Spaces:
Running
on
Zero
Running
on
Zero
Update beeper_model.py
Browse files- beeper_model.py +126 -222
beeper_model.py
CHANGED
@@ -1,38 +1,27 @@
|
|
1 |
# beeper.py
|
2 |
-
#
|
3 |
-
# Beeper Full Penta Controller — Rose-based tiny GPT (inference module with runtime pentachora influence)
|
4 |
-
# - Decoder-only GPT with SDPA (FlashAttention path on Ampere/Hopper)
|
5 |
-
# - Runtime "vertex pull" uses config["runtime_pentachora"] to bias hidden states toward
|
6 |
-
# pentachora vertices (coarse/topic/mood) exactly like training-time behavior, but non-destructive
|
7 |
-
# and fully toggleable.
|
8 |
-
# --------------------------------------------------------------------------------------------------
|
9 |
from __future__ import annotations
|
10 |
|
11 |
-
import math
|
12 |
-
import re
|
13 |
-
import inspect
|
14 |
from contextlib import nullcontext
|
15 |
-
from typing import Optional,
|
16 |
|
17 |
import torch
|
18 |
import torch.nn as nn
|
19 |
import torch.nn.functional as F
|
20 |
|
21 |
-
# --- Prefer high-throughput matmul where possible (Ampere/Hopper) ---
|
22 |
torch.set_float32_matmul_precision("high")
|
23 |
torch.backends.cuda.matmul.allow_tf32 = True
|
24 |
torch.backends.cudnn.allow_tf32 = True
|
25 |
|
26 |
-
# ----
|
27 |
try:
|
28 |
-
# PyTorch 2.3+ modern API
|
29 |
from torch.nn.attention import sdpa_kernel as _sdpa_kernel_modern
|
30 |
from torch.nn.attention import SDPBackend as _SDPBackend
|
31 |
_SDPA_SIG = inspect.signature(_sdpa_kernel_modern)
|
32 |
_sdpa_kernel = _sdpa_kernel_modern
|
33 |
except Exception:
|
34 |
try:
|
35 |
-
# Legacy API
|
36 |
from torch.backends.cuda import sdp_kernel as _sdpa_kernel_legacy
|
37 |
_SDPA_SIG = inspect.signature(_sdpa_kernel_legacy)
|
38 |
_SDPBackend = None
|
@@ -42,39 +31,31 @@ except Exception:
|
|
42 |
_SDPBackend = None
|
43 |
_sdpa_kernel = None
|
44 |
|
45 |
-
|
46 |
def sdpa_ctx_prefer_flash():
|
47 |
-
"""Bias SDPA toward FlashAttention where possible; otherwise no-op."""
|
48 |
if _sdpa_kernel is None or _SDPA_SIG is None:
|
49 |
return nullcontext()
|
50 |
params = {p.name for p in _SDPA_SIG.parameters.values()}
|
51 |
try:
|
52 |
if "backends" in params and _SDPBackend is not None:
|
53 |
-
return _sdpa_kernel(backends=[
|
54 |
-
_SDPBackend.FLASH_ATTENTION,
|
55 |
-
_SDPBackend.EFFICIENT_ATTENTION,
|
56 |
-
_SDPBackend.MATH
|
57 |
-
])
|
58 |
if "backend" in params and _SDPBackend is not None:
|
59 |
return _sdpa_kernel(backend=_SDPBackend.FLASH_ATTENTION)
|
60 |
-
if {"enable_flash",
|
61 |
return _sdpa_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=True)
|
62 |
-
if {"use_flash",
|
63 |
return _sdpa_kernel(use_flash=True, use_math=False, use_mem_efficient=True)
|
64 |
except Exception:
|
65 |
pass
|
66 |
return nullcontext()
|
67 |
|
68 |
-
|
69 |
-
# --------------------------------- Core blocks ------------------------------------------------------
|
70 |
class CausalSelfAttention(nn.Module):
|
71 |
-
"""Multi-head causal self-attention using PyTorch SDPA."""
|
72 |
def __init__(self, dim: int, n_heads: int, attn_dropout: float = 0.0):
|
73 |
super().__init__()
|
74 |
-
assert dim % n_heads == 0
|
75 |
-
self.nh =
|
76 |
-
self.hd = dim //
|
77 |
-
self.qkv = nn.Linear(dim, 3
|
78 |
self.proj = nn.Linear(dim, dim, bias=False)
|
79 |
self.attn_dropout = float(attn_dropout)
|
80 |
|
@@ -82,78 +63,59 @@ class CausalSelfAttention(nn.Module):
|
|
82 |
B, T, C = x.shape
|
83 |
qkv = self.qkv(x)
|
84 |
q, k, v = qkv.chunk(3, dim=-1)
|
85 |
-
q = q.view(B, T, self.nh, self.hd).transpose(1, 2)
|
86 |
k = k.view(B, T, self.nh, self.hd).transpose(1, 2)
|
87 |
v = v.view(B, T, self.nh, self.hd).transpose(1, 2)
|
88 |
-
|
89 |
if x.is_cuda:
|
90 |
with sdpa_ctx_prefer_flash():
|
91 |
-
y = F.scaled_dot_product_attention(
|
92 |
-
|
93 |
-
is_causal=True,
|
94 |
-
dropout_p=self.attn_dropout if self.training else 0.0,
|
95 |
-
)
|
96 |
else:
|
97 |
scale = 1.0 / math.sqrt(self.hd)
|
98 |
att = (q @ k.transpose(-2, -1)) * scale
|
99 |
-
mask = torch.full((1,
|
100 |
-
|
101 |
-
att = (att + mask).softmax(dim=-1)
|
102 |
-
y = att @ v
|
103 |
-
|
104 |
y = y.transpose(1, 2).contiguous().view(B, T, C)
|
105 |
return self.proj(y)
|
106 |
|
107 |
-
|
108 |
class MLP(nn.Module):
|
109 |
-
"""GELU MLP with dropout, sized by mlp_ratio."""
