Commit
·
132c5b7
1
Parent(s):
81ed588
Saving HF Model -- Step 0
Browse files- config.json +22 -0
- model.safetensors +3 -0
- pico_decoder.py +615 -0
config.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"activation_hidden_dim": 6144,
|
3 |
+
"architectures": [
|
4 |
+
"PicoHF"
|
5 |
+
],
|
6 |
+
"attention_n_heads": 12,
|
7 |
+
"attention_n_kv_heads": 4,
|
8 |
+
"auto_map": {
|
9 |
+
"AutoConfig": "pico.PicoHFConfig",
|
10 |
+
"AutoModelForCausalLM": "pico.PicoHF"
|
11 |
+
},
|
12 |
+
"batch_size": 1024,
|
13 |
+
"d_model": 1536,
|
14 |
+
"max_seq_len": 2048,
|
15 |
+
"model_type": "pico",
|
16 |
+
"n_layers": 12,
|
17 |
+
"norm_eps": 1e-06,
|
18 |
+
"position_emb_theta": 10000.0,
|
19 |
+
"torch_dtype": "float32",
|
20 |
+
"transformers_version": "4.48.1",
|
21 |
+
"vocab_size": 50304
|
22 |
+
}
|
model.safetensors
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:5533b7fc9ddfb2f170cff03059e363880a255b613e1d379c2cca8659e10c512b
|
3 |
+
size 2279246680
|
pico_decoder.py
ADDED
@@ -0,0 +1,615 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
Pico Decoder: A Lightweight Causal Transformer Language Model
|
3 |
+
|
4 |
+
Pico Decoder uses a simple LLAMA-style transformer architecture, written for clarity and educational purposes.
|
5 |
+
|
6 |
+
Everything is written with a modular design for easy modification and experimentation.
|
7 |
+
|
8 |
+
Key features:
|
9 |
+
- RMSNorm for layer normalization
|
10 |
+
- Rotary Positional Embeddings (RoPE)
|
11 |
+
- Multi-head attention with KV-cache support
|
12 |
+
- SwiGLU activation function
|
13 |
+
- Residual connections throughout
|
14 |
+
|
15 |
+
- KV-cache for faster autoregressive generation
|
16 |
+
|
17 |
+
References:
|
18 |
+
- RoPE: https://arxiv.org/abs/2104.09864
|
19 |
+
- SwiGLU: https://arxiv.org/abs/2002.05202
|
20 |
+
- LLAMA: https://arxiv.org/abs/2302.13971
|
21 |
+
|
22 |
+
Adapted from:
|
23 |
+
- OLMO: https://github.com/allenai/OLMo
|
24 |
+
- LLAMA: https://github.com/meta/llama
|
25 |
+
"""
|
26 |
+
|
27 |
+
import torch
|
28 |
+
import torch.nn as nn
|
29 |
+
import torch.nn.functional as F
|
30 |
+
from torch.nn.attention import sdpa_kernel, SDPBackend
|
31 |
+
|
32 |
+
from dataclasses import asdict
|
33 |
+
|
34 |
+
from transformers import PretrainedConfig, PreTrainedModel
|
35 |
+
from transformers.modeling_outputs import CausalLMOutputWithPast, CausalLMOutput
|
36 |
+
|
37 |
+
# typing imports
|
38 |
+
from typing import Union, Tuple, Optional, TYPE_CHECKING, Dict, Any
|
39 |
+
|
40 |
+
try:
|
41 |
+
if TYPE_CHECKING:
|
42 |
+
# We need to do this to avoid importing these when creating the HF-compatible models
|
43 |
+
from src.config import ModelConfig
|
44 |
+
except ImportError:
|
45 |
+
pass
|
46 |
+
|
47 |
+
########################################################
|
48 |
+
#
|
49 |
+
# Layer Normalization
|
50 |
+
#
|
51 |
+
########################################################
|
52 |
+
|
53 |
+
|
54 |
+
class RMSNorm(torch.nn.Module):
|
55 |
+
"""Root Mean Square Layer Normalization.
|
56 |
+
|
57 |
+
A variant of Layer Normalization that uses RMS statistics instead of mean/variance,
|
58 |
+
resulting in improved stability and performance.
