Merge branch 'main' into compile-tests

Author: matthewdouglas

.github/workflows/tests.yml

@@ -93,24 +93,32 @@ jobs:
           path: output/${{ matrix.os }}/${{ matrix.arch }}/*
           retention-days: 7
 
-  cpu-tests:
+  test-cpu:
     if: github.repository == 'bitsandbytes-foundation/bitsandbytes'
     needs: build-cpu
     strategy:
       fail-fast: false
       matrix:
         os: [ubuntu-22.04, ubuntu-22.04-arm, windows-2025, macos-15]
-        torch_version: ["2.6.0", "2.7.0"]
+        # Test with the oldest supported torch version and the two newest.
+        torch_version: ["2.2.2", "2.6.0", "2.7.0"]
         include:
           - os: ubuntu-22.04
             arch: x86_64
             runner: banb-aws-general-8-plus-use1-public-80
           - os: ubuntu-22.04-arm
             arch: aarch64
+          - os: ubuntu-22.04-arm
+            arch: aarch64
+            torch_version: "2.5.1"
           - os: windows-2025
             arch: x86_64
           - os: macos-15
             arch: arm64
+        exclude:
+          - os: ubuntu-22.04-arm
+            torch_version: "2.2.2"
+
     runs-on: ${{ matrix.runner || matrix.os }}
     env:
       BNB_TEST_DEVICE: cpu
@@ -135,6 +143,11 @@ jobs:
           pip install -e ".[test]"
           pip install pytest-cov
 
+      # We need to downgrade to numpy<2 for torch<2.3 compatibility.
+      - name: Downgrade NumPy
+        if: startsWith(matrix.torch_version, '2.2.')
+        run: pip install "numpy<2"
+
       - name: Show installed packages
         run: pip list
 
@@ -144,7 +157,7 @@ jobs:
       - name: Run tests
         run: pytest --durations=100
 
-  # cuda-aarch64-tests:
+  # test-cuda-aarch64:
   #   if: github.repository == 'bitsandbytes-foundation/bitsandbytes'
   #   needs: build-cuda
   #   strategy:
@@ -167,7 +180,7 @@ jobs:
 
 
 
-  cuda-tests:
+  test-cuda:
     if: github.repository == 'bitsandbytes-foundation/bitsandbytes'
     needs: build-cuda
     strategy:
@@ -179,7 +192,7 @@ jobs:
         cuda_version: ["11.8.0", "12.6.3", "12.8.1"]
         include:
           - cuda_version: "11.8.0"
-            torch_version: "2.4.1"
+            torch_version: "2.2.2"
             pypi_index: "https://download.pytorch.org/whl/cu118"
           - cuda_version: "12.6.3"
             torch_version: "2.6.0"
@@ -238,6 +251,11 @@ jobs:
           pip install -e ".[test]"
           pip install pytest-cov
 
+        # We need to downgrade to numpy<2 for torch<2.3 compatibility.
+      - name: Downgrade NumPy
+        if: startsWith(matrix.torch_version, '2.2.')
+        run: pip install "numpy<2"
+
       - name: Show installed packages
         run: pip list
 

benchmarking/int8/row_scale_benchmark.py

@@ -303,9 +303,9 @@ def get_config(self, gindex, pindex, group):
         config["eps"] = group["eps"]
         config["weight_decay"] = group["weight_decay"]
         config["lr"] = group["lr"]
-        config["alpha"] = group.get("alpha")
-        config["t_alpha"] = group.get("t_alpha")
-        config["t_beta3"] = group.get("t_beta3")
+        config["alpha"] = group.get("alpha", 0.0)
+        config["t_alpha"] = group.get("t_alpha", 0)
+        config["t_beta3"] = group.get("t_beta3", 0)
         config["optim_bits"] = self.args.optim_bits
         config["min_8bit_size"] = self.args.min_8bit_size
         config["percentile_clipping"] = self.args.percentile_clipping
@@ -530,7 +530,7 @@ def update_step(self, group, p, gindex, pindex):
                 state["state2"],
                 config["betas"][1],
                 config["betas"][2] if len(config["betas"]) >= 3 else 0.0,
-                config["alpha"],
+                config.get("alpha", 0.0),
                 config["weight_decay"],
                 gnorm_scale,
                 state["unorm_vec"] if config["max_unorm"] > 0.0 else None,
@@ -575,7 +575,7 @@ def update_step(self, group, p, gindex, pindex):
                 config["betas"][0],
                 config["betas"][1],
                 config["betas"][2] if len(config["betas"]) >= 3 else 0.0,
-                config["alpha"],
+                config.get("alpha", 0.0),
                 config["eps"],
                 step,
                 config["lr"],