File size: 12,553 Bytes
9e64e71 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 | """Live training visualization callbacks for GRPO/SFT notebooks."""
from __future__ import annotations
import html as _html
import logging
from typing import Any
_logger = logging.getLogger(__name__)
try:
from transformers import TrainerCallback
except ImportError:
TrainerCallback = object # type: ignore[assignment,misc]
class LiveVisualizationCallback(TrainerCallback):
"""TrainerCallback that plots reward and loss in place during training.
Updates a single plot via IPython display handle without clearing
the cell output.
"""
def __init__(self, **kwargs: Any) -> None:
# Accept and ignore extra kwargs for backward compat
_ = kwargs
self.log_steps: list[int] = []
self.log_rewards: list[float] = []
self.log_losses: list[float] = []
self._plot_handle = None
def on_train_begin(self, args, state, control, **kwargs):
try:
from IPython.display import HTML, display
self._plot_handle = display(
HTML("<em>Waiting for first log...</em>"),
display_id="viz_plot",
)
except Exception:
pass
def on_log(self, args, state, control, logs=None, **kwargs):
if not logs:
return
step = state.global_step
# Find reward (prefer mean)
reward = None
for key in sorted(logs.keys()):
if "reward" in key and "mean" in key:
reward = logs[key]
break
if reward is None:
for key in ("reward", "rewards/mean"):
if key in logs:
reward = logs[key]
break
loss = logs.get("loss")
has_data = False
if reward is not None:
self.log_rewards.append(float(reward))
has_data = True
if loss is not None:
self.log_losses.append(float(loss))
has_data = True
if has_data:
self.log_steps.append(step)
self._update_plot()
def _update_plot(self) -> None:
if self._plot_handle is None:
return
try:
import base64
import io
import matplotlib.pyplot as plt
from IPython.display import HTML
fig, ax = plt.subplots(1, 1, figsize=(8, 3.5))
if self.log_rewards:
ax.plot(
self.log_steps[: len(self.log_rewards)],
self.log_rewards,
"b-o",
markersize=3,
label="Reward",
)
ax.set_ylabel("Reward")
ax.legend(loc="upper left")
if self.log_losses:
# SFT-only: plot loss on primary axis
# GRPO: plot loss on secondary axis
if self.log_rewards:
ax2 = ax.twinx()
ax2.plot(
self.log_steps[: len(self.log_losses)],
self.log_losses,
"r-",
alpha=0.4,
label="Loss",
)
ax2.set_ylabel("Loss", color="r", alpha=0.6)
ax2.legend(loc="upper right")
else:
ax.plot(
self.log_steps[: len(self.log_losses)],
self.log_losses,
"r-o",
markersize=3,
label="Loss",
)
ax.set_ylabel("Loss")
ax.legend(loc="upper right")
ax.set_xlabel("Step")
latest = self.log_steps[-1] if self.log_steps else 0
ax.set_title(f"Training Progress (step {latest})")
ax.grid(True, alpha=0.3)
plt.tight_layout()
buf = io.BytesIO()
fig.savefig(buf, format="png", dpi=100, bbox_inches="tight")
plt.close(fig)
buf.seek(0)
img = base64.b64encode(buf.read()).decode("utf-8")
self._plot_handle.update(HTML(f'<img src="data:image/png;base64,{img}">'))
except Exception as exc:
_logger.debug("Plot update failed: %s", exc)
class SFTMonitorCallback(TrainerCallback):
"""Show sample completions and optional eval accuracy during SFT.
Every ``eval_every_steps`` training steps the callback generates
first-turn completions for a handful of prompts so the user can
watch the model learn tool-calling patterns in real time.
