config.json
CHANGED
|
@@ -3,5 +3,5 @@
|
|
| 3 |
"bos_token_id": 1,
|
| 4 |
"eos_token_id": 2,
|
| 5 |
"pad_token_id": 0,
|
| 6 |
-
"transformers_version": "4.40.
|
| 7 |
}
|
|
|
|
| 3 |
"bos_token_id": 1,
|
| 4 |
"eos_token_id": 2,
|
| 5 |
"pad_token_id": 0,
|
| 6 |
+
"transformers_version": "4.40.1"
|
| 7 |
}
|
model-00001-of-00021.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1aace34ee0da3bf95605bd150fff6d3e78110be4048a3c389b0a740354b2ccb7
|
| 3 |
+
size 4951761424
|
model-00002-of-00021.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:0ba1de67a86329431f14f7ffa165d84055d32ce57a6d2314e3b2464eac3732dc
|
| 3 |
+
size 4884669624
|
model-00003-of-00021.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1abc4f16865fb78241c9453292ee3b2ca2c1e2d54ee945631da625834b95c9b2
|
| 3 |
+
size 4992557120
|
model-00004-of-00021.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:45fab97739a58e924791572ea3d06f9c90b9ff2a299460aaa4bd87c6e9d424f3
|
| 3 |
+
size 4958853560
|
model-00005-of-00021.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c4b0ec6e8f33e6d7b1f837cd4c25818487dcc7e478734606da28110507e51c97
|
| 3 |
+
size 4975763832
|
model-00006-of-00021.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ed98d5c3c8d7ab7352944bea09b0d54d98066cf567ba3d069da12c05575d56ed
|
| 3 |
+
size 4884669616
|
model-00007-of-00021.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:735be2bc568711bf42a4caebcda8288dd300b31b48fa098b00df3cf1a98e10e2
|
| 3 |
+
size 4884669640
|
model-00008-of-00021.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d0c8d817b2b47661d361e8b520128b3194185f756cc2204a95d642e24895ee51
|
| 3 |
+
size 4992557176
|
model-00009-of-00021.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e50222cf865ca5678d22574b131294303c46b249478cf70113c701f70331e999
|
| 3 |
+
size 4932507176
|
model-00010-of-00021.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b1b4b69b24ae55827b6c8b1e4a10807aa3525bc85f4d34dc002ac7440757fbf4
|
| 3 |
+
size 4884669672
|
model-00011-of-00021.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:60213cac13b92ed34b93ce48e670434f22e3bf8b2b8df20c60b7bf8a9515c35c
|
| 3 |
+
size 4884669696
|
model-00012-of-00021.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:05805eacd3bb40cc9da802350409f1cb078e8b276da7e06c7a8a5ca5b26cc887
|
| 3 |
+
size 4884669688
|
model-00013-of-00021.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:201df979a1b34ced6cdbb7a790163412636779f1119e3845a704c489181d03d2
|
| 3 |
+
size 4932507176
|
model-00014-of-00021.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d0a7eb42a9ea3a385442c2e758dd5efd5dc5b913f1d10bfd37792cc963a33c93
|
| 3 |
+
size 4992557152
|
model-00015-of-00021.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a4b9afe4398000c28b36e3aa40c87086af673d4f8a64bfc5767941ab2008bcc9
|
| 3 |
+
size 4884669688
|
model-00016-of-00021.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:dd1ac6cc861971c43bdf0c9c6d4c9fe72d33e5227e054a621e2e68f001419763
|
| 3 |
+
size 4884669688
|
model-00017-of-00021.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:52d9eea696dd29ef413d617bbcb62a9f159e8fe8170d36e018932cef45ee281d
|
| 3 |
+
size 4908522856
|
model-00018-of-00021.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:77acada7c098e81280645ea0a9dbfa00196dca6da8946498b9907e9e376fb42d
|
| 3 |
+
size 4908654000
|
model-00019-of-00021.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:09e10dfd6c6459cd3460b1d667639717d3657274c1694c19a6fdbac1be6a76bf
|
| 3 |
+
size 4992557168
|
model-00020-of-00021.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2bd5c27b2cca6e06f7b4497ce8c9b1522a64846817a871bad274d08507960ed0
|
| 3 |
+
size 4884669696
|
model-00021-of-00021.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a47ef23db8deb5364da676a40dc3dcb011fb9d9ceef13ba044c176e9a83ac1e3
|
| 3 |
+
size 4647318576
|
model.safetensors.index.json
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
modeling_jamba.py
CHANGED
|
@@ -20,8 +20,6 @@
|
|
| 20 |
""" PyTorch Jamba model."""
|
| 21 |
import inspect
|
| 22 |
import math
|
| 23 |
-
import warnings
|
| 24 |
-
from dataclasses import dataclass, field
|
| 25 |
from typing import Any, Dict, List, Optional, Tuple, Union
|
| 26 |
|
| 27 |
import torch
|
|
@@ -31,10 +29,9 @@ from torch import nn
|
|
| 31 |
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
| 32 |
|
| 33 |
from transformers.activations import ACT2FN
|
| 34 |
-
from transformers.cache_utils import
|
| 35 |
from transformers.modeling_attn_mask_utils import (
|
| 36 |
-
|
| 37 |
-
_prepare_4d_causal_attention_mask_for_sdpa,
|
| 38 |
)
|
| 39 |
from transformers.modeling_outputs import (
|
| 40 |
MoeCausalLMOutputWithPast,
|
|
@@ -42,7 +39,6 @@ from transformers.modeling_outputs import (
|
|
| 42 |
SequenceClassifierOutputWithPast,
|
| 43 |
)
|
| 44 |
from transformers.modeling_utils import PreTrainedModel
|
| 45 |
-
from transformers.pytorch_utils import is_torch_greater_or_equal_than_1_13
|
| 46 |
from transformers.utils import (
|
| 47 |
add_start_docstrings,
|
| 48 |
add_start_docstrings_to_model_forward,
|
|
@@ -50,11 +46,15 @@ from transformers.utils import (
|
|
| 50 |
logging,
|
| 51 |
replace_return_docstrings,
|
| 52 |
)
|
| 53 |
-
from transformers.utils.import_utils import
|
|
|
|
|
|
|
|
|
|
|
|
|
| 54 |
from .configuration_jamba import JambaConfig
|
| 55 |
|
| 56 |
|
| 57 |
-
# try except block so it'll work with trust_remote_code.
|
| 58 |
try:
|
| 59 |
from flash_attn import flash_attn_func, flash_attn_varlen_func
|
| 60 |
from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input # noqa
|
|
@@ -63,22 +63,15 @@ try:
|
|
| 63 |
except ImportError:
|
| 64 |
pass
|
| 65 |
|
| 66 |
-
# This makes `_prepare_4d_causal_attention_mask` a leaf function in the FX graph.
|
| 67 |
-
# It means that the function will not be traced through and simply appear as a node in the graph.
|
| 68 |
-
if is_torch_fx_available():
|
| 69 |
-
if not is_torch_greater_or_equal_than_1_13:
|
| 70 |
-
import torch.fx
|
| 71 |
-
|
| 72 |
-
_prepare_4d_causal_attention_mask = torch.fx.wrap(_prepare_4d_causal_attention_mask)
|
| 73 |
|
| 74 |
-
# try except block so it'll work with trust_remote_code.
|
| 75 |
try:
|
| 76 |
from mamba_ssm.ops.selective_scan_interface import mamba_inner_fn, selective_scan_fn
|
| 77 |
from mamba_ssm.ops.triton.selective_state_update import selective_state_update
|
| 78 |
except ImportError:
|
| 79 |
selective_state_update, selective_scan_fn, mamba_inner_fn = None, None, None
|
| 80 |
|
| 81 |
-
# try except block so it'll work with trust_remote_code.
|
| 82 |
try:
|
| 83 |
from causal_conv1d import causal_conv1d_fn, causal_conv1d_update
|
| 84 |
except ImportError:
|
|
@@ -94,9 +87,12 @@ logger = logging.get_logger(__name__)
|
|
| 94 |
_CONFIG_FOR_DOC = "JambaConfig"
|
| 95 |
|
| 96 |
|
| 97 |
-
#
|
| 98 |
def load_balancing_loss_func(
|
| 99 |
-
|
|
|
|
|
|
|
|
|
|
| 100 |
) -> float:
|
| 101 |
r"""
|
| 102 |
Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.
|
|
@@ -106,7 +102,7 @@ def load_balancing_loss_func(
|
|
| 106 |
experts is too unbalanced.
|
| 107 |
|
| 108 |
Args:
|
| 109 |
-
|
| 110 |
Logits from the `router`, should be a tuple of model.config.num_hidden_layers tensors of
|
| 111 |
shape [batch_size X sequence_length, num_experts].
|
| 112 |
attention_mask (`torch.Tensor`, None):
|
|
@@ -118,16 +114,16 @@ def load_balancing_loss_func(
|
|
| 118 |
Returns:
|
| 119 |
The auxiliary loss.
|
| 120 |
"""
|
| 121 |
-
if
|
| 122 |
return 0
|
| 123 |
|
| 124 |
-
if isinstance(
|
| 125 |
-
compute_device =
|
| 126 |
-
|
| 127 |
-
[
|
| 128 |
)
|
| 129 |
|
| 130 |
-
routing_weights = torch.nn.functional.softmax(
|
| 131 |
|
| 132 |
_, selected_experts = torch.topk(routing_weights, top_k, dim=-1)
|
| 133 |
|
|
@@ -141,7 +137,7 @@ def load_balancing_loss_func(
|
|
| 141 |
router_prob_per_expert = torch.mean(routing_weights, dim=0)
|
| 142 |
else:
|
| 143 |
batch_size, sequence_length = attention_mask.shape
|
| 144 |
-
num_hidden_layers =
|
| 145 |
|
| 146 |
# Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask
|
| 147 |
expert_attention_mask = (
|
|
@@ -217,6 +213,82 @@ def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
|
| 217 |
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
| 218 |
|
| 219 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 220 |
# Adapted from transformers.models.mistral.modeling_mistral.MistralAttention with Mistral->Jamba
|
| 221 |
class JambaAttention(nn.Module):
|
| 222 |
"""
|
|
@@ -253,23 +325,16 @@ class JambaAttention(nn.Module):
|
|
| 253 |
self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
|
| 254 |
self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
|
| 255 |
|
| 256 |
-
def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
|
| 257 |
-
return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
|
| 258 |
-
|
| 259 |
def forward(
|
| 260 |
self,
|
| 261 |
hidden_states: torch.Tensor,
|
| 262 |
attention_mask: Optional[torch.Tensor] = None,
|
| 263 |
position_ids: Optional[torch.LongTensor] = None,
|
| 264 |
-
past_key_value: Optional[
|
| 265 |
output_attentions: bool = False,
|
| 266 |
use_cache: bool = False,
|
| 267 |
-
|
| 268 |
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
| 269 |
-
if "padding_mask" in kwargs:
|
| 270 |
-
warnings.warn(
|
| 271 |
-
"Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
|
| 272 |
-
)
|
| 273 |
bsz, q_len, _ = hidden_states.size()
|
| 274 |
|
| 275 |
query_states = self.q_proj(hidden_states)
|
|
@@ -280,16 +345,6 @@ class JambaAttention(nn.Module):
|
|
| 280 |
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 281 |
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 282 |
|
| 283 |
-
kv_seq_len = key_states.shape[-2]
|
| 284 |
-
if past_key_value is not None:
|
| 285 |
-
if self.layer_idx is None:
|
| 286 |
-
raise ValueError(
|
| 287 |
-
f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
|
| 288 |
-
"for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
|
| 289 |
-
"with a layer index."
|
| 290 |
-
)
|
| 291 |
-
kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
|
| 292 |
-
|
| 293 |
if past_key_value is not None:
|
| 294 |
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)
|
| 295 |
|
|
@@ -299,19 +354,9 @@ class JambaAttention(nn.Module):
|
|
| 299 |
|
| 300 |
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
|
| 301 |
|
| 302 |
-
if
|
| 303 |
-
|
| 304 |
-
|
| 305 |
-
f" {attn_weights.size()}"
|
| 306 |
-
)
|
| 307 |
-
|
| 308 |
-
if attention_mask is not None:
|
| 309 |
-
if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
|
| 310 |
-
raise ValueError(
|
| 311 |
-
f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
|
| 312 |
-
)
|
| 313 |
-
|
| 314 |
-
attn_weights = attn_weights + attention_mask
|
| 315 |
|
| 316 |
# upcast attention to fp32
|
| 317 |
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
|
|
@@ -357,37 +402,26 @@ class JambaFlashAttention2(JambaAttention):
|
|
| 357 |
hidden_states: torch.Tensor,
|
| 358 |
attention_mask: Optional[torch.Tensor] = None,
|
| 359 |
position_ids: Optional[torch.LongTensor] = None,
|
| 360 |
-
past_key_value: Optional[
|
| 361 |
output_attentions: bool = False,
|
| 362 |
use_cache: bool = False,
|
|
|
|
| 363 |
**kwargs,
|
| 364 |
):
|
| 365 |
-
if "padding_mask" in kwargs:
|
| 366 |
-
warnings.warn(
|
| 367 |
-
"Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
|
| 368 |
-
)
|
| 369 |
-
|
| 370 |
-
# overwrite attention_mask with padding_mask
|
| 371 |
-
attention_mask = kwargs.pop("padding_mask")
|
| 372 |
bsz, q_len, _ = hidden_states.size()
|
| 373 |
|
| 374 |
query_states = self.q_proj(hidden_states)
|
| 375 |
key_states = self.k_proj(hidden_states)
|
| 376 |
value_states = self.v_proj(hidden_states)
|
| 377 |
|
|
|
|
|
|
|
|
|
|
| 378 |
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 379 |
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 380 |
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 381 |
|
| 382 |
-
kv_seq_len =
|
| 383 |
-
if past_key_value is not None:
|
| 384 |
-
if self.layer_idx is None:
|
| 385 |
-
raise ValueError(
|
| 386 |
-
f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
|
| 387 |
-
"for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
|
| 388 |
-
"with a layer index."
|
| 389 |
-
)
|
| 390 |
-
kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
|
| 391 |
|
| 392 |
use_sliding_windows = (
|
| 393 |
_flash_supports_window_size
|
|
@@ -403,7 +437,7 @@ class JambaFlashAttention2(JambaAttention):
|
|
| 403 |
|
| 404 |
if past_key_value is not None:
|
| 405 |
# Activate slicing cache only if the config has a value `sliding_windows` attribute
|
| 406 |
-
cache_has_contents =
|
| 407 |
if (
|
| 408 |
getattr(self.config, "sliding_window", None) is not None
|
| 409 |
and kv_seq_len > self.config.sliding_window
|
|
@@ -505,7 +539,7 @@ class JambaFlashAttention2(JambaAttention):
|
|
| 505 |
attention_mask (`torch.Tensor`):
|
| 506 |
The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
|
| 507 |
position of padding tokens and 1 for the position of non-padding tokens.
