liger-kernel-nightly 0.6.3.dev20251105012545__tar.gz → 0.6.3.dev20251105224413__tar.gz

{liger_kernel_nightly-0.6.3.dev20251105012545 → liger_kernel_nightly-0.6.3.dev20251105224413}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.3.dev20251105012545
+Version: 0.6.3.dev20251105224413
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.3.dev20251105012545 → liger_kernel_nightly-0.6.3.dev20251105224413}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.6.3.dev20251105012545"
+version = "0.6.3.dev20251105224413"
 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-0.6.3.dev20251105012545 → liger_kernel_nightly-0.6.3.dev20251105224413}/src/liger_kernel/ops/fused_linear_cross_entropy.py

@@ -58,6 +58,9 @@ def fused_linear_cross_entropy_forward(
         else:
             grad_weight = torch.zeros_like(weight, dtype=accum_dtype, device=device) if weight.requires_grad else None
             grad_bias = torch.zeros_like(bias, dtype=accum_dtype, device=device) if bias is not None else None
+    else:
+        grad_weight = None
+        grad_bias = None
 
     loss_1d = torch.zeros(BT, dtype=torch.float32, device=device)
     z_loss_1d = torch.zeros(BT, dtype=_input.dtype, device=_input.device) if return_z_loss else None

{liger_kernel_nightly-0.6.3.dev20251105012545 → liger_kernel_nightly-0.6.3.dev20251105224413}/src/liger_kernel/transformers/monkey_patch.py

@@ -30,8 +30,6 @@ from liger_kernel.transformers.model.mixtral import lce_forward_deprecated as mi
 from liger_kernel.transformers.model.phi3 import lce_forward as phi3_lce_forward
 from liger_kernel.transformers.model.qwen2 import lce_forward as qwen2_lce_forward
 from liger_kernel.transformers.model.qwen2 import lce_forward_deprecated as qwen2_lce_forward_deprecated
-from liger_kernel.transformers.model.qwen3_vl import lce_forward as qwen3_vl_lce_forward
-from liger_kernel.transformers.model.qwen3_vl_moe import lce_forward as qwen3_vl_moe_lce_forward
 from liger_kernel.transformers.model.smollm3 import lce_forward as smollm3_lce_forward
 from liger_kernel.transformers.qwen2vl_mrope import liger_multimodal_rotary_pos_emb
 from liger_kernel.transformers.rms_norm import LigerRMSNorm
@@ -1679,6 +1677,8 @@ def apply_liger_kernel_to_qwen3_vl(
     from transformers.models.qwen3_vl.modeling_qwen3_vl import Qwen3VLModel
     from transformers.models.qwen3_vl.modeling_qwen3_vl import Qwen3VLTextModel
 
+    from liger_kernel.transformers.model.qwen3_vl import lce_forward as qwen3_vl_lce_forward
+
     if rope:
         modeling_qwen3_vl.apply_rotary_pos_emb = liger_rotary_pos_emb_with_cast
         modeling_qwen3_vl.apply_rotary_pos_emb_vision = liger_rotary_pos_emb_with_cast_and_leading_batch
@@ -1752,6 +1752,8 @@ def apply_liger_kernel_to_qwen3_vl_moe(
     from transformers.models.qwen3_vl_moe.modeling_qwen3_vl_moe import Qwen3VLMoeModel
     from transformers.models.qwen3_vl_moe.modeling_qwen3_vl_moe import Qwen3VLMoeTextModel
 
+    from liger_kernel.transformers.model.qwen3_vl_moe import lce_forward as qwen3_vl_moe_lce_forward
+
     if rope:
         modeling_qwen3_vl_moe.apply_rotary_pos_emb = liger_rotary_pos_emb_with_cast
         modeling_qwen3_vl_moe.apply_rotary_pos_emb_vision = liger_rotary_pos_emb_with_cast_and_leading_batch

{liger_kernel_nightly-0.6.3.dev20251105012545 → liger_kernel_nightly-0.6.3.dev20251105224413}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.3.dev20251105012545
+Version: 0.6.3.dev20251105224413
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.3.dev20251105012545 → liger_kernel_nightly-0.6.3.dev20251105224413}/test/convergence/fp32/test_mini_models_multimodal.py

@@ -1484,10 +1484,10 @@ def run_mini_model_multimodal(
             32,
             1e-4,
             torch.float32,
-            1e-8,
-            1e-5,
+            1e-7,
+            5e-4,
+            5e-2,
             5e-3,
-            1e-5,
             5e-3,
             1e-5,
             marks=[

{liger_kernel_nightly-0.6.3.dev20251105012545 → liger_kernel_nightly-0.6.3.dev20251105224413}/test/transformers/test_monkey_patch.py

@@ -1001,8 +1001,8 @@ def test_apply_liger_kernel_to_instance_for_qwen3_vl_moe_text():
 def test_qwen3_vl_rope_hooks_applied():
     # Ensure any monkey patching is cleaned up for subsequent tests
     with patch("transformers.models.qwen3_vl.modeling_qwen3_vl") as modeling_mod:
-        from liger_kernel.transformers.monkey_patch import _liger_qwen3_vl_apply_rotary_pos_emb_vision
-        from liger_kernel.transformers.monkey_patch import liger_rotary_pos_emb
+        from liger_kernel.transformers.monkey_patch import liger_rotary_pos_emb_with_cast
+        from liger_kernel.transformers.monkey_patch import liger_rotary_pos_emb_with_cast_and_leading_batch
 
         # Before applying, make sure attributes exist but are not the liger implementations
         setattr(modeling_mod, "apply_rotary_pos_emb", object())
@@ -1010,16 +1010,16 @@ def test_qwen3_vl_rope_hooks_applied():
 
         _apply_liger_kernel("qwen3_vl")
 
-        assert modeling_mod.apply_rotary_pos_emb is liger_rotary_pos_emb
-        assert modeling_mod.apply_rotary_pos_emb_vision is _liger_qwen3_vl_apply_rotary_pos_emb_vision
+        assert modeling_mod.apply_rotary_pos_emb is liger_rotary_pos_emb_with_cast
+        assert modeling_mod.apply_rotary_pos_emb_vision is liger_rotary_pos_emb_with_cast_and_leading_batch
 
 
 @pytest.mark.skipif(not is_qwen3_vl_moe_available(), reason="qwen3_vl_moe module not available")
 def test_qwen3_vl_moe_rope_hooks_applied():
     # Ensure any monkey patching is cleaned up for subsequent tests
     with patch("transformers.models.qwen3_vl_moe.modeling_qwen3_vl_moe") as modeling_mod:
-        from liger_kernel.transformers.monkey_patch import _liger_qwen3_vl_apply_rotary_pos_emb_vision
-        from liger_kernel.transformers.monkey_patch import liger_rotary_pos_emb
+        from liger_kernel.transformers.monkey_patch import liger_rotary_pos_emb_with_cast
+        from liger_kernel.transformers.monkey_patch import liger_rotary_pos_emb_with_cast_and_leading_batch
 
         # Before applying, make sure attributes exist but are not the liger implementations
         setattr(modeling_mod, "apply_rotary_pos_emb", object())
@@ -1027,8 +1027,8 @@ def test_qwen3_vl_moe_rope_hooks_applied():
 
         _apply_liger_kernel("qwen3_vl_moe")
 
-        assert modeling_mod.apply_rotary_pos_emb is liger_rotary_pos_emb
-        assert modeling_mod.apply_rotary_pos_emb_vision is _liger_qwen3_vl_apply_rotary_pos_emb_vision
+        assert modeling_mod.apply_rotary_pos_emb is liger_rotary_pos_emb_with_cast
+        assert modeling_mod.apply_rotary_pos_emb_vision is liger_rotary_pos_emb_with_cast_and_leading_batch
 
 
 @pytest.mark.skipif(not is_falcon_h1_available(), reason="falcon_h1 module not available")