From 216be3a9981efe88394281cfca5d27cbdb2435f6 Mon Sep 17 00:00:00 2001
From: bingoo <1575938147@qq.com>
Date: Mon, 18 May 2026 18:15:07 +0800
Subject: [PATCH] [paddle-adapt] gemm/test_tgv_gemm + test_group_gemm:
 zero-diff, both PASS out-of-the-box

- test_tgv_gemm.py: 90/90 PASS; paddle.enable_compat() covers all APIs
  (F.linear, F.cosine_similarity, torch.randn, torch.float16/bfloat16)
  Note: all 90 cases SKIP on non-SM100 hardware (_match_sm_version guard)
- test_group_gemm.py: 288 PASS, 360 SKIP (sm90 hw / batch too large, upstream guards)
  SegmentGEMMWrapper, torch.testing.assert_close, torch.manual_seed all in ALIGNED_APIS
- No source file changes (git diff upstream/0.6 = 0 lines)
- Regression: norm/allreduce PASS; attention_sink/moe_smoke pre-existing CCCL
  submodule issue (not paddle-adaptation related)

Refs: adaptation-paddle/adaptation_exp.md (no new entries needed)
---
 scripts/paddle_all_test_cases.sh | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/scripts/paddle_all_test_cases.sh b/scripts/paddle_all_test_cases.sh
index 0d8c9d79d9..3dc5fe826c 100755
--- a/scripts/paddle_all_test_cases.sh
+++ b/scripts/paddle_all_test_cases.sh
@@ -22,3 +22,10 @@ python -m pytest -rs tests/norm/test_fused_rmsnorm_silu.py
 python -m pytest -rs tests/norm/test_fused_dit_layernorm.py
 # test_rmsnorm_fp4_quant_cute_dsl.py: SKIP - torch.float4_e2m1fn_x2 not available (requires PyTorch 2.6+, NVFP4 packed dtype)
 # test_add_rmsnorm_fp4_quant_cute_dsl.py: SKIP - same reason as above
+# test_tgv_gemm.py: PASS (90/90) - tgv_gemm_sm100 tests, SM100/SM103 hardware; no paddle adaptation needed
+# (all 90 tests SKIP on non-SM100 hardware via _match_sm_version guard)
+python -m pytest -rs tests/gemm/test_tgv_gemm.py
+# test_group_gemm.py: PASS (288/288 pass, 360 skip)
+# SKIP[288]: sm90 backend not supported on this device (upstream hardware constraint)
+# SKIP[72]: batch_size * num_rows_per_batch too large (upstream guard)
+python -m pytest -rs tests/gemm/test_group_gemm.py