liger-kernel-nightly 0.4.2.dev20241121225747__tar.gz → 0.4.2.dev20241122175637__tar.gz

{liger_kernel_nightly-0.4.2.dev20241121225747/src/liger_kernel_nightly.egg-info → liger_kernel_nightly-0.4.2.dev20241122175637}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.4.2.dev20241121225747
+Version: 0.4.2.dev20241122175637
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.4.2.dev20241121225747 → liger_kernel_nightly-0.4.2.dev20241122175637}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.4.2.dev20241121225747"
+version = "0.4.2.dev20241122175637"
 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-0.4.2.dev20241121225747 → liger_kernel_nightly-0.4.2.dev20241122175637}/src/liger_kernel/chunked_loss/dpo_loss.py

@@ -9,15 +9,31 @@ from liger_kernel.chunked_loss.fused_linear_preference import (
 class LigerFusedLinearDPOFunction(LigerFusedLinearPreferenceBase):
 
     @staticmethod
-    def preference_loss_fn(chosen_logps, rejected_logps, beta=0.1):
+    def preference_loss_fn(
+        chosen_logps,
+        rejected_logps,
+        ref_chosen_logps=None,
+        ref_rejected_logps=None,
+        beta=0.1,
+    ):
         """
         Compute DPO loss (Direct Preference Optimization).
         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,).
+            ref_chosen_logps (torch.Tensor, optional): Reference log probabilities of chosen tokens. Shape: (batch_size,).
+            ref_rejected_logps (torch.Tensor, optional): Reference log probabilities of rejected tokens. Shape: (batch_size,).
             beta (float): Weight for the direct preference loss.
         """
-        logits_diff = beta * (chosen_logps - rejected_logps)
+        if ref_chosen_logps is None:
+            ref_chosen_logps = torch.tensor(0.0, device=chosen_logps.device)
+        if ref_rejected_logps is None:
+            ref_rejected_logps = torch.tensor(0.0, device=rejected_logps.device)
+
+        chosen_logratios = chosen_logps - ref_chosen_logps
+        rejected_logratios = rejected_logps - ref_rejected_logps
+
+        logits_diff = beta * (chosen_logratios - rejected_logratios)
         losses = -F.logsigmoid(logits_diff)
         return losses.sum()
 
@@ -28,10 +44,13 @@ class LigerFusedLinearDPOFunction(LigerFusedLinearPreferenceBase):
         weight,
         target,
         bias=None,
+        ref_weight=None,
+        ref_bias=None,
         ignore_index=-100,
         beta=0.1,
         compute_nll_loss=True,
         compiled=True,
+        use_ref_model=True,
     ):
         """
         Fused linear layer with DPO (Direct Preference Optimization) loss.
@@ -48,6 +67,9 @@ class LigerFusedLinearDPOFunction(LigerFusedLinearPreferenceBase):
             beta=beta,
             compute_nll_loss=compute_nll_loss,
             compiled=compiled,
+            use_ref_model=use_ref_model,
+            ref_weight=ref_weight,
+            ref_bias=ref_bias,
         )
 
     @staticmethod
@@ -55,7 +77,7 @@ class LigerFusedLinearDPOFunction(LigerFusedLinearPreferenceBase):
         # 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
+        return *grads, None, None, None, None, None, None, None
 
 
 class LigerFusedLinearDPOLoss(torch.nn.Module):
@@ -69,26 +91,36 @@ class LigerFusedLinearDPOLoss(torch.nn.Module):
         beta: float = 0.1,
         compute_nll_loss: bool = True,
         compiled: bool = True,
+        use_ref_model: 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 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.
         """
         super().__init__()
         self.ignore_index = ignore_index
         self.beta = beta
         self.compute_nll_loss = compute_nll_loss
         self.compiled = compiled
+        self.use_ref_model = use_ref_model
 
-    def forward(self, lin_weight, _input, target, bias=None):
+    def forward(
+        self, lin_weight, _input, target, bias=None, ref_weight=None, ref_bias=None
+    ):
         return LigerFusedLinearDPOFunction.apply(
             _input,
             lin_weight,
             target,
             bias,
+            ref_weight,
+            ref_bias,
             self.ignore_index,
             self.beta,
             self.compute_nll_loss,
             self.compiled,
+            self.use_ref_model,
         )

{liger_kernel_nightly-0.4.2.dev20241121225747 → liger_kernel_nightly-0.4.2.dev20241122175637}/src/liger_kernel/chunked_loss/fused_linear_preference.py

@@ -18,6 +18,42 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
         """
         raise NotImplementedError("Preference loss function must be implemented.")
 
+    @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:]
+        return chosen_logps, rejected_logps, chosen_nll_loss
+
     @staticmethod
     def forward(
         ctx,
@@ -32,6 +68,9 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
         beta=0.1,
         compute_nll_loss=True,
         compiled=True,
+        use_ref_model=False,
+        ref_weight=None,
+        ref_bias=None,
         **loss_kwargs,
     ):
         """
@@ -49,7 +88,11 @@ 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 odds ratio 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
@@ -61,7 +104,6 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
         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))
         loss_func_to_call = partial(
             LigerFusedLinearPreferenceBase._compute_loss,
             preference_loss_fn=loss_fn,
@@ -70,6 +112,9 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
             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,
         )
 
@@ -101,6 +146,7 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
             accumulate_chunk = torch.compile(accumulate_chunk)
 