"""
def __init__(
self,
tokenizer: Any,
sample_prompts: list[list[dict[str, str]]],
*,
tools: list[dict] | None = None,
train_dataset: Any = None,
eval_every_steps: int = 50,
max_new_tokens: int = 200,
) -> None:
self.tokenizer = tokenizer
self.sample_prompts = sample_prompts[:3]
self.tools = tools
self.train_dataset = train_dataset
self.eval_every_steps = eval_every_steps
self.max_new_tokens = max_new_tokens
self._model: Any = None
self._display_handle: Any = None
# ------------------------------------------------------------------
def on_train_begin(self, args, state, control, model=None, **kwargs):
self._model = model
try:
from IPython.display import HTML, display
# Always use canonical tools (avoid Dataset serialization artifacts)
tpl_tools = self.tools
# 1) Inference prompt β what the model sees at generation time
if self.sample_prompts:
tpl_kwargs: dict[str, Any] = {
"tokenize": False,
"add_generation_prompt": True,
}
if tpl_tools:
tpl_kwargs["tools"] = tpl_tools
preview = self.tokenizer.apply_chat_template(
self.sample_prompts[0],
**tpl_kwargs,
)
n_tok = len(self.tokenizer.encode(preview))
display(
HTML(
"<details><summary>"
f"<b>Inference prompt</b> ({n_tok} tok)"
" β system + tools + question, "
"as seen by model during GRPO generation"
"</summary>"
"<pre style='background:#2d2d2d;color:#e0e0e0;"
"padding:8px;border-radius:4px;font-size:12px;"
"white-space:pre-wrap;max-height:600px;"
"overflow-y:auto;'>"
f"{_html.escape(preview)}</pre></details>"
)
)
# 2) Training example β one per-turn example from the dataset
if self.train_dataset is not None and len(self.train_dataset) > 0:
row = self.train_dataset[0]
msgs = row.get("messages", [])
if msgs:
ex_kwargs: dict[str, Any] = {"tokenize": False}
if tpl_tools:
ex_kwargs["tools"] = tpl_tools
rendered_ex = self.tokenizer.apply_chat_template(
msgs,
**ex_kwargs,
)
n_ex_tok = len(self.tokenizer.encode(rendered_ex))
n_turns = sum(1 for m in msgs if m.get("role") == "assistant")
last_role = msgs[-1].get("role", "?")
display(
HTML(
"<details><summary>"
f"<b>SFT training example</b>"
f" ({n_ex_tok} tok, {n_turns} asst turn)"
" β history + one assistant tool_call, "
"exactly what the model learns to predict"
"</summary>"
"<pre style='background:#1a1a2e;color:#e0e0e0;"
"padding:8px;border-radius:4px;font-size:12px;"
"white-space:pre-wrap;max-height:600px;"
"overflow-y:auto;'>"
f"{_html.escape(rendered_ex)}</pre>"
f"<p style='color:#888;font-size:11px;'>"
f"Last message role: <b>{last_role}</b> "
f"| Loss is on this turn only</p>"
"</details>"
)
)
self._display_handle = display(
HTML("<em>SFT samples: waiting for first checkpoint...</em>"),
display_id="sft_samples",
)
except Exception:
pass
def on_log(self, args, state, control, logs=None, **kwargs):
step = state.global_step
if step == 0 or step % self.eval_every_steps != 0:
return
if self._model is None:
return
self._generate_and_display(step)
def on_train_end(self, args, state, control, **kwargs):
if self._model is not None:
self._generate_and_display(state.global_step, final=True)
# ------------------------------------------------------------------
def _generate_and_display(self, step: int, final: bool = False) -> None:
try:
import torch
was_training = self._model.training
self._model.eval()
header = "SFT Final Samples" if final else f"SFT Samples (step {step})"
parts = [f"<h4 style='color:#e0e0e0;'>{header}</h4>"]
with torch.no_grad():
for messages in self.sample_prompts:
question = messages[-1]["content"][:100]
tpl_kwargs: dict[str, Any] = {
"tokenize": False,
"add_generation_prompt": True,
}
if self.tools:
tpl_kwargs["tools"] = self.tools
rendered = self.tokenizer.apply_chat_template(
messages,
**tpl_kwargs,
)
inputs = self.tokenizer(rendered, return_tensors="pt")
device = next(self._model.parameters()).device
inputs = {k: v.to(device) for k, v in inputs.items()}
has_tools = "<tools>" in rendered
prompt_len = inputs["input_ids"].shape[1]
out = self._model.generate(
**inputs,
max_new_tokens=self.max_new_tokens,
do_sample=False,
)
new_tokens = out[0][prompt_len:]
raw = self.tokenizer.decode(new_tokens, skip_special_tokens=False)
# Show first tool call only (stop at </tool_call>)
end = raw.find("</tool_call>")
if end != -1:
raw = raw[: end + len("</tool_call>")]
text = self.tokenizer.decode(
self.tokenizer.encode(raw),
skip_special_tokens=True,
).strip()
badge = (
"<span style='color:#4caf50;'>β tools</span>"
if has_tools
else "<span style='color:#f44336;'>β no tools</span>"
)
parts.append(
"<pre style='background:#2d2d2d;color:#e0e0e0;"
"padding:8px;margin:4px 0;border-radius:4px;"
"font-size:13px;line-height:1.4;'>"
f"<b style='color:#82aaff;'>Q:</b> "
f"{_html.escape(question)}"
f" [{badge}, {prompt_len} tok]\n"
f"<b style='color:#c3e88d;'>β</b> "
f"{_html.escape(text)}</pre>"
)
if was_training:
self._model.train()
if self._display_handle is not None:
from IPython.display import HTML
self._display_handle.update(HTML("\n".join(parts)))
except Exception as exc:
_logger.debug("SFT sample generation failed: %s", exc)
|