liger-kernel-nightly 0.6.1.dev20250809233744__tar.gz → 0.6.1.dev20250819145841__tar.gz

{liger_kernel_nightly-0.6.1.dev20250809233744 → liger_kernel_nightly-0.6.1.dev20250819145841}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.1.dev20250809233744
+Version: 0.6.1.dev20250819145841
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.1.dev20250809233744 → liger_kernel_nightly-0.6.1.dev20250819145841}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.6.1.dev20250809233744"
+version = "0.6.1.dev20250819145841"
 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-0.6.1.dev20250809233744 → liger_kernel_nightly-0.6.1.dev20250819145841}/src/liger_kernel/transformers/model/loss_utils.py

@@ -13,6 +13,7 @@ def fixed_fused_linear_cross_entropy(
     num_items_in_batch: Optional[int] = None,
     ignore_index: int = -100,
     final_logit_softcapping: Optional[float] = None,
+    accum_dtype: Optional[torch.dtype] = None,
     **kwargs,
 ):
     reduction = "sum" if num_items_in_batch is not None else "mean"
@@ -23,6 +24,7 @@ def fixed_fused_linear_cross_entropy(
         reduction=reduction,
         ignore_index=ignore_index,
         softcap=final_logit_softcapping,
+        accum_dtype=accum_dtype,
     )
     if reduction == "sum":
         loss = loss / num_items_in_batch

{liger_kernel_nightly-0.6.1.dev20250809233744 → liger_kernel_nightly-0.6.1.dev20250819145841}/src/liger_kernel/transformers/model/mllama.py

@@ -190,7 +190,9 @@ def lce_forward(
         output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
     )
     return_dict = return_dict if return_dict is not None else self.config.use_return_dict
-
+    # Filter out accum_dtype from kwargs for model call as MllamaTextModel doesn't accept it in transformers 4.49.0
+    # but preserve it for loss function calls
+    model_kwargs = {k: v for k, v in kwargs.items() if k != "accum_dtype"}
     # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
     outputs = self.model(
         input_ids=input_ids,
@@ -206,7 +208,7 @@ def lce_forward(
         output_hidden_states=output_hidden_states,
         return_dict=return_dict,
         cache_position=cache_position,
-        **kwargs,
+        **model_kwargs,
     )
 
     hidden_states = outputs[0]

{liger_kernel_nightly-0.6.1.dev20250809233744 → liger_kernel_nightly-0.6.1.dev20250819145841}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.1.dev20250809233744
+Version: 0.6.1.dev20250819145841
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.1.dev20250809233744 → liger_kernel_nightly-0.6.1.dev20250819145841}/test/convergence/bf16/test_mini_models.py

@@ -926,7 +926,7 @@ def run_mini_model(
     for i in range(num_steps):
         batch = next(loader_iter).to(model.device)
         optimizer.zero_grad()
-        output = model(**batch)
+        output = model(**batch, accum_dtype=torch.float32)
         output.loss.backward()
         optimizer.step()
         print(f"Step {i}, Loss: {output.loss.item()}")

{liger_kernel_nightly-0.6.1.dev20250809233744 → liger_kernel_nightly-0.6.1.dev20250819145841}/test/convergence/bf16/test_mini_models_multimodal.py

@@ -557,6 +557,7 @@ if QWEN2_5_VL_AVAILABLE:
                 "hidden_size": 128,  # 1280
                 "num_heads": 16,
                 "in_chans": 3,
+                "out_hidden_size": 1024,
             },
             attn_implementation="sdpa",
         ),
@@ -860,7 +861,17 @@ def run_mini_model_multimodal(
     for i in range(num_steps):
         batch = next(loader_iter).to(model.device)
         optimizer.zero_grad()
-        output = model(**batch)
+        supports_accum = getattr(model, "_supports_accum_dtype", None)
+        if supports_accum is None:
+            import inspect
+
+            params = inspect.signature(model.forward).parameters
+            supports_accum = ("accum_dtype" in params) or any(
+                p.kind == inspect.Parameter.VAR_KEYWORD for p in params.values()
+            )
+            setattr(model, "_supports_accum_dtype", supports_accum)
+
+        output = model(**batch, accum_dtype=torch.float32) if supports_accum else model(**batch)
         output.loss.backward()
         optimizer.step()
 

{liger_kernel_nightly-0.6.1.dev20250809233744 → liger_kernel_nightly-0.6.1.dev20250819145841}/test/convergence/fp32/test_mini_models_multimodal.py

@@ -555,6 +555,7 @@ if QWEN2_5_VL_AVAILABLE:
                 "hidden_size": 128,  # 1280
                 "num_heads": 16,
                 "in_chans": 3,
+                "out_hidden_size": 1024,
             },
             attn_implementation="sdpa",
         ),