Adding support for apo losses, sppo_hard and nca_pair (#841)

## Summary
<!--- This is a required section; please describe the main purpose of
this proposed code change. --->
This pr adds support for apo zero, apo down, sppo_hard and nca_pair loss
just like in
(https://github.com/huggingface/trl/blob/main/trl/trainer/dpo_trainer.py).
<!---
## Details
This is an optional section; is there anything specific that reviewers
should be aware of?
--->

## Testing Done
<!--- This is a required section; please describe how this change was
tested. --->
python -m pytest test/chunked_loss/test_dpo_loss.py
<!-- 
Replace BLANK with your device type. For example, A100-80G-PCIe

Complete the following tasks before sending your PR, and replace `[ ]`
with
`[x]` to indicate you have done them. 
-->

- Hardware Type: H100
- [ ] run `make test` to ensure correctness
- [x] run `make checkstyle` to ensure code style
- [ ] run `make test-convergence` to ensure convergence

---------

Co-authored-by: Manan Shah <[email protected]>
Co-authored-by: Vaibhav Jindal <[email protected]>

Author: 3 people

src/liger_kernel/chunked_loss/dpo_loss.py

@@ -13,6 +13,7 @@ def preference_loss_fn(
         ref_chosen_logps=None,
         ref_rejected_logps=None,
         beta=0.1,
+        loss_type="sigmoid",
     ):
         """
         Paper: https://arxiv.org/pdf/2305.18290
@@ -48,8 +49,50 @@ def preference_loss_fn(
         chosen_rewards = beta * chosen_logratios
         rejected_rewards = beta * rejected_logratios
 
-        logits_diff = beta * (chosen_logratios - rejected_logratios)
-        loss = -F.logsigmoid(logits_diff).sum() / (full_target.shape[0] // 2)
+        if loss_type == "sigmoid":
+            logits_diff = beta * (chosen_logratios - rejected_logratios)
+            loss = -F.logsigmoid(logits_diff).sum() / (full_target.shape[0] // 2)
+
+        elif loss_type == "apo_zero":
+            # Eqn (7) of the APO paper (https://huggingface.co/papers/2408.06266)
+            # Use this loss when you believe the chosen outputs are better than your model's default output
+            losses_chosen = 1 - F.sigmoid(beta * chosen_logratios)  # Increase chosen likelihood
+            losses_rejected = F.sigmoid(beta * rejected_logratios)
+            losses = losses_chosen + losses_rejected
+            loss = losses.sum() / (full_target.shape[0] // 2)
+
+        elif loss_type == "apo_down":
+            # Eqn (8) of the APO paper (https://huggingface.co/papers/2408.06266)
+            # Use this loss when you believe the chosen outputs are worse than your model's default output.
+            # Decrease chosen likelihood and decrease rejected likelihood more
+            losses_chosen = F.sigmoid(beta * chosen_logratios)
+            losses_rejected = 1 - F.sigmoid(beta * (chosen_logratios - rejected_logratios))
+            losses = losses_chosen + losses_rejected
+            loss = losses.sum() / (full_target.shape[0] // 2)
+
+        elif loss_type == "sppo_hard":
+            # In the paper (https://huggingface.co/papers/2405.00675), SPPO employs a soft probability approach,
+            # estimated using the PairRM score. The probability calculation is conducted outside of the trainer class.
+            # The version described here is the hard probability version, where P in Equation (4.7) of Algorithm 1 is
+            # set to 1 for the winner and 0 for the loser.
+            a = chosen_logps - ref_chosen_logps
+            b = rejected_logps - ref_rejected_logps
+            losses = (a - 0.5 / beta) ** 2 + (b + 0.5 / beta) ** 2
+            loss = losses.sum() / (full_target.shape[0] // 2)
+
+        elif loss_type == "nca_pair":
+            losses = (
+                -F.logsigmoid(chosen_rewards)
+                - 0.5 * F.logsigmoid(-chosen_rewards)
+                - 0.5 * F.logsigmoid(-rejected_rewards)
+            )
+            loss = losses.sum() / (full_target.shape[0] // 2)
+
+        else:
+            raise ValueError(
+                f"Unsupported loss_type: {loss_type}. Supported types are: sigmoid, apo_zero, apo_down, sppo_hard, nca_pair"
+            )
+
         return loss, chosen_rewards, rejected_rewards
 
     @classmethod
@@ -70,6 +113,7 @@ def forward(
         use_ref_model=True,
         average_log_prob=False,
         chunk_size=1,
+        loss_type="sigmoid",
     ):
         """
         Fused linear layer with DPO loss.
@@ -108,12 +152,13 @@ def forward(
             ref_bias=ref_bias,
             average_log_prob=average_log_prob,
             chunk_size=chunk_size,
+            loss_type=loss_type,
         )
 
     @staticmethod
     def backward(ctx, *grad_output):
         grads = LigerFusedLinearPreferenceBase.backward(ctx, grad_output)[:4]
-        return *grads, None, None, None, None, None, None, None, None, None, None
+        return *grads, None, None, None, None, None, None, None, None, None, None, None
 
 
 class LigerFusedLinearDPOLoss(torch.nn.Module):
@@ -130,6 +175,7 @@ def __init__(
         use_ref_model: bool = True,
         average_log_prob: bool = False,
         chunk_size: int = 1,
+        loss_type: str = "sigmoid",
     ):
         """
         Args:
@@ -149,6 +195,10 @@ def __init__(
         self.use_ref_model = use_ref_model
         self.average_log_prob = average_log_prob
         self.chunk_size = chunk_size
+        self.loss_type = loss_type
+        supported_loss_types = {"sigmoid", "apo_zero", "apo_down", "sppo_hard", "nca_pair"}
+        if self.loss_type not in supported_loss_types:
+            raise ValueError(f"Unsupported loss_type: {self.loss_type}. Supported types are: {supported_loss_types}")
 
     def forward(
         self,
@@ -175,4 +225,5 @@ def forward(
             self.use_ref_model,
             self.average_log_prob,
             self.chunk_size,
+            self.loss_type,
         )