liger-kernel 0.4.1__py3-none-any.whl → 0.5.0__py3-none-any.whl

liger_kernel/chunked_loss/fused_linear_preference.py

@@ -0,0 +1,386 @@
+from abc import abstractmethod
+from functools import partial
+
+import torch
+from torch.nn import functional as F
+
+
+class LigerFusedLinearPreferenceBase(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,
+        bias=None,
+        loss_fn=None,
+        chunk_size=1,
+        ignore_index=-100,
+        alpha=1.0,
+        beta=0.1,
+        compute_nll_loss=True,
+        compiled=True,
+        use_ref_model=False,
+        # TODO: ref input
+        ref_weight=None,
+        ref_bias=None,
+        **loss_kwargs,
+    ):
+        """
+        Base class for fused linear layer with preference loss.
+        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.
+            alpha (float): Weight for the NLL loss.
+            beta (float): Weight for the preference loss.
+            compute_nll_loss (bool): Whether to compute NLL 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.
+            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_weight = torch.zeros_like(weight)
+        grad_chosen_inputs = []
+        grad_rejected_inputs = []
+        grad_bias = torch.zeros_like(bias) if bias is not None else None
+
+        # Loss to be accumulated
+        loss_acc = torch.zeros((), device=_input.device)
+
+        # Metrics to be recorded
+        policy_chosen_logps = []
+        policy_rejected_logps = []
+        policy_chosen_logits_mean = torch.zeros((), device=_input.device)
+        policy_rejected_logits_mean = torch.zeros((), device=_input.device)
+        policy_nll_loss = torch.zeros((), device=_input.device)
+        aggregated_aux_outputs = []  # aggregated aux outputs from all chunks
+
+        compute_loss = partial(
+            LigerFusedLinearPreferenceBase._compute_loss,
+            preference_loss_fn=loss_fn,
+            ignore_index=ignore_index,
+            alpha=alpha,
+            beta=beta,
+            compute_nll_loss=compute_nll_loss,
+            full_target=target,
+            use_ref_model=use_ref_model,
+            ref_weight=ref_weight,
+            ref_bias=ref_bias,
+            **loss_kwargs,
+        )
+
+        def fused_fwd_bwd(input_chunk, target_chunk):
+            """
+            Fused forward and backward pass for a chunk of input and target.
+            """
+            if bias is not None:
+                return torch.func.grad_and_value(
+                    compute_loss, argnums=(0, 1, 3), has_aux=True
+                )(input_chunk, weight, target_chunk, bias)
+            else:
+                return torch.func.grad_and_value(
+                    compute_loss, argnums=(0, 1), has_aux=True
+                )(input_chunk, weight, target_chunk)
+
+        def accumulate_chunk(input_chunk, target_chunk):
+            if bias is not None:
+                (chunk_grad_input, chunk_grad_weight, chunk_grad_bias), (
+                    chunk_loss,
+                    (
+                        chunk_chosen_logps,
+                        chunk_rejected_logps,
+                        chunk_chosen_logits_mean,
+                        chunk_rejected_logits_mean,
+                        chunk_nll_loss,
+                        *aux_outputs,
+                    ),
+                ) = fused_fwd_bwd(input_chunk, target_chunk)
+                grad_bias.add_(chunk_grad_bias)  # accumulate bias gradient
+            else:
+                (chunk_grad_input, chunk_grad_weight), (
+                    chunk_loss,
+                    (
+                        chunk_chosen_logps,
+                        chunk_rejected_logps,
+                        chunk_chosen_logits_mean,
+                        chunk_rejected_logits_mean,
+                        chunk_nll_loss,
+                        *aux_outputs,
+                    ),
+                ) = fused_fwd_bwd(input_chunk, target_chunk)
+
+            # Accumulate gradients