|
110 |
def __init__(self, dim: int, mlp_ratio: float = 4.0, dropout: float = 0.1):
|
111 |
super().__init__()
|
112 |
-
|
113 |
-
self.fc1 = nn.Linear(dim,
|
114 |
-
self.fc2 = nn.Linear(
|
115 |
self.drop = nn.Dropout(dropout)
|
116 |
-
|
117 |
-
|
118 |
-
x = self.fc1(x)
|
119 |
-
x = F.gelu(x, approximate="tanh")
|
120 |
x = self.drop(x)
|
121 |
x = self.fc2(x)
|
122 |
x = self.drop(x)
|
123 |
return x
|
124 |
|
125 |
-
|
126 |
-
# --------------------------------- Beeper Model -----------------------------------------------------
|
127 |
class BeeperRoseGPT(nn.Module):
|
128 |
"""
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
|
140 |
-
"mood_alpha": float # hidden blend strength for mood bank
|
141 |
-
}
|
142 |
-
Notes:
|
143 |
-
- Shares token embedding with LM head (tied weights).
|
144 |
-
- Includes Rose anchors and pentachora banks; at runtime we can apply a *non-destructive*
|
145 |
-
vertex pull to hidden states before the LM head using the above config.
|
146 |
"""
|
147 |
def __init__(self, cfg: dict):
|
148 |
super().__init__()
|
149 |
V, D, Ctx = cfg["vocab_size"], cfg["dim"], cfg["context"]
|
150 |
-
H, L, MR
|
151 |
-
RD, AD
|
152 |
self.grad_checkpoint = bool(cfg.get("grad_checkpoint", False))
|
153 |
self.runtime_cfg: Dict[str, Any] = dict(cfg.get("runtime_pentachora", {}) or {})
|
154 |
|
155 |
self.vocab_size, self.context = int(V), int(Ctx)
|
156 |
-
|
157 |
self.token_emb = nn.Embedding(V, D)
|
158 |
self.pos_emb = nn.Parameter(torch.zeros(1, Ctx, D))
|
159 |
self.drop = nn.Dropout(RD)
|
@@ -164,142 +126,123 @@ class BeeperRoseGPT(nn.Module):
|
|
164 |
"attn": CausalSelfAttention(D, H, attn_dropout=AD),
|
165 |
"norm2": nn.LayerNorm(D),
|
166 |
"mlp": MLP(D, mlp_ratio=MR, dropout=RD),
|
167 |
-
})
|
168 |
-
for _ in range(L)
|
169 |
])
|
170 |
-
|
171 |
-
self.norm = nn.LayerNorm(D)
|
172 |
self.lm_head = nn.Linear(D, V, bias=False)
|
173 |
-
self.lm_head.weight = self.token_emb.weight
|
174 |
|
175 |
-
# Rose
|
176 |
self.rose_proj = nn.Linear(D, D, bias=False)
|
177 |
-
self.rose_anchors = nn.Parameter(torch.randn(3, D) / (D
|
178 |
|
179 |
-
# Pentachora banks
|
180 |
self.register_buffer("pent_inited", torch.tensor(0, dtype=torch.uint8), persistent=False)
|
181 |
self.penta_coarse: Optional[nn.Parameter] = None # [C,5,D]
|
182 |
self.penta_medium: Optional[nn.Parameter] = None # [T,5,D]
|
183 |
self.penta_fine: Optional[nn.Parameter] = None # [M,5,D]
|
184 |
|
185 |
-
self.apply(self.
|
186 |
|
187 |
@staticmethod
|
188 |
-
def
|
189 |
if isinstance(m, nn.Linear):
|
190 |
nn.init.normal_(m.weight, mean=0.0, std=0.02)
|
191 |
-
if m.bias is not None:
|
192 |
-
nn.init.zeros_(m.bias)
|
193 |
elif isinstance(m, nn.Embedding):
|
194 |
nn.init.normal_(m.weight, mean=0.0, std=0.02)
|
195 |
|
196 |
-
# ---- Pentachora creation (must match sizes in checkpoint before strict load) -------------------
|
197 |
def ensure_pentachora(self, coarse_C: int, medium_C: int, fine_C: int, dim: int, device: torch.device):
|
198 |
-
"""Initialize pentachora banks if not already present."""
|
199 |
if self.pent_inited.item() == 1:
|
200 |
return
|
201 |
-
|
202 |
def bank(C: int) -> nn.Parameter:
|
203 |
-
if C <= 0:
|
204 |
-
return nn.Parameter(torch.zeros((0, 5, dim), device=device))
|
205 |
pts = torch.randn(C, 5, dim, device=device)
|
206 |
pts = F.normalize(pts - pts.mean(dim=1, keepdim=True), dim=-1)
|
207 |
return nn.Parameter(pts)
|
208 |
-
|
209 |
self.penta_coarse = bank(int(coarse_C))
|
210 |
self.penta_medium = bank(int(medium_C))
|
211 |
self.penta_fine = bank(int(fine_C))
|
212 |
self.pent_inited.fill_(1)
|
213 |
|
214 |
-
# ---- Runtime configuration helpers -------------------------------------------------------------
|
215 |
def set_runtime_pentachora(self, cfg: Dict[str, Any]) -> None:
|
216 |
-
"""Update runtime pentachora behavior (enable/alphas/temp/pool)."""
