ViperEkura
754624acf0
- RolloutRunner: generate + score responses with cached re-rollout trigger - BaseStrategy.__call__ switches online/offline via runner injection - GRPO/DPO implement prepare_from_rollout; aliases online_grpo/online_dpo - TrainConfig + train.py add rollout params and CLI flags - Tests cover generate_responses, RolloutRunner cache, shared __call__
273 lines
9.3 KiB
Python
273 lines
9.3 KiB
Python
"""Online rollout runner for RL training.
|
|
|
|
Provides:
|
|
- :class:`RolloutResult` — universal data container for online sampling
|
|
- :class:`BaseRewardModel` — pluggable reward interface
|
|
- :class:`RolloutRunner` — generates + scores batches for any RL strategy
|
|
"""
|
|
|
|
from abc import ABC, abstractmethod
|
|
from dataclasses import dataclass, field
|
|
from typing import Dict, List, Optional
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
from torch import Tensor
|
|
|
|
from astrai.inference.sample import SamplingPipeline
|
|
|
|
|
|
@dataclass
|
|
class RolloutResult:
|
|
"""Universal container produced by :class:`RolloutRunner`.
|
|
|
|
Fields are designed to cover all common RL algorithms:
|
|
GRPO, PPO, Online DPO, Rejection Sampling, etc.
|
|
"""
|
|
|
|
prompts: Tensor
|
|
"""Tokenized prompts, shape ``[B, P_len]``."""
|
|
|
|
responses: Tensor
|
|
"""Generated response token IDs, shape ``[B, G, R_max]``."""
|
|
|
|
response_mask: Tensor
|
|
"""Boolean mask for real (non-pad) response tokens, shape ``[B, G, R_max]``."""
|
|
|
|
rewards: Tensor
|
|
"""Reward per response, shape ``[B, G]``."""
|
|
|
|
logprobs_old: Tensor
|
|
"""Per-token log-probs under the behaviour policy, shape ``[B, G, R_max]``."""
|
|
|
|
prompt_texts: List[str] = field(default_factory=list)
|
|
"""Decoded prompt strings (for reward models that need text)."""
|
|
|
|
response_texts: List[List[str]] = field(default_factory=list)
|
|
"""Decoded response strings, shape ``[B, G]`` (for reward models)."""
|
|
|
|
|
|
class BaseRewardModel(ABC):
|
|
"""Pluggable reward model interface.
|
|
|
|
Subclasses should implement ``score()`` to return a ``[B, G]`` float
|
|
tensor of rewards. Implementations can be:
|
|
* A loaded reward model (e.g. ArmoRM, Skywork-Reward)
|
|
* An external API call
|
|
* A rule-based function (format, length, keyword matching)
|
|
"""
|
|
|
|
@abstractmethod
|
|
def score(self, prompts: List[str], responses: List[List[str]]) -> Tensor:
|
|
"""Score each generated response.
|
|
|
|
Args:
|
|
prompts: Raw prompt strings, length ``B``.
|
|
responses: Generated response strings, shape ``[B, G]``.
|
|
|
|
Returns:
|
|
Float tensor of shape ``[B, G]``.
|
|
"""
|
|
...
|
|
|
|
|
|
def generate_responses(
|
|
model: nn.Module,
|
|
input_ids: Tensor,
|
|
attention_mask: Tensor,
|
|
max_new_tokens: int,
|
|
sampling_pipeline: SamplingPipeline,
|
|
stop_ids: List[int],
|
|
) -> Dict[str, Tensor]:
|
|
"""Autoregressive generation with log-prob tracking.
|
|
|
|
Args:
|
|
model: Policy model (``forward`` returns ``{"logits": ...}``).
|
|
input_ids: ``[B, P_len]`` prompt token IDs.
|
|
attention_mask: ``[B, P_len]`` boolean mask.
|
|
max_new_tokens: Maximum tokens to generate.
|
|
sampling_pipeline: Composed sampling strategies.
|
|
stop_ids: Token IDs that stop generation (eos, etc.).