+            grad_weight.add_(chunk_grad_weight)
+            grad_chosen_inputs.append(chunk_grad_input[: chosen_target_chunk.shape[0]])
+            grad_rejected_inputs.append(
+                chunk_grad_input[chosen_target_chunk.shape[0] :]
+            )
+
+            # Accumulate loss
+            loss_acc.add_(chunk_loss)
+
+            # Accumulate metrics
+            policy_chosen_logps.append(chunk_chosen_logps)
+            policy_rejected_logps.append(chunk_rejected_logps)
+            policy_chosen_logits_mean.add_(chunk_chosen_logits_mean)
+            policy_rejected_logits_mean.add_(chunk_rejected_logits_mean)
+            policy_nll_loss.add_(chunk_nll_loss)
+
+            # aux_outputs
+            # Initialize storage for aux_outputs
+            if len(aggregated_aux_outputs) == 0:
+                for aux in aux_outputs:
+                    if aux.ndim == 0:
+                        aggregated_aux_outputs.append(
+                            torch.zeros((), device=aux.device)
+                        )
+                    else:
+                        aggregated_aux_outputs.append([])
+
+            # Process each aux_output
+            for i, aux in enumerate(aux_outputs):
+                if aux.ndim == 0:
+                    aggregated_aux_outputs[i].add_(aux)
+                else:
+                    aggregated_aux_outputs[i].append(aux)
+
+        if compiled:
+            fused_fwd_bwd = torch.compile(fused_fwd_bwd)
+
+        len_chosen = target.shape[0] // 2
+        chunks = max(1, _input.shape[0] // (2 * CHUNK_SIZE))
+        _chosen_input_chunks = torch.chunk(_input[:len_chosen], chunks=chunks, dim=0)
+        _chosen_target_chunks = torch.chunk(target[:len_chosen], chunks=chunks, dim=0)
+        _rejected_input_chunks = torch.chunk(_input[len_chosen:], chunks=chunks, dim=0)
+        _rejected_target_chunks = torch.chunk(target[len_chosen:], chunks=chunks, dim=0)
+
+        for (
+            chosen_input_chunk,
+            rejected_input_chunk,
+            chosen_target_chunk,
+            rejected_target_chunk,
+        ) in zip(
+            _chosen_input_chunks,
+            _rejected_input_chunks,
+            _chosen_target_chunks,
+            _rejected_target_chunks,
+        ):
+            input_chunk = torch.cat([chosen_input_chunk, rejected_input_chunk], dim=0)
+            target_chunk = torch.cat(
+                [chosen_target_chunk, rejected_target_chunk], dim=0
+            )
+
+            # mark input_chunk, target_chunk, and target dimension 1 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)
+
+            # accumulate loss, gradients, and metrics
+            accumulate_chunk(input_chunk, target_chunk)
+
+        # combine grad_chosen_inputs and grad_rejected_inputs
+        grad_inputs = grad_chosen_inputs + grad_rejected_inputs
+        policy_chosen_logps = torch.cat(policy_chosen_logps, dim=0)
+        policy_rejected_logps = torch.cat(policy_rejected_logps, dim=0)
+
+        # Aggregate aux outputs lists into tensors
+        for i, aux in enumerate(aggregated_aux_outputs):
+            if isinstance(aux, list):
+                aggregated_aux_outputs[i] = torch.cat(aux, dim=0)
+
+        ctx.save_for_backward(
+            torch.cat(grad_inputs, dim=0),
+            grad_weight,
+            grad_bias,
+        )
+        return_vars = (
+            policy_chosen_logps,
+            policy_rejected_logps,
+            policy_chosen_logits_mean,
+            policy_rejected_logits_mean,
+            policy_nll_loss,
+        )
+        return loss_acc, (*return_vars, *aggregated_aux_outputs)
+
+    @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, grad_bias, None, None, None
+
+    @staticmethod
+    def chunk_forward(
+        input_chunk,
+        weight,
+        target_chunk,
+        bias=None,
+        ignore_index=-100,
+        compute_nll_loss=True,
+    ):
+        len_chosen_chunk = target_chunk.shape[0] // 2
+        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)