|
217 |
self.runtime_cfg.update(cfg or {})
|
218 |
|
219 |
def _pool_hidden(self, h: torch.Tensor, mode: str) -> torch.Tensor:
|
220 |
return h.mean(dim=1) if mode == "mean" else h[:, -1, :]
|
221 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
222 |
@staticmethod
|
223 |
def _weighted_nearest_vertex_target(
|
224 |
-
pooled: torch.Tensor,
|
225 |
-
bank: torch.Tensor,
|
226 |
-
temp: float
|
|
|
227 |
) -> torch.Tensor:
|
228 |
"""
|
229 |
-
|
230 |
-
then compute a softmax over classes of -min_dists/temp and take the
|
231 |
-
weighted average of those nearest vertices => [B,D] target.
|
232 |
"""
|
233 |
B, D = pooled.shape
|
234 |
-
|
235 |
-
if C == 0:
|
236 |
return pooled
|
237 |
-
|
238 |
-
|
239 |
-
diffs = pooled[:, None, None, :] - bank[None, :, :, :]
|
240 |
-
dists = torch.norm(diffs, dim=-1)
|
241 |
-
|
242 |
-
min_dists
|
243 |
-
|
244 |
-
|
245 |
-
|
246 |
-
# gather nearest vertex vectors: [B,C,D]
|
247 |
-
bank_exp = bank.unsqueeze(0).expand(B, -1, -1, -1) # [B,C,5,D]
|
248 |
-
gather_idx = min_idx.unsqueeze(-1).unsqueeze(-1).expand(B, C, 1, D)
|
249 |
-
nearest = torch.gather(bank_exp, 2, gather_idx).squeeze(2) # [B,C,D]
|
250 |
-
|
251 |
-
target = (weights.unsqueeze(-1) * nearest).sum(dim=1) # [B,D]
|
252 |
return target
|
253 |
|
254 |
-
def _apply_runtime_vertex_pull(
|
255 |
-
self,
|
256 |
-
h: torch.Tensor, # [B,T,D]
|
257 |
-
runtime_cfg: Dict[str, Any]
|
258 |
-
) -> torch.Tensor:
|
259 |
-
"""
|
260 |
-
Apply non-destructive vertex pull to hidden states using banks selected by runtime_cfg.
|
261 |
-
We compute a pooled latent, a per-bank target vector, form a delta, and blend it back into h.
|
262 |
-
"""
|
263 |
if not runtime_cfg or not runtime_cfg.get("enable", False):
|
264 |
return h
|
265 |
-
|
266 |
pool_mode = str(runtime_cfg.get("pool", "mean"))
|
267 |
temp = float(runtime_cfg.get("temp", 0.10))
|
268 |
-
|
269 |
-
|
270 |
-
|
271 |
-
|
272 |
-
alpha_mood = float(runtime_cfg.get("mood_alpha", 0.0))
|
273 |
-
|
274 |
-
if (alpha_coarse <= 0 and alpha_topic <= 0 and alpha_mood <= 0):
|
275 |
return h
|
276 |
|
277 |
-
pooled = self._pool_hidden(h, pool_mode)
|
278 |
-
|
279 |
-
|
280 |
-
if
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
|
288 |
-
|
289 |
-
|
290 |
-
|
291 |
-
|
292 |
-
delta =
|
293 |
-
|
294 |
-
|
295 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
296 |
return h
|
|
|
297 |
|
298 |
-
|
299 |
-
h = h + total_delta.unsqueeze(1) # [B,T,D]
|
300 |
-
return h
|
301 |
-
|
302 |
-
# ---- Backbone / forward -----------------------------------------------------------------------
|
303 |
def _block_forward(self, blk: nn.ModuleDict, x: torch.Tensor) -> torch.Tensor:
|
304 |
x = x + blk["attn"](blk["norm1"](x))
|
305 |
x = x + blk["mlp"](blk["norm2"](x))
|
@@ -312,87 +255,53 @@ class BeeperRoseGPT(nn.Module):
|
|
312 |
if self.grad_checkpoint and self.training:
|
313 |
from torch.utils.checkpoint import checkpoint
|
314 |
for blk in self.blocks:
|
315 |
-
x = checkpoint(lambda _x: self._block_forward(blk, _x), x) # type: ignore
|
316 |
else:
|
317 |
for blk in self.blocks:
|
318 |
x = self._block_forward(blk, x)
|
319 |
return self.norm(x)
|
320 |
|
321 |
def forward(self, idx: torch.Tensor, runtime_cfg: Optional[Dict[str, Any]] = None) -> torch.Tensor:
|
322 |
-
"""
|
323 |
-
Forward pass with optional runtime pentachora influence.
|
324 |
-
If runtime_cfg is None, falls back to self.runtime_cfg set at init or via set_runtime_pentachora().
|
325 |
-
"""
|
326 |
h = self.backbone(idx)
|
327 |
cfg = self.runtime_cfg if runtime_cfg is None else {**self.runtime_cfg, **(runtime_cfg or {})}
|
328 |
h = self._apply_runtime_vertex_pull(h, cfg)
|
329 |
return self.lm_head(h)
|
330 |
|
331 |
-
#
|
332 |
def hidden_states(self, idx: torch.Tensor) -> torch.Tensor:
|
333 |
-
"""Return final hidden states (pre-LM head)."""
