torchrl-nightly 2025.6.20__cp313-cp313-win_amd64.whl → 2025.6.22__cp313-cp313-win_amd64.whl

torchrl/objectives/llm/sft.py

@@ -0,0 +1,465 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+from __future__ import annotations
+
+import contextlib
+import warnings
+
+from dataclasses import dataclass
+from typing import Literal
+
+import torch
+from tensordict import NestedKey, TensorClass, TensorDictBase
+from tensordict.nn import TensorDictModule
+from tensordict.utils import _zip_strict
+from torchrl.data import History
+from torchrl.modules.llm.policies.transformers_wrapper import TransformersWrapper
+from torchrl.objectives.common import LossModule
+
+
+def sft_loss(summed_log_probs: torch.Tensor, reduction: str) -> torch.Tensor:
+    """Compute the SFT loss."""
+    if reduction == "mean":
+        loss = -summed_log_probs.mean()
+    elif reduction == "sum":
+        loss = -summed_log_probs.sum()
+    elif reduction == "none":
+        loss = -summed_log_probs
+    else:
+        raise ValueError(f"Invalid reduction: {reduction}.")
+    return loss
+
+
+def minor_sft_loss(
+    log_probs: torch.Tensor,
+    ref_log_probs: torch.Tensor,
+    beta: float,
+    reduction: str,
+) -> torch.Tensor:
+    """Compute the MinorSFT loss.
+
+    This loss is inspired by DPO and is designed to be less aggressive than standard SFT.
+    It computes ``-log_sigmoid(beta * (log_probs - ref_log_probs))``.
+
+    Args:
+        log_probs (torch.Tensor): The log probabilities from the model being trained.
+        ref_log_probs (torch.Tensor): The log probabilities from the reference model.
+        beta (float): The beta parameter from DPO.
+        reduction (str): The reduction to apply to the loss.
+
+    Returns:
+        The MinorSFT loss.
+
+    References:
+        - Shiming Xie, Hong Chen, Fred Yu, Zeye Sun, Xiuyu Wu, 2024.
+          `"Minor SFT loss for LLM fine-tune to increase performance and reduce model deviation" <https://arxiv.org/abs/2408.10642>`_
+    """
+    if log_probs.shape != ref_log_probs.shape:
+        raise ValueError(
+            f"Current log probabilities and reference log probabilities have different shapes: {log_probs.shape=} vs {ref_log_probs.shape=}."
+        )
+    loss = -torch.nn.functional.logsigmoid(beta * (log_probs - ref_log_probs))
+    if reduction == "mean":
+        return loss.mean()
+    if reduction == "sum":
+        return loss.sum()
+    if reduction == "none":
+        return loss
+    raise ValueError(f"Invalid reduction: {reduction}")
+
+
+class SFTLossOutput(TensorClass["nocast"]):
+    """SFT Loss Output.
+
+    Attributes:
+        loss_sft (torch.Tensor): The loss for the SFT objective.
+        loss_kl_to_ref (torch.Tensor | None): The loss for the KL divergence to the reference model.
+        kl_to_ref (torch.Tensor | None): The KL divergence to the reference model.
+
+    .. note::
+        The loss components are kept separate to allow for logging and visualization.
+        Before backpropagation, the loss components are to be summed together. Since non-loss components are not differentiable
+        when the loss is constructed via :class:`~torchrl.objectives.llm.sft.SFTLoss`, summing
+        the :class:`~torchrl.objectives.llm.sft.SFTLossOutput` directly is a proper way of obtaining the total loss.
+
+            >>> loss_fn = SFTLoss(...)
+            >>> loss_output = loss_fn(td)
+            >>> loss = loss_output.loss_sft + loss_output.loss_kl_to_ref
+            >>> loss.backward()
+            >>> # or equivalently
+            >>> loss = loss_fn(td)
+            >>> loss.sum(reduce=True).backward()
+    """
+
+    loss_sft: torch.Tensor
+    loss_kl_to_ref: torch.Tensor | None = None
+    kl_to_ref: torch.Tensor | None = None
+
+
+class SFTLoss(LossModule):
+    r"""Supervised fine-tuning loss.
+
+    Args:
+        actor_network (TensorDictModule): the actor network. Usually a :class:`~torchrl.modules.llm.TransformersWrapper` instance,
+            with `return_log_prob=True` and `from_text=True`.
+        tokenizer (`Tokenizer`): the tokenizer to be used to tokenize the input and compute the assitant mask. If not provided, the tokenizer will be inferred from the `actor_network`.
+        tokenizer_kwargs (dict, optional): keyword arguments to pass to the tokenizer during :meth:`~torchrl.data.llm.chat.History.apply_chat_template`.
+            This can be used to override arguments such as the `chat_template` or `chat_template_name`.
+        reduction (Literal["mean", "sum", "none"], optional): the reduction to apply to the loss. Defaults to `"mean"`.
+        normalize_by_seq_length (bool, optional): whether to normalize the loss by the sequence length. Defaults to `True`.
+        kl_to_ref_coeff (float | None, optional): coefficient for KL divergence to reference model. Defaults to `None`.
+        loss_function (Literal["sft", "minor_sft"], optional): The loss function to use. Defaults to `"sft"`.
+        beta (float, optional): The beta parameter for MinorSFT loss. This is only used when `loss_function` is `"minor_sft"`.
+            Higher values of beta make the loss more aggressive (pushes the model to generate responses further from the reference model):
+
+            .. math::
+                \text{loss} = -\log\sigma(\beta \cdot (\text{log_probs} - \text{ref_log_probs}))
+
+            Defaults to `0.1`.
