liger-kernel 0.5.2__py3-none-any.whl → 0.5.4__py3-none-any.whl

liger_kernel/chunked_loss/fused_linear_unpaired_preference.py

@@ -0,0 +1,246 @@
+from abc import abstractmethod
+from functools import partial
+
+import torch
+
+from torch.nn import functional as F
+
+
+class LigerFusedLinearUnpairedPreferenceBase(torch.autograd.Function):
+    @abstractmethod
+    def preference_loss_fn(*args, **kwargs):
+        """
+        To be extended by subclasses.
+        """
+        raise NotImplementedError("Preference loss function must be implemented.")
+
+    @staticmethod
+    def forward(
+        ctx,
+        _input,
+        weight,
+        target,
+        preference_labels,
+        bias=None,
+        loss_fn=None,
+        chunk_size=1,
+        ignore_index=-100,
+        compiled=True,
+        use_ref_model=False,
+        ref_input=None,
+        ref_weight=None,
+        ref_bias=None,
+        **loss_kwargs,
+    ):
+        """
+        Base class for fused linear layer with unpaired preference loss like KTO
+        Expects _input to be stacked with chosen and rejected inputs on the batch dimension.
+
+        The mental model is:
+
+        forward()
+        ├── Loop over chunks
+            └── compute_loss()
+                ├── chunk_forward()  # Compute logits and log probs
+                └── prefer_loss()    # Calculate preference loss
+
+        Args:
+            _input (torch.Tensor): Input tensor. Shape: (batch_size, seq_len, hidden_size).
+            weight (torch.Tensor): Weight tensor. Shape: (vocab_size, hidden_size).
+            target (torch.Tensor): Target tensor. Shape: (batch_size, seq_len).
+            bias (torch.Tensor, optional): Bias tensor. Shape: (vocab_size,).
+            loss_fn (callable): Loss function to compute the loss on a chunk of input/target.
+            chunk_size (int): Size of a chunk (# of batches of stacked chosen and rejected inputs).
+            ignore_index (int): Index to ignore for loss computation.
+            beta (float): Weight for the preference loss.
+            compiled (bool): Whether to use torch compile for chunk accumulation.
+            use_ref_model (bool): Whether to use a reference model for the alignment loss.
+            preference_labels (torch.Tensor): Boolean tensor indicating chosen (True) vs rejected (False) examples.
+                Shape: (batch_size,).
+            ref_weight (torch.Tensor): Reference weight tensor. Shape: (vocab_size, hidden_size).
+            ref_bias (torch.Tensor, optional): Reference bias tensor. Shape: (vocab_size,).
+            loss_kwargs (dict): Other possible arguments that a loss function might need
+        """
+        # TODO: Tune CHUNK_SIZE to fully utilize the GPU
+        CHUNK_SIZE = chunk_size
+
+        # Gradients to be accumulated
+        grad_inputs = []
+        grad_weight = torch.zeros_like(weight)
+        grad_bias = torch.zeros_like(bias) if bias is not None else None
+
+        # Loss to be accumulated
+        loss_acc = torch.zeros((), device=_input.device)
+
+        compute_loss = partial(
+            LigerFusedLinearUnpairedPreferenceBase._compute_loss,
+            preference_loss_fn=loss_fn,
+            full_target=target,
+            ignore_index=ignore_index,
+            use_ref_model=use_ref_model,
+            ref_weight=ref_weight,
+            ref_bias=ref_bias,
+            **loss_kwargs,
+        )
+
+        def fused_fwd_bwd(input_chunk, target_chunk, preference_labels_chunk, ref_input_chunk):
+            """
+            Fused forward and backward pass for a chunk of input and target.
