liger-kernel-nightly 0.5.2.dev20241212030605__py3-none-any.whl → 0.5.2.dev20241212055403__py3-none-any.whl

liger_kernel/chunked_loss/dpo_loss.py

@@ -59,6 +59,7 @@ class LigerFusedLinearDPOFunction(LigerFusedLinearPreferenceBase):
         weight,
         target,
         bias=None,
+        ref_input=None,
         ref_weight=None,
         ref_bias=None,
         ignore_index=-100,
@@ -79,6 +80,7 @@ class LigerFusedLinearDPOFunction(LigerFusedLinearPreferenceBase):
             compute_nll_loss=compute_nll_loss,
             compiled=compiled,
             use_ref_model=use_ref_model,
+            ref_input=ref_input,
             ref_weight=ref_weight,
             ref_bias=ref_bias,
         )
@@ -86,7 +88,7 @@ class LigerFusedLinearDPOFunction(LigerFusedLinearPreferenceBase):
     @staticmethod
     def backward(ctx, *grad_output):
         grads = LigerFusedLinearPreferenceBase.backward(ctx, grad_output)[:4]
-        return *grads, None, None, None, None, None, None, None
+        return *grads, None, None, None, None, None, None, None, None
 
 
 class LigerFusedLinearDPOLoss(torch.nn.Module):
@@ -118,13 +120,21 @@ class LigerFusedLinearDPOLoss(torch.nn.Module):
         self.use_ref_model = use_ref_model
 
     def forward(
-        self, lin_weight, _input, target, bias=None, ref_weight=None, ref_bias=None
+        self,
+        lin_weight,
+        _input,
+        target,
+        bias=None,
+        ref_input=None,
+        ref_weight=None,
+        ref_bias=None,
     ):
         return LigerFusedLinearDPOFunction.apply(
             _input,
             lin_weight,
             target,
             bias,
+            ref_input,
             ref_weight,
             ref_bias,
             self.ignore_index,

liger_kernel/chunked_loss/fused_linear_preference.py

@@ -29,7 +29,7 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
         compute_nll_loss=True,
         compiled=True,
         use_ref_model=False,
-        # TODO: ref input
+        ref_input=None,
         ref_weight=None,
         ref_bias=None,
         **loss_kwargs,
@@ -97,20 +97,26 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
             **loss_kwargs,
         )
 
-        def fused_fwd_bwd(input_chunk, target_chunk):
+        def fused_fwd_bwd(input_chunk, target_chunk, ref_input_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)
+                )(
+                    input_chunk,
+                    weight,
+                    target_chunk,
+                    bias,
+                    ref_input_chunk=ref_input_chunk,
+                )
             else:
                 return torch.func.grad_and_value(
                     compute_loss, argnums=(0, 1), has_aux=True
-                )(input_chunk, weight, target_chunk)
+                )(input_chunk, weight, target_chunk, ref_input_chunk=ref_input_chunk)
 
-        def accumulate_chunk(input_chunk, target_chunk):
+        def accumulate_chunk(input_chunk, target_chunk, ref_input_chunk=None):
             if bias is not None:
                 (chunk_grad_input, chunk_grad_weight, chunk_grad_bias), (
                     chunk_loss,
@@ -122,7 +128,7 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
                         chunk_nll_loss,
                         *aux_outputs,
                     ),
-                ) = fused_fwd_bwd(input_chunk, target_chunk)
+                ) = fused_fwd_bwd(input_chunk, target_chunk, ref_input_chunk)
                 grad_bias.add_(chunk_grad_bias)  # accumulate bias gradient
             else:
                 (chunk_grad_input, chunk_grad_weight), (
@@ -135,7 +141,7 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
                         chunk_nll_loss,
                         *aux_outputs,
                     ),
-                ) = fused_fwd_bwd(input_chunk, target_chunk)
+                ) = fused_fwd_bwd(input_chunk, target_chunk, ref_input_chunk)
 
             # Accumulate gradients
             grad_weight.add_(chunk_grad_weight)
@@ -182,18 +188,43 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
         _rejected_input_chunks = torch.chunk(_input[len_chosen:], chunks=chunks, dim=0)
         _rejected_target_chunks = torch.chunk(target[len_chosen:], chunks=chunks, dim=0)
 
+        if use_ref_model:
+            _ref_chosen_input_chunks = torch.chunk(
+                ref_input[:len_chosen], chunks=chunks, dim=0
+            )
+            _ref_rejected_input_chunks = torch.chunk(
+                ref_input[len_chosen:], chunks=chunks, dim=0
+            )
+
         for (
             chosen_input_chunk,
             rejected_input_chunk,
             chosen_target_chunk,
             rejected_target_chunk,
+            ref_chosen_input_chunk,
+            ref_rejected_input_chunk,
         ) in zip(
             _chosen_input_chunks,
             _rejected_input_chunks,
             _chosen_target_chunks,
             _rejected_target_chunks,
+            (
+                _ref_chosen_input_chunks
+                if use_ref_model
+                else [None] * len(_chosen_input_chunks)
+            ),
+            (
+                _ref_rejected_input_chunks
+                if use_ref_model
+                else [None] * len(_rejected_input_chunks)
+            ),
         ):
             input_chunk = torch.cat([chosen_input_chunk, rejected_input_chunk], dim=0)
+            ref_input_chunk = (
+                torch.cat([ref_chosen_input_chunk, ref_rejected_input_chunk], dim=0)
+                if use_ref_model
+                else None
+            )
             target_chunk = torch.cat(
                 [chosen_target_chunk, rejected_target_chunk], dim=0
             )
@@ -202,9 +233,10 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
             torch._dynamo.mark_dynamic(input_chunk, 1)
             torch._dynamo.mark_dynamic(target_chunk, 1)
             torch._dynamo.mark_dynamic(target, 1)
+            torch._dynamo.mark_dynamic(ref_input_chunk, 1) if use_ref_model else None
 
