liger-kernel-nightly 0.6.2.dev20251020204513__tar.gz → 0.6.2.dev20251024142419__tar.gz

{liger_kernel_nightly-0.6.2.dev20251020204513 → liger_kernel_nightly-0.6.2.dev20251024142419}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.2.dev20251020204513
+Version: 0.6.2.dev20251024142419
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.2.dev20251020204513 → liger_kernel_nightly-0.6.2.dev20251024142419}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.6.2.dev20251020204513"
+version = "0.6.2.dev20251024142419"
 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-0.6.2.dev20251020204513 → liger_kernel_nightly-0.6.2.dev20251024142419}/src/liger_kernel/transformers/monkey_patch.py

@@ -1971,7 +1971,8 @@ def apply_liger_kernel_to_glm4v_moe(
     if rope:
         raise NotImplementedError("liger_rotary_pos_emb is not available for Glm4 models.")
     if rms_norm:
-        modeling_glm4v_moe.Glm4vRMSNorm = LigerRMSNormForGlm4
+        modeling_glm4v_moe.Glm4vMoeRMSNorm = LigerRMSNormForGlm4
+        modeling_glm4v_moe.Glm4vMoeTextRMSNorm = LigerRMSNormForGlm4
     if cross_entropy:
         from transformers.loss.loss_utils import nn
 

{liger_kernel_nightly-0.6.2.dev20251020204513 → liger_kernel_nightly-0.6.2.dev20251024142419}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.2.dev20251020204513
+Version: 0.6.2.dev20251024142419
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

liger_kernel_nightly-0.6.2.dev20251024142419/test/conftest.py

@@ -0,0 +1,11 @@
+import pytest
+import torch
+
+
+@pytest.fixture(autouse=True)
+def clear_gpu_cache():
+    yield
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+    elif torch.xpu.is_available():
+        torch.xpu.empty_cache()

{liger_kernel_nightly-0.6.2.dev20251020204513 → liger_kernel_nightly-0.6.2.dev20251024142419}/test/convergence/bf16/test_mini_models.py

@@ -1,3 +1,7 @@
+import os
+
+os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"  # Ensure deterministic behavior with CuBLAS
+
 import pytest
 import torch
 
@@ -47,6 +51,7 @@ from test.utils import MiniModelConfig
 from test.utils import assert_verbose_allclose
 from test.utils import get_logprobs
 from test.utils import get_topk
+from test.utils import require_deterministic
 from test.utils import revert_liger_kernel_to_falcon_h1
 from test.utils import revert_liger_kernel_to_gemma
 from test.utils import revert_liger_kernel_to_gemma2
@@ -1165,6 +1170,7 @@ def create_model(model_name="mini_llama4"):
     return model_class(model_config)
 
 
+@require_deterministic
 def run_mini_model(
     model_name="mini_llama4",
     num_steps=100,
@@ -1522,7 +1528,6 @@ def run_mini_model(
                     not GLM4V_AVAILABLE,
                     reason="Glm4v not available in this version of transformers",
                 ),
-                pytest.mark.skipif(device == "xpu", reason="skip for XPU"),
             ],
         ),
         pytest.param(
@@ -1542,7 +1547,6 @@ def run_mini_model(
                     not GLM4V_MOE_AVAILABLE,
                     reason="Glm4v_moe not available in this version of transformers",
                 ),
-                pytest.mark.skipif(device == "xpu", reason="skip for XPU"),
             ],
         ),
         pytest.param(

{liger_kernel_nightly-0.6.2.dev20251020204513 → liger_kernel_nightly-0.6.2.dev20251024142419}/test/convergence/bf16/test_mini_models_multimodal.py

@@ -1,6 +1,7 @@
 import functools
 import os
 
+os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"  # Ensure deterministic behavior with CuBLAS
 import pytest
 import torch
 
@@ -29,6 +30,7 @@ from test.utils import load_image_processing_config
 from test.utils import load_processor_config
 from test.utils import load_tokenizer_config
 from test.utils import multimodal_collate_fn
+from test.utils import require_deterministic
 from test.utils import revert_liger_kernel_to_gemma3
 from test.utils import revert_liger_kernel_to_internvl
 from test.utils import revert_liger_kernel_to_llama4
@@ -881,6 +883,7 @@ def create_model(model_name):
     return model_class(model_config)
 
 
+@require_deterministic
 def run_mini_model_multimodal(
     model_name="mini_qwen2_vl",
     num_steps=100,

{liger_kernel_nightly-0.6.2.dev20251020204513 → liger_kernel_nightly-0.6.2.dev20251024142419}/test/convergence/bf16/test_mini_models_with_logits.py

@@ -1,3 +1,7 @@
+import os
+
+os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"  # Ensure deterministic behavior with CuBLAS
+
 import pytest
 import torch
 
@@ -47,6 +51,7 @@ from test.utils import MiniModelConfig
 from test.utils import assert_verbose_allclose
 from test.utils import get_logprobs
 from test.utils import get_topk
+from test.utils import require_deterministic
 from test.utils import revert_liger_kernel_to_falcon_h1
 from test.utils import revert_liger_kernel_to_gemma
 from test.utils import revert_liger_kernel_to_gemma2
@@ -1164,6 +1169,7 @@ def create_model(model_name="mini_llama3"):
     return model_class(model_config)
 
