Fix actions (#31)

* Fix actions

* change to py39

* Fix black

* More fixes

* Fixes

* Fix torch version

* Fix cudnn

.flake8

+[flake8]
+show-source=true
+statistics=true
+max-line-length=80
+
+exclude =
+  .git,
+  .github,
+  setup.py,
+  build,

.github/scripts/install_torch.sh

@@ -14,8 +14,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-torch=$TORCH_VERSION
-cuda=$CUDA_VERSION
+echo "torch version: $torch"
+echo "cuda version: $cuda"
+
 case ${torch} in
   1.5.*)
     case ${cuda} in

.github/workflows/run_tests_cpu.yml

@@ -49,9 +49,9 @@ jobs:
       fail-fast: false
       matrix:
         os: [ubuntu-latest, macos-latest]
-        torch: ["1.13.1"]
-        torchaudio: ["0.13.1"]
-        python-version: ["3.11"]
+        torch: ["1.12.1"]
+        torchaudio: ["0.12.1"]
+        python-version: ["3.9"]
         build_type: ["Release", "Debug"]
 
     steps:
@@ -81,7 +81,7 @@ jobs:
         run: |
           python3 -m pip install -qq --upgrade pip
           python3 -m pip install -qq torch==${{ matrix.torch }}+cpu -f https://download.pytorch.org/whl/torch_stable.html
-          python3 -m pip install -qq torchaudio==${{ matrix.torchaudio }}+cpu -f https://download.pytorch.org/whl/cpu/torch_stable.html
+          python3 -m pip install -qq torchaudio==${{ matrix.torchaudio }} -f https://download.pytorch.org/whl/cpu/torch_stable.html
           python3 -c "import torch; print('torch version:', torch.__version__)"
 
           python3 -m torch.utils.collect_env
@@ -92,7 +92,7 @@ jobs:
         run: |
           python3 -m pip install -qq --upgrade pip
           python3 -m pip install -qq torch==${{ matrix.torch }}
-          python3 -m pip install -qq torch==${{ matrix.torchaudio }}
+          python3 -m pip install -qq torchaudio==${{ matrix.torchaudio }}
           python3 -c "import torch; print('torch version:', torch.__version__)"
 
           python3 -m torch.utils.collect_env

.github/workflows/run_tests_cuda.yml

@@ -47,10 +47,9 @@ jobs:
       fail-fast: false
       matrix:
         os: [ubuntu-latest]
-        cuda: ["11.7"]
-        torch: ["1.13.1"]
-        torchaudio: ["1.13.0"]
-        python-version: ["3.11"]
+        cuda: ["11.6"]
+        torch: ["1.12.1"]
+        python-version: ["3.9"]
         build_type: ["Release", "Debug"]
 
     steps:
@@ -103,7 +102,7 @@ jobs:
         env:
           cuda: ${{ matrix.cuda }}
         run: |
-          ./scripts/github_actions/install_cudnn.sh
+          ./.github/scripts/install_cudnn.sh
 
       - name: Configure CMake
         shell: bash

.github/workflows/style_check.yml

@@ -66,4 +66,4 @@ jobs:
         shell: bash
         working-directory: ${{github.workspace}}
         run: |
-          black --check --diff .
+          black -l 80 --check --diff .

CMakeLists.txt

@@ -11,6 +11,7 @@ endif()
 
 cmake_minimum_required(VERSION 3.8 FATAL_ERROR)
 
+set(CMAKE_DISABLE_FIND_PACKAGE_MKL TRUE)
 set(languages CXX)
 set(_FT_WITH_CUDA ON)
 

fast_rnnt/python/csrc/CMakeLists.txt

@@ -16,11 +16,15 @@ endif()
 pybind11_add_module(_fast_rnnt ${fast_rnnt_srcs})
 target_link_libraries(_fast_rnnt PRIVATE mutual_information_core)
 
-if(UNIX AND NOT APPLE)
+if(APPLE)
+  target_link_libraries(_fast_rnnt
+    PRIVATE
+      ${TORCH_DIR}/lib/libtorch_python.dylib
+  )
+elseif(UNIX)
   target_link_libraries(_fast_rnnt
     PRIVATE
       ${PYTHON_LIBRARY}
       ${TORCH_DIR}/lib/libtorch_python.so
   )
 endif()
-

fast_rnnt/python/fast_rnnt/__init__.py

@@ -13,5 +13,3 @@ from .rnnt_loss import rnnt_loss
 from .rnnt_loss import rnnt_loss_pruned
 from .rnnt_loss import rnnt_loss_simple
 from .rnnt_loss import rnnt_loss_smoothed
-
-

fast_rnnt/python/fast_rnnt/mutual_information.py

@@ -160,7 +160,8 @@ class MutualInformationRecursionFunction(torch.autograd.Function):
         if return_grad or px.requires_grad or py.requires_grad:
             ans_grad = torch.ones(B, device=px.device, dtype=px.dtype)
             (px_grad, py_grad) = _fast_rnnt.mutual_information_backward(
-                px, py, boundary, p, ans_grad)
+                px, py, boundary, p, ans_grad
+            )
             ctx.save_for_backward(px_grad, py_grad)
         assert len(pxy_grads) == 2
         pxy_grads[0] = px_grad
@@ -290,8 +291,9 @@ def mutual_information_recursion(
     px, py = px.contiguous(), py.contiguous()
 
     pxy_grads = [None, None]
-    scores = MutualInformationRecursionFunction.apply(px, py, pxy_grads,
-                                                      boundary, return_grad)
+    scores = MutualInformationRecursionFunction.apply(
+        px, py, pxy_grads, boundary, return_grad
+    )
     px_grad, py_grad = pxy_grads
     return (scores, (px_grad, py_grad)) if return_grad else scores
 