|
59 |
+
|
60 |
+
Args:
|
61 |
+
config (Union[ModelConfig, PicoHFConfig]): Configuration object containing normalization parameters
|
62 |
+
- config.norm_eps: Small constant for numerical stability
|
63 |
+
- config.d_model: Model dimension for the weight parameter
|
64 |
+
|
65 |
+
References:
|
66 |
+
https://arxiv.org/abs/1910.07467
|
67 |
+
"""
|
68 |
+
|
69 |
+
def __init__(self, config: Union["ModelConfig", "PicoDecoderHFConfig"]):
|
70 |
+
super().__init__()
|
71 |
+
self.eps = config.norm_eps
|
72 |
+
self.weight = nn.Parameter(torch.ones(config.d_model))
|
73 |
+
|
74 |
+
def _norm(self, x: torch.Tensor) -> torch.Tensor:
|
75 |
+
"""
|
76 |
+
Normalizes the input tensor by its RMS value.
|
77 |
+
"""
|
78 |
+
return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
|
79 |
+
|
80 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
81 |
+
"""
|
82 |
+
Applies RMS normalization to the input tensor and scales it by the weight parameter.
|
83 |
+
"""
|
84 |
+
output = self._norm(x.float()).type_as(x)
|
85 |
+
return output * self.weight
|
86 |
+
|
87 |
+
|
88 |
+
########################################################
|
89 |
+
#
|
90 |
+
# Positional Embedding
|
91 |
+
#
|
92 |
+
########################################################
|
93 |
+
|
94 |
+
|
95 |
+
class RoPE(nn.Module):
|
96 |
+
"""Rotary Positional Embeddings (RoPE).
|
97 |
+
|
98 |
+
Implements position-dependent rotation of keys and queries in attention mechanism,
|
99 |
+
allowing better modeling of relative positions in sequences. Uses complex number
|
100 |
+
operations for efficient rotation.
|
101 |
+
|
102 |
+
Args:
|
103 |
+
config (Union[ModelConfig, PicoHFConfig]): Model configuration containing:
|
104 |
+
- config.position_emb_theta: Base for frequency computation
|
105 |
+
- config.d_model: Model dimension
|
106 |
+
- config.attention_n_heads: Number of attention heads
|
107 |
+
- config.max_seq_len: Maximum sequence length
|
108 |
+
|
109 |
+
References:
|
110 |
+
https://arxiv.org/abs/2104.09864
|
111 |
+
"""
|
112 |
+
|
113 |
+
_freqs_cis_tensor: torch.Tensor | None = None
|
114 |
+
|
115 |
+
def __init__(self, config: Union["ModelConfig", "PicoDecoderHFConfig"]):
|
116 |
+
super().__init__()
|
117 |
+
|
118 |
+
self.theta = config.position_emb_theta
|
119 |
+
self.dim = config.d_model // config.attention_n_heads
|
120 |
+
|
121 |
+
max_seq_len = config.max_seq_len
|
122 |
+
|
123 |
+
# only gets set once, and then reused for all RoPE instances
|
124 |
+
if RoPE._freqs_cis_tensor is None:
|
125 |
+
RoPE._freqs_cis_tensor = self._setup_freqs_cis(
|
126 |
+
max_seq_len, self.theta, self.dim
|
127 |
+
)
|
128 |
+
|
129 |
+
# register _freqs_cis buffer
|
130 |
+
# can be easily recomputed so persistent=False
|
131 |
+
self.register_buffer("_freqs_cis", self._freqs_cis_tensor, persistent=False)
|
132 |
+
|
133 |
+
@classmethod
|
134 |
+
def _setup_freqs_cis(cls, seq_len: int, theta: float, dim: int) -> torch.Tensor:
|
135 |
+
"""Setup Frequency Tensor for RoPE Embeddings
|
136 |
+
|
137 |
+
Initializes the complex frequency tensor that is used to compute the RoPE embeddings.
