Enable Float4_e2m1fn_x2 to Float4_e2m1fn_x2 copy #2310
Conversation
There was a problem hiding this comment.
Pull Request Overview
This PR enables copying tensors with the Float4_e2m1fn_x2 data type on XPU devices by adding kernel support and extending test coverage. This addresses issue #2305 by implementing the missing copy operation for this float4 type.
Key changes:
- Added
float4_copy_kernel_xpufunction to handle Float4_e2m1fn_x2 copy operations - Extended test coverage to include Float4_e2m1fn_x2 dtype in copy/clone tests
Reviewed Changes
Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.
| File | Description |
|---|---|
| src/ATen/native/xpu/sycl/CopyKernel.cpp | Implements float4 copy kernel and integrates it into the main copy_kernel dispatch logic |
| test/regressions/test_copy.py | Adds Float4_e2m1fn_x2 to the test dtypes for copy and clone operations |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| void float4_copy_kernel_xpu(TensorIteratorBase& iter) { | ||
| ScalarType src_dtype = iter.dtype(1); | ||
|
|
||
| if (src_dtype == kFloat4_e2m1fn_x2) { | ||
| gpu_kernel_nocast(iter, CopyScalarFunc<Float4_e2m1fn_x2>()); | ||
| } else { | ||
| TORCH_CHECK(false, "Copy from ", src_dtype, " to Float4_e2m1fn_x2 has not been supported."); | ||
| } | ||
| } |
There was a problem hiding this comment.
The implementation only supports Float4_e2m1fn_x2 to Float4_e2m1fn_x2 copy, but the pattern in float8_copy_kernel_xpu shows support for casting from common types (float, Half, BFloat16). Consider adding similar casting support for Float4_e2m1fn_x2 to match the capabilities of the float8 implementation and provide a more complete copy operation.
CuiYifeng
force-pushed
the
yifeng/fp4_copy
branch
from
24cd909 to
48cd4bb
Compare
To solve #2305.
This PR adds support for copying tensors with the
Float4_e2m1fn_x2data type on XPU devices.