|
| 508 |
-
dropout (`
|
| 509 |
Attention dropout
|
| 510 |
softmax_scale (`float`, *optional*):
|
| 511 |
The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
|
|
@@ -580,6 +614,7 @@ class JambaFlashAttention2(JambaAttention):
|
|
| 580 |
|
| 581 |
return attn_output
|
| 582 |
|
|
|
|
| 583 |
def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
|
| 584 |
batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape
|
| 585 |
|
|
@@ -637,9 +672,10 @@ class JambaSdpaAttention(JambaAttention):
|
|
| 637 |
hidden_states: torch.Tensor,
|
| 638 |
attention_mask: Optional[torch.Tensor] = None,
|
| 639 |
position_ids: Optional[torch.LongTensor] = None,
|
| 640 |
-
past_key_value: Optional[
|
| 641 |
output_attentions: bool = False,
|
| 642 |
use_cache: bool = False,
|
|
|
|
| 643 |
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
| 644 |
if output_attentions:
|
| 645 |
# TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
|
|
@@ -666,21 +702,15 @@ class JambaSdpaAttention(JambaAttention):
|
|
| 666 |
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 667 |
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 668 |
|
| 669 |
-
kv_seq_len = key_states.shape[-2]
|
| 670 |
-
if past_key_value is not None:
|
| 671 |
-
kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
|
| 672 |
-
|
| 673 |
if past_key_value is not None:
|
| 674 |
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)
|
| 675 |
|
| 676 |
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
| 677 |
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
| 678 |
|
|
|
|
| 679 |
if attention_mask is not None:
|
| 680 |
-
|
| 681 |
-
raise ValueError(
|
| 682 |
-
f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
|
| 683 |
-
)
|
| 684 |
|
| 685 |
# SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
|
| 686 |
# Reference: https://github.com/pytorch/pytorch/issues/112577.
|
|
@@ -693,7 +723,7 @@ class JambaSdpaAttention(JambaAttention):
|
|
| 693 |
query_states,
|
| 694 |
key_states,
|
| 695 |
value_states,
|
| 696 |
-
attn_mask=
|
| 697 |
dropout_p=self.attention_dropout if self.training else 0.0,
|
| 698 |
# 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.
|
| 699 |
is_causal=self.is_causal and attention_mask is None and q_len > 1,
|
|
@@ -714,99 +744,6 @@ JAMBA_ATTENTION_CLASSES = {
|
|
| 714 |
}
|
| 715 |
|
| 716 |
|
| 717 |
-
class HybridMambaAttentionDynamicCache(DynamicCache):
|
| 718 |
-
"""
|
| 719 |
-
A dynamic cache that can handle both the attention cache (which has a seq_len dimension) and the mamba cache
|
| 720 |
-
(which has a constant shape regardless of seq_len).
|
| 721 |
-
|
| 722 |
-
It stores the Key and Value states as a list of tensors, one for each layer.
|
| 723 |
-
The expected shape for each tensor for attention layers is `[batch_size, num_heads, seq_len, head_dim]`.
|
| 724 |
-
For the mamba layers, the `key_cache` represents the convolution state and has a shape of `[batch_size, d_inner, 1, d_conv]`,
|
| 725 |
-
and the `value_cache` represents the ssm state and has a shape of `[batch_size, d_inner, 1, d_state]`. Mamba cache
|
| 726 |
-
shape[2] is a dummy "seqlen" dimension to match the number of attention cache dimensions. For mamba, the cache
|
| 727 |
-
doesn't grow with seqlen so this dimension is always 1.
|
| 728 |
-
"""
|
| 729 |
-
|
| 730 |
-
def __init__(self) -> None:
|
| 731 |
-
super().__init__()
|
| 732 |
-
self.attention_layer_idx = None # used to know which layer has data on seqlen in the cache shape
|
| 733 |
-
|
| 734 |
-
def update(
|
| 735 |
-
self,
|
| 736 |
-
key_states: torch.Tensor,
|
| 737 |
-
value_states: torch.Tensor,
|
| 738 |
-
layer_idx: int,
|
| 739 |
-
cache_kwargs: Optional[Dict[str, Any]] = None,
|
| 740 |
-
) -> Tuple[torch.Tensor, torch.Tensor]:
|
| 741 |
-
"""
|
| 742 |
-
Updates the cache with the new `key_states` and `value_states` for the layer `layer_idx`.
|
| 743 |
-
|
| 744 |
-
Parameters:
|
| 745 |
-
key_states (`torch.Tensor`):
|
| 746 |
-
The new key states to cache.
|
| 747 |
-
value_states (`torch.Tensor`):
|
| 748 |
-
The new value states to cache.
|
| 749 |
-
layer_idx (`int`):
|
| 750 |
-
The index of the layer to cache the states for.
|
| 751 |
-
cache_kwargs (`Dict[str, Any]`, `optional`):
|
| 752 |
-
Additional arguments for the cache subclass. No additional arguments are used in `HybridMambaAttentionDynamicCache`.
|
| 753 |
-
|
| 754 |
-
Return:
|
| 755 |
-
A tuple containing the updated key and value states.
|
| 756 |
-
"""
|
| 757 |
-
# Update the number of seen tokens
|
| 758 |
-
if self.attention_layer_idx is None and self._is_attn_layer(key_states, value_states):
|
| 759 |
-
self.attention_layer_idx = layer_idx
|
| 760 |
-
if self.attention_layer_idx is not None and layer_idx == self.attention_layer_idx:
|
| 761 |
-
if hasattr(self, "_seen_tokens"):
|
| 762 |
-
self._seen_tokens += key_states.shape[-2]
|
| 763 |
-
else:
|
| 764 |
-
self.seen_tokens += key_states.shape[-2]
|
| 765 |
-
|
| 766 |
-
# Update the cache
|
| 767 |
-
if len(self.key_cache) <= layer_idx:
|
| 768 |
-
self.key_cache.append(key_states)
|
| 769 |
-
self.value_cache.append(value_states)
|
| 770 |
-
else:
|
| 771 |
-
if self._is_attn_layer(self.key_cache[layer_idx], self.value_cache[layer_idx]):
|
| 772 |
-
# attention layer - append the new states to the existing cache on the seqlen dimension
|
| 773 |
-
self.key_cache[layer_idx] = torch.cat([self.key_cache[layer_idx], key_states], dim=-2)
|
| 774 |
-
self.value_cache[layer_idx] = torch.cat([self.value_cache[layer_idx], value_states], dim=-2)
|
| 775 |
-
else:
|
| 776 |
-
# mamba layer - replace the cache with the new states
|
| 777 |
-
self.key_cache[layer_idx] = key_states
|
| 778 |
-
self.value_cache[layer_idx] = value_states
|
| 779 |
-
|
| 780 |
-
return self.key_cache[layer_idx], self.value_cache[layer_idx]
|
| 781 |
-
|
| 782 |
-
def get_seq_length(self, layer_idx: Optional[int] = None) -> int:
|
| 783 |
-
"""Returns the sequence length of the cached states. A layer index can be optionally passed."""
|
| 784 |
-
if layer_idx is not None:
|
| 785 |
-
if len(self.key_cache) <= layer_idx:
|
| 786 |
-
return 0
|
| 787 |
-
if self._is_attn_layer(self.key_cache[layer_idx], self.value_cache[layer_idx]):
|
| 788 |
-
return self.key_cache[layer_idx].shape[-2]
|
| 789 |
-
else:
|
| 790 |
-
warnings.warn(
|
| 791 |
-
f"Asked to get the sequence length from cache of layer {layer_idx} which is not an attention layer. "
|
| 792 |
-
f"Ignoring that and using an attention layer cache"
|
| 793 |
-
)
|
| 794 |
-
if self.attention_layer_idx is None or len(self.key_cache) <= self.attention_layer_idx:
|
| 795 |
-
return 0
|
| 796 |
-
return self.key_cache[self.attention_layer_idx].shape[-2]
|
| 797 |
-
|
| 798 |
-
@staticmethod
|
| 799 |
-
def _is_attn_layer(key_states: torch.Tensor, value_states: torch.Tensor):
|
| 800 |
-
return key_states.shape[-1] == value_states.shape[-1]
|
| 801 |
-
|
| 802 |
-
|
| 803 |
-
@dataclass
|
| 804 |
-
class MambaCacheParams:
|
| 805 |
-
seqlen_offset: int = 0
|
| 806 |
-
conv_states: Dict[int, torch.Tensor] = field(default_factory=dict)
|
| 807 |
-
ssm_states: Dict[int, torch.Tensor] = field(default_factory=dict)
|
| 808 |
-
|
| 809 |
-
|
| 810 |
# Adapted from transformers.models.mamba.modeling_mamba.MambaMixer
|
| 811 |
class JambaMambaMixer(nn.Module):
|
| 812 |
"""
|
|
@@ -838,7 +775,6 @@ class JambaMambaMixer(nn.Module):
|
|
| 838 |
|
| 839 |
self.activation = config.hidden_act
|
| 840 |
self.act = ACT2FN[config.hidden_act]
|
| 841 |
-
self.apply_inner_layernorms = config.mamba_inner_layernorms
|
| 842 |
|
| 843 |
self.use_fast_kernels = config.use_mamba_kernels
|
| 844 |
|
|
@@ -858,14 +794,9 @@ class JambaMambaMixer(nn.Module):
|
|
| 858 |
self.D = nn.Parameter(torch.ones(self.intermediate_size))
|
| 859 |
self.out_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=self.use_bias)
|
| 860 |
|
| 861 |
-
|
| 862 |
-
|
| 863 |
-
|
| 864 |
-
self.C_layernorm = JambaRMSNorm(self.ssm_state_size, eps=config.rms_norm_eps)
|
| 865 |
-
else:
|
| 866 |
-
self.dt_layernorm = None
|
| 867 |
-
self.B_layernorm = None
|
| 868 |
-
self.C_layernorm = None
|
| 869 |
|
| 870 |
if not is_fast_path_available:
|
| 871 |
logger.warning_once(
|
|
@@ -874,145 +805,121 @@ class JambaMambaMixer(nn.Module):
|
|
| 874 |
" https://github.com/Dao-AILab/causal-conv1d. If you want to use the naive implementation, set `use_mamba_kernels=False` in the model config"
|
| 875 |
)
|
| 876 |
|
| 877 |
-
def
|
| 878 |
-
|
| 879 |
-
|
| 880 |
-
|
| 881 |
-
|
| 882 |
-
|
| 883 |
-
|
| 884 |
-
|
| 885 |
-
|
| 886 |
-
|
| 887 |
# 1. Gated MLP's linear projection
|
| 888 |
projected_states = self.in_proj(hidden_states).transpose(1, 2)
|
| 889 |
|
| 890 |
-
if
|
| 891 |
-
|
| 892 |
-
|
| 893 |
-
contextualized_states = mamba_inner_fn(
|
| 894 |
-
projected_states,
|
| 895 |
-
self.conv1d.weight,
|
| 896 |
-
self.conv1d.bias if self.use_conv_bias else None,
|
| 897 |
-
self.x_proj.weight,
|
| 898 |
-
self.dt_proj.weight,
|
| 899 |
-
self.out_proj.weight,
|
| 900 |
-
self.out_proj.bias.float() if self.use_bias else None,
|
| 901 |
-
-torch.exp(self.A_log.float()),
|
| 902 |
-
None, # input-dependent B
|
| 903 |
-
None, # input-dependent C
|
| 904 |
-
self.D.float(),
|
| 905 |
-
delta_bias=self.dt_proj.bias.float(),
|
| 906 |
-
delta_softplus=True,
|
| 907 |
-
)
|
| 908 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 909 |
else:
|
| 910 |
-
|
| 911 |
-
|
| 912 |
-
|
| 913 |
-
|
| 914 |
-
|
| 915 |
-
|
| 916 |
-
|
| 917 |
-
|
| 918 |
-
|
| 919 |
-
|
| 920 |
-
|
| 921 |
-
)
|
| 922 |
-
hidden_states = hidden_states.unsqueeze(-1)
|
| 923 |
-
else:
|
| 924 |
-
if cache_params is not None:
|
| 925 |
-
conv_states = nn.functional.pad(
|
| 926 |
-
hidden_states, (self.conv_kernel_size - hidden_states.shape[-1], 0)
|
| 927 |
-
)
|
| 928 |
-
cache_params.conv_states[self.layer_idx].copy_(conv_states)
|
| 929 |
-
hidden_states = causal_conv1d_fn(
|
| 930 |
-
hidden_states, conv_weights, self.conv1d.bias, activation=self.activation
|
| 931 |
-
)
|
| 932 |
|
| 933 |
-
|
| 934 |
-
|
| 935 |
-
|
| 936 |
-
|
| 937 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 938 |
)
|
| 939 |
-
|
| 940 |
-
|
| 941 |
-
|
| 942 |
-
|
| 943 |
-
|
| 944 |
-
# linear layers, and requires to call the forward pass directly.
|
| 945 |
-
# The original code here was: ```discrete_time_step = self.dt_proj.weight @ time_step.transpose(1, 2)```
|
| 946 |
-
if hasattr(self.dt_proj, "base_layer"):
|
| 947 |
-
# In case of LoRA, we need to access the base layer to get the weight
|
| 948 |
-
time_proj_bias = self.dt_proj.base_layer.bias
|
| 949 |
-
self.dt_proj.base_layer.bias = None
|
| 950 |
-
else:
|
| 951 |
-
time_proj_bias = self.dt_proj.bias
|
| 952 |
-
self.dt_proj.bias = None
|
| 953 |
-
discrete_time_step = self.dt_proj(time_step).transpose(1, 2)
|
| 954 |
-
if hasattr(self.dt_proj, "base_layer"):
|
| 955 |
-
self.dt_proj.base_layer.bias = time_proj_bias
|
| 956 |
-
else:
|
| 957 |
-
self.dt_proj.bias = time_proj_bias
|
| 958 |
-
|
| 959 |
-
A = -torch.exp(self.A_log.float())
|
| 960 |
-
# 3.c perform the recurrence y ← SSM(A, B, C)(x)
|
| 961 |
-
time_proj_bias = time_proj_bias.float() if time_proj_bias is not None else None
|
| 962 |
-
if cache_params is not None and cache_params.seqlen_offset > 0:
|
| 963 |
-
scan_outputs = selective_state_update(
|
| 964 |
-
cache_params.ssm_states[self.layer_idx],
|
| 965 |
-
hidden_states[..., 0],
|
| 966 |
-
discrete_time_step[..., 0],
|
| 967 |
-
A,
|
| 968 |
-
B[:, 0],
|
| 969 |
-
C[:, 0],
|
| 970 |
-
self.D,
|
| 971 |
-
gate[..., 0],
|
| 972 |
-
time_proj_bias,
|
| 973 |
-
dt_softplus=True,
|
| 974 |
-
).unsqueeze(-1)
|
| 975 |
-
else:
|
| 976 |
-
scan_outputs, ssm_state = selective_scan_fn(
|
| 977 |
-
hidden_states,
|
| 978 |
-
discrete_time_step,
|
| 979 |
-
A,
|
| 980 |
-
B.transpose(1, 2),
|
| 981 |
-
C.transpose(1, 2),
|
| 982 |
-
self.D.float(),
|
| 983 |
-
gate,
|
| 984 |
-
time_proj_bias,
|
| 985 |
-
delta_softplus=True,
|
| 986 |
-
return_last_state=True,
|
| 987 |
-
)
|
| 988 |
-
if ssm_state is not None and cache_params is not None:
|
| 989 |
-
cache_params.ssm_states[self.layer_idx].copy_(ssm_state)
|
| 990 |
|
| 991 |
-
# 4. Final linear projection
|
| 992 |
-
contextualized_states = self.out_proj(scan_outputs.transpose(1, 2))
|
| 993 |
return contextualized_states
|
| 994 |
|
| 995 |
# fmt: off
|
| 996 |
-
def slow_forward(self, input_states, cache_params:
|
| 997 |
batch_size, seq_len, _ = input_states.shape
|
| 998 |
dtype = input_states.dtype
|
| 999 |
# 1. Gated MLP's linear projection
|
| 1000 |
projected_states = self.in_proj(input_states).transpose(1, 2) # [batch, 2 * intermediate_size, seq_len]
|
| 1001 |
hidden_states, gate = projected_states.chunk(2, dim=1)
|
| 1002 |
|
|
|
|
| 1003 |
# 2. Convolution sequence transformation
|
| 1004 |
-
if cache_params
|
| 1005 |
if self.training:
|
| 1006 |
# In training mode, we don't want to perform in-place operations on ssm_state so we can compute the backwards pass
|
| 1007 |
ssm_state = cache_params.ssm_states[self.layer_idx].clone()
|
| 1008 |
else:
|
| 1009 |
ssm_state = cache_params.ssm_states[self.layer_idx]
|
| 1010 |
|
| 1011 |
-
if cache_params.