|
138 |
+
|
139 |
+
Note other implementations will use cos and sin directly, but using the complex
|
140 |
+
number representation is (probably?) more efficient:
|
141 |
+
|
142 |
+
e^(theta * i * t) = cos(theta * t) + i * sin(theta * t) [Euler's formula]
|
143 |
+
"""
|
144 |
+
_freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
|
145 |
+
positions = torch.arange(seq_len)
|
146 |
+
freqs = torch.outer(positions, _freqs)
|
147 |
+
return torch.polar(torch.ones_like(freqs), freqs) # complex64
|
148 |
+
|
149 |
+
def get_freqs_cis(
|
150 |
+
self, input_shape: torch.Size, start_pos: int, end_pos: int
|
151 |
+
) -> torch.Tensor:
|
152 |
+
"""Reshape Frequency Tensor for RoPE Embeddings
|
153 |
+
|
154 |
+
Makes the frequency tensor broadcastable with the input tensor.
|
155 |
+
"""
|
156 |
+
_freqs_cis = self._freqs_cis[start_pos:end_pos]
|
157 |
+
ndim = len(input_shape)
|
158 |
+
assert 0 <= 1 < ndim
|
159 |
+
assert _freqs_cis.shape == (input_shape[1], input_shape[-1])
|
160 |
+
|
161 |
+
# TODO: Check whether this is correct (might be able to remove this)
|
162 |
+
shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(input_shape)]
|
163 |
+
return _freqs_cis.view(*shape)
|
164 |
+
|
165 |
+
def forward(
|
166 |
+
self,
|
167 |
+
queries: torch.Tensor,
|
168 |
+
keys: torch.Tensor,
|
169 |
+
start_pos: int = 0,
|
170 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
171 |
+
"""Apply RoPE Embeddings to Queries and Keys
|
172 |
+
|
173 |
+
Applies the rotary positional embeddings to the input tensors via complex num multiplication
|
174 |
+
|
175 |
+
NOTE: The start_pos is used if we want to use the kv_cache in the attention mechanism.
|
176 |
+
"""
|
177 |
+
queries_ = torch.view_as_complex(
|
178 |
+
queries.float().reshape(*queries.shape[:-1], -1, 2)
|
179 |
+
)
|
180 |
+
keys_ = torch.view_as_complex(keys.float().reshape(*keys.shape[:-1], -1, 2))
|
181 |
+
|
182 |
+
input_shape = (
|
183 |
+
queries_.shape
|
184 |
+
) # same as keys: (batch_size, seq_len, n_heads, head_dim/2)
|
185 |
+
freqs_start_pos = start_pos
|
186 |
+
freqs_end_pos = freqs_start_pos + queries_.shape[1]
|
187 |
+
|
188 |
+
freqs_cis = self.get_freqs_cis(input_shape, freqs_start_pos, freqs_end_pos)
|
189 |
+
|
190 |
+
queries_rotated = torch.view_as_real(queries_ * freqs_cis).flatten(3)
|
191 |
+
keys_rotated = torch.view_as_real(keys_ * freqs_cis).flatten(3)
|
192 |
+
return queries_rotated.type_as(queries), keys_rotated.type_as(keys)
|
193 |
+
|
194 |
+
|
195 |
+
########################################################
|
196 |
+
#
|
197 |
+
# Attention
|
198 |
+
#
|
199 |
+
########################################################
|
200 |
+
|
201 |
+
|
202 |
+
class Attention(nn.Module):
|
203 |
+
"""Multi-head Attention with Group Query Attention support.