|
334 |
return self.backbone(idx)
|
335 |
-
|
336 |
def rose_hidden_pool(self, h: torch.Tensor, mode: str = "mean") -> torch.Tensor:
|
337 |
-
""
|
338 |
-
return h.mean(dim=1) if mode == "mean" else h[:, -1, :]
|
339 |
|
340 |
-
|
341 |
-
|
342 |
-
def prepare_model_for_state_dict(
|
343 |
-
model: BeeperRoseGPT,
|
344 |
-
state_dict: "dict[str, torch.Tensor]",
|
345 |
-
device: Optional[torch.device] = None,
|
346 |
-
) -> None:
|
347 |
-
"""
|
348 |
-
Ensure model has pentachora parameters sized to match the incoming state_dict,
|
349 |
-
so we can load with strict=True. No-op if checkpoint lacks penta_* keys.
|
350 |
-
"""
|
351 |
device = device or next(model.parameters()).device
|
352 |
-
need = all(k in state_dict for k in ("penta_coarse",
|
353 |
-
if not need:
|
354 |
-
return
|
355 |
-
|
356 |
-
pc, pt, pm = state_dict["penta_coarse"], state_dict["penta_medium"], state_dict["penta_fine"]
|
357 |
-
|
358 |
-
def dims_ok(t: torch.Tensor, D: int) -> bool:
|
359 |
-
return t.ndim == 3 and t.size(1) == 5 and t.size(2) == D
|
360 |
-
|
361 |
D = model.token_emb.embedding_dim
|
362 |
-
|
363 |
-
|
364 |
-
|
365 |
model.ensure_pentachora(pc.size(0), pt.size(0), pm.size(0), dim=D, device=device)
|
366 |
|
367 |
-
|
368 |
-
# --------------------------------- Generation -------------------------------------------------------
|
369 |
def _detok(text: str) -> str:
|
370 |
text = re.sub(r"\s+([,.;:!?%])", r"\1", text)
|
371 |
text = re.sub(r"\s+([\)\]\}])", r"\1", text)
|
372 |
text = re.sub(r"([\(\[\{])\s+", r"\1", text)
|
373 |
return text
|
374 |
|
375 |
-
|
376 |
@torch.no_grad()
|
377 |
-
def generate(
|
378 |
-
|
379 |
-
|
380 |
-
|
381 |
-
|
382 |
-
|
383 |
-
|
384 |
-
|
385 |
-
|
386 |
-
|
387 |
-
presence_penalty: Optional[float] = None,
|
388 |
-
frequency_penalty: Optional[float] = None,
|
389 |
-
device: Optional[torch.device] = None,
|
390 |
-
detokenize: bool = True,
|
391 |
-
runtime_cfg: Optional[Dict[str, Any]] = None, # <— NEW: pass-through to forward()
|
392 |
-
) -> str:
|
393 |
-
"""
|
394 |
-
Penalized nucleus sampling with optional runtime pentachora influence.
|
395 |
-
"""
|
396 |
temperature = cfg.get("temperature", 0.9) if temperature is None else float(temperature)
|
397 |
top_k = cfg.get("top_k", 40) if top_k is None else int(top_k)
|
398 |
top_p = cfg.get("top_p", 0.9) if top_p is None else float(top_p)
|
@@ -408,14 +317,12 @@ def generate(
|
|
408 |
V = int(cfg["vocab_size"])
|
409 |
counts = torch.zeros(V, dtype=torch.int32, device=device)
|
410 |
for t in ids:
|
411 |
-
if 0 <= t < V:
|
412 |
-
counts[t] += 1
|
413 |
|
414 |
for _ in range(int(max_new_tokens)):
|
415 |
logits = model(x[:, -cfg["context"]:], runtime_cfg=runtime_cfg)
|
416 |
logits = logits[:, -1, :]
|
417 |
|
418 |
-
# Repetition penalty
|
419 |
if repetition_penalty and repetition_penalty != 1.0:
|
420 |
mask = counts > 0
|
421 |
if mask.any():
|
@@ -423,9 +330,8 @@ def generate(
|
|
423 |
logits[:, mask][pos] /= repetition_penalty
|
424 |
logits[:, mask][~pos] *= repetition_penalty
|
425 |
|
426 |
-
# Presence/frequency penalties
|
427 |
if presence_penalty or frequency_penalty:
|
428 |
-
pen = counts.float() * (frequency_penalty or 0.0) + (counts
|
429 |
logits = logits - pen.unsqueeze(0)
|
430 |
|
431 |
logits = logits / max(1e-8, temperature)
|
@@ -433,8 +339,7 @@ def generate(
|
|
433 |
if top_k and top_k > 0:
|
434 |
k = min(top_k, logits.size(-1))
|
435 |
v, ix = torch.topk(logits, k, dim=-1)
|
436 |
-
|
437 |
-
logits = filt.scatter_(-1, ix, v)
|
438 |
|
439 |
if top_p and top_p < 1.0:
|
440 |
sl, si = torch.sort(logits, descending=True)
|
@@ -449,8 +354,7 @@ def generate(
|
|
449 |
next_id = torch.multinomial(probs, num_samples=1)
|
450 |
x = torch.cat([x, next_id], dim=1)
|
451 |
nid = next_id.item()
|
452 |
-
if 0 <= nid < V:
|
453 |
-
counts[nid] += 1
|
454 |
|
455 |
out = tok.decode(x[0].tolist())
|
456 |
return _detok(out) if detokenize else out
|
|
|
1 |
# beeper.py
|
2 |
+