+
+        chosen_nll_loss = 0.0
+        if compute_nll_loss:
+            chosen_nll_loss = F.nll_loss(
+                log_probs_chunk[:len_chosen_chunk].view(-1, log_probs_chunk.shape[-1]),
+                target_chunk[:len_chosen_chunk].view(-1),
+                reduction="sum",
+                ignore_index=ignore_index,
+            )
+
+        loss_mask = target_chunk != ignore_index
+        label_chunk = torch.where(loss_mask, target_chunk, 0)
+
+        per_token_logps = log_probs_chunk.gather(-1, label_chunk.unsqueeze(-1)).squeeze(
+            -1
+        )
+        average_log_prob = (per_token_logps * loss_mask).sum(-1) / loss_mask.sum(-1)
+
+        chosen_logps = average_log_prob[:len_chosen_chunk]
+        rejected_logps = average_log_prob[len_chosen_chunk:]
+
+        chosen_logits = logits_chunk[:len_chosen_chunk]
+        rejected_logits = logits_chunk[len_chosen_chunk:]
+
+        return (
+            chosen_logps,
+            rejected_logps,
+            chosen_logits,
+            rejected_logits,
+            chosen_nll_loss,
+        )
+
+    @staticmethod
+    def _compute_loss(
+        input_chunk,
+        weight,
+        target_chunk,
+        bias=None,
+        preference_loss_fn=None,
+        full_target=None,
+        ignore_index=-100,
+        alpha=1.0,
+        beta=0.1,
+        compute_nll_loss=True,
+        use_ref_model=False,
+        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.
+            alpha (float): Weight for the NLL loss.
+            beta (float): Weight for the preference loss.
+            compute_nll_loss (bool): Whether to compute NLL loss.
+            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.
+        """
+        (
+            chosen_logps,
+            rejected_logps,
+            chosen_logits,
+            rejected_logits,
+            chosen_nll_loss,
+        ) = LigerFusedLinearPreferenceBase.chunk_forward(
+            input_chunk,
+            weight,
+            target_chunk,
+            bias=bias,
+            ignore_index=ignore_index,
+            compute_nll_loss=compute_nll_loss,
+        )
+        chosen_nll_loss = (
+            chosen_nll_loss
+            / (full_target[: full_target.shape[0] // 2] != ignore_index).sum()
+        )
+        chosen_logits_mean = chosen_logits.sum() / (
+            full_target.shape[0] // 2 * input_chunk.shape[1] * weight.shape[0]
+        )
+        rejected_logits_mean = rejected_logits.sum() / (
+            full_target.shape[0] // 2 * input_chunk.shape[1] * weight.shape[0]
+        )
+
+        if use_ref_model:
+            with torch.no_grad():
+                (
+                    ref_chosen_logps,
+                    ref_rejected_logps,
+                    ref_chosen_logits,
+                    ref_rejected_logits,
+                    ref_chosen_nll_loss,
+                ) = LigerFusedLinearPreferenceBase.chunk_forward(
+                    input_chunk,
+                    ref_weight,
+                    target_chunk,
+                    ref_bias,
+                    ignore_index=ignore_index,
+                    compute_nll_loss=False,  # We don't need NLL loss for the reference model
+                )
+            loss_kwargs["ref_chosen_logps"] = ref_chosen_logps
+            loss_kwargs["ref_rejected_logps"] = ref_rejected_logps
+
+        preference_loss_outputs = preference_loss_fn(
+            chosen_logps, rejected_logps, full_target, beta=beta, **loss_kwargs
+        )
+        if isinstance(preference_loss_outputs, tuple):
+            preference_loss, *aux_outputs = preference_loss_outputs
+        else:
+            preference_loss, aux_outputs = preference_loss_outputs, []
+
+        loss = alpha * chosen_nll_loss - preference_loss
+        return_vars = (
+            chosen_logps,
+            rejected_logps,
+            chosen_logits_mean,
+            rejected_logits_mean,
+            chosen_nll_loss,
+        )
+        return loss, (*return_vars, *aux_outputs)