|
|
|
|
| 1012 |
conv_state = cache_params.conv_states[self.layer_idx] # [batch, intermediate_size, conv_kernel_size]
|
| 1013 |
conv_state = torch.roll(conv_state, shifts=-1, dims=-1)
|
| 1014 |
conv_state[:, :, -1] = hidden_states[:, :, 0]
|
| 1015 |
-
cache_params.conv_states[self.layer_idx]
|
| 1016 |
hidden_states = torch.sum(conv_state * self.conv1d.weight[:, 0, :], dim=-1)
|
| 1017 |
if self.use_conv_bias:
|
| 1018 |
hidden_states += self.conv1d.bias
|
|
@@ -1022,7 +929,7 @@ class JambaMambaMixer(nn.Module):
|
|
| 1022 |
hidden_states,
|
| 1023 |
(self.conv_kernel_size - hidden_states.shape[-1], 0)
|
| 1024 |
)
|
| 1025 |
-
cache_params.conv_states[self.layer_idx]
|
| 1026 |
hidden_states = self.act(self.conv1d(hidden_states)[..., :seq_len]) # [batch, intermediate_size, seq_len]
|
| 1027 |
else:
|
| 1028 |
ssm_state = torch.zeros(
|
|
@@ -1037,7 +944,11 @@ class JambaMambaMixer(nn.Module):
|
|
| 1037 |
time_step, B, C = torch.split(
|
| 1038 |
ssm_parameters, [self.time_step_rank, self.ssm_state_size, self.ssm_state_size], dim=-1
|
| 1039 |
)
|
| 1040 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1041 |
discrete_time_step = self.dt_proj(time_step) # [batch, seq_len, intermediate_size]
|
| 1042 |
discrete_time_step = nn.functional.softplus(discrete_time_step).transpose(1, 2) # [batch, intermediate_size, seq_len]
|
| 1043 |
|
|
@@ -1057,15 +968,15 @@ class JambaMambaMixer(nn.Module):
|
|
| 1057 |
scan_output = scan_output + (hidden_states * self.D[None, :, None])
|
| 1058 |
scan_output = (scan_output * self.act(gate))
|
| 1059 |
|
| 1060 |
-
if
|
| 1061 |
-
cache_params.ssm_states[self.layer_idx]
|
| 1062 |
|
| 1063 |
# 4. Final linear projection
|
| 1064 |
contextualized_states = self.out_proj(scan_output.transpose(1, 2)) # [batch, seq_len, hidden_size]
|
| 1065 |
return contextualized_states
|
| 1066 |
# fmt: on
|
| 1067 |
|
| 1068 |
-
def
|
| 1069 |
if self.use_fast_kernels:
|
| 1070 |
if not is_fast_path_available or "cuda" not in self.x_proj.weight.device.type:
|
| 1071 |
raise ValueError(
|
|
@@ -1074,64 +985,17 @@ class JambaMambaMixer(nn.Module):
|
|
| 1074 |
return self.cuda_kernels_forward(hidden_states, cache_params)
|
| 1075 |
return self.slow_forward(hidden_states, cache_params)
|
| 1076 |
|
| 1077 |
-
def forward(
|
| 1078 |
-
self,
|
| 1079 |
-
hidden_states: torch.Tensor,
|
| 1080 |
-
past_key_value: Optional[HybridMambaAttentionDynamicCache] = None,
|
| 1081 |
-
**kwargs,
|
| 1082 |
-
) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor]]]:
|
| 1083 |
-
if past_key_value is not None:
|
| 1084 |
-
cache_params = MambaCacheParams(
|
| 1085 |
-
seqlen_offset=0 if hidden_states.shape[1] > 1 else past_key_value.seen_tokens,
|
| 1086 |
-
)
|
| 1087 |
-
if len(past_key_value.key_cache) > self.layer_idx:
|
| 1088 |
-
# we already have cache for this layer, use it
|
| 1089 |
-
# remove the dummy seqlen dim (dim=2)
|
| 1090 |
-
cache_params.conv_states[self.layer_idx] = past_key_value.key_cache[self.layer_idx].squeeze(2)
|
| 1091 |
-
cache_params.ssm_states[self.layer_idx] = past_key_value.value_cache[self.layer_idx].squeeze(2)
|
| 1092 |
-
else:
|
| 1093 |
-
# we don't have cache for this layer, initialize it with zeros
|
| 1094 |
-
batch_size = hidden_states.shape[0]
|
| 1095 |
-
cache_params.conv_states[self.layer_idx] = torch.zeros(
|
| 1096 |
-
batch_size,
|
| 1097 |
-
self.intermediate_size,
|
| 1098 |
-
self.conv_kernel_size,
|
| 1099 |
-
device=hidden_states.device,
|
| 1100 |
-
dtype=hidden_states.dtype,
|
| 1101 |
-
)
|
| 1102 |
-
cache_params.ssm_states[self.layer_idx] = torch.zeros(
|
| 1103 |
-
batch_size,
|
| 1104 |
-
self.intermediate_size,
|
| 1105 |
-
self.ssm_state_size,
|
| 1106 |
-
device=hidden_states.device,
|
| 1107 |
-
dtype=hidden_states.dtype,
|
| 1108 |
-
)
|
| 1109 |
-
else:
|
| 1110 |
-
cache_params = None
|
| 1111 |
-
|
| 1112 |
-
res = self.mixer_forward(hidden_states, cache_params)
|
| 1113 |
-
|
| 1114 |
-
if past_key_value is not None:
|
| 1115 |
-
past_key_value.update(
|
| 1116 |
-
# add dummy seqlen dim (dim=2) to match the number of dimensions of the attention cache
|
| 1117 |
-
cache_params.conv_states[self.layer_idx].unsqueeze(2),
|
| 1118 |
-
cache_params.ssm_states[self.layer_idx].unsqueeze(2),
|
| 1119 |
-
self.layer_idx,
|
| 1120 |
-
)
|
| 1121 |
-
|
| 1122 |
-
return res, past_key_value
|
| 1123 |
-
|
| 1124 |
|
|
|
|
| 1125 |
class JambaMLP(nn.Module):
|
| 1126 |
-
def __init__(self, config
|
| 1127 |
super().__init__()
|
| 1128 |
-
self.
|
| 1129 |
-
self.
|
| 1130 |
-
|
| 1131 |
-
self.gate_proj = nn.Linear(self.
|
| 1132 |
-
self.
|
| 1133 |
-
self.
|
| 1134 |
-
|
| 1135 |
self.act_fn = ACT2FN[config.hidden_act]
|
| 1136 |
|
| 1137 |
def forward(self, x):
|
|
@@ -1151,39 +1015,20 @@ class JambaSparseMoeBlock(nn.Module):
|
|
| 1151 |
and memory on padding.
|
| 1152 |
"""
|
| 1153 |
|
| 1154 |
-
def __init__(self, config: JambaConfig
|
| 1155 |
super().__init__()
|
| 1156 |
self.hidden_dim = config.hidden_size
|
| 1157 |
self.ffn_dim = config.intermediate_size
|
|
|
|
|
|
|
| 1158 |
|
| 1159 |
-
|
| 1160 |
-
self.num_experts = num_experts
|
| 1161 |
-
self.top_k = num_experts_per_tok
|
| 1162 |
-
|
| 1163 |
-
if num_experts > 1:
|
| 1164 |
-
# expert routing
|
| 1165 |
-
self.router = nn.Linear(self.hidden_dim, self.num_experts, bias=False)
|
| 1166 |
-
else:
|
| 1167 |
-
self.router = None
|
| 1168 |
-
|
| 1169 |
self.experts = nn.ModuleList([JambaMLP(config) for _ in range(self.num_experts)])
|
| 1170 |
|
| 1171 |
def forward(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
|
| 1172 |
""" """
|
| 1173 |
batch_size, sequence_length, hidden_dim = hidden_states.shape
|
| 1174 |
|
| 1175 |
-
if self.num_experts == 1:
|
| 1176 |
-
# in this case we have a single MLP block and don't need to do any routing
|
| 1177 |
-
final_hidden_states = self.experts[0](hidden_states)
|
| 1178 |
-
router_logits = torch.ones(
|
| 1179 |
-
(batch_size * sequence_length, 1),
|
| 1180 |
-
device=hidden_states.device,
|
| 1181 |
-
dtype=hidden_states.dtype,
|
| 1182 |
-
requires_grad=hidden_states.requires_grad,
|
| 1183 |
-
)
|
| 1184 |
-
return final_hidden_states, router_logits
|
| 1185 |
-
|
| 1186 |
-
# in this case we have multiple experts and need to do routing
|
| 1187 |
hidden_states = hidden_states.view(-1, hidden_dim)
|
| 1188 |
# router_logits: (batch * sequence_length, n_experts)
|
| 1189 |
router_logits = self.router(hidden_states)
|
|
@@ -1208,15 +1053,11 @@ class JambaSparseMoeBlock(nn.Module):
|
|
| 1208 |
if top_x.shape[0] == 0:
|
| 1209 |
continue
|
| 1210 |
|
| 1211 |
-
# in torch it is faster to index using lists than torch tensors
|
| 1212 |
-
top_x_list = top_x.tolist()
|
| 1213 |
-
idx_list = idx.tolist()
|
| 1214 |
-
|
| 1215 |
# Index the correct hidden states and compute the expert hidden state for
|
| 1216 |
# the current expert. We need to make sure to multiply the output hidden
|
| 1217 |
# states by `routing_weights` on the corresponding tokens (top-1 and top-2)
|
| 1218 |
-
current_state = hidden_states[None,
|
| 1219 |
-
current_hidden_states = expert_layer(current_state) * routing_weights[
|
| 1220 |
|
| 1221 |
# However `index_add_` only support torch tensors for indexing so we'll use
|
| 1222 |
# the `top_x` tensor here.
|
|
@@ -1226,37 +1067,33 @@ class JambaSparseMoeBlock(nn.Module):
|
|
| 1226 |
|
| 1227 |
|
| 1228 |
class JambaAttentionDecoderLayer(nn.Module):
|
| 1229 |
-
def __init__(self, config: JambaConfig,
|
| 1230 |
super().__init__()
|
| 1231 |
-
|
| 1232 |
self.self_attn = JAMBA_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)
|
| 1233 |
|
| 1234 |
-
|
| 1235 |
-
self.
|
| 1236 |
self.input_layernorm = JambaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 1237 |
-
self.
|
| 1238 |
|
| 1239 |
def forward(
|
| 1240 |
self,
|
| 1241 |
hidden_states: torch.Tensor,
|
| 1242 |
attention_mask: Optional[torch.Tensor] = None,
|
| 1243 |
position_ids: Optional[torch.LongTensor] = None,
|
| 1244 |
-
past_key_value: Optional[
|
| 1245 |
output_attentions: Optional[bool] = False,
|
| 1246 |
output_router_logits: Optional[bool] = False,
|
| 1247 |
use_cache: Optional[bool] = False,
|
| 1248 |
-
|
| 1249 |
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
|
| 1250 |
-
if "padding_mask" in kwargs:
|
| 1251 |
-
warnings.warn(
|
| 1252 |
-
"Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
|
| 1253 |
-
)
|
| 1254 |
"""
|
| 1255 |
Args:
|
| 1256 |
hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
|
| 1257 |
attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
|
| 1258 |
`(batch, sequence_length)` where padding elements are indicated by 0.
|
| 1259 |
-
past_key_value (`
|
| 1260 |
output_attentions (`bool`, *optional*):
|
| 1261 |
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
|
| 1262 |
returned tensors for more detail.
|
|
@@ -1266,6 +1103,8 @@ class JambaAttentionDecoderLayer(nn.Module):
|
|
| 1266 |
use_cache (`bool`, *optional*):
|
| 1267 |
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
|
| 1268 |
(see `past_key_values`).
|
|
|
|
|
|
|
| 1269 |
"""
|
| 1270 |
|
| 1271 |
residual = hidden_states
|
|
@@ -1279,15 +1118,20 @@ class JambaAttentionDecoderLayer(nn.Module):
|
|
| 1279 |
past_key_value=past_key_value,
|
| 1280 |
output_attentions=output_attentions,
|
| 1281 |
use_cache=use_cache,
|
|
|
|
| 1282 |
)
|
| 1283 |
|
| 1284 |
# residual connection after attention
|
| 1285 |
hidden_states = residual + hidden_states
|
| 1286 |
|
| 1287 |
-
#
|
| 1288 |
residual = hidden_states
|
| 1289 |
-
hidden_states = self.
|
| 1290 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1291 |
hidden_states = residual + hidden_states
|
| 1292 |
|
| 1293 |
outputs = (hidden_states,)
|
|
@@ -1305,15 +1149,15 @@ class JambaAttentionDecoderLayer(nn.Module):
|
|
| 1305 |
|
| 1306 |
|
| 1307 |
class JambaMambaDecoderLayer(nn.Module):
|
| 1308 |
-
def __init__(self, config: JambaConfig,
|
| 1309 |
super().__init__()
|
| 1310 |
-
|
| 1311 |
self.mamba = JambaMambaMixer(config=config, layer_idx=layer_idx)
|
| 1312 |
|
| 1313 |
-
|
| 1314 |
-
self.
|
| 1315 |
self.input_layernorm = JambaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 1316 |
-
self.
|
| 1317 |
|
| 1318 |
def forward(
|
| 1319 |
self,
|
|
@@ -1324,18 +1168,14 @@ class JambaMambaDecoderLayer(nn.Module):
|
|
| 1324 |
output_attentions: Optional[bool] = False,
|
| 1325 |
output_router_logits: Optional[bool] = False,
|
| 1326 |
use_cache: Optional[bool] = False,
|
| 1327 |
-
|
| 1328 |
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
|
| 1329 |
-
if "padding_mask" in kwargs:
|
| 1330 |
-
warnings.warn(
|
| 1331 |
-
"Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
|
| 1332 |
-
)
|
| 1333 |
"""
|
| 1334 |
Args:
|
| 1335 |
hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
|
| 1336 |
attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
|
| 1337 |
`(batch, sequence_length)` where padding elements are indicated by 0.
|
| 1338 |
-
past_key_value (`
|
| 1339 |
output_attentions (`bool`, *optional*):
|
| 1340 |
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
|
| 1341 |
returned tensors for more detail.
|
|
@@ -1345,28 +1185,31 @@ class JambaMambaDecoderLayer(nn.Module):
|
|
| 1345 |
use_cache (`bool`, *optional*):
|
| 1346 |
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
|
| 1347 |
(see `past_key_values`).
|
|
|
|
|
|
|
| 1348 |
"""
|
| 1349 |
|
| 1350 |
residual = hidden_states
|
| 1351 |
|
| 1352 |
hidden_states = self.input_layernorm(hidden_states)
|
| 1353 |
|
| 1354 |
-
hidden_states
|
| 1355 |
hidden_states=hidden_states,
|
| 1356 |
-
|
| 1357 |
)
|
| 1358 |
-
|
| 1359 |
-
past_seqlen = self._get_past_seqlen(past_key_value, seqlen)
|
| 1360 |
-
num_attention_heads = self.mamba.config.num_attention_heads
|
| 1361 |
-
self_attn_weights = torch.empty(bs, num_attention_heads, seqlen, past_seqlen, device="meta")
|
| 1362 |
|
| 1363 |
# residual connection after mamba
|
| 1364 |
hidden_states = residual + hidden_states
|
| 1365 |
|
| 1366 |
-
#
|
| 1367 |
residual = hidden_states
|
| 1368 |
-
hidden_states = self.
