Text Generation
Transformers
Safetensors
English
fable
educational
tiny-model
from-scratch
byte-bpe
decoder-only
conversational
custom_code
Instructions to use dev-sandakelum/fable-mini with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use dev-sandakelum/fable-mini with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="dev-sandakelum/fable-mini", trust_remote_code=True) messages = [ {"role": "user", "content": "Who are you?"}, ] pipe(messages)# pip install -U transformers accelerate # Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("dev-sandakelum/fable-mini", trust_remote_code=True, device_map="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use dev-sandakelum/fable-mini with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "dev-sandakelum/fable-mini" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "dev-sandakelum/fable-mini", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/dev-sandakelum/fable-mini
- SGLang
How to use dev-sandakelum/fable-mini with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "dev-sandakelum/fable-mini" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "dev-sandakelum/fable-mini", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "dev-sandakelum/fable-mini" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "dev-sandakelum/fable-mini", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Docker Model Runner
How to use dev-sandakelum/fable-mini with Docker Model Runner:
docker model run hf.co/dev-sandakelum/fable-mini
File size: 10,452 Bytes
feb6b86 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 | """FableLMHeadModel — Hugging Face wrapper for the Fable miniature (model_type "fable").
Self-contained copy of the `fable_lm.FableLM` architecture (HF remote-code files
must not import the repo package):
pre-LN decoder-only transformer, learned absolute positions, GELU FF,
fused bias-free QKV, tied input/output embeddings.
Identical module tree and tensor names, so a checkpoint state dict loads
unchanged:
tok_emb.weight, pos_emb.weight,
blocks.{i}.ln1.{weight,bias}, blocks.{i}.attn.qkv.weight, blocks.{i}.attn.proj.weight,
blocks.{i}.ln2.{weight,bias}, blocks.{i}.ff.fc.{weight,bias}, blocks.{i}.ff.proj.{weight,bias},
ln_f.{weight,bias}, head.weight (tied to tok_emb.weight)
forward(input_ids, attention_mask, labels) -> CausalLMOutputWithPast, with an
optional DynamicCache KV cache for generate(). Supports temperature, top_p,
top_k, max_new_tokens and repetition_penalty through GenerationMixin.
"""
from __future__ import annotations
import math
from typing import Optional, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import GenerationMixin, PreTrainedModel
from transformers.cache_utils import Cache, DynamicCache
from transformers.modeling_outputs import CausalLMOutputWithPast
# transformers re-exports its own nn.init wrappers (they take a `generator`
# kwarg and advance the global RNG differently when it is explicitly None).
# Importing torch's originals by module path guarantees the same init draws
# as fable_lm.model.FableLM, so same-seed constructors are bit-identical.
_nn_init_normal = torch.nn.init.normal_
_nn_init_zeros = torch.nn.init.zeros_
from .configuration_fable import FableConfig
class CausalSelfAttention(nn.Module):
"""Fused-QKV multi-head self-attention, bias-free, with KV-cache support."""
def __init__(self, cfg: FableConfig):
super().__init__()
assert cfg.hidden_size % cfg.num_attention_heads == 0
self.n_heads = cfg.num_attention_heads
self.d_head = cfg.hidden_size // cfg.num_attention_heads
self.qkv = nn.Linear(cfg.hidden_size, 3 * cfg.hidden_size, bias=False)
self.proj = nn.Linear(cfg.hidden_size, cfg.hidden_size, bias=False)
def forward(self, x, attn_mask_4d, past_key_values=None, layer_idx=0, use_cache=True):
B, T, C = x.shape
qkv = self.qkv(x) # (B, T, 3C)
q, k, v = qkv.split(C, dim=2)
q = q.view(B, T, self.n_heads, self.d_head).transpose(1, 2) # (B, nh, T, dh)
k = k.view(B, T, self.n_heads, self.d_head).transpose(1, 2)
v = v.view(B, T, self.n_heads, self.d_head).transpose(1, 2)
if past_key_values is not None:
k, v = past_key_values.update(k, v, layer_idx)
att = (q @ k.transpose(-2, -1)) / math.sqrt(self.d_head) # (B, nh, T, kv)
if attn_mask_4d is not None:
att = att.masked_fill(~attn_mask_4d, float("-inf"))
att = F.softmax(att, dim=-1)
y = att @ v # (B, nh, T, dh)
y = y.transpose(1, 2).contiguous().view(B, T, C)
return self.proj(y)
class FeedForward(nn.Module):
def __init__(self, cfg: FableConfig):
super().__init__()
self.fc = nn.Linear(cfg.hidden_size, cfg.intermediate_size)
self.proj = nn.Linear(cfg.intermediate_size, cfg.hidden_size)
def forward(self, x):
return self.proj(F.gelu(self.fc(x)))
class Block(nn.Module):
"""Pre-LN block: x + attn(ln1(x)), then x + ff(ln2(x)) — as in fable_lm."""
def __init__(self, cfg: FableConfig):
super().__init__()
self.ln1 = nn.LayerNorm(cfg.hidden_size)
self.attn = CausalSelfAttention(cfg)
self.ln2 = nn.LayerNorm(cfg.hidden_size)
self.ff = FeedForward(cfg)
def forward(self, x, attn_mask_4d, past_key_values=None, layer_idx=0, use_cache=True):
x = x + self.attn(self.ln1(x), attn_mask_4d, past_key_values, layer_idx, use_cache)
x = x + self.ff(self.ln2(x))
return x
class FablePreTrainedModel(PreTrainedModel):
config_class = FableConfig
base_model_prefix = "fable"
supports_gradient_checkpointing = False
_tied_weights_keys = {"head.weight": "tok_emb.weight"} # 5.x: {target: source}
def _init_weights(self, module):
"""Match fable_lm.FableLM._init: normal(0, 0.02), zero biases.