+        device (torch.device | None, optional): the device to use for the loss, when tokenizing the input. Defaults to `None`.
+
+    .. note::
+        The input tensordict is expected to contain the following keys by default:
+            - ``("next", "history")``: The chat history
+            - ``("next", "ref_log_prob")`` (optional): Reference model log probabilities, required if kl_to_ref_coeff is set
+
+        These keys can be customized using the ``set_keys()`` method.
+
+    .. seealso:: :class:`~torchrl.envs.llm.transforms.RetrieveLogProb` for the KL divergence computation.
+
+    References:
+        - Shiming Xie, Hong Chen, Fred Yu, Zeye Sun, Xiuyu Wu, 2024.
+          `"Minor SFT loss for LLM fine-tune to increase performance and reduce model deviation" <https://arxiv.org/abs/2408.10642>`_
+
+    Examples:
+        >>> from torchrl.data.llm.chat import History, _CHAT_TEMPLATES
+        >>> from torchrl.modules.llm import TransformersWrapper
+        >>> from torchrl.objectives.llm.sft import SFTLoss
+        >>> from transformers import AutoTokenizer, OPTConfig, OPTForCausalLM
+        >>> from tensordict import TensorDict, lazy_stack
+        >>> import torch
+        >>>
+        >>> # Create chat data
+        >>> chats = [
+        ...     [
+        ...         {"role": "system", "content": "You are a helpful assistant."},
+        ...         {"role": "user", "content": "Hello, how are you?"},
+        ...         {"role": "assistant", "content": "I'm doing well, thank you!"},
+        ...     ],
+        ...     [
+        ...         {"role": "system", "content": "You are a helpful assistant."},
+        ...         {"role": "user", "content": "What's the weather like?"},
+        ...         {"role": "assistant", "content": "I can't check the weather for you."},
+        ...     ],
+        ... ]
+        >>> history = History.from_chats(chats)
+        >>>
+        >>> # Setup tokenizer and model
+        >>> tokenizer = AutoTokenizer.from_pretrained("facebook/opt-125m")
+        >>> tokenizer.pad_token = tokenizer.eos_token
+        >>> tokenizer.chat_template = _CHAT_TEMPLATES["chatml_format"]
+        >>> model = OPTForCausalLM(OPTConfig()).eval()
+        >>>
+        >>> # Create training and reference policies
+        >>> policy_train = TransformersWrapper(
+        ...     model,
+        ...     tokenizer=tokenizer,
+        ...     generate=False,
+        ...     from_text=True,
+        ...     chat_template_name="qwen",
+        ... )
+        >>> policy_ref = TransformersWrapper(
+        ...     model,
+        ...     tokenizer=tokenizer,
+        ...     generate=False,
+        ...     from_text=True,
+        ...     return_log_probs=True,
+        ...     chat_template_name="qwen",
+        ... )
+        >>>
+        >>> # Create the RetrieveLogProb transform
+        >>> transform = RetrieveLogProb(
+        ...     policy_ref,
+        ...     assistant_only=True,
+        ...     tokenizer_kwargs={"chat_template_name": "qwen"},
+        ...     tokenizer=tokenizer,
+        ... )
+        >>>
+        >>> # Prepare data
+        >>> text = history[:, :-1].apply_chat_template(
+        ...     tokenizer=tokenizer, chat_template_name="qwen", add_generation_prompt=True
+        ... )
+        >>> text_response = history.apply_chat_template(
+        ...     tokenizer=tokenizer, chat_template_name="qwen", add_generation_prompt=False
+        ... )
+        >>> text_response = [
+        ...     txt[len(txt_start):] for txt, txt_start in zip(text_response, text)
+        ... ]
+        >>> td = TensorDict(
+        ...     text=text,
+        ...     text_response=text_response,
+        ...     history=history,
+        ...     next=TensorDict(
+        ...         reward=torch.randn(2, 1),
+        ...         done=torch.zeros(2, dtype=torch.bool),
+        ...         history=history,
+        ...     ),
+        ...     batch_size=(2,),
+        ... )
+        >>> data = lazy_stack(list(td.unbind(0)))
+        >>>
+        >>> # Apply the transform to get reference log probabilities
+        >>> data = transform(data)
+        >>> assert "ref_log_prob" in data["next"].keys()
+        >>>
+        >>> # Use with SFTLoss for KL regularization
+        >>> loss = SFTLoss(
+        ...     actor_network=policy_train,
+        ...     tokenizer=tokenizer,
+        ...     reduction="mean",
+        ...     normalize_by_seq_length=True,
+        ...     kl_to_ref_coeff=0.1,
+        ...     tokenizer_kwargs={"chat_template_name": "qwen"},
+        ...     loss_function="sft",
+        ... )
+        >>> loss_vals = loss(data)
+        >>> print(f"SFT Loss: {loss_vals.loss_sft.item():.4f}")
+        >>> print(f"KL to Reference Loss: {loss_vals.loss_kl_to_ref.item():.4f}")
+
+    """
+
+    @dataclass
+    class _AcceptedKeys:
+        """Maintains default values for all configurable tensordict keys.
+
+        This class defines which tensordict keys can be set using '.set_keys(key_name=key_value)' and their
+        default values.
+
+        Attributes:
+            history (NestedKey): The input tensordict key where the chat history is expected.
+                Defaults to ``("next", "history")``.
+            ref_log_prob (NestedKey): The input tensordict key where the reference model log probabilities are expected.
