Sharded sweeps: prelu, softmax, sinh, softplus, relu_max and relu_min (…

…#16050) #11512 ### Ticket [Link to Github Issue](#11512) ### Problem description In sweep tests for sharded ops we add: - prelu - softmax - sinh - relu_max - relu_min - softplus ### What's changed Added new sweep tests ### Pass rates for new sweeps: `sweeps/eltwise/unary/sinh/sinh_sharded.py`: 65 fail, 317 pass (82.98%) `sweeps/eltwise/unary/softmax/softmax_sharded.py`: 382 fail, 0 pass (0%) `sweeps/eltwise/unary/relu_max/relu_max_sharded.py`: 65 fail, 317 pass (82.98%) `sweeps/eltwise/unary/prelu/prelu_sharded.py`: 260 fail, 1268 pass (82.98%) `sweeps/eltwise/unary/relu_min/relu_min_sharded.py`: 65 fail, 317 pass (82.98%) `sweeps/eltwise/unary/softplus/softplus_sharded.py`: 77 fail, 305 pass (79.84%). ### Checklist - [X] [Post commit CI passes](https://github.com/tenstorrent/tt-metal/actions/runs/12159246607) - [X] Sweep tests pass
tenstorrent · Jan 9, 2025 · 3dcd13d · 3dcd13d
1 parent 81f9494
commit 3dcd13d
Show file tree

Hide file tree

Showing 8 changed files with 687 additions and 1 deletion.
diff --git a/.github/workflows/ttnn-run-sweeps.yaml b/.github/workflows/ttnn-run-sweeps.yaml
@@ -125,17 +125,22 @@ on:
           - eltwise.unary.hardsigmoid.hardsigmoid_sharded
           - eltwise.unary.hardshrink.hardshrink
           - eltwise.unary.hardshrink.hardshrink_sharded
-          - eltwise.unary.softmax.softmax
+          - normalization.softmax.softmax
+          - normalization.softmax.softmax_sharded
           - eltwise.unary.identity.identity
           - eltwise.unary.identity.identity_sharded
           - eltwise.unary.neg.neg
           - eltwise.unary.neg.neg_sharded
           - eltwise.unary.sinh.sinh
+          - eltwise.unary.sinh.sinh_sharded
           - eltwise.unary.asinh.asinh
           - eltwise.unary.cosh.cosh
           - eltwise.unary.relu_min.relu_min
+          - eltwise.unary.relu_min.relu_min_sharded
           - eltwise.unary.relu_max.relu_max
+          - eltwise.unary.relu_max.relu_max_sharded
           - eltwise.unary.softplus.softplus
+          - eltwise.unary.softplus.softplus_sharded
           - eltwise.unary.selu.selu
           - eltwise.unary.selu.selu_sharded
           - eltwise.unary.softshrink.softshrink_sharded
@@ -236,6 +241,7 @@ on:
           - eltwise.unary.lez.lez
           - eltwise.unary.nez.nez
           - eltwise.unary.prelu.prelu
+          - eltwise.unary.prelu.prelu_sharded
           - eltwise.unary.hardswish.hardswish_pytorch2
           - eltwise.unary.hardtanh.hardtanh_pytorch2
           - eltwise.unary.leaky_relu.leaky_relu

diff --git a/tests/sweep_framework/sweeps/eltwise/unary/prelu/prelu_sharded.py b/tests/sweep_framework/sweeps/eltwise/unary/prelu/prelu_sharded.py
@@ -0,0 +1,114 @@
+# SPDX-FileCopyrightText: © 2024 Tenstorrent Inc.
+
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Optional, Tuple
+from functools import partial
+
+import json
+import torch
+import random
+import ttnn
+import math
+from tests.sweep_framework.sweep_utils.utils import gen_shapes, sanitize_shape_rm
+from tests.sweep_framework.sweep_utils.sharding_utils import (
+    gen_sharded_spec_unary,
+    parse_sharding_spec,
+    invalidate_vector_sharding,
+)
+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 = 120
+
+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_spec": gen_sharded_spec_unary(16, max_tensor_size_per_core=20 * 1024, layouts=["TILE_LAYOUT"]),
+        "input_a_dtype": [ttnn.bfloat16],
+        "weight": [-0.5, 0, 0.01, 0.5],
+    },
+}
+
+
+# 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]]:
+    input_layout = test_vector["input_spec"]["input_layout"]
+    sharding_invalidated, output_str = invalidate_vector_sharding(test_vector["input_spec"])
+
+    if input_layout == "ROW_MAJOR_LAYOUT":
+        return True, "Inputs to eltwise binary must be tilized"
+    if sharding_invalidated:
+        return sharding_invalidated, output_str
+    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'. Otherwise, it will be the default ttnn device opened by the infra.