Uses torch's original nn.init functions (not transformers' re-exports)
so the random stream matches fable_lm exactly for a given seed.
"""
if isinstance(module, nn.Linear):
_nn_init_normal(module.weight, mean=0.0, std=0.02)
if module.bias is not None:
_nn_init_zeros(module.bias)
elif isinstance(module, nn.Embedding):
_nn_init_normal(module.weight, mean=0.0, std=0.02)
class FableLMHeadModel(FablePreTrainedModel, GenerationMixin):
def __init__(self, config: FableConfig):
super().__init__(config)
self.tok_emb = nn.Embedding(config.vocab_size, config.hidden_size)
self.pos_emb = nn.Embedding(config.max_position_embeddings, config.hidden_size)
self.blocks = nn.ModuleList([Block(config) for _ in range(config.num_hidden_layers)])
self.ln_f = nn.LayerNorm(config.hidden_size)
# tied output head: logits = ln_f(x) @ tok_emb.weight.T
self.head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
# Tie BEFORE post_init, exactly like fable_lm.FableLM ties before its
# apply(_init): then the init draw for `head` overwrites the shared
# tok_emb storage, and same-seed construction is bit-identical to the
# original. post_init()'s tie_weights() then re-ties a no-op.
self.head.weight = self.tok_emb.weight
self.post_init()
# ---- embeddings plumbing (needed by PreTrainedModel.tie_weights) ----
def get_input_embeddings(self):
return self.tok_emb
def set_input_embeddings(self, value):
self.tok_emb = value
def get_output_embeddings(self):
return self.head
def set_output_embeddings(self, value):
self.head = value
# ---- mask helper ----
@staticmethod
def _make_attention_mask(
batch_size: int,
q_len: int,
kv_len: int,
past_len: int,
attention_mask: Optional[torch.Tensor],
device: torch.device,
) -> Optional[torch.Tensor]:
"""Combined causal + padding mask as a bool (B, 1, T, kv) keep-mask.
None (no padding) reduces to the pure causal mask, identical to
fable_lm's `masked_fill(triu(diagonal=1), -inf)` when past_len == 0.
"""
causal = torch.ones(q_len, kv_len, dtype=torch.bool, device=device).tril(diagonal=past_len)
keep = causal[None, None, :, :] # (1, 1, T, kv)
if attention_mask is not None:
pad = attention_mask.to(device=device, dtype=torch.bool)
if pad.ndim == 2: # (B, kv) -> (B, 1, 1, kv)
pad = pad[:, None, None, :]
keep = keep & pad
elif past_len == 0 and q_len == kv_len:
# exactly the original mask (triu diagonal=1 inverted) — skip the
# extra AND for bit-identical parity with fable_lm
return keep
return keep
def forward(
self,
input_ids: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.Tensor] = None,
past_key_values: Optional[Cache] = None,
inputs_embeds: Optional[torch.Tensor] = None,
labels: Optional[torch.Tensor] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
**kwargs,
) -> CausalLMOutputWithPast:
"""Causal LM forward. `labels` are shifted internally (HF convention):
loss = CE(logits[:, :-1], labels[:, 1:])."""
if input_ids is None and inputs_embeds is None:
raise ValueError("input_ids or inputs_embeds required")
use_cache = True if use_cache is None else use_cache
if inputs_embeds is not None:
x = inputs_embeds
B, T = x.shape[:2]
else:
x = self.tok_emb(input_ids)
B, T = input_ids.shape
past_len = past_key_values.get_seq_length() if isinstance(past_key_values, Cache) else 0
if T + past_len > self.config.max_position_embeddings:
raise ValueError(
f"sequence length {T + past_len} exceeds max_position_embeddings "
f"{self.config.max_position_embeddings}"
)
if position_ids is None:
position_ids = torch.arange(past_len, past_len + T, device=x.device)
x = x + self.pos_emb(position_ids) # (B, T, C); pos_emb(ids) broadcasts from (T, C)
attn_mask_4d = self._make_attention_mask(
B, T, T + past_len, past_len, attention_mask, x.device
)
if use_cache and past_key_values is None:
past_key_values = DynamicCache()
for i, blk in enumerate(self.blocks):
x = blk(x, attn_mask_4d, past_key_values, i, use_cache)
x = self.ln_f(x)
logits = self.head(x)
loss = None
if labels is not None:
shift_logits = logits[..., :-1, :].contiguous()
shift_labels = labels[..., 1:].contiguous()
loss = F.cross_entropy(
shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1), ignore_index=-100
)
return CausalLMOutputWithPast(
loss=loss,
logits=logits,
past_key_values=past_key_values if use_cache else None,
hidden_states=None,
attentions=None,
)
def num_params(self, non_embedding: bool = False) -> int:
"""fable_lm-compatible param count (tied head counted once)."""
n = sum(p.numel() for p in self.parameters())
if non_embedding:
n -= self.pos_emb.weight.numel()
return n
try:
# Local (non-remote-code) use: registering with the Auto classes makes
# from_pretrained work on a local hf_model/ dir without trust_remote_code.
from transformers import AutoModelForCausalLM
AutoModelForCausalLM.register(FableConfig, FableLMHeadModel, exist_ok=True)
except Exception: # pragma: no cover
pass
|