liger_kernel/chunked_loss/orpo_loss.py

@@ -0,0 +1,113 @@
+import torch
+import torch.nn.functional as F
+
+from liger_kernel.chunked_loss.fused_linear_preference import (
+    LigerFusedLinearPreferenceBase,
+)
+
+
+class LigerFusedLinearORPOFunction(LigerFusedLinearPreferenceBase):
+
+    @staticmethod
+    def preference_loss_fn(chosen_logps, rejected_logps, full_target, beta=0.1):
+        """
+        Paper: https://arxiv.org/pdf/2403.07691
+
+        Formula:
+        Compute odds-ratio loss: L_OR = -log(σ(log(odds_θ(y_w|x) / odds_θ(y_l|x))))
+        where odds_θ(y|x) = P_θ(y|x) / (1 - P_θ(y|x))
+
+        Where:
+        - P_θ(y|x): Policy (model) probability
+        - y_w: Chosen sequence
+        - y_l: Rejected sequence
+        - σ: Sigmoid function
+        - β: Weight for the odds ratio loss
+        - odds_θ: Odds function for the policy
+
+        Args:
+            chosen_logps (torch.Tensor): Avg log probabilities of chosen tokens. Shape: (batch_size,).
+            rejected_logps (torch.Tensor): Avg log probabilities of rejected tokens. Shape: (batch_size,).
+            full_target (torch.Tensor): Non chunked full target tensor
+            beta (float): Weight for the odds ratio loss.
+        """
+        log_odds = (chosen_logps - rejected_logps) - (
+            torch.log1p(-torch.exp(chosen_logps))
+            - torch.log1p(-torch.exp(rejected_logps))
+        )
+        ratio = F.logsigmoid(log_odds)
+        loss = beta * ratio.sum() / (full_target.shape[0] // 2)
+
+        chosen_rewards = beta * chosen_logps
+        rejected_rewards = beta * rejected_logps
+
+        log_odds_ratio = torch.sum(ratio) / (full_target.shape[0] // 2)
+        log_odds_chosen = torch.sum(log_odds) / (full_target.shape[0] // 2)
+
+        return loss, chosen_rewards, rejected_rewards, log_odds_ratio, log_odds_chosen
+
+    @staticmethod
+    def forward(
+        ctx,
+        _input,
+        weight,
+        target,
+        bias=None,
+        ignore_index=-100,
+        beta=0.1,
+        compute_nll_loss=True,
+        compiled=True,
+    ):
+        return LigerFusedLinearPreferenceBase.forward(
+            ctx=ctx,
+            _input=_input,
+            weight=weight,
+            target=target,
+            bias=bias,
+            loss_fn=LigerFusedLinearORPOFunction.preference_loss_fn,
+            ignore_index=ignore_index,
+            beta=beta,
+            compute_nll_loss=compute_nll_loss,
+            compiled=compiled,
+        )
+
+    @staticmethod
+    def backward(ctx, *grad_output):
+        grads = LigerFusedLinearPreferenceBase.backward(ctx, grad_output)[:4]
+        return *grads, None, None, None, None
+
+
+class LigerFusedLinearORPOLoss(torch.nn.Module):
+    """
+    Fused linear layer with ORPO (Odds-Ratio Preference Optimization) loss.
+    """
+
+    def __init__(
+        self,
+        ignore_index: int = -100,
+        beta: float = 0.1,
+        compute_nll_loss: bool = True,
+        compiled: bool = True,
+    ):
+        """
+        Args:
+            ignore_index (int): Index to ignore in the loss.
+            beta (float): Weight for the odds ratio loss.
+        """
+        super().__init__()
+        self.ignore_index = ignore_index
+        self.beta = beta
+        self.compute_nll_loss = compute_nll_loss
+        self.compiled = compiled
+
+    def forward(self, lin_weight, _input, target, bias=None):
+        return LigerFusedLinearORPOFunction.apply(
+            _input,
+            lin_weight,
+            target,
+            bias,
+            self.ignore_index,
+            self.beta,
+            self.compute_nll_loss,
+            self.compiled,
+        )

