liger-kernel-nightly 0.4.1.dev20241113011623__py3-none-any.whl → 0.4.1.dev20241114041219__py3-none-any.whl

liger_kernel/chunked_loss/fused_linear_preference.py

@@ -0,0 +1,107 @@
+import torch
+
+
+class LigerFusedLinearPreferenceBase(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        _input,
+        weight,
+        target,
+        bias=None,
+        loss_fn=None,
+        chunk_size=1,
+        compiled=True,
+    ):
+        """
+        Base class for fused linear layer with preference loss.
+        Expects _input to be stacked with chosen and rejected inputs on the batch dimension.
+
+        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).
+            compiled (bool): Whether to use torch compile for chunk accumulation.
+        """
+        # TODO: Tune CHUNK_SIZE to fully utilize the GPU
+        CHUNK_SIZE = chunk_size
+
+        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_acc = torch.zeros((), device=_input.device)
+
+        chunks = max(1, _input.shape[0] // (2 * CHUNK_SIZE))
+
+        def accumulate_chunk(input_chunk, target_chunk):
+            if bias is not None:
+                (chunk_grad_input, chunk_grad_weight, chunk_grad_bias), (
+                    chunk_loss,
+                    (chunk_or_loss, chunk_chosen_logps, chunk_rejected_logps),
+                ) = torch.func.grad_and_value(loss_fn, argnums=(0, 1, 3), has_aux=True)(
+                    input_chunk, weight, target_chunk, bias
+                )
+                grad_bias.add_(chunk_grad_bias)
+            else:
+                (chunk_grad_input, chunk_grad_weight), (
+                    chunk_loss,
+                    (chunk_or_loss, chunk_chosen_logps, chunk_rejected_logps),
+                ) = torch.func.grad_and_value(loss_fn, argnums=(0, 1), has_aux=True)(
+                    input_chunk, weight, target_chunk
+                )
+            grad_weight.add_(chunk_grad_weight)
+            loss_acc.add_(chunk_loss)
+            return chunk_grad_input
+
+        len_chosen = target.shape[0] // 2
+        _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
+            )
+
+            if compiled:
+                accumulate_chunk = torch.compile(accumulate_chunk)
+            grad_input = accumulate_chunk(input_chunk, target_chunk)
+
+            grad_chosen_inputs.append(grad_input[: chosen_target_chunk.shape[0]])
+            grad_rejected_inputs.append(grad_input[chosen_target_chunk.shape[0] :])
+
+        # combine grad_chosen_inputs and grad_rejected_inputs
+        grad_inputs = grad_chosen_inputs + grad_rejected_inputs
+
+        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, torch.tensor(1.0, device=grad_output.device)):
+            grad_input = grad_input * grad_output
+            grad_weight = grad_weight * grad_output
+            grad_bias = grad_bias * grad_output if grad_bias is not None else None
+
+        return grad_input, grad_weight, None, grad_bias, None, None, None

liger_kernel/chunked_loss/orpo_loss.py

@@ -0,0 +1,117 @@
+from functools import partial
+
+import torch
+import torch.nn.functional as F
+
+from liger_kernel.chunked_loss.fused_linear_preference import (
+    LigerFusedLinearPreferenceBase,
+)
+
+
+def odds_ratio_loss(chosen_logps, rejected_logps, beta=0.1):
+    """
+    Compute odds-ratio loss.
+    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,).
+        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)
+    return beta * ratio.sum()
+
+
+def _compute_orpo_loss(
+    input_chunk,
+    weight,
+    target_chunk,
+    bias=None,
+    full_target=None,
+    ignore_index=-100,
+    beta=0.1,
+    compute_nll_loss=True,
+):
+    """
+    Compute ORPO loss for a chunk of input and target.
+    Args:
+        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.
+        beta (float): Weight for the odds ratio loss.
+    """
+    len_chosen_chunk = target_chunk.shape[0] // 2
+
+    logits_chunk = input_chunk @ weight.t()  # chunk_size x V
+    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,
+        )
+        chosen_nll_loss = (
+            chosen_nll_loss
+            / (full_target[: full_target.shape[0] // 2] != ignore_index).sum()
+        )
+
+    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:]
+
+    or_loss = odds_ratio_loss(chosen_logps, rejected_logps, beta=beta)
+    or_loss = or_loss / (full_target.shape[0] // 2)
+
+    loss = chosen_nll_loss - or_loss
+    return loss, (or_loss, chosen_logps, rejected_logps)
+
+
+class LigerFusedLinearORPOFunction(LigerFusedLinearPreferenceBase):
+    @staticmethod
+    def forward(
+        ctx,
+        _input,
+        weight,
+        target,
+        bias=None,
+        ignore_index=-100,
+        beta=0.1,
+        compute_nll_loss=True,
+        compiled=True,
+    ):
+        """
+        Fused linear layer with ORPO (Odds-Ratio Preference Optimization) loss.
+        Handles both the forward and backward pass of the final linear layer with ORPO loss.
+        Inspired from LigerFusedLinearCrossEntropyFunction (https://arxiv.org/abs/2410.10989) which fuses final linear layer and CE loss.
+        """
+        orpo_loss_fn = partial(
+            _compute_orpo_loss,
+            full_target=target,
+            ignore_index=ignore_index,
+            beta=beta,
+            compute_nll_loss=compute_nll_loss,
+        )
+        return LigerFusedLinearPreferenceBase.forward(
+            ctx, _input, weight, target, bias, loss_fn=orpo_loss_fn
+        )
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        # Get gradients for _input, weight, bias, and target from the base class
+        grads = LigerFusedLinearPreferenceBase.backward(ctx, grad_output)[:4]
+        # Return these gradients, followed by None for the remaining inputs
+        return *grads, None, None, None, None

