liger-kernel-nightly 0.5.2.dev20241216214323__tar.gz → 0.5.2.dev20241218221959__tar.gz

{liger_kernel_nightly-0.5.2.dev20241216214323 → liger_kernel_nightly-0.5.2.dev20241218221959}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.2.dev20241216214323
+Version: 0.5.2.dev20241218221959
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.5.2.dev20241216214323 → liger_kernel_nightly-0.5.2.dev20241218221959}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.5.2.dev20241216214323"
+version = "0.5.2.dev20241218221959"
 description = "Efficient Triton kernels for LLM Training"
 urls = { "Homepage" = "https://github.com/linkedin/Liger-Kernel" }
 readme = { file = "README.md", content-type = "text/markdown" }

{liger_kernel_nightly-0.5.2.dev20241216214323 → liger_kernel_nightly-0.5.2.dev20241218221959}/src/liger_kernel/chunked_loss/cpo_loss.py

@@ -47,6 +47,7 @@ class LigerFusedLinearCPOFunction(LigerFusedLinearPreferenceBase):
         alpha=1.0,
         compute_nll_loss=True,
         compiled=True,
+        is_encoder_decoder=False,
     ):
         return LigerFusedLinearPreferenceBase.forward(
             ctx,
@@ -60,12 +61,13 @@ class LigerFusedLinearCPOFunction(LigerFusedLinearPreferenceBase):
             beta=beta,
             compute_nll_loss=compute_nll_loss,
             compiled=compiled,
+            is_encoder_decoder=is_encoder_decoder,
         )
 
     @staticmethod
     def backward(ctx, *grad_output):
         grads = LigerFusedLinearPreferenceBase.backward(ctx, grad_output)[:4]
-        return *grads, None, None, None, None, None
+        return *grads, None, None, None, None, None, None
 
 
 class LigerFusedLinearCPOLoss(torch.nn.Module):
@@ -80,11 +82,16 @@ class LigerFusedLinearCPOLoss(torch.nn.Module):
         alpha: float = 1.0,
         compute_nll_loss: bool = True,
         compiled: bool = True,
+        is_encoder_decoder: bool = False,
     ):
         """
         Args:
             ignore_index (int): Index to ignore in the loss.
             beta (float): Weight for the odds ratio loss.
+            alpha (float): Weight for the NLL loss.
+            compute_nll_loss (bool): Whether to compute NLL loss.
+            compiled (bool): Whether to compile the loss function.
+            is_encoder_decoder (bool): Whether the model is an encoder-decoder model.
         """
         super().__init__()
         self.ignore_index = ignore_index
@@ -92,6 +99,7 @@ class LigerFusedLinearCPOLoss(torch.nn.Module):
         self.alpha = alpha
         self.compute_nll_loss = compute_nll_loss
         self.compiled = compiled
+        self.is_encoder_decoder = is_encoder_decoder
 
     def forward(self, lin_weight, _input, target, bias=None):
         return LigerFusedLinearCPOFunction.apply(
@@ -104,4 +112,5 @@ class LigerFusedLinearCPOLoss(torch.nn.Module):
             self.alpha,
             self.compute_nll_loss,
             self.compiled,
+            self.is_encoder_decoder,
         )

{liger_kernel_nightly-0.5.2.dev20241216214323 → liger_kernel_nightly-0.5.2.dev20241218221959}/src/liger_kernel/chunked_loss/dpo_loss.py

@@ -67,6 +67,7 @@ class LigerFusedLinearDPOFunction(LigerFusedLinearPreferenceBase):
         compute_nll_loss=True,
         compiled=True,
         use_ref_model=True,
+        is_encoder_decoder=False,
     ):
         return LigerFusedLinearPreferenceBase.forward(
             ctx=ctx,
@@ -83,12 +84,13 @@ class LigerFusedLinearDPOFunction(LigerFusedLinearPreferenceBase):
             ref_input=ref_input,
             ref_weight=ref_weight,
             ref_bias=ref_bias,
+            is_encoder_decoder=is_encoder_decoder,
         )
 
     @staticmethod
     def backward(ctx, *grad_output):
         grads = LigerFusedLinearPreferenceBase.backward(ctx, grad_output)[:4]
-        return *grads, None, None, None, None, None, None, None, None
+        return *grads, None, None, None, None, None, None, None, None, None
 
 
 class LigerFusedLinearDPOLoss(torch.nn.Module):
@@ -103,6 +105,7 @@ class LigerFusedLinearDPOLoss(torch.nn.Module):
         compute_nll_loss: bool = True,
         compiled: bool = True,
         use_ref_model: bool = False,
+        is_encoder_decoder: bool = False,
     ):
         """
         Args:
@@ -111,6 +114,7 @@ class LigerFusedLinearDPOLoss(torch.nn.Module):
             compute_nll_loss (bool): Whether to compute the NLL loss.
             compiled (bool): Whether to use the torch compiled kernel.
             use_ref_model (bool): Whether to use a reference model for the DPO loss.
+            is_encoder_decoder (bool): Whether the model is an encoder-decoder model.
         """
         super().__init__()
         self.ignore_index = ignore_index
@@ -118,6 +122,7 @@ class LigerFusedLinearDPOLoss(torch.nn.Module):
         self.compute_nll_loss = compute_nll_loss
         self.compiled = compiled
         self.use_ref_model = use_ref_model
+        self.is_encoder_decoder = is_encoder_decoder
 
     def forward(
         self,
@@ -142,4 +147,5 @@ class LigerFusedLinearDPOLoss(torch.nn.Module):
             self.compute_nll_loss,
             self.compiled,
             self.use_ref_model,
+            self.is_encoder_decoder,
         )

{liger_kernel_nightly-0.5.2.dev20241216214323 → liger_kernel_nightly-0.5.2.dev20241218221959}/src/liger_kernel/chunked_loss/fused_linear_preference.py

@@ -26,6 +26,7 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
         ignore_index=-100,
         alpha=1.0,
         beta=0.1,
+        is_encoder_decoder=False,
         compute_nll_loss=True,
         compiled=True,
         use_ref_model=False,
@@ -56,6 +57,7 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
             ignore_index (int): Index to ignore for loss computation.
             alpha (float): Weight for the NLL loss.
             beta (float): Weight for the preference loss.
+            is_encoder_decoder (bool): Whether the model is an encoder-decoder model.
             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.
@@ -94,6 +96,7 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
             use_ref_model=use_ref_model,
             ref_weight=ref_weight,
             ref_bias=ref_bias,
+            is_encoder_decoder=is_encoder_decoder,
             **loss_kwargs,
         )
 
@@ -282,33 +285,48 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
         bias=None,
         ignore_index=-100,
         compute_nll_loss=True,
+        is_encoder_decoder=False,
     ):
-        len_chosen_chunk = target_chunk.shape[0] // 2
+        # Calculate logits and log probabilities
         logits_chunk = input_chunk @ weight.t()
         if bias is not None:
-            logits_chunk = logits_chunk + bias
+            logits_chunk += bias
         log_probs_chunk = F.log_softmax(logits_chunk.float(), dim=-1)
 
+        # Split chunk into chosen and rejected portions
+        len_chosen_chunk = target_chunk.shape[0] // 2
+
+        # Handle sequence shifting for non-encoder-decoder models
+        if not is_encoder_decoder:
+            logits_chunk = logits_chunk[:, :-1]
+            log_probs_chunk = log_probs_chunk[:, :-1]
+            target_chunk = target_chunk[:, 1:]
+
+        # Calculate NLL loss for chosen sequences
         chosen_nll_loss = 0.0
         if compute_nll_loss:
+            chosen_probs = log_probs_chunk[:len_chosen_chunk]
+            chosen_targets = target_chunk[:len_chosen_chunk]
             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),
+                chosen_probs.reshape(-1, chosen_probs.shape[-1]),
+                chosen_targets.reshape(-1),
                 reduction="sum",
                 ignore_index=ignore_index,
             )
 
+        # Calculate per-token log probabilities
         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:]
-
+        # Split results for chosen and rejected
+        chosen_logps, rejected_logps = (
+            average_log_prob[:len_chosen_chunk],
+            average_log_prob[len_chosen_chunk:],
+        )
         chosen_logits = logits_chunk[:len_chosen_chunk]
         rejected_logits = logits_chunk[len_chosen_chunk:]
 