|
204 |
+
|
205 |
+
Implements scaled dot-product attention and supports:
|
206 |
+
- Grouped Query Attention (GQA)
|
207 |
+
- Key-Value caching for efficient inference
|
208 |
+
- RoPE integration
|
209 |
+
|
210 |
+
Args:
|
211 |
+
config (Union[ModelConfig, PretrainedConfig]): Configuration containing:
|
212 |
+
- config.attention_n_heads: Number of attention heads
|
213 |
+
- config.attention_n_kv_heads: Number of key/value heads
|
214 |
+
- config.d_model: Model dimension
|
215 |
+
- config.batch_size: Maximum batch size
|
216 |
+
- config.max_seq_len: Maximum sequence length
|
217 |
+
|
218 |
+
Shape:
|
219 |
+
- Input: (batch_size, seq_len, d_model)
|
220 |
+
- Output: (batch_size, seq_len, d_model)
|
221 |
+
"""
|
222 |
+
|
223 |
+
def __init__(
|
224 |
+
self,
|
225 |
+
config: Union["ModelConfig", "PicoDecoderHFConfig"],
|
226 |
+
):
|
227 |
+
super().__init__()
|
228 |
+
|
229 |
+
self.n_heads = config.attention_n_heads
|
230 |
+
self.n_kv_heads = config.attention_n_kv_heads
|
231 |
+
|
232 |
+
self.batch_size = config.batch_size
|
233 |
+
self.max_seq_len = config.max_seq_len
|
234 |
+
|
235 |
+
d_model = config.d_model
|
236 |
+
self.head_dim = d_model // self.n_heads
|
237 |
+
|
238 |
+
self.n_rep = self.n_heads // self.n_kv_heads
|
239 |
+
|
240 |
+
self.q_proj = nn.Linear(d_model, self.n_heads * self.head_dim, bias=False)
|
241 |
+
self.k_proj = nn.Linear(d_model, self.n_kv_heads * self.head_dim, bias=False)
|
242 |
+
self.v_proj = nn.Linear(d_model, self.n_kv_heads * self.head_dim, bias=False)
|
243 |
+
self.o_proj = nn.Linear(self.n_heads * self.head_dim, d_model, bias=False)
|
244 |
+
|
245 |
+
self.rope = RoPE(config)
|
246 |
+
|
247 |
+
def forward(
|
248 |
+
self,
|
249 |
+
input: torch.Tensor,
|
250 |
+
mask: Optional[torch.Tensor] = None,
|
251 |
+
past_key_values: Optional[Tuple[torch.Tensor, ...]] = None,
|
252 |
+
use_cache: bool = False,
|
253 |
+
) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor, torch.Tensor]]]:
|
254 |
+
"""Forward pass for the attention mechanism.
|
255 |
+
|
256 |
+
Computes queries, keys, and values for the attention mechanism. Applies rotary positional
|
257 |
+
embeddings to the queries and keys, and then computes attention scores and outputs.
|
258 |
+
|
259 |
+
For an introduction to the attention mechanism, see:
|
260 |
+
https://arxiv.org/abs/1706.03762
|
261 |
+
|
262 |
+
A few things to note:
|
263 |
+
- The past_key_values is used to implement the KV cache, which is used to speed up
|
264 |
+
generation by caching the KV pairs from previous forward passes. This is useful when doing
|
265 |
+
tasks that require generating multiple tokens conditioned on previous tokens (e.g. language
|
266 |
+
modeling, text generation, etc.). The way the KV cache is implemented is that each layer has
|
267 |
+
its own KV cache - this KV cache is implemented as a tuple.
|
268 |
+
"""
|
269 |
+
bsz, seq_len, _ = input.shape
|
270 |
+
_queries, _keys, _values = (
|
271 |
+
self.q_proj(input),
|
272 |
+
self.k_proj(input),
|
273 |
+
self.v_proj(input),
|
274 |
+
)
|
275 |
+
|
276 |
+
# Reshaping for multi-head attention
|
277 |
+
queries = _queries.view(bsz, seq_len, self.n_heads, self.head_dim)
|
278 |
+
keys = _keys.view(bsz, seq_len, self.n_kv_heads, self.head_dim)
|
279 |
+
values = _values.view(bsz, seq_len, self.n_kv_heads, self.head_dim)