{liger_kernel_nightly-0.4.1.dev20241113011623.dist-info → liger_kernel_nightly-0.4.1.dev20241114041219.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.4.1.dev20241113011623
+Version: 0.4.1.dev20241114041219
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.4.1.dev20241113011623.dist-info → liger_kernel_nightly-0.4.1.dev20241114041219.dist-info}/RECORD

@@ -1,4 +1,7 @@
 liger_kernel/env_report.py,sha256=jye8RvUkmhqaIshdeIpoUABoAu7FPKJUib4FnAfvkpw,1132
+liger_kernel/chunked_loss/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+liger_kernel/chunked_loss/fused_linear_preference.py,sha256=WYsIW3t_DEPeC2VkKACimMI0LhZwOciLluDhKnaMmWE,4416
+liger_kernel/chunked_loss/orpo_loss.py,sha256=HgiWjS1ueVpi63ZGVveqwXkXY7EePfSzfbW-bbZlDxE,4220
 liger_kernel/ops/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 liger_kernel/ops/cross_entropy.py,sha256=sfUb7-jIZp0EKXjg1DYy2Wdzw_Mg-mHmGoR5bpdm4tw,15526
 liger_kernel/ops/fused_linear_cross_entropy.py,sha256=JPiQ0TgPjtQ-3F5ovC0b5ZnBk067XUmzyNuGO3KZv44,9963
@@ -43,9 +46,9 @@ liger_kernel/transformers/model/qwen2.py,sha256=EyhSSzQOskGjSnCsKMZpd1s5IAIlHd5P
 liger_kernel/transformers/model/qwen2_vl.py,sha256=j6xAhp9AG195dsZK5f8dFYVM9uKtWApZrggT5Y08jn4,7055
 liger_kernel/triton/__init__.py,sha256=yfRe0zMb47QnqjecZWG7LnanfCTzeku7SgWRAwNVmzU,101
 liger_kernel/triton/monkey_patch.py,sha256=5BcGKTtdqeYchypBIBopGIWPx1-cFALz7sOKoEsqXJ0,1584
-liger_kernel_nightly-0.4.1.dev20241113011623.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel_nightly-0.4.1.dev20241113011623.dist-info/METADATA,sha256=5XnD8ZeAAXku-4RiHLuQKYt2t_v8Pe20QVi_fK8glxY,21556
-liger_kernel_nightly-0.4.1.dev20241113011623.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel_nightly-0.4.1.dev20241113011623.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
-liger_kernel_nightly-0.4.1.dev20241113011623.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel_nightly-0.4.1.dev20241113011623.dist-info/RECORD,,
+liger_kernel_nightly-0.4.1.dev20241114041219.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel_nightly-0.4.1.dev20241114041219.dist-info/METADATA,sha256=o4-CpcoMQuNzIpWwfEGDLui8edp-sWcFlXXZFDSHmlw,21556
+liger_kernel_nightly-0.4.1.dev20241114041219.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel_nightly-0.4.1.dev20241114041219.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
+liger_kernel_nightly-0.4.1.dev20241114041219.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel_nightly-0.4.1.dev20241114041219.dist-info/RECORD,,