[TRTLLM-8777][feat] Update DeepGEMM to the latest commit to include optimizations for DeepSeek-v3.2 (#9380)

Signed-off-by: Fanrong Li <23290157+lfr-0531@users.noreply.github.com>

Author: lfr-0531

3rdparty/CMakeLists.txt

@@ -39,7 +39,7 @@ FetchContent_Declare(
 FetchContent_Declare(
   deepgemm
   GIT_REPOSITORY https://github.com/ruoqianguo/DeepGEMM
-  GIT_TAG 9fa5965e265e27995f539e0dd73a06351a8a9eaf
+  GIT_TAG 6cb8161516302550785d9af924d2778afef1f3f6 # swapab_sm100 branch
   GIT_SUBMODULES_RECURSE
   ON
   SOURCE_SUBDIR

tests/unittest/_torch/attention/sparse/test_dsa_indexer.py

@@ -308,9 +308,9 @@ def test_deepgemm_fp8_mqa_logits_basic():
     """
     torch.manual_seed(0)
 
-    num_heads, head_dim = 32, 128
-    seq_len = 512
-    seq_len_kv = 1024
+    num_heads, head_dim = 64, 128
+    seq_len = 2048
+    seq_len_kv = 4096
     #[seq_len, num_heads, head_dim]
     q = torch.randn(
         seq_len,
@@ -335,8 +335,8 @@ def test_deepgemm_fp8_mqa_logits_basic():
     )
     # ks[i] -> ke[i] for each q[i]
     ks = torch.zeros(seq_len, dtype=torch.int, device="cuda")
-    ke = torch.arange(seq_len, dtype=torch.int, device="cuda") + (
-        seq_len_kv - seq_len) + 1  # +1 for exclusive end
+    ke = torch.arange(seq_len, dtype=torch.int,
+                      device="cuda") + (seq_len_kv - seq_len)
 
     # Convert to FP8
     q_fp8 = q.to(torch.float8_e4m3fn)