@@ -331,6 +349,7 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
         ignore_index=-100,
         alpha=1.0,
         beta=0.1,
+        is_encoder_decoder=False,
         compute_nll_loss=True,
         use_ref_model=False,
         ref_input_chunk=None,
@@ -350,6 +369,7 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
             ignore_index (int): Index to ignore for loss computation.
             alpha (float): Weight for the NLL loss.
             beta (float): Weight for the preference loss.
+            is_encoder_decoder (bool): Whether the model is an encoder-decoder model.
             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).
@@ -369,33 +389,43 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
             bias=bias,
             ignore_index=ignore_index,
             compute_nll_loss=compute_nll_loss,
+            is_encoder_decoder=is_encoder_decoder,
         )
-        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 not is_encoder_decoder:
+            chosen_nll_loss = (
+                chosen_nll_loss
+                / (full_target[: full_target.shape[0] // 2, 1:] != ignore_index).sum()
+            )
+            chosen_logits_mean = chosen_logits.sum() / (
+                full_target.shape[0] // 2 * (input_chunk.shape[1] - 1) * weight.shape[0]
+            )
+            rejected_logits_mean = rejected_logits.sum() / (
+                full_target.shape[0] // 2 * (input_chunk.shape[1] - 1) * weight.shape[0]
+            )
+        else:
+            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(
-                    ref_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
+                (ref_chosen_logps, ref_rejected_logps, _, _, _) = (
+                    LigerFusedLinearPreferenceBase.chunk_forward(
+                        ref_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
+                        is_encoder_decoder=is_encoder_decoder,  # assume the ref model is the same family
+                    )
                 )
             loss_kwargs["ref_chosen_logps"] = ref_chosen_logps
             loss_kwargs["ref_rejected_logps"] = ref_rejected_logps

{liger_kernel_nightly-0.5.2.dev20241216214323 → liger_kernel_nightly-0.5.2.dev20241218221959}/src/liger_kernel/chunked_loss/orpo_loss.py

@@ -57,6 +57,7 @@ class LigerFusedLinearORPOFunction(LigerFusedLinearPreferenceBase):
         beta=0.1,
         compute_nll_loss=True,
         compiled=True,
+        is_encoder_decoder=False,
     ):
         return LigerFusedLinearPreferenceBase.forward(
             ctx=ctx,
@@ -69,12 +70,13 @@ class LigerFusedLinearORPOFunction(LigerFusedLinearPreferenceBase):
             beta=beta,
             compute_nll_loss=compute_nll_loss,
             compiled=compiled,
+            is_encoder_decoder=is_encoder_decoder,
         )
 
     @staticmethod
     def backward(ctx, *grad_output):
         grads = LigerFusedLinearPreferenceBase.backward(ctx, grad_output)[:4]
-        return *grads, None, None, None, None
+        return *grads, None, None, None, None, None
 
 
 class LigerFusedLinearORPOLoss(torch.nn.Module):
@@ -88,17 +90,22 @@ class LigerFusedLinearORPOLoss(torch.nn.Module):
         beta: float = 0.1,
         compute_nll_loss: bool = True,
         compiled: bool = True,
+        is_encoder_decoder: bool = False,
     ):
         """
         Args:
             ignore_index (int): Index to ignore in the loss.
             beta (float): Weight for the odds ratio loss.
+            compute_nll_loss (bool): Whether to compute NLL loss.
+            compiled (bool): Whether to compile the loss function.
+            is_encoder_decoder (bool): Whether the model is an encoder-decoder model.
         """
         super().__init__()
         self.ignore_index = ignore_index
         self.beta = beta
         self.compute_nll_loss = compute_nll_loss
         self.compiled = compiled
+        self.is_encoder_decoder = is_encoder_decoder
 
     def forward(self, lin_weight, _input, target, bias=None):
         return LigerFusedLinearORPOFunction.apply(
@@ -110,4 +117,5 @@ class LigerFusedLinearORPOLoss(torch.nn.Module):
             self.beta,
             self.compute_nll_loss,
             self.compiled,
+            self.is_encoder_decoder,
         )

{liger_kernel_nightly-0.5.2.dev20241216214323 → liger_kernel_nightly-0.5.2.dev20241218221959}/src/liger_kernel/transformers/model/mixtral.py

@@ -38,7 +38,7 @@ def lce_forward_deprecated(
     cache_position: Optional[torch.LongTensor] = None,
 ) -> Union[Tuple, MoeCausalLMOutputWithPast]:
     r"""
-    Copy paste Mixtral's forward from transfomers v4.44.2 but replace torch cross entropy with liger fused linear cross entropy
+    Copy paste Mixtral's forward from transformers v4.44.2 but replace torch cross entropy with liger fused linear cross entropy
 
