WOQ: fix reference kernel for binary fusion with odd M (#3490)

Author: Xia-Weiwen

csrc/cpu/aten/utils/woq.h

@@ -3454,11 +3454,21 @@ static at::Tensor woq_gemm_ref_impl(
     at::silu_(y);
   } else if (fusion_type == WOQ_FUSE_ADD || fusion_type == WOQ_FUSE_ADD_ADD) {
     for (auto& tin : others_list) {
-      y = at::add(y, tin.view(y.sizes()));
+      auto tin_view = tin.view({-1, y.size(-1)});
+      if (tin_view.size(0) < y.size(0)) {
+        tin_view = at::pad(
+            tin_view, {0, 0, 0, y.size(0) - tin_view.size(0)}, "constant", 0);
+      }
+      y = at::add(y, tin_view);
     }
   } else if (fusion_type == WOQ_FUSE_MUL) {
     for (auto& tin : others_list) {
-      y = at::mul(y, tin.view(y.sizes()));
+      auto tin_view = tin.view({-1, y.size(-1)});
+      if (tin_view.size(0) < y.size(0)) {
+        tin_view = at::pad(
+            tin_view, {0, 0, 0, y.size(0) - tin_view.size(0)}, "constant", 0);
+      }
+      y = at::mul(y, tin_view);
     }
   } else {
     TORCH_CHECK(

tests/cpu/test_quantization_default_recipe.py

@@ -1207,7 +1207,7 @@ def forward(self, x, others):
         ]
         bias_list = [False, True]
         bf16_list = [False, True]
-        batch_size_list = [4, 1024]
+        batch_size_list = [4, 1024, 63]
         cases = itertools.product(
             weight_dtype_list, bias_list, bf16_list, batch_size_list
         )