+                Only used when kl_to_ref_coeff is set. Defaults to ``("next", "ref_log_prob")``.
+            log_probs (NestedKey): The output tensordict key where the model's log probabilities will be written.
+                Defaults to ``"log_probs"``.
+        """
+
+        history: NestedKey = ("next", "history")
+        ref_log_prob: NestedKey = ("next", "ref_log_prob")
+        log_probs: NestedKey = "log_probs"
+
+    default_keys = _AcceptedKeys
+    tensor_keys: _AcceptedKeys
+
+    def __init__(
+        self,
+        actor_network: TensorDictModule | TransformersWrapper,
+        tokenizer: transformers.AutoTokenizer | None = None,  # noqa: F821
+        tokenizer_kwargs: dict | None = None,
+        reduction: Literal["mean", "sum", "none"] = "mean",
+        normalize_by_seq_length: bool = True,
+        kl_to_ref_coeff: float | None = None,
+        loss_function: Literal["sft", "minor_sft"] = "sft",
+        beta: float = 0.1,
+        device: torch.device | None = None,
+    ):
+        super().__init__()
+        self.in_keys = []
+        self.actor_network = actor_network
+        if tokenizer is None:
+            tokenizer = actor_network.tokenizer
+        self.tokenizer = tokenizer
+        if tokenizer_kwargs is None:
+            tokenizer_kwargs = {}
+        if tokenizer is None:
+            raise ValueError("Tokenizer must be provided.")
+        tokenizer_kwargs.setdefault("return_assistant_tokens_mask", True)
+        tokenizer_kwargs.setdefault("tokenize", True)
+        tokenizer_kwargs.setdefault("return_tensors", "pt")
+        tokenizer_kwargs.setdefault("padding", False)
+        tokenizer_kwargs.setdefault("add_generation_prompt", False)
+        self.tokenizer_kwargs = tokenizer_kwargs
+        self.reduction = reduction
+        self.normalize_by_seq_length = normalize_by_seq_length
+        self.kl_to_ref_coeff = kl_to_ref_coeff
+        self.loss_function = loss_function
+        if self.loss_function == "minor_sft" and kl_to_ref_coeff:
+            warnings.warn(
+                "kl_to_ref_coeff should not be set when using minor_sft loss, as KL regularization is implicit. Setting kl_to_ref_coeff to 0.0."
+            )
+            self.kl_to_ref_coeff = 0.0
+        self.beta = beta
+        self._set_in_keys()
+        self.device = device
+
+    def _set_in_keys(self) -> None:
+        """Sets the input keys for the loss module."""
+        in_keys = [self.tensor_keys.history]
+        if self.kl_to_ref_coeff is not None or self.loss_function == "minor_sft":
+            in_keys.append(self.tensor_keys.ref_log_prob)
+        self.in_keys = in_keys
+        self.out_keys = []  # Loss modules typically don't have out_keys
+
+    def _kl_to_ref(
+        self,
+        cur_log_prob: list[torch.Tensor],
+        ref_log_prob: list[torch.Tensor],
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        """Compute KL divergence to reference model.
+
+        Args:
+            cur_log_prob (List[torch.Tensor]): Log probabilities from current model. Must have shape [T] where T is the number of tokens in the assistant response.
+            ref_log_prob (List[torch.Tensor]): Log probabilities from reference model. Must have shape [T] where T is the number of tokens in the assistant response.
+
+        Returns:
+            tuple[torch.Tensor, torch.Tensor]: (KL loss term, KL penalty for logging)
+        """
+        # Apply mask
+        ref_log_prob = torch.cat(ref_log_prob)
+        cur_log_prob = torch.cat(cur_log_prob)
+        # ref_log_prob = ref_log_prob[mask]
+        # cur_log_prob = cur_log_prob[mask].squeeze()
+        if cur_log_prob.shape != ref_log_prob.shape:
+            raise ValueError(
+                f"Current log probabilities and reference log probabilities have different shapes: {cur_log_prob.shape=} vs {ref_log_prob.shape=}."
+            )
+        # Compute KL using same approximation as GRPO
+        diff = ref_log_prob - cur_log_prob
+
+        kl_penalty = (diff.expm1() - diff).mean()
+        return self.kl_to_ref_coeff * kl_penalty, kl_penalty
+
+    def forward(self, tensordict: TensorDictBase) -> TensorDictBase:
+        # Gather history
+        history: History = tensordict[self.tensor_keys.history]
+
+        # Apply tokenizer to history and gather mask
+        with torch.device(
+            self.device
+        ) if self.device is not None else contextlib.nullcontext():
+            token_struct = history.apply_chat_template(
+                tokenizer=self.tokenizer, **self.tokenizer_kwargs
+            )
+        if "assistant_masks" not in token_struct:
+            raise ValueError(
+                f"Assistant masks are not present in the token structure: {token_struct=}."
+            )
+        assistant_masks = token_struct.get(
+            "assistant_masks",
+            as_list=True,
+        )
+        assistant_masks = [mask.bool() for mask in assistant_masks]
+        attention_mask = token_struct.get("attention_mask", as_list=True)
+        attention_mask = [mask.bool() for mask in attention_mask]
+        assistant_masks = [
+            mask & a_mask for mask, a_mask in zip(assistant_masks, attention_mask)
+        ]
+
+        if not any(mask.any(-1).all() for mask in assistant_masks):
+            raise ValueError("Some inputs have no valid assistant masks.")