 
     Args:

{liger_kernel_nightly-0.5.2.dev20241216214323 → liger_kernel_nightly-0.5.2.dev20241218221959}/src/liger_kernel/transformers/trainer/orpo_trainer.py

@@ -17,7 +17,7 @@ class _FSDPForwardRedirection:
     This is needed in cases where we call a submodule of a FSDP module. For instance, when we want to call only
     the `LlamaModel` part out of a FSDP-wrapped `LlamaForCausalLM` to get the hidden states without involving
     GPU-memory-heavy `lm_head` and cross entropy computation, doing this directly (i.e. `model.model.forward()`)
-    will not work because the first `nn.Emebedding` layer is not independently wrapped as a FSDP module (because of
+    will not work because the first `nn.Embedding` layer is not independently wrapped as a FSDP module (because of
     the transformer-based wrapping policy), and not calling it through FSDP root module forward will not all-gather
     its parameter, thus resulting in "RuntimeError: 'weight' must be 2-D" error. Similarly, if we want to call just
     the `lm_head` part of a model, we need this trick too to properly get its params all-gathered.

{liger_kernel_nightly-0.5.2.dev20241216214323 → liger_kernel_nightly-0.5.2.dev20241218221959}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.2.dev20241216214323
+Version: 0.5.2.dev20241218221959
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.5.2.dev20241216214323 → liger_kernel_nightly-0.5.2.dev20241218221959}/test/utils.py

@@ -350,11 +350,13 @@ class HFAlignmentLoss:
         beta: float = 0.1,
         ignore_index: int = -100,
         use_ref_model: bool = False,
+        is_encoder_decoder: bool = False,
     ):
         self.alpha = alpha
         self.beta = beta
         self.ignore_index = ignore_index
         self.use_ref_model = use_ref_model
+        self.is_encoder_decoder = is_encoder_decoder
 
     @abstractmethod
     def alignment_loss(self):
@@ -372,7 +374,6 @@ class HFAlignmentLoss:
             logits: Logits of the model (unnormalized). Shape: (batch_size, sequence_length, vocab_size)
             labels: Labels for which to compute the log probabilities. Label tokens with a value of ignore_index are ignored. Shape: (batch_size, sequence_length)
             average_log_prob: If True, return the average log probability per (non-masked) token. Otherwise, return the sum of the log probabilities of the (non-masked) tokens.
-            is_encoder_decoder: Whether the model is an encoder-decoder model.
         Returns:
             A tensor of shape (batch_size,) containing the average/sum log probabilities of the given labels under the given logits.
         """
@@ -381,6 +382,9 @@ class HFAlignmentLoss:
                 "Logits (batch and sequence length dim) and labels must have the same shape."
             )
 
+        if not self.is_encoder_decoder:
+            logits = logits[..., :-1, :].contiguous()
+            labels = labels[..., 1:].contiguous()
         loss_mask = labels != self.ignore_index
 
         # dummy token; we'll ignore the losses on these tokens later
@@ -440,6 +444,9 @@ class HFAlignmentLoss:
         def cross_entropy_loss(logits, labels):
             # Flatten the tokens
             loss_fct = nn.CrossEntropyLoss(ignore_index=self.ignore_index)
+            if not self.is_encoder_decoder:
+                logits = logits[..., :-1, :].contiguous()
+                labels = labels[..., 1:].contiguous()
             logits = logits.view(-1, logits.shape[-1])
             labels = labels.view(-1)
             # Enable model parallelism
@@ -461,8 +468,12 @@ class HFAlignmentLoss:
         chosen_logps = all_logps[:len_chosen]
         rejected_logps = all_logps[len_chosen:]
 
-        chosen_logits = all_logits[:len_chosen]
-        rejected_logits = all_logits[len_chosen:]
+        if not self.is_encoder_decoder:
+            chosen_logits = all_logits[:len_chosen, :-1]
+            rejected_logits = all_logits[len_chosen:, :-1]
+        else:
+            chosen_logits = all_logits[:len_chosen]
+            rejected_logits = all_logits[len_chosen:]
 
         return (
             chosen_logps,