+            """
+            argnums = (0, 1, 4) if bias is not None else (0, 1)
+            return torch.func.grad_and_value(compute_loss, argnums=argnums, has_aux=False)(
+                input_chunk,
+                weight,
+                target_chunk,
+                preference_labels_chunk,
+                bias,
+                ref_input_chunk=ref_input_chunk,
+            )
+
+        def accumulate_chunk(
+            input_chunk,
+            target_chunk,
+            preference_labels_chunk=None,
+            ref_input_chunk=None,
+        ):
+            (chunk_grad_input, chunk_grad_weight, *chunk_grad_bias), (chunk_loss) = fused_fwd_bwd(
+                input_chunk, target_chunk, preference_labels_chunk, ref_input_chunk
+            )
+            if bias is not None:
+                grad_bias.add_(chunk_grad_bias[0])  # accumulate bias gradient
+
+            # Accumulate gradients
+            grad_weight.add_(chunk_grad_weight)
+            grad_inputs.append(chunk_grad_input)
+
+            # Accumulate loss
+            loss_acc.add_(chunk_loss)
+
+        if compiled:
+            fused_fwd_bwd = torch.compile(fused_fwd_bwd)
+
+        # When not paired, use labels to separate chosen and rejected
+        assert preference_labels is not None, "preference_labels must be provided for unpaired preference loss"
+
+        chunks = max(1, _input.shape[0] // CHUNK_SIZE)
+        _input_chunks = torch.chunk(_input, chunks=chunks, dim=0)
+        _target_chunks = torch.chunk(target, chunks=chunks, dim=0)
+        _preference_labels_chunks = torch.chunk(preference_labels, chunks=chunks, dim=0)
+
+        if use_ref_model:
+            _ref_input_chunks = torch.chunk(ref_input, chunks=chunks, dim=0)
+
+        for (
+            input_chunk,
+            target_chunk,
+            ref_input_chunk,
+            preference_labels_chunk,
+        ) in zip(
+            _input_chunks,
+            _target_chunks,
+            (_ref_input_chunks if use_ref_model else [None] * len(_input_chunks)),
+            _preference_labels_chunks,
+        ):
+            # mark input_chunk, target_chunk, and target dimension 1 (sequence length) as dynamic to prevent torch.compile recompilation
+            torch._dynamo.mark_dynamic(input_chunk, 1)
+            torch._dynamo.mark_dynamic(target_chunk, 1)
+            torch._dynamo.mark_dynamic(target, 1)
+            torch._dynamo.mark_dynamic(ref_input_chunk, 1) if use_ref_model else None
+            torch._dynamo.mark_dynamic(preference_labels_chunk, 1)
+
+            # accumulate loss, gradients, and metrics
+            accumulate_chunk(input_chunk, target_chunk, preference_labels_chunk, ref_input_chunk)
+
+        ctx.save_for_backward(
+            torch.cat(grad_inputs, dim=0),
+            grad_weight,
+            grad_bias,
+        )
+        return loss_acc
+
+    @staticmethod
+    def backward(ctx, *grad_output):
+        grad_input, grad_weight, grad_bias = ctx.saved_tensors
+        if torch.ne(grad_output[0][0], torch.tensor(1.0, device=grad_output[0][0].device)):
+            grad_input = grad_input * grad_output[0][0]
+            grad_weight = grad_weight * grad_output[0][0]
+            grad_bias = grad_bias * grad_output[0][0] if grad_bias is not None else None
+
+        return grad_input, grad_weight, None, None, grad_bias
+
+    @staticmethod
+    def chunk_forward(
+        input_chunk,
+        weight,
+        target_chunk,
+        bias=None,
+        ignore_index=-100,
+    ):
+        logits_chunk = input_chunk @ weight.t()
+        if bias is not None:
+            logits_chunk = logits_chunk + bias
+        log_probs_chunk = F.log_softmax(logits_chunk.float(), dim=-1)
+
+        loss_mask_chunk = target_chunk != ignore_index
+        label_chunk = torch.where(loss_mask_chunk, target_chunk, 0)
+
+        per_token_logps_chunk = log_probs_chunk.gather(-1, label_chunk.unsqueeze(-1)).squeeze(-1)
+        average_log_prob_chunk = (per_token_logps_chunk * loss_mask_chunk).sum(-1) / loss_mask_chunk.sum(-1)
+
+        return average_log_prob_chunk
+
+    @staticmethod
+    def _compute_loss(
+        input_chunk,
+        weight,
+        target_chunk,
+        preference_labels_chunk,
+        bias=None,
+        preference_loss_fn=None,
+        full_target=None,
+        ignore_index=-100,
+        use_ref_model=False,
+        ref_input_chunk=None,
+        ref_weight=None,
+        ref_bias=None,
+        **loss_kwargs,
+    ):
+        """
+        Compute the total loss for a chunk of input and target, while using an alignment/preference loss function.
+        Args:
+            preference_loss_fn (callable): Loss function to compute the loss on a chunk of input/target.
+            input_chunk (torch.Tensor): Chunk of input tensor. Shape: (2 * chunk_size, sequence_length, hidden_size).
+            weight (torch.Tensor): Weight tensor. Shape: (vocab_size, hidden_size).
+            target_chunk (torch.Tensor): Chunk of target tensor. Shape: (2 * chunk_size, sequence_length).
+            bias (torch.Tensor, optional): Bias tensor. Shape: (vocab_size,).
+            full_target (torch.Tensor): Full target tensor. Shape: (batch_size, sequence_length).
+            ignore_index (int): Index to ignore for loss computation.
+            use_ref_model (bool): Whether to use a reference model for the alignment loss.
+            ref_weight (torch.Tensor): Reference weight tensor. Shape: (vocab_size, hidden_size).
+            ref_bias (torch.Tensor, optional): Reference bias tensor. Shape: (vocab_size,).
+            loss_kwargs (dict): Additional arguments for the loss function.
+        """
+        average_log_prob_chunk = LigerFusedLinearUnpairedPreferenceBase.chunk_forward(
+            input_chunk,
+            weight,
+            target_chunk,
+            bias=bias,
+            ignore_index=ignore_index,
+        )
+
+        if use_ref_model:
+            with torch.no_grad():
+                ref_average_log_prob_chunk = LigerFusedLinearUnpairedPreferenceBase.chunk_forward(
+                    ref_input_chunk,
+                    ref_weight,
+                    target_chunk,
+                    ref_bias,
+                    ignore_index=ignore_index,
+                )
+            loss_kwargs["ref_average_log_prob_chunk"] = ref_average_log_prob_chunk
+
+        preference_loss_chunk = preference_loss_fn(
+            average_log_prob_chunk, preference_labels_chunk, full_target, **loss_kwargs
+        )
+
+        return preference_loss_chunk

liger_kernel/chunked_loss/grpo_loss.py

@@ -0,0 +1,160 @@
+import torch
+
+from liger_kernel.chunked_loss.fused_linear_rlhf import LigerFusedLinearRLHFBase
+
+
+class LigerFusedLinearGRPOFunction(LigerFusedLinearRLHFBase):
+    @staticmethod
+    def rlhf_loss_fn(
+        log_probs,
+        attention_mask,
+        rewards,
+        ref_log_probs=None,
+        beta=0.1,
+        **kwargs,
+    ):
+        """GRPO Loss Function matching GRPOTrainer implementation."""
+        # Get chosen token probabilities
+        chosen_tokens = log_probs.argmax(dim=-1)  # (batch_size, seq_len)
+        chosen_token_logprobs = log_probs.gather(dim=-1, index=chosen_tokens.unsqueeze(-1)).squeeze(
+            -1
+        )  # (batch_size, seq_len)
+
+        # Get reference model probabilities
+        if ref_log_probs is not None:
+            with torch.no_grad():
+                ref_token_logprobs = ref_log_probs.gather(dim=-1, index=chosen_tokens.unsqueeze(-1)).squeeze(-1)
+        else:
+            ref_token_logprobs = chosen_token_logprobs.detach()
+
+        # Compute advantages per batch entry in a grouped fashion
+        mean_grouped_rewards = rewards.mean()  # [batch_size,]
+        std_grouped_rewards = rewards.std()  # [batch_size,]
+
+        # Calculate advantages using the same epsilon as in GRPOTrainer
+        eps = 1e-4
+        advantages = (rewards - mean_grouped_rewards) / (std_grouped_rewards + eps)
+
+        # Compute policy gradient loss with importance sampling ratio
+        ratio = torch.exp(chosen_token_logprobs - chosen_token_logprobs.detach())
+        policy_loss = -ratio * advantages.unsqueeze(1)
+
+        # Compute KL penalty
+        kl_div = (
+            torch.exp(ref_token_logprobs - chosen_token_logprobs) - (ref_token_logprobs - chosen_token_logprobs) - 1.0
+        )
+
+        # Combine losses
+        per_token_loss = policy_loss + beta * kl_div
+
+        # Apply masking and normalize
+        masked_loss = per_token_loss * attention_mask
+        seq_lengths = attention_mask.sum()
+        seq_lengths = torch.clamp(seq_lengths, min=1.0)
+        loss = masked_loss.sum() / seq_lengths
+
+        # Calculate metrics
+        metrics = (
+            chosen_token_logprobs.mean(),  # mean log prob
+            chosen_token_logprobs.std(),  # std log prob
+            log_probs.mean(),  # mean all log probs
+            ((kl_div * attention_mask).sum(dim=1) / attention_mask.sum(dim=1)).mean(),  # mean KL div
+        )
+
+        return loss, metrics
+
+    @staticmethod
+    def forward(
+        ctx,
+        _input,
+        weight,
+        attention_mask,
+        rewards,
+        bias=None,
+        ref_input=None,
+        ref_weight=None,
+        ref_bias=None,
+        beta=0.1,
+        compiled=True,
+        use_ref_model=True,
+        num_generations=1,
+    ):
+        return LigerFusedLinearRLHFBase.forward(
+            ctx=ctx,
+            _input=_input,
+            weight=weight,
+            attention_mask=attention_mask,
+            loss_fn=LigerFusedLinearGRPOFunction.rlhf_loss_fn,
+            rewards=rewards,
+            bias=bias,
+            ref_input=ref_input,
+            ref_weight=ref_weight,
+            ref_bias=ref_bias,
+            beta=beta,
+            compiled=compiled,
+            use_ref_model=use_ref_model,
+            num_generations=num_generations,
+        )
+
+    @staticmethod
+    def backward(ctx, grad_output, *grad_metrics):
+        """Backward pass for GRPO loss.