# Beeper — Rose-based tiny GPT (inference, with runtime pentachora influence + class/topic/mood selection)
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
from __future__ import annotations
|
4 |
|
5 |
+
import math, re, inspect
|
|
|
|
|
6 |
from contextlib import nullcontext
|
7 |
+
from typing import Optional, Dict, Any, Iterable
|
8 |
|
9 |
import torch
|
10 |
import torch.nn as nn
|
11 |
import torch.nn.functional as F
|
12 |
|
|
|
13 |
torch.set_float32_matmul_precision("high")
|
14 |
torch.backends.cuda.matmul.allow_tf32 = True
|
15 |
torch.backends.cudnn.allow_tf32 = True
|
16 |
|
17 |
+
# ---- SDPA (FlashAttention) selection ----
|
18 |
try:
|
|
|
19 |
from torch.nn.attention import sdpa_kernel as _sdpa_kernel_modern
|
20 |
from torch.nn.attention import SDPBackend as _SDPBackend
|
21 |
_SDPA_SIG = inspect.signature(_sdpa_kernel_modern)
|
22 |
_sdpa_kernel = _sdpa_kernel_modern
|
23 |
except Exception:
|
24 |
try:
|
|
|
25 |
from torch.backends.cuda import sdp_kernel as _sdpa_kernel_legacy
|
26 |
_SDPA_SIG = inspect.signature(_sdpa_kernel_legacy)
|
27 |
_SDPBackend = None
|
|
|
31 |
_SDPBackend = None
|
32 |
_sdpa_kernel = None
|
33 |
|
|
|
34 |
def sdpa_ctx_prefer_flash():
|
|
|
35 |
if _sdpa_kernel is None or _SDPA_SIG is None:
|
36 |
return nullcontext()
|
37 |
params = {p.name for p in _SDPA_SIG.parameters.values()}
|
38 |
try:
|
39 |
if "backends" in params and _SDPBackend is not None:
|
40 |
+
return _sdpa_kernel(backends=[_SDPBackend.FLASH_ATTENTION, _SDPBackend.EFFICIENT_ATTENTION, _SDPBackend.MATH])
|
|
|
|
|
|
|
|
|
41 |
if "backend" in params and _SDPBackend is not None:
|
42 |
return _sdpa_kernel(backend=_SDPBackend.FLASH_ATTENTION)
|
43 |
+
if {"enable_flash","enable_math","enable_mem_efficient"} <= params:
|
44 |
return _sdpa_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=True)
|
45 |
+
if {"use_flash","use_math","use_mem_efficient"} <= params:
|
46 |
return _sdpa_kernel(use_flash=True, use_math=False, use_mem_efficient=True)
|
47 |
except Exception:
|
48 |
pass
|
49 |
return nullcontext()
|
50 |
|
51 |
+
# ---------------- Blocks ----------------
|
|
|
52 |
class CausalSelfAttention(nn.Module):
|
|
|
53 |
def __init__(self, dim: int, n_heads: int, attn_dropout: float = 0.0):
|
54 |
super().__init__()
|
55 |
+
assert dim % n_heads == 0
|
56 |
+
self.nh = n_heads
|
57 |
+
self.hd = dim // n_heads
|
58 |
+
self.qkv = nn.Linear(dim, 3*dim, bias=False)
|
59 |
self.proj = nn.Linear(dim, dim, bias=False)
|
60 |
self.attn_dropout = float(attn_dropout)
|
61 |
|
|
|
63 |
B, T, C = x.shape
|
64 |
qkv = self.qkv(x)
|
65 |
q, k, v = qkv.chunk(3, dim=-1)
|
66 |
+
q = q.view(B, T, self.nh, self.hd).transpose(1, 2)
|
67 |
k = k.view(B, T, self.nh, self.hd).transpose(1, 2)
|
68 |
v = v.view(B, T, self.nh, self.hd).transpose(1, 2)
|
|
|
69 |
if x.is_cuda:
|
70 |
with sdpa_ctx_prefer_flash():
|
71 |
+
y = F.scaled_dot_product_attention(q, k, v, is_causal=True,
|
72 |
+
dropout_p=self.attn_dropout if self.training else 0.0)
|
|
|
|
|
|
|
73 |
else:
|
74 |
scale = 1.0 / math.sqrt(self.hd)
|
75 |
att = (q @ k.transpose(-2, -1)) * scale
|
76 |
+
mask = torch.triu(torch.full((1,1,T,T), float("-inf"), device=x.device), diagonal=1)
|
77 |
+
y = (att + mask).softmax(dim=-1) @ v
|
|
|
|
|
|
|
78 |
y = y.transpose(1, 2).contiguous().view(B, T, C)
|
79 |
return self.proj(y)
|
80 |
|
|
|
81 |
class MLP(nn.Module):
|
|
|
82 |
def __init__(self, dim: int, mlp_ratio: float = 4.0, dropout: float = 0.1):
|
83 |
super().__init__()
|
84 |
+
h = int(dim*mlp_ratio)
|
85 |
+
self.fc1 = nn.Linear(dim, h)
|
86 |
+
self.fc2 = nn.Linear(h, dim)
|
87 |
self.drop = nn.Dropout(dropout)
|
88 |
+
def forward(self, x):
|
89 |
+
x = F.gelu(self.fc1(x), approximate="tanh")
|
|
|
|
|
90 |
x = self.drop(x)
|
91 |
x = self.fc2(x)
|
92 |
x = self.drop(x)
|
93 |
return x
|
94 |
|
95 |
+
# --------------- Model ---------------
|
|
|
96 |
class BeeperRoseGPT(nn.Module):
|
97 |
"""
|
98 |
+