|
| 1369 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1370 |
hidden_states = residual + hidden_states
|
| 1371 |
|
| 1372 |
outputs = (hidden_states,)
|
|
@@ -1375,25 +1218,13 @@ class JambaMambaDecoderLayer(nn.Module):
|
|
| 1375 |
outputs += (self_attn_weights,)
|
| 1376 |
|
| 1377 |
if use_cache:
|
| 1378 |
-
outputs += (
|
| 1379 |
|
| 1380 |
if output_router_logits:
|
| 1381 |
outputs += (router_logits,)
|
| 1382 |
|
| 1383 |
return outputs
|
| 1384 |
|
| 1385 |
-
def _get_past_seqlen(self, past_key_value, seqlen):
|
| 1386 |
-
if past_key_value is None:
|
| 1387 |
-
return seqlen
|
| 1388 |
-
past_seqlen = past_key_value.get_seq_length()
|
| 1389 |
-
if past_seqlen == 0:
|
| 1390 |
-
return seqlen
|
| 1391 |
-
if past_key_value.attention_layer_idx is None:
|
| 1392 |
-
return seqlen
|
| 1393 |
-
if self.mamba.layer_idx < past_key_value.attention_layer_idx:
|
| 1394 |
-
return past_seqlen + 1
|
| 1395 |
-
return past_seqlen
|
| 1396 |
-
|
| 1397 |
|
| 1398 |
JAMBA_START_DOCSTRING = r"""
|
| 1399 |
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
|
@@ -1416,7 +1247,6 @@ JAMBA_START_DOCSTRING = r"""
|
|
| 1416 |
"The bare Jamba Model outputting raw hidden-states without any specific head on top.",
|
| 1417 |
JAMBA_START_DOCSTRING,
|
| 1418 |
)
|
| 1419 |
-
# Adapted from transformers.models.mistral.modeling_mistral.MistralPreTrainedModel with Mistral->Jamba
|
| 1420 |
class JambaPreTrainedModel(PreTrainedModel):
|
| 1421 |
config_class = JambaConfig
|
| 1422 |
base_model_prefix = "model"
|
|
@@ -1438,42 +1268,6 @@ class JambaPreTrainedModel(PreTrainedModel):
|
|
| 1438 |
if module.padding_idx is not None:
|
| 1439 |
module.weight.data[module.padding_idx].zero_()
|
| 1440 |
|
| 1441 |
-
@staticmethod
|
| 1442 |
-
def _convert_to_standard_cache(
|
| 1443 |
-
past_key_value: Tuple[Tuple[torch.Tensor, torch.Tensor]], batch_size: int
|
| 1444 |
-
) -> Tuple[Tuple[torch.Tensor, torch.Tensor]]:
|
| 1445 |
-
"""
|
| 1446 |
-
Standardizes the format of the cache so as to match most implementations, i.e. have the seqlen as the third dim
|
| 1447 |
-
also for mamba layers
|
| 1448 |
-
"""
|
| 1449 |
-
attn_layer_index = [k.shape == v.shape for k, v in past_key_value].index(True)
|
| 1450 |
-
seqlen = past_key_value[attn_layer_index][0].shape[2]
|
| 1451 |
-
standard_past_key_value = ()
|
| 1452 |
-
for k, v in past_key_value:
|
| 1453 |
-
if k.shape != v.shape:
|
| 1454 |
-
# mamba layer
|
| 1455 |
-
# expand doesn't use more memory, so it's fine to do it here
|
| 1456 |
-
standard_past_key_value += ((k.expand(-1, -1, seqlen, -1), v.expand(-1, -1, seqlen, -1)),)
|
| 1457 |
-
else:
|
| 1458 |
-
standard_past_key_value += ((k, v),)
|
| 1459 |
-
return standard_past_key_value
|
| 1460 |
-
|
| 1461 |
-
@staticmethod
|
| 1462 |
-
def _convert_to_jamba_cache(
|
| 1463 |
-
past_key_value: Tuple[Tuple[torch.Tensor, torch.Tensor]],
|
| 1464 |
-
) -> Tuple[Tuple[torch.Tensor, torch.Tensor]]:
|
| 1465 |
-
"""
|
| 1466 |
-
Converts the cache to the format expected by Jamba, i.e. dummy seqlen dimesion with size 1 for mamba layers
|
| 1467 |
-
"""
|
| 1468 |
-
jamba_past_key_value = ()
|
| 1469 |
-
for k, v in past_key_value:
|
| 1470 |
-
if k.shape != v.shape:
|
| 1471 |
-
# mamba layer
|
| 1472 |
-
jamba_past_key_value += ((k[:, :, :1, :], v[:, :, :1, :]),)
|
| 1473 |
-
else:
|
| 1474 |
-
jamba_past_key_value += ((k, v),)
|
| 1475 |
-
return jamba_past_key_value
|
| 1476 |
-
|
| 1477 |
|
| 1478 |
JAMBA_INPUTS_DOCSTRING = r"""
|
| 1479 |
Args:
|
|
@@ -1510,17 +1304,14 @@ JAMBA_INPUTS_DOCSTRING = r"""
|
|
| 1510 |
config.n_positions - 1]`.
|
| 1511 |
|
| 1512 |
[What are position IDs?](../glossary#position-ids)
|
| 1513 |
-
past_key_values (`
|
| 1514 |
-
|
| 1515 |
-
|
| 1516 |
-
|
| 1517 |
-
|
| 1518 |
-
and
|
| 1519 |
-
|
| 1520 |
-
the
|
| 1521 |
-
|
| 1522 |
-
Contains pre-computed hidden-states (key and values in the self-attention blocks and convolution and
|
| 1523 |
-
ssm states in the mamba blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
|
| 1524 |
|
| 1525 |
If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that
|
| 1526 |
don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
|
|
@@ -1543,8 +1334,14 @@ JAMBA_INPUTS_DOCSTRING = r"""
|
|
| 1543 |
should not be returned during inference.
|
| 1544 |
return_dict (`bool`, *optional*):
|
| 1545 |
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1546 |
"""
|
| 1547 |
|
|
|
|
|
|
|
| 1548 |
|
| 1549 |
@add_start_docstrings(
|
| 1550 |
"The bare Jamba Model outputting raw hidden-states without any specific head on top.",
|
|
@@ -1565,35 +1362,10 @@ class JambaModel(JambaPreTrainedModel):
|
|
| 1565 |
self.vocab_size = config.vocab_size
|
| 1566 |
|
| 1567 |
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
|
| 1568 |
-
|
| 1569 |
-
# init each model layer, decide if it's mamba/attention and has experts or not
|
| 1570 |
decoder_layers = []
|
| 1571 |
for i in range(config.num_hidden_layers):
|
| 1572 |
-
|
| 1573 |
-
|
| 1574 |
-
|
| 1575 |
-
num_experts = self.config.num_experts if is_expert else 1
|
| 1576 |
-
if is_attn:
|
| 1577 |
-
decoder_layers.append(JambaAttentionDecoderLayer(config, num_experts=num_experts, layer_idx=i))
|
| 1578 |
-
else:
|
| 1579 |
-
decoder_layers.append(JambaMambaDecoderLayer(config, num_experts=num_experts, layer_idx=i))
|
| 1580 |
-
|
| 1581 |
-
if not any(isinstance(layer, JambaAttentionDecoderLayer) for layer in decoder_layers):
|
| 1582 |
-
raise ValueError("At least one layer in the decoder must be an attention layer")
|
| 1583 |
-
self._attn_layer_index = [isinstance(layer, JambaAttentionDecoderLayer) for layer in decoder_layers].index(
|
| 1584 |
-
True
|
| 1585 |
-
)
|
| 1586 |
-
|
| 1587 |
-
if not any(isinstance(layer, JambaMambaDecoderLayer) for layer in decoder_layers):
|
| 1588 |
-
raise ValueError("At least one layer in the decoder must be a Mamba layer")
|
| 1589 |
-
self._mamba_layer_index = [isinstance(layer, JambaMambaDecoderLayer) for layer in decoder_layers].index(True)
|
| 1590 |
-
|
| 1591 |
-
if (
|
| 1592 |
-
decoder_layers[self._mamba_layer_index].mamba.ssm_state_size
|
| 1593 |
-
== decoder_layers[self._mamba_layer_index].mamba.conv_kernel_size
|
| 1594 |
-
):
|
| 1595 |
-
raise ValueError("Mamba state size and convolution size must be different")
|
| 1596 |
-
|
| 1597 |
self.layers = nn.ModuleList(decoder_layers)
|
| 1598 |
|
| 1599 |
self._attn_implementation = config._attn_implementation
|
|
@@ -1609,20 +1381,20 @@ class JambaModel(JambaPreTrainedModel):
|
|
| 1609 |
def set_input_embeddings(self, value):
|
| 1610 |
self.embed_tokens = value
|
| 1611 |
|
| 1612 |
-
# Ignore copy
|
| 1613 |
@add_start_docstrings_to_model_forward(JAMBA_INPUTS_DOCSTRING)
|
| 1614 |
def forward(
|
| 1615 |
self,
|
| 1616 |
input_ids: torch.LongTensor = None,
|
| 1617 |
attention_mask: Optional[torch.Tensor] = None,
|
| 1618 |
position_ids: Optional[torch.LongTensor] = None,
|
| 1619 |
-
past_key_values: Optional[
|
| 1620 |
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 1621 |
use_cache: Optional[bool] = None,
|
| 1622 |
output_attentions: Optional[bool] = None,
|
| 1623 |
output_hidden_states: Optional[bool] = None,
|
| 1624 |
output_router_logits: Optional[bool] = None,
|
| 1625 |
return_dict: Optional[bool] = None,
|
|
|
|
| 1626 |
) -> Union[Tuple, MoeModelOutputWithPast]:
|
| 1627 |
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 1628 |
output_router_logits = (
|
|
@@ -1635,85 +1407,37 @@ class JambaModel(JambaPreTrainedModel):
|
|
| 1635 |
|
| 1636 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 1637 |
|
| 1638 |
-
|
| 1639 |
-
|
| 1640 |
-
|
| 1641 |
-
|
| 1642 |
-
batch_size, seq_length = input_ids.shape
|
| 1643 |
-
elif inputs_embeds is not None:
|
| 1644 |
-
batch_size, seq_length, _ = inputs_embeds.shape
|
| 1645 |
-
else:
|
| 1646 |
-
raise ValueError("You have to specify either input_ids or inputs_embeds")
|
| 1647 |
-
|
| 1648 |
-
past_key_values_length = 0
|
| 1649 |
-
|
| 1650 |
-
if self.gradient_checkpointing and self.training:
|
| 1651 |
-
if use_cache:
|
| 1652 |
-
logger.warning_once(
|
| 1653 |
-
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
|
| 1654 |
-
)
|
| 1655 |
-
use_cache = False
|
| 1656 |
-
|
| 1657 |
-
if use_cache:
|
| 1658 |
-
if isinstance(past_key_values, Cache) and not isinstance(
|
| 1659 |
-
past_key_values, HybridMambaAttentionDynamicCache
|
| 1660 |
-
):
|
| 1661 |
-
past_key_values = HybridMambaAttentionDynamicCache.from_legacy_cache(past_key_values.to_legacy_cache())
|
| 1662 |
-
use_legacy_cache = not isinstance(past_key_values, HybridMambaAttentionDynamicCache)
|
| 1663 |
-
if use_legacy_cache:
|
| 1664 |
-
past_key_values = HybridMambaAttentionDynamicCache.from_legacy_cache(past_key_values)
|
| 1665 |
-
past_key_values_length = past_key_values.get_usable_length(seq_length, self._attn_layer_index)
|
| 1666 |
|
| 1667 |
-
if
|
| 1668 |
-
|
| 1669 |
-
|
| 1670 |
-
past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
|
| 1671 |
)
|
| 1672 |
-
|
| 1673 |
-
else:
|
| 1674 |
-
position_ids = position_ids.view(-1, seq_length).long()
|
| 1675 |
|
| 1676 |
if inputs_embeds is None:
|
| 1677 |
inputs_embeds = self.embed_tokens(input_ids)
|
|
|
|
| 1678 |
|
| 1679 |
-
if
|
| 1680 |
-
|
| 1681 |
-
|
| 1682 |
-
|
| 1683 |
-
"You are attempting to perform batched generation with padding_side='right'"
|
| 1684 |
-
" this may lead to unexpected behaviour for Flash Attention version of Jamba. Make sure to "
|
| 1685 |
-
" call `tokenizer.padding_side = 'left'` before tokenizing the input. "
|
| 1686 |
-
)
|
| 1687 |
-
|
| 1688 |
-
if self._attn_implementation == "flash_attention_2":
|
| 1689 |
-
# 2d mask is passed through the layers
|
| 1690 |
-
attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
|
| 1691 |
-
elif self._attn_implementation == "sdpa" and not output_attentions:
|
| 1692 |
-
# output_attentions=True can not be supported when using SDPA, and we fall back on
|
| 1693 |
-
# the manual implementation that requires a 4D causal mask in all cases.
|
| 1694 |
-
attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
|
| 1695 |
-
attention_mask,
|
| 1696 |
-
(batch_size, seq_length),
|
| 1697 |
-
inputs_embeds,
|
| 1698 |
-
past_key_values_length,
|
| 1699 |
-
)
|
| 1700 |
-
else:
|
| 1701 |
-
# 4d mask is passed through the layers
|
| 1702 |
-
attention_mask = _prepare_4d_causal_attention_mask(
|
| 1703 |
-
attention_mask,
|
| 1704 |
-
(batch_size, seq_length),
|
| 1705 |
-
inputs_embeds,
|
| 1706 |
-
past_key_values_length,
|
| 1707 |
-
sliding_window=self.config.sliding_window,
|
| 1708 |
)
|
| 1709 |
|
| 1710 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1711 |
|
| 1712 |
-
# decoder layers
|
| 1713 |
all_hidden_states = () if output_hidden_states else None
|
| 1714 |
all_self_attns = () if output_attentions else None
|
| 1715 |
all_router_logits = () if output_router_logits else None
|
| 1716 |
-
next_decoder_cache = None
|
| 1717 |
|
| 1718 |
for decoder_layer in self.layers:
|
| 1719 |
if output_hidden_states:
|
|
@@ -1723,34 +1447,37 @@ class JambaModel(JambaPreTrainedModel):
|
|
| 1723 |
layer_outputs = self._gradient_checkpointing_func(
|
| 1724 |
decoder_layer.__call__,
|
| 1725 |
hidden_states,
|
| 1726 |
-
|
| 1727 |
position_ids,
|
| 1728 |
past_key_values,
|
| 1729 |
output_attentions,
|
| 1730 |
output_router_logits,
|
| 1731 |
use_cache,
|
|
|
|
| 1732 |
)
|
| 1733 |
else:
|
| 1734 |
layer_outputs = decoder_layer(
|
| 1735 |
hidden_states,
|
| 1736 |
-
attention_mask=
|
| 1737 |
position_ids=position_ids,
|
| 1738 |
past_key_value=past_key_values,
|
| 1739 |
output_attentions=output_attentions,
|
| 1740 |
output_router_logits=output_router_logits,
|
| 1741 |
use_cache=use_cache,
|
|
|
|
| 1742 |
)
|
| 1743 |
|
| 1744 |
hidden_states = layer_outputs[0]
|
| 1745 |
|
| 1746 |
-
if use_cache:
|
| 1747 |
-
next_decoder_cache = layer_outputs[2 if output_attentions else 1]
|
| 1748 |
-
|
| 1749 |
if output_attentions:
|
| 1750 |
-
|
|
|
|
|
|
|
| 1751 |
|
| 1752 |
if output_router_logits:
|
| 1753 |
-
|
|
|
|
|
|
|
| 1754 |
|
| 1755 |
hidden_states = self.final_layernorm(hidden_states)
|
| 1756 |
|
|
@@ -1758,9 +1485,10 @@ class JambaModel(JambaPreTrainedModel):
|
|
| 1758 |
if output_hidden_states:
|
| 1759 |
all_hidden_states += (hidden_states,)
|
| 1760 |
|
| 1761 |
-
|
| 1762 |
-
|
| 1763 |
-
|
|
|
|
| 1764 |
|
| 1765 |
if not return_dict:
|
| 1766 |
return tuple(
|
|
@@ -1776,6 +1504,41 @@ class JambaModel(JambaPreTrainedModel):
|
|
| 1776 |
router_logits=all_router_logits,
|
| 1777 |
)
|
| 1778 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1779 |
|
| 1780 |
# Adapted from transformers.models.mixtral.modeling_mixtral.MixtralForCausalLM with MIXTRAL->JAMBA, Mixtral->Jamba
|
| 1781 |
class JambaForCausalLM(JambaPreTrainedModel):
|
|
@@ -1818,7 +1581,7 @@ class JambaForCausalLM(JambaPreTrainedModel):
|
|
| 1818 |
input_ids: torch.LongTensor = None,
|
| 1819 |
attention_mask: Optional[torch.Tensor] = None,
|
| 1820 |
position_ids: Optional[torch.LongTensor] = None,
|
| 1821 |
-
past_key_values: Optional[
|
| 1822 |
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 1823 |
labels: Optional[torch.LongTensor] = None,
|
| 1824 |
use_cache: Optional[bool] = None,
|
|
@@ -1826,7 +1589,8 @@ class JambaForCausalLM(JambaPreTrainedModel):
|
|
| 1826 |
output_hidden_states: Optional[bool] = None,
|
| 1827 |
output_router_logits: Optional[bool] = None,
|
| 1828 |
return_dict: Optional[bool] = None,
|
| 1829 |
-
|
|
|
|
| 1830 |
) -> Union[Tuple, MoeCausalLMOutputWithPast]:
|
| 1831 |
r"""
|
| 1832 |
Args:
|
|
@@ -1835,12 +1599,28 @@ class JambaForCausalLM(JambaPreTrainedModel):
|
|
| 1835 |
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
|
| 1836 |
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
|
| 1837 |
|
| 1838 |
-
|
| 1839 |
-
|
| 1840 |
-
logits are needed for generation, and calculating them only for that token
|
| 1841 |
-
which becomes pretty significant for long sequences.
