liger-kernel-nightly 0.5.2.dev20250109023714__tar.gz → 0.5.2.dev20250110102924__tar.gz

{liger_kernel_nightly-0.5.2.dev20250109023714 → liger_kernel_nightly-0.5.2.dev20250110102924}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.2.dev20250109023714
+Version: 0.5.2.dev20250110102924
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.5.2.dev20250109023714 → liger_kernel_nightly-0.5.2.dev20250110102924}/benchmark/scripts/benchmark_cpo_loss.py

@@ -11,7 +11,6 @@ from utils import _test_memory
 from utils import parse_benchmark_script_args
 from utils import run_benchmarks
 
-from liger_kernel.chunked_loss.cpo_loss import LigerFusedLinearCPOFunction
 from liger_kernel.utils import infer_device
 
 device = infer_device()
@@ -27,7 +26,8 @@ sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "../.
 def bench_memory_fused_linear_cpo_loss(
     input: SingleBenchmarkRunInput,
 ) -> SingleBenchmarkRunOutput:
-    from test.chunked_loss.test_cpo_loss import LigerLMHeadCPO, TorchLMHeadCPO
+    from test.chunked_loss.test_cpo_loss import LigerLMHeadCPO
+    from test.chunked_loss.test_cpo_loss import TorchLMHeadCPO
 
     B = input.x
     T = input.extra_benchmark_config["T"]
@@ -36,12 +36,8 @@ def bench_memory_fused_linear_cpo_loss(
     dtype = input.extra_benchmark_config["dtype"]
     provider = input.kernel_provider
 
-    torch_lm_head_cpo = lambda x, target: TorchLMHeadCPO(H=H, V=V, dtype=dtype).to(
-        device
-    )(x, target)[0]
-    liger_lm_head_cpo = lambda x, target: LigerLMHeadCPO(H=H, V=V, dtype=dtype).to(
-        device
-    )(x, target)[0]
+    torch_lm_head_cpo = lambda x, target: TorchLMHeadCPO(H=H, V=V, dtype=dtype).to(device)(x, target)[0]
+    liger_lm_head_cpo = lambda x, target: LigerLMHeadCPO(H=H, V=V, dtype=dtype).to(device)(x, target)[0]
 
     _input = torch.randn(B, T, H, requires_grad=True, dtype=dtype, device=device)
     target = torch.randint(V, (B, T), dtype=torch.long, device=device)
@@ -72,7 +68,8 @@ def bench_memory_fused_linear_cpo_loss(
 def bench_speed_fused_linear_cpo_loss(
     input: SingleBenchmarkRunInput,
 ) -> SingleBenchmarkRunOutput:
-    from test.chunked_loss.test_cpo_loss import LigerLMHeadCPO, TorchLMHeadCPO
+    from test.chunked_loss.test_cpo_loss import LigerLMHeadCPO
+    from test.chunked_loss.test_cpo_loss import TorchLMHeadCPO
 
     B = input.x
     T = input.extra_benchmark_config["T"]
@@ -82,12 +79,8 @@ def bench_speed_fused_linear_cpo_loss(
     provider = input.kernel_provider
     mode = input.kernel_operation_mode
 
-    torch_lm_head_cpo = lambda x, target: TorchLMHeadCPO(H=H, V=V, dtype=dtype).to(
-        device
-    )(x, target)[0]
-    liger_lm_head_cpo = lambda x, target: LigerLMHeadCPO(H=H, V=V, dtype=dtype).to(
-        device
-    )(x, target)[0]
+    torch_lm_head_cpo = lambda x, target: TorchLMHeadCPO(H=H, V=V, dtype=dtype).to(device)(x, target)[0]
+    liger_lm_head_cpo = lambda x, target: LigerLMHeadCPO(H=H, V=V, dtype=dtype).to(device)(x, target)[0]
 
     _input = torch.randn(B, T, H, requires_grad=True, dtype=dtype, device=device)
     target = torch.randint(V, (B, T), dtype=torch.long, device=device)

{liger_kernel_nightly-0.5.2.dev20250109023714 → liger_kernel_nightly-0.5.2.dev20250110102924}/benchmark/scripts/benchmark_dpo_loss.py

@@ -4,7 +4,6 @@ import sys
 import torch
 import triton
 
-from test.chunked_loss.test_dpo_loss import HF_DPO_Loss
 from utils import QUANTILES
 from utils import SingleBenchmarkRunInput
 from utils import SingleBenchmarkRunOutput
@@ -12,7 +11,6 @@ from utils import _test_memory
 from utils import parse_benchmark_script_args
 from utils import run_benchmarks
 