Runtime pentachora control via self.runtime_cfg:
|
99 |
+
{
|
100 |
+
"enable": bool,
|
101 |
+
"pool": "mean"|"last",
|
102 |
+
"temp": 0.10,
|
103 |
+
"coarse_alpha": float, "topic_alpha": float, "mood_alpha": float,
|
104 |
+
# NEW: selection masks (ints or lists of ints)
|
105 |
+
"coarse_select": Optional[Iterable[int]],
|
106 |
+
"topic_select": Optional[Iterable[int]],
|
107 |
+
"mood_select": Optional[Iterable[int]],
|
108 |
+
}
|
|
|
|
|
|
|
|
|
|
|
|
|
109 |
"""
|
110 |
def __init__(self, cfg: dict):
|
111 |
super().__init__()
|
112 |
V, D, Ctx = cfg["vocab_size"], cfg["dim"], cfg["context"]
|
113 |
+
H, L, MR = cfg["n_heads"], cfg["n_layers"], cfg["mlp_ratio"]
|
114 |
+
RD, AD = cfg.get("resid_dropout", 0.1), cfg.get("dropout", 0.0)
|
115 |
self.grad_checkpoint = bool(cfg.get("grad_checkpoint", False))
|
116 |
self.runtime_cfg: Dict[str, Any] = dict(cfg.get("runtime_pentachora", {}) or {})
|
117 |
|
118 |
self.vocab_size, self.context = int(V), int(Ctx)
|
|
|
119 |
self.token_emb = nn.Embedding(V, D)
|
120 |
self.pos_emb = nn.Parameter(torch.zeros(1, Ctx, D))
|
121 |
self.drop = nn.Dropout(RD)
|
|
|
126 |
"attn": CausalSelfAttention(D, H, attn_dropout=AD),
|
127 |
"norm2": nn.LayerNorm(D),
|
128 |
"mlp": MLP(D, mlp_ratio=MR, dropout=RD),
|
129 |
+
}) for _ in range(L)
|
|
|
130 |
])
|
131 |
+
self.norm = nn.LayerNorm(D)
|
|
|
132 |
self.lm_head = nn.Linear(D, V, bias=False)
|
133 |
+
self.lm_head.weight = self.token_emb.weight
|
134 |
|
135 |
+
# Rose anchors (kept for compatibility)
|
136 |
self.rose_proj = nn.Linear(D, D, bias=False)
|
137 |
+
self.rose_anchors = nn.Parameter(torch.randn(3, D) / (D**0.5))
|
138 |
|
139 |
+
# Pentachora banks
|
140 |
self.register_buffer("pent_inited", torch.tensor(0, dtype=torch.uint8), persistent=False)
|
141 |
self.penta_coarse: Optional[nn.Parameter] = None # [C,5,D]
|
142 |
self.penta_medium: Optional[nn.Parameter] = None # [T,5,D]
|
143 |
self.penta_fine: Optional[nn.Parameter] = None # [M,5,D]
|
144 |
|
145 |
+
self.apply(self._init)
|
146 |
|
147 |
@staticmethod
|
148 |
+
def _init(m):
|
149 |
if isinstance(m, nn.Linear):
|
150 |
nn.init.normal_(m.weight, mean=0.0, std=0.02)
|
151 |
+
if m.bias is not None: nn.init.zeros_(m.bias)
|
|
|
152 |
elif isinstance(m, nn.Embedding):
|
153 |
nn.init.normal_(m.weight, mean=0.0, std=0.02)
|
154 |
|
|
|
155 |
def ensure_pentachora(self, coarse_C: int, medium_C: int, fine_C: int, dim: int, device: torch.device):
|
|
|
156 |
if self.pent_inited.item() == 1:
|
157 |
return
|
|
|
158 |
def bank(C: int) -> nn.Parameter:
|
159 |
+
if C <= 0: return nn.Parameter(torch.zeros((0,5,dim), device=device))
|
|
|
160 |
pts = torch.randn(C, 5, dim, device=device)
|
161 |
pts = F.normalize(pts - pts.mean(dim=1, keepdim=True), dim=-1)
|
162 |
return nn.Parameter(pts)
|
|
|
163 |
self.penta_coarse = bank(int(coarse_C))
|
164 |
self.penta_medium = bank(int(medium_C))
|
165 |
self.penta_fine = bank(int(fine_C))
|
166 |
self.pent_inited.fill_(1)
|
167 |
|
|
|
168 |
def set_runtime_pentachora(self, cfg: Dict[str, Any]) -> None:
|
|
|
169 |
self.runtime_cfg.update(cfg or {})
|
170 |
|
171 |
def _pool_hidden(self, h: torch.Tensor, mode: str) -> torch.Tensor:
|
172 |
return h.mean(dim=1) if mode == "mean" else h[:, -1, :]
|
173 |
|
174 |
+
@staticmethod
|
175 |
+
def _normalize_indices(sel: Optional[Iterable[int]], C: int) -> Optional[torch.Tensor]:
|
176 |
+
if sel is None: return None
|
177 |
+
if isinstance(sel, int): sel = [sel]
|
178 |
+
sel = [int(x) for x in sel if 0 <= int(x) < C]
|
179 |
+
if not sel: return None
|
180 |
+
return torch.as_tensor(sel, dtype=torch.long)
|
181 |
+
|
182 |
@staticmethod
|
183 |
def _weighted_nearest_vertex_target(
|
184 |
+
pooled: torch.Tensor, # [B,D]
|
185 |
+
bank: torch.Tensor, # [C,5,D]
|
186 |
+
temp: float,
|
187 |
+
restrict_idx: Optional[torch.Tensor] = None # [K] or None
|
188 |
) -> torch.Tensor:
|
189 |
"""
|
190 |
+
If restrict_idx is given, compute target within the selected classes only.
