This is an automated email from the ASF dual-hosted git repository.
tlopex pushed a commit to branch main
in repository https://gitbox.apache.org/repos/asf/tvm.git
The following commit(s) were added to refs/heads/main by this push:
new 1bc0ce8a23 [Fix][DLight] Decline GEMV reductions with only batch
spatial axes (#20985)
1bc0ce8a23 is described below
commit 1bc0ce8a237fbd892afe12bd3f43fc7885d37a84
Author: Julian Ng-Thow-Hing <[email protected]>
AuthorDate: Sun Nov 15 16:16:23 2026 -0500
[Fix][DLight] Decline GEMV reductions with only batch spatial axes (#20985)
Weighted centered reductions can look like GEMV when a broadcast operand
uses every spatial axis, leaving only batch axes and triggering an
assertion; decline those blocks so the reduction schedule handles them
instead.
The regression verifies rejection and fallback, all 15 GEMV tests pass
using the current pass wrapper with the retained compiler, and the
patched compiler builds ECAPA for WebGPU with eight HF parity cases
passing in Chrome.
---
python/tvm/s_tir/dlight/analysis/gemv.py | 6 +++--
tests/python/s_tir/dlight/test_gpu_gemv.py | 35 ++++++++++++++++++++++++++++++
2 files changed, 39 insertions(+), 2 deletions(-)
diff --git a/python/tvm/s_tir/dlight/analysis/gemv.py
b/python/tvm/s_tir/dlight/analysis/gemv.py
index 9c8477eeb6..4e1aa29bf0 100644
--- a/python/tvm/s_tir/dlight/analysis/gemv.py
+++ b/python/tvm/s_tir/dlight/analysis/gemv.py
@@ -158,8 +158,10 @@ def normalize(
else:
s_loops.append(loop)
- assert s_loops
- assert r_loops
+ # Broadcast operands can make a general reduction look like GEMV.
+ # GEMV needs a non-batch spatial axis as well as a reduction axis.
+ if not s_loops or not r_loops:
+ return None
if not c_loops:
c_loops = [sch.add_unit_loop(block_info.block_rv)]
if not batch_loops:
diff --git a/tests/python/s_tir/dlight/test_gpu_gemv.py
b/tests/python/s_tir/dlight/test_gpu_gemv.py
index 25d3aae3b3..00561edbe6 100644
--- a/tests/python/s_tir/dlight/test_gpu_gemv.py
+++ b/tests/python/s_tir/dlight/test_gpu_gemv.py
@@ -52,6 +52,41 @@ def test_gemv_rejects_composite_normalized_axis():
tvm_tirx.ir.assert_structural_equal(scheduled["main"], before)
+def test_gemv_rejects_all_batch_spatial_axes():
+ frames = T.dynamic("frames", "int64")
+
+ @Ts.function(private=True)
+ def before(
+ data: T.Tensor((T.int64(768), frames), "float32"),
+ mean: T.Tensor((T.int64(768),), "float32"),
+ weights: T.Tensor((T.int64(768), frames), "float32"),
+ result: T.Tensor((T.int64(768),), "float32"),
+ ):
+ for c, t in T.grid(T.int64(768), frames):
+ with Ts.sblock("weighted_variance"):
+ vc, vt = Ts.axis.remap("SR", [c, t])
+ with Ts.init():
+ result[vc] = T.float32(0)
+ result[vc] += weights[vc, vt] * (
+ (data[vc, vt] - mean[vc]) * (data[vc, vt] - mean[vc])
+ )
+
+ mod = tvm_tirx.IRModule({"main": before})
+ target = Target(
+ {
+ "kind": "webgpu",
+ "keys": ["webgpu", "gpu"],
+ "max_num_threads": 256,
+ "thread_warp_size": 1,
+ "max_shared_memory_per_block": 32768,
+ }
+ )
+ rejected = dl.ApplyDefaultSchedule(dl.gpu.GEMV(), target=target)(mod)
+ tvm_tirx.ir.assert_structural_equal(rejected["main"], before)
+ scheduled = dl.ApplyDefaultSchedule(dl.gpu.GEMV(), dl.gpu.Reduction(),
target=target)(mod)
+ assert scheduled["main"].attrs["tirx.is_scheduled"]
+
+
def test_gemv_basic():
# fmt: off
n = T.dynamic("n", "int32")