|
| 1842 |
|
| 1843 |
Returns:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1844 |
```"""
|
| 1845 |
|
| 1846 |
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
|
@@ -1864,14 +1644,15 @@ class JambaForCausalLM(JambaPreTrainedModel):
|
|
| 1864 |
output_attentions=output_attentions,
|
| 1865 |
output_hidden_states=output_hidden_states,
|
| 1866 |
output_router_logits=output_router_logits,
|
|
|
|
| 1867 |
return_dict=return_dict,
|
| 1868 |
)
|
| 1869 |
|
| 1870 |
hidden_states = outputs[0]
|
| 1871 |
-
if
|
| 1872 |
logits = self.lm_head(hidden_states)
|
| 1873 |
else:
|
| 1874 |
-
logits = self.lm_head(hidden_states[..., -
|
| 1875 |
logits = logits.float()
|
| 1876 |
|
| 1877 |
loss = None
|
|
@@ -1921,27 +1702,15 @@ class JambaForCausalLM(JambaPreTrainedModel):
|
|
| 1921 |
attention_mask=None,
|
| 1922 |
inputs_embeds=None,
|
| 1923 |
output_router_logits=False,
|
|
|
|
| 1924 |
**kwargs,
|
| 1925 |
):
|
| 1926 |
-
|
| 1927 |
-
if past_key_values is not None:
|
| 1928 |
-
# the cache may be in the stardard format (e.g. in contrastive search), convert to Jamba's format if needed
|
| 1929 |
-
if isinstance(past_key_values, Tuple):
|
| 1930 |
-
if past_key_values[self.model._mamba_layer_index][0].shape[2] > 1:
|
| 1931 |
-
past_key_values = self._convert_to_jamba_cache(past_key_values)
|
| 1932 |
-
|
| 1933 |
-
if isinstance(past_key_values, Cache):
|
| 1934 |
-
if not isinstance(past_key_values, HybridMambaAttentionDynamicCache):
|
| 1935 |
-
past_key_values = HybridMambaAttentionDynamicCache.from_legacy_cache(
|
| 1936 |
-
past_key_values.to_legacy_cache()
|
| 1937 |
-
)
|
| 1938 |
-
cache_length = past_key_values.get_seq_length()
|
| 1939 |
-
past_length = past_key_values.seen_tokens
|
| 1940 |
-
max_cache_length = past_key_values.get_max_length()
|
| 1941 |
-
else:
|
| 1942 |
-
cache_length = past_length = past_key_values[self.model._attn_layer_index][0].shape[2]
|
| 1943 |
-
max_cache_length = None
|
| 1944 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1945 |
# Keep only the unprocessed tokens:
|
| 1946 |
# 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
|
| 1947 |
# some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
|
|
@@ -1958,20 +1727,24 @@ class JambaForCausalLM(JambaPreTrainedModel):
|
|
| 1958 |
if (
|
| 1959 |
max_cache_length is not None
|
| 1960 |
and attention_mask is not None
|
| 1961 |
-
and
|
| 1962 |
):
|
| 1963 |
attention_mask = attention_mask[:, -max_cache_length:]
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1964 |
|
| 1965 |
position_ids = kwargs.get("position_ids", None)
|
| 1966 |
if attention_mask is not None and position_ids is None:
|
| 1967 |
# create position_ids on the fly for batch generation
|
| 1968 |
position_ids = attention_mask.long().cumsum(-1) - 1
|
| 1969 |
position_ids.masked_fill_(attention_mask == 0, 1)
|
| 1970 |
-
if
|
| 1971 |
position_ids = position_ids[:, -input_ids.shape[1] :]
|
| 1972 |
|
| 1973 |
# if `inputs_embeds` are passed, we only want to use them in the 1st generation step
|
| 1974 |
-
if inputs_embeds is not None and
|
| 1975 |
model_inputs = {"inputs_embeds": inputs_embeds}
|
| 1976 |
else:
|
| 1977 |
model_inputs = {"input_ids": input_ids}
|
|
@@ -1983,20 +1756,12 @@ class JambaForCausalLM(JambaPreTrainedModel):
|
|
| 1983 |
"use_cache": kwargs.get("use_cache"),
|
| 1984 |
"attention_mask": attention_mask,
|
| 1985 |
"output_router_logits": output_router_logits,
|
| 1986 |
-
"
|
|
|
|
| 1987 |
}
|
| 1988 |
)
|
| 1989 |
return model_inputs
|
| 1990 |
|
| 1991 |
-
@staticmethod
|
| 1992 |
-
def _reorder_cache(past_key_values, beam_idx):
|
| 1993 |
-
reordered_past = ()
|
| 1994 |
-
for layer_past in past_key_values:
|
| 1995 |
-
reordered_past += (
|
| 1996 |
-
tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
|
| 1997 |
-
)
|
| 1998 |
-
return reordered_past
|
| 1999 |
-
|
| 2000 |
|
| 2001 |
@add_start_docstrings(
|
| 2002 |
"""
|
|
|
|
| 20 |
""" PyTorch Jamba model."""
|
| 21 |
import inspect
|
| 22 |
import math
|
|
|
|
|
|
|
| 23 |
from typing import Any, Dict, List, Optional, Tuple, Union
|
| 24 |
|
| 25 |
import torch
|
|
|
|
| 29 |
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
| 30 |
|
| 31 |
from transformers.activations import ACT2FN
|
| 32 |
+
from transformers.cache_utils import DynamicCache # we need __iter__ and __len__ of pkv
|
| 33 |
from transformers.modeling_attn_mask_utils import (
|
| 34 |
+
AttentionMaskConverter,
|
|
|
|
| 35 |
)
|
| 36 |
from transformers.modeling_outputs import (
|
| 37 |
MoeCausalLMOutputWithPast,
|
|
|
|
| 39 |
SequenceClassifierOutputWithPast,
|
| 40 |
)
|
| 41 |
from transformers.modeling_utils import PreTrainedModel
|
|
|
|
| 42 |
from transformers.utils import (
|
| 43 |
add_start_docstrings,
|
| 44 |
add_start_docstrings_to_model_forward,
|
|
|
|
| 46 |
logging,
|
| 47 |
replace_return_docstrings,
|
| 48 |
)
|
| 49 |
+
from transformers.utils.import_utils import (
|
| 50 |
+
is_causal_conv1d_available,
|
| 51 |
+
is_flash_attn_2_available,
|
| 52 |
+
is_mamba_ssm_available,
|
| 53 |
+
)
|
| 54 |
from .configuration_jamba import JambaConfig
|
| 55 |
|
| 56 |
|
| 57 |
+
# try except block so it'll work with trust_remote_code.
|
| 58 |
try:
|
| 59 |
from flash_attn import flash_attn_func, flash_attn_varlen_func
|
| 60 |
from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input # noqa
|
|
|
|
| 63 |
except ImportError:
|
| 64 |
pass
|
| 65 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 66 |
|
| 67 |
+
# try except block so it'll work with trust_remote_code.
|
| 68 |
try:
|
| 69 |
from mamba_ssm.ops.selective_scan_interface import mamba_inner_fn, selective_scan_fn
|
| 70 |
from mamba_ssm.ops.triton.selective_state_update import selective_state_update
|
| 71 |
except ImportError:
|
| 72 |
selective_state_update, selective_scan_fn, mamba_inner_fn = None, None, None
|
| 73 |
|
| 74 |
+
# try except block so it'll work with trust_remote_code.
|
| 75 |
try:
|
| 76 |
from causal_conv1d import causal_conv1d_fn, causal_conv1d_update
|
| 77 |
except ImportError:
|
|
|
|
| 87 |
_CONFIG_FOR_DOC = "JambaConfig"
|
| 88 |
|
| 89 |
|
| 90 |
+
# Copied from transformers.models.mixtral.modeling_mixtral.load_balancing_loss_func with gate->router
|
| 91 |
def load_balancing_loss_func(
|
| 92 |
+
router_logits: torch.Tensor,
|
| 93 |
+
num_experts: torch.Tensor = None,
|
| 94 |
+
top_k=2,
|
| 95 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 96 |
) -> float:
|
| 97 |
r"""
|
| 98 |
Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.
|
|
|
|
| 102 |
experts is too unbalanced.
|
| 103 |
|
| 104 |
Args:
|
| 105 |
+
router_logits (Union[`torch.Tensor`, Tuple[torch.Tensor]):
|
| 106 |
Logits from the `router`, should be a tuple of model.config.num_hidden_layers tensors of
|
| 107 |
shape [batch_size X sequence_length, num_experts].
|
| 108 |
attention_mask (`torch.Tensor`, None):
|
|
|
|
| 114 |
Returns:
|
| 115 |
The auxiliary loss.
|
| 116 |
"""
|
| 117 |
+
if router_logits is None or not isinstance(router_logits, tuple):
|
| 118 |
return 0
|
| 119 |
|
| 120 |
+
if isinstance(router_logits, tuple):
|
| 121 |
+
compute_device = router_logits[0].device
|
| 122 |
+
concatenated_router_logits = torch.cat(
|
| 123 |
+
[layer_router.to(compute_device) for layer_router in router_logits], dim=0
|
| 124 |
)
|
| 125 |
|
| 126 |
+
routing_weights = torch.nn.functional.softmax(concatenated_router_logits, dim=-1)
|
| 127 |
|
| 128 |
_, selected_experts = torch.topk(routing_weights, top_k, dim=-1)
|
| 129 |
|
|
|
|
| 137 |
router_prob_per_expert = torch.mean(routing_weights, dim=0)
|
| 138 |
else:
|
| 139 |
batch_size, sequence_length = attention_mask.shape
|
| 140 |
+
num_hidden_layers = concatenated_router_logits.shape[0] // (batch_size * sequence_length)
|
| 141 |
|
| 142 |
# Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask
|
| 143 |
expert_attention_mask = (
|
|
|
|
| 213 |
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
| 214 |
|
| 215 |
|
| 216 |
+
class HybridMambaAttentionDynamicCache(DynamicCache):
|
| 217 |
+
"""
|
| 218 |
+
A dynamic cache that can handle both the attention cache (which has a seq_len dimension) and the mamba cache
|
| 219 |
+
(which has a constant shape regardless of seq_len).
|
| 220 |
+
|
| 221 |
+
This cache has two sets of lists of tensors: `key_cache` and `value_cache` for attention cache and `conv_states`
|
| 222 |
+
and `ssm_states` for mamba cache. Each of these lists has `num_layers` tensors. The expected shape for each tensor
|
| 223 |
+
For attention layers, `key_cache` and `value_cache` have a shape of `(batch_size, num_heads, seq_len, head_dim)`,
|
| 224 |
+
while `conv_states` and `ssm_states` have a shape of `(batch_size, 0)` (empty tensors).
|
| 225 |
+
For mamba layers, `key_cache` and `value_cache` have a shape of `(batch_size, 0)` (empty tensors),
|
| 226 |
+
while `conv_states` represents the convolution state and has a shape of `(batch_size, d_inner, d_conv)`,
|
| 227 |
+
and `ssm_states` represents the ssm state and has a shape of `(batch_size, d_inner, d_state)`.
|
| 228 |
+
"""
|
| 229 |
+
|
| 230 |
+
def __init__(self, config, batch_size, dtype=torch.float16, device=None):
|
| 231 |
+
self.dtype = dtype
|
| 232 |
+
self.layers_block_type = config.layers_block_type
|
| 233 |
+
self.has_previous_state = False # only used by mamba
|
| 234 |
+
intermediate_size = config.mamba_expand * config.hidden_size
|
| 235 |
+
ssm_state_size = config.mamba_d_state
|
| 236 |
+
conv_kernel_size = config.mamba_d_conv
|
| 237 |
+
self.conv_states = []
|
| 238 |
+
self.ssm_states = []
|
| 239 |
+
for i in range(config.num_hidden_layers):
|
| 240 |
+
if self.layers_block_type[i] == "mamba":
|
| 241 |
+
self.conv_states += [
|
| 242 |
+
torch.zeros(batch_size, intermediate_size, conv_kernel_size, device=device, dtype=dtype)
|
| 243 |
+
]
|
| 244 |
+
self.ssm_states += [
|
| 245 |
+
torch.zeros(batch_size, intermediate_size, ssm_state_size, device=device, dtype=dtype)
|
| 246 |
+
]
|
| 247 |
+
else:
|
| 248 |
+
self.conv_states += [torch.tensor([[]] * batch_size, device=device)]
|
| 249 |
+
self.ssm_states += [torch.tensor([[]] * batch_size, device=device)]
|
| 250 |
+
|
| 251 |
+
self.key_cache = [torch.tensor([[]] * batch_size, device=device) for _ in range(config.num_hidden_layers)]
|
| 252 |
+
self.value_cache = [torch.tensor([[]] * batch_size, device=device) for _ in range(config.num_hidden_layers)]
|
| 253 |
+
|
| 254 |
+
def update(
|
| 255 |
+
self,
|
| 256 |
+
key_states: torch.Tensor,
|
| 257 |
+
value_states: torch.Tensor,
|
| 258 |
+
layer_idx: int,
|
| 259 |
+
cache_kwargs: Optional[Dict[str, Any]] = None,
|
| 260 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
| 261 |
+
# Update the cache
|
| 262 |
+
if self.key_cache[layer_idx].shape[-1] == 0:
|
| 263 |
+
self.key_cache[layer_idx] = key_states
|
| 264 |
+
self.value_cache[layer_idx] = value_states
|
| 265 |
+
else:
|
| 266 |
+
self.key_cache[layer_idx] = torch.cat([self.key_cache[layer_idx], key_states], dim=2)
|
| 267 |
+
self.value_cache[layer_idx] = torch.cat([self.value_cache[layer_idx], value_states], dim=2)
|
| 268 |
+
|
| 269 |
+
return self.key_cache[layer_idx], self.value_cache[layer_idx]
|
| 270 |
+
|
| 271 |
+
def reorder_cache(self, beam_idx: torch.LongTensor):
|
| 272 |
+
"""Reorders the cache for beam search, given the selected beam indices."""