+        input_loss = tensordict.select(self.tensor_keys.history)
+        if (
+            isinstance(self.tensor_keys.history, tuple)
+            and self.tensor_keys.history[0] == "next"
+        ):
+            input_loss = input_loss["next"]
+
+        with torch.device(
+            self.device
+        ) if self.device is not None else contextlib.nullcontext():
+            output_loss = self.actor_network(input_loss)
+
+        # get log-probs
+        log_probs = output_loss.get(
+            self.tensor_keys.log_probs,
+            as_list=True,
+        )
+        # apply mask
+        if not all(
+            mask.shape == lp.shape
+            for mask, lp in _zip_strict(assistant_masks, log_probs)
+        ):
+            raise ValueError(
+                f"Assistant masks and log_probs have different shapes: {[mask.shape for mask in assistant_masks]} vs {[lp.shape for lp in log_probs]}. Tokens from current template: {[inp.shape for inp in token_struct.get('input_ids', as_padded_tensor=True)]}"
+            )
+
+        log_probs_masked = [
+            lp.masked_fill(~mask, 0.0)
+            for lp, mask in _zip_strict(log_probs, assistant_masks)
+        ]
+
+        # Sum log probs, optionally normalize by sequence length
+        summed_log_probs = torch.stack(
+            [lp.sum(tensordict.ndim - 1) for lp in log_probs_masked]
+        )
+        seq_lengths = torch.stack(
+            [mask.sum(tensordict.ndim - 1) for mask in assistant_masks]
+        )
+        if self.normalize_by_seq_length:
+            # Compute sequence lengths for normalization (number of assistant tokens)
+            summed_log_probs = summed_log_probs / seq_lengths.clamp(min=1)
+
+        # Compute main loss
+        if self.loss_function == "sft":
+            loss = sft_loss(summed_log_probs, self.reduction)
+            # Add KL divergence loss if reference model is provided
+            if self.kl_to_ref_coeff is not None:
+                ref_log_probs = tensordict.get(
+                    self.tensor_keys.ref_log_prob,
+                    default=None,
+                    as_list=True,
+                )
+                if ref_log_probs is None:
+                    raise ValueError(
+                        "Reference log probs not found in tensordict but kl_to_ref_coeff was set"
+                    )
+
+                loss_kl, kl_penalty = self._kl_to_ref(
+                    [lp[mask] for lp, mask in _zip_strict(log_probs, assistant_masks)],
+                    ref_log_probs,
+                )
+                output = SFTLossOutput(
+                    loss_sft=loss,
+                    loss_kl_to_ref=loss_kl,
+                    kl_to_ref=kl_penalty.detach(),
+                )
+            else:
+                output = SFTLossOutput(loss_sft=loss)
+        elif self.loss_function == "minor_sft":
+            ref_log_probs = tensordict.get(self.tensor_keys.ref_log_prob, as_list=True)
+            if ref_log_probs is None:
+                raise ValueError(
+                    f"Reference log probs not found at {self.tensor_keys.ref_log_prob=} in tensordict but loss_function is 'minor_sft'"
+                )
+
+            # we need to re-sum ref_log_probs as they are not summed per-sequence
+            summed_ref_log_probs = torch.stack([lp.sum() for lp in ref_log_probs]).to(
+                summed_log_probs.device
+            )
+            if self.normalize_by_seq_length:
+                summed_ref_log_probs = summed_ref_log_probs / seq_lengths.clamp(min=1)
+            loss = minor_sft_loss(
+                summed_log_probs, summed_ref_log_probs, self.beta, self.reduction
+            )
+            if self.kl_to_ref_coeff is not None:
+                with torch.no_grad():
+                    loss_kl, kl_penalty = self._kl_to_ref(
+                        [
+                            lp[mask]
+                            for lp, mask in _zip_strict(log_probs, assistant_masks)
+                        ],
+                        ref_log_probs,
+                    )
+                output = SFTLossOutput(
+                    loss_sft=loss,
+                    loss_kl_to_ref=loss_kl,
+                    kl_to_ref=kl_penalty.detach(),
+                )
+            else:
+                output = SFTLossOutput(loss_sft=loss)
+        else:
+            raise ValueError(f"Invalid loss function: {self.loss_function}")
+
+        return output

torchrl/version.py

@@ -1,2 +1,2 @@
-__version__ = '2025.6.20'
-git_version = 'e1e15d692a6df69bfdc80e85f45d37a4b967e625'
+__version__ = '2025.6.22'
+git_version = '77dbc6c9ffbce3d2ce3f26b659355cd46d8132c3'

{torchrl_nightly-2025.6.20.dist-info → torchrl_nightly-2025.6.22.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: torchrl-nightly
-Version: 2025.6.20
+Version: 2025.6.22
 Home-page: https://github.com/pytorch/rl
 Author: torchrl contributors
 Author-email: vmoens@fb.com

{torchrl_nightly-2025.6.20.dist-info → torchrl_nightly-2025.6.22.dist-info}/RECORD

@@ -3,11 +3,11 @@ build_tools/setup_helpers/__init__.py,sha256=l9zlK7Nm5bT7P_onQx-hZeIGzKKyCFm1PFk
 build_tools/setup_helpers/extension.py,sha256=ihV8jz8kqOvpqzuD006XqF1oNX5ukKGlwIOJRb1Vd-o,6075
 torchrl/__init__.py,sha256=76lKYwYKmAKORhyVt2tURmYAIRTifxxO3gWsskrHAXU,3054
 torchrl/_extension.py,sha256=x6Nqj2brF3VhlEwxmNA2fYbmpxq1HHGrHMnP0YnQwdc,2412
-torchrl/_torchrl.cp313-win_amd64.pyd,sha256=QAhUsKZIu8t4j0qCMXilS-3HZ_NP92sOLAVDTzizUkw,429056
+torchrl/_torchrl.cp313-win_amd64.pyd,sha256=CP0G1g9_FfwAaXAXkaznbN0USlRSkpOLIZeC6_V-QZE,429056
 torchrl/_utils.py,sha256=2N35rdD65U1khMi5gVIz8-nMjlZsoVq0kCiQftVRSxw,42297
-torchrl/version.py,sha256=HnpF56pzVs2b-GMFYIzT-B7z9-9fXZZkDTrkv-8dp-M,85
+torchrl/version.py,sha256=sFWjKZnDuj1wpE9ARKNe8cTHlKoOAExwF3ZTthG-Y-U,85
 torchrl/collectors/__init__.py,sha256=LzTyfxmkNGPSa5-3rS5unQK7HfT5ZEdr2NV291rAOlU,832
-torchrl/collectors/collectors.py,sha256=nWwPcWWdeEyj3Fe0bdFqN69Yw_jsjPhD49Iwc8H61D0,181298
+torchrl/collectors/collectors.py,sha256=Pz6VYYrekjBiVBQiyzp6zIyZrBdjSv4-FqlfrGYQz3E,181469
 torchrl/collectors/utils.py,sha256=aBmBLpphhfplqQjRCyn1jtWWJ-Wtc7TWvM0rOBN8SsE,11579
 torchrl/collectors/weight_update.py,sha256=Ydq5nJSTV3Q1uqLtJ_1Nj1JB5rwHwrG5StaLxymWFV4,21572
 torchrl/collectors/distributed/__init__.py,sha256=cKDWdNlwx2LoJkTwf-DKUXbq3Y-0Z1DctPYPcdgOSU0,730
@@ -18,12 +18,12 @@ torchrl/collectors/distributed/rpc.py,sha256=xta5tptC0mLlIY_AecLrARvFBYh7nMbclrr
 torchrl/collectors/distributed/sync.py,sha256=zjp0HEEcSMaDzq8xndoBWvyqYCdf7hp8urYRRiJP2GI,27912
 torchrl/collectors/distributed/utils.py,sha256=eY6M-vLCSzyACHRNBx5bHcieWsZfLg7DfNKGIv0IgHI,6625
 torchrl/collectors/llm/__init__.py,sha256=u03aQ97C3sb5-C0-s2tMBAGGs3kJTfZUSse29fHDkIk,365
-torchrl/collectors/llm/base.py,sha256=fs7YcEsGF8WtuS1my_4dyzAk94Yw4J25NnZxwnB7AzE,20842
-torchrl/collectors/llm/ray_collector.py,sha256=lxtHFkgmnMCNIUa9E3iDmpQfTs8tCuuZmydF6SJcnnM,11208