|
|
|
|
Returns:
|
|
``dict`` with keys:
|
|
- ``generated_ids``: ``[B, max_new_tokens]`` (padded to same length)
|
|
- ``generated_mask``: ``[B, max_new_tokens]``
|
|
- ``logprobs``: ``[B, max_new_tokens]`` per-token log-probs
|
|
"""
|
|
_PAD = 0
|
|
B, P_len = input_ids.shape
|
|
device = input_ids.device
|
|
stop_ids_set = set(stop_ids)
|
|
done = torch.zeros(B, dtype=torch.bool, device=device)
|
|
all_ids = input_ids.clone()
|
|
all_mask = attention_mask.clone()
|
|
logprob_list: List[Tensor] = []
|
|
|
|
for _ in range(max_new_tokens):
|
|
outputs = model(input_ids=all_ids, input_mask=all_mask)
|
|
logits = outputs["logits"][:, -1, :].float()
|
|
log_probs = F.log_softmax(logits, dim=-1)
|
|
|
|
logits = sampling_pipeline.apply(logits, input_ids=all_ids, input_mask=all_mask)
|
|
probs = torch.softmax(logits, dim=-1)
|
|
next_tokens = torch.multinomial(probs, num_samples=1).squeeze(-1)
|
|
|
|
next_tokens[done] = _PAD
|
|
chosen_logprobs = torch.gather(log_probs, -1, next_tokens.unsqueeze(-1))
|
|
logprob_list.append(chosen_logprobs)
|
|
|
|
all_ids = torch.cat([all_ids, next_tokens.unsqueeze(1)], dim=-1)
|
|
all_mask = torch.cat([all_mask, (~done).unsqueeze(1)], dim=-1)
|
|
|
|
done = done | torch.tensor(
|
|
[t.item() in stop_ids_set for t in next_tokens],
|
|
device=device,
|
|
)
|
|
if done.all():
|
|
break
|
|
|
|
logprobs = torch.cat(logprob_list, dim=-1)
|
|
if logprobs.size(1) < max_new_tokens:
|
|
pad_len = max_new_tokens - logprobs.size(1)
|
|
logprobs = F.pad(logprobs, (0, pad_len), value=0.0)
|
|
|
|
generated_ids = all_ids[:, P_len:]
|
|
if generated_ids.size(1) < max_new_tokens:
|
|
pad_len = max_new_tokens - generated_ids.size(1)
|
|
generated_ids = F.pad(generated_ids, (0, pad_len), value=_PAD)
|
|
|
|
generated_mask = generated_ids != _PAD
|
|
|
|
return {
|
|
"generated_ids": generated_ids,
|
|
"generated_mask": generated_mask,
|
|
"logprobs": logprobs,
|
|
}
|
|
|
|
|
|
class RolloutRunner:
|
|
"""Produces :class:`RolloutResult` from a prompt batch.
|
|
|
|
Maintains an internal cache so the same batch prompt can be replayed
|
|
for multiple gradient steps. A new rollout is triggered every
|
|
``rollout_interval`` calls to :meth:`step`.
|
|
|
|
Usage::
|
|
|
|
runner = RolloutRunner(policy, old_policy, tokenizer,
|
|
reward_model, sampling_pipeline, config)
|
|
result = runner(prompt_batch)
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
policy_model: nn.Module,
|
|
old_model: Optional[nn.Module],
|
|
tokenizer,
|
|
reward_model: BaseRewardModel,
|
|
sampling_pipeline: SamplingPipeline,
|
|
max_tokens: int = 1024,
|
|
group_size: int = 8,
|
|
rollout_interval: int = 512,
|
|
):
|
|
self.policy_model = policy_model
|
|
self.old_model = old_model
|
|
self.tokenizer = tokenizer
|
|
self.reward_model = reward_model
|
|
self.sampling_pipeline = sampling_pipeline
|
|
self.max_tokens = max_tokens
|
|
self.group_size = group_size
|
|
self.rollout_interval = rollout_interval
|
|
self.stop_ids = getattr(tokenizer, "stop_ids", []) or []
|
|
|
|
self._cache: Optional[RolloutResult] = None
|
|
self._steps_since_rollout: int = 0
|
|
|
|
def step(self):
|
|
"""Advance the internal counter (call once per optimizer step)."""