             # accumulate loss, gradients, and metrics
-            accumulate_chunk(input_chunk, target_chunk)
+            accumulate_chunk(input_chunk, target_chunk, ref_input_chunk)
 
         # combine grad_chosen_inputs and grad_rejected_inputs
         grad_inputs = grad_chosen_inputs + grad_rejected_inputs
@@ -301,6 +333,7 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
         beta=0.1,
         compute_nll_loss=True,
         use_ref_model=False,
+        ref_input_chunk=None,
         ref_weight=None,
         ref_bias=None,
         **loss_kwargs,
@@ -357,7 +390,7 @@ class LigerFusedLinearPreferenceBase(torch.autograd.Function):
                     ref_rejected_logits,
                     ref_chosen_nll_loss,
                 ) = LigerFusedLinearPreferenceBase.chunk_forward(
-                    input_chunk,
+                    ref_input_chunk,
                     ref_weight,
                     target_chunk,
                     ref_bias,

{liger_kernel_nightly-0.5.2.dev20241212030605.dist-info → liger_kernel_nightly-0.5.2.dev20241212055403.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.2.dev20241212030605
+Version: 0.5.2.dev20241212055403
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.5.2.dev20241212030605.dist-info → liger_kernel_nightly-0.5.2.dev20241212055403.dist-info}/RECORD

@@ -4,10 +4,10 @@ liger_kernel/utils.py,sha256=HJa-xVKOohDn6pLVIx-Fv0V9h0QAL3qZGQNRICI-OpI,249
 liger_kernel/chunked_loss/README.md,sha256=K6rucm6nqHpWCmxUOhBYcE3apwQxAy0TfRUippR7Icw,2243
 liger_kernel/chunked_loss/__init__.py,sha256=R2wCcz4Y0kTAve926DH3k182XKezpXeACMHj05g9Mm8,346
 liger_kernel/chunked_loss/cpo_loss.py,sha256=Qu1Ul2A12sp6CqIT-atPbHWFb_LLtINEA9mOpIRx_0g,3097
-liger_kernel/chunked_loss/dpo_loss.py,sha256=H9_RRhclckHYM2sd75tgbnf8IxC_PU2JCALbgtPQvwc,4222
+liger_kernel/chunked_loss/dpo_loss.py,sha256=9S67SzKkLyoBmHGx8bkmthSNHlCT2ikBy9CCdb7wGj0,4381
 liger_kernel/chunked_loss/functional.py,sha256=9Gr-YXIuEzEJkBUhDx3G2fuQayckLor7cC7svhmPML4,549
 liger_kernel/chunked_loss/fused_linear_distillation.py,sha256=2BH6DCPjsR2zS6zcwFPcIIZRhLF8SohjGdKsAJ_301o,10222
-liger_kernel/chunked_loss/fused_linear_preference.py,sha256=vlWfaaIECWvCQhY9PM7zRI0vKThIrydMf6P44bXn1EE,15114
+liger_kernel/chunked_loss/fused_linear_preference.py,sha256=AsovMdfsOjgWVxtDhZ_rXqpahMsKTg8YueXnZcHt1XQ,16376
 liger_kernel/chunked_loss/orpo_loss.py,sha256=ZuKGjbkIYzV4UzvupNdq6vyxCp7-BztQkUt8ZnFvKos,3531
 liger_kernel/chunked_loss/simpo_loss.py,sha256=Wa4LOlDG9PbJkOOkKg8hbKvnKgg7OTBz6-qIkwPK1yw,3275
 liger_kernel/ops/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -58,9 +58,9 @@ liger_kernel/transformers/trainer/__init__.py,sha256=c4OQVJmhNOloj0JYSEc0j_cQuBb
 liger_kernel/transformers/trainer/orpo_trainer.py,sha256=jko6oq_XQdBSmXubp05E-_YXOyhtB5Bj75dg5YNwOsE,7517
 liger_kernel/triton/__init__.py,sha256=yfRe0zMb47QnqjecZWG7LnanfCTzeku7SgWRAwNVmzU,101
 liger_kernel/triton/monkey_patch.py,sha256=5BcGKTtdqeYchypBIBopGIWPx1-cFALz7sOKoEsqXJ0,1584
-liger_kernel_nightly-0.5.2.dev20241212030605.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel_nightly-0.5.2.dev20241212030605.dist-info/METADATA,sha256=PeHGuXRXme-T4S249Fh6IWDCNH2-DMWzhyrs2i9MiyE,20260
-liger_kernel_nightly-0.5.2.dev20241212030605.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel_nightly-0.5.2.dev20241212030605.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
-liger_kernel_nightly-0.5.2.dev20241212030605.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel_nightly-0.5.2.dev20241212030605.dist-info/RECORD,,
+liger_kernel_nightly-0.5.2.dev20241212055403.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel_nightly-0.5.2.dev20241212055403.dist-info/METADATA,sha256=kgWj3JislmGGGRKJjny99tjjHfqgdim23iNGAeZhwfk,20260
+liger_kernel_nightly-0.5.2.dev20241212055403.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel_nightly-0.5.2.dev20241212055403.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
+liger_kernel_nightly-0.5.2.dev20241212055403.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel_nightly-0.5.2.dev20241212055403.dist-info/RECORD,,