+torchrl/collectors/llm/base.py,sha256=Wxdo4drsMk_i5u5DzDlikY4j5-TM9f6Ac4xjB6wJgPw,21132
+torchrl/collectors/llm/ray_collector.py,sha256=nk-i61ZAsYkDNZW2Y7vcDhddjkUyyIuiDYjU9iWYklE,11364
 torchrl/collectors/llm/utils.py,sha256=GnDY2cTu4XEdwqqhFCP4QWfS2tsgaLTy8nwpIaTEsQI,1184
 torchrl/collectors/llm/weight_update/__init__.py,sha256=ngbL_sPfXh8FMM3r_j0B9QEP_jQIVSOa8pZVouHg9ec,281
 torchrl/collectors/llm/weight_update/vllm.py,sha256=4kRlEBHb6093d9lkKVIqU8ZwiPoCFtmVVaADuhxKLL4,11571
-torchrl/data/__init__.py,sha256=LEjVJd1OcHNz60lnoDMwoFtp-ejbqhWhhkydH_bgCfE,4974
+torchrl/data/__init__.py,sha256=h6ZHGWzvWDfyw6tgo69gld5C_Kgg-T4DP-EVv1hy0Xk,5026
 torchrl/data/rlhf.py,sha256=_ENSvNe84snnFQG0jlTtOI419nIYtbBHvAw-pdFMiSs,1002
 torchrl/data/tensor_specs.py,sha256=PF-sta3dHy0UaDu26FS5-ZVlpAsagTUUMLmYHs5PQhk,254762
 torchrl/data/utils.py,sha256=tXBPxl5VHqPUfJF1VLqURmb066zDd9lipRDER4R1FY8,12444
@@ -39,12 +39,13 @@ torchrl/data/datasets/openx.py,sha256=0p2H3phnvsUgrFFfQulTyGEDcUrvXiFp-p18uFk8Xk
 torchrl/data/datasets/roboset.py,sha256=sLdDknyPj7f2NF5z5EbzEQ9QhD03VBKHEXXo_wxSRc4,17011
 torchrl/data/datasets/utils.py,sha256=tRZkarWl-BX_fnGEVNRt6Fjo0wmyKCNAuVmw5Le-0C4,350
 torchrl/data/datasets/vd4rl.py,sha256=YFjXvP-QGNzF7UWNKkGMKPFthcB0I8v6sJc7oESegYA,18694
-torchrl/data/llm/__init__.py,sha256=ZWoisCgT7JbAAHgcyrlp8o_q2mhgrgU4k-0kikLsTh8,941
-torchrl/data/llm/chat.py,sha256=z-PujCw0tVZXUktQqf8NUevKCSjL56_qKiouyse6nVo,31554
+torchrl/data/llm/__init__.py,sha256=X86bYW_uNAwKJcxK2AVQpJo56tDejtXSDFtOyZMNhHQ,1006
+torchrl/data/llm/chat.py,sha256=akIL6wj2QtpdWlFvOcRfPcrk7HE5NoSpgZxB8gtdOos,33962
 torchrl/data/llm/common.py,sha256=3Gb8sMojtNss6wi6hKGSIUDAwK1PC_8ve9W-bHCPjAk,2181
 torchrl/data/llm/dataset.py,sha256=GDuzflBq2ThgYn_V4bOr_1MHOhESnQ5jX1Wlcw69lfM,21194
 torchrl/data/llm/prompt.py,sha256=ikHWafhTIoCONCpuMHwIuGfpnPSpg5drZQTxTCmygQE,8578
 torchrl/data/llm/reward.py,sha256=QW1HWpNRORd3InwWLg-hAhjTlPqX4ffzAkYHEz0jQxo,8629
+torchrl/data/llm/topk.py,sha256=4MTxYTTdfSBM5vxDHnleY7FatanzHgneXs8rjgKwwqQ,8539
 torchrl/data/llm/utils.py,sha256=K2NQoEhBC6VWowsMeDHu2Q8vbg3ZPEWBBN6z4qifiNM,24143
 torchrl/data/map/__init__.py,sha256=bON0vqCksU7FPoWNqiNcdl60t7yWUh9SdLhNtglj7jI,576
 torchrl/data/map/hash.py,sha256=XRYdaFHQUm87fL9pWjhvi2LeZVaqJsASkCU-G_Gus8s,7437
@@ -56,11 +57,11 @@ torchrl/data/postprocs/__init__.py,sha256=fOyX5OMaDb5HGrQbn9W72_QwncNdh6l3DkVSqR
 torchrl/data/postprocs/postprocs.py,sha256=dpXOKWlhdKy4Um7HdzRKe42PJ_Q1jHC7AX5plR9AIiw,15509
 torchrl/data/replay_buffers/__init__.py,sha256=oINoSWKO3Ku6YIBF-0KnbVLZwelZbANN4nLU4q6Mir0,2455
 torchrl/data/replay_buffers/checkpointers.py,sha256=eizAw4W0tQ2EWgfx6-EUV_3EuZMcZVdoahLoux15Nr4,15186
-torchrl/data/replay_buffers/ray_buffer.py,sha256=_8ZFBCJKgEcPWSjMoMGoJzUfN5eF9UD79dXK6Lh7dhc,9310
-torchrl/data/replay_buffers/replay_buffers.py,sha256=fO9SyUjNl-VUNch3FkwVKIRPkprAcypIcpbQpvCleUI,90708
-torchrl/data/replay_buffers/samplers.py,sha256=tBDi6GN2oe9WrJpnbqHmAAkNT0Du0Ce7Ldw9wWZ_7Jk,107888
+torchrl/data/replay_buffers/ray_buffer.py,sha256=p8EkiXOP4EVMkkpjOyje7wiBfgbWOB0xhEJzDdelAxc,10135
+torchrl/data/replay_buffers/replay_buffers.py,sha256=YzNV543zDpvENbUnqjjghkHHq6IwyRms1I3DXl-ayq4,92567
+torchrl/data/replay_buffers/samplers.py,sha256=bFP8j3BahHULASAjeIGtxJX36GRsg3yBCNFu7vR6Zdo,112834
 torchrl/data/replay_buffers/scheduler.py,sha256=cGm4LZcZ2lo8azDMWKGTdhWApxjZFh0KfynApxAkVK4,10416
-torchrl/data/replay_buffers/storages.py,sha256=bP_pak6fi8X57OkXlQnbMX-ze7-VilRYKOTPr8KGkuA,61146