|
280 |
+
|
281 |
+
# The start position is used to apply the RoPE embeddings to only the new tokens
|
282 |
+
# when using the kv_cache in the attention mechanism.
|
283 |
+
# We want to start from the last position in the cache.
|
284 |
+
start_pos = past_key_values[0].shape[1] if past_key_values is not None else 0
|
285 |
+
|
286 |
+
# apply rotary positional embeddings
|
287 |
+
queries, keys = self.rope(queries, keys, start_pos)
|
288 |
+
|
289 |
+
if past_key_values is not None:
|
290 |
+
keys = torch.cat([past_key_values[0], keys], dim=1)
|
291 |
+
values = torch.cat([past_key_values[1], values], dim=1)
|
292 |
+
|
293 |
+
if use_cache:
|
294 |
+
cached_keys = keys
|
295 |
+
cached_values = values
|
296 |
+
else:
|
297 |
+
cached_keys = None
|
298 |
+
cached_values = None
|
299 |
+
|
300 |
+
queries = queries.transpose(1, 2)
|
301 |
+
keys = keys.transpose(1, 2)
|
302 |
+
values = values.transpose(1, 2)
|
303 |
+
|
304 |
+
apply_gqa = self.n_rep > 1
|
305 |
+
if apply_gqa and queries.device.type == "mps":
|
306 |
+
# NOTE: MPS does not support GQA in the SDPA kernel, but we can repeat the keys and values
|
307 |
+
# outside of the kernel to get the same effect.
|
308 |
+
# See: https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html
|
309 |
+
keys = keys.repeat_interleave(self.n_rep, dim=-3)
|
310 |
+
values = values.repeat_interleave(self.n_rep, dim=-3)
|
311 |
+
apply_gqa = False
|
312 |
+
|
313 |
+
backends = [SDPBackend.CUDNN_ATTENTION, SDPBackend.MATH]
|
314 |
+
|
315 |
+
with sdpa_kernel(backends=backends):
|
316 |
+
attn_output = F.scaled_dot_product_attention(
|
317 |
+
queries.contiguous(),
|
318 |
+
keys.contiguous(),
|
319 |
+
values.contiguous(),
|
320 |
+
attn_mask=mask.to(queries.dtype),
|
321 |
+
enable_gqa=apply_gqa,
|
322 |
+
)
|
323 |
+
|
324 |
+
attn_output = attn_output.transpose(1, 2).contiguous().view(bsz, seq_len, -1)
|
325 |
+
output = self.o_proj(attn_output)
|
326 |
+
|
327 |
+
return output, (cached_keys, cached_values)
|
328 |
+
|
329 |
+
|
330 |
+
########################################################
|
331 |
+
#
|
332 |
+
# SwiGLU (Combines MLP and Activation)
|
333 |
+
#
|
334 |
+
########################################################
|
335 |
+
|
336 |
+
|
337 |
+
class SwiGLU(nn.Module):
|
338 |
+
"""SwiGLU Activation Function with Linear Projections.
|
339 |
+
|
340 |
+
Implements the SwiGLU activation function combined with linear transformations,
|
341 |
+
serving as the feed-forward network in transformer blocks.
|
342 |
+
|
343 |
+
Args:
|
344 |
+
config (Union[ModelConfig, PicoDecoderHFConfig]): Configuration containing:
|
345 |
+
- config.d_model: Model dimension
|
346 |
+
- config.activation_hidden_dim: Hidden dimension (typically 4 * d_model)
|
347 |
+
|
348 |
+
References:
|
349 |
+
https://arxiv.org/abs/2002.05202
|
350 |
+
"""
|
351 |
+
|
352 |
+
def __init__(self, config: Union["ModelConfig", "PicoDecoderHFConfig"]):
|
353 |
+
super().__init__()
|
354 |
+
|
355 |
+
model_dim = config.d_model
|
356 |
+
act_hidden_dim = config.activation_hidden_dim # usually 4 * d_model
|
357 |
+
|
358 |
+
self.w_0 = nn.Linear(model_dim, act_hidden_dim, bias=False)
|
359 |
+
self.w_1 = nn.Linear(model_dim, act_hidden_dim, bias=False)
|
360 |
+
self.w_2 = nn.Linear(act_hidden_dim, model_dim, bias=False)
|
361 |
+
|
362 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
363 |
+
return self.w_2(F.silu(self.w_0(x)) * self.w_1(x))
|
364 |
+
|
365 |
+
|
366 |
+
########################################################
|
367 |
+
#
|
368 |
+
# PicoDecoderBlock
|
369 |
+
#
|
370 |
+
########################################################
|
371 |
+
|
372 |
+
|
373 |
+
class PicoDecoderBlock(nn.Module):
|
374 |
+
"""Single Transformer Block with Attention and Feed-forward layers.