+def run(
+    input_spec,
+    input_a_dtype,
+    weight,
+    *,
+    device,
+) -> list:
+    data_seed = random.randint(0, 20000000)
+    torch.manual_seed(data_seed)
+
+    (
+        input_shape,
+        core_grid,
+        sharding_strategy,
+        shard_orientation,
+        tensor_hw_as_shard_shape,
+        input_layout,
+        shard_height_mul_of_32,
+    ) = parse_sharding_spec(input_spec)
+
+    if input_layout == ttnn.ROW_MAJOR_LAYOUT:
+        input_shape = sanitize_shape_rm(input_shape)
+
+    torch_input_tensor_a = gen_func_with_cast_tt(
+        partial(torch_random, low=-100, high=100, dtype=torch.float32), input_a_dtype
+    )(input_shape)
+    golden_function = ttnn.get_golden_function(ttnn.prelu)
+
+    torch_output_tensor = golden_function(torch_input_tensor_a, weight)
+
+    sharded_config = ttnn.create_sharded_memory_config_(
+        shape=input_shape,
+        core_grid=core_grid,
+        strategy=sharding_strategy,
+        orientation=shard_orientation,
+        use_height_and_width_as_shard_shape=tensor_hw_as_shard_shape,
+        tile_layout=shard_height_mul_of_32,
+    )
+
+    input_tensor_a = ttnn.from_torch(
+        torch_input_tensor_a,
+        dtype=input_a_dtype,
+        layout=input_layout,
+        device=device,
+        memory_config=sharded_config,
+    )
+
+    start_time = start_measuring_time()
+    output_tensor = ttnn.prelu(input_tensor_a, weight, memory_config=sharded_config)
+    e2e_perf = stop_measuring_time(start_time)
+    output_tensor = ttnn.to_torch(output_tensor)
+
+    pcc = check_with_pcc(torch_output_tensor, output_tensor, 0.999)
+    return [pcc, e2e_perf]
diff --git a/tests/sweep_framework/sweeps/eltwise/unary/relu_max/relu_max_sharded.py b/tests/sweep_framework/sweeps/eltwise/unary/relu_max/relu_max_sharded.py
@@ -0,0 +1,113 @@
+# SPDX-FileCopyrightText: © 2024 Tenstorrent Inc.
+
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Optional, Tuple
+from functools import partial
+
+import json
+import torch
+import random
+import ttnn
+import math
+from tests.sweep_framework.sweep_utils.utils import gen_shapes, sanitize_shape_rm
+from tests.sweep_framework.sweep_utils.sharding_utils import (
+    gen_sharded_spec_unary,
+    parse_sharding_spec,
+    invalidate_vector_sharding,
+)
+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 = 120
+
+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_spec": gen_sharded_spec_unary(16, max_tensor_size_per_core=20 * 1024, layouts=["TILE_LAYOUT"]),
+        "input_a_dtype": [ttnn.bfloat16],
+    },
+}
+
+
+# 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]]:
+    input_layout = test_vector["input_spec"]["input_layout"]
+    sharding_invalidated, output_str = invalidate_vector_sharding(test_vector["input_spec"])
+
+    if input_layout == "ROW_MAJOR_LAYOUT":
+        return True, "Inputs to eltwise binary must be tilized"
+    if sharding_invalidated:
+        return sharding_invalidated, output_str
+    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'. Otherwise, it will be the default ttnn device opened by the infra.
+def run(
+    input_spec,
+    input_a_dtype,
+    *,
+    device,
+) -> list:
+    data_seed = random.randint(0, 20000000)
+    torch.manual_seed(data_seed)
+
+    (
+        input_shape,
+        core_grid,
+        sharding_strategy,
+        shard_orientation,
+        tensor_hw_as_shard_shape,
+        input_layout,
+        shard_height_mul_of_32,
+    ) = parse_sharding_spec(input_spec)
+
+    if input_layout == ttnn.ROW_MAJOR_LAYOUT:
+        input_shape = sanitize_shape_rm(input_shape)
+
+    torch_input_tensor_a = gen_func_with_cast_tt(
+        partial(torch_random, low=-100, high=100, dtype=torch.float32), input_a_dtype
+    )(input_shape)
+    golden_function = ttnn.get_golden_function(ttnn.relu_max)
+
+    upper_limit = torch.tensor(1, dtype=torch.bfloat16).uniform_(0, 100).item()
+
+    torch_output_tensor = golden_function(torch_input_tensor_a, upper_limit=upper_limit)
+
+    sharded_config = ttnn.create_sharded_memory_config_(
+        shape=input_shape,
+        core_grid=core_grid,
+        strategy=sharding_strategy,
+        orientation=shard_orientation,
+        use_height_and_width_as_shard_shape=tensor_hw_as_shard_shape,
+        tile_layout=shard_height_mul_of_32,
+    )
+
+    input_tensor_a = ttnn.from_torch(
+        torch_input_tensor_a,
+        dtype=input_a_dtype,
+        layout=input_layout,
+        device=device,
+        memory_config=sharded_config,
+    )
+
+    start_time = start_measuring_time()
+    output_tensor = ttnn.relu_max(input_tensor_a, upper_limit, memory_config=sharded_config)
+    e2e_perf = stop_measuring_time(start_time)
+    output_tensor = ttnn.to_torch(output_tensor)
+
+    pcc = check_with_pcc(torch_output_tensor, output_tensor, 0.999)
+    return [pcc, e2e_perf]
diff --git a/tests/sweep_framework/sweeps/eltwise/unary/relu_min/relu_min_sharded.py b/tests/sweep_framework/sweeps/eltwise/unary/relu_min/relu_min_sharded.py
@@ -0,0 +1,114 @@
+# SPDX-FileCopyrightText: © 2024 Tenstorrent Inc.
+
+# SPDX-License-Identifier: Apache-2.0
+
+
+from typing import Optional, Tuple
+from functools import partial
+
+import json
+import torch
+import random
+import ttnn
+import math
+from tests.sweep_framework.sweep_utils.utils import gen_shapes, sanitize_shape_rm
+from tests.sweep_framework.sweep_utils.sharding_utils import (
+    gen_sharded_spec_unary,
+    parse_sharding_spec,
+    invalidate_vector_sharding,
+)
+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 = 120
+
+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_spec": gen_sharded_spec_unary(16, max_tensor_size_per_core=20 * 1024, layouts=["TILE_LAYOUT"]),
+        "input_a_dtype": [ttnn.bfloat16],
+    },
+}
+
+
+# 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]]:
+    input_layout = test_vector["input_spec"]["input_layout"]
+    sharding_invalidated, output_str = invalidate_vector_sharding(test_vector["input_spec"])
+
+    if input_layout == "ROW_MAJOR_LAYOUT":
+        return True, "Inputs to eltwise binary must be tilized"
+    if sharding_invalidated:
+        return sharding_invalidated, output_str
+    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'. Otherwise, it will be the default ttnn device opened by the infra.
+def run(
+    input_spec,
+    input_a_dtype,
+    *,
+    device,
+) -> list:
+    data_seed = random.randint(0, 20000000)
+    torch.manual_seed(data_seed)
+
+    (
+        input_shape,
+        core_grid,
+        sharding_strategy,
+        shard_orientation,
+        tensor_hw_as_shard_shape,
+        input_layout,
+        shard_height_mul_of_32,
+    ) = parse_sharding_spec(input_spec)
+
+    if input_layout == ttnn.ROW_MAJOR_LAYOUT:
+        input_shape = sanitize_shape_rm(input_shape)
+
+    torch_input_tensor_a = gen_func_with_cast_tt(
+        partial(torch_random, low=-100, high=100, dtype=torch.float32), input_a_dtype
+    )(input_shape)
+    golden_function = ttnn.get_golden_function(ttnn.relu_min)
+
+    lower_limit = torch.tensor(1, dtype=torch.bfloat16).uniform_(0, 100).item()
+
+    torch_output_tensor = golden_function(torch_input_tensor_a, lower_limit=lower_limit)
+
+    sharded_config = ttnn.create_sharded_memory_config_(
+        shape=input_shape,
+        core_grid=core_grid,
+        strategy=sharding_strategy,
+        orientation=shard_orientation,
+        use_height_and_width_as_shard_shape=tensor_hw_as_shard_shape,
+        tile_layout=shard_height_mul_of_32,
+    )
+
+    input_tensor_a = ttnn.from_torch(
+        torch_input_tensor_a,
+        dtype=input_a_dtype,
+        layout=input_layout,
+        device=device,
+        memory_config=sharded_config,
+    )
+
+    start_time = start_measuring_time()
+    output_tensor = ttnn.relu_min(input_tensor_a, lower_limit, memory_config=sharded_config)
+    e2e_perf = stop_measuring_time(start_time)
+    output_tensor = ttnn.to_torch(output_tensor)
+
+    pcc = check_with_pcc(torch_output_tensor, output_tensor, 0.999)
+    return [pcc, e2e_perf]