|
|
self._steps_since_rollout += 1
|
|
|
|
def clear_cache(self):
|
|
"""Force next call to re-run rollout."""
|
|
self._cache = None
|
|
|
|
def _tokenize_prompts(self, raw_texts: List[str]) -> Dict[str, Tensor]:
|
|
ids_list = self.tokenizer.encode(raw_texts, out_ids=True)
|
|
B = len(ids_list)
|
|
P_max = max(len(ids) for ids in ids_list) if ids_list else 0
|
|
input_ids = torch.zeros(B, P_max, dtype=torch.long)
|
|
for i, ids in enumerate(ids_list):
|
|
input_ids[i, : len(ids)] = torch.tensor(ids[:P_max], dtype=torch.long)
|
|
attention_mask = input_ids != 0
|
|
return {"input_ids": input_ids, "attention_mask": attention_mask}
|
|
|
|
def _decode(self, token_ids: Tensor, mask: Tensor) -> List[List[str]]:
|
|
B, G, _ = token_ids.shape
|
|
texts = []
|
|
for i in range(B):
|
|
group_texts = []
|
|
for g in range(G):
|
|
ids = token_ids[i, g, mask[i, g]].tolist()
|
|
group_texts.append(self.tokenizer.decode(ids, skip_special_tokens=True))
|
|
texts.append(group_texts)
|
|
return texts
|
|
|
|
@torch.no_grad()
|
|
def _run(self, batch: Dict[str, Tensor]) -> RolloutResult:
|
|
"""Execute the actual generation + reward scoring."""
|
|
prompt_ids = batch["input_ids"] if "input_ids" in batch else batch["prompts"]
|
|
prompt_mask = (
|
|
batch["attention_mask"] if "attention_mask" in batch else (prompt_ids != 0)
|
|
)
|
|
B, P_len = prompt_ids.shape
|
|
G = self.group_size
|
|
device = prompt_ids.device
|
|
|
|
prompt_texts: List[str] = []
|
|
for i in range(B):
|
|
ids = prompt_ids[i, prompt_mask[i]].tolist()
|
|
prompt_texts.append(self.tokenizer.decode(ids, skip_special_tokens=True))
|
|
|
|
expanded_ids = prompt_ids.unsqueeze(1).expand(-1, G, -1).reshape(B * G, P_len)
|
|
expanded_mask = prompt_mask.unsqueeze(1).expand(-1, G, -1).reshape(B * G, P_len)
|
|
|
|
gen_out = generate_responses(
|
|
model=self.policy_model,
|
|
input_ids=expanded_ids,
|
|
attention_mask=expanded_mask,
|
|
max_new_tokens=self.max_tokens,
|
|
sampling_pipeline=self.sampling_pipeline,
|
|
stop_ids=self.stop_ids,
|
|
)
|
|
|
|
gen_ids = gen_out["generated_ids"].reshape(B, G, -1)
|
|
gen_mask = gen_out["generated_mask"].reshape(B, G, -1)
|
|
gen_logprobs = gen_out["logprobs"].reshape(B, G, -1)
|
|
|
|
response_texts = self._decode(gen_ids, gen_mask)
|
|
reward_tensor = self.reward_model.score(prompt_texts, response_texts)
|
|
rewards = reward_tensor.to(device=device)
|
|
|
|
return RolloutResult(
|
|
prompts=prompt_ids,
|
|
responses=gen_ids,
|
|
response_mask=gen_mask,
|
|
rewards=rewards,
|
|
logprobs_old=gen_logprobs,
|
|
prompt_texts=prompt_texts,
|
|
response_texts=response_texts,
|
|
)
|
|
|
|
def __call__(self, batch: Dict[str, Tensor]) -> RolloutResult:
|
|
"""Return cached or fresh :class:`RolloutResult`.
|
|
|
|
Triggers a new rollout when ``_steps_since_rollout >= rollout_interval``
|
|
or when the cache is empty.
|
|
"""
|
|
if self._cache is None or self._steps_since_rollout >= self.rollout_interval:
|
|
self._cache = self._run(batch)
|
|
self._steps_since_rollout = 0
|
|
return self._cache
|