liger_kernel/chunked_loss/simpo_loss.py

@@ -0,0 +1,115 @@
+import torch
+import torch.nn.functional as F
+
+from liger_kernel.chunked_loss.fused_linear_preference import (
+    LigerFusedLinearPreferenceBase,
+)
+
+
+class LigerFusedLinearSimPOFunction(LigerFusedLinearPreferenceBase):
+
+    @staticmethod
+    def preference_loss_fn(
+        chosen_logps, rejected_logps, full_target, beta=0.1, gamma=0.5
+    ):
+        """
+        Paper: https://arxiv.org/pdf/2405.14734
+
+        Formula:
+        L_SimPO(π_θ) = -E [log σ(β/|y_w| log π_θ(y_w|x) - β/|y_l| log π_θ(y_l|x) - γ)]
+
+        Where:
+        - π_θ(y|x): Policy (model) probability
+        - y_w: Chosen sequence
+        - y_l: Rejected sequence
+        - |y_w|, |y_l|: Sequence lengths
+        - σ: Sigmoid function
+        - β: beta weight
+        - γ: gemma margin term
+
+        Args:
+            chosen_logps (torch.Tensor): Avg log probabilities of chosen tokens. Shape: (batch_size,).
+            rejected_logps (torch.Tensor): Avg log probabilities of rejected tokens. Shape: (batch_size,).
+            full_target: Non chunked full target tensor
+            beta (float): beta weight
+            gamma (float): gemma margin term
+        """
+        logits = beta * (chosen_logps - rejected_logps) - gamma
+        loss = F.logsigmoid(logits).sum() / (full_target.shape[0] // 2)
+        return loss
+
+    @staticmethod
+    def forward(
+        ctx,
+        _input,
+        weight,
+        target,
+        bias=None,
+        ignore_index=-100,
+        beta=0.1,
+        alpha=1.0,
+        compute_nll_loss=False,
+        compiled=True,
+        gamma=0.5,
+    ):
+        return LigerFusedLinearPreferenceBase.forward(
+            ctx,
+            _input,
+            weight,
+            target,
+            bias,
+            loss_fn=LigerFusedLinearSimPOFunction.preference_loss_fn,
+            compute_nll_loss=compute_nll_loss,
+            ignore_index=ignore_index,
+            alpha=alpha,
+            beta=beta,
+            compiled=compiled,
+            gamma=gamma,
+        )
+
+    @staticmethod
+    def backward(ctx, *grad_output):
+        grads = LigerFusedLinearPreferenceBase.backward(ctx, grad_output)[:4]
+        return *grads, None, None, None, None, None, None
+
+
+class LigerFusedLinearSimPOLoss(torch.nn.Module):
+    """
+    Fused linear layer with SimPO loss.
+    """
+
+    def __init__(
+        self,
+        ignore_index: int = -100,
+        beta: float = 0.1,
+        alpha: float = 1.0,
+        compute_nll_loss: bool = True,
+        compiled: bool = True,
+        gamma: float = 0.5,
+    ):
+        """
+        Args:
+            ignore_index (int): Index to ignore in the loss.
+            beta (float): Weight for the odds ratio loss.
+        """
+        super().__init__()
+        self.ignore_index = ignore_index
+        self.beta = beta
+        self.alpha = alpha
+        self.compute_nll_loss = compute_nll_loss
+        self.compiled = compiled
+        self.gamma = gamma
+
+    def forward(self, lin_weight, _input, target, bias=None):
+        return LigerFusedLinearSimPOFunction.apply(
+            _input,
+            lin_weight,
+            target,
+            bias,
+            self.ignore_index,
+            self.beta,
+            self.alpha,
+            self.compute_nll_loss,
+            self.compiled,
+            self.gamma,
+        )

liger_kernel/env_report.py

@@ -1,5 +1,6 @@
 import platform
 import sys
+from importlib.metadata import version
 
 
 def print_env_report():
@@ -17,6 +18,11 @@ def print_env_report():
     print(f"Operating System: {platform.platform()}")
     print(f"Python version: {sys.version.split()[0]}")
 
+    try:
+        print(f"Liger Kernel version: {version('liger-kernel')}")
+    except ImportError:
+        print("Liger Kernel: Not installed")
+
     try:
         import torch
 
@@ -25,9 +31,17 @@ def print_env_report():
             torch.version.cuda if torch.cuda.is_available() else "Not available"
         )
         print(f"CUDA version: {cuda_version}")
+        hip_version = (
+            torch.version.hip
+            if torch.cuda.is_available() and torch.version.hip
+            else "Not available"
+        )
+        print(f"HIP(ROCm) version: {hip_version}")
+
     except ImportError:
         print("PyTorch: Not installed")
         print("CUDA version: Unable to query")
+        print("HIP(ROCm) version: Unable to query")
 
     try:
         import triton
@@ -43,6 +57,14 @@ def print_env_report():
     except ImportError:
         print("Transformers: Not installed")
 
+    try:
+        xpu_version = (
+            torch.version.xpu if torch.xpu.is_available() else "XPU Not Available"
+        )
+        print(f"XPU version: {xpu_version}")
+    except ImportError:
+        print("XPU version: Unable to query")
+
 
 if __name__ == "__main__":
     print_env_report()