+
+        Args:
+            grad_output: Gradient of the loss (scalar)
+            grad_metrics: Gradients of the metrics (not used in backward computation)
+        """
+        grads = LigerFusedLinearRLHFBase.backward(ctx, grad_output)
+        return (
+            *grads[:5],  # grad_input, grad_weight, grad_attention_mask, grad_rewards, grad_bias
+            None,  # grad_ref_input
+            None,  # grad_ref_weight
+            None,  # grad_ref_bias
+            None,  # grad_beta
+            None,  # grad_compiled
+            None,  # grad_use_ref_model
+            None,  # grad_num_generations
+        )
+
+
+class LigerFusedLinearGRPOLoss(torch.nn.Module):
+    """Fused linear layer with GRPO loss."""
+
+    def __init__(
+        self,
+        beta: float = 0.1,
+        compiled: bool = True,
+        use_ref_model: bool = True,
+        num_generations: int = 1,
+    ):
+        super().__init__()
+        self.beta = beta
+        self.compiled = compiled
+        self.use_ref_model = use_ref_model
+        self.num_generations = num_generations
+
+    def forward(
+        self,
+        _input,
+        lin_weight,
+        attention_mask,
+        rewards,
+        bias=None,
+        ref_input=None,
+        ref_weight=None,
+        ref_bias=None,
+    ):
+        return LigerFusedLinearGRPOFunction.apply(
+            _input,
+            lin_weight,
+            attention_mask,
+            rewards,
+            bias,
+            ref_input,
+            ref_weight,
+            ref_bias,
+            self.beta,
+            self.compiled,
+            self.use_ref_model,
+            self.num_generations,
+        )

liger_kernel/chunked_loss/jsd_loss.py

@@ -0,0 +1,154 @@
+import torch
+import torch.nn.functional as F
+
+from liger_kernel.chunked_loss.fused_linear_distillation import LigerFusedLinearDistillationBase
+
+
+class LigerFusedLinearJSDFunction(LigerFusedLinearDistillationBase):
+    @staticmethod
+    def distillation_loss_fn(student_logits, teacher_logits, beta=0.5):
+        """
+        Compute JSD loss (Jensen-Shannon Divergence Loss).
+        Args:
+            student_logits (torch.Tensor): Logits of student tokens. Shape: (batch_size * seq_len,).
+            teacher_logits (torch.Tensor): Logits of teacher tokens. Shape: (batch_size * seq_len,).
+            beta (float): Coefficient beta of generalized JSD in the interval [0, 1]. Default: `0.5`.
+        Returns:
+            torch.Tensor: Jensen-Shannon Divergence loss
+        """
+        student_log_probs = F.log_softmax(student_logits, dim=-1)
+        teacher_log_probs = F.log_softmax(teacher_logits, dim=-1)
+
+        # Compute probabilities (only required for mean calculation)
+        mean_probs = beta * student_log_probs.exp() + (1 - beta) * teacher_log_probs.exp()
+        log_mean_probs = mean_probs.log()
+
+        student_kl = F.kl_div(log_mean_probs, student_log_probs, reduction="sum", log_target=True)
+        teacher_kl = F.kl_div(log_mean_probs, teacher_log_probs, reduction="sum", log_target=True)
+
+        # JSD is the weighted average of the KL divergences
+        jsd_loss = beta * teacher_kl + (1 - beta) * student_kl
+        return jsd_loss
+
+    @staticmethod
+    def forward(
+        ctx,
+        student_input: torch.Tensor,
+        student_weight: torch.Tensor,
+        teacher_input: torch.Tensor,
+        teacher_weight: torch.Tensor,
+        true_labels: torch.LongTensor,
+        weight_hard_loss: float = 0.5,
+        weight_soft_loss: float = 0.5,
+        beta: float = 0.5,
+        ignore_index: int = -100,
+        temperature: float = 1.0,
+        compiled: bool = True,
+    ):
+        """
+        Fused linear layer with JSD distillation loss.