|
375 |
+
|
376 |
+
Implements a standard transformer block with:
|
377 |
+
- Multi-head attention with normalization and residual connection
|
378 |
+
- SwiGLU feed-forward network with normalization and residual connection
|
379 |
+
|
380 |
+
Args:
|
381 |
+
config (Union[ModelConfig, PicoDecoderHFConfig]): Model configuration; either a dataclass or
|
382 |
+
a HuggingFace PicoDecoderHFConfig
|
383 |
+
"""
|
384 |
+
|
385 |
+
def __init__(
|
386 |
+
self,
|
387 |
+
config: Union["ModelConfig", "PicoDecoderHFConfig"],
|
388 |
+
):
|
389 |
+
super().__init__()
|
390 |
+
|
391 |
+
self.attention = Attention(config)
|
392 |
+
self.swiglu = SwiGLU(config)
|
393 |
+
self.attention_norm = RMSNorm(config)
|
394 |
+
self.swiglu_norm = RMSNorm(config)
|
395 |
+
|
396 |
+
def forward(
|
397 |
+
self,
|
398 |
+
input: torch.Tensor,
|
399 |
+
mask: Optional[torch.Tensor] = None,
|
400 |
+
past_key_values: Optional[Tuple[torch.Tensor]] = None,
|
401 |
+
use_cache: bool = False,
|
402 |
+
) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor, torch.Tensor]]]:
|
403 |
+
attention_output, cached_key_values = self.attention(
|
404 |
+
self.attention_norm(input),
|
405 |
+
mask=mask,
|
406 |
+
past_key_values=past_key_values,
|
407 |
+
use_cache=use_cache,
|
408 |
+
)
|
409 |
+
# NOTE: cached_key_values is None if use_cache is False
|
410 |
+
|
411 |
+
h = input + attention_output
|
412 |
+
out = h + self.swiglu(self.swiglu_norm(h))
|
413 |
+
return out, cached_key_values
|
414 |
+
|
415 |
+
|
416 |
+
########################################################
|
417 |
+
#
|
418 |
+
# Pico Decoder (Causal Transformer Model)
|
419 |
+
#
|
420 |
+
########################################################
|
421 |
+
|
422 |
+
|
423 |
+
class PicoDecoder(nn.Module):
|
424 |
+
"""
|
425 |
+
Pico Decoder: combines the embedding, causal decoder blocks, and output projection into a
|
426 |
+
single autoregressive model.
|
427 |
+
|
428 |
+
For more information on the model, see the classes for the modules that make up the model.
|
429 |
+
"""
|
430 |
+
|
431 |
+
def __init__(
|
432 |
+
self,
|
433 |
+
model_config: Union["ModelConfig", "PicoDecoderHFConfig"],
|
434 |
+
):
|
435 |
+
super().__init__()
|
436 |
+
self.config = model_config
|
437 |
+
|
438 |
+
self.embedding_proj = nn.Embedding(self.config.vocab_size, self.config.d_model)
|
439 |
+
self.layers = nn.ModuleList(
|
440 |
+
[PicoDecoderBlock(self.config) for _ in range(self.config.n_layers)]
|
441 |
+
)
|
442 |
+
self.output_norm = RMSNorm(self.config)
|
443 |
+
self.de_embedding_proj = nn.Linear(
|
444 |
+
self.config.d_model, self.config.vocab_size, bias=False
|
445 |
+
)
|
446 |
+
|
447 |
+
def convert_to_hf_model(self) -> "PicoDecoderHF":
|
448 |
+
"""Convert the Lightning model to a HuggingFace model."""
|
449 |
+
# Create HF config without fabric-specific settings
|
450 |
+
hf_config = PicoDecoderHFConfig.from_dataclass(self.config)
|
451 |
+
|
452 |
+
# Create new HF model
|
453 |
+
hf_model = PicoDecoderHF(hf_config)
|
454 |
+
|
455 |
+
# Copy state dict, excluding fabric-specific keys
|
456 |
+
hf_model.load_state_dict(self.state_dict(prefix="pico_decoder."))
|
457 |
+
|
458 |
+
return hf_model
|
459 |
+
|
460 |
+
def forward(
|
461 |
+
self,
|
462 |
+
input_ids: torch.Tensor,
|
463 |
+
past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
|
464 |
+
use_cache: bool = False,
|
465 |
+
) -> Tuple[torch.Tensor, Optional[Tuple[Tuple[torch.Tensor, torch.Tensor]]]]:
|
466 |
+
"""
|
467 |
+
This is the forward pass for the entire Pico model. It boils down to:
|
468 |
+
- Embedding the input ids
|
469 |
+
- Creating a causal mask
|
470 |
+
- Processing through the pico layers
|
471 |
+
- Projecting the output to logits
|
472 |
+
|
473 |
+
NOTE: One feature that might be confusing is the KV cache. The KV cache is used to speed up
|
474 |
+
generation by caching the KV pairs from previous forward passes. This is useful when doing
|
475 |
+
tasks that require generating multiple tokens conditioned on previous tokens (e.g. language
|
476 |
+
modeling, text generation, etc.). The way the KV cache is implemented is that each layer has
|
477 |
+
its own KV cache which is stored as a tuple. The whole model then stores a tuple of these
|
478 |
+
KV caches (so a tuple of tuples).
|
479 |
+
"""
|
480 |
+
|
481 |
+
seq_len = input_ids.shape[-1]
|
482 |
+
h = self.embedding_proj(input_ids)
|
483 |
+
|
484 |
+
# Calculate start position from past cached KV pairs. Remember that each layer has its
|
485 |
+
# own KV Cache. So when we index past_key_values, we need to index into the KV pairs for the
|
486 |
+
# correct layer and then for either the keys or values.
|
487 |
+
start_pos = 0 if past_key_values is None else past_key_values[0][0].shape[1]
|
488 |
+
|
489 |
+
# Create causal mask for current sequence
|
490 |
+
mask = None
|
491 |
+
if seq_len > 1:
|
492 |
+
mask = torch.full((seq_len, seq_len), float("-inf"))
|
493 |
+
mask = torch.triu(mask, diagonal=1)
|
494 |
+
|
495 |
+
# If using KV cache, extend mask to cover cached sequence length
|
496 |
+
if past_key_values is not None:
|
497 |
+
# Add zeros for cached tokens (we can attend to all of them)
|
498 |
+
mask = torch.hstack([torch.zeros((seq_len, start_pos)), mask])
|
499 |
+
|
500 |
+
mask = mask.to(h.device)
|
501 |
+
|
502 |
+
# NOTE: If we are using the cache, we need to store the cached KV pairs for each layer
|
503 |
+
# in a tuple. Each layer will have its own cached KV pair which we aggregate in a tuple.
|
504 |
+
cached_key_values = () if use_cache else None
|
505 |
+
|
506 |
+
# Process through transformer blocks
|
507 |
+
for idx, layer in enumerate(self.layers):
|
508 |
+
layer_past_key_values = (
|
509 |
+
past_key_values[idx] if past_key_values is not None else None
|
510 |
+
)
|
511 |
+
|
512 |
+
h, layer_cached_key_values = layer(
|
513 |
+
h, mask=mask, past_key_values=layer_past_key_values, use_cache=use_cache
|
514 |
+
)
|
515 |
+
|
516 |
+
if use_cache:
|
517 |
+
cached_key_values += (layer_cached_key_values,)
|
518 |
+
|
519 |
+
# Final norm and projection
|
520 |
+
h = self.output_norm(h)
|
521 |
+
logits = self.de_embedding_proj(h).float()
|
522 |
+
|
523 |
+
return logits, cached_key_values
|
524 |
+
|
525 |
+
|
526 |
+
########################################################
|
527 |
+
#
|
528 |
+
# HuggingFace Wrapper
|
529 |
+
#
|
530 |
+
########################################################
|
531 |
+
|
532 |
+
"""
|
533 |
+
HuggingFace wrapper for the Pico model.