|
|
|
|
|
191 |
"""
|
192 |
B, D = pooled.shape
|
193 |
+
if bank.size(0) == 0:
|
|
|
194 |
return pooled
|
195 |
+
if restrict_idx is not None:
|
196 |
+
bank = bank.index_select(0, restrict_idx.to(bank.device)) # [K,5,D]
|
197 |
+
diffs = pooled[:, None, None, :] - bank[None, :, :, :] # [B,C|K,5,D]
|
198 |
+
dists = torch.norm(diffs, dim=-1) # [B,C|K,5]
|
199 |
+
min_dists = dists.min(dim=2).values # [B,C|K]
|
200 |
+
sims = -min_dists / max(1e-8, float(temp)) # [B,C|K]
|
201 |
+
weights = F.softmax(sims, dim=-1) # [B,C|K]
|
202 |
+
nearest = bank.unsqueeze(0).gather(2, dists.argmin(dim=2)[...,None,None].expand(B, weights.size(1), 1, D)).squeeze(2) # [B,C|K,D]
|
203 |
+
target = (weights.unsqueeze(-1) * nearest).sum(dim=1) # [B,D]
|
|
|
|
|
|
|
|
|
|
|
|
|
204 |
return target
|
205 |
|
206 |
+
def _apply_runtime_vertex_pull(self, h: torch.Tensor, runtime_cfg: Dict[str, Any]) -> torch.Tensor:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
207 |
if not runtime_cfg or not runtime_cfg.get("enable", False):
|
208 |
return h
|
|
|
209 |
pool_mode = str(runtime_cfg.get("pool", "mean"))
|
210 |
temp = float(runtime_cfg.get("temp", 0.10))
|
211 |
+
a_coarse = float(runtime_cfg.get("coarse_alpha", 0.0))
|
212 |
+
a_topic = float(runtime_cfg.get("topic_alpha", 0.0))
|
213 |
+
a_mood = float(runtime_cfg.get("mood_alpha", 0.0))
|
214 |
+
if a_coarse<=0 and a_topic<=0 and a_mood<=0:
|
|
|
|
|
|
|
215 |
return h
|
216 |
|
217 |
+
pooled = self._pool_hidden(h, pool_mode) # [B,D]
|
218 |
+
delta = None
|
219 |
+
|
220 |
+
if a_coarse>0 and getattr(self, "penta_coarse", None) is not None:
|
221 |
+
C = self.penta_coarse.size(0)
|
222 |
+
r = self._normalize_indices(runtime_cfg.get("coarse_select"), C)
|
223 |
+
tgt = self._weighted_nearest_vertex_target(pooled, self.penta_coarse, temp, r)
|
224 |
+
d = tgt - pooled
|
225 |
+
delta = a_coarse * d if delta is None else delta + a_coarse * d
|
226 |
+
|
227 |
+
if a_topic>0 and getattr(self, "penta_medium", None) is not None:
|
228 |
+
C = self.penta_medium.size(0)
|
229 |
+
r = self._normalize_indices(runtime_cfg.get("topic_select"), C)
|
230 |
+
tgt = self._weighted_nearest_vertex_target(pooled, self.penta_medium, temp, r)
|
231 |
+
d = tgt - pooled
|
232 |
+
delta = a_topic * d if delta is None else delta + a_topic * d
|
233 |
+
|
234 |
+
if a_mood>0 and getattr(self, "penta_fine", None) is not None:
|
235 |
+
C = self.penta_fine.size(0)
|
236 |
+
r = self._normalize_indices(runtime_cfg.get("mood_select"), C)
|
237 |
+
tgt = self._weighted_nearest_vertex_target(pooled, self.penta_fine, temp, r)
|
238 |
+
d = tgt - pooled
|
239 |
+
delta = a_mood * d if delta is None else delta + a_mood * d
|
240 |
+
|
241 |
+
if delta is None:
|
242 |
return h
|
243 |
+
return h + delta.unsqueeze(1) # broadcast across time
|
244 |
|
245 |
+
# ---- forward ----
|
|
|
|
|
|
|
|
|
246 |
def _block_forward(self, blk: nn.ModuleDict, x: torch.Tensor) -> torch.Tensor:
|
247 |
x = x + blk["attn"](blk["norm1"](x))
|
248 |
x = x + blk["mlp"](blk["norm2"](x))
|
|
|
255 |
if self.grad_checkpoint and self.training:
|
256 |
from torch.utils.checkpoint import checkpoint
|
257 |
for blk in self.blocks:
|
258 |
+
x = checkpoint(lambda _x: self._block_forward(blk, _x), x) # type: ignore
|
259 |
else:
|
260 |
for blk in self.blocks:
|
261 |
x = self._block_forward(blk, x)
|
262 |
return self.norm(x)
|
263 |
|
264 |
def forward(self, idx: torch.Tensor, runtime_cfg: Optional[Dict[str, Any]] = None) -> torch.Tensor:
|
|
|
|
|
|
|
|
|
265 |
h = self.backbone(idx)
|
266 |
cfg = self.runtime_cfg if runtime_cfg is None else {**self.runtime_cfg, **(runtime_cfg or {})}
|
267 |
h = self._apply_runtime_vertex_pull(h, cfg)
