Merge pull request #146 from BrandonGroth/no_triton

build: Move triton to an optional dependency

Author: chichun-charlie-liu

README.md

@@ -105,6 +105,7 @@ The following optional dependencies are available:
 - `mx`: `microxcaling` package for MX quantization
 - `opt`: Shortcut for `fp8`, `gptq`, and `mx` installs
 - `torchvision`: `torch` package for image recognition training and inference
+- `triton`: `triton` package for matrix multiplication kernels
 - `visualize`: Dependencies for visualizing models and performance data
 - `test`: Dependencies needed for unit testing
 - `dev`: Dependencies needed for development

fms_mo/custom_ext_kernels/triton_kernels.py

@@ -15,8 +15,20 @@
 """This file contains external kernels for FP and INT8 matmul written in triton."""
 
 # Third Party
-from triton.language.extra import libdevice
 import torch
+
+# Local
+from fms_mo.utils.import_utils import available_packages
+
+# Assume any calls to the file are requesting triton
+if not available_packages["triton"]:
+    raise ImportError(
+        "triton python package is not avaialble, please check your installation."
+    )
+
+# Third Party
+# pylint: disable=wrong-import-position
+from triton.language.extra import libdevice
 import triton
 import triton.language as tl
 

pyproject.toml

@@ -26,7 +26,6 @@ dependencies = [
 "accelerate>=0.20.3,!=0.34,<1.7",
 "transformers>=4.45,<4.53",
 "torch>=2.2.0,<2.6", 
-"triton>=3.0,<3.4",
 "tqdm>=4.66.2,<5.0",
 "datasets>=3.0.0,<4.0",
 "ninja>=1.11.1.1,<2.0",
@@ -47,6 +46,7 @@ mx = ["microxcaling>=1.1"]
 opt = ["fms-model-optimizer[fp8, gptq, mx]"]
 torchvision = ["torchvision>=0.17"]
 flash-attn = ["flash-attn>=2.5.3,<3.0"]
+triton = ["triton>=3.0,<3.4"]
 visualize = ["matplotlib", "graphviz", "pygraphviz"]
 dev = ["pre-commit>=3.0.4,<5.0"]
 test = ["pytest", "pillow"]

tests/triton_kernels/test_triton_mm.py

@@ -32,7 +32,7 @@
     )
 
 
-@pytest.mark.parametrize("mkn", [64, 256, 1024, 4096])
+@pytest.mark.parametrize("mkn", [64, 256, 1024])
 @pytest.mark.parametrize(
     "dtype_to_test",
     [
@@ -43,11 +43,12 @@
         torch.float8_e5m2,
     ],
 )
+@pytest.mark.skipif(
+    not torch.cuda.is_available(),
+    reason="test_triton_matmul_fp can only when GPU is available",
+)
 def test_triton_matmul_fp(mkn, dtype_to_test):
     """Parametric tests for triton matmul kernel using variety of tensor sizes and dtypes."""
-    if not torch.cuda.is_available():
-        # only run the test when GPU is available
-        return
 
     torch.manual_seed(23)
     m = n = k = mkn
@@ -81,12 +82,13 @@ def test_triton_matmul_fp(mkn, dtype_to_test):
     assert torch.norm(diff_trun_8b) / torch.norm(torch_output) < 1e-3
 
 
-@pytest.mark.parametrize("mkn", [64, 256, 1024, 4096])
+@pytest.mark.parametrize("mkn", [64, 256, 1024])
+@pytest.mark.skipif(
+    not torch.cuda.is_available(),
+    reason="test_triton_matmul_int8 can only when GPU is available",
+)
 def test_triton_matmul_int8(mkn):
     """Parametric tests for triton imatmul kernel using variety of tensor sizes."""
-    if not torch.cuda.is_available():
-        # only run the test when GPU is available
-        return
 
     torch.manual_seed(23)
     m = n = k = mkn
@@ -123,13 +125,14 @@ def test_triton_matmul_int8(mkn):
 
 @pytest.mark.parametrize("feat_in_out", [(64, 128), (256, 1024), (1024, 4096)])
 @pytest.mark.parametrize("trun_bits", [0, 8, 12, 16])
+@pytest.mark.skipif(
+    not torch.cuda.is_available(),
+    reason="test_linear_fpx_acc can only when GPU is available",
+)
 def test_linear_fpx_acc(feat_in_out, trun_bits):
     """Parametric tests for LinearFPxAcc. This Linear utilizes triton kernel hence can only be run
     on CUDA.
     """
-    if not torch.cuda.is_available():
-        # only run the test when GPU is available
-        return
 
     torch.manual_seed(23)
     feat_in, feat_out = feat_in_out