|
534 |
+
|
535 |
+
Many evaluation frameworks require a model be setup as a HuggingFace model, so we provide a simple
|
536 |
+
wrapper that does just that. When we save checkpoints of the Pico model, we save both the normal
|
537 |
+
Pico model as well as the model wrapped in this HuggingFace class.
|
538 |
+
|
539 |
+
This also lets you do cool things like:
|
540 |
+
|
541 |
+
`model = AutoModelForCausalLM.from_pretrained("path/to/checkpoint")`
|
542 |
+
"""
|
543 |
+
|
544 |
+
|
545 |
+
class PicoDecoderHFConfig(PretrainedConfig):
|
546 |
+
"""HuggingFace config for Pico model."""
|
547 |
+
|
548 |
+
model_type = "pico_decoder"
|
549 |
+
|
550 |
+
@classmethod
|
551 |
+
def from_dict(cls, config_dict: Dict[str, Any], **kwargs) -> "PicoDecoderHFConfig":
|
552 |
+
# NOTE The typical from_dict method doesn't actually set the attributes unless they are
|
553 |
+
# defined in the constructor.
|
554 |
+
|
555 |
+
pico_config = cls(**kwargs)
|
556 |
+
|
557 |
+
# Because this class is just a wrapper around the ModelConfig dataclass, we need to do
|
558 |
+
# a little extra work to ensure that the attributes are actually set.
|
559 |
+
for key, value in config_dict.items():
|
560 |
+
setattr(pico_config, key, value)
|
561 |
+
|
562 |
+
return_unused_kwargs = kwargs.pop("return_unused_kwargs", False)
|
563 |
+
unused_kwargs = {
|
564 |
+
key: value for key, value in kwargs.items() if not hasattr(pico_config, key)
|
565 |
+
}
|
566 |
+
|
567 |
+
if return_unused_kwargs:
|
568 |
+
return pico_config, unused_kwargs
|
569 |
+
return pico_config
|
570 |
+
|
571 |
+
@classmethod
|
572 |
+
def from_dataclass(cls, model_config: "ModelConfig"):
|
573 |
+
return cls.from_dict(asdict(model_config))
|
574 |
+
|
575 |
+
|
576 |
+
class PicoDecoderHF(PreTrainedModel):
|
577 |
+
"""HuggingFace wrapper for Pico model."""
|
578 |
+
|
579 |
+
config_class = PicoDecoderHFConfig
|
580 |
+
_no_split_modules = ["PicoBlock", "Attention", "SwiGLU", "RMSNorm"]
|
581 |
+
|
582 |
+
def __init__(self, config: PicoDecoderHFConfig):
|
583 |
+
super().__init__(config)
|
584 |
+
self.pico_decoder = PicoDecoder(config)
|
585 |
+
|
586 |
+
def forward(
|
587 |
+
self,
|
588 |
+
input_ids: torch.Tensor,
|
589 |
+
past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
|
590 |
+
use_cache: bool = False,
|
591 |
+
**kwargs,
|
592 |
+
) -> Union[CausalLMOutput, CausalLMOutputWithPast]:
|
593 |
+
"""HuggingFace forward pass wrapper.
|
594 |
+
|
595 |
+
Forwards pass for the HuggingFace version of the Pico Model. Basic wrapper around the
|
596 |
+
Pico model's forward pass, and returns the output as a HuggingFace CausalLMOutput.
|
597 |
+
"""
|
598 |
+
logits, past_key_values = self.pico_decoder(
|
599 |
+
input_ids, past_key_values, use_cache
|
600 |
+
)
|
601 |
+
if use_cache:
|
602 |
+
return CausalLMOutputWithPast(
|
603 |
+
logits=logits,
|
604 |
+
past_key_values=past_key_values,
|
605 |
+
)
|
606 |
+
else:
|
607 |
+
return CausalLMOutput(
|
608 |
+
logits=logits,
|
609 |
+
)
|
610 |
+
|
611 |
+
|
612 |
+
# Register for auto classes
|
613 |
+
PicoDecoderHFConfig.register_for_auto_class()
|
614 |
+
PicoDecoderHF.register_for_auto_class("AutoModel")
|
615 |
+
PicoDecoderHF.register_for_auto_class("AutoModelForCausalLM")
|