-from liger_kernel.chunked_loss.dpo_loss import LigerFusedLinearDPOFunction
 from liger_kernel.utils import infer_device
 
 device = infer_device()
@@ -21,7 +19,8 @@ sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "../.
 
 
 def bench_memory_dpo_loss(input: SingleBenchmarkRunInput) -> SingleBenchmarkRunOutput:
-    from test.chunked_loss.test_dpo_loss import LigerLMHeadDPO, TorchLMHeadDPO
+    from test.chunked_loss.test_dpo_loss import LigerLMHeadDPO
+    from test.chunked_loss.test_dpo_loss import TorchLMHeadDPO
 
     B = input.x
     T = input.extra_benchmark_config["T"]
@@ -70,7 +69,8 @@ def bench_memory_dpo_loss(input: SingleBenchmarkRunInput) -> SingleBenchmarkRunO
 
 
 def bench_speed_dpo_loss(input: SingleBenchmarkRunInput) -> SingleBenchmarkRunOutput:
-    from test.chunked_loss.test_dpo_loss import LigerLMHeadDPO, TorchLMHeadDPO
+    from test.chunked_loss.test_dpo_loss import LigerLMHeadDPO
+    from test.chunked_loss.test_dpo_loss import TorchLMHeadDPO
 
     B = input.x
     T = input.extra_benchmark_config["T"]

{liger_kernel_nightly-0.5.2.dev20250109023714 → liger_kernel_nightly-0.5.2.dev20250110102924}/benchmark/scripts/benchmark_orpo_loss.py

@@ -11,7 +11,6 @@ from utils import _test_memory
 from utils import parse_benchmark_script_args
 from utils import run_benchmarks
 
-from liger_kernel.chunked_loss.orpo_loss import LigerFusedLinearORPOFunction
 from liger_kernel.utils import infer_device
 
 device = infer_device()
@@ -27,7 +26,8 @@ sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "../.
 def bench_memory_fused_linear_orpo_loss(
     input: SingleBenchmarkRunInput,
 ) -> SingleBenchmarkRunOutput:
-    from test.chunked_loss.test_orpo_loss import LigerLMHeadORPO, TorchLMHeadORPO
+    from test.chunked_loss.test_orpo_loss import LigerLMHeadORPO
+    from test.chunked_loss.test_orpo_loss import TorchLMHeadORPO
 
     B = input.x
     T = input.extra_benchmark_config["T"]
@@ -36,12 +36,8 @@ def bench_memory_fused_linear_orpo_loss(
     dtype = input.extra_benchmark_config["dtype"]
     provider = input.kernel_provider
 
-    torch_lm_head_orpo = lambda x, target: TorchLMHeadORPO(H=H, V=V, dtype=dtype).to(
-        device
-    )(x, target)[0]
-    liger_lm_head_orpo = lambda x, target: LigerLMHeadORPO(H=H, V=V, dtype=dtype).to(
-        device
-    )(x, target)[0]
+    torch_lm_head_orpo = lambda x, target: TorchLMHeadORPO(H=H, V=V, dtype=dtype).to(device)(x, target)[0]
+    liger_lm_head_orpo = lambda x, target: LigerLMHeadORPO(H=H, V=V, dtype=dtype).to(device)(x, target)[0]
 
     _input = torch.randn(B, T, H, requires_grad=True, dtype=dtype, device=device)
     target = torch.randint(V, (B, T), dtype=torch.long, device=device)
@@ -73,7 +69,8 @@ def bench_memory_fused_linear_orpo_loss(
 def bench_speed_fused_linear_orpo_loss(
     input: SingleBenchmarkRunInput,
 ) -> SingleBenchmarkRunOutput:
-    from test.chunked_loss.test_orpo_loss import LigerLMHeadORPO, TorchLMHeadORPO
+    from test.chunked_loss.test_orpo_loss import LigerLMHeadORPO
+    from test.chunked_loss.test_orpo_loss import TorchLMHeadORPO
 
     B = input.x
     T = input.extra_benchmark_config["T"]
@@ -83,12 +80,8 @@ def bench_speed_fused_linear_orpo_loss(
     provider = input.kernel_provider
     mode = input.kernel_operation_mode
 
-    torch_lm_head_orpo = lambda x, target: TorchLMHeadORPO(H=H, V=V, dtype=dtype).to(
-        device
-    )(x, target)[0]
-    liger_lm_head_orpo = lambda x, target: LigerLMHeadORPO(H=H, V=V, dtype=dtype).to(
-        device
-    )(x, target)[0]
+    torch_lm_head_orpo = lambda x, target: TorchLMHeadORPO(H=H, V=V, dtype=dtype).to(device)(x, target)[0]
+    liger_lm_head_orpo = lambda x, target: LigerLMHeadORPO(H=H, V=V, dtype=dtype).to(device)(x, target)[0]
 