+        Args:
+            student_input (torch.Tensor): Student input tensor. Shape: (batch_size * seq_len, hidden_size_student)
+            student_weight (torch.Tensor): Student weight tensor. Shape: (vocab_size, hidden_size_student)
+            teacher_input (torch.Tensor): Teacher input tensor. Shape: (batch_size * seq_len, hidden_size_teacher)
+            teacher_weight (torch.Tensor): Teacher weight tensor. Shape: (vocab_size, hidden_size_teacher)
+            true_labels (torch.LongTensor): Target tensor. Shape: (batch_size * seq_len,)
+            weight_hard_loss (float): Weight for hard loss.
+            weight_soft_loss (float): Weight for soft loss.
+            beta (float): Coefficient beta of generalized JSD in the interval [0, 1]. Default: `0.5`.
+            ignore_index (int): Index to ignore in loss computation
+            temperature (float): Temperature for softening/sharpening distributions
+            compiled (bool): Whether to use torch compile
+        Returns:
+            torch.Tensor: Computed loss
+        """
+        return LigerFusedLinearDistillationBase.forward(
+            ctx=ctx,
+            student_input=student_input,
+            student_weight=student_weight,
+            teacher_input=teacher_input,
+            teacher_weight=teacher_weight,
+            target=true_labels,
+            loss_fn=LigerFusedLinearJSDFunction.distillation_loss_fn,
+            chunk_size=1,
+            weight_hard_loss=weight_hard_loss,
+            weight_soft_loss=weight_soft_loss,
+            beta=beta,
+            ignore_index=ignore_index,
+            temperature=temperature,
+            compiled=compiled,
+        )
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        grads = LigerFusedLinearDistillationBase.backward(ctx, grad_output)[:4]
+
+        return (*grads, None, None, None, None, None, None, None)
+
+
+class LigerFusedLinearJSDLoss(torch.nn.Module):
+    """
+    Fused linear layer with JSD distillation loss.
+    """
+
+    def __init__(
+        self,
+        weight_hard_loss: float = 0.5,
+        weight_soft_loss: float = 0.5,
+        beta: float = 0.5,
+        ignore_index: int = -100,
+        temperature: float = 1.0,
+        compiled: bool = True,
+    ):
+        """
+        Args:
+            weight_hard_loss (float): Weight for hard loss.
+            weight_soft_loss (float): Weight for soft loss.
+            ignore_index (int): Index to ignore in the loss
+            temperature (float): Temperature for softening distributions
+            compiled (bool): Whether to use torch compile
+            beta (float): Coefficient beta of generalized JSD in the interval [0, 1]. Default: `0.5`.
+        """
+        super().__init__()
+        assert temperature != 0, "Temperature cannot be 0."
+        self.weight_hard_loss = weight_hard_loss
+        self.weight_soft_loss = weight_soft_loss
+        self.ignore_index = ignore_index
+        self.temperature = temperature
+        self.compiled = compiled
+        self.beta = beta
+
+    def forward(
+        self,
+        student_input: torch.Tensor,
+        student_weight: torch.Tensor,
+        teacher_input: torch.Tensor,
+        teacher_weight: torch.Tensor,
+        true_labels: torch.LongTensor,
+    ) -> torch.Tensor:
+        """
+        Compute the JSD distillation loss.
+
+        Args:
+            student_input (torch.Tensor): Student input tensor
+            student_weight (torch.Tensor): Student weight tensor
+            teacher_input (torch.Tensor): Teacher input tensor
+            teacher_weight (torch.Tensor): Teacher weight tensor
+            true_labels (torch.LongTensor): Target labels tensor
+
+        Returns:
+            torch.Tensor: Computed loss
+        """
+        return LigerFusedLinearJSDFunction.apply(
+            student_input,
+            student_weight,
+            teacher_input,
+            teacher_weight,
+            true_labels,
+            self.weight_hard_loss,
+            self.weight_soft_loss,
+            self.beta,
+            self.ignore_index,
+            self.temperature,
+            self.compiled,
+        )