Skip to content

⚡️ Speed up function _create_device_sync_precompute_statements by 14% in PR #1015 (gpu-sync-instrumentation) #1040

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
codeflash-ai wants to merge 1 commit into from
+35 −48
Closed

⚡️ Speed up function _create_device_sync_precompute_statements by 14% in PR #1015 (gpu-sync-instrumentation) #1040

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
83 changes: 35 additions & 48 deletions codeflash/code_utils/instrument_existing_tests.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -768,35 +768,33 @@ def _create_device_sync_precompute_statements(used_frameworks: dict[str, str] |

precompute_statements: list[ast.stmt] = []

# PyTorch: pre-compute whether to sync CUDA or MPS
load_ctx = ast.Load()
store_ctx = ast.Store()

# PyTorch: pre-compute whether to sync CUDA or MPS
if "torch" in used_frameworks:
torch_alias = used_frameworks["torch"]
# _codeflash_should_sync_cuda = torch.cuda.is_available() and torch.cuda.is_initialized()
torch_name = ast.Name(id=torch_alias, ctx=load_ctx)

# Helper to create torch.cuda attribute access
torch_cuda = ast.Attribute(value=torch_name, attr="cuda", ctx=load_ctx)

# _codeflash_should_sync_cuda = torch.cuda.is_available() and torch.cuda.is_initialized()
precompute_statements.append(
ast.Assign(
targets=[ast.Name(id="_codeflash_should_sync_cuda", ctx=ast.Store())],
targets=[ast.Name(id="_codeflash_should_sync_cuda", ctx=store_ctx)],
value=ast.BoolOp(
op=ast.And(),
values=[
ast.Call(
func=ast.Attribute(
value=ast.Attribute(
value=ast.Name(id=torch_alias, ctx=ast.Load()), attr="cuda", ctx=ast.Load()
),
attr="is_available",
ctx=ast.Load(),
),
func=ast.Attribute(value=torch_cuda, attr="is_available", ctx=load_ctx),
args=[],
keywords=[],
),
ast.Call(
func=ast.Attribute(
value=ast.Attribute(
value=ast.Name(id=torch_alias, ctx=ast.Load()), attr="cuda", ctx=ast.Load()
),
attr="is_initialized",
ctx=ast.Load(),
),
func=ast.Attribute(value=torch_cuda, attr="is_initialized", ctx=load_ctx),
args=[],
keywords=[],
),
Expand All @@ -808,46 +806,37 @@ def _create_device_sync_precompute_statements(used_frameworks: dict[str, str] |
# _codeflash_should_sync_mps = (not _codeflash_should_sync_cuda and
# hasattr(torch.backends, 'mps') and torch.backends.mps.is_available() and
# hasattr(torch.mps, 'synchronize'))

# _codeflash_should_sync_mps = (not _codeflash_should_sync_cuda and
# hasattr(torch.backends, 'mps') and torch.backends.mps.is_available() and
# hasattr(torch.mps, 'synchronize'))
torch_backends = ast.Attribute(value=torch_name, attr="backends", ctx=load_ctx)
torch_mps_attr = ast.Attribute(value=torch_name, attr="mps", ctx=load_ctx)

precompute_statements.append(
ast.Assign(
targets=[ast.Name(id="_codeflash_should_sync_mps", ctx=ast.Store())],
targets=[ast.Name(id="_codeflash_should_sync_mps", ctx=store_ctx)],
value=ast.BoolOp(
op=ast.And(),
values=[
ast.UnaryOp(op=ast.Not(), operand=ast.Name(id="_codeflash_should_sync_cuda", ctx=ast.Load())),
ast.UnaryOp(op=ast.Not(), operand=ast.Name(id="_codeflash_should_sync_cuda", ctx=load_ctx)),
ast.Call(
func=ast.Name(id="hasattr", ctx=ast.Load()),
args=[
ast.Attribute(
value=ast.Name(id=torch_alias, ctx=ast.Load()), attr="backends", ctx=ast.Load()
),
ast.Constant(value="mps"),
],
func=ast.Name(id="hasattr", ctx=load_ctx),
args=[torch_backends, ast.Constant(value="mps")],
keywords=[],
),
ast.Call(
func=ast.Attribute(
value=ast.Attribute(
value=ast.Attribute(
value=ast.Name(id=torch_alias, ctx=ast.Load()), attr="backends", ctx=ast.Load()
),
attr="mps",
ctx=ast.Load(),
),
value=ast.Attribute(value=torch_backends, attr="mps", ctx=load_ctx),
attr="is_available",
ctx=ast.Load(),
ctx=load_ctx,
),
args=[],
keywords=[],
),
ast.Call(
func=ast.Name(id="hasattr", ctx=ast.Load()),
args=[
ast.Attribute(
value=ast.Name(id=torch_alias, ctx=ast.Load()), attr="mps", ctx=ast.Load()
),
ast.Constant(value="synchronize"),
],
func=ast.Name(id="hasattr", ctx=load_ctx),
args=[torch_mps_attr, ast.Constant(value="synchronize")],
keywords=[],
),
],
Expand All @@ -862,10 +851,10 @@ def _create_device_sync_precompute_statements(used_frameworks: dict[str, str] |
# _codeflash_should_sync_jax = hasattr(jax, 'block_until_ready')
precompute_statements.append(
ast.Assign(
targets=[ast.Name(id="_codeflash_should_sync_jax", ctx=ast.Store())],
targets=[ast.Name(id="_codeflash_should_sync_jax", ctx=store_ctx)],
value=ast.Call(
func=ast.Name(id="hasattr", ctx=ast.Load()),
args=[ast.Name(id=jax_alias, ctx=ast.Load()), ast.Constant(value="block_until_ready")],
func=ast.Name(id="hasattr", ctx=load_ctx),
args=[ast.Name(id=jax_alias, ctx=load_ctx), ast.Constant(value="block_until_ready")],
keywords=[],
),
lineno=1,
Expand All @@ -878,16 +867,14 @@ def _create_device_sync_precompute_statements(used_frameworks: dict[str, str] |
# _codeflash_should_sync_tf = hasattr(tf.test.experimental, 'sync_devices')
precompute_statements.append(
ast.Assign(
targets=[ast.Name(id="_codeflash_should_sync_tf", ctx=ast.Store())],
targets=[ast.Name(id="_codeflash_should_sync_tf", ctx=store_ctx)],
value=ast.Call(
func=ast.Name(id="hasattr", ctx=ast.Load()),
func=ast.Name(id="hasattr", ctx=load_ctx),
args=[
ast.Attribute(
value=ast.Attribute(
value=ast.Name(id=tf_alias, ctx=ast.Load()), attr="test", ctx=ast.Load()
),
value=ast.Attribute(value=ast.Name(id=tf_alias, ctx=load_ctx), attr="test", ctx=load_ctx),
attr="experimental",
ctx=ast.Load(),
ctx=load_ctx,
),
ast.Constant(value="sync_devices"),
],
Expand Down
Loading