 
+@require_deterministic
 def run_mini_model(
     model_name="mini_llama3",
     num_steps=100,

{liger_kernel_nightly-0.6.2.dev20251020204513 → liger_kernel_nightly-0.6.2.dev20251024142419}/test/convergence/fp32/test_mini_models.py

@@ -1,3 +1,7 @@
+import os
+
+os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"  # Ensure deterministic behavior with CuBLAS
+
 import pytest
 import torch
 
@@ -47,6 +51,7 @@ from test.utils import MiniModelConfig
 from test.utils import assert_verbose_allclose
 from test.utils import get_logprobs
 from test.utils import get_topk
+from test.utils import require_deterministic
 from test.utils import revert_liger_kernel_to_falcon_h1
 from test.utils import revert_liger_kernel_to_gemma
 from test.utils import revert_liger_kernel_to_gemma2
@@ -1160,6 +1165,7 @@ def create_model(model_name="mini_llama3"):
     return model_class(model_config)
 
 
+@require_deterministic
 def run_mini_model(
     model_name="mini_llama3",
     num_steps=100,
@@ -1436,7 +1442,7 @@ def run_mini_model(
             1e-4,
             torch.float32,
             1e-8,
-            1e-5,
+            1e-3,
             5e-3,
             1e-5,
             5e-3,
@@ -1446,7 +1452,6 @@ def run_mini_model(
                     not GLM4V_MOE_AVAILABLE,
                     reason="Glm4v_moe not available in this version of transformers",
                 ),
-                pytest.mark.skipif(device == "xpu", reason="skip for XPU"),
             ],
         ),
         ("mini_phi3", 32, 1e-4, torch.float32, 1e-8, 1e-5, 5e-3, 1e-5, 5e-3, 1e-5),

{liger_kernel_nightly-0.6.2.dev20251020204513 → liger_kernel_nightly-0.6.2.dev20251024142419}/test/convergence/fp32/test_mini_models_multimodal.py

@@ -1,6 +1,8 @@
 import functools
 import os
 
+os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"  # Ensure deterministic behavior with CuBLAS
+
 import pytest
 import torch
 
@@ -29,6 +31,7 @@ from test.utils import load_image_processing_config
 from test.utils import load_processor_config
 from test.utils import load_tokenizer_config
 from test.utils import multimodal_collate_fn
+from test.utils import require_deterministic
 from test.utils import revert_liger_kernel_to_gemma3
 from test.utils import revert_liger_kernel_to_internvl
 from test.utils import revert_liger_kernel_to_llama4
@@ -878,6 +881,7 @@ def create_model(model_name):
     return model_class(model_config)
 
 
+@require_deterministic
 def run_mini_model_multimodal(
     model_name="mini_qwen2_vl",
     num_steps=100,

{liger_kernel_nightly-0.6.2.dev20251020204513 → liger_kernel_nightly-0.6.2.dev20251024142419}/test/convergence/fp32/test_mini_models_with_logits.py

@@ -1,3 +1,7 @@
+import os
+
+os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"  # Ensure deterministic behavior with CuBLAS
+
 import pytest
 import torch
 
@@ -47,6 +51,7 @@ from test.utils import MiniModelConfig
 from test.utils import assert_verbose_allclose
 from test.utils import get_logprobs
 from test.utils import get_topk
+from test.utils import require_deterministic
 from test.utils import revert_liger_kernel_to_falcon_h1
 from test.utils import revert_liger_kernel_to_gemma
 from test.utils import revert_liger_kernel_to_gemma2
@@ -1161,6 +1166,7 @@ def create_model(model_name="mini_llama3"):
     return model_class(model_config)
 
 
+@require_deterministic
 def run_mini_model(
     model_name="mini_llama3",
     num_steps=100,

{liger_kernel_nightly-0.6.2.dev20251020204513 → liger_kernel_nightly-0.6.2.dev20251024142419}/test/utils.py

@@ -5,6 +5,7 @@ import random
 
 from abc import abstractmethod
 from dataclasses import dataclass
+from functools import wraps
 from typing import Any
 from typing import Dict
 from typing import List
@@ -59,6 +60,19 @@ def set_seed(seed=42):
     os.environ["PYTHONHASHSEED"] = str(seed)
 
 
+def require_deterministic(test_case):
+    @wraps(test_case)
+    def wrapper(*args, **kwargs):
+        original_state = torch.are_deterministic_algorithms_enabled()
+        try:
+            torch.use_deterministic_algorithms(True)
+            return test_case(*args, **kwargs)
+        finally:
+            torch.use_deterministic_algorithms(original_state)
+
+    return wrapper
+
+
 @torch.no_grad
 def get_logprobs(tensor):
     return torch.nn.functional.log_softmax(tensor, dim=-1, dtype=torch.float32)

liger_kernel_nightly-0.6.2.dev20251020204513/test/conftest.py

@@ -1,8 +0,0 @@
-import pytest
-import torch
-
-
-@pytest.fixture(autouse=True)
-def clear_cuda_cache():
-    yield
-    torch.cuda.empty_cache()