|
| 273 |
+
for layer_idx in range(len(self.key_cache)):
|
| 274 |
+
device = self.key_cache[layer_idx].device
|
| 275 |
+
self.key_cache[layer_idx] = self.key_cache[layer_idx].index_select(0, beam_idx.to(device))
|
| 276 |
+
device = self.value_cache[layer_idx].device
|
| 277 |
+
self.value_cache[layer_idx] = self.value_cache[layer_idx].index_select(0, beam_idx.to(device))
|
| 278 |
+
|
| 279 |
+
device = self.conv_states[layer_idx].device
|
| 280 |
+
self.conv_states[layer_idx] = self.conv_states[layer_idx].index_select(0, beam_idx.to(device))
|
| 281 |
+
device = self.ssm_states[layer_idx].device
|
| 282 |
+
self.ssm_states[layer_idx] = self.ssm_states[layer_idx].index_select(0, beam_idx.to(device))
|
| 283 |
+
|
| 284 |
+
def to_legacy_cache(self) -> Tuple[Tuple[torch.Tensor], Tuple[torch.Tensor]]:
|
| 285 |
+
raise NotImplementedError("HybridMambaAttentionDynamicCache does not have a legacy cache equivalent.")
|
| 286 |
+
|
| 287 |
+
@classmethod
|
| 288 |
+
def from_legacy_cache(cls, past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None) -> "DynamicCache":
|
| 289 |
+
raise NotImplementedError("HybridMambaAttentionDynamicCache does not have a legacy cache equivalent.")
|
| 290 |
+
|
| 291 |
+
|
| 292 |
# Adapted from transformers.models.mistral.modeling_mistral.MistralAttention with Mistral->Jamba
|
| 293 |
class JambaAttention(nn.Module):
|
| 294 |
"""
|
|
|
|
| 325 |
self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
|
| 326 |
self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
|
| 327 |
|
|
|
|
|
|
|
|
|
|
| 328 |
def forward(
|
| 329 |
self,
|
| 330 |
hidden_states: torch.Tensor,
|
| 331 |
attention_mask: Optional[torch.Tensor] = None,
|
| 332 |
position_ids: Optional[torch.LongTensor] = None,
|
| 333 |
+
past_key_value: Optional[HybridMambaAttentionDynamicCache] = None,
|
| 334 |
output_attentions: bool = False,
|
| 335 |
use_cache: bool = False,
|
| 336 |
+
cache_position: Optional[torch.LongTensor] = None,
|
| 337 |
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
|
|
|
|
|
|
|
|
|
|
|
|
| 338 |
bsz, q_len, _ = hidden_states.size()
|
| 339 |
|
| 340 |
query_states = self.q_proj(hidden_states)
|
|
|
|
| 345 |
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 346 |
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 347 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 348 |
if past_key_value is not None:
|
| 349 |
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)
|
| 350 |
|
|
|
|
| 354 |
|
| 355 |
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
|
| 356 |
|
| 357 |
+
if attention_mask is not None: # no matter the length, we just slice it
|
| 358 |
+
causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
|
| 359 |
+
attn_weights = attn_weights + causal_mask
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 360 |
|
| 361 |
# upcast attention to fp32
|
| 362 |
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
|
|
|
|
| 402 |
hidden_states: torch.Tensor,
|
| 403 |
attention_mask: Optional[torch.Tensor] = None,
|
| 404 |
position_ids: Optional[torch.LongTensor] = None,
|
| 405 |
+
past_key_value: Optional[HybridMambaAttentionDynamicCache] = None,
|
| 406 |
output_attentions: bool = False,
|
| 407 |
use_cache: bool = False,
|
| 408 |
+
cache_position: Optional[torch.LongTensor] = None,
|
| 409 |
**kwargs,
|
| 410 |
):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 411 |
bsz, q_len, _ = hidden_states.size()
|
| 412 |
|
| 413 |
query_states = self.q_proj(hidden_states)
|
| 414 |
key_states = self.k_proj(hidden_states)
|
| 415 |
value_states = self.v_proj(hidden_states)
|
| 416 |
|
| 417 |
+
# Flash attention requires the input to have the shape
|
| 418 |
+
# batch_size x seq_length x head_dim x hidden_dim
|
| 419 |
+
# therefore we just need to keep the original shape
|
| 420 |
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 421 |
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 422 |
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 423 |
|
| 424 |
+
kv_seq_len = cache_position[-1]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 425 |
|
| 426 |
use_sliding_windows = (
|
| 427 |
_flash_supports_window_size
|
|
|
|
| 437 |
|
| 438 |
if past_key_value is not None:
|
| 439 |
# Activate slicing cache only if the config has a value `sliding_windows` attribute
|
| 440 |
+
cache_has_contents = cache_position[0] > 0
|
| 441 |
if (
|
| 442 |
getattr(self.config, "sliding_window", None) is not None
|
| 443 |
and kv_seq_len > self.config.sliding_window
|
|
|
|
| 539 |
attention_mask (`torch.Tensor`):
|
| 540 |
The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
|
| 541 |
position of padding tokens and 1 for the position of non-padding tokens.
|
| 542 |
+
dropout (`float`, *optional*):
|
| 543 |
Attention dropout
|
| 544 |
softmax_scale (`float`, *optional*):
|
| 545 |
The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
|
|
|
|
| 614 |
|
| 615 |
return attn_output
|
| 616 |
|
| 617 |
+
# Copied from transformers.models.mixtral.modeling_mixtral.MixtralFlashAttention2._upad_input
|
| 618 |
def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
|
| 619 |
batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape
|
| 620 |
|
|
|
|
| 672 |
hidden_states: torch.Tensor,
|
| 673 |
attention_mask: Optional[torch.Tensor] = None,
|
| 674 |
position_ids: Optional[torch.LongTensor] = None,
|
| 675 |
+
past_key_value: Optional[HybridMambaAttentionDynamicCache] = None,
|
| 676 |
output_attentions: bool = False,
|
| 677 |
use_cache: bool = False,
|
| 678 |
+
cache_position: Optional[torch.LongTensor] = None,
|
| 679 |
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
| 680 |
if output_attentions:
|
| 681 |
# TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
|
|
|
|
| 702 |
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 703 |
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 704 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 705 |
if past_key_value is not None:
|
| 706 |
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)
|
| 707 |
|
| 708 |
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
| 709 |
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
| 710 |
|
| 711 |
+
causal_mask = attention_mask
|
| 712 |
if attention_mask is not None:
|
| 713 |
+
causal_mask = causal_mask[:, :, :, : key_states.shape[-2]]
|
|
|
|
|
|
|
|
|
|
| 714 |
|
| 715 |
# SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
|
| 716 |
# Reference: https://github.com/pytorch/pytorch/issues/112577.
|
|
|
|
| 723 |
query_states,
|
| 724 |
key_states,
|
| 725 |
value_states,
|
| 726 |
+
attn_mask=causal_mask,
|
| 727 |
dropout_p=self.attention_dropout if self.training else 0.0,
|
| 728 |
# 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.
|
| 729 |
is_causal=self.is_causal and attention_mask is None and q_len > 1,
|
|
|
|
| 744 |
}
|
| 745 |
|
| 746 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 747 |
# Adapted from transformers.models.mamba.modeling_mamba.MambaMixer
|
| 748 |
class JambaMambaMixer(nn.Module):
|
| 749 |
"""
|
|
|
|
| 775 |
|
| 776 |
self.activation = config.hidden_act
|
| 777 |
self.act = ACT2FN[config.hidden_act]
|
|
|
|
| 778 |
|
| 779 |
self.use_fast_kernels = config.use_mamba_kernels
|
| 780 |
|
|
|
|
| 794 |
self.D = nn.Parameter(torch.ones(self.intermediate_size))
|
| 795 |
self.out_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=self.use_bias)
|
| 796 |
|
| 797 |
+
self.dt_layernorm = JambaRMSNorm(self.time_step_rank, eps=config.rms_norm_eps)
|
| 798 |
+
self.b_layernorm = JambaRMSNorm(self.ssm_state_size, eps=config.rms_norm_eps)
|
| 799 |
+
self.c_layernorm = JambaRMSNorm(self.ssm_state_size, eps=config.rms_norm_eps)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 800 |
|
| 801 |
if not is_fast_path_available:
|
| 802 |
logger.warning_once(
|
|
|
|
| 805 |
" https://github.com/Dao-AILab/causal-conv1d. If you want to use the naive implementation, set `use_mamba_kernels=False` in the model config"
|
| 806 |
)
|
| 807 |
|
| 808 |
+
def cuda_kernels_forward(self, hidden_states: torch.Tensor, cache_params: HybridMambaAttentionDynamicCache = None):
|
| 809 |
+
batch_size, seq_len, _ = hidden_states.shape
|
| 810 |
+
use_precomputed_states = (
|
| 811 |
+
cache_params is not None
|
| 812 |
+
and cache_params.has_previous_state
|
| 813 |
+
and seq_len == 1
|
| 814 |
+
and cache_params.conv_states[self.layer_idx].shape[0]
|
| 815 |
+
== cache_params.ssm_states[self.layer_idx].shape[0]
|
| 816 |
+
== batch_size
|
| 817 |
+
)
|
| 818 |
# 1. Gated MLP's linear projection
|
| 819 |
projected_states = self.in_proj(hidden_states).transpose(1, 2)
|
| 820 |
|
| 821 |
+
# We can't use `mamba_inner_fn` even if in training and without cache params because we have the
|
| 822 |
+
# inner layernorms which isn't supported by this fused kernel
|
| 823 |
+
hidden_states, gate = projected_states.chunk(2, dim=1)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 824 |
|
| 825 |
+
# 2. Convolution sequence transformation
|
| 826 |
+
conv_weights = self.conv1d.weight.view(self.conv1d.weight.size(0), self.conv1d.weight.size(2))
|
| 827 |
+
if use_precomputed_states:
|
| 828 |
+
hidden_states = causal_conv1d_update(
|
| 829 |
+
hidden_states.squeeze(-1),
|
| 830 |
+
cache_params.conv_states[self.layer_idx],
|
| 831 |
+
conv_weights,
|
| 832 |
+
self.conv1d.bias,
|
| 833 |
+
self.activation,
|
| 834 |
+
)
|
| 835 |
+
hidden_states = hidden_states.unsqueeze(-1)
|
| 836 |
else:
|
| 837 |
+
if cache_params is not None:
|
| 838 |
+
conv_states = nn.functional.pad(hidden_states, (self.conv_kernel_size - hidden_states.shape[-1], 0))
|
| 839 |
+
cache_params.conv_states[self.layer_idx].copy_(conv_states)
|
| 840 |
+
hidden_states = causal_conv1d_fn(hidden_states, conv_weights, self.conv1d.bias, activation=self.activation)
|
| 841 |
+
|
| 842 |
+
# 3. State Space Model sequence transformation
|
| 843 |
+
# 3.a. input varying initialization of time_step, B and C
|
| 844 |
+
ssm_parameters = self.x_proj(hidden_states.transpose(1, 2))
|
| 845 |
+
time_step, B, C = torch.split(
|
| 846 |
+
ssm_parameters, [self.time_step_rank, self.ssm_state_size, self.ssm_state_size], dim=-1
|
| 847 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 848 |
|
| 849 |
+
time_step = self.dt_layernorm(time_step)
|
| 850 |
+
B = self.b_layernorm(B)
|
| 851 |
+
C = self.c_layernorm(C)
|
| 852 |
+
|
| 853 |
+
# Here we need to apply dt_proj without the bias, as the bias is added in the selective scan kernel.
|
| 854 |
+
# This is a hack to apply dt_proj while still using the forward pass of `torch.nn.Linear`, which is needed
|
| 855 |
+
# in order to make quantization work. Quantization code replaces `torch.nn.Linear` layers with quantized
|
| 856 |
+
# linear layers, and requires to call the forward pass directly.
|
| 857 |
+
# The original code here was: ```discrete_time_step = self.dt_proj.weight @ time_step.transpose(1, 2)```
|
| 858 |
+
time_proj_bias = self.dt_proj.bias
|
| 859 |
+
self.dt_proj.bias = None
|
| 860 |
+
discrete_time_step = self.dt_proj(time_step).transpose(1, 2)
|
| 861 |
+
self.dt_proj.bias = time_proj_bias
|
| 862 |
+
|
| 863 |
+
A = -torch.exp(self.A_log.float())
|
| 864 |
+
# 3.c perform the recurrence y ← SSM(A, B, C)(x)
|
| 865 |
+
time_proj_bias = time_proj_bias.float() if time_proj_bias is not None else None
|
| 866 |
+
if use_precomputed_states:
|
| 867 |
+
scan_outputs = selective_state_update(
|
| 868 |
+
cache_params.ssm_states[self.layer_idx],
|
| 869 |
+
hidden_states[..., 0],
|
| 870 |
+
discrete_time_step[..., 0],
|
| 871 |
+
A,
|
| 872 |
+
B[:, 0],
|
| 873 |
+
C[:, 0],
|
| 874 |
+
self.D,
|
| 875 |
+
gate[..., 0],
|
| 876 |
+
time_proj_bias,
|
| 877 |
+
dt_softplus=True,
|
| 878 |
+
).unsqueeze(-1)
|
| 879 |
+
else:
|
| 880 |
+
scan_outputs, ssm_state = selective_scan_fn(
|
| 881 |
+
hidden_states,
|
| 882 |
+
discrete_time_step,
|
| 883 |
+
A,
|
| 884 |
+
B.transpose(1, 2),
|
| 885 |
+
C.transpose(1, 2),
|
| 886 |
+
self.D.float(),
|
| 887 |
+
gate,
|
| 888 |
+
time_proj_bias,
|
| 889 |
+
delta_softplus=True,
|
| 890 |
+
return_last_state=True,
|
| 891 |
)
|
| 892 |
+
if ssm_state is not None and cache_params is not None:
|
| 893 |
+
cache_params.ssm_states[self.layer_idx].copy_(ssm_state)
|
| 894 |
+
|
| 895 |
+
# 4. Final linear projection
|
| 896 |
+
contextualized_states = self.out_proj(scan_outputs.transpose(1, 2))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 897 |
|
|
|
|
|
|
|
| 898 |
return contextualized_states
|
| 899 |
|
| 900 |
# fmt: off
|
| 901 |
+
def slow_forward(self, input_states, cache_params: HybridMambaAttentionDynamicCache = None):
|
| 902 |
batch_size, seq_len, _ = input_states.shape
|
| 903 |
dtype = input_states.dtype
|
| 904 |
# 1. Gated MLP's linear projection
|
| 905 |
projected_states = self.in_proj(input_states).transpose(1, 2) # [batch, 2 * intermediate_size, seq_len]
|
| 906 |
hidden_states, gate = projected_states.chunk(2, dim=1)
|
| 907 |
|
| 908 |
+
use_cache = isinstance(cache_params,HybridMambaAttentionDynamicCache)
|
| 909 |
# 2. Convolution sequence transformation
|
| 910 |
+
if use_cache and cache_params.ssm_states[self.layer_idx].shape[0] == batch_size:
|
| 911 |
if self.training:
|
| 912 |
# In training mode, we don't want to perform in-place operations on ssm_state so we can compute the backwards pass
|
| 913 |
ssm_state = cache_params.ssm_states[self.layer_idx].clone()
|
| 914 |
else:
|
| 915 |
ssm_state = cache_params.ssm_states[self.layer_idx]
|
| 916 |
|
| 917 |
+
if cache_params.has_previous_state and seq_len == 1 and \
|
| 918 |
+
cache_params.conv_states[self.layer_idx].shape[0] == batch_size:
|
| 919 |
conv_state = cache_params.conv_states[self.layer_idx] # [batch, intermediate_size, conv_kernel_size]
|
| 920 |
conv_state = torch.roll(conv_state, shifts=-1, dims=-1)
|
| 921 |
conv_state[:, :, -1] = hidden_states[:, :, 0]
|
| 922 |
+
cache_params.conv_states[self.layer_idx] = conv_state
|
| 923 |
hidden_states = torch.sum(conv_state * self.conv1d.weight[:, 0, :], dim=-1)
|
| 924 |
if self.use_conv_bias:
|
| 925 |
hidden_states += self.conv1d.bias
|
|
|
|
| 929 |
hidden_states,
|
| 930 |
(self.conv_kernel_size - hidden_states.shape[-1], 0)
|
| 931 |
)
|
| 932 |
+
cache_params.conv_states[self.layer_idx] = conv_state
|
| 933 |
hidden_states = self.act(self.conv1d(hidden_states)[..., :seq_len]) # [batch, intermediate_size, seq_len]
|
| 934 |
else:
|
| 935 |
ssm_state = torch.zeros(
|
|
|
|
| 944 |
time_step, B, C = torch.split(
|
| 945 |
ssm_parameters, [self.time_step_rank, self.ssm_state_size, self.ssm_state_size], dim=-1
|
| 946 |
)
|
| 947 |
+
|
| 948 |
+
time_step = self.dt_layernorm(time_step)
|
| 949 |
+
B = self.b_layernorm(B)
|
| 950 |
+
C = self.c_layernorm(C)
|
| 951 |
+
|
| 952 |
discrete_time_step = self.dt_proj(time_step) # [batch, seq_len, intermediate_size]
|
| 953 |
discrete_time_step = nn.functional.softplus(discrete_time_step).transpose(1, 2) # [batch, intermediate_size, seq_len]
|
| 954 |
|
|
|
|
| 968 |
scan_output = scan_output + (hidden_states * self.D[None, :, None])
|
| 969 |
scan_output = (scan_output * self.act(gate))
|
| 970 |
|
| 971 |
+
if use_cache:
|
| 972 |
+
cache_params.ssm_states[self.layer_idx] = ssm_state
|
| 973 |
|
| 974 |
# 4. Final linear projection
|
| 975 |
contextualized_states = self.out_proj(scan_output.transpose(1, 2)) # [batch, seq_len, hidden_size]
|
| 976 |
return contextualized_states
|
| 977 |
# fmt: on
|
| 978 |
|
| 979 |
+
def forward(self, hidden_states, cache_params: HybridMambaAttentionDynamicCache = None):
|
| 980 |
if self.use_fast_kernels:
|
| 981 |
if not is_fast_path_available or "cuda" not in self.x_proj.weight.device.type:
|
| 982 |
raise ValueError(
|
|
|
|
| 985 |
return self.cuda_kernels_forward(hidden_states, cache_params)
|
| 986 |
return self.slow_forward(hidden_states, cache_params)
|
| 987 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 988 |
|
| 989 |
+
# Copied from transformers.models.mistral.modeling_mistral.MistralMLP with Mistral->Jamba
|
| 990 |
class JambaMLP(nn.Module):
|
| 991 |
+
def __init__(self, config):
|
| 992 |
super().__init__()
|
| 993 |
+
self.config = config
|
| 994 |
+
self.hidden_size = config.hidden_size
|
| 995 |
+
self.intermediate_size = config.intermediate_size
|
| 996 |
+
self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
|
| 997 |
+
self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
|
| 998 |
+
self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
|
|
|
|
| 999 |
self.act_fn = ACT2FN[config.hidden_act]
|
| 1000 |
|
| 1001 |
def forward(self, x):
|
|
|
|
| 1015 |
and memory on padding.
|
| 1016 |
"""
|
| 1017 |
|
| 1018 |
+
def __init__(self, config: JambaConfig):
|
| 1019 |
super().__init__()
|
| 1020 |
self.hidden_dim = config.hidden_size
|
| 1021 |
self.ffn_dim = config.intermediate_size
|
| 1022 |
+
self.num_experts = config.num_experts
|
| 1023 |
+
self.top_k = config.num_experts_per_tok
|
| 1024 |
|
| 1025 |
+
self.router = nn.Linear(self.hidden_dim, self.num_experts, bias=False)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1026 |
self.experts = nn.ModuleList([JambaMLP(config) for _ in range(self.num_experts)])
|
| 1027 |
|
| 1028 |
def forward(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
|
| 1029 |
""" """
|
| 1030 |
batch_size, sequence_length, hidden_dim = hidden_states.shape
|
| 1031 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1032 |
hidden_states = hidden_states.view(-1, hidden_dim)
|
| 1033 |
# router_logits: (batch * sequence_length, n_experts)
|
| 1034 |
router_logits = self.router(hidden_states)
|
|
|
|
| 1053 |
if top_x.shape[0] == 0:
|
| 1054 |
continue
|
| 1055 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1056 |
# Index the correct hidden states and compute the expert hidden state for
|
| 1057 |
# the current expert. We need to make sure to multiply the output hidden
|
| 1058 |
# states by `routing_weights` on the corresponding tokens (top-1 and top-2)
|
| 1059 |
+
current_state = hidden_states[None, top_x].reshape(-1, hidden_dim)
|
| 1060 |
+
current_hidden_states = expert_layer(current_state) * routing_weights[top_x, idx, None]
|
| 1061 |
|
| 1062 |
# However `index_add_` only support torch tensors for indexing so we'll use
|
| 1063 |
# the `top_x` tensor here.
|
|
|
|
| 1067 |
|
| 1068 |
|
| 1069 |
class JambaAttentionDecoderLayer(nn.Module):
|
| 1070 |
+
def __init__(self, config: JambaConfig, layer_idx: int):
|
| 1071 |
super().__init__()
|
| 1072 |
+
num_experts = config.layers_num_experts[layer_idx]
|
| 1073 |
self.self_attn = JAMBA_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)
|
| 1074 |
|
| 1075 |
+
ffn_layer_class = JambaSparseMoeBlock if num_experts > 1 else JambaMLP
|
| 1076 |
+
self.feed_forward = ffn_layer_class(config)
|
| 1077 |
self.input_layernorm = JambaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 1078 |
+
self.pre_ff_layernorm = JambaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 1079 |
|
| 1080 |
def forward(
|
| 1081 |
self,
|
| 1082 |
hidden_states: torch.Tensor,
|
| 1083 |
attention_mask: Optional[torch.Tensor] = None,
|
| 1084 |
position_ids: Optional[torch.LongTensor] = None,
|
| 1085 |
+
past_key_value: Optional[HybridMambaAttentionDynamicCache] = None,
|
| 1086 |
output_attentions: Optional[bool] = False,
|
| 1087 |
output_router_logits: Optional[bool] = False,
|
| 1088 |
use_cache: Optional[bool] = False,
|
| 1089 |
+
cache_position: Optional[torch.LongTensor] = None,
|
| 1090 |
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1091 |
"""
|
| 1092 |
Args:
|
| 1093 |
hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
|
| 1094 |
attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
|
| 1095 |
`(batch, sequence_length)` where padding elements are indicated by 0.
|
| 1096 |
+
past_key_value (`HybridMambaAttentionDynamicCache`, *optional*): cached past key and value projection states
|
| 1097 |
output_attentions (`bool`, *optional*):
|
| 1098 |
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
|
| 1099 |
returned tensors for more detail.
|
|
|
|
| 1103 |
use_cache (`bool`, *optional*):
|
| 1104 |
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
|
| 1105 |
(see `past_key_values`).
|
| 1106 |
+
cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
|
| 1107 |
+
Indices depicting the position of the input sequence tokens in the sequence.
|
| 1108 |
"""
|
| 1109 |
|
| 1110 |
residual = hidden_states
|
|
|
|
| 1118 |
past_key_value=past_key_value,
|
| 1119 |
output_attentions=output_attentions,
|
| 1120 |
use_cache=use_cache,
|
| 1121 |
+
cache_position=cache_position,
|
| 1122 |
)
|
| 1123 |
|
| 1124 |
# residual connection after attention
|
| 1125 |
hidden_states = residual + hidden_states
|
| 1126 |
|
| 1127 |
+
# feed-forward (experts/MLP)
|
| 1128 |
residual = hidden_states
|
| 1129 |
+
hidden_states = self.pre_ff_layernorm(hidden_states)
|
| 1130 |
+
ff_outputs = self.feed_forward(hidden_states)
|
| 1131 |
+
if isinstance(ff_outputs, tuple):
|
| 1132 |
+
hidden_states, router_logits = ff_outputs
|
| 1133 |
+
else:
|
| 1134 |
+
hidden_states, router_logits = ff_outputs, None
|
| 1135 |
hidden_states = residual + hidden_states
|
| 1136 |
|
| 1137 |
outputs = (hidden_states,)
|
|
|
|
| 1149 |
|
| 1150 |
|
| 1151 |
class JambaMambaDecoderLayer(nn.Module):
|
| 1152 |
+
def __init__(self, config: JambaConfig, layer_idx: int):
|
| 1153 |
super().__init__()
|
| 1154 |
+
num_experts = config.layers_num_experts[layer_idx]
|
| 1155 |
self.mamba = JambaMambaMixer(config=config, layer_idx=layer_idx)
|
| 1156 |
|
| 1157 |
+
ffn_layer_class = JambaSparseMoeBlock if num_experts > 1 else JambaMLP
|
| 1158 |
+
self.feed_forward = ffn_layer_class(config)
|
| 1159 |
self.input_layernorm = JambaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 1160 |
+
self.pre_ff_layernorm = JambaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 1161 |
|
| 1162 |
def forward(
|
| 1163 |
self,
|
|
|
|
| 1168 |
output_attentions: Optional[bool] = False,
|
| 1169 |
output_router_logits: Optional[bool] = False,
|
| 1170 |
use_cache: Optional[bool] = False,
|
| 1171 |
+
cache_position: Optional[torch.LongTensor] = None,
|
| 1172 |
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1173 |
"""
|
| 1174 |
Args:
|
| 1175 |
hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
|
| 1176 |
attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
|
| 1177 |
`(batch, sequence_length)` where padding elements are indicated by 0.
|
| 1178 |
+
past_key_value (`HybridMambaAttentionDynamicCache`, *optional*): cached past key and value projection states
|
| 1179 |
output_attentions (`bool`, *optional*):
|
| 1180 |
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
|
| 1181 |
returned tensors for more detail.
|
|
|
|
| 1185 |
use_cache (`bool`, *optional*):
|
| 1186 |
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
|
| 1187 |
(see `past_key_values`).
|
| 1188 |
+
cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
|
| 1189 |
+
Indices depicting the position of the input sequence tokens in the sequence.
|
| 1190 |
"""
|
| 1191 |
|
| 1192 |
residual = hidden_states
|
| 1193 |
|
| 1194 |
hidden_states = self.input_layernorm(hidden_states)
|
| 1195 |
|
| 1196 |
+
hidden_states = self.mamba(
|
| 1197 |
hidden_states=hidden_states,
|
| 1198 |
+
cache_params=past_key_value,
|
| 1199 |
)
|
| 1200 |
+
self_attn_weights = None
|
|
|
|
|
|
|
|
|
|
| 1201 |
|
| 1202 |
# residual connection after mamba
|
| 1203 |
hidden_states = residual + hidden_states
|
| 1204 |
|
| 1205 |
+
# feed-forward (experts/MLP)
|
| 1206 |
residual = hidden_states
|
| 1207 |
+
hidden_states = self.pre_ff_layernorm(hidden_states)
|
| 1208 |
+
ff_outputs = self.feed_forward(hidden_states)
|
| 1209 |
+
if isinstance(ff_outputs, tuple):
|
| 1210 |
+
hidden_states, router_logits = ff_outputs
|
| 1211 |
+
else:
|
| 1212 |
+
hidden_states, router_logits = ff_outputs, None
|
| 1213 |
hidden_states = residual + hidden_states
|
| 1214 |
|
| 1215 |
outputs = (hidden_states,)
|
|
|
|
| 1218 |
outputs += (self_attn_weights,)
|
| 1219 |
|
| 1220 |
if use_cache:
|
| 1221 |
+
outputs += (past_key_value,)
|
| 1222 |
|
| 1223 |
if output_router_logits:
|
| 1224 |
outputs += (router_logits,)
|
| 1225 |
|
| 1226 |
return outputs
|
| 1227 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1228 |
|
| 1229 |
JAMBA_START_DOCSTRING = r"""
|
| 1230 |
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
|
|
|
| 1247 |
"The bare Jamba Model outputting raw hidden-states without any specific head on top.",
|
| 1248 |
JAMBA_START_DOCSTRING,
|
| 1249 |
)
|
|
|
|
| 1250 |
class JambaPreTrainedModel(PreTrainedModel):
|
| 1251 |
config_class = JambaConfig
|
| 1252 |
base_model_prefix = "model"
|
|
|
|
| 1268 |
if module.padding_idx is not None:
|
| 1269 |
module.weight.data[module.padding_idx].zero_()
|
| 1270 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1271 |
|
| 1272 |
JAMBA_INPUTS_DOCSTRING = r"""
|
| 1273 |
Args:
|
|
|
|
| 1304 |
config.n_positions - 1]`.
|
| 1305 |
|
| 1306 |
[What are position IDs?](../glossary#position-ids)
|
| 1307 |
+
past_key_values (`HybridMambaAttentionDynamicCache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
|
| 1308 |
+
A HybridMambaAttentionDynamicCache object containing pre-computed hidden-states (keys and values in the
|
| 1309 |
+
self-attention blocks and convolution and ssm states in the mamba blocks) that can be used (see
|
| 1310 |
+
`past_key_values` input) to speed up sequential decoding.
|
| 1311 |
+
Key and value cache tensors have shape `(batch_size, num_heads, seq_len, head_dim)`.
|
| 1312 |
+
Convolution and ssm states tensors have shape `(batch_size, d_inner, d_conv)` and
|
| 1313 |
+
`(batch_size, d_inner, d_state)` respectively.
|
| 1314 |
+
See the `HybridMambaAttentionDynamicCache` class for more details.
|
|
|
|
|
|
|
|
|
|
| 1315 |
|
| 1316 |
If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that
|
| 1317 |
don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
|
|
|
|
| 1334 |
should not be returned during inference.
|
| 1335 |
return_dict (`bool`, *optional*):
|
| 1336 |
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
|
| 1337 |
+
cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
|
| 1338 |
+
Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
|
| 1339 |
+
this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
|
| 1340 |
+
the complete sequence length.
|
| 1341 |
"""
|
| 1342 |
|
| 1343 |
+
ALL_DECODER_LAYER_TYPES = {"attention": JambaAttentionDecoderLayer, "mamba": JambaMambaDecoderLayer}
|
| 1344 |
+
|
| 1345 |
|
| 1346 |
@add_start_docstrings(
|
| 1347 |
"The bare Jamba Model outputting raw hidden-states without any specific head on top.",
|
|
|
|
| 1362 |
self.vocab_size = config.vocab_size
|
| 1363 |
|
| 1364 |
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
|
|
|
|
|
|
|
| 1365 |
decoder_layers = []
|
| 1366 |
for i in range(config.num_hidden_layers):
|
| 1367 |
+
layer_class = ALL_DECODER_LAYER_TYPES[config.layers_block_type[i]]
|
| 1368 |
+
decoder_layers.append(layer_class(config, layer_idx=i))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1369 |
self.layers = nn.ModuleList(decoder_layers)
|
| 1370 |
|
| 1371 |
self._attn_implementation = config._attn_implementation
|
|
|
|
| 1381 |
def set_input_embeddings(self, value):
|
| 1382 |
self.embed_tokens = value
|
| 1383 |
|
|
|
|
| 1384 |
@add_start_docstrings_to_model_forward(JAMBA_INPUTS_DOCSTRING)
|
| 1385 |
def forward(
|
| 1386 |
self,
|
| 1387 |
input_ids: torch.LongTensor = None,
|
| 1388 |
attention_mask: Optional[torch.Tensor] = None,
|
| 1389 |
position_ids: Optional[torch.LongTensor] = None,
|
| 1390 |
+
past_key_values: Optional[HybridMambaAttentionDynamicCache] = None,
|
| 1391 |
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 1392 |
use_cache: Optional[bool] = None,
|
| 1393 |
output_attentions: Optional[bool] = None,
|
| 1394 |
output_hidden_states: Optional[bool] = None,
|
| 1395 |
output_router_logits: Optional[bool] = None,
|
| 1396 |
return_dict: Optional[bool] = None,
|
| 1397 |
+
cache_position: Optional[torch.LongTensor] = None,
|
| 1398 |
) -> Union[Tuple, MoeModelOutputWithPast]:
|
| 1399 |
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 1400 |
output_router_logits = (
|
|
|
|
| 1407 |
|
| 1408 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 1409 |
|
| 1410 |
+
if (input_ids is None) ^ (inputs_embeds is not None):
|
| 1411 |
+
raise ValueError(
|
| 1412 |
+
"You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
|
| 1413 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1414 |
|
| 1415 |
+
if self.gradient_checkpointing and self.training and use_cache:
|
| 1416 |
+
logger.warning_once(
|
| 1417 |
+
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
|
|
|
|
| 1418 |
)
|
| 1419 |
+
use_cache = False
|
|
|
|
|
|
|
| 1420 |
|
| 1421 |
if inputs_embeds is None:
|
| 1422 |
inputs_embeds = self.embed_tokens(input_ids)
|
| 1423 |
+
hidden_states = inputs_embeds
|
| 1424 |
|
| 1425 |
+
if use_cache and past_key_values is None:
|
| 1426 |
+
logger.warning_once(
|
| 1427 |
+
"Jamba requires an initialized `HybridMambaAttentionDynamicCache` to return a cache. None was "
|
| 1428 |
+
"provided, so no cache will be returned."
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1429 |
)
|
| 1430 |
|
| 1431 |
+
if cache_position is None:
|
| 1432 |
+
cache_position = torch.arange(hidden_states.shape[1], device=hidden_states.device)
|
| 1433 |
+
if position_ids is None:
|
| 1434 |
+
position_ids = cache_position.unsqueeze(0)
|
| 1435 |
+
|
| 1436 |
+
causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position)
|
| 1437 |
|
|
|
|
| 1438 |
all_hidden_states = () if output_hidden_states else None
|
| 1439 |
all_self_attns = () if output_attentions else None
|
| 1440 |
all_router_logits = () if output_router_logits else None
|
|
|
|
| 1441 |
|
| 1442 |
for decoder_layer in self.layers:
|
| 1443 |
if output_hidden_states:
|
|
|
|
| 1447 |
layer_outputs = self._gradient_checkpointing_func(
|
| 1448 |
decoder_layer.__call__,
|
| 1449 |
hidden_states,
|
| 1450 |
+
causal_mask,
|
| 1451 |
position_ids,
|
| 1452 |
past_key_values,
|
| 1453 |
output_attentions,
|
| 1454 |
output_router_logits,
|
| 1455 |
use_cache,
|
| 1456 |
+
cache_position,
|
| 1457 |
)
|
| 1458 |
else:
|
| 1459 |
layer_outputs = decoder_layer(
|
| 1460 |
hidden_states,
|
| 1461 |
+
attention_mask=causal_mask,
|
| 1462 |
position_ids=position_ids,
|
| 1463 |
past_key_value=past_key_values,
|
| 1464 |
output_attentions=output_attentions,
|
| 1465 |
output_router_logits=output_router_logits,
|
| 1466 |
use_cache=use_cache,
|
| 1467 |
+
cache_position=cache_position,
|
| 1468 |
)
|
| 1469 |
|
| 1470 |
hidden_states = layer_outputs[0]
|
| 1471 |
|
|
|
|
|
|
|
|
|
|
| 1472 |
if output_attentions:
|
| 1473 |
+
if layer_outputs[1] is not None:
|
| 1474 |
+
# append attentions only of attention layers. Mamba layers return `None` as the attention weights
|
| 1475 |
+
all_self_attns += (layer_outputs[1],)
|
| 1476 |
|
| 1477 |
if output_router_logits:
|
| 1478 |
+
if layer_outputs[-1] is not None:
|
| 1479 |
+
# append router logits only of expert layers. Regular MLP layers return `None` as the router logits
|
| 1480 |
+
all_router_logits += (layer_outputs[-1],)
|
| 1481 |
|
| 1482 |
hidden_states = self.final_layernorm(hidden_states)
|
| 1483 |
|
|
|
|
| 1485 |
if output_hidden_states:
|
| 1486 |
all_hidden_states += (hidden_states,)
|
| 1487 |
|
| 1488 |
+
if past_key_values and not past_key_values.has_previous_state:
|
| 1489 |
+
past_key_values.has_previous_state = True
|
| 1490 |
+
|
| 1491 |
+
next_cache = None if not use_cache else past_key_values
|
| 1492 |
|
| 1493 |
if not return_dict:
|
| 1494 |
return tuple(
|
|
|
|
| 1504 |
router_logits=all_router_logits,
|
| 1505 |
)
|
| 1506 |
|
| 1507 |
+
def _update_causal_mask(self, attention_mask, input_tensor, cache_position):
|
| 1508 |
+
if self.config._attn_implementation == "flash_attention_2":
|
| 1509 |
+
if attention_mask is not None and 0.0 in attention_mask:
|
| 1510 |
+
return attention_mask
|
| 1511 |
+
return None
|
| 1512 |
+
|
| 1513 |
+
dtype, device = input_tensor.dtype, input_tensor.device
|
| 1514 |
+
min_dtype = torch.finfo(dtype).min
|
| 1515 |
+
sequence_length = input_tensor.shape[1]
|
| 1516 |
+
target_length = cache_position[-1] + 1
|
| 1517 |
+
|
| 1518 |
+
causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
|
| 1519 |
+
if sequence_length != 1:
|
| 1520 |
+
causal_mask = torch.triu(causal_mask, diagonal=1)
|
| 1521 |
+
causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
|
| 1522 |
+
causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
|
| 1523 |
+
if attention_mask is not None:
|
| 1524 |
+
causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit
|
| 1525 |
+
if attention_mask.dim() == 2:
|
| 1526 |
+
mask_length = attention_mask.shape[-1]
|
| 1527 |
+
padding_mask = causal_mask[..., :mask_length].eq(0.0) * attention_mask[:, None, None, :].eq(0.0)
|
| 1528 |
+
causal_mask[..., :mask_length] = causal_mask[..., :mask_length].masked_fill(padding_mask, min_dtype)
|
| 1529 |
+
|
| 1530 |
+
if (
|
| 1531 |
+
self.config._attn_implementation == "sdpa"
|
| 1532 |
+
and attention_mask is not None
|
| 1533 |
+
and attention_mask.device.type == "cuda"
|
| 1534 |
+
):
|
| 1535 |
+
# Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
|
| 1536 |
+
# using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
|
| 1537 |
+
# Details: https://github.com/pytorch/pytorch/issues/110213
|
| 1538 |
+
causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
|
| 1539 |
+
|
| 1540 |
+
return causal_mask
|
| 1541 |
+
|
| 1542 |
|
| 1543 |
# Adapted from transformers.models.mixtral.modeling_mixtral.MixtralForCausalLM with MIXTRAL->JAMBA, Mixtral->Jamba
|
| 1544 |
class JambaForCausalLM(JambaPreTrainedModel):
|
|
|
|
| 1581 |
input_ids: torch.LongTensor = None,
|
| 1582 |
attention_mask: Optional[torch.Tensor] = None,
|
| 1583 |
position_ids: Optional[torch.LongTensor] = None,
|
| 1584 |
+
past_key_values: Optional[HybridMambaAttentionDynamicCache] = None,
|
| 1585 |
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 1586 |
labels: Optional[torch.LongTensor] = None,
|
| 1587 |
use_cache: Optional[bool] = None,
|
|
|
|
| 1589 |
output_hidden_states: Optional[bool] = None,
|
| 1590 |
output_router_logits: Optional[bool] = None,
|
| 1591 |
return_dict: Optional[bool] = None,
|
| 1592 |
+
cache_position: Optional[torch.LongTensor] = None,
|
| 1593 |
+
num_logits_to_keep: Optional[Union[int, None]] = None,
|
| 1594 |
) -> Union[Tuple, MoeCausalLMOutputWithPast]:
|
| 1595 |
r"""
|
| 1596 |
Args:
|
|
|
|
| 1599 |
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
|
| 1600 |
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
|
| 1601 |
|
| 1602 |
+
num_logits_to_keep (`int` or `None`, *optional*):
|
| 1603 |
+
Calculate logits for the last `num_logits_to_keep` tokens. If `None`, calculate logits for all
|
| 1604 |
+
`input_ids`. Only last token logits are needed for generation, and calculating them only for that token
|
| 1605 |
+
can save memory, which becomes pretty significant for long sequences.
|
| 1606 |
|
| 1607 |
Returns:
|
| 1608 |
+
|
| 1609 |
+
Example:
|
| 1610 |
+
|
| 1611 |
+
```python
|
| 1612 |
+
>>> from transformers import AutoTokenizer, JambaForCausalLM
|
| 1613 |
+
|
| 1614 |
+
>>> model = JambaForCausalLM.from_pretrained("ai21labs/Jamba-v0.1")
|
| 1615 |
+
>>> tokenizer = AutoTokenizer.from_pretrained("ai21labs/Jamba-v0.1")
|
| 1616 |
+
|
| 1617 |
+
>>> prompt = "Hey, are you conscious? Can you talk to me?"
|
| 1618 |
+
>>> inputs = tokenizer(prompt, return_tensors="pt")
|
| 1619 |
+
|
| 1620 |
+
>>> # Generate
|
| 1621 |
+
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
|
| 1622 |
+
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
| 1623 |
+
"Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
|
| 1624 |
```"""
|
| 1625 |
|
| 1626 |
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
|
|
|
| 1644 |
output_attentions=output_attentions,
|
| 1645 |
output_hidden_states=output_hidden_states,
|
| 1646 |
output_router_logits=output_router_logits,
|
| 1647 |
+
cache_position=cache_position,
|
| 1648 |
return_dict=return_dict,
|
| 1649 |
)
|
| 1650 |
|
| 1651 |
hidden_states = outputs[0]
|
| 1652 |
+
if num_logits_to_keep is None:
|
| 1653 |
logits = self.lm_head(hidden_states)
|
| 1654 |
else:
|
| 1655 |
+
logits = self.lm_head(hidden_states[..., -num_logits_to_keep:, :])
|
| 1656 |
logits = logits.float()
|
| 1657 |
|
| 1658 |
loss = None
|
|
|
|
| 1702 |
attention_mask=None,
|
| 1703 |
inputs_embeds=None,
|
| 1704 |
output_router_logits=False,
|
| 1705 |
+
cache_position=None,
|
| 1706 |
**kwargs,
|
| 1707 |
):
|
| 1708 |
+
empty_past_kv = past_key_values is None
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1709 |
|
| 1710 |
+
# Omit tokens covered by past_key_values
|
| 1711 |
+
if not empty_past_kv:
|
| 1712 |
+
past_length = cache_position[0] if cache_position is not None else attention_mask.shape[1]
|
| 1713 |
+
max_cache_length = self.config.sliding_window
|
| 1714 |
# Keep only the unprocessed tokens:
|
| 1715 |
# 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
|
| 1716 |
# some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
|
|
|
|
| 1727 |
if (
|
| 1728 |
max_cache_length is not None
|
| 1729 |
and attention_mask is not None
|
| 1730 |
+
and past_length + input_ids.shape[1] > max_cache_length
|
| 1731 |
):
|
| 1732 |
attention_mask = attention_mask[:, -max_cache_length:]
|
| 1733 |
+
else:
|
| 1734 |
+
past_key_values = HybridMambaAttentionDynamicCache(
|
| 1735 |
+
self.config, input_ids.shape[0], self.dtype, device=self.device
|
| 1736 |
+
)
|
| 1737 |
|
| 1738 |
position_ids = kwargs.get("position_ids", None)
|
| 1739 |
if attention_mask is not None and position_ids is None:
|
| 1740 |
# create position_ids on the fly for batch generation
|
| 1741 |
position_ids = attention_mask.long().cumsum(-1) - 1
|
| 1742 |
position_ids.masked_fill_(attention_mask == 0, 1)
|
| 1743 |
+
if not empty_past_kv:
|
| 1744 |
position_ids = position_ids[:, -input_ids.shape[1] :]
|
| 1745 |
|
| 1746 |
# if `inputs_embeds` are passed, we only want to use them in the 1st generation step
|
| 1747 |
+
if inputs_embeds is not None and empty_past_kv:
|
| 1748 |
model_inputs = {"inputs_embeds": inputs_embeds}
|
| 1749 |
else:
|
| 1750 |
model_inputs = {"input_ids": input_ids}
|
|
|
|
| 1756 |
"use_cache": kwargs.get("use_cache"),
|
| 1757 |
"attention_mask": attention_mask,
|
| 1758 |
"output_router_logits": output_router_logits,
|
| 1759 |
+
"num_logits_to_keep": self.config.num_logits_to_keep,
|
| 1760 |
+
"cache_position": cache_position,
|
| 1761 |
}
|
| 1762 |
)
|
| 1763 |
return model_inputs
|
| 1764 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1765 |
|
| 1766 |
@add_start_docstrings(
|
| 1767 |
"""
|
special_tokens_map.json
CHANGED
|
@@ -1,6 +1,30 @@
|
|
| 1 |
{
|
| 2 |
-
"bos_token":
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 6 |
}
|
|
|
|
| 1 |
{
|
| 2 |
+
"bos_token": {
|
| 3 |
+
"content": "<|startoftext|>",
|
| 4 |
+
"lstrip": false,
|
| 5 |
+
"normalized": false,
|
| 6 |
+
"rstrip": false,
|
| 7 |
+
"single_word": false
|
| 8 |
+
},
|
| 9 |
+
"eos_token": {
|
| 10 |
+
"content": "<|endoftext|>",
|
| 11 |
+
"lstrip": false,
|
| 12 |
+
"normalized": false,
|
| 13 |
+
"rstrip": false,
|
| 14 |
+
"single_word": false
|
| 15 |
+
},
|
| 16 |
+
"pad_token": {
|
| 17 |
+
"content": "<|pad|>",
|
| 18 |
+
"lstrip": false,
|
| 19 |
+
"normalized": false,
|
| 20 |
+
"rstrip": false,
|
| 21 |
+
"single_word": false
|
| 22 |
+
},
|
| 23 |
+
"unk_token": {
|
| 24 |
+
"content": "<|unk|>",
|
| 25 |
+
"lstrip": false,
|
| 26 |
+
"normalized": false,
|
| 27 |
+
"rstrip": false,
|
| 28 |
+
"single_word": false
|
| 29 |
+
}
|
| 30 |
}
|