+torchrl/data/replay_buffers/storages.py,sha256=VdEYOQ29FWGwDeHfLafZuobMlmuagxHVHnntKnU-yX4,62298
 torchrl/data/replay_buffers/utils.py,sha256=vlGfyHVKUAMKBR0l7fJM9NI47ZinS18Qzf8lpwoo6pI,39644
 torchrl/data/replay_buffers/writers.py,sha256=-aI6Y28oisuFDutMVlPp4e8wTe6x0wlY0MY1OUKHl4Q,28466
 torchrl/envs/__init__.py,sha256=2eVr8StUSMiNd-IoD5BQAFFuV10pAtO926b6QzRzB_M,6082
@@ -96,8 +97,8 @@ torchrl/envs/libs/smacv2.py,sha256=pr03oGHE2G_fc86qHeSQjSz3S6IH_l2hX0J2umb020M,2
 torchrl/envs/libs/unity_mlagents.py,sha256=vszCYjEX0S9AmIwLvGsoqc0Jr7jvlBAqZ1HQ1uqesjM,50558
 torchrl/envs/libs/utils.py,sha256=Ce8nAYc2MQOBTYCV17Yswk98pg3PStnaGPFVW2jqARQ,5354
 torchrl/envs/libs/vmas.py,sha256=giTORg2AqYzyjrazdD94fD2dNYwX7qe5TFnr-E1mjIg,37140
-torchrl/envs/llm/__init__.py,sha256=_srpJ1x42TqTbsMvJpTObVeyAlRhvG6XHpbRqk8qYi4,1274
-torchrl/envs/llm/chat.py,sha256=2ZN5fef0cYUGwLPHbDK3YNb9bI2pewIApQKzong-LKA,17993
+torchrl/envs/llm/__init__.py,sha256=Iz5HtLoVy8O4u1mrPmyql4G8SU9S-MCinP_Gh8sbUWo,1320
+torchrl/envs/llm/chat.py,sha256=YvADxo11RKkjD06rBvbbljch3Jb_H4snaBcgkU2Q-7w,18171
 torchrl/envs/llm/envs.py,sha256=wphbzLwDKYO_OTV63WYW4iTK5Ek4vmb1zNv5gehzodY,35450
 torchrl/envs/llm/datasets/__init__.py,sha256=6-x0WlKD7lpMVLKA4W1AktvgUs6adMuaGAqYYhgQ_hk,490
 torchrl/envs/llm/datasets/gsm8k.py,sha256=MfCFu0U7uetDtLdzUdvqX4rENXPsL8msnkMT98Q29jE,15624
@@ -105,18 +106,18 @@ torchrl/envs/llm/datasets/ifeval.py,sha256=sJ4bvXEWBzzNnDDbKkj6yz_1zemDStVFNfxqo
 torchrl/envs/llm/libs/__init__.py,sha256=zvUe6oe3pjZwGefV-_x4MAC6K89TMqxh3TZs5s3ADkI,274
 torchrl/envs/llm/libs/mlgym.py,sha256=TMaoV9P5w5EGgBSmLiw42_DOyKEh7ZGf3mDf-LaZ9W0,32237
 torchrl/envs/llm/reward/__init__.py,sha256=KYNJxyDOe2mZkjyH4CSuQ8qM0_Zu3EAaIGocYhLduPQ,380
-torchrl/envs/llm/reward/gsm8k.py,sha256=gkDQu0Xa8r4PEyuhH146M3dnlJbI7XAVzThIt3Opm1M,7899
+torchrl/envs/llm/reward/gsm8k.py,sha256=TW2lACMLXHRlcTTRfTcFTLl7NIJA0TNh6qfSJiC52QI,8066
 torchrl/envs/llm/reward/ifeval/__init__.py,sha256=vvh7JSUQaEiMjNeMeJvWlcFb2-6_J1LfM6l4mENn4Zg,324
 torchrl/envs/llm/reward/ifeval/_instructions.py,sha256=jlNvIO3dykk8fBFXC35PQSDJ9vLF3knS-ywq2ILfF00,63362
 torchrl/envs/llm/reward/ifeval/_instructions_main.py,sha256=DEc7QqfujGxYvqcm2y_zPasBqB7FgSfXRt3QQ4HQUz0,4244
 torchrl/envs/llm/reward/ifeval/_instructions_registry.py,sha256=bY8R51RgjKJYiim67j5IXSfYhtWtvZrRF61yuqi0Tzs,3914
 torchrl/envs/llm/reward/ifeval/_instructions_util.py,sha256=63ZJbqUKaqMA_SDhnYT7VppULbible8udHGihiamxKc,27719
 torchrl/envs/llm/reward/ifeval/_scorer.py,sha256=iv-316dBYlz4fz6WUtzP7151y4xEwuwOq-Wf7Qazgmc,14928
-torchrl/envs/llm/transforms/__init__.py,sha256=P6-KVptxmsweLvwmnTl5dkfLS9JNAizwFPWOl8Hr3So,773
+torchrl/envs/llm/transforms/__init__.py,sha256=BnVW7WVCYPlaNPd4cEyXIUI1Qfd7YQkZVsXvj5q5UVw,814
 torchrl/envs/llm/transforms/browser.py,sha256=d0JIUZ3TfgmqBcci5ihzzTqZA9KeTrs1iCProRWQQK8,10715
 torchrl/envs/llm/transforms/dataloading.py,sha256=Zl--I6bT2AqWDZmM6RMQq4ds3b1PFGqilAaIbXBJuNc,25054
 torchrl/envs/llm/transforms/format.py,sha256=tME390wkG0h2V5DAWHZa7EhJ5Or-6cga6AIjxPuy1l8,2592
-torchrl/envs/llm/transforms/kl.py,sha256=T-ab6P2PCpue0N4Pkz3ExttcEkTuPZylLlK5MGKAXuY,11992
+torchrl/envs/llm/transforms/kl.py,sha256=GrANICxnF-FC_yVkdO4EU66bj7aTfkB8OXG7wA2uxDo,23251
 torchrl/envs/llm/transforms/policy_version.py,sha256=fko23hsQrAMmUqFwKjV_CQVavDhixXFUeVE0lJBASOA,7080
 torchrl/envs/llm/transforms/tokenizer.py,sha256=Nest15FD1iPLNZuw0rAobyb7n3ce6KFX00qfN3dUE2M,14274
