Description
On XPU (device B60), all tests in test/inductor/test_snode_runtime.py fail with
the same error, on torch 2.14:
AssertionError: AssertionError not raised
Impacted cases
| testfile |
classname |
name |
| test/inductor/test_snode_runtime.py |
ComputeBoundedTests |
test_addmm |
| test/inductor/test_snode_runtime.py |
ComputeBoundedTests |
test_bmm |
| test/inductor/test_snode_runtime.py |
ComputeBoundedTests |
test_conv1d |
| test/inductor/test_snode_runtime.py |
ComputeBoundedTests |
test_conv2d |
| test/inductor/test_snode_runtime.py |
ComputeBoundedTests |
test_conv2d_transpose |
| test/inductor/test_snode_runtime.py |
ComputeBoundedTests |
test_conv3d |
| test/inductor/test_snode_runtime.py |
ComputeBoundedTests |
test_mm |
| test/inductor/test_snode_runtime.py |
MemoryBoundedTests |
test_dynamic |
| test/inductor/test_snode_runtime.py |
MemoryBoundedTests |
test_horizontal_reduction_pointwise |
| test/inductor/test_snode_runtime.py |
MemoryBoundedTests |
test_pointwise |
| test/inductor/test_snode_runtime.py |
MemoryBoundedTests |
test_relu |
Root cause (brief)
These tests assert assertNotZero(calculate_runtime(...)), where
calculate_runtime sums metrics.node_runtimes from the inductor scheduler's
runtime estimator. On XPU the estimate comes back as 0.0, so assertNotZero
(-> assertNotEqual(x, 0.0)) fails, surfacing as "AssertionError not raised".
The scheduler runtime estimate depends on per-device info (peak FLOPS and DRAM
GB/s). The XPU device-info support added in #189819 only populates data for PVC.
On B60 (Battlemage) the device is not covered, so FLOPS/bandwidth resolve to 0
and every node's estimated runtime is 0 -- failing both the compute-bound and
memory-bound tests.
Fix direction
Add B60 (and other non-PVC XPU devices) to the device-info table / GBps data-sheet
lookup so the runtime estimator returns nonzero FLOPS and DRAM bandwidth.
Environment
Description
On XPU (device B60), all tests in
test/inductor/test_snode_runtime.pyfail withthe same error, on torch 2.14:
Impacted cases
Root cause (brief)
These tests assert
assertNotZero(calculate_runtime(...)), wherecalculate_runtimesumsmetrics.node_runtimesfrom the inductor scheduler'sruntime estimator. On XPU the estimate comes back as 0.0, so
assertNotZero(->
assertNotEqual(x, 0.0)) fails, surfacing as "AssertionError not raised".The scheduler runtime estimate depends on per-device info (peak FLOPS and DRAM
GB/s). The XPU device-info support added in #189819 only populates data for PVC.
On B60 (Battlemage) the device is not covered, so FLOPS/bandwidth resolve to 0
and every node's estimated runtime is 0 -- failing both the compute-bound and
memory-bound tests.
Fix direction
Add B60 (and other non-PVC XPU devices) to the device-info table / GBps data-sheet
lookup so the runtime estimator returns nonzero FLOPS and DRAM bandwidth.
Environment