[Operator] Init NLL_LOSS

benchmark/test_reduction_perf.py

@@ -114,6 +114,15 @@ def cross_entropy_loss_input_fn(shape, cur_dtype, device):
         }
 
 
+def nll_loss_input_fn(shape, cur_dtype, device):
+    inp = generate_tensor_input(shape, cur_dtype, device)
+    target = torch.randint(0, shape[-1], (shape[0],), device=device)
+    yield inp, target
+    if Config.bench_level == BenchLevel.COMPREHENSIVE:
+        weight = torch.randn(shape[-1], dtype=cur_dtype, device=device)
+        yield inp, target, {"weight": weight, "ignore_index": 1, "reduction": "none"}
+
+
 def cumsum_input_fn(shape, cur_dtype, device):
     inp = generate_tensor_input(shape, cur_dtype, device)
     yield inp, 1
@@ -162,6 +171,13 @@ def cumsum_input_fn(shape, cur_dtype, device):
                 ),
             ],
         ),
+        pytest.param(
+            "nll_loss",
+            torch.nn.functional.nll_loss,
+            nll_loss_input_fn,
+            FLOAT_DTYPES,
+            marks=pytest.mark.NLLLoss,
+        ),
     ],
 )
 def test_generic_reduction_benchmark(op_name, torch_op, input_fn, dtypes):

src/flag_gems/__init__.py

@@ -163,6 +163,10 @@ def enable(lib=aten_lib, unused=None, registrar=registrar):
             ("log_softmax.int", log_softmax, Autograd.enable),
             ("outer", outer, Autograd.enable),
             ("cross_entropy_loss", cross_entropy_loss, Autograd.enable),
+            ("nll_loss_forward", nll_loss_forward, Autograd.disable),
+            ("nll_loss_backward", nll_loss_backward, Autograd.disable),
+            ("nll_loss2d_forward", nll_loss2d_forward, Autograd.disable),
+            ("nll_loss2d_backward", nll_loss2d_backward, Autograd.disable),
             ("scatter.src", scatter, Autograd.disable),
             ("scatter.reduce", scatter, Autograd.disable),
             ("gather", gather, Autograd.disable),

src/flag_gems/ops/__init__.py

@@ -76,6 +76,12 @@
 from .mv import mv
 from .ne import ne, ne_scalar
 from .neg import neg
+from .nllloss import (
+    nll_loss2d_backward,
+    nll_loss2d_forward,
+    nll_loss_backward,
+    nll_loss_forward,
+)
 from .nonzero import nonzero
 from .normal import normal_float_tensor, normal_tensor_float, normal_tensor_tensor
 from .ones import ones
@@ -289,4 +295,8 @@
     "logical_xor",
     "logical_not",
     "sort",
+    "nll_loss_forward",
+    "nll_loss_backward",
+    "nll_loss2d_forward",
+    "nll_loss2d_backward",
 ]

src/flag_gems/ops/cross_entropy_loss.py

@@ -462,19 +462,13 @@ def celoss_indices_smooth_bwd(
 
 
 @libentry()
-@triton.autotune(
-    configs=runtime.get_tuned_config("cross_entropy_loss_sum_and_scale"),
-    key=[
-        "N",
-    ],
-)
 @triton.jit
 def sum_and_scale(
     inp_ptr,
     out_ptr,
     N,
     scalebyw: tl.constexpr,
-    BLOCK_N: tl.constexpr,
+    BLOCK_N: tl.constexpr = 128,
     scale=1.0,
     mean_num=None,
 ):

src/flag_gems/ops/embedding.py

@@ -148,9 +148,11 @@ def backward(ctx, grad_outputs):
         grad_inputs = torch.zeros(
             (ctx.num_weights, grad_outputs.shape[-1]),
             device=grad_outputs.device,
-            dtype=torch.float32
-            if grad_outputs.dtype is torch.bfloat16
-            else grad_outputs.dtype,
+            dtype=(
+                torch.float32
+                if grad_outputs.dtype is torch.bfloat16
+                else grad_outputs.dtype
+            ),
         )
 
         if ctx.scale_grad_by_freq:
@@ -191,9 +193,11 @@ def backward(ctx, grad_outputs):
                     grad_inputs, indice_freq, ctx.num_weights, ctx.N, BLOCK_SIZE
                 )
         return (
-            grad_inputs.to(torch.bfloat16)
-            if grad_outputs.dtype is torch.bfloat16
-            else grad_inputs,
+            (
+                grad_inputs.to(torch.bfloat16)
+                if grad_outputs.dtype is torch.bfloat16
+                else grad_inputs
+            ),
             None,
             None,
             None,