Prod backward sweep (#14884)

New sweeps:
- prod_bw

.github/workflows/ttnn-run-sweeps.yaml

@@ -263,6 +263,7 @@ on:
           - eltwise.ternary_backward.addcmul_bw
           - eltwise.ternary_backward.addcdiv_bw
           - embedding.embedding
+          - reduction.backward.prod_bw.prod_bw
           - reduction.topk.topk
           - reduction.argmax.argmax
           - reduction.prod

tests/sweep_framework/sweeps/reduction/backward/prod_bw/prod_bw.py

@@ -0,0 +1,112 @@
+# SPDX-FileCopyrightText: © 2024 Tenstorrent Inc.
+
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Optional, Tuple
+from functools import partial
+
+import torch
+import random
+import ttnn
+from tests.sweep_framework.sweep_utils.utils import gen_shapes, sanitize_shape_rm
+from tests.tt_eager.python_api_testing.sweep_tests.generation_funcs import gen_func_with_cast_tt
+
+from tests.ttnn.utils_for_testing import check_with_pcc, start_measuring_time, stop_measuring_time
+from models.utility_functions import torch_random
+
+# Override the default timeout in seconds for hang detection.
+TIMEOUT = 30
+
+random.seed(0)
+
+
+# Parameters provided to the test vector generator are defined here.
+# They are defined as dict-type suites that contain the arguments to the run function as keys, and lists of possible inputs as values.
+# Each suite has a key name (in this case "suite_1" and "suite_2") which will associate the test vectors to this specific suite of inputs.
+# Developers can create their own generator functions and pass them to the parameters as inputs.
+parameters = {
+    "xfail": {
+        "input_shape": gen_shapes([1, 1, 32, 32], [6, 12, 256, 256], [1, 1, 32, 32], 16)
+        + gen_shapes([1, 1, 1, 1], [6, 12, 256, 256], [1, 1, 1, 1], 2),
+        "dim": [0, 1, 2, 3],
+        "grad_dtype": [ttnn.bfloat16, ttnn.bfloat8_b],
+        "input_a_dtype": [ttnn.bfloat16],
+        "input_layout": [ttnn.TILE_LAYOUT],
+        "grad_memory_config": [ttnn.DRAM_MEMORY_CONFIG, ttnn.L1_MEMORY_CONFIG],
+        "input_a_memory_config": [ttnn.DRAM_MEMORY_CONFIG, ttnn.L1_MEMORY_CONFIG],
+        "output_memory_config": [ttnn.DRAM_MEMORY_CONFIG, ttnn.L1_MEMORY_CONFIG],
+    }
+}
+
+
+# Invalidate vector is called during the generation phase where each vector will be passed in.
+# If invalidated, the vector will still be stored but will be skipped.
+# Returns False, None if the vector is valid, and True, str with a reason for invalidation if it is invalid.
+def invalidate_vector(test_vector) -> Tuple[bool, Optional[str]]:
+    if test_vector["input_layout"] == ttnn.ROW_MAJOR_LAYOUT:
+        return True, "Unary operation requires tensor to be in Tile layout when working with non-sharded input tensor"
+    if test_vector["input_layout"] == ttnn.ROW_MAJOR_LAYOUT and (
+        test_vector["grad_dtype"] == ttnn.bfloat8_b or test_vector["input_a_dtype"] == ttnn.bfloat8_b
+    ):
+        return True, "bfloat8_b is only supported on tiled layout"
+    return False, None
+
+
+# This is the run instructions for the test, defined by the developer.
+# The run function must take the above-defined parameters as inputs.
+# The runner will call this run function with each test vector, and the returned results from this function will be stored.
+# If you defined a mesh_device_fixture above, the object you yielded will be passed into this function as 'device'. pOtherwise, it will be the default ttnn device opened by the infra.
+def run(
+    input_shape,
+    dim,
+    grad_dtype,
+    input_a_dtype,
+    input_layout,
+    grad_memory_config,
+    input_a_memory_config,
+    output_memory_config,
+    *,
+    device,
+) -> list:
+    data_seed = random.randint(0, 20000000)
+    torch.manual_seed(data_seed)
+
+    torch_input_tensor_a = gen_func_with_cast_tt(
+        partial(torch_random, low=0, high=100, dtype=torch.float32), input_a_dtype
+    )(input_shape)
+    torch_input_tensor_a.requires_grad = True
+    torch_input_tensor_a.retain_grad()
+
+    max_dim = len(input_shape) - 1
+    dim = random.randint(-max_dim - 1, max_dim)
+
+    intermediate_result = torch.prod(torch_input_tensor_a, dim=dim, keepdim=True)
+    torch_grad_tensor = gen_func_with_cast_tt(
+        partial(torch_random, low=-100, high=100, dtype=torch.float32), grad_dtype
+    )(intermediate_result.shape)
+
+    intermediate_result.backward(gradient=torch_grad_tensor)
+    torch_output_tensor = torch_input_tensor_a.grad
+
+    grad_tensor = ttnn.from_torch(
+        torch_grad_tensor,
+        dtype=grad_dtype,
+        layout=input_layout,
+        device=device,
+        memory_config=grad_memory_config,
+    )
+
+    input_tensor_a = ttnn.from_torch(
+        torch_input_tensor_a,
+        dtype=input_a_dtype,
+        layout=input_layout,
+        device=device,
+        memory_config=input_a_memory_config,
+    )
+
+    start_time = start_measuring_time()
+    output_tensor = ttnn.prod_bw(grad_tensor, input_tensor_a, dim=dim, memory_config=output_memory_config)[0]
+    output_tensor = ttnn.to_torch(output_tensor)
+    e2e_perf = stop_measuring_time(start_time)
+
+    return [check_with_pcc(torch_output_tensor, output_tensor, 0.999), e2e_perf]