         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)
@@ -159,6 +205,9 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
         alpha=1.0,
         beta=0.1,
         compute_nll_loss=True,
+        use_ref_model=False,
+        ref_weight=None,
+        ref_bias=None,
         **loss_kwargs,
     ):
         """
@@ -173,38 +222,41 @@ 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 odds ratio 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.
         """
-        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",
+        chosen_logps, rejected_logps, 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()
-            )
-
-        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_nll_loss = (
+            chosen_nll_loss
+            / (full_target[: full_target.shape[0] // 2] != ignore_index).sum()
+        )
 
-        chosen_logps = average_log_prob[:len_chosen_chunk]
-        rejected_logps = average_log_prob[len_chosen_chunk:]
+        if use_ref_model:
+            with torch.no_grad():
+                ref_chosen_logps, ref_rejected_logps, _ = (
+                    LigerFusedLinearPreferenceBase.chunk_forward(
+                        input_chunk,
+                        ref_weight,
+                        target_chunk,
+                        ref_bias,
+                        ignore_index=ignore_index,
+                        compute_nll_loss=False,
+                    )
+                )
+            loss_kwargs["ref_chosen_logps"] = ref_chosen_logps
+            loss_kwargs["ref_rejected_logps"] = ref_rejected_logps
 
         alignment_loss = preference_loss_fn(
             chosen_logps, rejected_logps, beta=beta, **loss_kwargs

{liger_kernel_nightly-0.4.2.dev20241121225747 → liger_kernel_nightly-0.4.2.dev20241122175637}/src/liger_kernel/ops/cross_entropy.py

@@ -92,8 +92,8 @@ def liger_cross_entropy_kernel(
     # 3. [Online softmax] first pass: find max + sum
     m = float("-inf")  # m is the max value. use the notation from the paper
     d = 0.0  # d is the sum. use the notation from the paper
-    ori_X_y = tl.load(
-        X_ptr + y
+    ori_X_y = tl.load(X_ptr + y).cast(
+        tl.float32
     )  # we need to store the original value of X_y for the loss calculation
     if HAS_SOFTCAPPING:
         ori_X_y = softcap * tanh(ori_X_y / softcap)
@@ -106,8 +106,11 @@ def liger_cross_entropy_kernel(
     for i in range(0, n_cols, BLOCK_SIZE):
         X_offsets = i + tl.arange(0, BLOCK_SIZE)
         X_block = tl.load(
-            X_ptr + X_offsets, mask=X_offsets < n_cols, other=float("-inf")
-        )
+            X_ptr + X_offsets,
+            mask=X_offsets < n_cols,
+            other=float("-inf"),
+            # Ensure float32 precision for softmax calculation
+        ).cast(tl.float32)
         if HAS_SOFTCAPPING:
             X_block = softcap * tanh(X_block / softcap)
         block_max = tl.max(X_block)
@@ -141,8 +144,11 @@ def liger_cross_entropy_kernel(
     for i in range(0, n_cols, BLOCK_SIZE):
         X_offsets = i + tl.arange(0, BLOCK_SIZE)
         X_block = tl.load(
-            X_ptr + X_offsets, mask=X_offsets < n_cols, other=float("-inf")
-        )
+            X_ptr + X_offsets,
+            mask=X_offsets < n_cols,
+            other=float("-inf"),
+            # Ensure float32 precision for softmax calculation
+        ).cast(tl.float32)
         if HAS_SOFTCAPPING:
             intermediate = tanh(X_block / softcap)
             X_block = softcap * intermediate

{liger_kernel_nightly-0.4.2.dev20241121225747 → liger_kernel_nightly-0.4.2.dev20241122175637}/src/liger_kernel/ops/fused_linear_cross_entropy.py

@@ -26,7 +26,6 @@ def fused_linear_cross_entropy_forward(
     reduction="mean",
     softcap=None,
 ):
-    dtype = _input.dtype
     device = _input.device
 
     # inputs have shape: BT x H
@@ -74,9 +73,6 @@ def fused_linear_cross_entropy_forward(
         loss_1d_slice = loss_1d[start_idx:end_idx]  # chunk_size,
         n_non_ignore = (target_chunk != ignore_index).sum().item()
 
-        # when doing CE, use the upcasted precision
-        logits_chunk = logits_chunk.float()
-
         # ensure _input and target are contiguous
         logits_chunk = logits_chunk.contiguous()
         target_chunk = target_chunk.contiguous()
@@ -103,13 +99,6 @@ def fused_linear_cross_entropy_forward(
             num_warps=32 if not is_hip() else 16,
         )
 
-        # gradient of logits_chunk is computed in-place by the above triton kernel.
-        # Following HuggingFace model source code, we do the forward and backward
-        # w.r.t. logits in fp32 for numerical stability especially as the num classes (vocab size) is huge.
-        # (reference: https://github.com/huggingface/transformers/blob/v4.42.4/src/transformers/models/llama/modeling_llama.py#L1194)
-        # Propagating to lm_head's backward, we'll switch back to the original dtype.
-        logits_chunk = logits_chunk.to(dtype)
-
         # gradient of logits_chunk is computed in-place by the above triton kernel and is of shape: chunk_size x V
         # thus grad_input[start_idx: end_idx] should be of shape: chunk_size x H
         # additionally, since we are chunking the inputs, observe that the loss and gradients are calculated only

{liger_kernel_nightly-0.4.2.dev20241121225747 → liger_kernel_nightly-0.4.2.dev20241122175637/src/liger_kernel_nightly.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.4.2.dev20241121225747
+Version: 0.4.2.dev20241122175637
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation