Skip to content

[AMD] Add check for cooperative groups support in driver #8935

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

Open
nithinsubbiah wants to merge 3 commits into
base: main
Choose a base branch
from
+10 −2
Open

[AMD] Add check for cooperative groups support in driver #8935

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
a9aa06a
[AMD] Launch hipModuleLaunchCooperativeKernel only for multiple ctas
nithinsubbiah Dec 8, 2025
bc52eb4
Add check for cooperative groups support
nithinsubbiah Dec 8, 2025
7f10bb8
Remove num_ctas check for launcing cooperative grid
nithinsubbiah Dec 8, 2025
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
5 changes: 3 additions & 2 deletions third_party/amd/backend/driver.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -309,12 +309,13 @@ static PyObject *getDeviceProperties(PyObject *self, PyObject *args) {

// create a struct to hold device properties
return Py_BuildValue(
"{s:i, s:i, s:i, s:i, s:i, s:i, s:s, s:i, s:i}", "max_shared_mem",
"{s:i, s:i, s:i, s:i, s:i, s:i, s:s, s:i, s:i, s:i}", "max_shared_mem",
props.sharedMemPerBlock, "max_num_regs", props.regsPerBlock,
"multiprocessor_count", props.multiProcessorCount, "sm_clock_rate",
props.clockRate, "mem_clock_rate", props.memoryClockRate, "mem_bus_width",
props.memoryBusWidth, "arch", props.gcnArchName, "warpSize",
props.warpSize, "max_threads_per_sm", props.maxThreadsPerMultiProcessor);
props.warpSize, "max_threads_per_sm", props.maxThreadsPerMultiProcessor,
"cooperativeLaunch", props.cooperativeLaunch);
}

static PyObject *loadBinary(PyObject *self, PyObject *args) {
Expand Down
7 changes: 7 additions & 0 deletions third_party/amd/backend/driver.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -809,6 +809,13 @@ def __init__(self, src, metadata):
mod = compile_module_from_src(src=src, name="__triton_launcher", include_dirs=include_dirs)
self.launch = wrap_handle_tensordesc(mod.launch, signature, tensordesc_meta)
self.launch_cooperative_grid = metadata.launch_cooperative_grid
# Check if cooperative groups are supported on the device.
if self.launch_cooperative_grid:
driver = triton.runtime.driver.active
assert isinstance(driver, HIPDriver)
device = driver.get_current_device()
device_properties = driver.utils.get_device_properties(device)
self.launch_cooperative_grid = device_properties['cooperativeLaunch']
Copy link
Collaborator

@antiagainst antiagainst Dec 11, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We should assert the device supports cooperative launch here instead of silently overwriting what the developer said. Rationale being that the developer indicates they want cooperative launch. That's likely due to some forward progress needs which demands certain number of blocks to co-launch. By silently overwriting that it can be a big suprise for developers to see hangs or incorrect results. In general, if reality doesn't match with what the developer explicitly said, better to be noisy and error out than "correct" it under the hood for developer.

self.profile_scratch_size = metadata.profile_scratch_size
self.profile_scratch_align = metadata.profile_scratch_align

Expand Down