 torchrl/envs/llm/transforms/tools.py,sha256=WoNgUN1Me4mhbqH5ef9XNc5qKXE0H-g3ZobjEKDM_kw,30308
@@ -147,8 +148,8 @@ torchrl/modules/llm/backends/__init__.py,sha256=ABKK4mJeRtoLXEqfnMvIuiovs7VJoCxn
 torchrl/modules/llm/backends/vllm.py,sha256=5P78jEtAIytgYHzEkOrg-wwqh1ryhiMVy4M_AxNQ9JQ,9649
 torchrl/modules/llm/policies/__init__.py,sha256=CK7VEdfShjkeNu_-TmYOobrCEjKTIb2aw2hE6s5RBNs,439
 torchrl/modules/llm/policies/common.py,sha256=GXzmVRa0SJvQ8iPMeuNjwV7EaZDOPrVy5k_LlJ10QXY,3111
-torchrl/modules/llm/policies/transformers_wrapper.py,sha256=yylsatdc63hodGcjhfO79CAA2-kQmuvRqiFT38mbPCQ,23177
-torchrl/modules/llm/policies/vllm_wrapper.py,sha256=iDG0593FfZfzf8-JiQQPANQ6O-WnF8ifQJw1-O4vDHs,30488
+torchrl/modules/llm/policies/transformers_wrapper.py,sha256=0wDYGpC1T5T8ZVyZOi5S2Qoa2wtg5Oix2W9Y_bKMKs8,25988
+torchrl/modules/llm/policies/vllm_wrapper.py,sha256=LBoFbrTyGiEiilYVmU7Ze-WBpwYinvAVIEtbU1QKajw,32226
 torchrl/modules/models/__init__.py,sha256=Y1XTkBOB5EMj6IaMru6V3CDwFLnkUtxzsHcqzeqq_4Y,1829
 torchrl/modules/models/batchrenorm.py,sha256=bR4ZhaJ5E1cSK5o8L2dNX5KVLIb-bgrYxcq6yhx0I1A,4869
 torchrl/modules/models/decision_transformer.py,sha256=ANFTOm3k9_3Uv1vKGdXumRy3meBPnDdT8HqhVvJ2RCo,6783
@@ -194,8 +195,9 @@ torchrl/objectives/sac.py,sha256=gKOgCU399miKgpgu7Bmzs1bkIF8JTm_lybHn8V4wDuk,654
 torchrl/objectives/td3.py,sha256=Rq2q5gXo3AMuHm2OjRZvpfvKsAl1lIK5ALh2_sZM1ZE,23743
 torchrl/objectives/td3_bc.py,sha256=1pjB8mjCT2CLvQzjnqwAfZoc7yhjMB9UQjuJ5wZfTUY,26558
 torchrl/objectives/utils.py,sha256=Vrjj07SjMYANfFyn3n1xS7izBIs5Mq9mCvyITMzifZs,24705
-torchrl/objectives/llm/__init__.py,sha256=hlj2mKgz0BJR1ob1uObF8IkjyrDBnF24fXQf2zKEqaw,337
+torchrl/objectives/llm/__init__.py,sha256=tZmIz3rkeclw3MzJoOWEs2gkewjx2USKrKJbWdyiiaQ,406
 torchrl/objectives/llm/grpo.py,sha256=nT3Ukjaz7nZZnkS5tnb-pDnRzvZ3L1edpcNCzi5WZRs,17164
+torchrl/objectives/llm/sft.py,sha256=9fzX9Qo0Goyjxuwca6eLN1PUQ24F0LZGRpjzTDLFfs4,20572
 torchrl/objectives/multiagent/__init__.py,sha256=5uebDe5KrvlzeYV_BSd5vdmfruJQYMeDVVbU4iHErEg,245
 torchrl/objectives/multiagent/qmixer.py,sha256=yttOxc5FNylKw4iMnYSG1qO8EbHvx8imAhxNxW9_iLw,17362
 torchrl/objectives/value/__init__.py,sha256=QkSnenYVqe_3FVtwGr_D86N52unnpBvRXfcC5JFTBOw,589
@@ -221,8 +223,8 @@ torchrl/trainers/helpers/losses.py,sha256=rWKure02dl8hLBzLUs-jhNJV8L3QHWtFbl3HbX
 torchrl/trainers/helpers/models.py,sha256=VujBq9H92sEzpCtU1iTrJQNlwvyOO-Rho4bzsMonX6s,22465
 torchrl/trainers/helpers/replay_buffer.py,sha256=RaZqXnHimmadiibvDBcLbtIhpPaVMTPhYMOBvX4v3CA,2060
 torchrl/trainers/helpers/trainers.py,sha256=hB1FtHtP-S0PBQ4LF6WPy37caaLpacyaLThj1BNl5Ho,12372
-torchrl_nightly-2025.6.20.dist-info/LICENSE,sha256=PGO-oZsq4EzhE1-WQS2xGiEF3UCVb9YawfQ09cIMV_8,1119
-torchrl_nightly-2025.6.20.dist-info/METADATA,sha256=jj9lUe-poxEhkdNE2J7ExqB1sKG47hV8vCKY1Bn2b_k,40113
-torchrl_nightly-2025.6.20.dist-info/WHEEL,sha256=34WDYMHIz_-H9l4AW0T3KmyzuxEGEmufYB-iQzXMn2g,101
-torchrl_nightly-2025.6.20.dist-info/top_level.txt,sha256=JeTJ1jV7QJwLcUS1nr21aPn_wb-XlAZ9c-z_EH472JA,20
-torchrl_nightly-2025.6.20.dist-info/RECORD,,
+torchrl_nightly-2025.6.22.dist-info/LICENSE,sha256=PGO-oZsq4EzhE1-WQS2xGiEF3UCVb9YawfQ09cIMV_8,1119
+torchrl_nightly-2025.6.22.dist-info/METADATA,sha256=hHTi-fhdpfc3I2ipKsmcjKtiCzkK_mWcn_7nZwnH-0U,40113
+torchrl_nightly-2025.6.22.dist-info/WHEEL,sha256=34WDYMHIz_-H9l4AW0T3KmyzuxEGEmufYB-iQzXMn2g,101
+torchrl_nightly-2025.6.22.dist-info/top_level.txt,sha256=JeTJ1jV7QJwLcUS1nr21aPn_wb-XlAZ9c-z_EH472JA,20
+torchrl_nightly-2025.6.22.dist-info/RECORD,,