[CI]:Reduce test shapes to avoid OOM errors during CI.

Author: yyttt6

examples/blocksparse_attention/test_example_blocksparse_attention.py

@@ -24,11 +24,27 @@ def test_example_tilelang_sparse_gqa_decode_varlen_mask():
 
 
 def test_example_triton_sparse_gqa_decode_varlen_indice():
-    example_triton_sparse_gqa_decode_varlen_indice.main()
+    example_triton_sparse_gqa_decode_varlen_indice.main(
+        batch=16,
+        heads=16,
+        heads_kv=8,
+        max_cache_seqlen=4096,
+        dim=128,
+        dim_v=128,
+        sparse_ratio=0.8,
+        block_size=32)
 
 
 def test_example_triton_sparse_gqa_decode_varlen_mask():
-    example_triton_sparse_gqa_decode_varlen_mask.main()
+    example_triton_sparse_gqa_decode_varlen_mask.main(
+        batch=16,
+        heads=16,
+        heads_kv=8,
+        max_cache_seqlen=4096,
+        dim=128,
+        dim_v=128,
+        sparse_ratio=0.8,
+        block_size=32)
 
 
 if __name__ == "__main__":

examples/cast/example_group_per_split_token_cast_to_fp8.py

@@ -161,8 +161,7 @@ def ref_program(x: torch.Tensor, batch_sizes: torch.Tensor) -> \
     return x_fp8
 
 
-def main():
-    M, N, BG, blk_m = 8192, 8192, 2, 8
+def main(M=8192, N=8192, BG=2, blk_m=8):
     if dtype == "float":
         x = torch.randn(M, N, device="cuda", dtype=torch.float32)
     elif dtype == "float16":

examples/cast/example_per_token_cast_to_fp8.py

@@ -79,8 +79,7 @@ def ref_program(x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
     return x_fp8, (x_amax / 448.0).view(m, -1)
 
 
-def main():
-    M, N, blk_m = 8192, 8192, 8
+def main(M=8192, N=8192, blk_m=8):
     kernel = per_token_cast_to_fp8(M, N, blk_m)
     print(kernel.get_kernel_source())
     profiler = kernel.get_profiler(tensor_supply_type=tilelang.TensorSupplyType.Randn)

examples/cast/test_example_cast.py

@@ -4,11 +4,11 @@
 
 
 def test_example_group_per_split_token_cast_to_fp8():
-    example_group_per_split_token_cast_to_fp8.main()
+    example_group_per_split_token_cast_to_fp8.main(M=8192, N=2048, BG=2, blk_m=8)
 
 
 def test_example_per_token_cast_to_fp8():
-    example_per_token_cast_to_fp8.main()
+    example_per_token_cast_to_fp8.main(M=8192, N=2048, blk_m=8)
 
 
 if __name__ == "__main__":

examples/deepseek_v32/test_tilelang_example_deepseek_v32.py

@@ -13,7 +13,7 @@ def test_example_topk_selector():
 
 
 def test_example_fp8_lighting_indexer():
-    test_fp8_lighting_indexer()
+    test_fp8_lighting_indexer(S=1024, SKV=2048, H=32, HKV=1, D=64, kv_stride=1)
 
 
 @tilelang.testing.requires_cuda

examples/dynamic_shape/example_dynamic.py

@@ -96,8 +96,7 @@ def ref_program(A, B):
     print(f"Latency: {latency} ms")
 
 
-def main():
-    M, N, K = 16384, 16384, 16384
+def main(M=16384, N=16384, K=16384):
     block_M, block_N, block_K = 128, 128, 32
     trans_A, trans_B = False, False
     in_dtype, out_dtype = "float16", "float16"

examples/dynamic_shape/test_example_dynamic.py

@@ -3,7 +3,7 @@
 
 
 def test_example_dynamic():
-    example_dynamic.main()
+    example_dynamic.main(M=1024, N=1024, K=1024)
 
 
 if __name__ == "__main__":

examples/flash_attention/test_example_flash_attention.py

@@ -44,12 +44,14 @@ def test_example_mha_bwd_wgmma_pipelined():
 @tilelang.testing.requires_cuda
 @tilelang.testing.requires_cuda_compute_version_ge(9, 0)
 def test_example_gqa_fwd_bshd_wgmma_pipelined():
-    example_gqa_fwd_bshd_wgmma_pipelined.main()
+    example_gqa_fwd_bshd_wgmma_pipelined.main(
+        batch=1, heads=16, seq_len=1024, dim=128, is_causal=False, groups=16, tune=False)
 
 
 @tilelang.testing.requires_cuda
 def test_example_gqa_fwd_bshd():
-    example_gqa_fwd_bshd.main()
+    example_gqa_fwd_bshd.main(
+        batch=1, heads=16, seq_len=1024, dim=128, is_causal=False, groups=16, tune=False)
 
 
 @tilelang.testing.requires_cuda

testing/python/issue/test_tilelang_issue_96.py

@@ -53,7 +53,7 @@ def run_gemm_pipeline_test(N, block_M=128, block_N=128, block_K=32):
 
 def test_pipeline_large_matrix():
     """Test pipeline stages with large matrix multiplication (8192x8192)"""
-    run_gemm_pipeline_test(8192)
+    run_gemm_pipeline_test(4096)
 
 
 def test_pipeline_small_matrix():