|
268 |
return self.lm_head(h)
|
269 |
|
270 |
+
# Utilities
|
271 |
def hidden_states(self, idx: torch.Tensor) -> torch.Tensor:
|
|
|
272 |
return self.backbone(idx)
|
|
|
273 |
def rose_hidden_pool(self, h: torch.Tensor, mode: str = "mean") -> torch.Tensor:
|
274 |
+
return h.mean(dim=1) if mode=="mean" else h[:, -1, :]
|
|
|
275 |
|
276 |
+
# ---- Loader helper ----
|
277 |
+
def prepare_model_for_state_dict(model: BeeperRoseGPT, state_dict: Dict[str, torch.Tensor], device: Optional[torch.device] = None) -> None:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
278 |
device = device or next(model.parameters()).device
|
279 |
+
need = all(k in state_dict for k in ("penta_coarse","penta_medium","penta_fine"))
|
280 |
+
if not need: return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
281 |
D = model.token_emb.embedding_dim
|
282 |
+
pc, pt, pm = state_dict["penta_coarse"], state_dict["penta_medium"], state_dict["penta_fine"]
|
283 |
+
ok = lambda t: (t.ndim==3 and t.size(1)==5 and t.size(2)==D)
|
284 |
+
if not (ok(pc) and ok(pt) and ok(pm)): return
|
285 |
model.ensure_pentachora(pc.size(0), pt.size(0), pm.size(0), dim=D, device=device)
|
286 |
|
287 |
+
# ---- Generation ----
|
|
|
288 |
def _detok(text: str) -> str:
|
289 |
text = re.sub(r"\s+([,.;:!?%])", r"\1", text)
|
290 |
text = re.sub(r"\s+([\)\]\}])", r"\1", text)
|
291 |
text = re.sub(r"([\(\[\{])\s+", r"\1", text)
|
292 |
return text
|
293 |
|
|
|
294 |
@torch.no_grad()
|
295 |
+
def generate(model: BeeperRoseGPT, tok, cfg: dict, prompt: str,
|
296 |
+
max_new_tokens: int = 120, temperature: float | None = None,
|
297 |
+
top_k: int | None = None, top_p: float | None = None,
|
298 |
+
repetition_penalty: float | None = None,
|
299 |
+
presence_penalty: float | None = None,
|
300 |
+
frequency_penalty: float | None = None,
|
301 |
+
device: Optional[torch.device] = None,
|
302 |
+
detokenize: bool = True,
|
303 |
+
runtime_cfg: Optional[Dict[str, Any]] = None) -> str:
|
304 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
305 |
temperature = cfg.get("temperature", 0.9) if temperature is None else float(temperature)
|
306 |
top_k = cfg.get("top_k", 40) if top_k is None else int(top_k)
|
307 |
top_p = cfg.get("top_p", 0.9) if top_p is None else float(top_p)
|
|
|
317 |
V = int(cfg["vocab_size"])
|
318 |
counts = torch.zeros(V, dtype=torch.int32, device=device)
|
319 |
for t in ids:
|
320 |
+
if 0 <= t < V: counts[t] += 1
|
|
|
321 |
|
322 |
for _ in range(int(max_new_tokens)):
|
323 |
logits = model(x[:, -cfg["context"]:], runtime_cfg=runtime_cfg)
|
324 |
logits = logits[:, -1, :]
|
325 |
|
|
|
326 |
if repetition_penalty and repetition_penalty != 1.0:
|
327 |
mask = counts > 0
|
328 |
if mask.any():
|
|
|
330 |
logits[:, mask][pos] /= repetition_penalty
|
331 |
logits[:, mask][~pos] *= repetition_penalty
|
332 |
|
|
|
333 |
if presence_penalty or frequency_penalty:
|
334 |
+
pen = counts.float() * (frequency_penalty or 0.0) + (counts>0).float() * (presence_penalty or 0.0)
|
335 |
logits = logits - pen.unsqueeze(0)
|
336 |
|
337 |
logits = logits / max(1e-8, temperature)
|
|
|
339 |
if top_k and top_k > 0:
|
340 |
k = min(top_k, logits.size(-1))
|
341 |
v, ix = torch.topk(logits, k, dim=-1)
|
342 |
+
logits = torch.full_like(logits, float("-inf")).scatter(-1, ix, v)
|
|
|
343 |
|
344 |
if top_p and top_p < 1.0:
|
345 |
sl, si = torch.sort(logits, descending=True)
|
|
|
354 |
next_id = torch.multinomial(probs, num_samples=1)
|
355 |
x = torch.cat([x, next_id], dim=1)
|
356 |
nid = next_id.item()
|
357 |
+
if 0 <= nid < V: counts[nid] += 1
|
|
|
358 |
|
359 |
out = tok.decode(x[0].tolist())
|
360 |
return _detok(out) if detokenize else out
|