@@ -388,16 +390,18 @@ def joint_mutual_information_recursion(
     p = torch.empty(B, S + 1, T + 1, device=px_tot.device, dtype=px_tot.dtype)
 
     # note, tot_probs is without grad.
-    tot_probs = _fast_rnnt.mutual_information_forward(px_tot, py_tot, boundary, p)
+    tot_probs = _fast_rnnt.mutual_information_forward(
+        px_tot, py_tot, boundary, p
+    )
 
     # this is a kind of "fake gradient" that we use, in effect to compute
     # occupation probabilities.  The backprop will work regardless of the
     # actual derivative w.r.t. the total probs.
     ans_grad = torch.ones(B, device=px_tot.device, dtype=px_tot.dtype)
 
-    (px_grad,
-     py_grad) = _fast_rnnt.mutual_information_backward(px_tot, py_tot, boundary, p,
-                                                ans_grad)
+    (px_grad, py_grad) = _fast_rnnt.mutual_information_backward(
+        px_tot, py_tot, boundary, p, ans_grad
+    )
 
     px_grad = px_grad.reshape(1, B, -1)
     py_grad = py_grad.reshape(1, B, -1)

fast_rnnt/python/fast_rnnt/rnnt_loss.py

@@ -170,7 +170,7 @@ def get_rnnt_logprobs(
         am.transpose(1, 2),  # (B, C, T)
         dim=1,
         index=symbols.unsqueeze(2).expand(B, S, T),
-    ) #  (B, S, T)
+    )  #  (B, S, T)
 
     if rnnt_type == "regular":
         px_am = torch.cat(
@@ -291,7 +291,9 @@ def rnnt_loss_simple(
         T = T0 if rnnt_type != "regular" else T0 - 1
         if boundary is None:
             offset = torch.tensor(
-                (T - 1) / 2, dtype=px.dtype, device=px.device,
+                (T - 1) / 2,
+                dtype=px.dtype,
+                device=px.device,
             ).expand(B, 1, 1)
         else:
             offset = (boundary[:, 3] - 1) / 2
@@ -495,7 +497,9 @@ def rnnt_loss(
         T = T0 if rnnt_type != "regular" else T0 - 1
         if boundary is None:
             offset = torch.tensor(
-                (T - 1) / 2, dtype=px.dtype, device=px.device,
+                (T - 1) / 2,
+                dtype=px.dtype,
+                device=px.device,
             ).expand(B, 1, 1)
         else:
             offset = (boundary[:, 3] - 1) / 2
@@ -770,9 +774,7 @@ def do_rnnt_pruning(
     lm_pruning = torch.gather(
         lm,
         dim=1,
-        index=ranges.reshape(B, T * s_range, 1).expand(
-            (B, T * s_range, C)
-        ),
+        index=ranges.reshape(B, T * s_range, 1).expand((B, T * s_range, C)),
     ).reshape(B, T, s_range, C)
     return am_pruning, lm_pruning
 
@@ -1057,7 +1059,9 @@ def rnnt_loss_pruned(
         T = T0 if rnnt_type != "regular" else T0 - 1
         if boundary is None:
             offset = torch.tensor(
-                (T - 1) / 2, dtype=px.dtype, device=px.device,
+                (T - 1) / 2,
+                dtype=px.dtype,
+                device=px.device,
             ).expand(B, 1, 1)
         else:
             offset = (boundary[:, 3] - 1) / 2
@@ -1248,7 +1252,7 @@ def get_rnnt_logprobs_smoothed(
         am.transpose(1, 2),  # (B, C, T)
         dim=1,
         index=symbols.unsqueeze(2).expand(B, S, T),
-    ) #  (B, S, T)
+    )  #  (B, S, T)
 
     if rnnt_type == "regular":
         px_am = torch.cat(
@@ -1413,7 +1417,9 @@ def rnnt_loss_smoothed(
         T = T0 if rnnt_type != "regular" else T0 - 1
         if boundary is None:
             offset = torch.tensor(
-                (T - 1) / 2, dtype=px.dtype, device=px.device,
+                (T - 1) / 2,
+                dtype=px.dtype,
+                device=px.device,
             ).expand(B, 1, 1)
         else:
             offset = (boundary[:, 3] - 1) / 2

setup.py

@@ -39,7 +39,7 @@ class BuildExtension(build_ext):
             cmake_args = "-DCMAKE_BUILD_TYPE=Release -DFT_BUILD_TESTS=OFF"
 
         if make_args == "" and system_make_args == "":
-            make_args = ' -j '
+            make_args = " -j "
 
         if "PYTHON_EXECUTABLE" not in cmake_args:
             print(f"Setting PYTHON_EXECUTABLE to {sys.executable}")
@@ -89,17 +89,17 @@ def get_package_version():
     latest_version = latest_version.strip('"')
     return latest_version
 
+
 def get_requirements():
     with open("requirements.txt", encoding="utf8") as f:
         requirements = f.read().splitlines()
 
     return requirements
 
+
 package_name = "fast_rnnt"
 
-with open(
-    "fast_rnnt/python/fast_rnnt/__init__.py", "a"
-) as f:
+with open("fast_rnnt/python/fast_rnnt/__init__.py", "a") as f:
     f.write(f"__version__ = '{get_package_version()}'\n")
 
 setuptools.setup(