     _input = torch.randn(B, T, H, requires_grad=True, dtype=dtype, device=device)
     target = torch.randint(V, (B, T), dtype=torch.long, device=device)

{liger_kernel_nightly-0.5.2.dev20250109023714 → liger_kernel_nightly-0.5.2.dev20250110102924}/benchmark/scripts/benchmark_simpo_loss.py

@@ -11,7 +11,6 @@ from utils import _test_memory
 from utils import parse_benchmark_script_args
 from utils import run_benchmarks
 
-from liger_kernel.chunked_loss.simpo_loss import LigerFusedLinearSimPOFunction
 from liger_kernel.utils import infer_device
 
 device = infer_device()
@@ -27,7 +26,8 @@ sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "../.
 def bench_memory_fused_linear_simpo_loss(
     input: SingleBenchmarkRunInput,
 ) -> SingleBenchmarkRunOutput:
-    from test.chunked_loss.test_simpo_loss import LigerLMHeadSimPO, TorchLMHeadCPO
+    from test.chunked_loss.test_simpo_loss import LigerLMHeadSimPO
+    from test.chunked_loss.test_simpo_loss import TorchLMHeadCPO
 
     B = input.x
     T = input.extra_benchmark_config["T"]
@@ -36,12 +36,8 @@ def bench_memory_fused_linear_simpo_loss(
     dtype = input.extra_benchmark_config["dtype"]
     provider = input.kernel_provider
 
-    torch_lm_head_simpo = lambda x, target: TorchLMHeadCPO(H=H, V=V, dtype=dtype).to(
-        device
-    )(x, target)[0]
-    liger_lm_head_simpo = lambda x, target: LigerLMHeadSimPO(H=H, V=V, dtype=dtype).to(
-        device
-    )(x, target)[0]
+    torch_lm_head_simpo = lambda x, target: TorchLMHeadCPO(H=H, V=V, dtype=dtype).to(device)(x, target)[0]
+    liger_lm_head_simpo = lambda x, target: LigerLMHeadSimPO(H=H, V=V, dtype=dtype).to(device)(x, target)[0]
 
     _input = torch.randn(B, T, H, requires_grad=True, dtype=dtype, device=device)
     target = torch.randint(V, (B, T), dtype=torch.long, device=device)
@@ -72,7 +68,8 @@ def bench_memory_fused_linear_simpo_loss(
 def bench_speed_fused_linear_simpo_loss(
     input: SingleBenchmarkRunInput,
 ) -> SingleBenchmarkRunOutput:
-    from test.chunked_loss.test_simpo_loss import LigerLMHeadSimPO, TorchLMHeadCPO
+    from test.chunked_loss.test_simpo_loss import LigerLMHeadSimPO
+    from test.chunked_loss.test_simpo_loss import TorchLMHeadCPO
 
     B = input.x
     T = input.extra_benchmark_config["T"]
@@ -82,12 +79,8 @@ def bench_speed_fused_linear_simpo_loss(
     provider = input.kernel_provider
     mode = input.kernel_operation_mode
 
-    torch_lm_head_simpo = lambda x, target: TorchLMHeadCPO(H=H, V=V, dtype=dtype).to(
-        device
-    )(x, target)[0]
-    liger_lm_head_simpo = lambda x, target: LigerLMHeadSimPO(H=H, V=V, dtype=dtype).to(
-        device
-    )(x, target)[0]
+    torch_lm_head_simpo = lambda x, target: TorchLMHeadCPO(H=H, V=V, dtype=dtype).to(device)(x, target)[0]
+    liger_lm_head_simpo = lambda x, target: LigerLMHeadSimPO(H=H, V=V, dtype=dtype).to(device)(x, target)[0]
 
     _input = torch.randn(B, T, H, requires_grad=True, dtype=dtype, device=device)
     target = torch.randint(V, (B, T), dtype=torch.long, device=device)

{liger_kernel_nightly-0.5.2.dev20250109023714 → liger_kernel_nightly-0.5.2.dev20250110102924}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.5.2.dev20250109023714"
+version = "0.5.2.dev20250110102924"
 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-0.5.2.dev20250109023714 → liger_kernel_nightly-0.5.2.dev20250110102924}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.2.dev20250109023714
+Version: 0.5.2.dev20250110102924
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation